Illumination-Robust Camera-Based Heart-Rate Estimation for Physiological Sensing in Robots

Researchers present a transformer-based framework for non-contact heart-rate estimation using RGB cameras, addressing the challenge of varying illumination conditions. The system achieves 0.79 bpm mean absolute error and 0.982 correlation on illumination-varied datasets, significantly outperforming existing baselines and enabling practical physiological sensing for service robots.

This research tackles a fundamental challenge in robotic perception: reliable physiological monitoring across real-world lighting conditions. Heart-rate estimation through remote photoplethysmography (rPPG) has long promised non-invasive health monitoring for human-robot interaction, but illumination variability has prevented practical deployment. The proposed spatial-temporal transformer architecture directly addresses this limitation through multiple technical innovations, including illumination-aware data augmentation and a hybrid loss function balancing temporal waveform accuracy with frequency-domain precision.

The breakthrough emerges from combining three complementary approaches: PRNet-based facial alignment for robust face tracking, spectral-domain supervision that guides the model toward physiologically valid heart-rate frequencies, and controlled weighting between waveform and frequency losses. By achieving 93.6% error reduction compared to PhysFormer—a significant margin—the system demonstrates that lighting robustness is achievable through careful architectural design rather than simply collecting more data.

For the robotics industry, this advancement enables assistive and social robots to monitor user health status during extended interactions, opening applications in elder care, rehabilitation, and stress monitoring. The solution's camera-based approach requires no wearables or additional sensors, reducing implementation friction in real-world deployments. The 0.79 bpm accuracy threshold approaches clinical utility levels, particularly for detecting abnormal heart-rate states rather than absolute measurements.

Future work should focus on cross-dataset generalization, real-time processing constraints for embedded robot systems, and validation with diverse skin tones and facial geometries. Real-world deployment testing with actual service robots will reveal whether controlled laboratory improvements translate to practical robustness in dynamic home and healthcare environments.

• →New spatial-temporal transformer framework reduces heart-rate estimation error by 93.6% compared to PhysFormer baseline on illumination-varied datasets.
• →Hybrid loss function combining temporal waveform accuracy and frequency-domain guidance proves critical for robust physiological signal extraction.
• →System achieves 0.79 bpm MAE and 0.982 correlation, approaching clinical utility for non-contact health monitoring in robots.
• →Illumination augmentation during training significantly improves real-world robustness across varied lighting conditions without additional hardware.
• →Technology enables camera-based physiological sensing for service robots without wearable requirements, expanding assistive robotics applications.