Eye Tracking: Principles and Applications | 眼动追踪实验原理与应用

📚 Eye Tracking: Principles and Applications | 眼动追踪实验原理与应用

Eye tracking is a technology that measures where a person is looking, how their eyes move, and how long they fixate on specific objects. It has evolved from a laboratory research tool into a versatile technology used in psychology, human-computer interaction, marketing, and medical diagnostics. This article explains the scientific principles behind eye tracking and explores its major applications in computer science and beyond.

眼动追踪是一种测量人的视线方向、眼球运动模式以及在特定对象上注视时长的人机交互技术。它已从实验室研究工具发展为广泛应用于心理学、人机交互、市场营销和医学诊断的通用技术。本文将从计算机科学的角度，系统阐述眼动追踪的科学原理及其主要应用场景。

---

1. What Is Eye Tracking? | 什么是眼动追踪

Eye tracking refers to the process of measuring either the point of gaze (where one is looking) or the motion of an eye relative to the head. An eye tracker is a device that records eye position and eye movement at a high sampling rate, typically between 30 Hz and 2000 Hz. The output data includes gaze coordinates, pupil dilation, blink rate, and saccade velocity — all of which can be mapped onto a stimulus such as a screen image, a webpage, or a real-world scene.

眼动追踪是指测量注视点（即眼睛看向何处）或眼球相对于头部的运动过程。眼动仪是一种以高采样率记录眼睛位置和运动的设备，典型采样率在30赫兹到2000赫兹之间。输出数据包括注视坐标、瞳孔直径、眨眼频率和扫视速度，这些数据可以映射到屏幕图像、网页或真实场景等刺激物上。

From a computer science perspective, eye tracking is fundamentally a signal acquisition and pattern recognition problem. The device captures raw video frames of the eye, extracts geometric features, applies machine-learning models to classify eye movements, and converts the results into a real-time gaze stream that applications can consume.

从计算机科学的角度看，眼动追踪本质上是一个信号采集和模式识别问题。设备采集眼部原始视频帧，提取几何特征，应用机器学习模型对眼球运动进行分类，并将结果转化为可供应用程序消费的实时注视数据流。

Eye trackers can be classified into two broad categories: head-mounted (mobile) eye trackers worn on the user’s head, and remote (screen-based) eye trackers placed on or near a display. Each type has its own calibration requirements, accuracy characteristics, and use cases, which we will examine in Section 4.

眼动仪可分为两大类：佩戴在用户头上的头戴式（移动式）眼动仪，以及放置在显示器上或附近的桌面式眼动仪。每种类型都有各自的校准要求、精度特性和使用场景，我们将在第4节详细介绍。

---

2. Human Visual System and Eye Movement Types | 人类视觉系统与眼动类型

To understand how eye tracking works, we must first understand the basic anatomy of the human eye. The retina contains two main types of photoreceptor cells: rods, which are sensitive to light intensity and work well in dim conditions, and cones, which are responsible for color vision and high-acuity central vision. Cones are densely concentrated in a small central region of the retina called the fovea, which spans only about 1 to 2 degrees of the visual field.

要理解眼动追踪的工作原理，首先需要了解人眼的基本解剖结构。视网膜包含两类主要的感光细胞：视杆细胞对光强敏感，在昏暗条件下表现良好；视锥细胞负责颜色视觉和高精度中心视觉。视锥细胞密集分布在视网膜中央一个称为中央凹的小区域内，该区域仅覆盖视野中约1到2度的范围。

Because the fovea provides the highest visual acuity, humans must continuously move their eyes to place objects of interest onto this small region. This gives rise to three fundamental types of eye movements that eye-tracking algorithms must recognise and distinguish:

由于中央凹提供最高视觉敏锐度，人类必须不断移动眼球，使感兴趣的物体落在这一小块区域上。这产生了眼动追踪算法必须识别和区分的三种基本眼动类型：

• Fixations — periods when the eye remains relatively still on a target, typically lasting 150–600 milliseconds. During a fixation, visual information is extracted and processed by the brain.

