Research On Object Matching With Multiple Cameras Based On Features Fusion And SVM

With the development of information society, intelligent monitoring system has played an important role in all aspects of human life. Especially the multi-camera, multi-target tracking technologies, has become a hot topic at home and abroad. With respect to the single camera target tracking, whos’ technology is relative mature, realization of multi-camera target tracking technology is more difficult. Target matching plays a connecting role in tracking process, which makes the pedestrian object matching a key technology of intelligent surveillance system. Improving the efficiency and accuracy of target matching become one of the hotspots of intelligent monitoring.In this thesis, the local features and color characteristics of the target are fused in the pedestrians target matching. The main contents and innovations are as follows:(1) HOG feature is selected to achieve the pedestrians target matching. To address the problem of time-consuming in HOG feature extraction, two improvements are made. Firstly, the HOG feature based on sub-cell structure is proposed, which improves the tri-linear interpolation process by effective reducing the complexity of tri-linear interpolation, and ensures the effectiveness of the features. In addition, G-HOG feature scheme is presented for Color image use.(2) Research on the object matching based on the fusion of local features and global features. In order to meet the needs of different environments under multi-camera, the method to fuse multi-scale Multi Level-HOG feature and color coherence vector with spatial information is discussed, which improves the robustness and accuracy of the object matching.(3) Research on the object matching based on BPSO and SVM combined optimization. To address the problem of high dimensionality after feature fusion, support vector machine and binary PSO are combined to make feature selection. Furthermore, we propose a multiple feature selection algorithm. Firstly, make feature selection to HOG feature and color feature respectively and obtain two binary feature mask respectively. Then fuse the subset of HOG feature and color feature and make feature selection to the fusion feature again. Improve the efficiency furthermore.The thesis presents detailed simulation analysis to verify the correctness of the proposed algorithms. Finally, it summarizes the work of this study, and gives the prospect of further research.