{"responseHeader":{"status":0,"QTime":15,"params":{"q":"{!q.op=AND}id:\"196231\"","hl":"true","hl.simple.post":"","hl.fragsize":"5000","hl.fl":"ocr_t","hl.method":"unified","wt":"json","hl.simple.pre":""}},"response":{"numFound":1,"start":0,"docs":[{"ark_t":"ark:/87278/s6m93ht9","setname_s":"ir_etd","subject_t":"Device-free localization; Hidden Markov model; Radio tomographic imaging; Range estimation; Received signal strength; Ultra-wideband","restricted_i":0,"department_t":"Electrical & Computer Engineering","format_medium_t":"application/pdf","identifier_t":"etd3/id/2656","date_t":"2013-12","mass_i":1515011812,"publisher_t":"University of Utah","description_t":"This work seeks to improve upon existing methods for device-free localization (DFL) using radio frequency (RF) sensor networks. Device-free localization is the process of determining the location of a target object, typically a person, without the need for a device to be with the object to aid in localization. An RF sensor network measures changes to radio propagation caused by the presence of a person to locate that person. We show how existing methods which use either wideband or narrowband RF channels can be improved in ways including localization accuracy, energy efficiency, and system cost. We also show how wideband and narrowband systems can combine their information to improve localization. A common assumption in ultra-wideband research is that to estimate the bistatic delay or range, \"background subtraction\" is effective at removing clutter and must first be performed. Another assumption commonly made is that after background subtraction, each individual multipath component caused by a person's presence can be distinguished perfectly. We show that these assumptions are often not true and that ranging can still be performed even when these assumptions are not true. We propose modeling the difference between a current set of channel impulse responses (CIR) and a set of calibration CIRs as a hidden Markov model (HMM) and show the effectiveness of this model over background subtraction. The methods for performing device-free localization by using ultra-wideband (UWB) measurements and by using received signal strength (RSS) measurements are often considered separate topic of research and viewed only in isolation by two different communities of researchers. We consider both of these methods together and propose methods for combining the information obtained from UWB and RSS measurements. We show that using both methods in conjunction is more effective than either method on its own, especially in a setting where radio placement is constrained. It has been shown that for RSS-based DFL, measuring on multiple channels improves localization accuracy. We consider the trade-o s of measuring all radio links on all channels and the energy and latency expense of making the additional measurements required when sampling multiple channels. We also show the benefits of allowing multiple radios to transmit simultaneously, or in parallel, to better measure the available radio links.","rights_management_t":"Copyright © Merrick K. McCracken 2013","title_t":"Utilization of narrowband and wideband radio frequency measurements for device-free localization","id":196231,"publication_type_t":"dissertation","parent_i":0,"type_t":"Text","dissertation_institution_t":"University of Utah","thumb_s":"/e6/37/e63757c220df84c6066e0a3d5efee54a1577a27c.jpg","oldid_t":"etd3 2656","author_t":"McCracken, Merrick K.","metadata_cataloger_t":"CLR","format_t":"application/pdf","modified_tdt":"2017-11-07T18:32:09Z","dissertation_name_t":"Doctor of Philosophy","school_or_college_t":"College of Engineering","language_t":"eng","file_s":"/02/38/02386716b9422a515e3fc982f558f57b5679a7ce.pdf","format_extent_t":"1,415,303 bytes","created_tdt":"2014-01-14T00:00:00Z","permissions_reference_url_t":"https://collections.lib.utah.edu/details?id=1278879","doi_t":"doi:10.26053/0H-Z830-4VG0","_version_":1619873475386671104,"ocr_t":"UTILIZATION OF NARROWBAND AND WIDEBAND RADIO FREQUENCY MEASUREMENTS FOR DEVICE-FREE LOCALIZATION by Merrick K. McCracken A dissertation submitted to the faculty of The University of Utah in partial ful llment of the requirements for the degree of Doctor of Philosophy Department of Electrical and Computer Engineering The University of Utah December 2013 Copyright c Merrick K. McCracken 2013 All Rights Reserved Th e Uni v e r s i t y o f Ut a h Gr a dua t e S cho o l STATEMENT OF DISSERTATION APPROVAL The dissertation of Merrick K. McCracken has been approved by the following supervisory committee members: Neal Patwari , Chair 10/31/2013 Date Approved Rong Rong Chen , Member 10/31/2013 Date Approved Behrouz Farhang , Member 10/31/2013 Date Approved Thomas Schmid , Member 10/31/2013 Date Approved Suresh Venkatasubramanian , Member 10/31/2013 Date Approved and by Gianluca Lazzi , Chair/Dean of the Department/College/School of Electrical and Computer Engineering and by David B. Kieda, Dean of The Graduate School. ABSTRACT This work seeks to improve upon existing methods for device-free localization (DFL) using radio frequency (RF) sensor networks. Device-free localization is the process of determining the location of a target object, typically a person, without the need for a device to be with the object to aid in localization. An RF sensor network measures changes to radio propagation caused by the presence of a person to locate that person. We show how existing methods which use either wideband or narrowband RF channels can be improved in ways including localization accuracy, energy e ciency, and system cost. We also show how wideband and narrowband systems can combine their information to improve localization. A common assumption in ultra-wideband research is that to estimate the bistatic delay or range, \\background subtraction\" is e ective at removing clutter and must rst be performed. Another assumption commonly made is that after background subtraction, each individual multipath component caused by a person's presence can be distinguished perfectly. We show that these assumptions are often not true and that ranging can still be performed even when these assumptions are not true. We propose modeling the di erence between a current set of channel impulse responses (CIR) and a set of calibration CIRs as a hidden Markov model (HMM) and show the e ectiveness of this model over background subtraction. The methods for performing device-free localization by using ultra-wideband (UWB) measurements and by using received signal strength (RSS) measurements are often consid- ered separate topic of research and viewed only in isolation by two di erent communities of researchers. We consider both of these methods together and propose methods for combining the information obtained from UWB and RSS measurements. We show that using both methods in conjunction is more e ective than either method on its own, especially in a setting where radio placement is constrained. It has been shown that for RSS-based DFL, measuring on multiple channels improves localization accuracy. We consider the trade-o s of measuring all radio links on all channels and the energy and latency expense of making the additional measurements required when sampling multiple channels. We also show the bene ts of allowing multiple radios to transmit simultaneously, or in parallel, to better measure the available radio links. CONTENTS ABSTRACT : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : iii LIST OF TABLES: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : vii CHAPTERS 1. INTRODUCTION : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 1 1.1 Radio Frequency Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1.1 Wideband E ects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.1.2 Narrowband E ects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 Radio Tomographic Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2. ESTIMATION OF BISTATIC RANGE FROM CLUTTERED ULTRA- WIDEBAND IMPULSE RESPONSES: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 9 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2.1 Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2.2 Quanti cation of Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.2.3 CIR Changes as a Hidden Markov Model . . . . . . . . . . . . . . . . . . . . . . . . 16 2.2.4 Continuous Observation Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.2.5 HMM Solving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2.6 First Threshold Crossing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.2.7 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.2.8 Proposed Localization Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.3 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.3.1 In-room Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.3.2 Through-wall Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.4.1 First Threshold Crossing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.4.2 HMM-based Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.4.3 Through-wall experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.4.4 False Positives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.4.5 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 3. JOINT ULTRA-WIDEBAND AND SIGNAL STRENGTH-BASED THROUGH- BUILDING TRACKING FOR TACTICAL OPERATIONS : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 39 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.2 Radio Tomographic Imaging (RTI) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.2.1 Attenuation-based RTI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.2.2 Variance-based RTI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 3.2.3 RTI Image Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 3.3 Ultra-wide Band Range Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.3.1 Changes to the CIR as a Hidden Markov Model . . . . . . . . . . . . . . . . . . 46 3.3.2 Variance-based UWB Range Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.3.3 UWB Image Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.4 Combining RTI and UWB Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.4.1 Image Combination by Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.4.2 Linear Inversion with UWB Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.4.3 Estimating X by Using Y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.5 Localization and Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.6 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.7.1 Study Room . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.7.2 Hotel Room . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.7.3 Area of Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4. PARALLEL MULTICHANNEL TRANSMISSION FOR RSS-BASED RADIO TOMOGRAPHY : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 62 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 4.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 4.3 Device-free Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 4.3.1 Sampling Schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 4.4.1 Scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 4.4.2 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 4.4.3 Solution Space Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 4.4.4 Choosing Transmitters and Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 4.4.5 Alternative Score Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 4.4.6 Receiver Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 4.4.7 Simple Alternative Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 4.4.8 Radio Tomographic Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 4.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 4.6.1 Fade Level Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 4.6.2 Channel Diversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 4.6.3 Energy Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 v 5. CONCLUSIONS : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 88 5.1 Engineering Trade-o s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 5.2 Future Areas of Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 REFERENCES : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 93 vi LIST OF TABLES 2.1 RMS Localization Error (cm) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.2 Median Localization Error (cm) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.1 RTI image estimation parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 3.2 UWB estimation parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 3.3 Mean RMS localization error for the second experiment over all 18 trials for the methods described. Gating was used for all methods except random selection. Units given in meters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 3.4 Percent reduction of AoU by including UWB data. . . . . . . . . . . . . . . . . . . . . . 61 4.1 Error of each method for all experiments given as the RMS `2 error in units of cm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 5.1 A comparison of a typical UWB localization system and a typical RSS-based localization system. Accuracy is given in units of cm. Energy is given in units of Watts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 CHAPTER 1 INTRODUCTION Device-free localization (DFL) is the process of determining the location of a target object, typically a person, without the need for a device to be with the object to aid in localization. Knowing where a person or multiple people are located in a room or building is valuable information for building security. The information provided by device-free localization can also be used to help caretakers of the elderly by knowing their movements within their home and their inactivity as well, which could signal their need for help. In either of these scenarios, it is very impractical to require the people to carry with them a device that would be used to aid localization. Localization is performed by employing the information provided by sensors placed such that the the object or person can be located in a target area or region. The sensors detect and quantify changes in the environment caused by the person's presence or movement. By knowing or modeling how a person may change sensor data as a function of the person's location, an estimate of the person's location can be made using the available sensor data. Device-free localization can be done with a number of di erent sensing techniques including cameras and image processing [1], reading radio-frequency identi cation (RFID) tags [2], and wideband [3] and narrowband [4]{[7] radio frequency (RF) sensing as will be discussed in Sections 1.1.1 and 1.1.2, respectively. 1.1 Radio Frequency Sensing The human body a ects radio propagation, both in narrowband and in wideband trans- missions. For a radio signal there exist multiple transmission paths, or multipath, between radio transmitters and receivers. The presence of a person can both cause new multipath to form, from re ection or refraction, and block or attenuate existing multipath. The presence of a person, therefore, causes a change in the signal arriving at the receiving radio compared to the signal that would have arrived had the person not been there. 