Predicting the adjacency matrix to construct lane topology of signalised intersections with aerial imagery and deep learning
(2026) In Transportation Research Interdisciplinary Perspectives 35.- Abstract
Access to detailed network data is often constrained by access-controlled datasets or by extensive manual labour. Using aerial imagery and a simplified OpenStreetMap road-level graph, a novel approach is proposed to extract lane-level road topology graphs at signalized intersections based on detection of common road objects, classification of lane attributes, and a simultaneous lane-to-lane connectivity classifier. The approach bridges the gap in methodologies focused on image segmentation or direct graph extraction. Based on individual intersection arms, we achieve an F1 score of 0.88 in predicting the number of approach lanes, 0.87 number of turn movements, and 0.63 in predicting the number of lane types using seven lane type classes.... (More)
Access to detailed network data is often constrained by access-controlled datasets or by extensive manual labour. Using aerial imagery and a simplified OpenStreetMap road-level graph, a novel approach is proposed to extract lane-level road topology graphs at signalized intersections based on detection of common road objects, classification of lane attributes, and a simultaneous lane-to-lane connectivity classifier. The approach bridges the gap in methodologies focused on image segmentation or direct graph extraction. Based on individual intersection arms, we achieve an F1 score of 0.88 in predicting the number of approach lanes, 0.87 number of turn movements, and 0.63 in predicting the number of lane types using seven lane type classes. Although key topology aspects are predicted successfully, performance is comparable or inferior to a simpler rule-based model. The results demonstrate the viability of the approach, but confirming practical usefulness will require training and evaluation on a larger and more varied dataset across cities and intersection types.
(Less)
- author
- Nilsson, Hugo and Oucheikh, Rachid LU
- organization
- publishing date
- 2026-01
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- Computer Vision, Deep Learning, Graph Extraction, Object Detection, Remote Sensing, Road topology
- in
- Transportation Research Interdisciplinary Perspectives
- volume
- 35
- article number
- 101779
- publisher
- Elsevier
- external identifiers
-
- scopus:105024467940
- ISSN
- 2590-1982
- DOI
- 10.1016/j.trip.2025.101779
- language
- English
- LU publication?
- yes
- additional info
- Publisher Copyright: © 2025 The Author(s).
- id
- 574b87c4-0634-4bbc-a2d0-3c5eb7369618
- date added to LUP
- 2026-03-30 16:14:24
- date last changed
- 2026-03-30 16:15:17
@article{574b87c4-0634-4bbc-a2d0-3c5eb7369618,
abstract = {{<p>Access to detailed network data is often constrained by access-controlled datasets or by extensive manual labour. Using aerial imagery and a simplified OpenStreetMap road-level graph, a novel approach is proposed to extract lane-level road topology graphs at signalized intersections based on detection of common road objects, classification of lane attributes, and a simultaneous lane-to-lane connectivity classifier. The approach bridges the gap in methodologies focused on image segmentation or direct graph extraction. Based on individual intersection arms, we achieve an F1 score of 0.88 in predicting the number of approach lanes, 0.87 number of turn movements, and 0.63 in predicting the number of lane types using seven lane type classes. Although key topology aspects are predicted successfully, performance is comparable or inferior to a simpler rule-based model. The results demonstrate the viability of the approach, but confirming practical usefulness will require training and evaluation on a larger and more varied dataset across cities and intersection types.</p>}},
author = {{Nilsson, Hugo and Oucheikh, Rachid}},
issn = {{2590-1982}},
keywords = {{Computer Vision; Deep Learning; Graph Extraction; Object Detection; Remote Sensing; Road topology}},
language = {{eng}},
publisher = {{Elsevier}},
series = {{Transportation Research Interdisciplinary Perspectives}},
title = {{Predicting the adjacency matrix to construct lane topology of signalised intersections with aerial imagery and deep learning}},
url = {{http://dx.doi.org/10.1016/j.trip.2025.101779}},
doi = {{10.1016/j.trip.2025.101779}},
volume = {{35}},
year = {{2026}},
}