Human Action Recognition Based On Multi-Feature Fusion

With the development of science and technology,the coverage of monitoring equipment is also increasing.The resulting massive data puts forward higher requirements for security systems and relevant practitioners.In the new era,video based human behavior recognition,as an independent and human centered recognition technology,has greatly alleviated the pressure of security system and has a very broad application prospect.In this paper,different solutions are proposed for the problems that may occur in the practical application of human behavior recognition,such as background noise interference,insufficient feature representation ability and difficulty in constructing video long-term information.The research content of this paper mainly includes the following three points:(1)Aiming at the problem of background noise interference in the application of behavior recognition,a motion region extraction method embedded in the attention pool layer is proposed to enhance the performance of the target motion region.Methods the attention pooling layer was used to replace the full connection layer in the conventional convolutional neural network,and the hard attention weight vector and soft attention weight vector were trained to capture the region with the highest correlation with the classification results.Experiments show that this method can capture the main moving area of the target,and then improve the accuracy of behavior recognition.(2)Aiming at the lack of fine-grained motion representation ability in behavior recognition,a behavior feature extraction method based on flow field physical characteristics and manifold learning is proposed to construct behavior features focusing on fine-grained performance.Methods firstly,the divergence and curl values of each sampling point of discrete optical flow field are calculated,and the appearance and motion features are constructed.Then,the method uses the ndimensional unit hypersphere to map the features at different times,and then calculates the change rate of the features on the time scale,so as to amplify the motion performance of finegrained behavior.Experimental results show that the features extracted by this method have good distinguishability and robustness.(3)Aiming at the difficulty of long-term information learning in behavior process,a feature fusion method based on cascaded convolution fusion strategy is proposed to fuse features and learn the long-term information of video.Firstly,the first level fusion network is used to splice the quaternion feature groups at different times,and extract their spatial information.Then,the two-level feature fusion network is used to fuse the spliced features on all time scales,and the features are sent to the classifier to realize the task of human behavior recognition.The experimental results show that compared with other feature fusion methods,this method can calculate the features of the same length according to the videos of different time lengths,and maintain a high recognition rate without destroying the spatial and temporal relationship of the original features.There are 26 figures,12 tables,and 81 citations in this thesis.