Abstract:Recently, many studies have actively dealt with the issue of driver-gaze tracking for monitoring the forward gaze and physical condition. Driver-gaze tracking is an effective method of measuring a driver's inattention that is one of the major causes of traffic accidents. Among many gaze-tracking methods, the corneal specular reflection (SR)-based method becomes ineffective, unlike in an indoor environment, when a driver's head rotates, which makes SR disappear from input images or disperses SR in the lachrymal gland or eyelid, thereby increasing the gaze-tracking error. Besides, since a driver's eyes in a vehicle environment need to be captured in a wide range covering his head rotation, the eye region is captured in a relatively low resolution compared to face-only images taken in indoor environments at the same resolution, making pupil and corneal SR difficult to detect accurately. To solve these problems, we propose a fuzzy-system-based method for detecting a driver's pupil and corneal SR for gaze tracking in a vehicle environment. Unlike existing studies detecting pupil and corneal SR in both eyes, the method proposed in this research uses the results of a fuzzy system based on two features considering the symmetrical characteristics of face and facial feature points to determine the status of a driver's head rotation. Based on the output of the fuzzy system, the proposed method excludes the eye region, which is very likely to have a high error rate of detection due to excessive head rotation, from the detection process of the pupil and corneal SR. Accordingly, the proposed method detects pupil and corneal SR only in the eye region that apparently has a low detection error rate, thereby achieving accurate detection. We use 20,654 images capturing 15 subjects (including subjects wearing glasses), who gaze at pre-set fifteen regions in a vehicle, to measure the detection accuracy of the pupil and corneal SR for each region and the gaze tracking accuracy. Our experimental results show that the proposed method performs better than existing methods.