Human Abnormal Behavior Recognition Algorithm Research And Implementation

Intelligent monitoring system is based on image processing, artificial intelligence and pattern recognition technologies. The studies of this system have the following aspects: extraction of moving foreground; target segmentation and identification; target tracking; behavior identification and classification. In this dissertation, the target is human body, and the author mainly researches the identification of abnormal behavior.A brief introduction of the major digital image processing techniques related to the entire monitoring system is introduced. Because the results of the traditional de-noising methods are not satisfactory, adaptive Gaussian background updating model is used to extract the foreground target, and presents an improving background difference algorithm based on three-channel separation.And then, abnormal behavior of human body in videos is studied. According to different scene, it can be divided to three situations: single-person's behavior; violent behavior among several persons and clustering incident. Because there is different feature information between different images, different solutions are adopted.Towards single-person, close-camera is adopted to extract target's external structure and geometrical features. Due to computer's strong learning ability, support vector machine theory is used to train sample videos, then classifies behaviors and identifies them.Towards abnormal behavior among several persons, fighting video between two persons is selected. At first, two motion features must be obtained: velocity and the change-rate of direction-angle. Target's dramatic and irregular action can be reflected by these features. This dissertation takes them as kinetic energy of image. On the other hand, abnormal behavior is more likely to happen if there is a shorter distance between them. According to gravitation theory, potential energy of image can be obtained. The sum of these two parts is the final energy function, and it is used to determine whether there is abnormal behavior or not.In square environment, camera monitors a wider range, human body's feature is not very clear. Based on Markov theory, every pixel in an image is considered as a point in Markov Random Field, from this, time-filed and space operator can be obtained. The MRF method based on time-space can be used to determine clustering incident.