Driver's gaze prediction in dynamic automotive scenes
Julian Schwehr, Volker Willert · 2017
For future Advanced Driver Assistance Systems (ADAS), knowledge about what the driver perceived in his surrounding environment is important to estimate the driver's situation awareness. This estimate can then be used for example to adapt the systems' warning and intervention strategies according to the driver's needs. We propose a Dynamic Bayesian Network (DBN) which operates in the ground plane at pixel level and simultaneously tracks two gaze motion models to model the driver's focus of attention. We introduce a new time variant transition probability for motion hypotheses of fixations and saccades combining spatial and temporal domain motivated by human gaze motion characteristics. For environment perception, we solely rely on series sensors, while for gaze tracking, a commercial eye tracker is employed. Our system efficiently smooths the measured gaze target point during estimated fixations while preserving the characteristics of saccadic jump behavior. Thereby, the driver's gaze target in the world is effectively extracted.