Research And Application Of Abnormal Behavior Detection Based On Deep Learnin

The emergence of video equipment has provided a certain guarantee for the public security problem which has been concerned by all walks of life.However,the comprehensive deployment of monitoring equipment in every corner brings a new problem,that is,how to timely find out whether there are abnormal events in the monitoring area in the massive video data,and effectively make countermeasures,that is,how to carry out intelligent crowd abnormal behavior detection based on the monitoring video.This problem has also become a research hotspot in the field of computer vision,attracting the attention of many researchers.In recent years,the detection technology of crowd abnormal behavior based on surveillance video has been developed to a certain extent.However,there are still many difficulties to be solved,such as detection accuracy,model stability and uneven distribution of positive and negative samples.This thesis mainly studies the application of deep convolutional neural network to detect abnormal behavior in population.Firstly the crowd behavior detection of related is introduced,and then two kinds of abnormal behavior detection algorithm are proposed in this thesis.Experimental results on public datasets UMN and VIF verify the effectiveness of the proposed methods,and demonstrate they are superior to the state-of-the-art abnormal behavior detection algorithm.Finally,a software system is developed based on the algorithm of crowd behavior detection.The main content of this thesis is as follows.1.From the perspective of feature representation method and deep learning network model,the relevant algorithms of crowd anomaly detection used in this thesis were studied in detail.Firstly,the optical flow method with excellent performance in the underlying feature representation is introduced.Then,the representation methods based on flow characteristics are discussed,including pathline,streamline and streakline,as well as their advantages and disadvantages.Finally,commonly used deep learning methods are studied,in which neural network,convolutional neural network,dual-stream CNN network model and GANomaly network model are introduced in detail.2.In view of the method lack of static characteristic and dynamic characteristic of video data detection and thereby leads to low accuracy and stability of the model,this thesis presents a streak flow convolution neural network method(SFCNN),which is composed of two-stream CNN and ResNet101 networks.Our methods the apparent characteristics and streakline flow characteristics in the video,and fuses of two kinds of extracted features,to complete the detection of crowd behavior.Experimental results show that,compared with some state-of-the-art algorithms,this method has achieved 93.33%detection accuracy and 89.99 ROC offline area on the VIF data set,and 98.78%detection accuracy and 98.58 ROC offline area on the UMN data set.3.Aiming at solving the uneven samples problem in abnormal crowd behavior detection,this article adopts the method of unsupervised learning designed a model based on flow acceleration streakline flow space-time against network model(SFA-TGANomaly)to detect abnormal human behavior,by introducing a set of weights of encoder,horizontal acceleration encoder and the results of the vertical acceleration son encoder to get more potential variable flow space-time encoder,through the decoder to build image,according to the norm constraint adjustment model,at the same time introducing parameters the restructuring test data for anomaly detection.The experimental results show that,compared with other algorithms,this method achieved 70%VIF is based in the data set of detection accuracy and ROC 89.70 in UMN data set the detection accuracy of 86.3%and 99.10 ROC.4.Based on the proposed two kinds of crowd behavior detection algorithm,design and implement a crowd behavior detection based on video software,this system has realized the video file to import and export,video frame serialization,optical flow and flow characteristics of streakline extraction,optical flow acceleration and streakline flow acceleration feature extraction method,based on SFCNN and SFA-TGANomaly abnormal behavior detection,and other functions.