Heart Rate Estimation Method Based on Frame Difference Fusion and Uncertainty Aware Perception Network

Remote photoplethysmography (rPPG) represents an emergings non-contact approach utilizing camera technology, showing great potential in the field of remote cardiac health monitoring. In the field of remote heart rate detection using rPPG technology, to address the limitations in spatiotemporal representation of traditional algorithms, we propose a dual-pathway spatiotemporal representation learning network based on frame difference fusion. The network consists of a dynamic frame-difference path and an RGB path. The dynamic frame-difference path takes a sequence of differential frames from facial video as input, then extracts changing features within the residual frame sequence through the use of a 3D convolutional network. The RGB path extracts spatial details like facial texture and background lighting from face frames, compensating for static elements lost in frame differencing. To mitigate heart rate label uncertainties due to head motion, lighting variations, and sensor noise, we adopt a parameterizable Gaussian distribution for modeling, enhanced with temporal supervision to refine network training. Experiments on the UBFC-rPPG and PURE datasets indicate that our proposed network model achieves better performance compared with the benchmark.