8 Head Pose as an Indicator of Drivers’ Visual Attention

Estimating the direction of visual attention is very important to predict driver’s distraction, assess his/her situational awareness, and improve in-vehicle dialog systems. Tracking eye movement can be an accurate measure to identify the exact location of the gaze of a driver. However, robustly measuring gaze in a driving environment is challenging due to changes in illuminations, occlusions, and changes in head poses from the drivers. Head pose provides a coarse estimate of the gaze direction, which might be helpful in determining the visual attention of the drivers for most applications. To what extent can the head pose be a useful indicator of visual attention? In addition to head pose, gaze activities are characterized by eye movements. Therefore, the relation between head pose and gaze is not deterministic. This chapter summarizes our effort to understand and model the relation between head pose and visual attention. We are interested in understanding how much the head pose deviates from the actual gaze of the driver, and how much the head pose varies for a given gaze direction. We observe that the deviation is much higher in the vertical direction compared to the horizontal direction, making it more difficult to estimate vertical gaze. We observe that as the direction of visual attention is directed further away from the frontal direction, the deviation between gaze direction and head pose increases. Given that the relations between gaze and head pose are not deterministic, we propose probabilistic maps to describe visual attention. Instead of estimating the exact direction of the gaze, this formulation provides confidence regions that are mapped to either the windshield or the road scene. We describe a parametric probabilistic map built with Gaussian process regression (GPR), and a nonparametric probabilistic map built by upsampling convolutional neural networks (CNNs).