Pothole Detection.

A small computer-vision system that sees road damage before a council inspector does — built as a teaching exercise in object-detection from scratch.

Introduction

This project implemented an object-detection system using Faster-RCNN in MATLAB to automatically identify potholes on road surfaces. The goal was to develop a robust solution that could assist in road maintenance and improve driving safety by accurately pinpointing areas requiring repair.

Project goals & features

• Accurate pothole identification. A model that distinguishes potholes from other road features and anomalies.
• Faster-RCNN implementation. Leveraging the architecture for efficient and precise object detection.
• MATLAB integration. Using MATLAB as the primary environment for training and deployment.
• Image data processing. Handling and preprocessing image datasets of varying road conditions to train the detector.

Technologies used

Challenges & solutions

A significant challenge was acquiring a diverse and representative dataset of potholes under varying lighting and weather conditions. This was addressed by augmenting existing datasets and carefully selecting images that captured a wide range of pothole characteristics. Another hurdle was optimising the Faster-RCNN model for real-time or near real-time detection, which involved fine-tuning hyperparameters and exploring model compression techniques within MATLAB’s capabilities.

Learnings & future enhancements

Through this project I gained a deeper understanding of object-detection algorithms, particularly Faster-RCNN, and practical experience implementing them in MATLAB. I also sharpened my image-preprocessing and model-evaluation skills. Future enhancements include integrating the system with GPS for precise location mapping of potholes, deploying the model on embedded systems for in-vehicle detection, and exploring other deep learning architectures for improved accuracy and speed.