Use Case Title
Pothole Detection on Roads Using YOLOvX
Team Members
- Sai Mandekar
- Gayatri More
- Laxmi Navhi
Use Case Importance
Potholes pose serious risks to vehicles and pedestrians, often going unnoticed until accidents or complaints occur. Our solution automates pothole detection, helping institutions and municipalities take timely action and improve road safety.
Data Collection and Annotation
Data Collection: Captured 600+ images using a mobile phone under various lighting and weather conditions.
Annotation Classes: One class — “Pothole”
Annotation Tool Used: Roboflow
Total Images:
303 for training
100 for validation
76 for testing
Sample Images:
Model Training and Validation
Model Used: YOLOv8
Monitored Metrics: Precision, Recall, and mAP
Data Augmentation: Conducted after low-confidence results in night/shadow conditions; included new night-time images for better generalization.
Outcome: Accuracy improved significantly, especially under normal daylight conditions.
Training Details: 25 epochs, batch size 16, image size 400×400.
Model Deployment and Demo Video
Real-World Performance:
We deployed the model into the YOLOvX mobile app
Accuracy: ~90% in clear lighting, ~80% at night, ~45% in shadows
Deployed Models: Single trained YOLOvX model integrated into Android mobile app using a Flask-based API backend.
Demo Video Link: Google Drive - Pothole Detection Demo
https://drive.google.com/drive/folders/1JXMYxN6LFN8soQ_TllxvrG-bsKQYhicE?usp=sharing
Conclusion
Our model achieved high accuracy in detecting potholes from both live video and dashcam footage. Real-time mobile integration was successful, although low-light and shadow scenarios remain areas for improvement. We learned that fine-tuning parameters and augmenting datasets significantly improves performance in challenging conditions.