2 1.1.1 Wideband E ects For wideband or ultra-wideband (UWB) transmissions, a time-domain representation of the arriving signal is often available from the receiving UWB radio. Assume that an UWB transmitter sends pulse (t). Due to multipath propagation, the received signal is described by h(t) = X i i (t i); (1.1) where i and i are the complex amplitude and time delay of the ith path, respectively. The value 0 corresponds to the UWB impulse following the line-of-sight path between the two radios, which are separated from each other by a distance on the order of meters or more. As the time progresses, more UWB impulses arrive at the receiver that re ected o of objects and scatterers in the environment. These impulses traveled paths longer than the line-of-sight path, each with its corresponding i. The receiver radio approximately measures the channel impulse response convolved with the pulse shape. Fig. 1.1 is an example of how the transmitted pulse may follow many di erent paths to arrive at the receiver. The number of multipath components seen by the receiver depends on the environment around the radios. The presence of a person will have two primary e ects. The rst is that the person's body will cause a new multipath impulse to arrive at the receiver. When a person appears at point x0 in the environment with the transmitter at xt and receiver at xr, he causes an additional path with path length kxt x0k + kx0 xrk, where k k is the L2 norm. This is true if we assume the impulse will re ect or bounce only o of the person and no other object while traveling from the transmitter to the receiver. This path length is greater than the line-of-sight path length. The di erence in length between these two paths is the excess path length. The excess path length of the multipath component caused by the person can be used to estimate the sum of the distances to each radio from the person's location. The sum of these distances together is known as the bistatic range. The delay associated with this new multipath component is , which we refer to as the bistatic delay. The second e ect the person has is that it will a ect existing multipath. As illustrated in Fig. 1.2, the presence of a person will a ect multipath components that the receiver would have measured had the person not been there. The person's body can a ect any multipath component whose path 3 length is greater than or equal to that path length of the person's multipath component, that is, the person a ects many i for i . 1.1.2 Narrowband E ects Narrowband radio signals are similarly a ected by the presence of a person's body. Unlike UWB impulse measurements, however, a time-domain representation of the arriving signal is unavailable. For packet-based radio protocols, such as 802.15.4, a link quality measurement is often provided by the receiving radio for each packet reception. This measurement is often a received signal strength (RSS) measurement. This RSS measurement can be used to determine the channel fading, or fade level, or the wireless communication channel. Similar to UWB measurements, received narrowband signals can be modeled as H = X i iejfc i : (1.2) This is similar to equation (1.1) but considered in the frequency domain, where (t i) becomes ejfc i in the frequency domain. This is done because the time required to transmit the narrowband signals is much greater than individual multipath delays, which is unlike UWB measurements. Therefore, we can consider all multipath as if they are arriving simultaneously. The RSS measurements are the received power and are proportional to jHj2. RSS is measured in units of dBm. These arriving multipath can interfere constructively or destructively. The level of constructive or destructive interference is know as fading or is describe by the fade level of the signal. This interference can be illustrated in the following way. In Figs. 1.3 and 1.4, the individual complex amplitudes of each arriving multipath component are shown on the complex plane. In both gures, the black arrows represent the complex amplitudes of individual multipath components and the red arrow represents the sum of the individual components. In Fig. 1.3, the components interfere constructively. A link with a received signal such as this would be considered an antifade link, or a link where there is little or no fading. The received signal is strong, and e ects to the link's multipath components generally cause the RSS value to decrease. In Fig. 1.4, the components interfere destructively. This signal would be considered a deep-fade link, or a link where the fading is strong. Changes to the multipath components may cause the RSS value to increase or decrease in value. It is because RSS is measured in a log scale that there is a higher variability is RSS measurements for deep-fade links than for those in antifade links. 4 Generally, when a person a ects one or more multipath components of an antifade link, the overall signal is attenuated. For a deep-fade link, however, it becomes more likely that an a ected multipath could result in constructive interference and an increase in RSS. The changes in RSS values relative to their fade level, which is measured during a calibration period, can be used to determine where a person is within the environment. This is done by employing a model of how the person a ects a link's RSS measurements depending on where the person is and what the fade level of the link is. This e ect has been modeled in a number of ways [5], [8]{[10]. 1.2 Radio Tomographic Imaging Radio tomographic imaging (RTI) is the process of estimating the changes to the RF propagation eld. These changes are estimated over a discretized target area, which is the area where the radio sensors can e ectively monitor the changes caused by the presence of a person. Each discretized partition of the target area corresponds to one pixel in the localization image. Image maxima indicate possible person locations. Critical to performing RTI is a model of how a person a ects radio propagation as a function of the person's location relative to the link. For narrowband systems, the measured changes to RSS values serve as a measure of changes in radio propagation. A number of di erent models of how a person a ects radio propagation have been proposed based on changes in attenuation [5], signal variance [8], and based on both the change in attenuation and the direction of change [9]. Instead of applying a model of how a link is a ected by the presence of a person, an alternative method for estimating a person's location is by using ngerprint-based methods [4], [11]. These methods require additional training data beyond a calibration measurement of the empty environment. For these methods, training measurements are made for many discrete locations where the person may be by having a person stand in each of these locations while measurements are made. Then when a person's location estimate is to be made, the current measurements are compared to the set of training data. The training data that most closely resemble the measured data give their corresponding training location as the estimated location of the person. These methods require much more training data, which grow linearly with the number of discrete training locations and exponentially with the number of possible people in the target environment. New training data may also be necessary if the environment changes signi cantly. For ultra-wideband systems, changes in the radio propagation eld for RTI are modeled by changes to the received, time-domain, channel impulse responses (CIR). This information 5 is used to estimate the bistatic range of people or objects a ecting the CIR, producing iso-range contours around the UWB radios. These contours, shaped as ellipses, are estimates of where changes occur in the radio propagation eld due to the presence of a person. A number of ways to perform localization using UWB measurements is further discussed in Section 2.2.7. 1.3 Contributions A common assumption in ultra-wideband research is that to estimate the bistatic delay or range, \\background subtraction\" must rst be performed. This means that a prior measurement of the CIR is subtracted from any current CIR measurement. These prior mea- surements are presumed to be made when the area is empty (i.e., with a static background). Some work in UWB-based impulse response radar assumes that background subtraction is completely e ective in removing the response due to the static environment [12]{[15]. Another assumption commonly made is that after background subtraction, each individual multipath component caused by a person's presence can be distinguished perfectly from the impulses caused by other people and the environment [13], [14]. In Chapter 2, we show that these assumptions are often not true and that ranging can still be performed even when these assumptions are not true. The methods for performing device-free localization by using UWB measurements and by using RSS measurements are often considered a separate topic of research and viewed only in isolation by two di erent communities of researchers. In Chapter 3, we consider both of these methods together and propose methods for combining the information obtained from UWB and RSS measurements. We show that using both methods in conjunction is more e ective than either method on its own, especially in a setting where radio placement is constrained. It has been shown that for RSS based DFL, measuring on multiple channels improves localization accuracy [9], [10]. In Chapter 4, we consider the trade-o s of measuring all radio links on all channels and the energy and latency expense of making the additional measurements required when sampling multiple channels. This work can be used to allow for large-scale deployed systems of RF sensors to use multiple channels simultaneously. The following publications are a result of this work. M. McCracken and N. Patwari, \\Hidden Markov Estimation of Bistatic Range from Cluttered Ultra-wideband Impulse Responses,\" in 2nd IEEE Topical Meeting on Wireless Sensors and Sensor Networks (WiSNet 2012), 2012, pp. 17-20. [16] 6 M. McCracken and N. Patwari, \\Hidden Markov Estimation of Bistatic Range from Cluttered Ultra-wideband Impulse Responses,\" IEEE Transactions on Mobile Com- puting, vol. PP, no. 99, pp. 1-1, 2013. [3] M. McCracken, M. Bocca, and N. Patwari, \\Joint Ultra-wideband and Signal Strength- based Through-building Tracking for Tactical Operations,\" in 2013 IEEE Interna- tional Conference on Sensing, Communications and Networking (SECON), 2013, pp. 309-317. [17] M. McCracken, M. Bocca, and N. Patwari. \\Selection of Links in Multichannel RSS Measurements for Radio Tomography\", (to be submitted to arxiv.org and to be included in \\Large Scale, Device Free Localization\" by the A. Luong, M. McCracken, M. Bocca, and N. Patwari to be submitted to 2014 IEEE International Conference on Sensing, Communications and Networking (SECON)). 7 Figure 1.1. The transmitted radio signal takes many paths to arrive at the receiving radio due to the environment. Figure 1.2. The presence of a person in the environment both creates new multipath signals and a ects existing multipath signals. How the signal is a ected depends on the signal arriving at the receiver without the presence of a person and on where the person is within the environment. 8 Figure 1.3. Multipath components arrive at the receiver with amplitudes and phase o sets such that the overall received signal strength is high due to constructive interference. These are antifade links. Magnitude contours are drawn on a log scale. Figure 1.4. Multipath components arrive at the receiver with amplitudes and phase o sets such that the overall received signal strength is very low due to destructive interference. These are deep-fade links. Magnitude contours are drawn on a log scale. CHAPTER 2 ESTIMATION OF BISTATIC RANGE FROM CLUTTERED ULTRA- WIDEBAND IMPULSE RESPONSES UWB multistatic radar can be used for target detection and tracking in buildings and rooms. Target detection and tracking relies on accurate knowledge of the bistatic delay. Noise, measurement error, and the problem of dense, overlapping multipath signals in the measured UWB CIR all contribute to make bistatic delay estimation challenging. It is often assumed that a calibration CIR, that is, a measurement from when no person is present, is easily subtracted from a newly captured CIR. We show this is often not the case. We propose modeling the di erence between a current set of CIRs and a set of calibration CIRs as a hidden Markov model (HMM). Multiple experimental deployments are performed to collect CIR data and test the performance of this model and compare its performance to existing methods. Our experimental results show an RMSE of 2.85 ns and 2.76 ns for our HMM-based approach, compared to a thresholding method which, if the ideal threshold is known a priori, achieves 3.28 ns and 4.58 ns. By using the Baum{Welch algorithm, the HMM-based estimator is shown to be very robust to initial parameter settings. Localization performance is also improved using the HMM-based bistatic delay estimates.1 2.1 Introduction A useful application of UWB impulse radio is detection and tracking of people2 in buildings. In particular, bistatic and multistatic radar systems are used for this application 1 c 2013 IEEE. Reprinted, with permission, from M. McCracken and N. Patwari, \\Hidden Markov Estimation of Bistatic Range from Cluttered Ultra-wideband Impulse Responses,\" IEEE Transactions on Mobile Computing, vol. PP, no. 99, pp. 1-1, 2013. 2In this paper, we use \\people\" or \\person\" to indicate the object being tracked. 10 [18]. This is done by capturing the CIR, h(t), between transmitter/receiver pairs and detecting changes to the CIR. This paper describes a contribution to bistatic delay (or equivalently, bistatic range) estimation. A person induces changes in the CIR starting at the bistatic delay, that is, the earliest time delay at which changes occur in the CIR due to the person being tracked. If the bistatic delay is denoted , then the bistatic range is simply the distance this multipath component has traveled (i.e., c where c is the speed of light). If RF energy traveled from the transmitter to the person and then to the receiver, with no additional scattering, then the bistatic range de nes an ellipse on which the person is located. Thus bistatic range estimation is a key primitive of UWB tracking systems. The primary contribution of this work is to develop a method which considers the changes which occur in a CIR at all time delays in order to estimate bistatic delay. Current published research, as described in Section 2.1.1, generally are rst threshold-crossing methods, that is, they estimate the bistatic delay as the rst delay in which a metric exceeds a threshold. As a result, they are (a) sensitive to noise in the CIR prior to the true bistatic delay, and (b) sensitive to the correct setting of the threshold parameter. Our proposed method uses a HMM to model the changes to the CIR as a function of time delay. The Markov chain is a progression between two states: X = 0, meaning that a person in the environment is not causing changes at the current time delay, or X = 1, meaning that a person is causing changes at the current time delay. The state of the system is observable only indirectly via the CIR because of noise and the variability in the multipath channel. The distribution of the observations is dependent on the current state of the system, thus the system is a HMM. Using the observations and the system model, the forward-backward algorithm solves for the most likely state at any given time. The bistatic delay estimate is the time delay at which the system transitions from state 0 to state 1. When solving for the bistatic delay, our proposed method considers all of the available data and, as we show, the error in bistatic delay estimation is reduced compared to the best thresholding scheme. Further, using a Baum{Welch algorithm, we avoid the requirement of knowing a priori the correct parameters. 