Object Tracking for Multiple Non-overlapping Cameras Based on TLD Framework

Abstract

Wide-area video surveillance system can track object in a wide range. However, environments of different cameras are generally quite different, so object handoff and data integration among cameras are research difficulties. To solve these problems, a kind of non-overlapping multi-camera tracking system based on TLD framework is proposed. TLD framework maintains a unified sample classifier which uses affine transformation to generate new samples and updates classifier parameters through online learning which fuses data among cameras. To achieve object handoff, detection module scans video frame of a certain range of the camera. Then object result is obtained by comparing the similarity. Tracking module of origin framework uses optical flow based on feature points. In order to enhance tracking robustness of framework in complex environment, MeanShift and particle filter tracking algorithm based on color feature are used to replace of it and experiments are carried out. Experimental results show that system can achieve continuous tracking in non-overlapping multi-camera and the TLD framework based on MeanShift tracking algorithm has better robust and accuracy.