Temporal Step Supervision for Robust Step Counting from Wearable Accelerometers.

Step-counting remains difficult in irregular walking, where step timing and signal patterns are less consistent than in controlled walking. We propose a step-counting model that uses step timing annotations as temporal supervision during training. Each step timestamp is expanded into a short temporal label interval, and a U-Net is used to localize step events because its dense prediction structure is well suited to temporal localization. The model is trained to estimate both where steps occur in the window and how many steps the window contains. The final count is obtained by summing the predicted values over time and dividing by the label width. On the Clemson wrist-worn dataset, the model achieves MAPE values of 4.61%, 9.90%, and 15.78% under regular, semi-regular, and irregular walking, respectively. Compared with classical step-detection algorithms and reproduced CNN, WaveNet, and LSTM baselines, the proposed model shows substantially lower error under irregular walking. On OxWalk, the model achieves 13.06% MAPE using only OxWalk training data, reaching performance comparable to fine-tuned Stepcount SSL, which relies on large-scale self-supervised pretraining. These results suggest that temporal step supervision improves robustness under irregular walking and can reduce dependence on large-scale external pretraining.