Human behavior recognition algorithm based on improved 3D ResNet and GRU

In order to solve the problem that traditional CNN lacks feature extraction for the time dimension, and the traditional RNN is easy to ignore local information when processing long sequences. In this paper, we propose a spatiotemporal feature extraction network (3D ResNet-Double Gated Recurrent Unit-Convolutional Block Attention Module, RDCnet) fused with convolutional attention mechanism, which first uses the improved 3D ResNet to extract the preliminary spatiotemporal features, and then sends the DGRU module with GRU as the basic unit constructed in this paper to further model the temporal information of the behavior. In this process, the convolutional attention module is introduced to capture the channel correlation and spatial correlation of the behavior. Finally, the classification results are obtained through the softmax layer. The proposed method is evaluated on the publicly available datasets UCF101 and HMDB51. Experimental results show that the average accuracy of the proposed method on the two datasets is 91.3% and 63.4%, and the recognition performance is high.