Computer vision research / USF Research
Road Asset Detection with YOLO
A computer vision pipeline for detecting and mapping roadway infrastructure from stitched 360-degree imagery collected for Boone County.
Python / YOLOv8 / Roboflow / AWS / React / Computer Vision
Highlights
- Evaluated YOLO, LaneNet, and DINOv2 across a 7,650-image roadway dataset to determine the most effective approach for detecting road infrastructure.
- Trained YOLO models to distinguish visually similar assets including lane markings, shoulders, barriers, streetlights, and other roadway features.
- Built a 3D visualization and validation tool for inspecting detections in the context of the roadway rather than reviewing predictions as isolated images.
- Developed the resulting workflow to 92% mAP while reducing the amount of manual labeling required for roadway asset detection.
Problem
Boone County roadway imagery contains infrastructure that is useful for mapping and maintenance, but extracting those assets manually from thousands of stitched 360-degree images is expensive and time-consuming.
The computer vision problem is also more difficult than simply detecting visually distinct objects.
Many roadway assets share similar shapes, colors, and surroundings. Lane markings, shoulders, barriers, and other infrastructure can appear differently depending on camera position, road geometry, lighting, distance, and image stitching.
The goal was to build a workflow that could identify these assets reliably enough to turn large collections of roadway imagery into usable infrastructure data.
What I built
I worked on the computer vision pipeline from model evaluation through validation.
The project began by evaluating multiple approaches, including YOLO, LaneNet, and DINOv2, on a dataset containing 7,650 roadway images.
Based on those experiments, the final detection workflow was built around YOLO, with models trained for multiple categories of road infrastructure.
The pipeline covered assets including:
- Lane markings
- Shoulders
- Barriers
- Streetlights
- Other visually similar roadway features
The work also extended beyond model training.
I built a 3D visualization and validation tool that made it possible to inspect model detections in the context of the roadway and verify whether predicted infrastructure aligned with the underlying imagery.
That created a feedback loop between model predictions, human validation, and dataset improvement rather than treating model evaluation as a single offline metric.
How it works
The workflow moves from raw roadway imagery to validated infrastructure detections.
- Stitched 360-degree roadway imagery is prepared for the detection pipeline.
- Trained YOLO models identify relevant infrastructure within the images.
- Predictions are converted into structured detection results.
- Those results are surfaced through the validation tooling so detections can be inspected in context.
- Incorrect or uncertain detections can be identified and used to improve the dataset and subsequent model iterations.
- Validated detections can then be used as part of the roadway mapping workflow.
This combination of model inference and validation tooling made the system more useful than a standalone object detector.
Technical challenge
A major challenge was distinguishing between visually similar roadway classes.
Road infrastructure does not always have clean visual boundaries, and the same asset can look substantially different depending on viewing angle, road conditions, image stitching, and distance from the camera.
That made the work as much about dataset quality and validation as model architecture.
Evaluating multiple model approaches helped determine which direction was most practical, while the validation workflow provided a way to inspect where the model was succeeding or failing rather than relying only on aggregate metrics.
Result
The final workflow reached 92% mAP and reduced the amount of manual labeling required for roadway infrastructure detection.
More importantly, the project produced a complete workflow around the model:
roadway imagery -> detection -> validation -> usable infrastructure data
rather than stopping at a trained computer vision model.