以眼角定位為基礎之眼球模型與凝視點估計

Abstract

可見光眼動儀無須額外的紅外光源，並且提供比一般紅外光眼動儀系統更好的使用者體驗。近年來，可見光眼動儀系統於人機系統互動的領域中變得更加重要，一般支援頭部移動的紅外光眼動儀系統已可在市場上看到，可見光系統卻還未有相關產品。可見光眼動儀系統在讓使用者自由地移動頭部的開發中仍留有許多進步空間。這是因為其系統精準度在頭部移動的情況下會降低。然而，頭動補償的問題並不能直接被簡化為眼球定位方面的問題，由於眼球定位可能不夠準確，眼球模型亦跟著建構的不準確，這導致了後續的凝視點估計結果無法接受。在本研究中，我們探討以內眼角為錨點的立體眼球模型，並分析了眼球中心和內眼角之間的相對關係，利用這個數學關係，可以改善眼球模型的建構、虹膜匹配和頭動補償等階段的精準度。除此之外，於頭動補償的階段，本研究可以準確地估計相機，屏幕和人眼之間的相對位置。實驗結果顯示本研究提出的方法能夠容忍眼球定位些微不準，並且提高凝視點估計精準度。

The visible-spectrum gaze tracker (VSGT), which is designed without the extra Infra-ray (IR) illumination, has provided a more superior user experience than the traditional IR-based gaze tracker. It has become an important human machine interface, while it remains challenging for allowing the users to move their heads. That is, the system performance appears significantly inferior due to head movement. However, the head movement compensation cannot be simply formulated as an eye detection problem. The minor error of the eye detection algorithm leads to an unacceptable result for the gaze estimation due to the fact that an incorrect eyeball model will be adopted. In this thesis, we further explore the 3-D eyeball model anchored by the inner eye corner point. The relative location between the eyeball center and the inner eye corner is analyzed. This feature is used to improve the eyeball model construction, the limbus circle matching, and the head motion compensation. In addition, the proposed approach can accurately estimate the relative positions/poses among the camera, the screen, and the human eyes. The experimental results show the proposed approach tolerates a wide range of the estimation error for the eye detection. Thus, the gaze point estimation performance could be remarkably improved.