Driver drowsiness detection relying on infrared thermal imaging: a machine learning approach

One of the main factors contributing to road accidents all around the world is driver fatigue. For safety reasons, it is essential to identify drowsiness episodes as soon as possible. Numerous studies have shown that PERCLOS, i.e. the percentage of eyelid closure over the pupil across time, is one of the most accurate parameters for drowsiness state assessment. The evaluation of PERCLOS is, however, dependent on the lighting conditions because it is normally computed from the visible video of the subjects. The goal of this study is to get around these constraints by assessing sleepy states using inexpensive, high-resolution thermal infrared technologies. Twelve sleep-deprived participants were chosen for the study, which involved an hour-long driving activity on a driving simulator. Thermal camera Device Alab SmartIr640 was used to capture facial skin temperature throughout the trial, along with facial visible videos of the subjects. Relevant thermal features were estimated from facial regions of interest and extracted over a 30 second time span. A data-driven multivariate machine learning approach based on a three-level Decision Tree Classification of the drowsy state (AWAKE class: PERCLOS<0.15, FATIGUE class: 0.15<0.23, and SLEEPY class: PERCLOS>0.23) was developed. The average classification accuracy was 0.72±0.05 (mean ± standard deviation). These preliminary results point to the possibility of determining drivers' drowsiness based on facial thermal traits, circumventing the limitations associated with lighting and eye detection that are typical of conventional approaches.