A Tracking-Learning-Detection (TLD) method with local binary pattern improved
Chun-Xiao Jia, Zhongli Wang, Xian Wei Wu, Baigen Cai, Zhenhui Huang, Guiguo Wang, Tianbai Zhang, Dezhong Tong · 2015
Tracking-Learning-Detection (TLD) is an excellent visual tracking method, it decomposes the long-term tracking into three sub-tasks: tracking, learning and detecting. Each sub-task is addressed by a single component and operates simultaneously, all three sub-tasks are unified in a tracking-learning-detection framework. But our experiments show that it is sensitive to the illumination changing. In this paper, we try to improve its performance in such case by enhancing the nearest neighbor (NN) classifier with Local Binary Pattern (LBP) algorithm. The modified NN classifier can get the bounding boxes which are closer to the tracking target. Moreover, the LBP algorithm has good performance on texture feature, so when the target has the good property of texture feature, the modified NN classifier has better performance. So a distinguish module is designed to select the right classifier. The experiments show that compared with conventional TLD algorithms, the proposed modification can improve the accuracy rate and robustness of the tracking results.