Multi-target Tracking Based On Trajectory Optimization

Multi-moving object detection and tracking is receiving increasing concern in the field of computer vision. It combines the knowledge of computer applications, image processing, artificial intelligence, pattern recognition, mathematics and other related fields. Now it has caused high attention of the majority of scientists, the national related departments and enterprises, and has now been widely used in video surveillance, medical image analysis, intelligent transportation and other fields. However, the research of moving object detection and tracking still faces a great quantity of challenging issues because of the complex background, diverse shapes, mutual occlusion of many objects and so on.In this paper, we first analysis and make a comparison of the commonly used target detection and tracking methods, then aimed at the main problems of multi-moving target detection and tracking, we carry on further research. The main content of article is as followed:1. We propose an improved target detection algorithm which is detection of secondary. First of all, find the target area by VIBE algorithm which is optimized by Otsu guidelines. After that, we accurately segment the objects by the SVM classifier based on HOG feature in foreground regions which has been extracted by VIBE. The experiment results show that this algorithm can effectively solve the non-adaptive parameters and objectives adhesion problems which VIBE algorithm can not solve. The algorithm has good robustness, and can accurately detect moving objects.2. In this paper, we propose a new multi-target tracking algorithm which based on trajectory optimization. First, using the location information of each moving object which is obtained by detection algorithm, we establish the initial trajectory of each target. Second, we build an energy function that can describe the current trajectory space in details. Next, based on the premise that the total energy does not increase, we continue to optimize the trajectory space until generating an ultimate accurate trajectory. Finally, we mark each frame to complete the multi-target tracking. Experiment results shows that this method can solve the problems existing in the process of moving target tracking, such as shelter, merge, miss, etc., At the same time, it keep track of multi-target, and has a high tracking accuracy.