三维注视点的建模和隐式标定方法研究

Point of gaze is the position where human visual attention locates, it is an important channel for obtaining cognitive information or implementing natural human computer interaction. Currently, video-oculography methods are constructed based on the population with normal vision,thus it does not have much suitability. Existing methods have some disadvantages, they fail to meet all the requirements for convenient operation, large-scale head movements and high spatial accuracy. This project firstly extends the Gullstrand eye model by establishing a general model to handle different cornea curvatures, distortions and refractions. The general model supports 3D to 3D point of gaze localization. Next, a 3D saliency field is analyzed, the Bayesian learning method is used to implicitly calibrate the model's parameters. The spatial saliency features are analyzed to progressively correct the errors and improve accuracy. Thirdly, motion equations are built according to the synchronization features of the eye movements and the Listing law, methods in data association and trajectory filtering are applied to eliminate the image noise interferences in target localization. This project aims to expand the scope of target population and applications, to achieve performance optimization in calibration, tracking and anti-interference ability. It is expected to supply new ideas and solutions in modeling, calibration and image feature tracking in the research fields, and to fine down the relevant theories.

注视点即人类视觉注意的空间位置所在，它是获取心理认知信息、实现自然的人机交互的重要渠道。目前，注视点图像视频法基于视力正常人群建立，不具普适性；且现有的方法存在较多弊端，未能同时满足标定操作便捷、支持头部运动、空间定位精度高等要求。本项目首先扩展Gullstrand眼球模型，建立一个通用的适合不同角膜曲率、不受畸变和折射影响，且支持3D到3D定位的注视点模型；开展三维视觉显著场研究，应用贝叶斯学习方法对注视点模型进行隐式参数标定，并通过空间视觉敏感特征分析渐进修正以提高精度；根据双眼运动同步特征及Listing定律构建眼球图像运动方程，依靠数据关联和轨迹滤波等技术消除图像噪声对目标定位的干扰。本项目旨在拓展注视点研究的人群适用范围和应用领域，实现标定操作、跟踪精度和视觉抗干扰等相关性能的优化。项目有望为注视点建模、标定和图像特征跟踪提供新的研究思路和解决方案，并完善相关理论。

通过对人眼部图像的处理、分析和跟踪，注视点技术可以获取人眼的注视位置，为进一步探索人机交互和心理认知提供重要信息渠道。本研究在搭建头戴式和远程式注视点图像采集实验装置的基础上，完成了如下几个方面的研究，首先设计了一款开源的、跨多个操作系统平台的可视化图像调试器，该调试器可在图像算法调试过程中可视化地查看内存变量，因此提高了算法开发和调试效率；其次项目提出了一种结合凸区域投票和模型约束的瞳孔定位算法，通过对图像中候选瞳孔边缘的梯度进行投票，有效抑制图像中的各种干扰，提高了瞳孔中心的定位精度；在人眼参数估计方面，项目提出一种双眼瞳孔间距参数隐式估计方法，利用双眼同步运动规律简化了双眼运动模型，运用Kalman滤波器渐进估计运动参数，在图像测量存在噪声时，滤波器输出平稳；利用AAM模型和POSIT算法，估计人脸的三维姿态，进而修正注视点的估计位置，减小因人脸姿态旋转导致的注视点定位误差。项目为注视点建模、人眼的相关参数标定和图像特征跟踪提供了新的算法和研究思路，为注视点技术应用和推广提供了一定的理论基础和实践经验。