Tracking Multiple-person using Sparse Stereo Information

Abstract

In this study, we address the problem of multi-person detection and tracking in challenging scenes using sparse stereo information.In each frame, only a sparse set of object feature points are extracted.All these feature points are then projected onto a plan-view map, and grouped into several clusters by employing the biometric information, the optical flow information of object feature points, as well as the width of a person.By producing clusters, the location of a possible person can be determined.In addition, a Modified Joint Probabilistic Data Association Filter (MJPDAF) is proposed for improving the performance of measurements association during the people tracking process.Compared to the traditional JPDAF, the methods for the construction of the validation matrix and the calculation of association probabilities are improved.Experiments on challenging datasets demonstrate that the proposed algorithm is robust for people detection and tracking through fixed stereo vision.