Collaborative object tracking with motion similarity measure

This paper presents a new approach to tracking objects using motion similarity measure for pairs of objects, with an emphasis on collaborative tracking. Assuming the distribution of motion in a scene can be decomposed into multiple Gaussian distributions, the motion of each object can be computed and matched. An object sharing similar motion characteristics with the tracked object can be used as a prior belief. Hence, the problem of object tracking can be formulated as an estimation problem based on two components, namely self information and information from other objects based on a motion similarity measure. Most existing object-tracking approaches perform tracking based on only self information. Collaborative tracking uses both information in an optimal manner under a Bayesian framework. Experimental results show that the hypothesis of motion decomposition is valid in many real-world scenarios. Moreover, information from other objects based on a motion similarity measure is especially useful in tracking when the self information is not reliable or not available because of the occlusion of tracked object in the scene.