Research On Behavior Recognition Based On Local Information Fusion

Human vision and behavior recognition technology is an important academic research topic in the fields of artificial intelligence video technology,computer science and visual technology.Due to the wide application of video human action recognition in video surveillance,security monitoring systems,human-computer interaction and other fields,the research on human action recognition has important academic research value and significance.Due to the complexity of human movement and the variability of the external environment,human behavior recognition has some challenges.The academic community has continuously innovated the research methods of human behavior recognition,and related theoretical research has gradually deepened.It has made great achievements in application fields such as video processing and big data analysis.The current human behavior recognition methods have a better effect on human behavior recognition in short videos,such as dual-stream neural network,3D convolutional neural network,spatiotemporal convolutional neural network and other recognition methods.The video input of these behavior recognition networks is either an original RGB image randomly selected,or a group of dense RGB images,or a group of optical flow diagrams.But for long videos,intensive selection of a set of image frames cannot objectively represent global information.Therefore,from the perspective of short video theory and technical practice,this paper proposes a segmented video recognition network(Key Frame Segment Network,KFSN)based on local information fusion of key frames for human behavior recognition in long videos.The long video is divided into multiple videos of equal length to recognize human behaviors respectively,and then the recognition results of the segmented short videos are merged.This kind of network is based on the idea of long-term modeling,and it combines the strategy of sparse time video sampling,so that the entire action video can be efficiently learned.The recognition method proposed in this thesis has been tested many times on public datasets UCF101 and HMDB51.The experimental results obtained show that the KFSN network proposed in this thesis can achieve a good behavior recognition effect,which can achieve a recognition rate of 95.1% in UCF101 and 70.1% in HMDB51,better than some existing behavior recognition network performance.