The feature image for the project Intelligent Road Safety Monitoring System.
The feature image for the project Intelligent Road Safety Monitoring System.
The feature image for the project Intelligent Road Safety Monitoring System.
The feature image for the project Intelligent Road Safety Monitoring System.

UNIVERSITY PROJECT | 2024

Intelligent Road Safety Monitoring System

A drone-based road defect detection system that deploys a UAV to autonomously survey roads, uses YOLOv8 to classify potholes, cracks, manholes, and blurred markings, then pins every finding onto an interactive map — giving both road authorities and commuters real-time awareness of hazards before they cause accidents.

YOLOv8
Python

OVERVIEW

IRSMS was built to address the inefficiency of manual road inspection in Malaysia, where road defects account for over 11% of traffic fatalities. A DJI Air 3 drone captures aerial images at fixed intervals while its embedded GPS logs each shot's coordinates. A Python pipeline on Google Colab then extracts the GPS metadata, renames images with their coordinates, and feeds them through a fine-tuned YOLOv8 model trained on the RDD 2020 dataset — augmented with locally collected Malaysian road defects. Detected defects are plotted onto a Folium leaflet map as an HTML file, accessible from any internet-connected device by both maintenance authorities and road users.

IMPLEMENTATION

TECH STACK

Drone

DJI Air 3

MlModel

finetuned YOLOv8

Language

Python

Libraries

Pillow (PIL), Folium, OS

Cloud

Google Drive, Google Colab (Nvidia Tesla T4 GPU)

Dataset

Road Damage Dataset (RDD) 2020

FEATURES

UAV aerial image capture with embedded GPS metadata at fixed time intervals

Automated YOLOv8 detection of 4 defect classes: potholes, cracks, manholes, and blurred road markings

GPS EXIF extraction and coordinate-based image renaming for traceability

Interactive virtual map with geotagged defect markers and image previews via Folium

End-to-end cloud pipeline: DJI Fly App → Google Drive → Google Colab → HTML map

Dataset augmented with Malaysia-specific road defects for local accuracy

CHALLENGES & SOLUTIONS

Selecting the right detection model for drone-captured aerial imagery

Conducted a systematic literature review comparing YOLOv5, YOLOv7, YOLOv8, and RCNN variants across accuracy, inference speed, and suitability for real-time aerial detection. Selected YOLOv8 based on its leading mAP of 91.1% and lightweight inference time of 8.8ms per image while remaining compatible with Google Colab GPU resources.