Human Action Recognition Method Based on Video-Level Features and Attention Mechanism

In order to capture the spatiotemporal information in the video and improve the long-term modeling capability of the network, two-stream network usually adopts the method of sparse sampling. However, there are two problems in the process of feature extraction on the sample. One is the unreasonableness of the snippet-level features corresponding to the video labels, and the other is that the salient features are not highlighted. In view of the above two points, we propose an action recognition method based on video-level features and attention mechanism (VFAM), which combines snippet-level features to generate video-level features, and adds attention mechanism to give effective features with greater weight, and good experimental results have been achieved on the dataset HMDB51, reflecting the superiority and robustness of our method.