Sharing-Net: lightweight feedforward network for skeleton-based action recognition based on information sharing mechanism

With the development of metaverse, augmented reality and human–robot teleoperation, action recognition plays an increasingly important role. In this work, we propose a Lightweight Feedforward Cross-channel Information Sharing Network (Sharing-Net) for action recognition. A multi-feature input module is constructed, which includes Cartesian Motion features, Global Joint Distances (GJD), and Global Joint Angles (GJA). The three types of features can tackle the problems of velocity differentiation, viewpoint diversification and object-distance variation, respectively. In order to take full use of the restricted parameters caused by the lightweight structure to enhance the accuracy under the premise of guaranteeing high speed, a multi-feature cross-channel information sharing mechanism is proposed. Dynamic nonlinear composite mapping between feature channels uses cross-channel residual blocks to share data information and establish coupling relationships. Extensive experiments on 3 public datasets and a self-built dataset verify the effectiveness of proposed methods. Compared with the state-of-the-art (SOAT) methods, Sharing-Net achieves the best accuracy with high speed on JHMDB and SHREC and performs superior balance of accuracy and computational cost on NTU RGB+D.