Regression-based occlusion handling using silhouette data in video surveillance

Tins paper aims at addressing the occlusion problem, which is the challenging part iii a video tracking system. Our algorithm is composed of two steps: loca tion and recognition. In the location step, we first apply a non-linear regression after segnientation to model the human body silhouette. Afterwards, we identify peak points of the regressioli curve which are then used as candidate positions of occluded objects. hi the recognition step, we find the correct correspondence between candidate models and object color models stored prior to occlusion, so that we cast derive tile position of each person during occlusion. We have com pared our tracking method with two other popular tracking algorithms: meanshift and particle-filter. Experimental results show that tile correctness of our method is much lugher than the mean-shift algorithm and basically the same as a particle-filter, however, with the major benefit of being a factor of 10-20 faster in computing