Research On Human Action Recognition Algorithm Based On Joint Features

Human action recognition has become an important area in computer vision research due to their application prospects and huge market in video surveillance,human computer interaction,video retrieval,etc.However,it is one of the challenging problems in computer vision for some reasons,such as the variety of motion types,perspective changes,background disturbances,intra-class variations in the same action class,and the unclear definition of the action categories in the research.The feature representation of motion information is the most critical steps to achieve accurate human action recognition.This paper focuses on two aspect: the feature representation and multi-feature fusion strategies.In feature extraction,the paper proposed the human action recognition method based on dense sampling of motion boundary and motion gradient histogram to enhance the feature expression.Since different features have different focuses in the representation of action information,a multi-feature fusion strategy based on multi-kernel learning was proposed to highlight the role of different features in the overall action recognition to achieve optimal recognition results.The main contents of this paper are as follows:(1)We proposed a human action recognition method based on dense sampling of motion boundary and histogram of motion gradient.In order to avoid the interference of camera movement,background information redundancy and other issues in the process of sampling feature points,the sampling method based on moving boundary is incorporated into the improved dense sampling,eliminating a large number of invalid sampling points and reducing the number of trajectories.Next,in order to fully exploit the inner relationship of human action in time and space,motion gradient histogram based on time and space is used to capture motion information,integration with dense features,to enhance the feature expression.Finally,the encoded feature vectors are merged and input into the classifier to realize the recognition of human motion.The experiment results show that the proposed method effectively improves the human action recognition accuracy in the case of accelerating the speed of algorithm.(2)We study a feature fusion strategy based on multi-kernel learning for human action recognition.In the feature representation,the human motion information includes the appearance,trajectory,and spatio-temporal of the motion,which requires different feature to be characterized.However,the contribution of different features exist some difference in the recognition effect.In order to realize the most effective feature fusion,the multi-kernel learning method was adopted in this paper.The combination of features is converted into a linear combination of the kernel matrix corresponding to each feature.The weight value of the kernel matrix highlights the importance of different features in the representation of human movements,and the role of the features that contain more action information is exerted.The analysis of experimental results shows that this method can deeper the mining of the relationships between various features,and achieve the optimal combination of features in the data-driven way,thereby further improving the effect of motion recognition.