STC-rPPG: A Remote Heart Rate Estimation Network Based on Spatiotemporal-Channel Feature Collaborative Learning

Remote photoplethysmography (rPPG) measurement has emerged as a prominent technique for heart rate (HR) estimation. However, its accuracy is often affected by information redundancy, illumination variations, and motion artifacts present in raw video frames, as well as insufficient collaborative learning among spatiotemporal and channel features. To address these limitations, we propose the STC-rPPG network, which simultaneously models spatiotemporal dynamics and inter-channel correlations in facial videos to improve HR estimation accuracy. Specifically, global dependencies are captured by the facial shallow feature extraction (FSFE) module to suppress noise interference. The core innovation lies in the spatio-temporal feature extraction (STFE) module, which integrates multi-scale temporal modeling with spatial dimension. This enhances discriminative feature extraction while reducing redundant information. Furthermore, a hybrid loss function combining time- and frequency-domain constraints is introduced to optimize feature representation learning. Extensive experiments on the PURE and UBFC-rPPG datasets demonstrate the superiority of STC-rPPG, achieving mean absolute error (MAE) of 0.23 bpm and 0.49 bpm, and root mean square error (RMSE) of 0.63 bpm and 0.96 bpm, respectively. These results outperform most existing state-of-the-art models. This study provides new solution ideas for the development of practical remote HR measurement. The codes are available at https://github.com/LeguangWang/STC-rPPG.