Research On Human Behavior Recognition Method Based On Convolutional Neural Network

In these years,the domain of computer vision has continuously improved,and human behavior recognition technology has gradually become a popular research direction by experts in this field.This technology has wide application value in many fields,ranging from human-computer interaction to video retrieval,Big to intelligent security and safe driving.In recent years,deep learning technology has been well developed.Deep learning recognition methods have gradually replaced traditional methods,and human behavior recognition technology has made breakthrough progress.Deep learning has two main problems: one is the large amount of network parameters and high calculation complexity;the second is that the network is prone to overfitting.In response to these problems,the research work in this paper is as described below :First of all,based on the Dense Net network,it uses three-dimensional convolution to expand to extract the motion information in the video,while effectively improving the accuracy and reducing the amount of parameters.Dense Net adopts the dense connection method,which has the advantage of feature reuse.It reuses the features in the image,improves the utilization rate of the features,reduces the amount of parameters of the overall network,reduces the amount of calculation during the network training process,and solves the gradient to a certain extent.The problem disappeared.3D Dense Net can better identify the motion information in the video and improve the accuracy of behavior recognition.On the basis of the 3D-Dense Net-BC network used,replace the Re LU activation function with the Maxout activation function,and use the Maxout activation function with the Dropout method,which can effectively reduce overfitting during network training and improve the network.On the basis of the 3D-Dense Net-BC network used,replace the Re LU activation function with the Maxout activation function,and use the Maxout activation function with the Dropout method to effectively reduce overfitting during network training.At the same time,it can improve the convergence speed of the network during trainingIn order to improve the generalization ability of the network model,this paperreplaces the Average Pooling in the network model with Stochastic Pooling.The random pooling and the average pooling are similar in the average sense;in the local sense,Then obey the principle of Max Pooling.In reducing overfitting,Stochastic Pooling performs better than Max Pooling,effectively improving the generalization ability;in addition,the network using Stochastic Pooling has a better recognition effect than the network using Average Pooling.Through experimental verification and testing,the improved human behavior recognition method based on 3D-Dense Net-BC proposed in this paper has achieved very good recognition results on the human behavior recognition data sets UCF-101 and MERL shopping data sets,with recognition rates of 95.62% respectively and 88.4%,effectively reduce the amount of network parameters and improve the generalization ability of the network.