Multiple Objects Tracking With Improved Sparse Representation and Rank Based Dynamic Estimation

This paper presents a novel multiple objects tracking method by constructing an improved tracklet affinity function to enhance the performance of multiple object tracking (MOT) within a network flow optimization framework. The aims of this paper are to improve the appearance affinity model and motion affinity model of the tracklet affinity function, both models being key factors to limit MOT performance in tracking by detection framework. In the proposed method, the tracklet association is considered as a generalized linear assignment problem relied on tracklets affinity that is calculated by both improved sparse representation appearance model and rank-based motion model. This paper uses the detection confidence to identify the important targets and, hence, to form the templates set. Then, a weighted target templates set is used for sparse representation of the appearance model. The rank-based motion model exploits motion estimation based on spatial information to interpolate the missing objects during tracklets association. Furthermore, we consider the inconsistency of two cues in different situations and set weights for them based on their properties to describe affinity between two tracklets. Our method has been evaluated on three challenging data sets, and the experimental results show that the proposed method has a good tracking performance compared with several state-of-the-art multi-object trackers.