This repository presents a focused empirical study of classical computer vision methods for edge and line detection in road-like images.
The goal is not to build a production-ready system, but to analyze robustness, parameter sensitivity, and failure modes of widely used techniques.
Despite the dominance of deep learning, classical vision pipelines remain relevant for:
- interpretability
- low-compute environments
- geometric reasoning
This project explores how traditional methods behave under controlled perturbations, providing intuition about when and why they succeed or fail.
The pipeline consists of three main stages:
-
Canny Edge Detection
Automatic threshold selection is used to study the trade-off between noise sensitivity and edge preservation. -
Probabilistic Hough Transform (HoughLinesP)
Line segments are detected and evaluated using an edge coverage metric to assess how well dominant structures are captured. -
RANSAC-based Dominant Line Estimation
A robust fitting approach is applied to edge points to recover a single dominant scene structure under noise and fragmented edges.
- Canny performance is highly sensitive to threshold selection, with moderate ranges providing the best balance.
- HoughLinesP reliably detects strong, continuous road lines but struggles with weak or fragmented boundaries.
- RANSAC is robust to noise and clutter, producing a stable global line model, but is limited to capturing a single dominant structure.
These observations align with known theoretical properties of each method.
This project is analytical rather than benchmark-driven.
It focuses on qualitative and simple quantitative insights rather than optimizing performance metrics or comparing against deep models.
- Quantitative evaluation with ground truth lane annotations
- Multi-model RANSAC for multiple line hypotheses
- Integration with learning-based edge detectors
- Hybrid pipelines combining classical and deep methods
This repository is intended for educational and exploratory research purposes only.