• 注视 — 眼睛相对静止地停留于某一目标的时段，通常持续150至600毫秒。注视期间，大脑提取并处理视觉信息。

• Saccades — rapid, ballistic jumps of the eye between two fixation points, reaching speeds of up to 900 degrees per second. Visual perception is suppressed during saccades.

• 扫视 — 眼睛在两个注视点之间快速跳跃的运动，速度可达每秒900度。扫视期间视觉感知被抑制。

• Smooth pursuit — slow tracking movements (typically below 30 degrees per second) that occur when the eye follows a moving object. This movement is largely involuntary and cannot be executed without a moving stimulus.

• 平滑追踪 — 当眼睛跟随移动物体时产生的缓慢追踪运动（通常低于每秒30度）。这种运动在很大程度上是不自主的，没有移动刺激就无法执行。

Other important eye events include blinks, micro-saccades (tiny involuntary movements during fixations), and nystagmus (a rhythmic oscillation of the eyes). Each of these events has a distinctive velocity and spatial signature that signal-processing algorithms use to classify the raw data stream.

其他重要的眼动事件包括眨眼、微扫视（注视期间微小的不自主运动）和眼球震颤（眼球的节律性振荡）。每个事件都具有独特的速度和空间特征，信号处理算法利用这些特征对原始数据流进行分类。

---

3. The Core Principle: Pupil–Corneal Reflection (PCCR) | 核心原理：瞳孔角膜反射法（PCCR）

Most modern eye trackers rely on a technique called pupil–corneal reflection (PCCR), which is also known as video-oculography. The method works as follows. An infrared light source illuminates the eye, and a camera captures high-resolution images of the eye at a high frame rate. The infrared light generates two key features in the image: the pupil (which appears as a dark circle in infrared images) and the corneal reflection (a bright spot, called the first Purkinje image, caused by light reflecting off the anterior surface of the cornea).

现代大多数眼动仪都依赖于一种称为瞳孔角膜反射法（PCCR）的技术，也称为视频眼动图法。其工作原理如下：红外光源照射眼睛，同时相机以高帧率捕捉眼部高分辨率图像。红外光在图像中产生两个关键特征：瞳孔（在红外图像中呈现为黑色圆形区域）和角膜反射（由光在角膜前表面反射产生的一个亮点，称为第一浦肯野像）。

The vector connecting the center of the pupil to the center of the corneal reflection changes systematically as the eye rotates. When the eye rotates upward, the pupil center shifts relative to the corneal reflection in a predictable way. By computing the 2D vector between these two features and applying a geometric calibration model, the system can determine the direction of gaze with sub-degree accuracy. The relationship is often expressed by the following geometric model:

瞳孔中心到角膜反射中心之间的向量会随着眼球转动而发生系统性变化。当眼球向上转动时，瞳孔中心相对于角膜反射点会产生可预测的位移。通过计算这两个特征之间的二维向量并应用几何校准模型，系统可以确定注视方向，精度可达到亚度级别。其几何关系通常表示为：

GazeDirection = f(θ_horizontal, θ_vertical) = M · (PupilCenter − CornealReflection) + B

where M is a scaling-and-rotation matrix determined by calibration, B is a bias offset, and θ_horizontal, θ_vertical represent the horizontal and vertical angles of the eye with respect to the camera. The advantage of PCCR is that it is relatively robust to small head movements, since both the pupil and the corneal reflection move together when the head translates.

其中M是由校准确定的缩放旋转矩阵，B是偏置偏移量，θ_horizontal和θ_vertical分别表示眼睛相对于相机的水平和垂直角度。PCCR的好处在于它对小幅头部运动较为鲁棒，因为当头部平移时，瞳孔和角膜反射会一起移动。

A key limitation of PCCR is that it requires the eye to be visible and well-illuminated. Glasses, contact lenses, eyelashes, and strongly angled gaze can degrade the quality of the corneal reflection. Additionally, PCCR measures the optical axis of the eye, not the true visual (foveal) axis. The angle between the two, known as the kappa angle, varies between individuals and must be corrected through a participant-specific calibration procedure.

