Deep Learning Approach for the Future Advanced Driver Assistance Systems

Technological development made vehicle safer thanks to the advanced driving assistance systems (ADAS). There exists passive systems, like ABS that boosts vehicle performance during an emergency braking, and active systems, like active cruise control that actively drives car during normal cruise. However, ADAS need to further evolve to continue the risk reduction and to increase automation level. Physiological sensing and deep learning are promising options at these regards. This paper will present a methodology to estimate the driver drowsiness from photoplethysmography (PPG) waveform, acquired by no-invasive probe embedded in the vehicle steering wheel. PPG signal is treated using the innovative hyper-filtering process, then it is elaborated by a convolutional deep architecture. A driver sample, made by 70 people distributed by gender and age, was selected to validate the approach, monitoring PPG and confirming the actual level of drowsiness by electroencephalography. The proposed network makes a classification task to distinguish a wakeful drive from a drowsy one. The developed method has an accuracy of 99%, which were higher than other networks considered for the benchmark. The evaluation of drowsiness permits to evaluate the actual attention level of the driver, augmenting the interaction between human pilot and ADAS.