Study On Human Behavior Recognition Technology Based On Deep Learning

With the development of computer information technology and artificial intelligence technology,human behavior recognition has become one of the hottest research directions,and is widely used in the fields of human-machine interaction,autonomous driving,smart home,intelligent monitoring and healthcare.Traditional human behavior recognition algorithms are manually designed for specific tasks,and the process of extracting features is complicated,leading to low efficiency and poor generalization ability.In recent years,three dimensional convolutional neural network algorithm based on deep learning has replaced traditional algorithms with the advantages of extracting behavioral features in an end-to-end manner and high accuracy.But there are also some shortcomings.Firstly,3D convolutional neural network has a large number of parameter and complexity,which inadequately extracts spatio-temporal features;Secondly,when the input video information is too long,the network inadequately captures the interdependence of human behavioral features in long time sequences,losing important feature information.Based on the above problems,this paper focuses on the following research work:To address the problem that C3 D network has a large number of parameter and complexity,which inadequately extracts spatio-temporal features,this paper designs a human behavior recognition algorithm based on SR3 D network.The network places the BN layer and Re Lu activation function before the 3D convolutional layer to improve the learning ability of the network.Meanwhile,the SE module is extended to 3D and combined with the improved 3D residual block to increase the weight proportion of the important feature channels,which enables the network to extract more important spatio-temporal features.Experimental analysis was conducted on the UCF101 dataset,and the results showed that compared with C3 D network,SR3 D network has a small number of parameter and complexity,which can adequately extract spatio-temporal features and improve the accuracy of human behavior recognition.To address the problem that the SR3 D network inadequately captures the interdependence of human behavioral features in long time sequences,losing important feature information,a VT3 D network is designed in this paper.The network uses multi-scale convolution to extract long,medium and short time sequences information,which can adequately capture the interdependencies of human behavior features in long time sequences and avoid the loss of important feature information.Meanwhile,this network uses a spatio-temporal separation technique to realize the lightweight,and the influence of parameters such as resolution and sampling interval on the training effect of the network is explored.Besides,a VTR3 D network is designed by combining VT3 D network and residual connection,which speeds up the training of the network and solves the model degradation problem of VT3 D network,and conducts a comparative analysis with the existing network on UCF101 dataset.The results show that the VTR3 D network not only reduces the number of parameters and computations,but also improves the real-time performance and accuracy of the network.Finally,the VTR3 D network is tested on the self-produced human behavior dataset by using migration learning,and verifies that the proposed algorithm has good generalization ability.At the same time,the human behavior recognition system is built and visualized,which lays the foundation for the subsequent research.