For skeleton-based action recognition, graph convolutional networks (GCN) have absolute advantages. Existing state-of-the-art (SOTA) methods tended to focus on extracting and identifying features from all bones and joints. However, they ignored many new input features which could be discovered. Moreover, many GCN-based action recognition models did not pay sufficient attention to the extraction of temporal features. In addition, most models had swollen structures due to too many parameters. In order to solve the problems mentioned above, a temporal feature cross-extraction graph convolutional network (TFC-GCN) is proposed, which has a small number of parameters. Firstly, we propose the feature extraction strategy of the relative displacements of joints, which is fitted for the relative displacement between its previous and subsequent frames. Then, TFC-GCN uses a temporal feature cross-extraction block with gated information filtering to excavate high-level representations for human actions. Finally, we propose a stitching spatial–temporal attention (SST-Att) block for different joints to be given different weights so as to obtain favorable results for classification. FLOPs and the number of parameters of TFC-GCN reach 1.90 G and 0.18 M, respectively. The superiority has been verified on three large-scale public datasets, namely NTU RGB + D60, NTU RGB + D120 and UAV-Human.

In recent years, great progress has been made in the recognition of skeletal behaviors based on graph convolutional networks (GCNs). In most existing methods, however, the fixed adjacency matrix and fixed graph structure are used for skeleton data feature extraction in the spatial dimension, which usually leads to weak spatial modeling ability, unsatisfactory generalization performance, and an excessive number of model parameters. Most of these methods follow the ST-GCN approach in the temporal dimension, which inevitably leads to a number of non-key frames, increasing the cost of feature extraction and causing the model to be slower in terms of feature extraction and the required computational burden. In this paper, a gated temporally and spatially adaptive graph convolutional network is proposed. On the one hand, a learnable parameter matrix which can adaptively learn the key information of the skeleton data in spatial dimension is added to the graph convolution layer, improving the feature extraction and generalizability of the model and reducing the number of parameters. On the other hand, a gated unit is added to the temporal feature extraction module to alleviate interference from non-critical frames and reduce computational complexity. A channel attention mechanism based on an SE module and a frame attention mechanism are used to enhance the model’s feature extraction ability. To prevent model degradation and ensure more stable training, residual links are added to each feature extraction module. The proposed approach was ultimately able to achieve 0.63% higher accuracy on the X-Sub benchmark with 4.46 M fewer parameters than GAT, one of the best SOTA methods. Inference speed of our model reaches as fast as 86.23 sequences/(second × GPU). Extensive experimental results further validate the effectiveness of our proposed approach on three large-scale datasets, namely, NTU RGB+D 60, NTU RGB+D 120, and Kinetics Skeleton.