Robust Eye Feature Points Tracking

Eye is an important part of facial features, including lot of information. Eye detection and tracking is an important topic and has become a focus in the field of computer vision. Simultaneously, eye tracking is applied in many applications, such as driver fatigue detection, gaze estimation, facial expression analysis and human-computer communication.However, due to the variability in appearance caused by pose, illumination, and facial expression change, it's very difficult to track robustly and rapidly. Some approaches only localize the key points but do not describe shape information. Some methods only depict the contour in near-front view but do not consider other views. Considering these factors, our work can robustly track the eye, in detail describing the eye contour and the visible iris center.In this paper, we propose a novel approach for eye feature points tracking. The whole architecture is composed of two modules, eye image patch tracking and location of feature points. For eye image patch tracking, IVT tracker is utilized to track robustly. Note that, the first stage defines the unified eye sub-image and the image-plane origin that are used in the second stage.For eye contour control points tracking, a novel on-line affine manifold model is presented, in which the sequentially learning shape and texture are modeled in the first stage and non-iterative recovering estimation in the second stage. Upon this, the point local feature matching, in which combination of SIFT and RGB feature is created regarding the local information, is also utilized to learn the eye contour points. In addition, the visible iris center is confirmed by the fusion result of the adaptive black round template matching and point local feature matching. Finally, Kalman filtering is required to refine the feature points.Experiments indicate that the proposed method can track eye feature points accurately in the PC or domestic camera captured image streams with different head pose, scale change and eyeball rotation. The average RMS value of the difference between tracking result and hand marked result is approximately 1.5.