Research On Abnormal Behavior Of Human Body In Video Surveillance Based On Deep Learning

Video surveillance plays an important role in maintaining the security order of public places.How to deal with a large amount of information collected in video surveillance,2 analyze and understand the abnormal information is an important research direction in this field.At present,the main methods for abnormal behavior recognition in video surveillance based on computer vision are:(1)BP neural network,support vector machine and other traditional pattern recognition methods;(2)deep learning method represented by CNN and RNN.However,the traditional pattern recognition method relies on the characteristics of manual selection,the degree of intelligence is low,and the recognition accuracy is not high.Deep learning model has strong generalization ability and can extract the feature automatically.In this paper,the deep learning method is used to study the two abnormal behaviors of fall and fight in video surveillance,which are summarized as follows:(1)Aiming at the accuracy of recognition of fall behavior in video surveillance is not high,and the existing deep learning method can not effectively combine the spatial and temporal characteristics of fall behavior in video surveillance,and propose the recognition of human fall behavior based on CNN and LSTM(long-short term memory)hybrid model.The model adopts a two-layer structure,First,the video is input into the network in groups of 5 frames.Then,using CNN to extract the spatial features of the video sequence,LSTM extracts the features in the time dimension of the video data.Finally,the softmax classifier obtained the classification result.Experiments show that this method can effectively improve the accuracy of fall behavior recognition.(2)Aiming at the problem that the accuracy of fight behavior recognition in video surveillance is not high and the existing deep learning methods can not make good use of the time dimension information,a fight behavior recognition model combining Two-stream CNN and LSTM is proposed.First,the model inputs sequential optical flow images and sequential RGB images,respectively,to extract temporal and spatial features.Then,the time series features and the spatial features are merged.Finally,the fusion feature is input into the LSTM network to learn the motion characteristics of the long-time fighting behavior in the time dimension and combined with the softmax classifier to realize the recognition of the fighting behavior in the video surveillance.The comparison experiment with traditional pattern recognition method and the existing deep learning method shows that the method can effectively improve the accuracy of fight behavior recognition.