Deep Learning-based rPPG Models Towards Automotive Applications: A Benchmark Study

Remote photoplethysmography (rPPG) has the potential to significantly enhance driver safety systems by enabling the detection of critical conditions, such as driver drowsiness and sudden illness, through non-invasive monitoring of cardio-respiratory functions. However, the dynamic environment within a vehicle, characterized by motion artifacts and varying illumination, presents unique challenges for accurate rPPG estimation. In this study, we conducted a comprehensive benchmark of various supervised and unsupervised rPPG algorithms using the MR-NIRP car dataset to assess their performance in automotive settings. Qualitative and quantitative experiments were performed to evaluate and compare several rPPG models designed in stable, noise-controlled environments, highlighting the impact of real-world conditions on model performance. Our findings highlight the promise of machine learning approaches, particularly neural network-based models, in overcoming these challenges and accurately estimating heart and res-piration rates in real-world driving scenarios. This study underscores the potential for integrating rPPG-based mon-itoring systems into vehicles to enhance driver safety and well-being.