In wearable sensor-based human activity recognition (HAR), the traditional sliding window method encounters the challenge of multiclass windows in which multiple actions are combined within a single window. To address this problem, an approach that predicts activities at each point in time within a sequence has been proposed, and U-Net-based models have proven to be effective owing to their excellent space-time feature restoration capabilities. However, these models have limitations in that they are prone to overfitting owing to their large number of parameters and are not suitable for deployment. In this study, a lightweight U-Net was designed by replacing all standard U-Net convolutions with depthwise separable convolutions to implement dense prediction. Compared with existing U-Net-based models, the proposed model reduces the number of parameters by 57–89%. When evaluated on three benchmark datasets (MHEALTH, PAMAP2, and WISDM) using subject-independent splits, the performance of the proposed model was equal to or superior to that of all comparison models. Notably, on the MHEALTH dataset, which was collected in an uncontrolled environment, the proposed model improved accuracy by 7.89%, demonstrating its applicability to real-world wearable HAR systems.