2.1.1 Related Work Generally, methods to estimate the bistatic delay or range rst perform \\background subtraction.\" This means that a prior measurement, or an average of many prior measure- ments, of the CIR is subtracted from any current CIR measurement. These prior measure- ments are presumed to be made when the area is empty (i.e., with a static background). 11 Some work in UWB-based impulse response radar assumes that background subtraction is completely e ective in removing the response due to the static environment [12]{[15]. Some work additionally assumes that, after background subtraction, each single multipath component caused by a person's presence can be distinguished perfectly from the impulses caused by other people and the environment [13], [14]. In this paper, we show that ranging can still be performed when these assumptions are not true, as is often the case in a cluttered multipath environment. One way to estimate the bistatic delay is rst to perform \\background subtraction,\" and then to threshold on the amplitude of the di erence. Zetik et al. [15] describe a thresholding method that uses a simple formula for choosing an appropriate threshold value for accurate range estimation after background subtraction has been performed. Each UWB module has one transmitting and two receiving directional antennas, all relatively close to one another. This makes each UWB module approach a monostatic radar con guration. All of the sensor nodes were pointed inward toward an empty room using directional horn antennas for their experiments. In contrast, our measurements are performed in furnished o ce environments, and the additional clutter can make background subtraction less e ective. The estimation methods described in [15] will be used in this work for comparison. Another way to estimate the delay is to perform a cross correlation of the received signal with a known target scattering pro le and then to threshold the correlation values. Chang et al. approach detection by modeling a human body's scattering as a spectral multipath model and cross correlating this model with the received CIRs [19], [20]. Detection is then performed using an adaptive threshold on the cross correlation. In their work they used a UWB radio similar to those used in this work but in a monostatic radar con guration. The human body spectral multipath model was obtained using empirically collected data from their UWB radio. They collected data of a moving human subject in an open eld where there was little or no multipath propagation to validate their detection method [19]. They expanded the method to tracking a human target and tested it using additional data collected from the UWB radio [20]. The experimental data for tracking was also collected in an open eld. In contrast, we use measured data from cluttered environments to show that our method is robust to the indoor multipath channel. The work done by Giorgetti and Chiani o ers a method of performing time-of-arrival estimation in UWB signals without performing thresholding on the signal [21]. They describe a \\non-linear excision lter\" to remove binned data that is purely noise to leave only noise-plus-data bins. They use information theoretic criteria to arrive at an estimator. 12 They o er simulation results using the 802.15.4a channel models. It is unknown how these models change with the presence of a person. The tests performed in this work depend on the presence of a person to a ect the channel. Our work is not the rst to propose using HMMs for tracking, however, it is the rst, to our knowledge, to propose using a HMM for UWB impulse radar bistatic delay estimation. Nijsure et al. used a HMM to model movement in a UWB radar-based tracking system and simulated its performance [22]. In their work, the states of the model are nonoverlapping geographic regions near the radios rather than changes to the received signal. The measurements in [22] are unambiguous power delay pro les. In contrast, our HMM is used to estimate the bistatic range, with only two states, whether or not the CIR is impacted by a person at a given time delay or not. Two-state HMMs have been used in other applications, for example in detecting channel use in dynamic spectrum access [23]. The work in [23] simulated channel access by primary users and the performance of detection by secondary users, who would use the channel opportunistically, using a HMM-based estimator to detect whether a primary user is currently transmitting. Simulations showed improved detection performance for the HMM-based method compared to a threshold-based method. 2.1.2 Organization This paper is organized as follows. Section 2.2 describes the methods proposed in this work to estimate using hidden Markov models. Section 2.3 describes the data collection campaigns carried out to test the proposed methods empirically. Results for our proposed methods as well as those from performing simple thresholding and the thresholding method described in [15] are reported in Section 2.4. Finally, conclusions are discussed in Section 2.6. 2.2 Methods 2.2.1 Measurements Assume that an UWB transmitter sends pulse (t). Due to multipath propagation, the received signal is described by h(t) = X i i (t i); (2.1) where i and i are the complex amplitude and time delay of the ith path, respectively. The line of sight path delay is 0. The receiver radio approximately measures the channel 13 impulse response convolved with the pulse shape. Fig. 2.1(a) is an example of how the transmitted pulse may follow many di erent paths to arrive at the receiver. The number of multipath components seen by the receiver depends on the environment around the radios. When a person enters the environment, the person's body will cause a new multipath component at the receiver as well as a ect existing multipath components. This is illustrated in Fig. 2.1(b). The delay associated with this new multipath component is , which we refer to as the bistatic delay. The person also a ects many i for i . In bistatic or multistatic radar systems, the bistatic delay, described by , is used to locate and track objects near the radio transmitters and receivers. Assuming component i is a single-bounce path (i.e., the path is a ected by only one scatter as it travels from transmitter, to the target, and then to the receiver), the scatter is located on an ellipse with foci at the transmitter and receiver locations. That is, the locations where the scatter may be located are points S where the distances from S to the transmitter and receiver, St and Sr, sum to: St + Sr = c i; (2.2) where c is the speed of light. This work seeks to accurately estimate the bistatic delay , that of the path created by the person, particularly in environments with \\cluttered\" impulse responses (i.e., those where individual multipath components arrive closely in time and become di cult to separate from the CIR). Estimation of is a key primitive operation for UWB impulse radar systems|estimates from multiple transmitter and receiver pairs can be used to determine possible scatter locations under a single-bounce assumption, as we explore in Section 2.2.8. The IEEE 802.15.4a channel modeling subgroup performed a large measurement cam- paign to help develop an ultra-wideband channel model for many indoor and outdoor environments [24]. These models are useful for estimating the line-of-sight time of arrival for UWB pulses. They are not useful, however, for simulating the channel impulse response of the environment with the presence of a person or with respect to the person's location within the environment. Estimating under these channel models in simulation would not be possible without additional models of how the channel is a ected as a function of a person's location. As described in Section 2.1.1, background subtraction is a standard method for removing the static background CIR from a current CIR measurement. However, we have found that background subtraction is not e ective in cluttered environments. An example is shown 14 in Fig. 2.2, which shows the true bistatic delay and a captured CIR subtracted from a calibration CIR over about 20 ns of time. Both CIRs were averaged over ten measurements, each set of ten captured with a signal-to-noise ratio (SNR) of 31 dB. Individual multipath components are indistinguishable and the signal is very noisy. If background subtraction were e ective, the amplitudes prior to would be signi cantly lower than the amplitudes after , however, this is not the case. Better methods than simple subtraction to quantify the changes in the CIR are needed. 2.2.2 Quanti cation of Change We describe in this section an alternative to background subtraction. We introduce a divergence measure which quanti es the change between the signal energy measured during the period when the environment is static and the current period. We consider a discrete-sampled version of the signal energy, rk, given by rk = Z (k+1=2)T (k1=2)T jh(t)j2dt; (2.3) where T is the sampling period. For example, in our experimental work, we use T = 1 ns which covers between 62 and 63 discrete samples for the CIRs captured using the UWB radios. Essentially, rk is the energy in multipath components contained within a T-duration window near time delay kT. We call this T duration window \\range-bin k.\" The vector r = [r1; : : : ; rn]T is the sequence of rk samples. The unit of time k describes the \\fast time\" of the radar signal. We choose to estimate the energy in each range-bin rather than using deconvolution to nd the CIR. Performing deconvolution to determine the number and arrival times of mulitpaths arriving at the receiver will give incorrect multipath quantities and delays when the multipath experience frequency distortion, which is common for UWB signals [25]. Moreover, the memory and computational burdens are reduced by considering T duration windows rather than all samples. In this work we use the Kullback{Leibler (KL) divergence to quantify the change in the signal energy rk at each time k. The KL divergence is a measure of how many additional bits would be required to encode the samples of one distribution relative to another distribution. This is also known as relative entropy [26]. For continuous distributions the asymmetric KL divergence is de ned as D(p(x)kq(x)) = Z p(x) log p(x) q(x) dx; (2.4) 15 where p(x) and q(x) are the probability densities of rk for the calibration measurements and for those under test, respectively. These measurements are with respect to the \\slow time\" of the captured CIRs. The symmetric KL divergence is de ned as D(p(x)kq(x)) + D(q(x)kp(x)). The observation signal, Ok, in this model represents the di erence between rk and rk of the empty room, that is, the calibration samples. In this work, this di erence was calculated as the symmetric KL divergence. For the observed signal, Ok, we use the symmetric KL divergence assuming Gaussian distributions for rk. This measure is given in closed form by, Ok = 1 2 2 p 2 q + 2 q 2 p + ( p q)2 2 p + 2 q 2 p 2 q ! 1; (2.5) where p and 2 p are the mean and variance of rk during calibration, and q and 2 q are the mean and variance of rk from the CIR measurements collected for testing. These mean and variance estimates are for a xed k over several CIRs, which is the \\slow time.\" This closed form solution for Ok is non-negative and the pdf fO;i will allow us to estimate Xk by applying our hidden Markov model. The unit of measurement for the KL divergence, or Ok, is bits. The assumption that rk is Gaussian with respect to the \\slow time\" of the measurements is important to the closed form solution of Ok given in equation 2.5. To show that rk follows a Gaussian distribution, each set of 10 samples of rk for the empty room was normalized to have a mean of 0 and a variance of 1. The SNR for these empty room measurements was 31 dB. These samples were then aggregated for testing. With 10 sets of 90 samples of rk for the six radio pairs gives 5400 samples. To reduce sampling instability, each CIR was interpolated by a factor of four and then cross correlated with each other to align them in time. A histogram of the normalized samples is given in Fig 2.3. Submitting these samples to a Kolmogorov{Smirnov test fails to reject the null hypothesis that they come from a standard normal distribution with p = 0:198. These results are similar for other CIRs captured throughout this experiment. An example of an observation vector O of KL divergences is given in Fig. 2.4. This particular example is one where a rst threshold-crossing method would be unable to correctly estimate the true bistatic delay, k where k = b T c, of 15. For this particular set of empty-room and target CIRs, the SNRs were, respectively, 27 and 28 dB. This example shows how the assumption of easily being able to discern the background signal from the 16 changes to the CIR can sometimes be wrong. In this case, there is a very large divergence at a time when the signals should have shown little di erence. The majority of observation vectors will not exhibit this behavior but the errors of vectors such as these signi cantly impact the overall error for an environment as will be shown in Section 2.4. Other distance measures or distributions could be applied. However, the KL-divergence and Gaussian assumption provide a standard approach for this proof-of-concept study. 2.2.3 CIR Changes as a Hidden Markov Model A hidden Markov model is a special case of a Markov chain. The states of a HMM are not directly observable but may be inferred. Other signals available for observation help determine the past and current states of the system. Let i be the probability of initially starting the HMM in state i, Pij is the probability of transitioning from state i to state j, and fO;i is the probability of observing signal O given the HMM is in state i, that is, f(OjXk = i). A simple illustration of a hidden Markov model is shown in Fig. 2.5. In the case when the observations are continuous, we use the probability density function (pdf) conditioned on the state, fO;i, for a continuous valued random variable. This is the typical way to describe a HMM for continuous-valued observations [27]. By knowing fO;i, Pij , and i, a best estimate of the current state at each time, ^X k, can be calculated. This is found by applying the forward-backward algorithm to the sequence of observation signals. When the estimated states transition from ^X k = 0 to ^X k+1 = 1, this gives an estimate for k and indicates the presence of a person due to the changes to the observation vector. Estimation of k , where k = b T c, is equivalent to estimating . Due to multipath scattering and the person's impact on those later-arriving signals, rk will experience changes, or Xk = 1, for many k k . The advantage of applying a HMM is that information over all k is considered when solving for Xk rather than considering values at each k independently of changes at all other k. A more thorough introduction to hidden Markov models and the algorithms used to infer information about them can be found in [27]. 