A Homography-Based Multiple-Camera Person-Tracking Algorithm
It is easy to install multiple inexpensive video surveillance cameras around an area. However, multiple-camera tracking is still a developing field. Surveil lance products that can be produced with multiple video cameras include cam era cueing, wide-area traffic analysis, tracking in the presence of occlusions, and tracking with in-scene entrances. All of these products require solving the consistent labelling problem. This means giving the same meta-target tracking label to all projections of a realworld target in the various cameras. This thesis covers the implementation and testing of a multiple-camera peopletracking algorithm. First, a shape-matching single-camera tracking algorithm was partially re-implemented so that it worked on test videos. The outputs of the single-camera trackers are the inputs of the multiple-camera tracker. The al gorithm finds the feet feature of each target: a pixel corresponding to a point on a ground plane directly below the target. Field of view lines are found and used to create initial meta-target associations. Meta-targets then drop a series of mark ers as they move, and from these a homography is calculated. The homographybased tracker then refines the list of meta-targets and creates new meta-targets as required. Testing shows that the algorithm solves the consistent labelling problem and requires few edge events as part of the learning process. The homography-based matcher was shown to completely overcome partial and full target occlusions in one of a pair of cameras.