Remote photoplethysmography (rPPG) provides a non‑contact heart rate (HR) estimation from facial videos. However, real‑world deployment is distorted by motion artifacts, illumination variations, and the high computational cost of deep learning and transformer models. In this paper, we propose a lightweight multi‑scale temporal convolutional network (TCN) designed for CPU‑based edge devices. The framework dynamically extracts green‑channel signals from three facial regions (forehead, left cheek, right cheek) and a background reference using MediaPipe landmarks. These 1‑D temporal traces are processed by a four‑block TCN with parallel multi-scale kernels, exponential dilation, and squeeze‑and‑excitation attention, enabling the model to capture both fast systolic peaks and slower diastolic trends. Regularization via ROI dropout further improves generalization. The entire system uses only 0.1 million parameters and requires 5.5 MFLOPs. Evaluated on two public datasets under a subject‑disjoint protocol, our method achieves a mean absolute error (MAE) of 0.86 BPM on UBFC‑rPPG and 1.95 BPM on COHFACE, outperforming traditional and many deep learning baselines while being an order of magnitude more efficient than transformer‑based models. The results demonstrate that compact, landmark‑guided temporal networks can deliver near‑state‑of‑the‑art accuracy with minimal computational resources, making them suitable for real‑time, on‑device physiological monitoring. |
