Action recognition based on R2.5D-GRU networks

Abstract

The field of body action recognition is a research hotspot in computer vision. Due to the complex calculation process of traditional recognition algorithms and the limitations of the data set to be processed, action recognition algorithms based on deep learning have gradually attracted attention. Various network frameworks have been proposed, which greatly improved the recognition Accuracy. In view of some problems in the action recognition algorithm of deep learning at this stage, this paper proposes a new R2.5D-GRU network. First, the 3D convolution is decomposed into a two-dimensional spatial convolution and a one-dimensional time convolution, and the low-level spatio-temporal features are extracted, and the high-level temporal features are extracted using GRU for temporal modeling. Experimental results show that the algorithm proposed in this paper performs better than some existing mainstream algorithms in the UCF101 data set.