DWnet: Deep-wide network for 3D action recognition

Action recognition plays an important role in human–robot cooperation and interaction. By recognizing human actions, robots can imitate or reproduce human actions and obtain skills. Recently, convolutional neural networks (CNNs) have been widely used to recognize actions based on 3D skeleton. Good performance has been achieved due to the approximation capability gained from the depth of the model. Unfortunately, in the mainstream deep structures, dropout and fully connected layers are usually used to classify actions. That is to say, ensemble is used to guarantee the recognition performance, which decreases the computational efficiency. In order to improve the computational efficiency, we propose in this paper a deep-wide network (DWnet) to recognize human actions based on 3D skeleton. Specifically, we modify the decision-making mechanism of the deep CNN with a shallow structure, which improves the computational efficiency. The state-of-the-art deep CNN is used to extract spatial–temporal features from the skeletal sequence. Then features are transformed into a higher dimensional feature space to obtain global information and classified by the modified decision making mechanism. Experiments on two skeletal datasets demonstrate the advantage of the proposed model on testing efficiency and the effectiveness of the novel model to recognize the action. The code has been publicly available at https://github.com/YHDang/DWnet. •We propose a novel deep-wide network to recognize actions based on 3D skeletal data.•The proposed method is efficient while maintaining the testing accuracy.•A novel decision-making mechanism is proposed to improve the testing efficiency.•Experiments show that the DWnet is the most efficient while maintaining the accuracy.