Abnormal Crowd Behavior Detection Based On Video Surveillance

In recent years,the terrorist attack and critical group incident have challenged the bottom line of the city's defense constantly;the public security has caused widespread concern in the world.Traditional video monitoring mainly relies on manual monitoring,inefficient and wasteful in time.The effective measure to reduce personal injury and property damage in public places is the realization of intelligent video surveillance system.Therefore,it has important application value and far-reaching practical significance to deeply research the system of abnormal crowd behavior detection based on video monitoring.In the traditional methods of detecting the abnormal behavior of the small and medium crowd,the detected type of behavior and real-time cannot achieve balance.Starting from the perspective of machine learning,this paper proposed an abnormal crowd behavior recognition method based on the crowd feature of density and motion.This paper estimates the feature of crowd density and extracts the motion feature of crowd,finally,uses the random forests to complete the recognition of abnormal crowd behavior.Specific research work as follows:Firstly,estimate of crowd density.Extracting the the number of pixels updated by perspective correction algorithm in the range of ROI,using FAST corner density set the weight of population density to finish the foreground pixel normalization.then,Estimating population density in different weight range by respective least square fitting curve.This paper reduces misjudgment by eliminating the situation of Motion vector's chaotic in sparse groups.Secondly,extract the crowd feature of motion.By using the two value foreground mask,the foreground image is obtained.The optical flow information is extracted by calculating the local dense HS optical flow in optical flow mask,and then the average kinetic energy and the motion direction entropy are extracted.Distance potential energy is extracted by calculating the Euclidean distance between FAST corners in optical flow mask.The change rate of crowd number is calculated by using the normalized foreground pixel number.The four characteristics of crowd are respectively described the violent degree of group movement,the degree of confusion in the direction of movement,the degree of dispersion among groups and the change of the number of people in the range of video capture.Thirdly,recognize and classify the crowd behavior.The extracted crowd motion features are used as the basis for classification,the RF classifier is trained by the training samples,then realizing the recognition and classification of crowd abnormal behavior by the properties of various attributes in the test samples.This paper realizes the recognition and classification of four kinds of crowd behavior,they are normal,outdoor panic,scattered in the same direction,emergency run and stop.The crowd abnormal behavior detection method of this paper mainly are effectively improved in scene adaptability and real-time,Experimental results on the UMN and PETS2009 datasets show that the method can detect and recognize different crowd abnormal behaviors with 97 percent accuracy,which is higher than other methods,the proposed method is real-time and robust.