Different Encoding of High-Low Level Semantic in Point Cloud Video for 3D Human Action Recognition

3D action recognition is crucial in a variety of fields, including human-computer interaction, sports analytics, and smart home systems. However, in previous work, low-level semantic information and high-level semantic information in human action data failed to be treated differently, which limits the information feature extraction capability of deep neural networks. To address the above problems, we propose a multiscale spatio-temporal frequency action recognition network (MSTF-net). The MSTF-net is mainly composed of multiscale low-level action semantic module and frequency enhancement high-level action semantic module. In the multiscale low-level action semantic module, we decompose the original point cloud sequences in both spatial and temporal dimensions, and adopt a multiscale strategy in order to comprehensively enrich the low-level semantic information in the spatial dimension. The frequency enhancement high-level action semantic module adaptively learns high-level semantic information for each channel based on the frequency information in the temporal dimension, thus improving the discriminative power of action spatio-temporal features. Comprehensive evaluations on commonly utilized 3D action recognition datasets reveal the network's perfect performance.