NeuralPOS: Physiological Measurement via Remote Photoplethysmography for Driver Monitoring

Physiological measurement of the cardiovascular activities of the driver could contribute greatly to driving safety by empowering advanced driver status monitoring functions such as fatigue detection, distraction warning and early heart failure diagnosis. However, traditional contact-based physiological measurement methods are not suitable for the driver monitoring scenarios. In this study, we leverage the remote pho-toplethysmography (rPPG) technology and propose NeuralPOS, a neural network specifically designed for driver physiological measurement task. First, we give a detailed analysis on the unique challenges of driver rPPG measurement, identify that the instability of ambient light intensity is the main cause of these challenges, and propose to use plane-orthogonal-to-skin (POS) features as a countermeasure to alleviate these challenges. Second, we also identify the data collection quality problem of driver rPPG datasets, and adopt the transfer learning (TL) technique to mitigate the data quality problem. Finally, we conduct extensive comparative experiments and ablation studies on driver rPPG measurement dataset, which shows that our proposed NeuralPOS model outperforms previous state-of-the-art models by a significant margin. NeuralPOS demonstrates promising potential for driver fatigue and distraction detection as well as acute heart diseases warning.