Pothole Detection using YOLOv8 with Road Severity Grading

Potholes damage vehicles, slow traffic and cause accidents, yet most road surveys still depend on engineers driving the network and noting damage by eye — slow, inconsistent, and impossible to repeat across a whole city. Computer vision changes the economics of this task: a camera on a survey vehicle captures the road surface, and a trained detector flags damage in every frame. The RDD2022 dataset — 47,420 road images from six countries, released for the CRDDC 2022 challenge — gives this a standard benchmark, and YOLOv8's single-shot architecture predicts bounding boxes, classes and confidence in one pass. But finding potholes is only half the job: a road authority also needs to know which stretch to fix first. This project pairs a YOLOv8 detector fine-tuned on RDD2022's D40 pothole class with a documented severity-grading rule that converts each detection's box area into HIGH, MEDIUM or LOW maintenance priority, producing a ranked repair list.

1. The RDD2022 release (47,420 road images, four damage classes D00, D10, D20, D40) is organized into YOLO label format; the D40 class is the primary target and a stratified train/validation split is created.
2. Images are resized to 640×640 and normalized; training-time augmentation (mosaic, horizontal flip, brightness/HSV jitter) simulates varied lighting and camera angles.
3. A YOLOv8n or YOLOv8s checkpoint pre-trained on COCO is fine-tuned for a documented number of epochs with early stopping on validation mAP; the best checkpoint (best.pt) is saved.
4. The validation split is evaluated once: precision, recall, mAP@50 and mAP@50–95, the confusion matrix and PR curves are generated for the report.
5. Severity grading is applied to every detection: box area as a fraction of the frame maps to HIGH (≥ 3%), MEDIUM (≥ 0.9%) or LOW priority, and detections are ranked with their confidence scores into the maintenance-priority list.
6. In the web demo, an uploaded road photo is analyzed and the app renders D40 bounding boxes with class label, confidence and severity badge, plus damaged-area statistics and the exportable report. The browser page is a classical-vision preview of this flow; the trained YOLOv8 weights run in the project's Python pipeline.

Tech stack:

• Python 3, Ultralytics YOLOv8
• PyTorch
• OpenCV (image I/O & annotation)
• RDD2022 dataset (CRDDC 2022)
• Jupyter Notebook (training & evaluation)
• NumPy, Matplotlib (metrics & plots)
• HTML · CSS · JavaScript (web demo)

Dataset & model details

• Dataset: RDD2022 — Road Damage Dataset 2022, released through the Crowdsensing-based Road Damage Detection Challenge (CRDDC 2022). 47,420 road images from 6 countries (Japan, India, Czech Republic, Norway, United States, China), 55,000+ annotated instances across 4 official classes: D00 (longitudinal crack), D10 (transverse crack), D20 (alligator crack), D40 (pothole).
• Task: Object detection; input = 640×640×3 road image, output = bounding boxes with class labels and confidence scores. This project specializes on the D40 class (the dataset's pothole-containing class).
• Model: YOLOv8n or YOLOv8s (Ultralytics), anchor-free single-shot detector pre-trained on COCO and fine-tuned on RDD2022; inference applies non-maximum suppression and a configurable confidence threshold.
• Metrics: Precision, recall, mAP@50 and mAP@50–95 (the standard Ultralytics validation metrics), confusion matrix, PR curves. Design target: approximately 0.75 mAP@50 for the D40 class on the validation split — the report carries the training run's measured figure, never a pre-filled claim.
• Severity grading: A documented design rule of this project (not a dataset label): detection box area as a fraction of the frame — ≥ 3% → HIGH, ≥ 0.9% → MEDIUM, below → LOW maintenance priority.

• [YOLOv8 fine-tune on RDD2022] A YOLOv8n / YOLOv8s detector, pre-trained on COCO and fine-tuned on the RDD2022 road-damage dataset with the D40 class (rutting, bump, pothole, separation) as the primary target — documented train/validation split and augmentation policy.
• [Documented severity-grading rule] A design rule applied to every detection: bounding-box area ≥ 3% of the frame → HIGH priority, ≥ 0.9% → MEDIUM, below that → LOW; combined with the confidence score into a ranked maintenance-priority list.
• [Standard detection metrics] Evaluation with precision, recall, mAP@50 and mAP@50–95 on the validation split, plus a confusion matrix and PR curves — the report carries only the figures the training run actually produces.
• [Batch inference pipeline] Point it at a single road photo or a folder of survey images and get annotated images (D40 boxes with class, confidence and severity badge) plus the ranked priority report.
• [Interactive web demo] Upload a road photo or pick from the bundled samples, watch detections appear with severity badges, review the damaged-area statistics, and export the maintenance-priority list as a text report.
• [Reproducible training notebook] Dataset organization into YOLO label format, preprocessing, augmentation, training configuration, early stopping and evaluation in one notebook with every choice documented.
• [Complete viva kit] Project report PDF, PPT presentation and a Q&A document covering the RDD2022 taxonomy, YOLOv8's anchor-free detection head, mAP-style metrics and the severity rule.

• Reproducible training & evaluation Jupyter notebook (data prep, YOLO labels, augmentation, training, metrics)
• Trained YOLOv8 weights (best.pt) with the inference pipeline script
• Batch inference: annotated images + ranked maintenance-priority report
• Interactive single-file web demo with sample road images and exportable report
• Evaluation plots: PR curves, confusion matrix, training curves
• Project report PDF (dataset, methodology, results, severity rule, limitations)
• PPT presentation for final review
• Viva Q&A preparation document (RDD2022 taxonomy, YOLOv8, mAP metrics, grading rule)