Research On Human Behavior Recognition Based On Deep Learning

With the popularity of Internet technology and intelligent photo equipment,tens of thousands of videos are shared in the network every day.How to effectively find out the illegal videos has become the top priority.The traditional methods of artificial recognition were limited in the face of massive video.With the deepening of deep learning research,human behavior recognition based on video has developed rapidly,and was widely used in network monitoring,intelligent driving,smart city,intelligent medical,human-computer interaction and other fields.There is a great demand for human behavior recognition in the market,so the research on human behavior recognition has certain economic value and significance.This paper mainly studied human behavior recognition from two aspects: one is key frame extraction technology,the other is to study and improve the existing human behavior recognition model.Firstly,the key frame extraction technology was studied.Combined with the existing technology,a new key frame extraction algorithm was proposed to solve the problems of key redundancy and missing key frames in key frame extraction.Secondly,based on the classical dual stream convolution neural network model,the gating cycle unit network was added to extract the time dimension information of video frame time.Finally,the loss function was optimized to improve the recognition accuracy.The main research work of this paper is as follows.(1)Optimized the key frame extraction algorithm.By calculating the optical flow between adjacent frames,the optical flow difference in the neighborhood was calculated.According to the preset threshold,the video frame was divided into key frame set and candidate key frame set.Secondly,the mutual information entropy in the key frame set was calculated,and the minimum mutual information entropy was selected as the comparison.According to the comparison results,the greater than the threshold was put into the key frame combination,and then the inter frame similarity ratio was passed compared with deleting redundant video frames,the remaining video frames are the key frames to be extracted.The optimized key frame extraction accuracy and F1 uniformity are significantly improved,and the extracted key frame can more effectively reflect the video content.(2)Improved the model structure.With the rapid development of deep learning,the depth and dimension of neural network also became deepened.In order to better extract behavior features,improve the accuracy of behavior recognition,this paper improved on the basis of Two-Stream CNN.Firstly,VGG16 net was used as the neural network to extract the spatiotemporal feature vector;secondly,in order to better extract the temporal dimension information of video,the weighted fusion was carried out after the feature information was extracted,which is used as the input of gating loop network to extract the temporal feature information between video frames;finally,the extracted feature information vector was input into the classification function to realize the recognition of human behavior.The recognition rate of the proposed model is higher than that of the UCF101.