Recognition And Analysis Of Abnormal Human Behavior Based On Neural Network

Nowadays,video surveillance systems have been installed in most of domestic pubilc area.However,most of those systems only have one function which is what we called‘Afterwards Verification'.So it is necessary to conduct recognition,analysis and prediction of crowd behavior through surveillance video.In this paper,we propose a method of the abnormal behavior of crowd based on neural network to contrapose the problems such as low recognition rate,low recognition speed and the flaw of unable to locate abnormal area further.This method can determine whether any abnormal behavior is happening followed by alter for security personal to eavcuate crowd to pervent any tragedy.In this paper we mainly focus on crowd scattering and crowd gathering to conduct abnormal behavior detection and analyze.Specific methods are as followed:?.Improved Multi-scale Convolutional Neural Network.In this paper,based on the idea of VGG network model,we add more layers to the previous network,so that more pixel features can be extracted and the effect of detection is improved.In this paper,the improved multi-scale convolutional neural network is used to predict the coordinate of pedestrian head in the surveillance video and obtain the crowd density map.At present,most of the methods are using optical flow method to extract feature points and then calculate their states of motion.Now that optical flow method extracts more feature points and solely calculates the motion and direction values of feature points,but its calculation is still large.In this paper,three kinds of eigenvalues of crowd movement are calculated based on the predicted pedestrian coordinate points to make a dynamic analysis of crowd behavior,and the dynamic changes of the three eigenvalues are used to illustrate the overall movement of the crowd.The three eigenvalues of crowd state are crowd average kinetic energy,crowd density value and crowd distribution entropy.Due to the unique pedestrian coordinates,the calculation amount of the three eigenvalues is reduced to a certain extent,and the detection speed is also improved.?.The classification algorithm of the abnormal behavior.This paper mainly appraises the population as a whole.Aiming at the cases of low classification accuracy and lack of real-time performance,this paper adopts the Extreme Learning Machine to classify anomalies,and conducts parameter optimization on the connection weight and hidden layer deviation of the Extreme Learning Machine through difference and particle swarm optimization algorithm to improve the classification accuracy.Due to the number of hidden layer nodes need to manually,generally,the more hidden layer nodes,the anomaly classification accuracy will be higher.So we can increase accuracy by constantly setting the number of nodes,and this paper by adding the time complexity of the calculation,at the time of highest accuracy,we find the suitable number of hidden nodes for this data set and reduce the amount of calculation.The experimental results show that the optimized Extreme Learning Machine has loftier accuracy and real-time performance.?.Analysis of abnormal behavior of crowd gathering.This paper uses the YOLOV3 target detection network to detect and make spot pedestrians in the surveillance video security system,identify and track their lanes,and update their posts in real time.After identifying the target pedestrian,the detected pedestrian center point is continuously updated to keep track its position,so as to obtain the pedestrian track in the surveillance video.The experimental results show that through real-time analysis of the surveillance video,each pedestrian is detected and tracked,and the crowd gathering behavior is detected according to the continuous reduction of the crowd range.In view of the inability to locate the clustering anomaly area,this paper finally realizes the determination of the clustering anomaly area by calculating the mean value and variance of the coordinate points of the crowd trajectory.The above methods are used to detect and analyze the abnormal behaviors in the surveillance video security system.The experimental results show that the above methods are feasible,at meanwhile,the speed and accuracy of detection can also be improved according to the comparative experiments.