Research Of Several Detection And Tracking Algorithms In Intelligent Video Surveillance System

With development of informantion and computer hardware-software technologies, video surveillance system has revealed its importance in the urbanization process. The range of its applications includes various fields such as politics, military, culture, finance and technology. Specially, the application of computer vision algorithms in surveillance system is a crucial premise of its intellectualization. It is also the main reason of the superiority of video-based surveillance system compared with other surveillance system, which has shown higher cost performance and more appropriateness for modern development. In conclusion, the research to computer vision in surveillance system has very important significance.In this paper, we choose several most representative intelligent applications in video surveillance field and carry on research about the algorithms to solve with them. It includes three aspects:fire detection, head detection and human upper body detection and segmentation, objects tracking. In details:1. In the field of fire detection, we propose a novel video fire detection method based on artificial neural network. Except for analyzing fire’s motion and color features, the proposed method researches and utilizes tempral and special features such as fire’s flickering frequency and geometry. All these extracted features are fed into an artificial neural network and the network outputs an integrated result. Moreover, we propose GPU based fast Fourier transformation algorithm. The proposed method can distinguish between flickering vehicle light and real fire, which results higher detection rate.2. In the field of head detection and human upper body detection and segmentation, we propose histogram of oriented gradient and shape2D histogram features respectively. On that basis, for head detection, we further propose filter method of motion and appearance likelihoods based on Bayesian theory. For human upper body detection and segmentation, we further propose foreground segmentation method combined background subtraction and energy function optimization. Moreover, we design a GPU acceleration algorithm based on CUDA in computing HOG feature. The proposed head detection method reduce false positives effectively while proposed human upper body detection and segmentationi method can achieve extraction of upper body regions correctly.3. In the field of objects tracking, we propose an objects tracking method based earth mover’s distance (EMD) and SURF feature points. We introduce the idea of reducing the problem of tracking objects with SURF feature points to the linear programming problem which solves EMD. Otherwise, we propose two phases tracking strategy, which means coarse-to-fine idea and the solution of multi-objects occlusion based on Bayesian framework. The proposed tracking method can locate multiple objects for a long time and achieve robustness and reliability.Experiments have proven that the three proposed methods are available and have excellent performance. We have integrated them into real surveillance systems and achieved applicable results.