A P2P Camera System with New Consistent Labeling Method Involving Only Simple Geometric Operations

摘要

Consistent labeling is a fundamental problem in multiple-camera-multiple-human tracking. We present a new consistent labeling method called CPSPL that is used in our P2P human tracking system with multiple calibrated cameras, media servers, human trackers and trajectory mergers. The proposed method differs from most existent ones in that it only uses projective lines for establishing the correspondence between the tracked objects in multiple camera views, as well as correct the object location errors in a single camera view. In the world coordinates, the projective lines passing through the same person from separate cameras are often skew lines. We use the common perpendicular segments of projective lines to determine whether they pass through the same person. Although CPSPL method involves only simple geometric operations, the experiment results show its surprising accuracy in multiple-camera-multiple-human tracking.