Improving Generalization of Drowsiness Classification From Facial Skin Temperature Distribution Using Data Preprocessing in CNN

The demand for remote biometric technologies to monitor drowsiness levels is expected to increase in the future of autonomous driving. Facial skin is constantly exposed to the outside world and its temperature is an autonomic index that can be measured while driving. Previous studies have attempted to estimate drowsiness from the facial skin temperature distribution (FSTD) using a convolutional neural network (CNN). However, only models specific to a single individual have been constructed. The objective of this study is to improve the generalization of drowsiness classification from the FSTD using data preprocessing in a CNN. The three main approaches of this study are as follows: 1) The input is not the entire face, but only the skin area without the background. 2) Input image preprocessing focuses on the spatial temperature gradient of facial thermal images. 3) Normalization methods are improved to emphasize the spatial variation of skin temperature. In this study, two and three levels of drowsiness were estimated, resulting in 83%–92% and 79%–85% accuracy, respectively. Thus, the generalization of drowsiness classification from the FSTD is considered to have improved.