Self-calibrated driver gaze estimation via gaze pattern learning

Driver’s eye gaze is regarded as an important clue for evaluating the driver’s awareness in the Advanced Driver Assistance Systems and the Automated Driving Systems. Existing driver gaze estimation methods have poor performance on uncalibrated drivers or camera views due to the tedious calibration process. To solve this problem, a knowledge-based solution based on domain prior of typical driver gaze patterns is proposed to realize self-calibration for real driving in naturalistic driving scenarios, which implicitly select the representative time samples of the forward-view gaze zone, the left-side mirror, the right-side mirror, and the rear-view mirror, the speedometer, and the center stack as calibration points. Supporting by the gaze pattern learning algorithm, the proposed method can extract the relevant driver status features gradually and update the estimation parameters periodically. Experimental results demonstrate that the auto-calibrated gaze estimation method can achieve automatic gaze calibration for gaze tracking during in-the-wild on-road driving, and does not require any cooperation from the driver.