Conventional remote photoplethysmography (rPPG) methods suffer from limited robustness in complex scenarios, while deep learning approaches often involve high computational cost and poor interpretability. To address these issues, a lightweight solution that integrates the advantages of machine learning and linear modeling is developed.The proposed method first partitions the facial region into multiple regions of interest (ROI) to extract multi-channel (RGB) temporal signals, followed by correlation-based filtering to achieve temporal alignment between video signals and reference pulse signals. A lightweight network is then employed to learn the optimal linear fusion weights across channels, and pulse signal reconstruction is performed under joint constraints in both the time and frequency domains. In this process, the model preserves the interpretability of traditional linear combination methods while introducing a data-driven weight learning mechanism, enabling adaptive fusion of multi-channel signals. Finally, high-resolution spectral analysis and ridge-based time-frequency tracking are applied for heart rate estimation and dynamic tracking. Experimental results on multiple public datasets, including VitalVideo, MMPD, UBFC-Phys, iBVP, and PhyRec, demonstrate consistent performance across different scenarios. On the VitalVideo dataset, the proposed method achieves a signal-to-noise ratio (SNR) of 6.15 dB and a mean absolute error (MAE) of 1.45 min-1. On the PhyRec dataset, the MAE reaches 0.29 min-1. In the overall evaluation, the method achieves an MAE of 8.45 min-1, a root mean square error (RMSE) of 10.43 min-1, an SNR of -0.55 dB, and an accuracy (ACC) of 0.84. These results indicate that the proposed multi-channel linear fusion strategy improves signal reconstruction quality and heart rate estimation stability while maintaining model interpretability. The method effectively balances interpretability and performance under complex conditions, making it suitable for non-contact physiological signal measurement.