Simultaneous Visual Recognition and Tracking Based on Joint Decision and Estimation

Visual target tracking and recognition have been increasingly important in video surveillance. Conventional works deal with tracking and recognition as separate steps, whereas tracking and recognition are closely interrelated and can help each other potentially and significantly. To tackle this problem, based on the joint decision and estimation (JDE) model which guarantees the general decision (recognition) and estimation (tracking) arriving at the global optimization, a simultaneous visual recognition and tracking method is provided. Besides, the structured sparse representation (SSR) model shows great efficiency and robustness in exploiting both holistic and local information of the target appearance. We show that constructing the appearance model with SSR can improve the performance of the proposed algorithm. Then, the contribution of each test candidate is considered into the learning procedure by a kernel function. Furthermore, the new joint weights of the kernel function provide flexibility with appearance changes and thus robustness to the dynamic scene. The experimental results demonstrate that the proposed method performs well in terms of accuracy and robustness.