👁️ Computer Vision Course Portfolio

This repository showcases a comprehensive set of projects implemented in MATLAB, covering fundamental to advanced topics in image processing, geometric computer vision, and pattern recognition.

---

📂 Projects

1. 📐 Camera Calibration and Homography

Focus: Geometric Vision & 3D Projections

This project explores how cameras perceive the 3D world and how to project virtual objects onto a 2D image plane with high precision.

• 3D Robot Simulation: Representation of virtual 3D robots (cubes) with Aruco markers on top, allowing for precise spatial localization.
• Intrinsic & Extrinsic Parameters: Implementation of focal length, principal point, and rotation/translation matrices to simulate realistic camera physics.
• Homography-Based Projection: Mathematical mapping of coordinates to project the Aruco markers and robot bodies onto the table with the chosen perspective.
• Dynamic Simulations: Scripts to simulate moving 3D robots and changing camera views within an augmented reality environment.

Topics covered:

• Pinhole Camera Model: Understanding intrinsic parameters (K) and extrinsic transformations ([R|t]).
• Homography Calculation: Implementing SVD-based Direct Linear Transform (DLT) to solve the homography matrix H.
• 3D-to-2D Projection: Transforming world coordinates into image pixels for augmented reality.
• Image Composition: Blending virtual objects with real-world backgrounds using transparency and perspective.

Execution:

1. Open MATLAB.
2. Navigate to the 01_Camera_Calibration_and_Homography/Matlab folder.
3. Run the main script main_execution_script.m section by section to see the results.

📊 Extraction Results:

🧊 Dinamic 3D Rendering	🎞️ Perspective Motion
3D cubes and camera frame correctly rendered on the table.	Dynamic perspective update from the camera frame.

🎥 Visual Demos:

• 🧊 3D Cube Projection & Shading - Demonstration of overlapping 3D volumes on a planar surface.
• 🏁 Moving Aruco Markers - Real-time animation of markers with perspective updates.
• 🎥 Virtual Camera Flight - Simulation of a moving observer view.

---

2. 🥦 Vegetable Segmentation and Classification

Focus: Image Processing & Feature Extraction

Development of an automated pipeline to detect, segment, and identify different types of vegetables within a scene using advanced image processing.

• Intelligent Segmentation: Using HSL and RGB color-space thresholds to isolate specific vegetable species from complex backgrounds.
• Morphological Refinement: Applying cleanup operations (imopen, imclose) and binary labeling to isolate individual objects.
• Automated Classification: Extracting geometric features (Area, Eccentricity, Bounding Boxes) to distinguish between different types of vegetables.
• Multi-Instance Detection: Capability to detect and count multiple instances of the same vegetable type or classify a diverse set of species simultaneously.

Topics covered:

• Color Segmentation: Exploring RGB, HSL, and Lab spaces to isolate objects by chromaticity.
• Adaptive Thresholding: Handling non-uniform lighting conditions with local binarization techniques.
• Morphological Ops: Using Erosion, Dilation, and Fill to refine segmented object masks.
• Connected Component Analysis: Extracting region properties (Area, Centroid, Bounding Box) for labeling.

Execution:

1. Open MATLAB.
2. Navigate to the 02_Vegetable_Segmentation_and_Classification/Matlab folder.
3. Run the main script vegetable_classification.m with the desired image path to see the results.

📊 Extraction Results:

Classification of Different Vegetables	Multiple Objects of the Same Type
Automatic classification of various species (Kiwis, Tomatoes, Peppers, etc.).	Segmentation and individual count of multiple instances of the same vegetable.

---

3. 🛡️ Marker Detection and Homography Solve

Focus: Template Matching & Image Rectification

Advanced system for the detection and identification of planar markers (similar to Aruco/QR) to reconstruct scene geometry.

• Marker Identification: Scanning complex scenes for potential candidates and verifying their identity using binary template matching.
• Perspective Rectification: Using 4-point homography to "flatten" tilted or distorted markers into a canonical view for reliable identification.
• Geometric Verification: High-precision corner extraction and coordinate mapping to handle lighting variations and occlusions.

Topics covered:

• Marker Geometry: Identification of square patterns and corner extraction in noisy images.
• Homography-based Rectification: Warping tilted markers into a canonical frontal view for identification.
• Template Correlation: Matching detected patches against known marker definitions.
• Inverse Projection: Mapping marker IDs back into the scene with accurate spatial alignment.

Execution:

1. Open MATLAB.
2. Navigate to the 03_Homography_and_Marker_Detection/Matlab folder.
3. Run the script marker_detection_basic.m or marker_detection_advanced.m with the desired image path to see the results.

📍 Detection Output:

Aruco Intersection Logic	Final Marker Mapping
Detail of the geometric intersection points used for homography estimation.	Visual confirmation of detected markers with their corresponding IDs and accuracy indicators.

---

4. 🔢 MNIST Handwritten Digit Recognition

Focus: Pattern Recognition & Machine Learning

Implementation of a complete pattern recognition system for the classification of handwritten digits using the historical MNIST dataset.

• Dataset Processing: Loading and pre-processing 70,000 samples (60k train / 10k test) of diverse handwritten numbers.
• Statistical Classifiers: Implementation of a Mahalanobis Distance classifier to handle high-dimensional feature distributions.
• Feature Optimization: Using Fisher's Linear Discriminant Analysis (LDA) to project data into an optimal space that maximizes class separability and accuracy.

Topics covered:

• Feature Vector Construction: Flattening 2D images into high-dimensional vectors for processing.
• Supervised Learning: Training models on labeled data and evaluating on unseen test sets.
• Mahalanobis Distance: Statistical classification considering the covariance of the data distribution.
• Fisher's LDA: Maximizing class separability in reduced feature spaces for higher accuracy.

Execution:

1. Open MATLAB.
2. Navigate to the 04_MNIST_Handwritten_Digit_Recognition/Matlab folder.
3. Run the script mnist_recognition_basic.m or mnist_recognition_advanced.m with the desired dataset path to see the results.

📈 Recognition Performance:

✅ Success Sample	❌ Failure Sample
Correctly classified digit with high confidence.	Failure due to the digit being poorly drawn.

• 📼 Full Digit Recognition Demo

---

🛠️ Getting Started

Prerequisites

• MATLAB (R2021a or newer recommended).
• Image Processing Toolbox.

---

Developed by Fernando Román and Andrés Martínez | Computer Vision Course, University of Seville