In the present world, unsafe and unharmonious factors seriously threaten the safeties of country, society and people. Video surveillance system is one of the main technical means available to effectively prevent and eliminate hidden danger in fields of public security by real-time monitoring. However, traditional video surveillance system has some shortcomings and is unable to cope with complex scenes and behaviors. So it is imperative to develop intelligent video surveillance system (IVSS) which is based on the techniques of behavior recognition. In this thesis, we focus on key techniques of behavior recognition based on video, including moving object detection, static object detection and object tracking. Based on analysis of the techniques, a preliminary behavior recognition system can be built. The main contributions of this thesis are summarized as: A method of kernel density estimation for adaptive motion detection (AKDE) is presented. To begin with, an approach for adaptive selecting thresholds of foreground and background was proposed by analyzing the probability histogram to classify pixels. Meanwhile, a background model updated according to probability is also provided. The background model of inter-frame difference incorporated with results of AKDE can solve deadlock problem in updating background model. It is also used to detect abruptly changed background and update background model. The gray, color and texture information are used to eliminate the disturbance of shadow, which improves performance of the algorithm. A method for static object detection is proposed by combining motion and statistical features. The static object here refers to the object in relatively static or in tiny movement. Compared with object detection method based on background subtraction, our method succeeds in overcoming the problem of detection objects when they are in long time static. Compared with object detection method based on shape, the proposed method is independent of samples and achieves real-time performance. The image of inter-frame difference, which is formed by one image subtracted form the image with a pixel offset in row and column, is used to obtain the motion feature. The statistic feature of the target region and candidate region is used to detect the target. The template of the target is updated according to motion feature and the similarity between target and candidate region. In order to improve the real-time performance of the algorithm, int...
修改评论