Research Of Gaze Estimation Based On Image Processing

Gaze estimation research has experienced significant progress due to thetechnique of high definition camera and image processing. Auxiliaries such asreference light sources were adopted by many related researches since they improvedthe accuracy and the tolerance of the head movement. But the problem is that theyneed complex setup and restricted light environment to make sure the system workwhich significantly limits the application. This paper analyzes the physical features ofeyeball and set up a proper geometric model for computing. By analyzing the relationbetween eyeball, camera and the plane of gaze, a method which only needs one singlecamera is introduced. It needs no reference light and reduces the requirement tominimum.Based on the geometric model we set up, we proved the reasonableness of seeingthe front side of eyeball as a plane. A homography matrix can describes the relationbetween the photo and gaze plane so that the point of gaze can be figure out by thefeature points on the photo. If the user’s head moves, which is inevitable, thehomography matrix will not be valid. So we proposed a method to compensate theerror of gaze estimation caused by the head movement, which allows3degrees-of-freedom out of6of head movement.To verify the accuracy and tolerance of head movement of the system, wepropose an experiment. After the homography matrix is acquired, we test severalgroups of points to gaze while user’s head is at different positions. We output the gazeestimation results before and after the head movement compensation. The final resultof each group has error of less than1cm (about2°) which is practical for manyapplications.