Human Action Recognition Method Based On Feature Fusion

Human behavior recognition technology is one of the important computer vision technologies that has developed rapidly with the maturity and popularization of deep learning.It is a key technology with great practical and economic needs.Compared with single-frame image-based image recognition,image fusion,and image style conversion,human behavior recognition has the characteristics of larger data volume,more complex tasks,and more diverse scenes,so human behavior recognition is more difficult.Based on the existing human behavior recognition research,the text selects specific frames in the video and some image blocks in specific frames to highlight the semantic subject information of motion in human behavior.Based on deep learning,it uses traditional algorithms Features are fused to achieve the purpose of improving video feature representation capabilities,reducing network complexity,and improving behavior recognition accuracy.Based on the existing network,this paper integrates the characteristics of deep learning and traditional algorithms to increase the discrimination between actions,reduce the data size,and improve the speed and accuracy of human behavior recognition.The main tasks are as follows:Firstly,a fast human behavior recognition method based on key frame selection is proposed.In 3DCNN,network frames are directly trained by taking fixed frame lengths of video image frame sequences,and inspired by this,an algorithm based on key frame selection is used to output fixed-length video feature information for behavior recognition.The algorithm first extracts features from the frame sequence of the behavior recognition image frame by frame and sorts the key frames to obtain fixed-length video features,and then restores the timedomain information of the image through secondary sorting,and then constructs fixed-length structured features through feature rearrangement.Finally,the rearranged features are learned by LSTM to improve the accuracy of behavior recognition.Secondly,a human behavior recognition method based on background suppression is proposed.Since selecting only a part of the frames is sufficient to complete the characterization of the video,it should have the same effect by selecting the limited image blocks in the image frames.This chapter draws on the research ideas in target detection,by first performing low-rank sparse reconstruction on the video in the input network,and then using Unet to detect the portrait area,suppressing the background information of the action,on the one hand,to improve the discrimination between actions,on the other hand The role of weighting in the whole image is emphasized in the action that occupies a relatively small area of the image.Finally,the motion history graph is used to fuse the characteristics of the action over a period of time,so as to improve the accuracy of behavior recognition.Thirdly,a human behavior recognition method based on background fusion is proposed.The image of the input network is directly processed by the background suppression method,which destroys the context information of the action and is not visually interpretable.In this chapter,based on the background suppression-based human behavior recognition algorithm,first use the dynamic image algorithm to complete the fusion of the background image and the background image,and finally send the dynamic image with the background information to Res Net for learning.On the one hand,the network has the foreground action effect protruding in the background suppression,on the other hand,it maintains the complete background information of the action context,thereby improving the accuracy of behavior recognition and classification.