An improved compressive tracker for multiple pedestriansin surveillance videos

This paper proposes an improved online framework based on Compressive Tracker (CT) for multiple pedestrian tracking in surveillance videos. The CT method proposed by Zhang et al was originally used for single object tracking, and fails to make use of context information during the tracking process. To overcome the crucial drawbacks of CT, our method implements multi-scale tracking and fuse CT with Kalman Filter to take advantage of the spatio-temporal context information. Additionally, incorporated with the detection of foreground blobs and an online learned detector, this paper introduces a supplementary mechanism to handle the inter-target occlusion. Experimental results on realistic sequences demonstrate the effectiveness of our approach.