A multi-object tracking system for surveillance video analysis

We present a novel and robust clustering based multi-object tracking system for surveillance video analysis. It is designed to extract the trajectory data of vehicles in crowded traffic scenes and can be extended to other applications of surveillance and sports video analysis. In our system, a fast accurate fuzzy clustering algorithm is employed, and the feature space is constructed by extracting the position, color and velocity information of foreground pixels. By using growing and predictive adaptation, fixed linkages are expected between meaningful targets and corresponding active cluster centroids. In this way the motion classifier and tracker are combined seamlessly. Experimental results suggest the efficiency and robustness of the proposed method with severe occlusions and clutter effect.