The optimization of sparse modeling for drowsiness estimation based on general facial skin temperature distribution

In recent years, there has been a growing demand for drivers’ drowsiness detection technology in driver monitoring systems for level 3 automated vehicles. In a previous study, sparse modeling was applied to the facial skin temperature distribution which can be remotely measured with a thermography camera, achieving an estimation accuracy of about 74% in three stages. However, the constructed drowsiness estimation model had low generalization performance because it was an individual model, and the sparse modeling had the property of easily decomposing behavioral indicators other than drowsiness. Therefore, in the construction of a general model using sparse modeling, individual features among subjects may be preferentially decomposed, and a method to absorb these effects is needed. Thus, in this study, we devised an attempt to reduce the influence of decomposition of individual features by applying averaged faces. Partial averaging of images across subjects is thought to remove such effects. In this study, we attempted to construct a general model for drowsiness estimation by applying sparse modeling to facial thermal images averaged across subjects, and to examine the possibility of using averaged facial thermal images in constructing general model. As a result, we obtained an estimation accuracy of approximately 54.6% in average by applying averaging, which is 7% higher than that using the original images, and succeeded in reducing its standard deviation by 4–6.9%. As the result, modeling with averaged images was shown to be effective in improving generality.