Research Of Human Abnormal Behavior Detection In Surveillance Video

Anomaly detection is an important research direction in the field of artificial intelligence. The detection of human abnormal behavior in Surveillance video can effectively reduce the cost of artificial detection and timely response to emergencies, so it has very extensive application scenario in the field of security, transportation, etc. Now, the identification and detection of human abnormal behavior are attracting more and more attention from researchers.The thesis focuses on the study of human abnormal behavior detection methods in surveillance video, so it is necessary to detect the moving body and filter out the noise firstly.Firstly, the thesis expounds the current commonly used moving object detection methods, and analyses the advantages and disadvantages of these methods. As a result, we choose the method of background subtraction combined with background modeling to detect moving body. In the aspect of background modeling, we finally choose Gaussian Mixed Model after comparing the several current commonly used background modeling methods.Secondly, in the study of human abnormal behavior detection, we use Minimum Enclosing Rectangle to indicate the moving human body, use gray-gravity of the Minimum Enclosing Rectangle to indicate the gravity center of moving human body, and then extract several morphological characteristics those can express motion behaviors of human body, define some characteristic values to represent these morphological characteristics, so can we quantize the motion behaviors of human body. As a result, it is feasible to determine whether the behavior of moving human body is abnormal or not through the change of these characteristic values. In the fall detection of human body, we use the rate of the height of gravity center of human body change and the height-length ratio of the Minimum Enclosing Rectangle as judgment standards. In the run detection of human body, we use the distance of human gravity center between adjacent five frames as judgment standards. In the squat detection of human body, we use the same characteristic values as in the fall detection.Finally, in order to validate the performance of the algorithms of human abnormal behavior detection proposed in this thesis, we did some tests in Matlab213 a programming environment. The good experimental results prove that the three algorithms of human abnormal behavior detection proposed in this thesis are feasible.