The Abnormal Detection Of Crowd Based On The Surveillance

Along with the development of China's economy and the acceleration of urbanization,more and more attention has been paid to the safety of important public places.In order to maintain national and social stability,and ensure the safety of life and property.In 2015,China has launched the construction of video surveillance about public security relying on the internet.The traditional video surveillance systems depends on the eye of human,which will expend large amount of manpower and material resources.It can no longer meet the requirements of provessing mass video data.Therefore,in order to meet the social demand.The intelligent video surveillance technology emerges at the historic moment.As an important part of intelligent video surveillance,anomaly detection can provide the real-time early warning and the detection of anomalies in surveillance scenarios without human intervention.In this paper,the main research object is the crowd in video surveillance scene.A crowd anomaly detection algorithm based on the temporal and spatial gradient features of of activity map is proposed.The main contents of this paper are as follows:(1)In this paper,the crowd is regarded as a whole for studying.Then each pixel of the image is regarder as a moving particle,and the velocity of each particle is calculated by the method of optical flow.Next,we compare the crowd movement system to the thermodynamic system.Finally we obtain the activity map of crowd by applying the enthalpy model to crowd motion system.(2)Based on the fluid characteristics of moving particles,the concept of fluid energy gradient is introduced into the activity map.At the same time,we combine with the field theory,proposing the concept of space-time gradient features of activity map.Then we extract the temporal and spatial gradient features of activity map.After clustering the feature,we judge whether there are abnormal crowd events in the secne by training the normal crowd in the scene.(3)In the process of calculating particle velocity,it is necessary to normalize the particle velocity by weighting,because of the influence of camera perspective deformation.In this paper,with the help of physical model of momentum conservation.We transforme the weight of particle velocity into the ratio of the foreground area of the object.Then the different weights are given to the particles according to the hypothesis that the foreground area changes linearly with the distance of the camera.(4)In the end,we carried out the experiments on multiple groups of videos in common data sets UMN and PETS09.The algorithm is evaluated by AUC,EER,accuracy and other reasonable indicators.Then we compared the algorithm in this paper with other excellent algorithms.The experimental results show that the detection results of the algorithm in this paper are basically consistent with human perception.Compared with other algorithms,the algorithm in this paper has better ability of anomaly recognition,and the judgment of anomalies is more accurate and the error rate is lower.