Non-intrusive Automatic 3D Gaze Ground-truth System

Abstract

Driver distraction has surfaced as a significant safety issue worldwide, and the capacity to track a driver's attention via monitoring its gaze direction is one of the most critical features in the modern Driver Monitoring System (DMS). Deep learning based gaze estimation has grown in popularity due to its robustness across operating conditions. Though appropriate network structure design and parameters tuning are important, accurate ground-truth estimation for millions of gaze training images to build the model also plays a critical role in achieving high-quality gaze estimation results. This paper proposes a non-intrusive automatic 3D ground-truth data collection system for large-scale on-bench and in-car data collection, using gamified camera calibration, occlusion invariant mirror-based camera localization, and noise-robust 3D reconstruction algorithms. Experimental results are provided to demonstrate the system's accuracy and robustness even in challenging conditions.