Evaluation and Validation of Distraction Detection Algorithms on Multiple Data Sources

Abstract

Many researchers have developed algorithms to detect distraction, but they have yet to be validated on multiple data sources. This study aims to evaluate these algorithms by comparing their ability to detect distraction and predict event likelihood. Four algorithms that use measures of cumulative glance, past glance behavior, and glance eccentricity were used to understand the distracted state of the driver and were validated on two separate data sources: naturalistic and experimental data. Results showed that there was a higher likelihood of event detection when cumulative glances were considered. Glance eccentricity was best for predicting distraction. Future research can use these findings to design mitigation systems that give drivers feedback in instances of high crash likelihood.