On driver gaze estimation: Explorations and fusion of geometric and data driven approaches

Gaze direction is important in a number of applications such as active safety and driver's activity monitoring. However, there are challenges in estimating gaze robustly in real world driving situations. While performance of personalized gaze estimation models has improved significantly, performance improvement of universal gaze estimation is lagging behind; one reason being, learning based methods do not exploit the physical constraints of the car. In this paper, we propose a system to estimate driver's gaze from head and eye cues projected on a multi-plane geometrical environment and a system which fuses the geometric with data driven learning method. Evaluations are conducted on naturalistic driving data containing different drivers in different vehicles in order to test the generalization of the methods. Systematic evaluations on this data set are presented for the proposed geometric based gaze estimation method and geometric plus learning based hybrid gaze estimation framework, where exploiting the geometrical constraints of the car shows promising results of generalization.