PCCR的一个关键局限性是要求眼睛可见且照明良好。眼镜、隐形眼镜、睫毛和大幅度的斜视会降低角膜反射的质量。此外，PCCR测量的是眼睛的光轴，而非真实的视觉（中央凹）轴。光轴与视轴之间的夹角称为kappa角，因人而异，必须通过针对个体的校准程序进行校正。

---

4. Hardware Types of Eye Trackers | 眼动仪的硬件类型

Eye-tracking hardware has diversified significantly over the past two decades. Each hardware configuration has distinct trade-offs between accuracy, invasiveness, portability, and cost. The following table summarizes the three main categories of eye trackers:

在过去二十年中，眼动仪硬件已显著多元化。每种硬件配置在精度、侵入性、便携性和成本方面各有取舍。下表总结了三类主要的眼动仪：

Type | 类型	Typical Accuracy | 典型精度	Sampling Rate | 采样率	Key Pros & Cons | 主要优缺点
Tower-mounted | 塔式固定	0.2°–0.5°	250–2000 Hz	Highest precision; head must remain in a chinrest | 精度最高；头部需固定在颌托中
Remote / Screen-based | 桌面式/屏幕式	0.4°–1.0°	30–120 Hz	Free head movement, convenient for UX testing; less accurate at distance | 头部可自由移动，适合用户体验测试；远距离精度较低
Head-mounted / Mobile | 头戴式/移动式	0.5°–1.5°	60–200 Hz	Enables real-world experiments; heavier, more obtrusive | 可进行真实世界实验；较沉重，侵入性更强

A fundamental distinction also exists between video-based systems that require head stabilization, and inertial-based systems that combine an eye camera with accelerometers and gyroscopes. Modern mobile eye trackers, such as those used in augmented-reality glasses, fuse eye images with IMU data to reconstruct the gaze vector in 3D space. This sensor-fusion approach is a classic example of multi-modal signal processing in computer science.

另一项重要区别存在于需要头部固定的视频系统和将眼动相机与加速度计、陀螺仪相结合的惯性系统之间。现代移动眼动仪（如用于增强现实眼镜的型号）将眼部图像与惯性测量单元（IMU）数据进行融合，以在三维空间中重建注视向量。这种传感器融合方法是计算机科学中多模态信号处理的典型示例。

---

5. Calibration and Data Quality | 校准流程与数据质量

Calibration is the process of mapping raw pupil-and-corneal-reflection coordinates to actual gaze coordinates on the stimulus. The standard procedure involves a “grid calibration,” where the participant looks at a sequence of target points (typically 5, 9, or 13 points) displayed at known positions on a screen. The eye tracker records the corresponding raw feature vectors at each point and fits a mathematical function — usually a quadratic polynomial or a homography — to compress the differences between predicted and known positions.

校准是将瞳孔-角膜反射原始坐标映射到刺激上真实注视坐标的过程。标准流程采用“网格校准”，即让参与者依次注视屏幕上一系列已知位置的目标点（通常为5、9或13个点）。眼动仪在每个点记录对应的原始特征向量，并拟合一个数学函数——通常是二次多项式或单应矩阵——以最小化预测位置与已知位置之间的差异。

During calibration, two metrics are of paramount importance: accuracy and precision. Accuracy refers to the angular distance between the true gaze point and the measured gaze point, averaged over time. Precision refers to the spatial variability of repeated measurements when the gaze is fixed on a single point. A good eye tracker should achieve accuracy below 0.5° and precision below 0.1° root-mean-square (RMS).