2.2.4 Continuous Observation Densities The observations Ok are continuous valued and their probability distribution is described by fO;i, the probability density function of Ok given Xk = i, i 2 f0; 1g. The HMM parameters fO;i, i, and Pij are estimated using the data D collected in one room and are used as initial estimates of the HMM parameters when estimating k for the other room. 17 The data sets Di, for each state i, are made using the knowledge of k by D0 = fOkjk < k g; (2.6) D1 = fOkjk k g: (2.7) Dividing the observation signals in this way assumes that there will only be one transition from state 0 to state 1 and no transitions back to state 0, that is P10 = 0 and P11 = 1. Under the assumption that Xk = 1 given k k , one may also assume that P1;0 = 0 and P1;1 = 1, that is, P(OkjXk = 1) remains constant as k increases. This assumption may not be true|a person's e ect will eventually diminish for large k. Also, a probability of 0 leaves little opportunity for change during optimization. To improve the model, we allow a small probability of returning from state 1 to state 0 (i.e., set P10 = where is a small value greater than 0). In [16], no assumptions were made regarding the distribution the observations took on. The distribution was estimated by performing an Expectation Maximization algorithm to t the data to a Gaussian mixture model. This operation was computationally expensive but e ective. In this work we utilize our observation that the densities are similar to a log-normal distribution. Under this assumption, well known maximum-likelihood estimates are used for the distribution parameters. Fig. 2.6 shows the empirical cumulative distribution functions (CDFs) of the aggregate samples before and after k for one room. The natural log is applied to Ok in these distributions. This log-normal approximation reduces the computational load without sacri cing solving accuracy. Initial estimates for i and Pij are given by [27, eq. (40a{b)] using the training data. 2.2.5 HMM Solving The HMM parameters are described by as = [ i; Pij ; fO;i] : (2.8) The data from one room is used as training data to obtain an initial estimate of to begin solving for k with the other room's data, or that of the measurement room. The following describes how k is estimated for the measurement room once is estimated from the training data, as described previously. 18 Finding ^X k, the estimate of Xk, for the measurement room is done by solving the forward-backward algorithm. This algorithm nds the most likely state X at each range-bin k [27]. ^X k = arg max i P(Xk = ijO; ): (2.9) The forward-backward algorithm is di erent than the Viterbi algorithm, which nds the most likely state sequence over all k. It may seem more appropriate to use the Viterbi algorithm to estimate when the state change occurs. The Viterbi algorithm, however, only returns a state sequence. By using the forward-backward algorithm, the additional uncer- tainty information of P(Xk = ijO; ) is available for each k when performing localization. It should be noted that estimates for k are not constrained by the room boundaries or by any prior information about where the person might be located. After estimates for Xk are obtained, the Baum{Welch algorithm uses these estimates to update the set of HMM parameters such that P(Oji; n+1) > P(Oji; n): (2.10) This algorithm is an iterative optimization on the space of to maximize P(Oji; n). The HMM parameters are updated over all sets of D as described by Rabiner [27]. Also, fO;i is again found by estimating the distribution as log-normal using Di. However, Di is now found as Di = fOkj ^X k = ig: (2.11) The algorithm continues for a predetermined number of iterations or until P(Oj n) no longer increases more than a given tolerance with each iteration, that is, P(Oj n) P(Oj n1) < . The nal estimate for k is ^k = fk j ^X k 6= 18k < k g: (2.12) This nds a local maximum in the space of possible but may not nd the global maximum. The e ectiveness of this algorithm is dependent on the initial values of the HMM parameters and the data itself. Other optimization algorithms exist but were not explored in this research. 19 2.2.6 First Threshold Crossing A standard method to determine the bistatic delay, k , is simply to nd the rst time at which Ok is greater than a threshold. We refer to this method as rst threshold crossing (FTC). Speci cally the estimate of k in rst threshold crossing is given by ^kFTC = arg min k Ok > ; (2.13) where is a threshold. We show the performance of this method in Fig. 2.7 as a function of . To show how the method would perform with training, we assume that is set by using the that achieves the lowest root mean squared error (RMSE) in one room, and test performance with that in the other room. The work presented by Zetik et al. in [15] gives another method for thresholding the received CIR to estimate . This method is also used for comparison in Section 2.4.1. 2.2.7 Localization Multiple range estimates allow localization to be performed. In this section, we describe methods for merging bistatic range estimates to obtain a position estimate. Clearly, range estimates contain errors, and any location estimator must deal with these noisy inputs. The works of Paolini et al. [28] and Bartoletti et al. [29] provide important information about uncertainty and error in performing localization in an UWB network. One advantage of the HMM-based approach we propose in this paper is that it provides a \\soft\" decision on the bistatic range estimate. The forward-backward algorithm quanti es the probability of each state i at each time index k, P(Xk = ijO; ). If the conditional probability of state 1 increases from zero to one very quickly at time k, then this delay bin k is very likely to have been the true bistatic delay. If the conditional probability increases slowly from zero to one over several delay bins, then it becomes more di cult to estimate k over those bins when the probability increases. Essentially, a quanti cation of the probability of each delay bin k being the true bistatic delay is given by the rate at which the conditional probability changes. The forward-backward algorithm nds the conditional probability of being in a given state at time k. To simplify notation going forward, we will let k = P(Xk = 1jO; ). Since there are only two states, k fully describes the probability of being in a given state at time k. Also, let (x)+ be de ned by 20 (x)+ = ( x if x 0 0 if x < 0: (2.14) Assuming a single-bounce model, each time delay measurement corresponds to a region on the plane given an ellipsoid with the transmitting and receiving radios at the foci. For a location estimate on a 2D plane, at least three radio pairs must give range estimates for the overlapping elliptical regions to produce a unique solution, assuming noise-free range estimates. Due to the cluttered environment, whose background UWB re ections are often much stronger than the ones caused by a person, the range estimates cannot be assumed to be noise-free. For this work, to mitigate the e ect of having range estimate inaccuracy, we obtained data from six radio pairs. One way localization can be solved is as an inverse problem, described by Chang et al. as a semilinear algorithm (SLA) [13] which models the radio locations and range estimates as a linear function [13, eq. (4)]. SLA is solved using a linear least squares method. Where range estimates alone are available, solving the problem as an inverse problem makes the most sense since these estimates will often not converge perfectly due to errors and noise. The output of the HMM, however, is more than a simple range estimate. Additional information about the probability of being in one of the two HMM states is available. This additional uncertainty at each time k can be used to improve localization accuracy. The work of Ergut et al. [30] o ers another localization method complementary to the range estimation method proposed in this work. They propose an arti cial neural network to localize a target within a sensor network which returns range estimates as inputs to the neural network. Their localization method, however, requires localization training and makes the assumption of high SNR range estimates. The work also fails to show how their method performs using empirically measured channel impulse responses. The work of Chiani et al. performs localization using a soft image derived from the time-of-arrival estimates [31]. This is similar to our proposed localization method in that no single range estimate is used to perform localization but an intermediate set of data, which in our case is the output of the forward-backward algorithm. No detection methods will be applied, however, to the UWB image. Chiani et al. also o er a number of ways to obtain pixel values for the soft image. By nature of the experiments performed, tracking algorithms were not applied. The work of Bartoletti et al. apply tracking algorithms that may aid in the localization of a moving person [32]. 21 2.2.8 Proposed Localization Method In this work, localization is solved as a forward problem as follows. We discretize space into P pixels containing the area being monitored. We denote li to be a quanti cation of the \\presence\" of a person in pixel i. The image vector is then L = [l1; : : : ; lP ]T ; (2.15) where pixel i is centered at coordinate zi = (xi; yi). A person in pixel i would, assuming the single-bounce model, be measured to be in range-bin km i for transmitter/receiver pair m, where m 2 f1; : : : ;Mg, km i = ktm zik + kzi rmk ktm rmk dk ; (2.16) where tm and rm are the transmitter and receiver coordinates for link m and dk is the distance light travels during one time bin. The value li is given by li = \" MX m=1 [Am]p i #1 p ; (2.17) where A is the non-negative di erence function of at km i , [Am]i = ( km i km i 1)+; (2.18) with 0 = 0. Equation (2.17) is the p-norm of fAmg for all radio pairs m = 1; : : : ;M at pixel i. A p-norm of 0 (i.e., p = 0) gives a count of nonzero values and a p-norm of 1 is a sum of the elements. In this work, p = 0:2 was found to give the best performance and was the value used for the results given in Section 2.4.2. This p-value weights the elements of A such that, qualitatively, lower values are weighted more and higher values are weighted less. Rather than using k, localization can also be done using estimates ^k . This would change the way A is calculated from what is given in Equation (2.18) to: [Am]i = ( 1 if i = ^k 0 otherwise: (2.19) 22 Results for both of these methods for solving localization as a forward problem as well as solving using SLA are given in Section 2.4.5. To understand pixel value li more intuitively, we recall that km i km i 1 is a soft metric for the probability that pixel i is at the same bistatic range as the person, as indicated by the measurement on link m. Due to the p-norm in (2.17), li is a type of average of these probabilities over all links. This method is especially useful when the measurements from a link are ambiguous, and thus k for that link doesn't change from zero to one suddenly. The uncertainty in f kgk is re ected in the presence image L. For purposes of noise reduction, we apply a 2-D Gaussian lter to image L. For experiments with one person in the area, we take the coordinate of the pixel with highest li (after the ltering) as the location of the person. 2.3 Experiment We conduct two types of experiments for evaluation of our proposed algorithms. First, we conduct in-room experiments where transmitters and receivers are in the same room as the person being located. Second, we conduct an experiment in which the transmitter and receiver are on the other side of an interior wall of the room in which the person is located. In all experiments, we use two P220 UWB impulse radios from Time Domain, Inc., to capture CIR measurements with an antenna height of 0.9 m. The radios transmitted at power level of 16:13 dBm with a transmitted center frequency of 4.7 GHz and a 10 dB radiated bandwidth of 3.2 GHz. Pulses are transmitted at an average rate of 9.6 million pulses per second. Additional information about the P220 radios can be obtained from Time Domain [33]. 2.3.1 In-room Experiments We rst conduct measurements in rooms 3325 and 1280 in the Merrill Engineering Building. Two rooms are measured so that one room can be used as a training room while the other is used as an experiment room. Figs. 2.8(a) and 2.8(b) describe the positions of the radios and where the person stands in each room. Room 3325 contains typical o ce furniture; desks, chairs, bookshelves, and computers. It measures 6.2m by 6.2m with the ceiling 2.5m from the oor. Room 1280 is a classroom and all of the desks and furnishings were removed from the room for the experiment. Room 1280 measures 8m by 8.2m with the ceiling 2.7m from the oor. We collect both empty-room (i.e., no person in the room) calibration measurements and measurements which represent all measurements possible in a four UWB transceiver 23 multistatic network when a person is standing at any of the possible grid points in the two rooms. Since we have only two UWB transceivers, we conduct these measurements as follows. The two radios are placed in any of the four locations designated for the radios in the room. Ten calibration measurements of the CIR, or rk, are taken when the room is empty. Then, at each of the designated points, a person stands and remains as motionless as possible while ten more measurements of the CIR are taken. Each set of ten measurements was captured over approximately ve seconds. After collecting measurements at all points, the two radios are moved. This process is repeated for the M = 6 pair-wise radio locations. The full process is repeated in the second room. These sets of ten are those used to estimate and 2 for rk for the calibration and measurement sets. Experiment A uses the data collected in room 1208 as the training room data and the data collected in room 3325 as the data for the experiment room. Experiment B swaps the data used for the training and experiment rooms and performs the estimation again. 2.3.2 Through-wall Experiment In addition to ranging and localizing a person that is in the same room as the radios, one data set is also collected to test ranging through an interior wall. Two radios are placed 1 m apart from one another and 18 cm from the wall in room 3220 separating it from room 3230 in the Merrill Engineering Building at the University of Utah. Room 3220 and 3230 both measure 6.7m by 4.9m with a ceiling height of 2.7 m. The separating wall is approximately 10 cm thick and is constructed of a metal support with internal insulating material and covered in drywall. We also report the power loss due to wall penetration, in order to characterize the experiment condition. To estimate the penetration loss of the wall, the CIR is measured with the radios 4.5m apart with both radios in room 3220. The transmitting radio is then placed on the other side of the wall in room 3230, and the receiving radio is also moved to maintain a 4.5m separation. The CIR is measured again and the line-of-sight component of two measured CIRs are compared. The measured power loss of the wall is approximately 5 dB over the 3{5 GHz band. The measurements are made as follows. A person stands at 30 di erent locations in the adjacent room 3230 while the CIR was captured 20 times per location. Fig. 2.9 shows these two rooms with their corresponding person and radio locations. Both before and after all of these CIRs are sampled with a person present, the CIR for the empty room is captured 100 times. UWB pulse integration is also increased by a factor of 8 from what was used 24 in the other experiments. This increases the SNR of each CIR at the cost of lowering the maximum possible sampling rate. This through wall experiment is performed for just one radio pair, which is insu cient for localization. Instead, the purpose of this through-wall experiment is to allow us to quantify the performance of UWB impulse radio bistatic delay estimation. 