Automatic extraction of relevant road infrastructure using connected vehicle data and deep learning model

Adu-Gyamfi, Kojo Konadu · 2024 · Crossref

DOI: 10.31274/td-20240617-211

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This thesis addresses the inefficiency and high cost of traditional road infrastructure mapping, which relies on manual surveys or specialized sensors. The author proposes an automated method to extract road infrastructure information, specifically identifying straight roads and intersections, using connected vehicle trajectory data and deep learning. The motivation stems from the need for scalable, real-time updates to road networks to support traffic management, urban planning, and autonomous vehicle navigation, overcoming the limitations of static databases like OpenStreetMap or labor-intensive methods like the Model Inventory of Roadway Elements. The methodology combines geohashing with the YOLOv5 object detection algorithm. Connected vehicle GPS trajectories were collected in Ames, Iowa, and segmented into small rectangular grid cells using geohashing. These spatial segments were converted into image representations, where vehicle trajectories were plotted to visualize road patterns. The images were preprocessed using techniques such as grayscale conversion, flipping, rotation, shearing, blurring, and noise addition to enhance model robustness. The YOLOv5 model was then trained to classify these images into two categories: straight roads and intersections. The experimental design focused on validating the feasibility of this image-based classification approach for automated infrastructure extraction. The results demonstrate high accuracy in classifying road segments. The model achieved an overall classification accuracy of 95%. Specifically, straight roads attained an F1 score of 97%, while intersections achieved an F1 score of 90%. The study compared performance metrics between uncolored and colored trajectory images, noting that visual representations based on average speed helped distinguish features. Confusion matrices and sample outputs were used to analyze false classifications, such as misidentified intersections or straight roads, providing insight into model limitations. The findings validate that geohashing combined with YOLOv5 can effectively identify key road infrastructure elements from raw trajectory data. The significance of this work lies in providing a scalable, data-driven framework for updating road networks without manual intervention. By leveraging ubiquitous connected vehicle data, the approach offers a cost-effective alternative to traditional mapping, enabling more frequent updates and comprehensive coverage. This method supports smarter transportation systems by facilitating better traffic flow optimization, infrastructure maintenance, and autonomous navigation. The thesis concludes that integrating spatial indexing with deep learning is a robust strategy for extracting valuable insights from large-scale mobility datasets, paving the way for more adaptive and efficient transportation management.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• exposure measurement