毛云丰 沈文忠 滕童



摘 要: 视线跟踪是人机互动技术中重要的组成部分,可以广泛地应用在机器人、手机、笔记本等设备中。针对传统方法在低分辨率图像上的准确率不高和设计步骤繁琐等问题,提出一种基于深度神经网络的视线跟踪算法。利用卷积神经网络,能快速且准确地定位输入图片上的人眼区域和虹膜中心位置,再利用多项式拟合估计视线方向。实验结果表明,该算法在MPIIGaze数据库中对虹膜中心和眼角位置的标定有着98%的准确性,在Swith数据库中对视线的估计准确率达到了90%左右,证明了该算法在低分辨率图像上有着良好的可行性。
关键词: 视线跟踪; 深度神经网络; 人眼定位; 虹膜中心; 多项式拟合; 视线方向估计
Abstract: Sight tracking is an important part of human?computer interaction technology. It can be widely used in robots, mobile phones, notebooks and other equipments. In allusion to the low accuracy of traditional methods dealing with low?resolution images and their cumbersome design steps, a line?of?sight tracking algorithm based on deep neural network is proposed. With the convolutional neural network, the human eye area and the iris center position on an input image can be located quickly and accurately. The polynomial fitting is used to estimate the line?of?sight direction. The experimental results show that the algorithm has the accuracy of 98% in calibration of the iris center and the eye corner position in the MPIIGaze database, and the accuracy of 90% in estimation of the line?of?sight in the Swith database, which prove that the algorithm has good feasibility for sight tracking in low?resolution images.
Keywords: sight tracking; deep neural network; eye positioning; iris center; polynomial fitting; sight direction estimation
0 引 言
随着科学技术的发展,深度神经网络的兴起,使得计算机视觉技术的发展有了较大的进步。如人脸识别、虹膜识别、视线追踪等技术在深度神经网络技术的应用下有了较为突破的发展。而视线跟踪技术是人机互动领域的核心技术,是人机互动中不可缺少的一环。通常视线跟踪技术分为基于模型的方法和基于外观的方法。
基于模型的方法(也称3D视线跟踪法)是构建几何眼睛模型,利用入射光线在眼球角膜表面时发生的折射与反射现象,通过反射与折射的几何模型求解出眼球角膜中心与瞳孔中心,进而重构眼球的视线方向[1?3]。最近几年,国内外还构建头部姿势的模型用来辅助视线跟踪技术[3],这种方法可以排除头部姿势的干扰,大大提高视线跟踪的准确性。……