Experimental results from using a rank and fuse approach for multi-target tracking in CCTV surveillance

We study a novel approach to the problem of fusion of sensory information in tracking multiple targets in CCTV surveillance video. The approach, called "rank and fuse" (RAF) is based on multiple feature ranking and merging as opposed to a more typical combination of all scores (similarity or probability) in a single ranking. This has the advantages of low computational complexity, easy scalability to multiple features, and low-latency. Experimental results are presented to illustrate two aspects of the RAF approach for a "difficult" example from CCTV surveillance: the advantage of rank versus score combination, and the use of the rank versus score curve to decide which features to fuse.