| Description |
There has been much research on how to get unmanned aerial vehicles (UAVs) to perch on many different types of surfaces and objects, including flat surfaces, ramps, tree branches, power lines, etc. Many of these surfaces are static and it is easy to detect falls using inertial sensors such as accelerometers or gyroscopes. However, some perches, such as tree branches and power lines, are not static. When the UAV is perched on these perches, it will move with them, making the detection of falls from such a perch much more difficult than simply trying to sense motion. This thesis proposes two methods for fall detection of a UAV perched on such a dynamic perch. Computer vision is used on a feed from a camera mounted on the bottom of the UAV. Optical flow is used in conjunction with a filter that segments the perch in the image from the background to estimate the relative motion between the UAV and the perch. If the motion exceeds certain bounds, the UAV is considered falling. The second method tries to find the instantaneous center of rotation (ICR) of the UAV utilizing accelerometers and a gyroscope mounted to the UAV frame. Two methods are proposed to do this, one based on integrating the accelerometers to find the velocity at a point, the other finds the distance between the ICR and three points on the rigid frame of the UAV. The ICR estimates from these two methods are compared to an ICR estimate derived from data from an external Vicon motion capture system. The estimated ICR is then compared to the ICR of the perch that the UAV has perched on, if the two diverge enough, the perch is considered to be falling. The proposed methods were tested experimentally by placing a test quadrotor fitted with the appropriate sensors on three different test perches: a painted PVC pipe, a PVC pipe with a swirl pattern on it, and a tree branch. The quadrotor and perch are then actuated in three different tests. The first test has the quadrotor rotating about the perch while the perch is static. The second test has the quadrotor swinging on the perch without slipping. In the final test, the swing perch angle is increased until the quadrotor falls off. Each of the 9 tests is performed 5 times. Accelerometer, gyroscope, and vision data are gathered during these tests and analyzed using the methods described in this thesis. These experiments show that the vision method works fairly well, and that the ICR method works to a degree, but there is more work to be done in that area. |
