Study On Fighting Action Detection Technology Based On Surveillance Scene

Fighting action may infringe upon the personal safety and property security of citizens, and even lead to major social public safety problems. As the closely-watched front subjects of computer vision field,one of the difficulties is how quickly and effectively extracts the motion feature from fighting actions, and improve the detection speed and accuracy. Motion vector (MV) received from Block-Matching Algorithm is not only rich in motion information, such as trajectory, speed and orientation, etc.; but also has small variation influenced by the background, object color and noise. It has strong capacity to adapt the external environments and the light variation. This paper studies on how to extract features from the motion vectors to detect fighting action in surveillance scene. Work mainly includes the following aspects:1. The imaging height and angle will affect the size of the object imaging, resulting in the change of the number and magnitude of motion vectors, which will seriously affect the accuracy of characteristic parameters extracted from motion vectors. Focus on this problem, a method to normalize motion vector base on pinhole imaging model is presented in this paper. The method first reduces or enlarges the size of the image to reduce the difference in the size of object imaging caused by the camera installation height, and then based on pinhole imaging model, introduce a parameter named Normalization Factor (NF) to correct the magnitude deviation of motion vector, using the camera's pitch angle and view angle. Simulation experimental results show that the method can quickly and efficiently to amend the magnitude deviation of motion vector.2. Focus on the problems in fighting action feature extraction, which is slow, an method to extract motion feature of fighting action based on motion vector is presented in this paper. The differences in statistical properties of the motion vector between fighting action and normal behavior are analyzed firstly. And then we use adaptive threshold three frame difference method to segment motion region, and use an improved three-step search algorithm to computing motion vectors, and use Normalization Factor (NF) to correct the magnitude deviation of motion vector. Finally, we extract features from magnitude and orientation information of motion vector. Simulation experimental results show that this method can rapid exact motion features, has better robustness in the number of moving object in the scene, illumination change and noise, and so on, so that suits for indoor and outdoor surveillance environment.3. Based on fuzzy theory and artificial neural network, two judgment methods of fighting action is presented. One is the method based on fuzzy pattern recognition method. The method use Joint Gaussian Function to normalize the feature parameters, and use the weighted sum method to fusion feature parameters, and introduce the concept of Average Maximum Violence Index (AMVI) to judge fighting action in surveillance scene by a fixed threshold. Experiment results show that the method is efficient, fast, and can obtain higher detection accuracy. Another method is the method based on dynamic fuzzy neural network. This method uses motion vectors to track motion regions, and use dynamic fuzzy neural network to judge if there is fighting action in surveillance scene. Simulation experimental results show that the method can effectively reduce the workload in establishing motion feature model of fighting action, and further improve the detection accuracy.