基于Kinect传感器的两眼模型精确凝视估计方法

4 浏览量更新于2024-08-27 收藏 1.06MB PDF 举报

"这篇研究论文探讨了如何使用Kinect传感器基于两眼模型进行精确且有效的凝视估计，即使在头部自由移动的情况下也能实现。作者提出了一种新的方法，包括使用两眼模型提高精度，改进的梯度中心定位3D空间中的虹膜中心，以及仅需一个校准点的新个人校准方法。该方法通过将视觉轴近似为从虹膜中心到凝视点的直线来确定眼球中心和Kappa角，从而计算出最终的凝视点。实验在11个受试者上验证了该方法，平均估计精度约为1.99度，优于许多现有的先进方法。" 本文的核心知识点如下： 1. **凝视估计**：这是一种技术，旨在确定一个人的视线方向，即他们正在注视的位置。这对于人机交互、心理分析、医疗诊断等领域有重要应用。 2. **Kinect传感器**：由微软开发的设备，主要用于捕捉人体运动和深度信息，常用于游戏和交互式应用。在本文中，它被用于获取用户的面部和眼睛数据。 3. **两眼模型**：该模型考虑了两个眼睛的位置和相互关系，以更准确地估计凝视点。这有助于纠正单眼模型可能引入的误差，尤其是在头部运动时。 4. **虹膜中心定位**：确定虹膜中心是凝视估计的关键步骤。文中提出的改进的卷积基梯度方法能够在3D空间中精确定位虹膜中心，提高了估计的准确性。 5. **个人校准**：传统的校准方法可能需要多个校准点，而本文提出的新方法只需要一个校准点，简化了校准过程，降低了用户参与的复杂性。 6. **Kappa角**：这是描述眼球旋转中心与瞳孔中心之间角度的术语，对理解眼球运动和正确计算视觉轴至关重要。 7. **视觉轴**：从虹膜中心到凝视点的直线代表了视线方向，是计算凝视点的基础。 8. **实验评估**：通过在11个受试者上进行实验，论文展示了所提方法的高精度和优越性，平均误差小于2度，证明了其在实际应用中的潜力。这些知识点共同构成了基于Kinect传感器的两眼模型凝视估计方法，该方法不仅提高了估计精度，还降低了校准的复杂性，对于理解和改进人机交互技术具有重要意义。

Two-Eye Model-Based Gaze Estimation from A Kinect Sensor

Xiaolong Zhou

, Haibin Cai

, Youfu Li

, and Honghai Liu

2,4

Abstract— In this paper, we present an effective and accurate

gaze estimation method based on two-eye model of a subject

with the tolerance of free head movement from a Kinect sensor.

To accurately and efﬁciently determine the point of gaze, i) we

employ two-eye model to improve the estimation accuracy; ii)

we propose an improved convolution-based means of gradients

method to localize the iris center in 3D space; iii) we present a

new personal calibration method that only needs one calibration

point. The method approximates the visual axis as a line from

the iris center to the gaze point to determine the eyeball

centers and the Kappa angles. The ﬁnal point of gaze can be

calculated by using the calibrated personal eye parameters. We

experimentally evaluate the proposed gaze estimation method

on eleven subjects. Experimental results demonstrate that our

gaze estimation method has an average estimation accuracy

around 1.99

◦

, which outperforms many leading methods in the

state-of-the-art.

I. INTRODUCTION

Gaze estimation is to determine the point of regard of

a person, which plays an important role in understanding

human attention, feelings, and desires. It has been widely

explored in many intelligent systems for virtual reality,

human-computer interaction, human-robot interaction, hu-

man behavior analysis and so on. Some gaze estimation

researchers concentrated on using the pupil center corneal re-

ﬂection technique. This kind of technique normally requires

one or multiple infrared lights and high-quality cameras,

which limits the system’s potential for broader applications.

Moreover, most of the existing gaze estimation systems have

low tolerance toward head movement, which hinders them

from being widely used.

