Use Fukunaga-Koontz Transform to Solve Occlusion Problems in Multitarget Tracking

Abstract

For multitarget tracking problems, occlusions between targets are quite tough tasks. We present a novel algorithm to solve such problems. For the two targets in occlusions, Fukunaga-Koontz transform is exploited to achieve the projection matrix, with which the two targets are projected into a low dimensional space where they are quite distinguishing. To solve the problem of the change of target appearance, the eigenspace model is used as the probabilistic observation model, with which the algorithm can learn the changes of the target appearance online. These two procedures are evaluated in the particle filter based tracking framework. Experimental results demonstrated the effectiveness of our algorithm.