A Temporal-Channel-Spatial Attention Network for Skiing Recognition Based on Multigraph Generation

Sports activity recognition using wearable sensors has become a hot research point, promoting the development of intelligent sports. Skiing activity recognition can enhance both skiers’ performance and safety. However, existing sports activity recognition studies lack sufficient focus on skiing, and most existing studies suffer from issues, such as difficulty in outdoor data collection, low accuracy in recognition, and neglect of falling recognition. To address these challenges, this article develops a new multigraph generation temporal-channel–spatial attention (MGG-TCSA) network for human skiing activity recognition using inertial measurement units (IMUs). The MGG-TCSA network comprises two key components: the multigraph generation (MGG) block, which represents temporal and spatial correlations within sensor data, and the temporal-channel–spatial attention (TCSA) block, which combines recurrent and convolutional neural networks (CNNs) with attention mechanisms. Remarkably, this study pioneers the use of attention mechanisms across temporal, channel, and spatial dimensions for skiing activity recognition. We evaluate the proposed MGG-TCSA network through ablation and comparison experiments on both the self-collected indoor Skiing dataset and the publicly available Opportunity dataset. The results demonstrate that the MGG-TCSA network outperforms baseline models and most existing models. In addition, we validate the model’s effectiveness in indoor and outdoor skiing, demonstrating the feasibility of utilizing indoor skiing data for outdoor skiing research and overcoming the challenges of outdoor data collection.