Label Similarity Based Graph Network for Badminton Activity Recognition

The sensor-based human activity recognition is a key technology for modern intelligent sports. However, the complexity of sport activities and lacking of large-scale dataset give rise to the challenges on training effective deep neural networks for it. On image/video-based computer vision tasks, deep learning models can be pretrained on large-scale datasets which are semantically similar with specific tasks. However, we cannot pretrain deep learning models for sensor-based human activity recognition due to lacking public large-scale datasets. To get rid of this problem, we propose a similarity-based graph network for the sensor-based human activity recognition. Specifically, it is a Convolutional Neural Network (CNN) being enhanced with an embedded Graph Neural Network (GNN) for learning the label relationship in terms of two proposed similarity measures. The experimental results on BSS-V2 dataset demonstrate that our proposed network outperforms prior state-of-the-art work by 10.3% in accuracy and 13.3% better than backbone CNN model.