Moving object detection and tracking are the two fundamental problems in computer vision. They have broad application prospects in video surveillance, human-computer interface, intelligent transportation and motion analysis, etc. Although with the efforts on the research for many years, the technology of moving object detection and tracking has made great progress, there are still lots of problems to achieve robust moving object detection and tracking in complex scenes. In this thesis, we study the moving object detection and tracking in complex scenes and focus on to solve the problems of illumination changes, cluttered background, complex motion, etc. We have made progress in complex scene modeling, appearance modeling, tracking strategy and severe blurred object tracking. The main contributions of this thesis include the following issues: 1 We propose an online and unsupervised foreground segmentation algorithm in the maximum a posterior-Markov random field framework to improve the performance of background modeling based moving object detection in complex scenes. In this framework, the foreground segmentation problem can be considered as a labeling problem, then it is converted to be an energy minimization problem. Color, locality, temporal coherence and spatial consistency are fused together to improve the accuracy and robustness. We also propose new online and unsupervised learning techniques to learn the distributions of background and foreground pixels based on these features. The energy minimization problem is solved by the max-flow algorithm, which guarantees the result is global optimal. The experiments also demonstrate that our method can reduce the noise and holes in the foreground segmentation result. And the robustness of the algorithm in dynamic scenes is also promoted. 2 Inspired by the semi-supervised interactive segmentation algorithm, we propose a two-stage optimization for the foreground segmentation. First reliable foreground and background pixels are extracted by using the color and motion information. These pixels can provide reliable constraints for the next stage of foreground extraction. In the maximum a posterior-Markov random field framework, accurate foreground region are extracted by using the interactive graph cut algorithm. Our method absorbs the advantages of multi-stage filtering and the weakly learning algorithm, thus the accuracy and robustness are promoted dramatically. 3 To solve the problems o...
修改评论