RoadVision
AI-Based Road Damage Detection and Severity Assessment
Overview
RoadVision is a computer vision-based road inspection system that uses deep learning to automatically detect and classify visible road damage from images and video. Built around a YOLOv8-based object detection pipeline, the system identifies road defects such as potholes and cracks and provides visual detection results along with severity assessment.
The project demonstrates how AI-based image analysis can support faster road-condition monitoring by reducing the effort required for manual inspection and helping prioritize areas that may require maintenance.
RoadVision was developed as a mini project to explore how computer vision and deep learning can automate the initial detection of visible road damage. It is not a production deployment or full road maintenance management system.
The Problem
Traditional road inspections often require manual surveys to identify potholes, cracks, surface deterioration, and other visible road damage across large road networks.
Manual inspection is time-consuming and difficult to scale, especially when monitoring extensive infrastructure where defects may be missed or reported inconsistently.
RoadVision was designed as a mini project to explore how computer vision and deep learning can automate the initial detection of visible road damage and support faster preliminary inspection workflows.
Key Features
Road Damage Detection
The system analyzes uploaded road images or video using a YOLOv8 deep-learning object detection model. It identifies visible road damage including potholes, longitudinal cracks, transverse cracks, and alligator or surface cracking, marking affected areas for review.
Damage Classification
After detecting damage, the system categorizes each road defect by type. The pipeline moves from raw image input through model inference to structured damage classification, making outputs easier to interpret than manual image review.
Severity Assessment
RoadVision estimates the severity of detected damage to help distinguish between relatively minor and more serious road defects. Severity categories such as low, medium, and high support maintenance prioritization decisions.
Visual Detection Results
Detection results are displayed visually with bounding boxes highlighting damaged areas on road images or video frames. Class labels and confidence scores make model outputs easier to interpret at a glance.
Road Condition Insights
The project aggregates detected damage information to provide a broader overview of road conditions, including number of defects, damage types, severity distribution, and overall condition indicators from analyzed inputs.
Streamlit Inspection Interface
A Streamlit-based interface allows users to upload road images or video, run the detection pipeline, and review classification and severity results through a simple interactive workflow.
How It Works
- 1
Road image or video input
A user uploads a road image or video clip for analysis through the Streamlit interface.
- 2
Image processing
OpenCV handles preprocessing and frame extraction to prepare inputs for the detection model.
- 3
YOLOv8 detection
The YOLOv8 deep learning model runs object detection to identify visible road damage and localize defect regions.
- 4
Damage classification
Detected defects are categorized by type, such as potholes, longitudinal cracks, transverse cracks, or surface cracking.
- 5
Severity assessment
The system assigns severity levels to detected damage to help distinguish minor defects from more serious road conditions.
- 6
Visual results and insights
Bounding-box overlays, classification labels, and aggregated condition insights are displayed for inspection and maintenance prioritization support.
System Architecture
Detection Model (YOLOv8)
YOLOv8 object detection model trained and evaluated on road damage data to identify and localize visible defects in images and video frames.
Training Data (RDD2022)
The Road Damage Dataset 2022 (RDD2022) provides labeled road damage examples for model training, validation, and evaluation using standard object-detection metrics.
Image & Video Processing (OpenCV)
OpenCV handles image loading, preprocessing, frame extraction from video, and preparation of inputs for the detection pipeline.
Application Interface (Streamlit)
Streamlit provides the interactive front end for uploading inputs, triggering inference, and reviewing detection, classification, and severity results.
Analysis & Aggregation Layer
Post-detection logic summarizes defect counts, damage types, and severity distribution to provide road condition insights from analyzed inputs.