Assessing Driver Gaze Location in a Dynamic Vehicle Environment

Measuring the human gaze is an important area of research due to this measurement's ability to give insight into what or where a person is focused and/or paying attention to. However, gaze has been very challenging to measure effectively and then convert into a metric. The problem of measuring human gaze is challenging in the context of dynamic environments with motion of the subject or the environment itself. One domain of research that has sought these gaze-related attention metrics has been the area of automotive driver assessment. Being able to understand if a driver is looking at relevant areas as well as scanning the road for hazards is a valuable metric to evaluate if an individual is fit to drive. Eye-tracking glasses measure where a person is looking relative to their head position but do not map this information against important regions within the visual field. This paper provides a computationally scalable method to identify relevant regions within a dynamic visual field and allow for the measurement of what a driver is focused on, reducing the need for extensive manual segmentation. The paper provides a method of identifying the windshield and other key regions within a motor vehicle typical for a driver's field of view. The identification of key regions was accomplished through the application of convolutional neural networks (CNNs) with a Dice score of 0.9404 The model is then shown to allow for the assessment of visual focus for drivers.