Coordinate transformation and connection feature for Skeleton-based action recognition

Abstract

Graph structure is an important part of Graph convolutional networks (GCNs), which can reflect the connection between each nodes of non-Euclidean data. A connection feature between nodes is hidden in graph structure, which can provide additional spatial features that represent the relationship between human joints. However many GCNs-based methods ignore these spatial features. We put forward a connection feature extraction module, which can obtain implicit connection between human joints, and extract the implicit spatial features from the structural connection and implicit connection of human joints. In order to enhance the temporal representation, we propose a long-range frame-difference feature extraction module. Furthermore, we also propose a coordinate transformation module, which can map joint from Cartesian coordinates to spherical coordinates to extract more representative features. Experiments show that our method outperforms several advanced methods.