3D Template-Based Single Camera Multiple Object Tracking

This paper describes a 3D template-based tracking method that allows simultaneous tracking of multiple objects of different type. The assumption is that movements of all objects are constrained to a ground plane so that the tracking functionality can be accomplished using just one properly calibrated camera. The tracking algorithm is based on particle filtering where each particle represents one hypothetical configuration of the scene. Tracked objects are modeled by instances of 3D templates, positions and dimensions of which are continually updated from frame to frame. A novel method for initialization of new objects has been developed that can easily be adopted in any particle-based approach, where the measurement step is done on the pixel level. Numerous experiments have been conducted that indicate the presented approach copes with occlusions in an efficient way. The results for scenes of various complexity also demonstrate that the real-time performance can be achieved on a standard PC.