Research On Human Abnormal Behavior Detection Based On ARM9

Human behavior recognition under video surveillance is a comprehensive research topic with high theoretical and applied value, which involves many fields such as image processing, machine learning and artificial intelligence. The monitoring system not only greatly reduces the consumption of human resources in the traditional video monitoring, but also can effectively and quickly identify the abnormal situation,which has been concerned by the researchers at home and abroad.Design of video surveillance is in the ARM9 platform to achieve, a combination of hardware and software, the wireless network data transmission, server transplanted mjpg streamer, using its plug-in can complete the image acquisition, compression and display. Client is combined with the QT software and Open CV visual library to complete.Abnormal behavior analysis for outdoor scene will walk definition for normal behavior, in addition the sit-up, push-up, crawling as abnormal behavior.Moving object detection test is used to combine the traditional optical flow method with the frame difference method. The results show that the algorithm can greatly reduce the time of optical flow computation and improve the real-time performance of the system. Also in order to overcome the effects of illumination changes, shadows moving objects on the target detection, RGB space into HSV color space, and in the traditional shadow of the shadow detection method based on proposed an improved algorithm.Target tracking contrast the CAMSHIFT algorithm, LK optical flow method, and the template matching method, and analyze their respective advantages and disadvantages, finally choose the Kalman filtering method for target tracking, the filtering process is continuous prediction- correction recursive process. At the same time, the adaptive template update of Calman filter method can solve the problem of attitude change, object occlusion and so on.Feature selection plays a very important role in behavior recognition. After analyzing the advantages and disadvantages of different features, the local space-time features are selected as the features of human behavior recognition, which combines the advantages of static and dynamic features, and the robustness is strong. The3D-Harris and HOG/HOF detection algorithm to detect interest points, detectedinterest points are discretised into interest point of the word, by clustering algorithm for constructing a codebook, namely "bag of words". At the same time to build statistical word histogram generated feature vector and the input to the classifier for classification.Experimental results show that the algorithm can be used to classify the 4 kinds of behavior in video, get the behavior label, can detect the abnormal behavior, the paper also gives the algorithm of detection accuracy and recognition efficiency.