55 Rides: Attention annotated head and gaze data during naturalistic driving

Trained eye patterns are essential for safe driving. Whether for exploration of the surrounding traffic or to make sure that a lane is clear through a shoulder check - quick and effective perception is the key to driving safety. Surprisingly though, free and open access data on gaze behavior during driving are yet extremely sparse. The environment inside a vehicle is challenging for eye-tracking technology due to rapidly changing illumination conditions, such as exiting a tunnel to brightest sunlight, proper calibration and safety. So far, available data exhibits environments that likely influence the viewing behavior, sometimes dramatically (e.g., driving simulators without mirrors, limited field of view). We propose crowd-sourced eye-tracking data collected during real-world driving using NIR-cameras and illuminators that were placed within the driver's cabin. We analyze this data using a deep learning appearance-based gaze estimation, with raw videos not being part of the data set due to legal restrictions. Our data set contains four different drivers in their habitual cars and 55 rides of an average of 30 minutes length. At least three human raters rated each ride continuously with regard to driver attention and vigilance level on a ten-point scale. From the recorded videos we extracted drivers' head and eye movements as well as eye opening angle. For this data, we apply a normalization with respect to different placement of the driver monitoring camera and demonstrate a baseline for driver attention monitoring based on eye gaze and head movement features.