Tracking Based On Sub-Block Color Histogram,Hog And Particle Filter

With the increasing pressure of security and traffic management, more and more surveillance cameras are being installed at roads, stations and other public zones. Currently most of the video surveillance system requires human monitoring. Due to the fact that the surveillance cameras heavily outnumbered the management personnel, lots of videos have been overlooked, lots of information can't be utilized effectively. Using computer vision technology to process these surveillance videos will enable machine to automatically detect abnormal incidents. Making machines intelligent is the trend of the future. In the effort to realize such system, detection and tracking in videos are two key technologies. In this article, a tracking method based on sub-block color histogram and particle filter is presented. Also it tries to melt the HOG feature plus SVM detection into PF framework.Firstly, several mature tracking methods and their application fields are introduced. These methods include Gaussian background modeling, optic-flow, feature-clustering and mean-shift. Then improvement is made to the traditional color histogram. A novel sub-block color histogram as the feature of PF tracking is presented. According to the new sub-block color histogram, the weight calculation and template update method in PF framework are improved. Sub-block color histogram preserves the robustness to objects' minor change in shape as traditional color histogram but with better ability to recognize the objects with similar color. Finally HOG feature and SVM method are introduced and melt into tracking framework. When system finds that the tracking result is wrong it will perform a local mean shift search to further increase the accuracy in pedestrian tracking.The presented system above is realized with C++ code. HOG pedestrian model is trained on INRIA human database and integrated into the program. LIBSVM SDK is used in SVM training. Experiments are conducted in different scenes and videos which prove the presented method outperforms the traditional method.