RoadGuardian_AI

Inspiration

RoadGuardian AI was inspired by the need to make road safety monitoring more efficient using Computer Vision. Potholes and other road hazards can be difficult to identify manually, especially when monitoring large areas or long road-survey videos. I wanted to build a system that could automatically detect these hazards from road footage and make the monitoring process faster and more consistent.

What I Built

I developed RoadGuardian AI, a real-time road hazard detection system using a custom-trained YOLO26 model. The system takes road or dashcam video as input and detects potholes across video frames.

I used Python, YOLO26, Ultralytics, OpenCV, and ByteTrack to build the pipeline. ByteTrack helps maintain consistent IDs for detected potholes across frames, while temporal smoothing and a frame-hold mechanism help reduce flickering detections.

The model was trained using a pothole detection dataset containing more than 50,000 images with different road conditions, viewing angles, lighting conditions, and pothole sizes. I trained the model using Kaggle GPU resources and evaluated its performance using metrics such as precision, recall, F1-score, and mAP.

Challenges

One of the main challenges was maintaining stable detections across consecutive video frames. A pothole could sometimes disappear or change slightly between frames, causing flickering results. I addressed this using tracking and temporal smoothing techniques.

Another challenge was handling different road conditions and lighting environments. Training with a diverse dataset helped the model become more robust to these variations.

What I Learned

Through this project, I gained practical experience in training custom YOLO models, video-based object detection, object tracking, model evaluation, and building real-time Computer Vision pipelines. I also learned how important temporal consistency and proper evaluation are when deploying object detection models on real-world video.

Future Improvements

Future versions could include pothole severity estimation, road-damage segmentation, GPS-based hazard mapping, and integration with a mobile or web dashboard for road monitoring.

Built With

• bytetrack
• computervision
• opencv
• python
• pytorch
• ultralytics
• yolo
• yolo26