Research On Abnormal Human Behavior Detection In Surveillance Videos

With the rapid development of science and the continuous improvement of people's living standard, people gradually strengthen the awareness of their own safety. Intelligent monitoring equipments have already appeared in the lift, at the airport, in the bank, in the square or any other public places. Along with the development in the field of Computer Vision, the detection and analysis of abnormal behavior has become a hot spot in current academic research. It has a broad prospect in the future application. However, in the traditional video monitoring, the work needs to be done manually. The relevant staff must be trained and devote all the day to operate the machine, such as adjusting the monitor angle, capturing the abnormal images and issuing commands etc. It costs a lot of time and money. Thus, it is significent to precisely detect and analyse abnormal behavior of human body on the intelligent monitoring platform.This paper introduces the related theory and technology involved in the process of abnormal behavior detection. The abnormal behavior detection process can be roughly divided into three parts: moving target extraction, moving target tracking and analysis of behavior detection. In the part of moving target extraction, compared with the effect of foreground extraction and execution efficiency among different background modeling algorithm, this paper chooses the adaptive mixture gaussian model, which is proposed by Zivkovic. It can get the ideal foreground target. In the part of moving target tracking, aimed at the advantages and disadvantages of several kinds of tracking algorithm, this paper puts forward an improved method which uses Camshift combined with kalman filter method. This method ca n solve the following two problems effectively. One is the color similarity between the moving target and the background in large area. The other is the area overlap between the targets.This paper is mainly talking about the human abnormal behavior detect ion in the fewer interactive behaviors surveillance scenarios which divide into indoor and outdoor scenarios. In indoor scenarios, this paper puts forward the regional optical flow energy method using Horn&Schunck optical flow feature. It can detect the abnormal behavior by setting behavior threshold from the five kinds of behavior videos. It can improve the identification accuracy compared with the traditional optical flow energy. In outdoor scenarios, the paper uses the human geometrical features such as the body's centroid and external tectangular box, and the movment trajectory feature to detect human falling and wandering behavior. By testing different behavior types of videos, the results show that the method proposed in this paper can achieve the purpose of detecting abnormal behavior.