Crowd Abnormal Behavior Recognition Based On Deep Learning And Sparse Optical Flow

To solve the problems such as fewer definitions of abnormal types,low detection accuracy,mostly abnormal running and some abnormalities cannot be recognized in public places,this paper proposes a recognition algorithm for abnormal behaviors of people based on YOLOv3 and sparse optical flow.By detecting small group abnormalities,i.e.the inducement of group abnormalities,it can provide early warning and recognition for group abnormalities.Take appropriate emergency measures to provide adequate time.Firstly,in order to locate the abnormal area conveniently,the video was divided into several sub-areas and the image samples of sub-areas are obtained.Secondly,small group abnormal detection was carried out to induce group abnormality.Aiming at the difficulty of traditional algorithm in detecting pedestrian stick,gun,knife and face occlusion,these abnormalities were detected by YOLOv3 neural network.Finally,because of other reasons such as herd psychology,people will also have abnormal evacuation when no abnormal inducement of the above groups was found.For this reason,sparse optical flow method was used to obtain the average kinetic energy and direction of motion entropy of the crowd,and the obtained characteristic data are sent to DE-PSO-ELM for classification,to distinguish normal behavior from isotropic evacuation or irregular evacuation.The experimental results show that,compared with the existing algorithms,the proposed algorithm can effectively detect small group anomalies such as pedestrian abnormalities and facial occlusion anomalies,and locate the abnormal areas,which provides more time for early warning and emergency measures,and its accuracy is up to 98.18%.There are few methods for detecting abnormal behaviors of crowd massing in public places.Besides,most of these methods have defects such as low detection accuracy and time effectiveness.And they are generally performed after the crowd massing abnormity has formed.In light of this,a prediction method of crowd massing abnormity based on multi-scale convolutional neural network(MCNN)is proposed in this paper.Firstly,a crowd counting model was built through MCNN,which would be used for testing the video of crowd mass abnormity.In the process,the number of crowd and the coordinate points of their heads were acquired.Then,the crowd density,crowd distance potential energy and crowd distribution entropy were calculated.Finally,the predictive model was built through the eigenvalues of three crowd motion state by PSO-ELM.Through the change of characteristic data,the prediction was completed.According to the experimental result,compared with existing algorithms,the algorithm in this paper can effectively achieve the early warning of abnormal behaviors in crowd massing.With a prediction accuracy rate of 97.17%,it's more time-sensitive and provides more time for taking corresponding emergency measures.