Research On Human Behavior Recognition Based On Deep Learning

Human behavior recognition is a hot research direction at present,and has been widely applied to various intelligent domain based on video.For the past few years,deep learning has developed at express speed in plenty of domain.Many scholars have obtained a number of achievements in human behavior recognition based on deep learning,but the video scene is complex and changeable,and the human behavior recognition technology still faces many challenges.At present,the behavior recognition method based on deep learning still has some problems,such as insufficient feature extraction,features contain a large amount of redundant information,and single-mode information is difficult to fully express video behavior.Therefore,this paper studies two human behavior recognition models based on dual stream network to improve the performance of behavior recognition.The primary work and innovative points of this paper are described below:1.Aiming at the problems of poor modeling effect and insufficient feature extraction of long video sequences based on dual stream network,a human behavior recognition model based on video segmentation and improved residual network is proposed.In order to realize the modeling of long video,the idea of time segmentation is introduced,the input video is isochronous segmented,and the prediction scores of each segment are fused in the later stage to form video level prediction.For the sake of promote the capability of network feature extraction,the residual module of Res Net-34 network is improved,and the attention on multi-channel residual module is proposed.Instead of using the relu activation function in the original residual module,the mish activation function with better gradient descent effect is replaced.On the basis of a single path,another two parallel feature extraction paths are added,and the attention mechanism is embedded to make the network pay attention to important features.Furthermore,to go further enhance the capability of behavior recognition,the key frame with the largest amount of image information in each segment is selected as the input of spatial convolution network.Experimental comparison was carried out on UCF101 and HMDB51 datasets,the experimental results manifest that the network can validly enhance the accuracy of behavior recognition.2.Aiming at the problem that a single modal information can not fully express the human behavior in video,a human behavior recognition model based on GRU and multimodal information fusion is proposed.Based on the dual flow network model,bone graph is introduced as the third mode to supplement RGB graph and optical flow graph.Appearance features,optical flow features and bone features are extracted from RGB graph,optical flow graph and bone graph respectively through improved Res Net-34 network.In order to capture the long time sequence information of the video sequence,the gating cycle unit GRU is used to process the three feature sequences respectively,and different fusion methods are considered to fuse the prediction scores of the three features to realize the complementarity of multi-modal information.Experimental comparison was carried out on UCF101 and HMDB51 datasets,the experimental results manifest that the network can validly enhance the accuracy of behavior recognition.