DSAFN: Dual-Stream Attention Temporal-Spatial Fusion Network for Human Activity Recognition

Human activity recognition (HAR) technology plays a vital role in the field of mobile computing. With the advancement of mobile device intelligence, the number of sensors carried by wearable devices such as smart watches is also increasing, which greatly improves the model's ability to recognize intricate human behaviors. Nevertheless, this advancement also raises the problem of data heterogeneity, which causes the model to encounter the difficulty of data distribution alignment when fusing multi-modal data across different individuals. This paper proposes a dualstream attention temporal-spatial fusion network for Human Activity Recognition (DSFAN) to fully extract the spatio-temporal features of activity data, and then fuses the features of different modalities based on the attention mechanism to extract domain-invariant features to achieve unsupervised domain adaption (UDA). In addition, we conducted experiments on three public HAR datasets to verify the superiority of our proposed model over other models.