Eye-tracking solutions are a mainstay in driver modeling research; however, their limitations motivate exploration of complementary methods. One such is available in a class of machine vision models that estimate head pose, eye gaze, facial landmarks, and facial action units. In this article, we share lessons learned from using an open source tool to measure driver behavior. We selected a tool called OpenFace and applied it to an existing drowsy driving dataset that was collected with a commercial camera-based driver monitoring system. Forty participants each drove for around three hours in the early morning. Variables from the driver monitoring system combined with other simulator data and a set of videos from different cameras were recorded as part of the simulator video capture software. OpenFace was used to process the driver-facing video and integrate its output into the dataset. We considered a set of measures appropriate for modeling driver drowsiness, including eye closure, gaze dispersion, head movements, head stillness, and facial contortions (e.g., yawns, brow furrowing, etc.). We present comparisons of similar data between the commercial system and open-source model. Several lessons learned on topics including camera placement, ambient lighting, and data synchronization are shared, along with recommendations for future work.
Read full abstract