The Abnormal Detection Of Crowd Based On The Change Of Energy-level Distribution

With the rapid development of artificial intelligence and the continuous improvement of the human cost,the intelligent monitoring technology has attracted more and more researcher’s attention.As well as the core content of intelligent monitoring system is the detection and analysis for the crowd behavior,and its main task is to detect and extract the crowd movement characteristics from the consecutive frames in the image,so as to realize the description and analysis for the crowd motion.Once detected abnormal behavior occurs to alarm prompt in a timely manner,so can reduce casualties and property losses in the largest extent.Thus,how to detect the abnormal crowd behavior accurately is the main focus of this paper.Nowadays,there are two main approaches in modeling the crowds : Microscopic approach and Macroscopic approach.The first method needs to detect and track the individuals,but when the number of people is large there will be overlapping phenomenon,which will produce a larger error.So according to the macroscopic approach this article puts forward a crowd abnormal detection method based on the change of energy-level distribution.The main content consists of kinetic model establishment,energy level co-occurrence matrix calculation and abnormal behavior detection.Firstly,we treat pixels in the image as particles,and use the optical flow method to extract the velocity of them.Then we make a linear interpolation calculation to pedestrian’s foreground area who the farthest and the nearest to the camera,and obtain the particle’s quality of the different location;Secondly,according to the information of velocity and quality,a kinetic model of the particle is established.Then we grade the kinetic of the particle to different levels,and describe the energy-level distribution with the co-occurrence matrix;Finally,according to the change of the consistency,entropy and contrast the three descriptors of co-occurrence matrix to analyze the crowd behavior,and determining the time of the crowd abnormal happened by setting a threshold.In this paper,multiple sets of video from the three different scenes in UMN dataset are carried on the experiment,and the experiment results are compared with two different methods.The results show that our algorithm is an outstanding way to characterize anomalies in videos,and the test results are more close to the true value than the other two methods.