权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Research on eye contact recognition with active camera

主动摄像头眼神接触识别研究

基本信息

批准号：
16200014
负责人：
WADA Toshikazu
金额：
$ 30.78万
依托单位：
Wakayama University
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (A)
财政年份：
2004
资助国家：
日本
起止时间：
2004 至 2006
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/en/grant/KAKENHI-PROJECT-16200014/
关键词：
View direction estimation Two circle method PaLM-tree K-means tracker Distinctiveness Active stereo tracking K-d decision tree Principal component tree PaLm-tree 色ターゲット検出・追跡色弁別度能動ステレオカメラアイモデル PaLM-Tree 視線認識顔検出対象追跡能動

项目摘要

In this research, we address "eye contact recognition problem", i.e., estimating a human gaze direction from a single image. This problem can be positioned in the field of Computer Vision, which was initiated by imitating human visual functions, because this problem can also be regarded as an imitation of a visual function that people are sensitive to those who watching them.The possible approaches for realizing this peculiar function are classified into two types : 1) gaze recognition based on the iris contour shape and 2) gaze recognition based on the positional arrangement of pupils and other face organs. We investigated and compared both approaches. According to the first approach, we developed a gaze estimation algorithm based on ellipse fitting to the iris contours using an eye-model. Of course, each person has different eye ball diameter and distance between eye balls. These personal parameters can be estimated from short image sequence less than 5 seconds. After that, precise e … More ye direction estimation can be performed within NTSC video interval (33[ms]). On the other hand, we developed a gaze estimation algorithm according to the second approach. For this algorithm, we developed a learning algorithm for non-linear mapping named "PaLM-tree". By using this algorithm, we can estimate the viewing direction from the positional arrangement of face organs, to visual direction, it requires large training data and the accuracy of the vertical direction is less than the first one.Also, we investigated problems of 3) active visual tracking for capturing high quality eye image of moving objective person and 4) recognition method for eye contact recognition. For the third problem, we developed a visual tracking algorithm adaptive to color change named "k-means tracker". This algorithm is robust against the color shift and shape deformation of the object. And we extend it to "reliability based k-means tracker" for further improvement of the robustness. We also developed another very fast tracking algorithm based on the distinctiveness of target colors. By using these algorithms, high performance active stereo tracking system have been constructed. For the fourth problem, we developed an accelerated algorithm of nearest neighbor classification named "k-d decision tree". Currently, this is the world fastest nearest neighbor classifier. Less

在这项研究中，我们解决了“目光接触识别问题”，即从单个图像中估计人类的凝视方向。这个问题可以定位在计算机视觉领域，它是从模仿人类的视觉功能开始的，因为这个问题也可以看作是对人们对观看者敏感的视觉功能的模仿。实现这一特殊功能的可能方法分为两种：1)基于虹膜轮廓形状的凝视识别和2)基于瞳孔和其他面部器官的位置排列的凝视识别。我们调查并比较了这两种方法。在第一种方法的基础上，提出了一种基于椭圆拟合虹膜轮廓的注视估计算法。当然，每个人的眼球直径和眼球之间的距离是不同的。这些个人参数可以从不到5秒的短图像序列中估计出来。在NTSC视频间隔（33[ms]）内可以进行更精确的方向估计。另一方面，我们根据第二种方法开发了一种注视估计算法。针对该算法，我们开发了一种非线性映射的学习算法“PaLM-tree”。使用该算法，我们可以从人脸器官的位置排列，到视觉方向估计观看方向，需要大量的训练数据，并且垂直方向的精度低于第一种方法。此外，我们还研究了3)捕捉高质量运动目标人眼图像的主动视觉跟踪问题和4)眼神接触识别的识别方法问题。针对第三个问题，我们开发了一种自适应颜色变化的视觉跟踪算法“k-means tracker”。该算法对物体的颜色偏移和形状变形具有较强的鲁棒性。并将其扩展为“基于可靠性的k-means跟踪器”，进一步提高了鲁棒性。我们还开发了另一种基于目标颜色独特性的快速跟踪算法。利用这些算法，构建了高性能的主动立体跟踪系统。对于第四个问题，我们开发了一种加速的最近邻分类算法“k-d决策树”。目前，这是世界上最快的最近邻分类器。少