Multi-Hypothesis Multi-Model Driver's Gaze Target Tracking

For a safe handover of the driving task or driver-adaptive warning strategies the driver's situation awareness is a helpful source of information. In order to estimate and track the driver's focus of attention over time in a dynamic automotive scene, a Multi-Hypothesis Multi-Model probabilistic tracking framework was developed in which we postulate consistency between machine and human perception during gaze fixations. Within this framework, we explicitly included target object motion in the spatial transition step and integrated spatiotemporal models of human-like gaze behavior for fixations and saccades in the motion transition. This elaborate design makes the target estimation robust and yet flexible. At the same time, the representation in continuous 2D coordinates makes the algorithm run in real time on a standard laptop. By incorporating dynamic and static potential gaze targets from an object list and a free space spline, the algorithm is in principle independent from the applied sensor setup. The benefit of the proposed model is presented on real world data where the filter's tracking performance as well as the driver's visual sampling are presented based on an exemplary scene.