2.4 Results In this section, we apply the methods proposed in Section 2.2 to the data collected as described in Section 2.3. We measure the performance of our proposed HMM-based bistatic delay estimator in three ways: (1) the RMSE of the bistatic delay estimator, (2) the false negative and false positive rates, and (3) the performance of localization using our bistatic delay estimates. We compare the results of our method of estimating bistatic delay to simple thresholding as well as the thresholding method given in [15]. The bistatic delay error is the di erence between the person's actual bistatic delay and the estimated bistatic delay, \" = T ^k k : (2.20) We use RMSE across all experiments to quantify average performance. We report false negative and false positive rates for the methods studied. For bistatic delay estimation, a false negative is when there was no person's bistatic delay detected when a person is actually present. For our HMM-based method, this corresponds to the forward-backward algorithm detecting no transition from state 0 to state 1 for the measured CIR. A false positive is when there was a bistatic delay estimated when no person was present. In all results, we chose a delay-bin duration T of 1 ns. The choice of T is a trade-o between computational requirements and quantization noise. We note that 1 ns of time corresponds to about 30 cm of distance traveled at the speed of light, approximately the width of an adult human body. Further, our results show errors signi cantly higher than 1 ns, and thus it has not been necessary for us to reduce T further. 2.4.1 First Threshold Crossing First, we test the performance of the FTC estimator as described in Section 2.2.6. We nd the threshold that is optimal (for minimum RMSE) for the training room and then use that threshold in the testing room. From this method, a minimum RMSE of 5:25 ns 25 is achieved for Experiment A and 5:20 ns for Experiment B. Next, we see what minimum could have been obtained for the testing room even if the optimal threshold for that room had been known. These absolute minimums achieved are 3:28 ns and 4:58 ns, respectively. Fig 2.7 shows how the RMSE varies as a function of the threshold. Clearly, the optimal threshold would not be known a priori for each room. Fig. 2.7 shows the sensitivity of the RMSE to the chosen threshold. For Experiment A there is a large change in the estimates with a small change to . This large change to the RMSE, occurring near values of 65 and 99 bits, are due primarily to one set of CIRs for one point and radio pair. Without knowledge of the true values for k , one would still notice the large change to ^k with small changes to . The e ect on RMSE due to this one particular person-location and radio-pair combination, which we also describe as an outlier, is shown in Fig. 2.10. There were no false negatives for the range of tested in Fig 2.7 for either experiment using the rst threshold crossing method. The work done by Zetik et al. [15] gives a somewhat di erent method for thresholding the signals. The background is continually updated for each UWB node, which would correspond to a radio pair in our work, as: bi = bi1 + (1 )mi; (2.21) where b is the background estimate and m is the newly measured CIR. The signal s then used for thresholding is: si = mi bi: (2.22) This removes the static background signal from the time-varying signal, which is what we wish to detect and range. The threshold is calculated as: ti = 0:3 + 0:7 ni jjsijj1 si 1 ; (2.23) where ni is the peak noise level of mi. Using the method of [15], described in Equations (2.21), (2.22), and (2.23), and the data collected, we obtain an RMSE of 6:5 ns and 10:6 ns for experiments A and B, respectively. When rst threshold crossing is performed on the through-wall experiment data, a plot of RMSE versus threshold is obtained and shown in Fig. 2.11. This is comparable to 26 those shown in Fig. 2.7. Notice that the that achieves the optimal estimation of k is di erent for each experiment and varies signi cantly. In other words, the optimal cannot be determined from data measured in a di erent location. 2.4.2 HMM-based Method The HMM and process described in Section 2.2.5 are applied to the two in-room exper- imental data sets. The changes to RMSE for each iteration of the Baum{Welch algorithm is shown in Fig 2.12. The RMSE achieved after 15 iterations is 2:85 ns and 2:76 ns for Experiments A and B, respectively. This corresponds to less than 90 cm of range error. Others have described how range error corresponds to localization error in UWB sensor networks [28], [29]. There were no false negatives. The bias, E[^k k ], was 0:3 ns for Experiment A and 0:2 ns for Experiment B. For all points on the line-of-sight of two radios, the estimated range was always ^rk = 1 ns, which is between times 0 n and 1 n because T = 1 ns, which is the correct estimate. Although the line-of-sight signal is blocked by the person, there is still a signi cant amount of energy arriving at the receiving radio from di raction and re ection in the environment. The observations Ok from the received CIRs with the line-of-sight path blocked indicate that ^X k = 1 over all k rather than transitioning from Xk = 0 to Xk+1 = 1 at some k, giving an estimated range of ^rk = 1 ns. The marked improvement in RMSE from using a HMM over energy detection also comes without foreknowledge of an ideal threshold value. Although an initial estimate for is required, the Baum{Welch algorithm eliminates much of the error due to a poor estimate, as will be shown with the through-wall results 2.4.3. The HMM, unlike a simple threshold, takes into account the data across all time values to estimate k . The stopping condition used for the given results is to continue the Baum{Welch al- gorithm until there is little change to P(Oj ) from one iteration to the next. That is P(Oj n) P(Oj n1) < . Experiment A converges, using this metric, after 9 iterations and Experiment B after 14 iterations. 2.4.3 Through-wall experiment Our proposed HMM method is also applied to data captured through a wall dividing two rooms as described in Section 2.3.2. Observation vectors are calculated using all of the available empty room CIRs and CIRs with a person present. With the observation vectors and an initial estimate for the HMM parameters , estimates for k can be found. 27 Using the that is found to be optimum for any one of the three environments as the initial for any of the other environments results in the same solution for from the Baum{Welch algorithm. This is illustrated using the through wall data. For the through wall data, there are three choices of , two obtained from the data collected from the two in-room experiments described in Section 2.3.1 and one from the data and known locations of this through wall data. The obtained from the through wall data could not be used in a production system because it is derived using a knowledge of k . If k is known, there is no reason to use it to nd to then estimate k . It is used here solely for illustrative purposes. Fig. 2.13 shows the bistatic delay RMSE at each iteration of the Baum{Welch algorithm for the three di erent choices for at the rst iteration. The choice of greatly in uences the RMSE at rst, but the e ect of the choice is ultimately negated by the Baum{Welch algorithm. The nal RMSE in all three cases is 1.33 ns. This nal error is better than the results obtained with the subject in the same room as the radios. There are several reasons for this. 1. Number of samples: Many more samples of the empty room were collected and used in determining the KL-divergences in the through-wall experiment (200) compared to the in-room experiments (10). These additional samples help to reduce the noise in the observation vectors. The e ect of choosing di erent empty room samples is explored further below. 2. Additional integration: Additional signal integration was done in sampling to reduce noise in the CIRs because of the additional path loss in the through-wall experiment. To show the e ect of the number of empty room samples on the performance of the ranging estimation (item 1 above), we run an experiment in which we reduce the number of empty-room samples used in the through-wall experiment. Here, we calculate observation vectors of KL-divergences using sets of 20 sequential empty room samples. From the two sets of 100 empty room samples, this leads to 162 sets of sequential samples. The initial choice of was the same used in Experiment A. The overall RMSE was calculated for each of these sets of empty room samples. Two of the 30 person locations had a wide variation in their range estimate depending on which set of empty room samples was chosen. Fig. 2.14 shows the empirical CDF of the nal RMSE obtained using each of these sets of empty room samples both with and without these two person locations. For the trials using all person locations, 12.3% of the trials resulted in an RMSE better than the 1.33 ns achieved using all of the empty room samples together. The overall RMSE 28 for all of the trials using 20 empty room samples is 4.19 ns. This illustrates that, on average, using a fewer number of empty room samples degrades performance. 2.4.4 False Positives Testing for false positives, or nonzero estimates of k in empty room samples, was also performed. The Baum{Welch algorithm was not performed on these samples, that is, no updating of the HMM parameters was done for re-estimation of k . False positives were tested by randomly dividing the set of empty-room samples into the known empty-room and possible point sample sets. Due to the limited sample sizes for empty-room samples, this random set division allows us to simulate how false positive tests might perform using di erent sample sets that are not available. For each radio pair, the available samples were divided evenly between the known empty-room sample set and the possible point sample set. These two sets were used to nd the observation vector of KL divergences, which the HMM uses to estimate k . For each of the six transmitter/receiver pairs for each of the two rooms, 1000 trials were performed using the random subset division described for a total of 12 000 trials. Of these a total of 50 trials resulted in false positives, that is, a 4:2 103 false positive rate. We note that over half of the false positives come from a single transmitter/receiver pair in one of the rooms. Notably, this pair had just 10 empty-room samples available for testing. This is the fewest number of empty-room samples for any transmitter receiver pair. 2.4.5 Localization Results for localization are given for both the forward method described in Section 2.2.8 and the SLA described described by Chang et al. [13]. The forward solving method is done in two ways, rst using k where k = P(Xk = 1jO; ) and second using only the range estimates, ^k , without the additional information of the probability of being in a given state. The SLA described by Chang et al. only uses range estimates for localization. A summary of the results of each localization method with its available information is given in Tables 2.1 and 2.2. All values are given in centimeters. The forward solving method described here gives location estimates that are signi cantly better than those from the SLA described by Chang et al. Taking into account k rather than using ^k alone also improves the location estimates for the forward solving method. Figs. 2.15 and 2.16 describe the true person locations, as shown previously in Fig 2.8, and the estimates for those locations using the forward solving method with all available information. 29 Range estimates from three unique radio pairs is su cient to unambiguously localize a target. In this work there are estimates from six radios pairs for each localization image. Having information from additional radio pairs aids in accurate localization because some of the estimates may be inaccurate. As long as the information from at least three radio pairs are accurate, the target's location can be accurately estimated. Also, as shown in the results given, using all of the available information from the forward-backward algorithm, and not using the range estimate alone, increases localization accuracy. Chang et al. [13] o er a Cram er{Rao lower bound on location estimation accuracy assuming range estimates are corrupted by Gaussian noise. They nd asymptotic bounds as the number of transmitters and receivers grows for a few di erent scenarios. For an object at the origin, they nd the bound to be 2 2 NM ; (2.24) where NM is the number of UWB links in the network, which in our case is 6. Range errors of 2:85 ns and 2:76 ns correspond to a of 85 cm and 83 cm, respectively. This gives lower bounds of 24 cm and 23 cm for localization error. Notice that in Table 2.1 the results for using only range estimates results in errors of 155 cm and 75 cm, respectively. The actual error is higher than these bounds. This may be due to the fact that the noise in the range estimates may not be Gaussian, as is assumed by Chang et al. This may also be because the assumption that range errors are independent on di erent links may not be valid. Also note that much better performance is achieved when all of the additional information available from the forward-backward algorithm is used rather than using range estimates alone. However, no lower bound on performance is available in this case. 2.5 Discussion One primary limitation of the algorithm as proposed is that it assumes only one person is causing changes to the CIR. To account for more people, future work must expand the HMM-based estimator to estimate a bistatic delay for each person in the environment. Research must determine what methods to use in the multiple person case. For example, more states could be added to the Markov model that account for more than the channel impulse response simply being a ected or not. Additional states could estimate the number of targets to have a ected the channel. Another possibility is if joint estimation of the num- ber of people and their bistatic delays improves performance. Once localization of multiple 30 targets has been performed, several multitarget tracking methods have been developed for RSS-based device free localization [34]{[39]. Another limitation of the proposed algorithm is its reliance on calibration CIRs. We have shown that the proposed method is robust to variances in the initial choice of but it is not to signi cant changes to the environment. New calibration samples would be needed each time a change in the environment requires it. One possible way to eliminate the calibration requirement is to use the CIRs that were sampled immediately preceding the CIRs with a possible target. This, however, may introduce bias and make static targets harder to detect. 