A Novel Approach for Driver Eye Gaze Estimation using Epipolar Geometry

The objective of this research is to estimate the gaze point of a driver using epipolar geometry. The system requires two video inputs from internal and external views of a vehicle. We use SIFT image descriptor with L-2 norm distance to find putative correspondences, and then we use RANSAC algorithm to estimate robust fundamental matrix between two views. Once the underlaying scene geometry is estimated, we can further use affine transformation to model the distribution of the reliable correspondence points across two views. Finally, eye gaze point is estimated on the external reference image plane by mapping the center point in the internal view to the external view using the estimated affine transformation. Experiment results show that epipolar geometry with affine transformation can perform accurate driver’s gaze point estimation for Computer Vision and Robotics Research Laboratory (CVRR) video dataset and produce highest accuracy than pure homography approach and epipolar line searching approach. In addition, our system can also handle the exceptional case as a driver looks at a mirror or a steering wheel which increases the robustness of the gaze estimation system.