Deep Learning-Based Road Damage Detection and Severity Prioritization

An automated, camera-based pipeline that detects road defects with YOLOv8, scores each frame with a composite Road Severity Index (SI), and visualises the results on an interactive Folium map for maintenance prioritisation.

Trained on the Road Damage Dataset 2022 (RDD2022) across four defect classes: longitudinal cracks, transverse cracks, alligator cracks, and potholes.

Final model (YOLOv8s, 50 epochs): mAP@50 57.5% · mAP@50–95 30.1% · Precision 61.3% · Recall 55.1%.

---

Project Structure

RoadImageDetection/
├── pipeline.py              # ★ Master integration: images → SI → ranked CSV → map
├── evaluate.py              # ★ Option A re-evaluation (hi-res + TTA, no retraining)
├── generate_csv.py          # Human-validation template generator
├── apply_real_si.py         # Builds the severity_index notebook
│
├── backend/                 # Data + model + API + map backend
│   ├── audit_and_split.py       # (Rubin) dataset audit + stratified 80/10/10 split
│   ├── augmentation_pipeline.py # (Rubin) Albumentations augmentation pipeline
│   ├── severity_map.py          # (Sandesh) Folium map generation
│   ├── severity_points.json     # (Sandesh) geocoded detection store
│   ├── api.py                   # FastAPI service
│   ├── requirements.txt
│   └── model/                   # (Arjit) training + weights
│       ├── train_colab.ipynb    # YOLOv8 training notebook (Colab)
│       ├── rdd2022.yaml         # dataset config
│       ├── predict_sample.py    # minimal inference example
│       └── weights/best.pt      # final trained model
│
├── severity/                # (Adarsha) Severity Index
│   ├── si_utils.py              # single source of truth: weights, grading, smoothing
│   ├── generate_segment_report.py  # per-folder ranked CSV
│   ├── severity_index.ipynb     # SI derivation + validation notebook
│   └── *.png                    # validation / sensitivity / distribution figures
│
├── RDD2022_runs/            # training runs + validation outputs (curves, matrices)
├── frontend/                # Vite + React frontend (separate app)
└── main.tex                 # final project report (LaTeX)

Team / Domain Ownership

Member	Domain	Key files
Rubin	Data	backend/audit_and_split.py, backend/augmentation_pipeline.py
Arjit	Model	backend/model/, train_colab.ipynb, weights/best.pt
Adarsha	Severity	severity/si_utils.py, severity/severity_index.ipynb
Sandesh	Maps	backend/severity_map.py, backend/map/

---

Setup

python -m venv .venv
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate

pip install -r backend/requirements.txt

---

Usage

1. Train the model (Colab / GPU)

Open backend/model/train_colab.ipynb, or run directly:

yolo train model=yolov8s.pt data=backend/model/rdd2022.yaml \
    epochs=50 imgsz=640 batch=16 lr0=0.01 cos_lr=True

2. Evaluate — with Option A inference-time boost (no retraining)

Re-validates the trained weights at higher resolution with test-time augmentation, targeting recall on the thin-crack classes. Regenerates all PR/F1/P/R curves and the confusion matrix.

python evaluate.py --model backend/model/weights/best.pt \
    --data backend/model/rdd2022.yaml --imgsz 1280

Outputs metrics to stdout and figures to runs/detect/val_optionA/.

3. Run the full pipeline

End-to-end: YOLO inference → per-frame SI → EWMA smoothing → attach GPS coordinates → ranked CSV → interactive map.

python pipeline.py --images_dir <segment_frames> \
    --model backend/model/weights/best.pt \
    --coords backend/severity_points.json \
    --csv_out ranked_severity.csv \
    --map_out maps/severity_map.html

Produces ranked_severity.csv (segments ranked worst-first) and an interactive severity_map.html. Open the HTML in any browser to view the colour-coded markers, heatmap, and layer toggles.

4. Frontend

cd frontend
npm install
npm run dev

---

Severity Index

Each frame is scored as:

SI = Σ ( W_i × Confidence_i × A_rel,i )

where W is the class severity weight (Pothole 1.0, Alligator 0.8, Longitudinal 0.5, Transverse 0.3), Confidence is the detection confidence, and A_rel is the bounding-box area relative to the frame. Per-frame scores are smoothed (EWMA, α=0.3) into segment scores and graded:

Grade	SI Range	Map Colour
Good	SI < 0.005	Green
Fair	0.005 ≤ SI < 0.02	Light Green
Poor	0.02 ≤ SI < 0.05	Orange
Critical	SI ≥ 0.05	Red

---

Scope Note

The project is trained and evaluated entirely on RDD2022.