███████╗██╗   ██╗███████╗
  ██╔════╝╚██╗ ██╔╝██╔════╝
█████╗   ╚████╔╝ █████╗
██╔══╝    ╚██╔╝  ██╔══╝
  ███████╗   ██║   ███████╗
  ╚══════╝   ╚═╝   ╚══════╝
  camera-only gaze estimation — ~98% accuracy

---

No eye-tracking hardware. No depth sensor. No IR emitter.
Just a standard webcam, a custom vision pipeline, and geometry — and it works.

A fully software-driven, real-time gaze estimation system built from scratch in MATLAB. Given only a regular photograph of a person's face, the system detects the iris, localizes it with sub-pixel precision, and maps its position to exact screen coordinates — no calibration rig, no specialized equipment required.

B.Sc. Thesis project — Electrical Engineering.

---

🔍 What This System Does

Full pipeline, end to end:

1. Face & eye region detection — Viola-Jones cascade classifier isolates the face, then a second cascade finds both eye regions with geometric validation (filters false positives by checking relative eye positions)
2. Iris localization — Custom Laplacian-sharpened, Canny-edge-detected pipeline feeds into imfindcircles (Hough circle transform) with adaptive polarity selection
3. Pupil detection — Separate concentric-circle search with radius constraints calibrated to the iris boundary
4. Gaze coordinate prediction — Calibration-frame geometric model maps iris displacement across 4 reference gaze directions to absolute screen pixel coordinates
5. Multi-hypothesis averaging — 17 gaze estimates computed across all combinations of calibration reference points; their mean suppresses noise

---

🧠 Technical Highlights

Capability	Implementation
Face detection	vision.CascadeObjectDetector (Viola-Jones)
Eye region validation	Geometric constraint: Δy < 68px, Δx > 40px between detected regions
Edge enhancement	Custom Laplacian kernel [1,1,1; 1,-8,1; 1,1,1] sharpening + Canny edge detection
Iris detection	Hough circle transform (imfindcircles) on Laplacian-enhanced Canny edges
Polarity selection	Confidence-metric comparison between dark and bright iris candidates
Pupil detection	Constrained Hough search at ≈1/3 iris radius
Gaze mapping	screen_x = 100 + (screen_width × Δx_eye) / eye_span_x
Noise reduction	17-hypothesis ensemble averaging across all calibration reference combinations
Dataset support	Custom dataset + CASIA iris database (switchable via parameter tuning)

---

🏗️ System Architecture

Input: Face Photo(s)
        │
        ▼
┌─────────────────────┐
│  Sep.m              │  ← Face crop → Eye ROI extraction (Cascade + geometry)
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  irisfinder1.m      │  ← Full iris detection on first calibration frame
│  irisfinder2.m      │  ← Locked-ROI iris detection for subsequent frames
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  pupiliris.m        │  ← Concurrent pupil + iris detection (CASIA mode)
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  Predictor.m        │  ← 4-point calibration → gaze coordinate mapping
│                     │    17-hypothesis ensemble → mean (x, y) on screen
└────────┬────────────┘
         │
         ▼
Output: Screen gaze point marked on display board

---

🚀 Running the System

Requirements: MATLAB with Computer Vision Toolbox and Image Processing Toolbox

% Load 4 calibration images (gaze at screen corners) + 1 target image
ImUR = imread('UpRight.png');
ImUL = imread('UpLeft.png');
ImDR = imread('DownRight.png');
ImDL = imread('DownLeft.png');
Desired = imread('Target.png');

% Run gaze prediction
[gaze_x, gaze_y] = Predictor(ImUR, ImUL, ImDL, ImDR, Desired);

Dataset toggle — Switch between custom and CASIA parameters in irisfinder1.m:

Min_Radii = 10;  Sensitivity = 0.96;  % Custom dataset
% Min_Radii = 90; Sensitivity = 0.98; % CASIA dataset

---

💡 Key Design Decisions

Why Laplacian pre-sharpening? Raw images have soft iris boundaries. The custom Laplacian kernel amplifies the circular iris edge before Canny detection, dramatically improving imfindcircles reliability on low-contrast eye photos.

Why 17 hypotheses? Single-point calibration is brittle under lighting variation or micro-movement. Computing gaze estimates across all combinations of the 4 reference corner positions and averaging makes the system significantly more robust to per-frame noise.

Why locked ROI in irisfinder2? Once the eye region is located from the first calibration frame, all subsequent frames use the same crop coordinates — eliminating re-detection jitter across all calibration images.

---

🛠️ Skills Demonstrated

Computer Vision · Image Processing · MATLAB · Hough Transform · Cascade Classifiers · Canny Edge Detection · Laplacian Filtering · Geometric Calibration · Gaze Estimation · Iris/Pupil Segmentation