Action Recognition Method Research Based On Multi-granular Feature Parallel Processing

Due to the rapid development of artificial intelligence in recent years,the application of video action recognition has been widely used,and this technology has been focused on by scholars.However,the existing RGB image information of the original video used in the video action recognition research has a complex scene and the human body may be blocked.Moreover,as the video scene becomes more complex,the interactive motion of human body increases day by day,and the single feature cannot describe all the motion information of the video.In addition,when describing the video information,researchers usually only use a single action class and ignore the shared features in the action class group.However,there are similarities in the shape of people in many different motion categories,which is easy to be confused.The problem will affect the accuracy of action recognition.Therefore,in order to solve the above problems,this paper focuses on the research of multiple feature fusion and multi-granularity feature refinement,proposes the action recognition method based on the parallel convolutional recurrent neural network and the action recognition method based on the parallel multi-granularity feature refinement network,and conducts the action recognition from two perspectives.The main contents are as follows:1.To solve the problem of insufficient expression of motion information by single feature in video,a feature fusion action recognition based on parallel convolutional recurrent neural network is proposed.In this method,the RGB image features and the skeleton features of human body joints are input into the CNN part and the RNN+LSTM part of the parallel convolutional recurrent neural network,respectively.Then the two features are connected to a joint space-time feature vector for feature fusion after feature extraction,and finally the action recognition is realized.The experimental results show that the accuracy of the proposed method in UCF101 data set is better than that of other mainstream action recognition methods,which verifies the effectiveness of the proposed method in action recognition.2.The action recognition method based on parallel multi-granularity feature refinement network is proposed to solve the problem of insufficient motion information expression by single feature in video and the problem that a single action class cannot describe the all motion information and the shared feature in action class group is ignored.The method loosens the requirement of action recognition and describes themotion information of a video with multiple action labels.We obtain the granularity features of the three action classes through the three action class tag groups,and then integrate the the precise features of the RGB image and the skeleton information of the joint nodes for action recognition.In this paper,experiments were conducted on the UCF101 data set,and the accuracy was higher than the traditional mainstream action recognition method,which proved that the method is effective in action recognition.At the same time,the accuracy of the two action recognition methods in this paper on the UCF101 data set is compared.Because the action recognition method based on parallel multi-granularity feature refinement network is more detailed in feature extraction and refinement,the accuracy of this method is 1.1% higher than the feature fusion action recognition based on parallel convolutional recurrent neural network.