A Robust Extrinsic Calibration Method for Non-Contact Gaze Tracking in the 3-D Space

In general, 3-D gaze tracking methods employ both a frontal-viewing camera and an eye-capturing camera facing the opposite direction to precisely estimate the point-of-regard (POR) in the 3-D space. The extrinsic calibration of these two cameras for accurate 3-D gaze tracking is a challenging task. This paper presents a robust extrinsic calibration method for non-contact gaze tracking in the 3-D space. Even in a noisy environment, the extrinsic calibration parameters are precisely estimated by minimizing the proposed cost function consisting of both the angular and Euclidean errors. Furthermore, using the estimated parameters, the 3-D POR is exactly determined based on the two-view geometry. Compared with the conventional methods, the proposed method provides superior results in experiments considering various factors such as the noise level, head movement, and camera configuration. In real experiments, we achieved an average Euclidean error of 12.6 cm and the average angular error of 0.98° when estimating the 3-D coordinates of PORs that were 4-8 m away from the user.