Deven Varu

Computer vision research / USF Research

Road Asset Detection with YOLO

Computer vision pipeline for Boone County that detects and maps road infrastructure from stitched 360-degree roadway imagery.

University of San Francisco research / Computer vision research assistant / Production-ready

Python / YOLOv8 / Roboflow / AWS / React / Computer Vision

Road Asset Detection with YOLO project media preview

Highlights

  • Evaluated YOLO, LaneNet, and DINOv2 on a 7,650-image dataset.
  • Trained YOLO models for lane markings, shoulders, barriers, streetlights, and other visually similar assets.
  • Built a 3D visualization and validation tool, improving accuracy to 92% mAP and reducing manual labeling.

Context

This University of San Francisco research project supported Boone County roadway analysis using stitched 360-degree roadway imagery.

What I built

  • A computer vision pipeline for lane markings, shoulders, barriers, streetlights, and other visually similar road assets.
  • YOLO model training and evaluation alongside LaneNet and DINOv2 on a 7,650-image dataset.
  • A 3D visualization and validation tool for inspecting detections and reducing manual labeling.

Result

The workflow improved accuracy to 92% mAP and reduced manual labeling work for road infrastructure detection.

Related projects