2.6 Conclusions In this paper, we introduce and experimentally verify a hidden Markov model-based algorithm for estimating the bistatic delay in an UWB impulse radar system. We show the proposed algorithm achieves a lower RMSE than rst threshold crossing methods for highly cluttered multipath environments. Applying the Baum{Welch algorithm allows the proposed estimator to adapt its parameters to be best for the particular environment. We show the algorithm is robust to initialization parameters derived from a di erent environment. Compared to using the rst threshold crossing estimate of , our method reduces error by almost half. Since these estimates of the person's bistatic delay are used directly in tracking algorithms, we expect to similarly improve UWB-based localization performance. The forward solving method described here for localization using the probabilities k was very e ective, achieving a median error of 18 cm. 31 (a) Static Environment (b) Person's E ect Figure 2.1. When a person appears at x0 in the environment between the transmitter at xt and receiver at xr, an additional path is caused with path length kxt x0k + kx0 xrk, where k k is the L2 norm, and also a ects multipath components with longer path lengths. Figure 2.2. The di erence between a calibration CIR and a new CIR gives a noisy signal with multipath components that are indistinguishable from one another. The red, dashed line is the actual bistatic delay (i.e., ). Table 2.1. RMS Localization Error (cm) Forward SLA All Info Range Only Range Only Rm 3325 36 155 165 Rm 1208 24 75 194 32 Figure 2.3. Empty room samples normalized to have zero mean and a variance of 1 exhibit a Gaussian distribution. Figure 2.4. An example of an observation vector where no threshold can nd the true k , which is 15 in this case. The HMM correctly estimated k for this vector. Table 2.2. Median Localization Error (cm) Forward SLA All Info Range Only Range Only Rm 3325 16 67 159 Rm 1208 16 29 172 33 P01 P10 P00 P11 X =0 Xk =1 k f(O|X =0) k f(O|X =1) k Figure 2.5. The change in CIR measurement we observe at range-bin k, Ok, has a distribution dependent on the state, Xk, of a hidden Markov chain. Figure 2.6. Empirical CDFs of the log of Ok for one room. Although these distributions are not precisely log-normal, this assumption is reasonable for the solving methods. 34 Figure 2.7. Performance of rst threshold crossing method given by equation (2.13) as a function of threshold . (a) Room 3325 (b) Room 1280 Figure 2.8. Circles are points where the person would stand and squares are radio locations. Gray rectangles are furniture. Neighboring points are spaced 90 cm apart. 35 Figure 2.9. Squares represent radio locations in room 3220 and circles represent person locations in room 3230. Person locations are spaced 60 and 120 cm apart. Figure 2.10. RMSE for Experiment A with and without the outlier point. 36 Figure 2.11. Performance of rst threshold crossing method for the through-wall experi- ment Figure 2.12. Performance of HMM-based estimator of k as a function of iteration count. 37 Figure 2.13. The RMSE for the through-wall experiment converges to 1.33 ns for each of the initial choices of derived from the data for each of the three rooms. Figure 2.14. Variance of the estimator based on which set of empty room samples is used. 38 Figure 2.15. MEB 3325 actual person positions (O) and localization estimates (X) using the forward solving method. Figure 2.16. MEB 3325 actual person positions (O) and localization estimates (X) using the forward solving method.. CHAPTER 3 JOINT ULTRA-WIDEBAND AND SIGNAL STRENGTH-BASED THROUGH- BUILDING TRACKING FOR TACTICAL OPERATIONS Accurate DFL-based on RSS measurements requires placement of radio transceivers on all sides of the target area. Accuracy degrades dramatically if sensors do not surround the area. However, law enforcement o cers sometimes face situations where it is not possible or practical to place sensors on all sides of the target room or building. For example, for an armed subject barricaded in a motel room, police may be able to place sensors in adjacent rooms, but not in front of the room, where the subject would see them. In this paper, we show that using two UWB impulse radios, in addition to multiple RSS sensors, improves the localization accuracy, particularly on the axis where no sensors are placed (which we call the x-axis). We introduce three methods for combining the RSS and UWB data. By using UWB radios together with RSS sensors, it is still possible to localize a person through walls even when the devices are placed only on two sides of the target area. Including the data from the UWB radios can reduce the localization area of uncertainty by more than 60%.1 3.1 Introduction Device free localization systems can be used in tactical operations or crisis situations to help emergency personnel know where people are in a room or building before they enter [8]. These systems do not require people to participate in the localization e ort by wearing or carrying sensors or radio devices. Systems based on radio frequency measurements are 1 c 2013 IEEE. Reprinted, with permission, from M. McCracken, M. Bocca, and N. Patwari, \\Joint Ultra-wideband and Signal Strength-based Through-building Tracking for Tactical Operations,\" in 2013 10th Annual IEEE International Conference on Sensing, Communications and Networking (SECON), 2013. 40 particularly appropriate for e.g. hostage or barricade situations because RF penetrates (nonmetal) walls. However, in many such situations, it is not possible to place sensors on all sides of the building or area. For example, some sides of a building might have windows where an armed subject may be watching, and deploying sensors on that side could expose police to harm or escalating the situation. As another example, a room on an upper oor of a building may have some accessible interior walls (e.g., in a hallway), but the exterior wall may be unaccessible simply because of its height. This paper presents a system that expands the possibilities for RF-based DFL systems where an area cannot be surrounded with sensors by combining RSS-based DFL methods with bistatic UWB impulse radar methods. We are particularly motivated by discussions with our local SWAT team, who have unfortunately faced three situations in as many years in our metro area [40]{[42] in which hostages were taken by a barricaded subject in a hotel or motel room. Knowing the location of the suspect represents very valuable information in planning the actions (e.g., forced entry) required to bring the stando to an end safely. In such situations, sensors could be placed in adjacent rooms to the barricaded room, but rooms have front windows, and sometimes back windows; thus, front and back walls are potentially o limits. A DFL system based on RSS measurements [4]{[7] typically has radio transceivers, which we call RSS sensors placed on all four sides of a target area. RSS measurements of the links connecting every pair of sensors are used to estimate the location of the person in the room in real time. The localization process is based on models for the change in RSS introduced by the presence of a person on or near the link line, that is, the straight imaginary line connecting the transmitter and receiver [5], [34], [43]. When RSS sensors are placed only on two opposite sides of a room, the links cross the monitored area along one axis but not the other. This signi cantly degrades the localization accuracy of the system, especially along the axis with no crossing links [5]. UWB radios can be used for DFL through walls and can be accurate on the order of centimeters or tens of centimeters [12], [44]. Multiple UWB radios cooperating in a multistatic radar con guration can provide an unambiguous localization estimate [12]. A transmitter broadcasts a UWB impulse and receivers capture the time-domain CIR of the environment. Changes to the CIR are detected, and the time delay beyond the line-of-sight (LoS) pulse for each of these changes is used to estimate the range of the target from the radios [3]. These radios, however, can be prohibitively expensive to install on a permanent basis: a single UWB impulse radio can cost thousands of dollars, and using only a single 41 pair of radios provides insu cient information to unambiguously localize a target. In this paper, we introduce a joint DFL system that uses the changes measured in RSS and CIR to localize and track a target, such as a person, through walls. We demonstrate, in particular, the localization accuracy of a system which deploys sensors only on two opposite sides of a room. We call the axis parallel to the sides of the room without sensors the X axis and the axis parallel to the sides of the room with sensors the Y axis (see Figs. 3.1 and 3.2). The RSS sensors primarily provide the information about the target's y coordinate, while the UWB radios primarily provide information about the target's x coordinate. This removes the need to have RSS sensors on all four sides of a target room and reduces the number of UWB radios required for localization. In this paper we introduce methods to process and combine the RSS and CIR data in order to provide a unique position estimate. The experimental results collected in two deployments, i.e., a study room at the University of Utah and a motel room in Salt Lake City, show that the joint RSS-UWB DFL system can accurately localize a noncooperative target through walls. Even when the number of deployed devices is low, e.g, only two UWB radios and six (three per side) RSS sensors, the system can still provide a position estimate accurate enough to reliably indicate in which part of the room the person is located. In tactical situations where the only opportunity to have access to the target room is to open a breach in a wall with an explosive frame, this information can be used by police forces to decide which wall has to be detonated and avoid hurting or killing the suspect. In tactical operations or crisis situations, law enforcement may not have the possibility of calibrating the systems used for DFL in stationary conditions (i.e., when no person is located in the target area). Thus the methods used to process the data coming from the RSS sensors and UWB radios should be able to localize and track the suspect in the room from the start, making DFL a plug-and-play type of system. In this paper, we propose novel variance-based methods for RSS and CIR measurements that can localize the person without requiring an initial calibration of the system in stationary conditions. This is a proof of concept study to show the performance capabilities of a system that combines UWB information with RSS based localization techniques. In order to be practical for law enforcement personnel, the system should be able to be quickly deployed, and as such, we also study the performance of the proposed methods as a function of the number of sensors required to be deployed. This work does not address multiple target tracking. This is a future area of research. This has been a topic of research for RSS based localization [34]{[39], but the UWB 42 techniques used in this work have not been designed for multiple target range estimation. At the time of writing, there are several commercially available through-wall radio technologies that can help law enforcement determine the position of people inside a room. The Prism200 from Cambridge Consultants [45] is a through-wall radar system for deter- mining the location and movement of people for law enforcement or emergency personnel. The XaverTM products from Camero are also through-wall UWB solutions that provide similar capabilities [46]. Time Domain is another company that o ers solutions for target localization and tracking using UWB radios [47]. The UWB radios used in this work are a pair of P220 UWB radios from Time Domain. Compared to these products, the joint RSS-UWB DFL system described in this paper is considerably less expensive, as the RSS sensors cost a few tens of dollars each and only two UWB radios are required. Moreover, the compact size and low weight of the RSS sensors and UWB radios make our system easier to be installed. The paper is organized as follows. In Section 3.2, we describe the radio RTI technique used to process the RSS measurements coming from the RSS sensors. In Section 3.3, we describe the methods used for estimating the bistatic range of a target using UWB radios by modeling the changes to the CIR as a hidden Markov model. Section 3.5 describes a target tracking scheme. Section 3.4 introduces three methods to combine the RSS and CIR data in order to provide a unique position estimate. Section 3.6 describes the experiments carried out, while Section 3.7 presents the results and compares the performance of the di erent methods. Conclusions are given in Section 3.8. 3.2 Radio Tomographic Imaging (RTI) In RTI, originally introduced in [5], static radio transceivers placed at known positions form a wireless mesh network and collect RSS measurements that can be used to localize and track a person in real time without requiring the person to wear or carry any sensor or radio device. RTI can provide submeter localization accuracy, also in through-wall scenarios [8], [10], [48]. The RSS measurements of all the links of the network are processed in order to estimate a discretized image x of the change in the propagation eld of the monitored area caused by the presence of a person. The estimation problem can be de ned as: y = Wx+ n; (3.1) in which y and n are L 1 vectors of the RSS measurements and noise of the L links of the network, respectively, and x is the N 1 image to be estimated, where N is the number of 43 voxels of the image. Each element xn of x represents the change in the propagation eld due to the presence of a person in voxel n. The L N weight matrixWrepresents a spatial impact model between the L links of the network and the N voxels of the image. The model used in RTI [5], [8], [10], [43] is an ellipse having the foci located at the transmitter and receiver of the the link. The voxels located within the ellipse have their weight set to a constant which is inversely proportional to the root distance between the transmitter and receiver, while the voxels located outside of the ellipse have their weight set to zero. 3.2.1 Attenuation-based RTI For attenuation-based RTI (AB-RTI) we use the method introduced in [10]. In this section, we brie y present this method and the terminology that will be used also in the following sections. The RSS of link l on channel c at time instant k, rl;c(k), can be modeled as: rl;c(k) = Pc Ll;c Sl;c(k) + Fl;c(k) l;c(k); c 2 F; (3.2) where Pc is the transmit power, Ll;c the large scale path loss, Sl;c the shadowing loss, Fl;c the fading gain (or fade level [34]), l;c the measurement noise, and F = f1; : : : ;Hg is the set of measured frequency channels. Both the large scale path loss Ll;c and the shadowing loss Sl;c change very slowly with the center frequency. In our experiments, we use IEEE 802.15.4-compliant transceivers [49] which may transmit in one of 16 channels across the 2:4 GHz ISM band. Because the band, 80 MHz, is small compared to 2.4 GHz, we can assume that both Ll;c and Sl;c are independent of the frequency channel c. Consequently, Fl;c can be calculated as: Fl;c(k) = rl;c(k) Pc + l;c(k): (3.3) Due to the measurement noise l;c, the fade level can not be measured directly. Thus, we estimate it by using the average RSS, rl;c;, measured during an initial calibration of the system performed when no person is in the monitored area: Fl;c = rl;c min c rl;c: (3.4) In [34], the links are divided in antifade and deep-fade links depending on the change in RSS measured when a person crosses the link line, i.e. the imaginary straight line 44 connecting the transmitter and receiver. A link is in a deeper fade on channel c1 than on channel c2 if rl;c1 < rl;c2 . By de nition, Fl;c 0 and Fl;c = 0 for one channel c on each link. Antifade links are the most informative for localization, since their spatial impact area is limited around the link line, while deep-fade links measure a consistent change in RSS even when the person is far from the link line. For this reason, for each link l we calculate the fade level in (3.4) of each channel c 2 F, and we rank the measured frequency channels from the most antifade to the most deep-fade. If Ai is the set of size m containing the indices of the m top channels in the fade level ranking, the link RSS measurement yl at time k is calculated as: yl(k) = 1 m X c2Ai rl;c(k); (3.5) where rl;c(k) = rl;c(k) rl;c, i.e., rl;c(k) is the di erence between the current RSS measurement of link l on channel c and the average RSS measured during the initial calibration phase. 