A deep-learning system that auto-detects and measures road-crack position, length, and width from dashcam footage. Achieves mAP 0.639 detection accuracy at 10 FPS, delivering objective and scalable infrastructure inspection.
Camera angles, mounting positions, and resolutions vary by vehicle, producing inconsistent image quality that degrades detection accuracy.
SolutionCombined literature research with multiple approaches and exposed flexible configuration options. A preprocessing pipeline normalizes images across environments.
Input footage lacked GPS and camera parameters, limiting position and distance-measurement accuracy.
SolutionFine-tuned pretrained models, collaborated with customers to fill in data gaps, and built fallback estimation logic for missing fields.
Commercially-usable road-crack datasets were extremely limited.
SolutionGenerated and augmented proprietary datasets on top of public ones, with rigorous license auditing of every source.
