基于暗瞳图像的人眼视线估计

虹膜外边缘受眼睑遮挡较为严重时, 会给虹膜中心的准确提取造成很大的困难. 为此, 提出利用放置在相机轴外的红外光源产生的暗瞳图像估计瞳孔中心, 该方法避免了提取虹膜外边缘遇到的遮挡问题. 首先利用角膜反射光斑在相机像面中的位置估计角膜所在球体中心的三维空间坐标, 作为眼球的平动信息; 然后考察瞳孔中心与角膜球体中心在相机成像面投影位置的相对偏移, 作为眼球的转动信息; 最后利用人工神经网络完成视线特征向量与注视点坐标间的映射. 在人眼区域定位的问题上, 利用两部大视场相机, 采用自适应增强算法和主动表观模型算法实现眼部区域的准确定位, 该步骤可以将提取反射光斑和瞳孔中心需要考虑的图像区域限定在较小的范围内. 实验结果表明, 本文视线估计方法在水平方向上的平均误差为0.62°, 在竖直方向上的平均误差为1.05°, 是解决视线点估计的有效方法. 关键词： 暗瞳 人工神经网络 自适应增强 主动表观模型

Eye gaze tracking based on dark pupil image

The accurate localization of iris center is difficult since the outer boundary of iris is often occluded significantly by the eyelids. In order to solve this problem, an infrared light source un-coaxial with the camera is used to produce dark pupil image for pupil center estimation. Firstly, the 3D position of the center of cornea curvature, which is used as translational movement information of eyeball, is computed using two cameras and the coordinates of two cornea reflections on the cameras' imaging planes. Then, the relative displacement of pupil center from the projection of the cornea curvature center on 2D image is extracted, describing the rotational movement of the eyeball. Finally, the feature vector is mapped into coordinates of gazing point on the screen using artificial neural network. As for the eye region detection problem, two wide-view webcams are used, and adaptive boosting+active appearance model algorithm is adopted to limit the region of interest within a small area. The result of our experiment shows that the average root-mean-square error is 0.62° in horizontal direction and 1.05° in vertical direction, which demonstrates the effectiveness of our solution in eye gaze tracking.
Keywords： dark pupil artificial neural network adaptive boosting active appearance model