Abnormal Behavior Detection Of Moving Object In Video

Currently most monitoring systems only detect and track moving target. However,the purpose of monitoring is that when an exception occurs, the monitoring personnel can make timely decisions. In real life, manual monitoring results in a waste of human and financial resources since there are little chance of abnormal conditions. It is likely to cause unpredictable losses if the monitoring personnel misses alarm. Intelligent video surveillance system can detect abnormal conditions, and notify the relevant personnel timely, thus avoiding the waste of human and financial resources.In recent years, the proportion of elderly people is rising and the ratio between nursing staff and the patient's is imbalanced seriously, falls especially falls at night will bring a serious threat to people's health. If they cannot get timely help and rescue when falls occur, serious consequences are likely to cause. Therefore, we use video analysis technology to automatically detect the abnormal behavior – fall.In order to solve the low image recognition rate caused by lighting, occlusion and other factors in traditional video surveillance, this paper introduces the Kinect device.This paper mainly detects abnormal behavior in the small range. To solve the problem of low detection rate in single feature, this paper uses feature detected fusion method based on shape feature and model feature. In order to improve detection rate of abnormal behavior, this paper proposes an improved algorithm of dynamic adjustment learning factor weight of PSO to optimize the parameters of SVM classifier, and design an abnormal behavior detection system. The main contents of this paper is organized as follows:Firstly, research the working principle of Kinect, Kinect depth image principle,Kinect skeleton tracking and joints extraction principle in-depth. The experiment is designed to analyze the depth perception of Kinect, and the working environment is determined.Secondly, this paper using a method based on the height and skeletal tracking technology of data fusion joints feature based on Kinect depth data image. Because of the interference data with joint points of other acts in the data, it is not convenient for feature extraction. In order to solve this problem, it is proposed sliding window algorithm to locate abnormal behaviors(squat, bend and sit, fall). Meanwhile, It isextracted the moving object height changed sequence of abnormal behavior and the min height of the head and spine center from the ground and the speed of the head.Thirdly, in order to improve the detection rate of abnormal behavior in video surveillance system, the particle swarm optimization algorithm is adopted to optimize SVM parameters. Because of the shortcomings of the particle swarm optimization algorithm, such as low precision, easy precocious, excessive parameters, this paper proposes an improved algorithm of dynamic adjustment learning factor weight of PSO.It improves the accuracy and convergence rate of PSO algorithm.Finally, this paper uses the improved PSO algorithm to optimize the parameters of SVM classifier and use the mixture kernels function as the kernel function of SVM to design human behavior anomaly detection system.