校准过程中，有两个指标至关重要：准确度和精密度。准确度是指真实注视点与被测量注视点之间的角距离，按时间平均计算；精密度是指当注视固定在单个点上时，重复测量的空间变异性。好的眼动仪应实现低于0.5°的准确度和低于0.1°均方根（RMS）的精密度。

Several factors can degrade data quality even after a successful calibration. Pupil diameter changes due to lighting or cognitive load can shift the pupil center estimate. The kappa angle changes as the pupil dilates, causing systematic offset. Dry eyes or partial eyelid closure reduce image contrast. Therefore, most eye-tracking software includes a validation phase — a second grid test performed immediately after calibration — to quantify the residual error and decide whether recalibration is necessary.

即使在成功校准之后，多个因素仍可能降低数据质量。因光照或认知负荷导致的瞳孔直径变化会移动瞳孔中心估计值；瞳孔扩张时kappa角发生改变，引起系统性偏移；干眼或部分眼睑闭合会降低图像对比度。因此，大多数眼动追踪软件包含验证阶段——校准后立即进行的第二遍网格测试——用来量化残余误差并判断是否需要重新校准。

---

6. Eye-Movement Classification and Metrics | 眼动分类算法与关键指标

Once raw gaze coordinates are acquired, they must be segmented into meaningful eye events. The most widely used algorithms include the velocity-threshold identification (I-VT) method, the dispersion-threshold identification (I-DT) method, and the hidden Markov model (HMM) based classifiers. The I-VT method computes the instantaneous angular velocity between consecutive gaze samples; if the velocity exceeds a threshold (e.g., 30 degrees per second), the sample is labeled as part of a saccade; otherwise, it is a fixation candidate. Velocity thresholds must be carefully tuned because fixation noise in low-end trackers can exceed 30°/s.

获得原始注视坐标后，必须将其分割为有意义的眼动事件。最广为使用的算法包括速度阈值识别法（I-VT）、离散度阈值识别法（I-DT）和基于隐马尔可夫模型（HMM）的分类器。I-VT方法计算连续注视样本之间的瞬时角速度，如果速度超过阈值（例如每秒30度），则该样本标记为扫视的一部分；否则标记为注视候选。速度阈值需要仔细调整，因为低端眼动仪的注视噪声可能超过30度/秒。

After classification, a rich set of metrics can be derived for each region of interest (AOI). Common metrics include:

分类完成后，可以为每个感兴趣区域（AOI）导出丰富的度量指标。常用指标包括：

• Fixation count — the total number of fixations within an AOI. A higher count generally indicates greater cognitive processing or attention.

• 注视次数 — 在一个AOI内的注视总数。较高的注视次数通常意味着更多的认知加工或更强的注意。

• Total dwell time — the sum of all fixation durations in an AOI, reflecting overall engagement.

• 总停留时间 — 一个AOI内所有注视时长的总和，反映总体投入程度。

• Time to first fixation — the time elapsed before a participant first looks at a specific region, used to measure attention attraction.

• 首次注视时间 — 参与者首次看向某个特定区域之前经过的时间，用于衡量注意吸引程度。

• Saccade amplitude — the angular distance of a saccade, which is related to the information content of the visual scene.

• 扫视幅度 — 扫视的角距离，与视觉场景的信息含量相关。

• Pupil dilation — a physiological correlate of cognitive load and emotional arousal.

• 瞳孔扩张 — 认知负荷和情绪唤醒的生理相关指标。

These metrics are often visualized as heatmaps (gaze distribution over a two-dimensional stimulus), scanpaths (a sequence of fixations connected by saccades), and AOI transition matrices (which quantify how frequently participants switch attention between regions).

这些指标常通过热力图（注视在二维刺激上的分布）、扫描路径（由扫视连接的一系列注视）以及AOI转移矩阵（量化参与者在不同区域之间切换注意力的频率）来可视化呈现。

---

7. AI and Machine Learning in Eye-Tracking Data Analysis | 人工智能与机器学习在眼动数据分析中的应用

Modern eye-tracking research increasingly relies on machine learning for both low-level feature detection and high-level intent prediction. At the low level, convolutional neural networks (CNNs) are used to estimate pupil centers and corneal reflections directly from raw camera images, achieving better robustness to occlusion and lighting variation than classical image-processing pipelines. At the high level, recurrent neural networks (RNNs) and transformers are applied to gaze sequences to detect user states such as fatigue, confusion, or task disengagement.