Recently, Kinect-based 3D gaze estimation [1], [2], [3],

[4], [5], [6], [7], [8] has attracted increasing attention since

it is low-cost, non-intrusive, simple-setup and it allows free

head movements. Generally, Kinect-based gaze estimation

methods can be roughly classiﬁed into non-eye model-based

methods and eye model-based methods. Non-eye model-

based methods are typically appearance-based or regression-

based. For example, Mora and Odobez [1] estimated 3D gaze

This work was supported in part by the National Natural Science Foun-

dation of China (61403342, 61673329, U1509207, 61325019, 51575338)

and Hubei Key Laboratory of Intelligent Vision Based Monitoring for

Hydroelectric Engineering (2014KLA09).

Xiaolong Zhou is with the College of Computer Science and

Technology, Zhejiang University of Technology, Hangzhou, China.

zxl@zjut.edu.cn

Haibin Cai and Honghai Liu are with the School of Computing,

University of Portsmouth, Portsmouth, UK.

Youfu Li is with the Department of Mechanical and Biomedical Engi-

neering, City University of Hong Kong, Hong Kong, China.

Honghai Liu is with the State Key Laboratory of Mechanical Systems

and Vibration, School of Mechanical Engineering, Shanghai Jiao Tong

University, Shanghai, China.

from multimodal Kinect data and achieved an estimation ac-

curacy with average error around 7.6

◦

−12.6

◦

. Furthermore,

they proposed a geometric generative 3D gaze estimation

method [2] based on an appearance generative process that

modeled head-pose rectiﬁed eye images recovered by using

of RGB-D cameras, which improved the estimation accuracy

to 6.3

◦

. Cazzato et al. [3] incorporated the 3D head pose to

estimate the ﬁnal gaze direction according to the geomet-

ric relations among the sensor, observer and target. They

reported the estimation errors for unaware users with 6.9

◦

while for informed users with 3.6

◦

. The main beneﬁt of non-

eye model-based methods are speciﬁc personal calibration

free. However, the estimation accuracy of this kind of method

is low (generally above 6

◦

Different from the non-eye model-based methods that

estimate the gaze using appearance or regression technique,

3D eye model-based methods directly determine the gaze

using the geometric relationship among human eyes, sensors

and gazing points. For example, J. Li and S. Li [4] proposed

an eye-model-based 3D gaze estimation method from a

Kinect sensor. They built a head model based on the Kinect

sensor and calibrated the eyeball center by gazing at a target

in 3D space. The gaze direction was estimated after the

calibration and the reported average error of estimation was

around 6

◦

. Recently, they estimated the gaze from color

image based on an eye model with known head pose [5].

They ﬁrst determined the 3D eyeball center in calibration

manner by gazing at the center of the color image camera,

and then estimated the 3D iris center using the information

of its contour and projection. They reported the average

estimation errors for seven subjects with 5.9

◦

vertically and

4.4

◦

horizontally. Sun et al. [6] estimated the gaze direction

based on a 3D geometric eye model by considering the

head movement and deviation of the visual axis from the

optical axis. They reported a high estimation accuracy of

1.4

◦

-2.7

◦

. However, the proposed method involved many

calibration procedures like screen-camera calibration and

personal calibration with multiple calibration points.

Although eye model-based gaze estimation methods can

achieve a higher accuracy (below 6

◦

), this kind of method

normally require speciﬁc personal calibration, which in-

volves human interactions. Moreover, the estimation ac-

curacy greatly relies on the number of calibration points.

Generally, more calibration points will lead to higher esti-

mation accuracy while at the same time require more human

interactions.

Besides the personal calibration, the 3D location of hu-

man’s iris is another key technique that affects the ﬁnal gaze

estimation accuracy. Currently, a large number of iris center

2017 IEEE International Conference on Robotics and Automation (ICRA)

Singapore, May 29 - June 3, 2017

下载后可阅读完整内容，剩余7页未读，立即下载

weixin_38670297

粉丝: 7
资源: 927

基于Kinect传感器的两眼模型精确凝视估计方法

基于Kinect传感器动态手势识别

Kinect传感器的彩色和深度相机标定_郭连朋1

基于Kinect传感器的摔倒检测研究

基于Kinect传感器的康复训练系统

基于kinect传感器进行手势控制X80机器

基于Kinect传感器的跌倒行为的检测与分析

基于双Kinect传感器定标和配准的研究1

基于kinect传感器的全方位运输平台控制系统研究

基于Kinect传感器的在线连续人体动作识别算法。

基于Kinect传感器的机器人室内环境检测方法.pdf

最新资源