Research On Video Violence Behavior Detection Algorithm Of Deep Convolutional Network

In recent years,in order to continuously promote the construction of safe cities,security monitoring equipment has been widely used throughout the country.How to use computer vision technology to realize effective intelligent analysis of violence in video surveillance is essential for maintaining social stability and safeguarding people's lives and property.With the rapid development of deep learning,this paper studies the detection of violent behavior in surveillance video based on the deep learning convolutional neural network.The main research contents include the following points:Firstly,in order to solve the problem of low detection rate,and high computational complexity caused by the hand-crafted descriptors,a violence detection algorithm based on three-dimensional convolutional neural network is proposed.The three-dimensional convolution and three-dimensional pooling operations on the input original video sequence can effectively obtain deep spatial and temporal features in the video,and achieve end-to-end violent behavior detection.In order to improve the generalization ability of the network,and accelerate the network training,batch normalization is performed after each convolution operation.In order to avoid the network over-fitting phenomenon caused by the small sample of the dataset,dropout will be added after the two fully connected layers.Experiments show that the algorithm effectively improved the performance of video violence detection.Secondly,in order to make full use of video spatio-temporal feature information and improve detection accuracy,a violence detection algorithm based on two-stream three-dimensional convolution feature fusion is proposed.The three-dimensional convolution neural network is used to extract the features of the continuous video frame sequence,which can not only obtain the video spatial information but also effectively extract the temporal and spatial fusion information of the video.The VGG network(Visual Geometry Group Network,VGGNet-16)is used to extract the motion information of the optical flow images.Finally,the prediction results of the two models are averaged at the classification layer to obtain the final classification result.Experiments show that thealgorithm can effectively improve the performance of video violence detection.Finally,in order to solve the problem of large parameter and high computational complexity of network model,based on densely connected convolutional network,a violent detection algorithm of three-dimensional convolution densely connected convolutional network is proposed.The network model is constructed to obtain the spatio-temporal features of the video frame sequence by three-dimensional convolution and three-dimensional pooling operations.The network layer adopts a dense connection method,which enhances feature propagation,improves feature reuse rate,reduces model parameter quantity and effectively reduces over-fitting of the network.Then,in order to reduce the output feature dimension,the transition layers are used between the three-dimensional dense blocks.Finally,the softmax layer obtains the final classification result.Experiments show that the algorithm has certain effectiveness in video violence behavior detection tasks.