现代眼动追踪研究日益依赖机器学习，既用于底层特征检测，也用于高层意图预测。在底层，卷积神经网络（CNN）被用于直接从原始相机图像估算瞳孔中心和角膜反射点，相比经典图像处理流水线，对遮挡和光照变化的鲁棒性更强。在高层，循环神经网络（RNN）和Transformer被应用于注视序列，以检测疲劳、困惑或任务脱离等用户状态。

One prominent application of AI to eye tracking is the estimation of visual attention in free-viewing tasks. The task can be formulated as saliency prediction: given an image, predict a probability map that indicates where a human would look. This is inherently a supervised deep-learning problem, trained on large-scale datasets of human gaze data such as SALICON and MIT300. Modern saliency models achieve human-level performance on several benchmarks, and the learned features have been found to correlate with object boundaries, text regions, and faces.

人工智能在眼动追踪中的一个重要应用是在自由观看任务中估计视觉注意。该任务可表述为显著性预测：给定一张图像，预测一个人会看向哪里的概率图。这本质上是一个有监督的深度学习问题，在SALICON和MIT300等大规模人类注视数据集上训练。现代显著性模型在多个基准测试上达到人类水平，且学到的特征已被发现与物体边界、文本区域和人脸相关联。

Another emerging research direction is gaze-based user identification and biometric authentication. The temporal dynamics of eye movements — including saccade velocity profiles and micro-saccade patterns — are highly individualistic. Researchers have demonstrated that CNN-RNN hybrid models can identify individuals from their eye-movement traces alone with accuracy above 90%, which raises both exciting security applications and alarming privacy concerns.

另一个新兴研究方向是基于注视的身份识别和生物特征认证。眼动的时间动力学特征——包括扫视速度轮廓和微扫视模式——具有高度的个体独特性。研究人员已证明，CNN-RNN混合模型仅凭眼动轨迹就能以90%以上的准确率识别个体，这既带来了令人兴奋的安全应用，也引发了令人担忧的隐私问题。

---

8. Application: Usability Testing and Human–Computer Interaction | 应用一：可用性测试与人机交互

One of the most commercially successful applications of eye tracking is in usability testing of websites, software, and mobile applications. Eye-tracking data reveals exactly where users look on a user interface, which elements attract attention, and which areas are ignored. This objective behavioral evidence complements subjective self-report data derived from interviews and questionnaires, exposing the gap between what users say they did and what they actually did.

眼动追踪在商业上最成功的应用之一是网站、软件和移动应用的可用性测试。眼动数据准确地揭示了用户在用户界面上看向何处、哪些元素吸引了注意、哪些区域被忽略。这种客观的行为证据补充了通过访谈和问卷获得的主观自我报告数据，暴露出用户所称行为与实际行为之间的差距。

Consider the classic “banner blindness” phenomenon in web design. Heatmap analyses have shown that users systematically ignore banner-shaped areas at the top of a webpage, even when the banners contain important navigation content. This finding directly informs layout design: interface designers now place critical content in the upper-left region of the viewport, following the F-pattern reading behavior that eye-tracking studies have repeatedly confirmed for text-heavy pages.

以网页设计中的经典“横幅盲视”现象为例。热力图分析表明，用户会系统性地忽略网页顶部横幅形状的区域，即使横幅中包含重要的导航内容。这一发现直接为布局设计提供了依据：界面设计师现在会将关键内容放置在视口左上角区域，遵循眼动研究在文本密集型页面中反复证实的F型阅读模式。

In mobile UX research, a well-known eye-tracking finding is the “thumb zone” concept: users predominantly interact with the lower half of a smartphone screen with their thumbs, but their eyes first glance at the upper portion of the screen. This decoupling between visual attention and touch interaction can lead to mis-swipes and interface errors. Eye-tracking-informed design thus encourages placing action buttons in the visual-attention zone while reserving the thumb zone for swiping gestures and scrollable content.

