An improved tracking-learning-detection method

As a novel tracking framework that explicitly decomposes the long-term tracking task into tracking, learning and detection, TLD still exists some drawbacks and challenges that have to be addressed in order to get a more reliable and general system, such as the manual initialization of tracking region and the bad adaptation in case of full out-of-plane rotation and strong deformation. In this paper, we put forward a framework of motion detection and recognition to solve the manual initialization problem. In addition, the components of the tracking points of original tracker have been transformed into partial ORB feature points at more reliable position, which could also develop the performance of detector and learning in turn. Experiments show that the improved TLD achieves higher precision, especially for out-of-plane rotation and strong deformation.