Action Recognition Based on Parallel Multi-Granularity Feature Refinement Networks

Video action recognition is an important research content in the field of computer vision. However, single feature motion information is under-represented and cannot completely describe the motion information. In this paper, an action recognition method based on parallel multi-granularity feature refinement network was propose to improve the action recognition accuracy. This method relaxes the requirement restriction on action recognition by describing the motion information of a video with multiple action class labels and shared features in the different action class group. Three action class granularity features was obtained by three action class label groups and integrate them to obtain the exact feature fusion of RGB image features and joint point skeleton information for action recognition. In order to verify the effective of our proposed network, a series of experiments were conducted on the UCF101 dataset. The experimental results show that the accuracy rate of proposed approach is higher than the traditional mainstream action recognition methods, which proves that the method is effective in action recognition.