在移动端用户体验研究中，一个著名的眼动发现是“拇指区域”概念：用户主要用拇指操作手机屏幕的下半部分，但目光首先扫向屏幕的上半部分。视觉注意与触摸交互之间的这种脱节可能导致误滑和界面错误。因此，基于眼动信息的设计鼓励将操作按钮放在视觉注意区域，而将拇指区域保留给滑动手势和可滚动内容。

From a computer-science examination perspective, it is important to understand how eye-tracking data integrates into an iterative design cycle. The typical workflow is: (1) record gaze data during task-oriented user testing, (2) identify problematic areas via fixation-density visualization, (3) redesign the interface, and (4) re-test using the same protocol. This mirrors the fundamental concept of iterative user-centered design, which is a core topic in A-level computer science units on human-computer interaction.

从计算机考试的角度而言，重要的是理解眼动数据如何融入迭代设计周期。典型流程是：(1) 在任务导向的用户测试期间记录注视数据；(2) 通过注视密度可视化识别问题区域；(3) 重新设计界面；(4) 使用相同协议重新测试。这反映了以用户为中心的迭代设计的基本概念，也是A-level计算机科学中人机交互单元的核心考点。

---

9. Application: Cognitive Psychology and Reading Research | 应用二：认知心理学与阅读研究

Eye tracking has been instrumental in understanding reading processes. In reading experiments, researchers record where the eyes land on lines of text and how long they dwell on each word. Established models such as the E-Z Reader model and the SWIFT model use eye-tracking data to test predictions about lexical access, syntactic parsing, and sentence comprehension.

眼动追踪在理解阅读过程方面发挥了重要作用。在阅读实验中，研究人员记录目光落在文本行的位置以及在每个单词上停留的时长。既有的E-Z Reader模型和SWIFT模型使用眼动数据来检验有关词汇提取、句法解析和句子理解的理论预测。

A fundamental finding from reading studies is the relationship between word frequency and fixation duration. Fixations on high-frequency words (such as “the” and “and”) are consistently shorter than fixations on low-frequency words (such as “meticulous” or “algorithm”). Additionally, the length of a saccade during reading is typically 7 to 9 characters, and the landing position tends to be slightly left of the center of the word — a phenomenon described by the optimal-viewing-position hypothesis.

阅读研究的一个基本发现是高频率词与注视时长之间的关系。对高频词（如“的”和“是”）的注视时长始终短于对低频词（如“缜密”或“算法”）的注视时长。此外，阅读期间扫视的长度通常为7到9个字符，落点位置往往在单词中心的稍左侧——这一现象由最佳注视位置假说描述。

Eye-tracking data also reveals the phenomenon of “regression” — a backward saccade that re-reads previously processed text. Regression frequency increases when readers encounter syntactically ambiguous sentences or semantic inconsistencies. By measuring regression patterns, cognitive scientists can infer the point at which comprehension fails, providing a powerful window into the online processes of language comprehension that offline measures such as comprehension questions cannot capture.

眼动数据还揭示了“回视”现象——即重新阅读已加工文本的向后的扫视。当读者遇到句法歧义或语义不一致时，回视频率会增加。通过测量回视模式，认知科学家可以推断理解失败的节点，为语言理解的在线过程提供了一个离线测量工具（如理解性问题）无法捕捉的强有力窗口。

In clinical psychology, eye tracking is used to diagnose and monitor disorders such as autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD). Children with ASD show reduced visual attention to social cues, spending less time looking at the eye region of human faces. ADHD patients exhibit more frequent saccades and shorter fixation durations during attention-demanding tasks. These quantitative biomarkers support earlier diagnosis and more objective treatment evaluation.

