||Currently, few methods exist to accurately model a human motion inside a monitored area. Most of the approaches that exist depend on some kind of boolean data from sensors that tell the presence or absence of person a at a given instant of time near a particular sensor. Using that information, some systems can then track a person across the area at di erent timestamps. Furthermore, for most existing approaches, the accuracy drops rapidly as the number of persons in the image increases. The sensors used in such settings are usually expensive. Not much work has been done to build a similar system based on inexpensive radio sensors. As there is no way for our radio sensors to provide information as to whether a person is present at a location, we need to extract it from the data using computer vision and machine learning techniques. However, it is not easy in such a system to model the noise component accurately. Therefore, we provide a probabilistic model to decide whether a detected blob is noise or an actual person. In our work, we exploit the fact that images do not change by much between successive timeframes and use this to detect and track multiple persons in a monitored area with a reasonably high accuracy. We use location and count of persons in historical images, and their similarity with the current image to calculate the new locations and count.