Research On Video Behavior Recognition Based On Dual-stream Convolutional Neural Network

In recent years,with the continuous development of deep learning technology and the continuous improvement of computing power of computer hardware(GPU),the research field based on video understanding is now receiving more and more attenotion.Video-based action recognition is a very important direction in the field of video understanding.Video action recognition cross-cuts theories of image processing,pattern recognition,machine learning,computer vision and other disciplines.Its main task is to start from an unknown video or the video image sequence automatically analyzes the ongoing behavior and gives the correct classification.Human action recognition has high research value and application prospects in many fields,and is widely used in intelligent monitoring,human-computer interaction,virtual reality,video retrieval and other fields.Since the development of video behavior recognition technology,many excellent algorithms have appeared,which can more accurately recognize video,and can be applied in specific scenarios.However,due to inconsistent length of video segments,camera shake,diversity of human action,complexity and diversity of video scenes,video actionr recognition technology is still a difficult and hot topic in computer vision.This article focuses on the method of video action recognition based on deep learning.The specific research content is as follows:Firstly,the basic theory related to deep learning is introduced,and the back-propagation algorithm of deep learning,the structure and principle of convolutional neural network,the common optimization algorithms and data preprocessing methods are introduced in detail.Secondly,the current classic video action recognition methods are introduced,and the video action methods based on traditional manual feature extraction and the video action methods based on deep learning are described respectively,three mainstream video action recognition methods based on deep learning are emphatically introduced.In this paper,the initial two-stream convolution neural network model is studied deeply.Aiming at the problem that the network model feature extraction ability is insufficient and the sampling strategy cannot deal with the inconsistency of video length,a video action recognition model based on two-stream residual network structure is proposed by combining the two-stream network and residual network structure,by using the residual network structure to replace the original network structure and adopting the strategy of segmented sampling,the problem of inconsistent video length and the recognition performance of themodel can be improved,and explored the effect of data enhancement,pre-training initialization and some hyper parameter settings on the model recognition performance.The experimental results show that the improvement can improve the recognition performance to a certain extent.Based on the original two-stream residual network structure,this paper improves the two-stream residual network model by using two different methods for the first time,one of the methods is to solve the problem that the residual network structure can not make full use of the receptive field to extract the feature information of the video clip,the Res2 net module is used to replace the original residual network module,which can increase the receptive field of network layer and obtain multi-scale feature information.The other is to use the i Resn Net method to improve the residual network structure,which can promote the flow of network information and reduce the loss of information.Experimental results show that the two methods can further improve the recognition performance of the model while maintaining the original similar computational complexity.