System Of Fighting Detection Based On Video

Detection of fighting behaviors is an important topic in the field of security; a lot of scholars have conducted extensive study. It is mainly divided into two kinds: one is based on the color information(color, blood), sound information(scream, explosion) on the fight violence, there are certain limitations on the accuracy of the warning; The second kind is based on microcosmic thought, namely the use of human action library but the construction process is too complex, and it is too difficult to ensure accuracy. This paper focuses on fighting detection under the relatively open static background. The main work of this paper is as follows:(1) Human target detection. A simple count of the background model is proposed, that is, in a period of continuous video, through the pixel value of each point of statistics; the highest frequency pixel value is regarded as background pixel value of the current point. The main steps of targets detection: 1) With mode background model, we model the video image RGB three channel; 2) The video images of the three channel background cancellation, each point from difference in the three channel maximum value as the difference map(single channel) difference in the point; 3) The difference image of local threshold segmentation, finally get the accurate mass; 4) In order to improve the accuracy of detection, we establish three-dimensional fixed-point distance function using ground target. This method can realize monocular measurement, estimation of each target size. This method can eliminate the small animal scene and vehicle interference, finally it get accurate targets.(2) Moving object tracking. We are tracking moving object in real time with the centroid and area. In the process of tracking, we can calculate the velocity of the moving target, direction and acceleration etc.(3) Fighting detection. We analyze the characteristics of the human in the scene of fighting, such as the velocity of movement, the track of the movement, the characteristic of the position, the acceleration and so on. Through the above analysis, we design the classifier. We extract the characteristics of the characters in the middle of the fight and non fight video, and train the classifier parameters. The experimental results demonstrate that using the above method can get better result.This topic aim at the situation of the relative dispersion of moving objects.It has not yet involved the complex scenes such as the mutual occlusion of moving objects and the fast changing of light. Identification number there is a certain limit. This paper achieves a higher recognition rate, but the false alarm rate is somewhat high. These are the focus of continuing research in the future.