Estimating a Driver's Gaze Point by a Remote Spherical Camera

Abstract

In this paper we propose a novel method of detecting a driver's gaze point by a single remote spherical camera. A spherical camera is fixed at the upside of the windshield of a vehicle so that it can observes outside scenes and an inside driver simultaneously. We propose a simplified model of a driver's gaze point detection under this setting-up by approximating the gaze direction vector as that starting from the viewpoint of the spherical camera by omitting the distance of the driver from the spherical camera. Furthermore, two algorithms are proposed and evaluated: one is gaze point estimation based on the geometric model using OpenFace [1], and the other is to use a neural network, which uses the gaze vector and head pose parameters obtained from OpenFace as input of the neural network, to estimate the gaze point in the spherical image. As shown in the experimental results, using a neural network to compensate the measuring errors of OpenFace can achieves a much better result.