Driver head pose and gaze estimation based on multi-template ICP 3-D point cloud alignment

Abstract

Head movements, combined with the line of gaze, play a fundamental role in predicting the driver's actions and in inferring his intention. However, a gaze tracking system for automotive applications needs to satisfy high demands: It must not disturb the driver in his freedom of movements, it must cover large and fast head turns in yaw and pitch, be resistant to changing illumination conditions, be fast enough to recognize fast mirror checks, which are performed almost exclusively through eye rather than head movements, and be accurate and reliable enough to derive high quality information for driver assistance systems relying on their output. In this work a multi-template, ICP-based gaze tracking system is introduced. The system determines the head pose and subsequently estimates the driver's line of gaze by analyzing the angles of the eyes. Due to a fast search of correspondences, and switching between point-to-point and point-to-plane alignment, real-time performance and high accuracy can be achieved. The system is compared with other state of the art head pose estimation systems based on a publicly available benchmark database, where a classification rate of 92% at a tolerance of 10 degrees in yaw could be achieved. We further show in the experiments section, that head rotations up to 4 radians per seconds can be handled. Taking the angles of the eyes into account, rather than the head pose only, the driver's line of sight could be successfully mapped to particular regions of interest.