在临床心理学中，眼动追踪被用于诊断和监测自闭症谱系障碍（ASD）和注意缺陷/多动障碍（ADHD）等疾病。ASD儿童对社交线索的视觉注意减少，花在看人脸眼睛区域的时间更短。ADHD患者在需要集中注意的任务中表现出更频繁的扫视和更短的注视时长。这些定量生物标志物支持更早的诊断和更客观的治疗评估。

---

10. Application: Advertising, Marketing, and Scene Perception | 应用三：广告、营销与场景感知

In advertising research, eye tracking answers a fundamental question: how do consumers allocate visual attention to products, packaging, and promotional displays? The concept of “attention capture” is quantified by time-to-first-fixation and the total number of fixations on a target product. Research consistently shows that the center of a display shelf receives more attention than the periphery, and that products at eye level (roughly 140–160 cm from the floor) are viewed earlier and more often.

在广告研究中，眼动追踪回答一个基本问题：消费者如何将视觉注意分配给产品、包装和促销陈列？“注意捕获”的概念通过首次注视时间和目标产品上的注视总数进行量化。研究一致表明，货架中央区域比边缘区域获得更多注意，视线水平高度（距离地面约140至160厘米）的产品能被更早且更频繁地看到。

Eye-tracking evaluations of online advertisements reveal the “banner blindness” effect more broadly: users voluntarily suppress attention to display ads that resemble the typical rectangular banner format. Transitioning to video ads, gaze heatmaps show that attention is strongly driven by faces (especially the eyes) and by text overlays, and that attention decreases monotonically over the duration of a 30-second video advertisement.

对在线广告的眼动评估更广泛地揭示了“横幅盲视”效应：用户会主动抑制对类似典型方形横幅格式的展示型广告的注意。在视频广告方面，注视热力图的显示表明注意力强烈地被面孔（尤其是眼睛）和文字叠加层所驱动，并且随着30秒视频广告的播放，注意力呈单调递减趋势。

Within scene perception research, eye tracking has produced the well-established finding that human observers exhibit a robust central fixation bias — even when images contain highly salient objects in the periphery, people disproportionately fixate near the center of a screen. Gist extraction, the ability to recognize the overall meaning of a scene within 100 milliseconds, is accomplished by this central vision bias supplemented by a few targeted peripheral saccades.

在场景感知研究中，眼动追踪产生了一个已被充分论证的发现：人类观察者表现出稳健的中心注视偏差——即使图像在外围包含极具显著性的物体，人们仍会不成比例地注视屏幕中央附近。在100毫秒内识别场景整体意义的能力，即“主旨提取”，是通过这种中心视觉偏差并由少数有针对性的周边扫视来辅助完成的。

In the field of visual attention modeling, researchers use eye-tracking data to train saliency models and computational attention systems. The benchmark datasets for saliency prediction are built by collecting gaze data from hundreds of observers viewing natural images. These models, in turn, are used in computer vision applications such as image compression (coding only the attended regions at high bitrate), video summarization (selecting frames that attract attention), and visual search systems.

在视觉注意建模领域，研究人员使用眼动数据来训练显著性模型和计算性注意系统。显著性预测的基准数据集是通过收集数百名观察者观看自然图像时的注视数据构建的。这些模型反过来又应用于计算机视觉任务，如图像压缩（仅以高比特率编码被注意区域）、视频摘要（选择吸引注意的帧）和视觉检索系统。

---

11. Application: Driving Safety, Assistive Technology, and VR/AR | 应用四：驾驶安全、辅助技术与虚拟/增强现实

Eye tracking is a key enabling technology for driver-monitoring systems in modern vehicles. Cameras mounted on the steering column track the driver’s gaze direction, eyelid closure rate, and blink frequency to assess fatigue and distraction. If the system detects that gaze has been

Published by TutorHao | Computer Science Revision Series | aleveler.com

更多咨询请联系16621398022（同微信）