---
id: wiki-2026-0508-eye-tracking
title: Eye Tracking
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Gaze Tracking, Eye Gaze, Foveated Rendering Input]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [eye-tracking, xr, vr, ar, accessibility, ux]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: typescript
  framework: webxr
---

# Eye Tracking

## 매 한 줄
> **"매 eye tracking 은 user gaze direction 의 measure — pupil/cornea reflection 또는 ML model 의 사용"**. 2026 의 mainstream: Apple Vision Pro, Meta Quest 3S Pro, PSVR2. 매 foveated rendering, gaze-based UI selection, accessibility (ALS), UX research 의 enables.

## 매 핵심

### 매 Sensing Methods
- **PCCR (Pupil Center Corneal Reflection)**: IR LED illumination + IR camera. Sub-degree accuracy, dedicated HW (Tobii, Vision Pro).
- **Webcam-based ML**: WebGazer.js, MediaPipe Iris. ~3-5° accuracy, no special HW.
- **Electrooculography (EOG)**: skin electrodes; medical/lab use.
- **Scleral search coil**: highest precision, invasive (research only).

### 매 Output Data
- **Gaze point**: 2D screen coord or 3D ray.
- **Pupil diameter**: cognitive load proxy.
- **Fixations**: stable gaze (>100ms) — what user actually reads.
- **Saccades**: rapid eye movement between fixations.
- **Smooth pursuit**: tracking moving target.

### 매 응용
1. Foveated rendering (high-res only at gaze point — VR perf 4-10x).
2. Gaze + pinch UI selection (Apple Vision Pro paradigm).
3. Accessibility: gaze-typing for ALS/locked-in patients.
4. UX research: heatmaps, attention analysis.
5. Driver monitoring (drowsiness detection).

## 💻 패턴

### WebGazer.js (browser, no HW)
```js
import webgazer from 'webgazer';

await webgazer.setRegression('ridge')
  .setGazeListener((data, timestamp) => {
    if (!data) return;
    console.log('gaze:', data.x, data.y);  // screen px
  })
  .begin();

// User must calibrate by clicking around
```

### MediaPipe Iris (TF.js)
```js
import * as faceLandmarksDetection from '@tensorflow-models/face-landmarks-detection';

const detector = await faceLandmarksDetection.createDetector(
  faceLandmarksDetection.SupportedModels.MediaPipeFaceMesh,
  { runtime: 'tfjs', refineLandmarks: true }  // refine = iris landmarks
);

const faces = await detector.estimateFaces(video);
const irisLeft = faces[0].keypoints.filter(k => k.name?.includes('leftIris'));
const irisRight = faces[0].keypoints.filter(k => k.name?.includes('rightIris'));
```

### WebXR Eye Tracking (Vision Pro Safari, Quest)
```ts
const session = await navigator.xr.requestSession('immersive-vr', {
  requiredFeatures: ['eye-tracking'],
});

session.requestAnimationFrame(function onFrame(t, frame) {
  const eyeSpace = session.requestReferenceSpace('viewer');
  const gaze = frame.getViewerPose(eyeSpace);
  // gaze.transform.position / orientation = gaze ray
  session.requestAnimationFrame(onFrame);
});
```

### Fixation Detection (I-DT algorithm)
```ts
// Identification by Dispersion Threshold
function detectFixations(gazePoints: {x:number, y:number, t:number}[],
                        dispersionPx = 50, minDurationMs = 100) {
  const fixations = [];
  let window: typeof gazePoints = [];
  for (const p of gazePoints) {
    window.push(p);
    const xs = window.map(g => g.x), ys = window.map(g => g.y);
    const disp = (Math.max(...xs) - Math.min(...xs))
               + (Math.max(...ys) - Math.min(...ys));
    if (disp > dispersionPx) {
      const dur = window[window.length-2].t - window[0].t;
      if (dur >= minDurationMs) fixations.push(window.slice(0, -1));
      window = [p];
    }
  }
  return fixations;
}
```

### Foveated Rendering Hint (WebGPU)
```ts
// VRS / foveated rendering via GPU extension (browser-dependent)
const passEncoder = encoder.beginRenderPass({
  colorAttachments: [...],
  // implementation-specific: shading rate texture pointing at gaze
});
```

### Heatmap from Fixations
```ts
function buildHeatmap(fixations, w, h, sigma = 30) {
  const map = new Float32Array(w * h);
  for (const f of fixations) {
    for (let y = 0; y < h; y++) for (let x = 0; x < w; x++) {
      const d2 = (x-f.x)**2 + (y-f.y)**2;
      map[y*w+x] += Math.exp(-d2 / (2*sigma*sigma)) * f.duration;
    }
  }
  return map;
}
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| VR/AR headset (2026) | Native eye-tracking API (visionOS, OpenXR ext) |
| Browser, accuracy critical | Tobii / dedicated HW + SDK |
| Browser, no HW | WebGazer.js or MediaPipe Iris (lower accuracy) |
| Accessibility input | Tobii Eye Tracker 5 + Windows Eye Control |
| UX research | Tobii Pro / GazePoint + heatmap tool |

**기본값**: 매 platform 의 native eye-tracking API. Browser fallback: MediaPipe Iris.

## 🔗 Graph
- 부모: [[Human Computer Interaction]] · [[XR]]
- 응용: [[Accessibility (A11y)|Accessibility]]
- Adjacent: [[WebXR]] · [[MediaPipe]]

## 🤖 LLM 활용
**언제**: gaze-based UI 의 design, foveated rendering question, accessibility input.
**언제 X**: physiological eye-disease diagnosis (clinical 의 ophthalmologist).

## ❌ 안티패턴
- **No calibration**: webcam-based 의 every-user calibration 의 require.
- **Click-on-gaze without dwell time**: Midas touch problem — every glance triggers.
- **Storing raw gaze without consent**: PII / biometric (GDPR Art. 9).
- **Webcam ML for safety-critical**: insufficient accuracy (driver assist 의 dedicated HW).
- **Ignoring head movement**: gaze ≠ pupil position; head pose 의 compensate.

## 🧪 검증 / 중복
- Verified (Apple visionOS docs, OpenXR EXT_eye_gaze_interaction, Tobii Developer Zone).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — eye tracking full content |