3.2.2 Variance-based RTI We present a new multichannel version of variance-based RTI (VB-RTI) extending and improving the results of [8]. In this new method, we also include the concept of fade level. The attenuation-based RTI method in [10] requires an initial calibration of the system in stationary conditions, i.e., when the monitored area is empty. Moreover, if the environment changes, e.g., when the suspect in the room moves furniture or other objects, the RTI system would need to be recalibrated or would otherwise lose accuracy. The work in [48] addresses this issue and introduce methods capable of estimating the baseline RSS of the links online. In tactical operations, such as when an armed person has barricaded himself in a house or motel room before the arrival of police forces on the scene, we cannot expect to require an empty area. Variance-based RTI can be applied in this scenario. The change in RSS due to the presence of a person on the link line can be quanti ed as the unbiased sample variance of the last Ns RSS measurements: ^sl;c(k) = 1 Ns 1 NXs1 p=0 (rl;c(k p) l;c(k))2 ; (3.6) where 45 l;c(k) = 1 N NX 1 p=0 rl;c(k p) (3.7) is the mean of the last N RSS measurements of link l on channel c, where N > Ns to be able to estimate the mean RSS of each channel on a longer time window and to lter the changes due to the person crossing the link line. Variance-based RTI does not require an initial calibration of the system and can adapt at run-time to eventual changes in the environment. For each link l, l;c(k) in (3.7) provides an estimate of the fade level of channel c at time k. As for attenuation-based RTI in Section 3.2.1, the channels are ranked from the most antifade to the most deep-fade. The link measurement yl at time k is calculated as: yl(k) = 1 m X c2Ai ^sl;c(k): (3.8) 3.2.3 RTI Image Estimation Since the number of links L is considerably smaller than the number of voxels N, the estimation of the image x is an ill-posed inverse problem that can be solved through regularization. In this work, we use a regularized least-squares approach [10], [43], [48], [50]. The discretized image of the change in the propagation eld of the monitored area is calculated as: ^x = y; (3.9) where y = [y1; : : : ; yL]T , and = (WTW+ C1 x 2N ) 1 WT ; (3.10) in which N is the regularization parameter. The elements of the a priori covariance matrix Cx are calculated by using an exponential spatial decay: [Cx]ji = 2x ekvjvik= c ; (3.11) where 2x is the variance of voxel measurements, and c is the voxels' correlation distance. The linear transformation is computed only once before the system starts operating in 46 real time. The calculation of ^x in (3.9) requires L N operations and can be performed in real time. Table 3.1 indicates the values of the parameters of the RTI image estimation process. 3.3 Ultra-wide Band Range Estimation Assuming an UWB transmitter sends a pulse (t), each CIR is measured as: h(t) = X i i (t i); (3.12) where i and i are the complex amplitude and time delay of the ith multipath component, respectively. The line of sight path delay is 0. The path delay of the target, which we wish to estimate, is . We will consider a discrete-sampled version of the signal energy, rk: rk = Z (k+1=2)T (k1=2)T jh(t)j2dt; (3.13) where T is the sampling period and k ranges from 1 : : :M discrete periods. In this work, T = 1 ns. From now on, CIR time delays will be considered only over discrete time intervals k rather than continuously on t. 3.3.1 Changes to the CIR as a Hidden Markov Model The changes to the UWB CIR are modeled as a hidden Markov chain. We will refer to this method as HMM-UWB or hidden Markov model (HMM) based UWB. A hidden Markov chain is one whose states, Xk = i, are not directly observable but are inferred from other observation signals, Ok, available from the system. The distribution of the observation signals is dependent on the state of the system, i.e., fO;i = P(OjX = i). To estimate the probability the system is in a given state at any time k, i.e., P(Xk = ijO; ), we need to know the distributions of the observation signals, the initial state probabilities i, and the state transition probabilities, Pi;j , all of which are described by = [ i; Pij ; fO;i] [27]. The observations, Ok, are the di erence between the CIRs of the static environment and the CIRs of when a person is located in the monitored area. This di erence is calculated as the symmetric Kullback-Leibler divergence, also known as relative entropy [26]. The distribution of the observations is approximately a log-normal distribution [16]. If the changes to the CIR are modeled as a hidden Markov chain, the CIR goes from an unchanged state, X = 0, to a changed state, X = 1, at the time delay corresponding to the 47 time traveled by the UWB pulse from the transmitter to re ect o of the target and then arrive at the receiving radio, i.e., k which is equivalent to . By applying this model to the system, standard HMM solving algorithms, such as the forward-backward algorithm [27], can be used to estimate when the system state changes and, thus, when changes to the CIR occur. The forward-backward algorithm determines the most likely state of the system at any given time as: ^X k = arg max i P(Xk = ijO; ): (3.14) These state estimates are used to estimate k as ^k = fk j ^X k 6= 18k < k g: (3.15) The work in [3] describes in further detail the method for estimating UWB bistatic range and its improved performance over other methods. From now on, we will let k = P(Xk = 1jO; ). k describes the probability those CIRs possibly a ected by a person at time k are in the a ected state. These probabilities are used to form the UWB localization image. When solving the forward-backward algorithm, accurate estimates of when state changes occur are dependent on how well models the true system parameters. A known from another environment can be used as an initial estimate for when solving for the state estimates. The Baum-Welch algorithm can then help tune to more closely match the true parameters and improve range estimates [3], [27]. In this work, we assume there are no major changes to the environment throughout each trial that would require new calibration CIRs to be captured. This allows us to use just one calibration period for estimating k . One possible way to eliminate the calibration requirement for HMM-UWB is to use the CIRs that immediately precede the CIRs with a possible target. This, however, may introduce bias and make static targets harder to detect. 3.3.2 Variance-based UWB Range Estimation An alternative method is to use the short-term variance of the CIR for each rk. We refer to this method as variance-based UWB (VB-UWB). k is calculated as: k = 2 rk grk ; (3.16) 48 where the variance 2 rk is the unbiased sample variance of rk over the NU most recent CIRs. In this work, we let NU = 5, corresponding to the number of CIRs captured in approximately 0:5 s. The normalization coe cient g is calculated as: g = g(1 ) + rk : (3.17) This is equivalent to applying a low-pass in nite impulse response (IIR) lter to rk. In this work, = 1 N . Because the variance of rk is high when the mean of rk is high and vice versa, we normalize the variance 2 rk by the mean of rk. In this way, k increases only when the person moves. This method is used in conjunction with the variance-based RTI method described in Section 3.2.2. The primary advantage of this method is that no calibration is required to solve for k. A disadvantage is that the target can disappear if it remains motionless over a long period of time. We alleviate this problem by applying the tracking method in Section 3.5. 3.3.3 UWB Image Estimation When estimating the UWB image, the image space is constrained to contain only the inner dimensions of the target room plus one additional voxel on each image edge. Discretizing the image space into N voxels, the image vector is: lu = [lu 1 ; : : : ; luN]T ; (3.18) where each voxel lun has a bistatic range to the UWB transmitter and receiver described by its path delay kn. The value of each voxel, lun , is calculated as the non-negative di erence function: lun = ( kn kn1)+; (3.19) where the non-negative di erence function is de ned as: (x)+ = ( x if x 0 0 if x < 0; (3.20) and assuming 0 = 0. 49 3.4 Combining RTI and UWB Information In this section we introduce three methods to combine the RSS and UWB data. We compare the results of the di erent methods in Section 3.7. 3.4.1 Image Combination by Product An RTI image is formed as described in Section 3.2.1 after every RSS sensor has transmitted a packet on all channels in F, i.e., after RSS measurements have been collected on all the links and channels. A UWB image is formed for every new CIR captured. In this method, the two images are combined to form the new image Lc by performing a voxel-wise product, Lc = lr ^ lu; (3.21) where lr = ^x from (3.9) and lu is from the UWB image Lu. We de ne MLc = max (Lc). When no person is located in the monitored area, MLc has a very low value. We use a threshold Te to avoid further processing images not showing the presence of a person in the target area: if MLc Te, we discard the current combined image and wait for the next one formed by the system. Otherwise, we normalize the values of the voxels of lr and lu such that their minimum value is zero and the sum of all voxels is one: [^l r]n = lr n PN i=1 lr i ; (3.22) and similarly for lu: [^l u]n = lu n PN i=1 lu i : (3.23) The normalization brings the two images in the same range of values and weights them equally. The normalized combined image ^L c is calculated again by performing a voxel-wise product of ^l r and ^lu: ^L c = ^l r ^^l u: (3.24) The voxel-wise product is used because both images cover the same geographic region. If we consider the normalized values of the images as probabilities, the product of the two 50 values for each voxel pair would give the probability of the both UWB and RSS \\events\" occurring in that voxel. This can be done because the error in the estimates are statistically independent. The RSS and UWB data collected by the two systems are time stamped to allow synchronizing the two images. Images are formed at the same rate as the higher of the two sampling rates. In our case, since the UWB CIRs are sampled more frequently than each RTI cycle, a combined image is formed for each new UWB sample. This image will then be the combination of the most recently formed RTI with the new UWB image. From the normalized combined image, ^L c, the position of the person is estimated as: ^p = arg max n2N ^L c; (3.25) i.e., the person's position estimate is at the voxel n having the highest value. 3.4.2 Linear Inversion with UWB Data An alternative method to form a combined image is to modify the weight matrix W in (3.1) to include the UWB measurements in the inversion process. We de ne a new matrix WU as an M N matrix where M is the maximum value of k and N is the number of voxels of the image. The n-th column of WU represents the ideal vector of k if the target were located at voxel i. The vector yU is the estimated vector of k from the results of the forward-backward algorithm. Equation (3.1) then becomes: yR yU = WR WU x + nR nU (3.26) where the subscripts R and U correspond to the matrices derived from the RSS or UWB data, respectively. The inversion matrix is calculated as in (3.10) using the combined matrixWC. A combined localization image ^L c is then formed by multiplying the combined inversion matrix C to the combined RSS and UWB measurement matrix yc. The position of the person is estimated as in (3.25). 3.4.3 Estimating X by Using Y The third method we propose for combining the UWB and RTI images is to derive one coordinate of the position estimate of a target from each image. First, we estimate the target location from the RTI image formed as described in Section 3.2.1. The y-coordinate from this position estimate is then used to derive an x -coordinate from the UWB image, 51 which is calculated as described in 3.3.3. If the target location estimate from the RTI image is at coordinates (^xR; ^yR), we consider the row of the UWB image corresponding to ^yR. The target position estimate ^p is set at the voxel having the maximum value in that row, i.e., ^p = (^xU; ^yR). 3.5 Localization and Tracking The position estimate ^p is used for updating an already existing track of a person or for initiating a new one if the target area is empty. To this purpose, we use track con rmation and deletion rules [51]. If at time k the set of candidate tracks, Td, and the set of con rmed tracks, Tf , are both empty, the position estimate ^p(k) is used to start a new candidate track, which is added to Td. A candidate track becomes a con rmed track only if its position has been updated at least happ times in the last H formed images (happ H). If this condition is not ful lled, the candidate track is deleted. A circular gating area of radius ! is centered at the target's position estimate ^p. The radius ! is de ned as an integer multiple of the voxel width p. We de ne Tg (Tf [ Tc) as the set of tracks (either candidate or con rmed) located within the gating area. Only the tracks in Tg are considered. The con rmed tracks in Tg are given priority over the candidate tracks: the current position estimate is used to update the closest con rmed track. Otherwise, if no con rmed track exists, the current position estimate is used to update the closest candidate track. If the set Tg is empty, the current position estimate is used to start a new candidate track. By using the gating area, we avoid the position estimate of the person to have large sudden changes in correspondence of noisy RSS and CIR measurements from the two systems. 3.6 Experiments The rst experiment was conducted in a 27m2 study room on the second oor of the Warnock Engineering Building at the University of Utah. A total of 33 RSS sensors were placed outside of the room along two opposite walls, 17 on one side and 16 on the other. The sensors were 30:5 cm apart. Two UWB radios were placed on one of the two sides of the room where the RSS sensors were positioned. The UWB radios were 1 m apart. A person walked along a prede ned path six times, three times counterclockwise and three times clockwise. The person entered and exited the room in each of the six trials. With the help of a metronome and markings on the oor, the person walked at a constant speed of 0:5 m/s. Fig. 3.1 shows the setup of the tests carried out in the study room. 52 The second experiment was conducted in a 28m2 room of a motel in Salt Lake City, Utah. The layout of this room is described in Fig. 3.2. This time, ten RSS sensors were placed along each of the walls separating the room from the adjacent ones. Two UWB radios were placed outside one wall of the target room. The experiments were conducted with the UWB radios at two di erent distances, 0:9 m and 2:7 m apart. A person walked along a prede ned path at a constant speed of 0:5 m/s, entering and exiting the room each trial. There were no other rooms adjacent to the target room besides the two where sensors were placed. For the second experiment, a person walked the target path 18 times. Six of the trials were done with the UWB radios in con guration A and twelve in con guration B, represented by white stars and black stars, respectively, in Fig. 3.2. 