Real-Time Target Tracking With Approximate Topological Isomorphism In Video

Video target tracking is one of the hottest topics in the field of computer vision, which has a wide application in military and civil areas. Thus, it attracts widespread attention of many researchers in China or at aboard. Researchers have put forward a lot of effective tracking algorithms, and the performance of tracking has been improved continuously. However, many problems still exist and need better solutions in practical applications. There are many influential factors especially under the circumstance of target nonlinear deformation, pose change, occlusion of foreground or background, overlapping, interference by complex background noise, etc. Therefore, it is still a challenging task to improve the robustness and accuracy of video tracking algorithm. In addition, the real-time performance of the tracking algorithm also needs to be taking into account.The thesis mainly researches on the following three innovative aspects of the video tracking algorithm: 1. According to the problem that the background is complex and hard to extract directly in target detection and tracking, the thesis presents a video target detection method based on background reconstructed dynamically. The method combines the background of dynamic reconstruction with the target direction and tracking, and makes the target detection and tracking based on the new background. This method solves the problem that foreground detection relies on the known background. Besides, the method strengths the anti-interference ability to slight jitter, background brightness changes and other complex background by doing morphological operations to background, and the method also improves accuracy and efficiency of the target detection.2. According to the problem of target overlap in video target tracking, the thesis presents a new target tracking method which combines topological relations of target color information. The topological relations of the the target color position are used as a kind of new target feature fused with other target features to track targets. The method can solve the problem effectively while background is partly similar to targets. By referring to the approximate topological isomorphism of sequence similarity, the method also can effectively solve the tracking error leaded by the overlap of part of the target color information. The thesis provides a new idea for this kind of feature fusion target tracking algorithm. In addition, establish a target decision model by combining multiple feature weak classifiers to a strong cascade classifier, can effectively solve the problem which often occurs that tracking cannot continue with the single feature is masking. The method improves the accuracy and efficiency of the video target tracking.3. According to the problem that the computation of traditional multiple moving target tracking method usually is very large, and traditional target tracking method cannot ensure the real-time performance. The thesis presents a distributed real-time method of of video multiple targets tracking, and designs a reasonable algorithm of distributed scheduling. The method tracks target by segmenting it into a number of sub targets according to its connectivity, and effectively reduces time consumption of tracking caused by target size or target number. The method can effectively improve the efficiency of the tracking of multiple moving targets, and can keep real-time.In conclusion, this thesis researches several important problems of moving target tracking in video sequence. These problems contain background reconstruction, target recognition with partly overlapped and real-time requirement. The thesis greatly improves tracking effectiveness by get rid of dependence of exist background in moving recognition method. Also, we apply background reconstruction to avoid negative influence such as complex background, quiver of imaging device, luminance variations, etc. Second, the thesis extracts approximate topological isomorphic information of blocks with single color as a feature, and fuses it with other features in target recognition. These multi-features report characteristics of moving target well and can avoid partly loss of features by overlap and other reasons. Lastly, in order to solve overtime problem, the study of distributed moving multi-targets tracking is researched. The tracking is processed by multi-computers dynamically and parallel, which is applied to achieve better computational time and to make our distributed multi-targets tracking method real-time.