3D gaze estimation without explicit personal calibration. (July 2018)

• Record Type: Journal Article
• Authors: Wang, Kang; Ji, Qiang
• Is Part Of: Pattern recognition. Volume 79(2018:Jul.)
• Page Start: 216
• Page End: 227
• Publication Date: 2018-07
• Subjects: Gaze estimation -- Implicit calibration -- Natural constraints -- Human computer interaction; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2018.01.031
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A non-intrusive and user-friendly eye gaze tracking system is proposed. Personal eye parameters can be implicitly calibrated with natural constraints. Propose the hard-EM algorithm to solve the constrained unsupervised regression problem. The proposed method achieves comparable gaze estimation accuracy with state-of-the-art implicit calibration methods, while is less restricted and can be applied to a wider range of practical applications. Abstract: Model-based 3D gaze estimation represents a dominant technique for eye gaze estimation. It allows free head movement and gives good estimation accuracy. But it requires a personal calibration, which may significantly limit its practical utility. Various techniques have been proposed to replace intrusive and subject-unfriendly calibration methods. In this paper, we introduce a new implicit calibration method that takes advantage of four natural constraints during eye gaze tracking. The first constraint is based on two complementary gaze estimation methods. The underlying assumption is that different gaze estimation methods, though based on different principles and mechanisms, ideally predict exactly the same gaze point at the same time. The second constraint is inspired by the well-known center prior principle, it is assumed that most fixations are concentrated on the center of the screen with natural viewing scenarios. The third constraint arises from the fact that for console based eye tracking, human's attention/gaze are always within the screen region. The final constraint comes from eye anatomy, where the value of eye parameters must be within certain regions. The four constraints are integrated jointly and help formulate the implicit calibration as a constrained unsupervised regression problem, which can be effectively solved through the proposed iterative hard EM algorithm. Experiments on two everyday interactions Web-browsing and Video-watching demonstrate the effectiveness of the proposed implicit calibration method.