Selective attentional point-tracking through a head-mounted stereo gaze tracker based on trinocular epipolar geometry

This paper proposes a selective-gaze tracking method for a head-mounted stereo gaze tracker using trinocular epipolar geometry. As the tracker consists of one front-view camera and stereo eye-view camera to extract three-dimensional attentional points, the relationship between three cameras can be considered as a kind of trinocular vision system. A process of extracting a line of sight from each eyeball movement due to selective attention is modeled as an imaging process of an attentional point in 3D space. First, two line-of-sights are calculated from eyeball images of stereo eye-view cameras. Secondly, two epipolar lines are projected simultaneously on the forward-looking image from front-view camera. Then, the intersection point between them is determined as a current point-of-interests in the image. This selective attentional point is tracked continuously in dynamic images in a frame rate. Although image distortion problem due to the eyeball shapes makes the trinocular relation difficult, the proposed sensor calibration algorithms resolves it effectively. In experiments, the proposed gaze tracking method is applied to car systems operated by the driver, and its accuracy and effectiveness are tested and verified under outdoor environments. As the proposed method doesn´t use any restriction of regarding target object distance from the tracker in three-dimensional space, this system has the advantage of being applicable to moving objects in real 3D space.