3.7 Results The following results are derived from data collected empirically in the study room and hotel room. This data collection was described in Section 3.6. The methods described previously were applied to this data. In the case of the study room, there were at least 16 RSS-based radios, or RSS sensors, available on each side of the room. There were 10 available on each side of the hotel room. To better understand how the performance varied with the number of available RSS sensors, the methods described were applied to the data multiple times, each time using a subset of the collected data. The results were averaged for a given number of RSS sensors. This was done to simulate the performance of the system using a fewer number of radios than were actually used. All simulations in the following results are performed in this manner (i.e., using subsets of the available empirical data). Performance is measured by the root mean square (RMS) error of the target's location estimate relative to the true location in units of meters. For AB-RTI and HMM-UWB, calibration measurements are required. For VB-RTI and VB-UWB, no calibration mea- surements are required. 3.7.1 Study Room For the rst experiment, 50 simulations were run using randomly selected subsets of S RSS sensors available on each side of the room. The density of sensors on each side of the target room is higher than what would be used in a typical deployment. Subset sizes for these simulations ranged from 3 to 10 sensors per side. The same subset of sensors was used for each of the six trials and remained the same when UWB radio data was included for a given simulation. The gating algorithm described in Section 3.5 was applied in all simulations. Simulations were performed using AB-RTI, AB-RTI with HMM-UWB, 53 VB-RTI, and VB-RTI with VB-UWB. Fig. 3.3 shows the mean RMS localization error for each of the methods used. Each point on the gure is the error averaged over the 50 simulations and 6 trials, measured using S sensors. The Y-axis error improves signi cantly with each additional sensor used on each side of the room. There is also little improvement in the Y-axis error as a result of including the UWB information. Variance-based methods show improvement in reducing Y error over attenuation-based methods. The X-axis error improves as a result of including more RSS sensors on each side of the room but not as greatly as does the Y-axis error. The improvement as a result of including UWB information, however, is much more signi cant and is also almost constant with the number of RSS sensors. The localization error, that is, the Euclidean distance (L2), improves overall by 51 cm and 33 cm, on average, for attenuation and variance-based methods, respectively. For comparison, if a point in the room is selected at random at each time, the RMS L2 error is 2.94 m on average over the 6 trials. Errors for the X and Y axes by selecting random locations are 1.65m and 2.44 m, respectively. The tracking algorithm is not applied when using random coordinates. 3.7.2 Hotel Room For the second experiment, 50 simulations were also run using randomly selected subsets of S RSS sensors on each side of the room for each simulation. When S = 10, however, only one simulation was performed because there was only one possible combination of S = 10 radios per side. For each simulation, localization was performed using AB-RTI, AB-RTI with HMM-UWB, VB-RTI, and VB-RTI with VB-UWB. The tracking algorithm described in Section 3.5 was also applied to each of these methods. Fig. 3.4 shows, from left to right, the L2, X, and Y errors when applying these four methods to the data collected over the 18 trials performed in the motel room. The reason the Y error degrades when including VB-UWB to VB-RTI is that VB-RTI gave noisier range estimates than HMM-UWB did. This is due to the greater signal attenuation in the hotel versus the study room and the additional environmental variations of furnishings. One noticeable di erence between the results of the two experiments is that the Y error in the second experiment decreases signi cantly by including VB-UWB with VB-RTI whereas for the rst experiment the Y error was e ectively the same. Generally, however, the same trends are visible in the results for the second experiment. The Y error improves with increasing S, and including UWB data signi cantly improves X error. 54 For the second experiment, the error using 10 sensors per side is higher than the error using 7 sensors, in many cases. There were only 10 sensors on each side of the room and, therefore, only one unique simulation could be performed using 10 sensors. By performing many simulations using subsets of the available sensors, the e ect of sensor placement on localization error could be minimized. This was not possible in the case where S = 10 for the second experiment. Table 3.3 shows the mean RMS error over the 18 trials performed for this experiment using all 20 RSS sensors. For comparison and as an estimate of the upper bound on error for a given environment and target path, random image coordinates are selected as the target location estimate. At each time when a combined image would be formed, X and Y coordinates are randomly selected and are used as the location estimate at that time. The gating algorithm described in Section 3.5 is not applied when randomly choosing location estimates. The results from applying the methods described in Sections 3.4.2 and 3.4.3 are also given in Table 3.3. Note in Table 3.3 that when performing localization using AB-RTI or VB-RTI, the X-axis error is about the same as that obtained from randomly guessing an X coordinate for each image. This is critically important for tactical operations. Having some knowledge about the person's coordinate in each axis is essential for law enforcement personnel to be able to make tactical decisions. Of the three combination methods described in Section 3.4, the image product method proposed in Section 3.4.1 performed the best. 3.7.3 Area of Uncertainty We de ne the area of uncertainty (AoU) as the ratio of the L2 mean squared error (MSE) to the total area of the monitored room: AoU = L2 MSE Room Area : (3.27) Table 3.4 shows the percent reduction in the AoU by adding UWB data to AB-RTI and VB-RTI for S = 3 and S = 10 sensors. The percent reduction in the AoU is signi cant except for VB-RTI in the motel room using 3 sensors. This may be due to the particular subsets of sensors used in the simulations when S = 3. The reduction in the AoU con rms that by adding UWB data the system can more accurately indicate to law enforcement personnel in which part of the room the person is located. 55 3.8 Conclusions In this work, we present a joint DFL system that uses the changes measured in RSS and UWB CIR to localize and track a person through walls. We target tactical operations and crisis situations where it is not possible for the police forces to place sensors on all sides of the area to be monitored. Experimental results show that including UWB with RSS data signi cantly improves localization accuracy when RSS sensors are only available on two sides of the target area. Where RSS sensors have been placed along the Y axis, improvements in accuracy along the X axis by including UWB data are especially signi cant. Without including UWB data, the accuracy along the X axis can be as bad as randomly guessing an X coordinate. We introduce three methods to combine the information from the UWB and RSS systems and we compare their performance. The multichannel variance-based RTI method proposed in this work, which does not require an initial calibration in stationary conditions, is as e ective or more e ective than attenuation-based RTI for through-wall localization. The improvements in localization accuracy and the reduction in the AoU demonstrate that UWB data should be included in a DFL system for tactical operations where RSS sensors may only be placed on two sides of a room. 56 Figure 3.1. Layout of the study room located in the Warnock Engineering Building at the University of Utah used for the experiments. Xs represent the 33 RSS sensors. Stars represent the 2 UWB radios. Circles represent the steps taken by the person at one second intervals. Grey rectangles represent furniture. The target room's inner dimensions are 3:82 m by 5:49 m (21 m2 area). Figure 3.2. Layout of the room of a motel located in Salt Lake City, Utah. Xs represent the RSS sensors. White and black stars represent the UWB radios in con gurations A and B, respectively. Circles represent the steps taken by the person at one second intervals. Grey rectangles represent furniture. The target room's inner dimensions are 3.96 m by 7.11 m (28 m2 area). 57 Table 3.1. RTI image estimation parameters Description Parameter Value Voxel width [m] p 0.15 Ellipse excess path length [m] 0.02 Voxels' variance [dB] 2x 0.05 Noise standard deviation [dB] N 1 Voxels' correlation distance c 4 Number of selected channels m 3 Short-term RSS variance window Ns 5 Long-term RSS mean window N 50 Empty area intensity threshold Te 0.05 Number of updates for con rmation happ 8 Con rmation window H 15 Gating area radius [m] ! 1.2 Table 3.2. UWB estimation parameters Description Parameter Value Voxel width [m] p 0.15 Sampling Period [ns] T 1 Short-term CIR variance window NU 5 Variance normalization parameter 1/NU 58 Figure 3.3. From left to right, the mean RMS L2, X, and Y errors over the 6 trials and 50 simulations using random subsets of S sensors per side of the study room. 59 Figure 3.4. From left to right, the mean RMS L2, X, and Y errors over the 18 trials and 50 simulations using random subsets of S sensors per side of the motel room. When S = 10, only 1 simulation was performed. 60 Table 3.3. Mean RMS localization error for the second experiment over all 18 trials for the methods described. Gating was used for all methods except random selection. Units given in meters. Random AB-RTI AB-RTI & HMM-UWB VB-RTI VB-RTI & VB-UWB Inversion with UWB X from Y L2 3.31 2.10 1.59 2.07 1.91 1.84 1.76 X 1.53 1.54 0.73 1.61 1.16 1.31 0.98 Y 2.94 1.42 1.41 1.30 1.51 1.28 1.44 61 Table 3.4. Percent reduction of AoU by including UWB data. Study Room Motel Room AB-RTI VB-RTI AB-RTI VB-RTI S = 3 40.2% 32.4% 26.3% 0.2% S = 10 61.8% 43.2% 41.3% 14.9% CHAPTER 4 PARALLEL MULTICHANNEL TRANSMISSION FOR RSS-BASED RADIO TOMOGRAPHY 4.1 Introduction Device-free localization using RTI uses radios surrounding a target environment to sense and locate people within the environment. These radios take RSS measurements as sensor data. These measurements can be used to estimate the location of a person or multiple people within the environment [4]{[7]. A number of models have been developed that describe how a person a ects RSS measurements depending on the location of the person and depending on the RSS levels normally measured when no person is present [4], [5], [8]{[10].1 Typically the method for collecting RSS measurements employs a round-robin or time division multiple access (TDMA) protocol [4], [5], [8]{[10], sometimes called spin [52] or multispin [10] which uses multiple channels. This protocol is for one radio to transmit while all other radios receive. Once the transmission ends, the next radio transmits while all others receive. This TDMA protocol continues until all radios get a turn to transmit, at which point the process repeats. The number of turns, or time slots, needed to complete one cycle of spin or multispin is N or NM, respectively, where N is the number of radios and M is the number of channels selected for measuring. The radios in this work use about 3.5 ms per time slot. For a typical network with 25 radios using 4 channels, one complete cycle of multispin 1Reprinted, with permission, from M. McCracken, M. Bocca, and N. Patwari. \\Selection of Links in Multichannel RSS Measurements for Radio Tomography\", (to be submitted to arxiv.org and to be included in \\Large Scale, Device Free Localization\" by the A. Luong, M. McCracken, M. Bocca, and N. Patwari to be submitted to The 12th International Conference on Mobile Systems, Applications, and Services (MobiSys 2014)). 63 would require approximately 350 ms while spin, which uses only one channel, would require approximately 87.5 ms. It has been shown that multispin achieves a lower localization error than does spin, which uses just one channel [10]. A trade o is that the latency of location estimates is larger for multispin due to the larger number of measurements that are required. The spin and multispin protocols are illustrated in Figs. 4.1 and 4.2, respectively. In these examples there are 8 radios in the sensor network. Each row represents a single radio. Each column represents a time slot in the protocol. The transmitting radio is represented with a \\T\" and all other radios in that time slot are receiving. The transmitting or receiving channel for each radio at each time slot is represented by the color. Once each protocol ends, it repeats from the beginning. The links between the transmitter and the receivers in the spin and multispin protocols can be visualized spatially as shown in Fig. 4.3. The information transmitted in each transmission packet is not critical to the mea- surement of RSS values. Any packet transmission can be used to measure RSS. The transmission, however, is leveraged as a method to report received RSS measurements from all other radios to a central location where processing is performed. Packet length is proportional to the number of nodes, N, in the network. The duration of each time slot must be greater than a packet duration, so the slot duration is also O(N). Longer packets require longer transmission times, which necessarily makes the time required to complete one cycle of spin or multispin longer. Energy consumption could be measured as the energy required by the sensor network to produce one localization estimate. By this measure, the energy consumption is proportional to the number of time slots. In this work, a Texas Instruments CC2531 system-on-chip is used for each radio. The power required during either a transmission or reception is approximately 100 mW. Multispin, in this case, requires M times more energy than spin to produce a localization estimate. The central question answered in this chapter is the following: \\In one slot, is it best to measure every link from one transmitter, or is it better to allow multiple transmitters to operate in parallel, even though doing so reduces the number of link-channels measured?\" We show that using parallel transmitters can be used to dramatically improve accuracy compared to spin for the same latency and energy consumption, or to dramatically reduce either energy consumption or latency for similar accuracy compared to multispin. This is achieved by modifying the protocol, or sensor sampling sc"}]},"highlighting":{"196231":{"ocr_t":[]}}}