Research On The Approach Of Human Action Recognition Based On Spatio-temporal Features

With the development of science and technology, human action recognition in videoshas been widely applied in the research field of human-computer interaction, intelligentmonitoring, animation-simulation, content analysis and retrieval of videos. Actionrecognition has been an important research hotspot which is an extremely challengingdifficulty as well. How to recognize the human action in videos exactly is one of the keylink to solve these problems. Based on analysis of the present situation of the domesticand foreign research, this paper do in-depth research on the approach of human actionrecognition by using Support Vector Machine classifier as technical foundation andcomprehensively using many aspects of theories knowledge, such as spatio-temporalfeatures, kernel function, bag-of-features and so on.Firstly, detection approaches of spatio-temporal feature points, formation of motiontrajectories and commonly used descriptors for spatio-temporal features are introduced,and the basic idea of bag-of-features and the technology of Support Vector Machine whichis used for classification are also introduced.Secondly, in connection with the shortcomings of existing approaches, a novel actionrecognition approach based on feature combination and histogram intersection kernelfunction is proposed. In this approach, after the joint feature for each trajectory iscomputed, equivalently sample methods is proposed to quantify the joint feature matrices.Then input these quantified feature matrices to Support Vector Machine which selects thehistogram intersection kernel function as the kernel function to do the training and testingfor classifier.Thirdly, an approach of action recognition based on spatio-temporal features foredges is proposed which is in connection with the number of spatio-temporal featurepoints. In this approach, Canny detection for edges is applied to filter the spatio-temporalfeature points, and then the reservations which are on or besides to edges are followed toform trajectories. According to the scale invariant feature computation for edge motion trajectories, the enhanced joint feature matrices are obtained. Classifier’s training andtesting is done at last after the quantification by bag-of-features.Finally, experimental results show that the proposed approaches for actionrecognition are feasible and effective, and the analysis of experimental results are made.