A Multi-target Tracking Problem Of Occlusion

Moving target tracking is the core content in the area of computer vision. It has extensive practical value and broad application prospects. It combines advanced technology and research from multi-disciplinary, which include image processing, automatic control, pattern recognition, artificial intelligence and computer applications. Moving target tracking is a process, which automatically identifies a target, determines target location, and automatically tracks the target. Image signal loss of information collection process and complex application environments makes tracking become a challenging task. Especially in the tracking process, the occlusion is the key factor which limits the effects of tracking algorithm.This paper studies about the occlusion problem of the target tracking. This thesis is organized as follows:it begins with an overview of classic methods about moving target tracking and their basic principal. Then through the introduction of independent component analysis algorithm, we used the target partition based on improved independent component analysis in this paper; followed by introducing the application of data association in the multi-target tracking process, we proposed a kind of data association used in the multi-target tracking; At the end, by introducing greedy algorithms, creating target feature modeling and occlusion relationship model, we propose a greedy algorithm used in the target tracking process.By introducing a variety of related technologies and important theories, we elaborate the algorithm chapter by chapter, which consists of three parts:(1) Target detection:For the input image sequences, this thesis uses the improved independent component analysis techniques to detect moving foreground target, and to segment multiple target motion regions. At the end, the appearance of the model uses the movement bounding rectangle to represent the region.(2) Data association and target merging:This thesis uses the detected target to carry out data association and determines whether to merge target. When no occlusion occurs, the target trajectory is managed with respect to the results of data associated. And then the target features are updated. (3) Using greedy algorithm to determine the occluding target location:when occluding targets among targets, this paper uses a greedy algorithm to determine the specific location of each target in the foreground block.Based on the above algorithm, this paper has done an experiment, whose content is about the target tracking in the video occlusion. The simulation experiment used Visual Studio 2005 and OpenCV platform designing and developing multi-target occlusion tracking system. Experimental results showed that using the method can be effectively addressed in the occlusion problem in the target tracking process, improve the tracking results.