Cross-Person Human Activity Recognition Method Based on the HARG Model
Xinyang Zhang, Peng Xin, Jing Zhang, Miao Liu, Hengyi Yue, Minghui Hou · 2023
In recent years, Sensor-based human activity recognition technology has found widespread applications in various fields. Recently, Transformer and its variant models, further improving the effectiveness of activity recognition models. However, almost all established models assume that the training and test data maintain data distribution consistency. In practical applications, due to differences in body types, activity styles, and habits, test data distribution changes, leading to a sharp drop in model recognition performance. In this paper, we integrate the Domain-Invariant and the Domain-Specific Feature Module, proposing the Deep Adaptive Domain Feature Module that focuses on deep-level domain features. This is a Domain Generalization-based representation learning that can effectively improve the model’s generalization performance. Specifically, we use Transformer and CNN as the backbone networks, adopt a novel parallelized network structure, proposed the Human Activity Recognition Generalization (HARG) cross-Person human activity recognition model. We systematically evaluate the proposed model on UCI-HAR and Opportunity public datasets, the experiments show that the HARG model has an absolute advantage in cross-domain activity recognition, outperforming current models in recognition performance.