Multi-partitioned Spatiotemporal Graph Convolutional Network for Skeletal Action Recognition

Compared with RGB video data， human bone point data has better environmental adaptability and motion expression ability. Therefore， motion recognition algorithms based on bone point data have been paid more and more attention and research. Recently， the motion recognition model of bone points based on graph convolutional network （GCN） has shown good performance， but most of the models based on GCN often use a fixed spatial configuration partitioning strategy and manually set the connection relationship between each bone point， which can not better adapt to the changing characteristics of different movements. To solve this problem， this paper proposes an adaptive spatiotemporal graph convolutional network with multiple configuration partitions for bone point motion recognition. By searching for a more reasonable number of configuration partitions and adaptive acquisition of node connection relations， the motion features of bone points can be more fully utilized. Experiments on NTU-RGBD datasets and Kinetics-Skeleton datasets show that the proposed method can achieve higher accuracy of motion recognition than most existing literatures.