The multi-target tracking in complex scenes is one of the key issues of the computer vision field. There is very board applications in both civil and military field. However, in the video surveillance systems by monocular stationary cameras, occlusions are common phenomenon because of the viewing angle and other reasons. It significantly impacts the accuracy of target tracking algorithms, even seriously affects the application of video surveillance systems. Therefore, how to effectively solve the occlusions in complex scenes, especially short-term occlusions caused by interactive non-rigid targets, becomes one of the most urgent problems in multi-target tracking. This thesis focuses on the occlusion handling algorithms and applications in multi-target tracking in complex scenes. This thesis takes the occlusion handling as the classification of foreground pixels: Firstly, the traditional pixel-based background model is improved, and an optical flow based background modeling approach is proposedto extract both foreground and their motion information; Secondly, two specific occlusion handling algorithms are proposed ——stepwise KNN occlusion segmentation and skeleton points assignment based occlusion segmentation; Finally, shortly occluded targets are compensated by the historical data, and object tracking is handled by a particle filter-based tracking framework. Experiments show that more accurate target information can be obtained by occlusion segmentation and compensation, which effectively reduces the mistakes in the process of multi-target tracking and enhances the tracking robustness. In addition, an embedded video surveillance system based on IT’s DSP DM642 has been developed. After porting and optimization, the occlusion handling and multi-target tracking have been implemented in embedded environment and been widely used in intelligent security field, generating practical application values. The mainly work and contribution of this thesis are as follows: 1. In the aspect of occlusion theory of multi-target tracking in complex scenes. A formal expression of occlusion handling based on foreground classification has been proposed. With Bayesian theory and the comprehensive utilization of appearance, color, movement, location and other information, the likelihood between foreground pixel and each target are constructed. Thus, occlusion segmentation could be completed base on the MAP criteria. 2. In the aspect of occlusion handling algorithm des...
修改评论