Research On Human Action Recognition Based On Convolutional Neural Network

As a research hotspot and difficulty in the field of computer vision,human action recognition has been favored by researchers and scholars.Although the accuracy of the traditional human motion recognition method is still acceptable,it is necessary to manually design and extract the action feature,so it is greatly affected by personal differences.The convolutional neural network can automatically extract low-level features to high-level features,completely free from human characteristics,and has made breakthroughs in the field of image recognition.Therefore,the purpose of this article is to propose a human action recognition method with higher recognition rate and better robustness based on convolutional neural networks.In this article,the LRCN algorithm based on two-dimensional convolutional neural network is studied in depth.It is found that LRCN uses a shallower AlexNet convolution network for feature extraction,which makes it impossible to extract more effective and more semantic features.However,the memory nerve LSTM unit used has structural redundancy and many parameters,which makes the model run longer and is more prone to over-fitting.In view of the above problems,this paper proposes an improved LRCN algorithm,which replaces the AlexNet network with a deeper ResNet-34 network,and replaces the LSTM unit with GRU units,and compares the improved LRCN algorithm model to improve not only through experiments.The human body action recognition rate also shortens the running speed of the model to some extent.Although LRCN increases timing learning compared to convolutional neural networks,the LRCN algorithm cannot simultaneously learn spatial features and temporal features.Therefore,this thesis deeply studies the human motion recognition algorithm of 3D convolutional neural network which can learn two features simultaneously.However,although the 3D convolutional neural network can synchronously study spatial and temporal features,the algorithm model only uses two layers of linear convolutional layers,so its model abstraction ability is not strong enough to extract more abstract features.Based on this,this paper redesigned a human motion recognition network based on 3D convolutional neural network.Compared with the original network,the new network not only adds two layers of MLP convolutional layer,but also uses dropout technology to prevent over-fitting of the network.And the parameter calculation is reduced,and finally the last fully connected layer is replaced with GAP.Therefore,the new network has stronger abstraction capabilities and better robustness.Finally,in order to verify the advantages of the new network,this article makes a synchronous comparison in the UCF101 public data set.The experiment proves that the new network performs better in the accuracy of human motion recognition than the original network.