The Problem

Mapping traffic signs and public-lighting poles usually means reviewing field recordings, identifying each object by hand, and transferring those observations into a spatial inventory. The work is repetitive and difficult to keep consistent as the amount of recorded material grows.

This prototype tested whether the full path from street video to reviewable geospatial data could be automated while still keeping the source evidence available to a human operator.

System Design

The prototype connects four technical stages into one processing workflow:

1. Frame and telemetry ingestion: Street-level frames are processed together with the camera's recorded position and direction of travel.
2. Object detection and classification: YOLO-based models detect traffic signs and public-lighting poles. Traffic signs are also assigned to operationally useful classes such as stop and priority-road signs.
3. Spatial consolidation: Observations of the same object across multiple frames are combined with the capture metadata to estimate a real-world position.
4. Human review: A Streamlit interface presents the resulting objects on a Leaflet/OpenStreetMap map, alongside coordinates, class information, and the frames supporting each detection.

Technical Output

The interface was designed as a validation tool rather than a black-box model demo. For every mapped object, a reviewer can inspect:

• its predicted class and coordinates;
• how many source frames contributed to the result;
• up to three evidence frames from the original video; and
• the object's position along the recorded route.

The resulting inventory can be exported as CSV data or as a standalone interactive HTML map, making it usable outside the prototype application.

What Made It Interesting

The detection model was only one part of the problem. A useful system also had to handle repeated sightings of the same object, connect image-space detections with geographic metadata, and expose enough evidence for someone to verify the result quickly.

Local operating conditions mattered as well: traffic-sign appearance, recording angles, weather, image quality, and variation in public-lighting infrastructure all affect detection quality. The prototype therefore emphasized validation on locally representative street imagery and kept the original evidence frames available for review.

Outcome

Testing showed that the approach could substantially reduce repetitive footage review while producing a more consistent, inspectable infrastructure inventory. The architecture can also be extended with additional object classes or connected to other spatial data sources.

Developed as part of an EDIH Adria Test Before Invest initiative. The project is documented in the official EDIH Adria success story. The figure above is an original reconstruction and does not reproduce imagery from that article.