Multi-depth cross-calibration of remote eye gaze trackers and stereoscopic scene systems

Abstract

We present a robust and accurate technique for the cross-calibration of 3D remote gaze trackers with stereoscopic scene vision systems between which no common imaging area exists. We empirically demonstrate that a multi- depth calibration approach yields remarkably superior results for obtaining 3D Point-of-Gaze (PoG) when compared with traditional methods using monocular scene cameras and co- planar eye gaze calibration points. I. INTRODUCTION Remote gaze trackers have been in use for various ap- plications together with scene cameras to determine the point of gaze (PoG) of human subjects on an imaged scene. Several types of applications benefit from the use of such systems including vehicle driver training and advanced driver assistance systems, the context in which the results herein have been obtained. The task of projecting back the 3D gaze direction onto the imaged scene requires a cross-calibration between the remote gaze tracking device and the scene. In most if not all of commercially available systems, this type of calibration is performed by requiring that test subjects fixate specific, pre- selected image points on a planar surface placed at a known distance such as on a computer screen or, by using a scene image from a monocular camera and treating it essentially as a 2D object (co-planar fixation calibration points). Such approaches are dependable when the subject's eye center is not highly offset from the scene camera(s). In other words, because the origin of the reference system of the scene cameras and the subject's eye center approximately coincide, the projection ray of any fixated object will also approximately lie on the line of sight regardless of the depth of the object. In such cases, the calibration process may be performed correctly. Otherwise, objects with different depths along the line of sight correspond to different image locations, and must be calibrated for as such. Our primary goal is to determine whether driver intent and driving-related actions can be predicted from qualitative and quantitative analyses of driver behavior. Toward this end, it is necessary to establish the correspondence between cephalo- ocular behavior and visual stimuli in such a way as to identify the elements in the visual field to which driver attention turns to. This type of information in turn may facilitate the task