Established Deep Learning Object Detectors Remain Adequate for Road Infrastructure Monitoring
Technical Annals · 28.01.2026 · 2 (3)
DOI: 10.12681/ta.45596 ↗Kurz gefasst
Fachartikel in der Zeitschrift Technical Annals, online erschienen am 28.01.2026. Verfasst von Anastasios Doulamis und 5 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-SA 4.0).
Abstract (Originalsprache)
The continued advancement of deep learning has introduced a succession of increasingly complex object detection architectures, raising a practical question for road infrastructure (RI) monitoring deployments: are established, well-understood detectors still adequate, or does every new generation mandate migration? This paper addresses that question by systematically benchmarking seven detectors spanning a broad maturity range — SSD MobileNet, Faster R-CNN, EfficientDet, YOLOv5, YOLOv6, YOLOv7, and YOLOv8 — across five RI monitoring scenarios using heterogeneous datasets from Unmanned Aerial Vehicles (UAVs) and ground inspection vehicles. The scenarios encompass UAV-based crack and pothole detection, UAV-based vehicle/traffic monitoring, and ground-vehicle-based crack, pothole, faded-marking, traffic-sign, and traffic-cone detection. Performance is measured using mAP50 and mAP50:95 metrics with default model parameters, deliberately avoiding task-specific fine-tuning to yield reproducible baselines. Results reveal that established architectures remain highly competitive: EfficientDet achieves the best combined accuracy for UAV defect detection (mAP50 60.8%) and leads traffic cone localisation (mAP50 97.0%, mAP50:95 88.9%); older single-stage models such as SSD Mo-bileNet attain 88.8% mAP50 for cone detection; and YOLOv5 nearly matches YOLOv8 for aerial pothole detection (87.6% vs. 87.0%). While YOLOv8 sets the top score for traffic sign recognition (mAP50 98.7%) and aerial vehicle detection (mAP50 77.7%), the performance margin over its predecessors is modest in most scenarios. These findings support the practical case for considering mature detectors for constrained inspection platforms, while controlled device-level profiling remains necessary for final deployment decisions.
Zitieren
APA
Doulamis, A., Katsamenis, I., Protopapadakis, E. P., Bakalos, N., Doulamis, N., & Kalogeras, D. (2026). Established Deep Learning Object Detectors Remain Adequate for Road Infrastructure Monitoring. Technical Annals, 2(3). https://doi.org/10.12681/ta.45596
Vancouver
Doulamis A, Katsamenis I, Protopapadakis EP, Bakalos N, Doulamis N, Kalogeras D. Established Deep Learning Object Detectors Remain Adequate for Road Infrastructure Monitoring. Technical Annals. 2026;2(3). doi:10.12681/ta.45596
BibTeX
@article{Doulamis2026,
author = {Doulamis, Anastasios and Katsamenis, Iason and Protopapadakis, Eftychios Protopapadakis and Bakalos, Nikolaos and Doulamis, Nikolaos and Kalogeras, Dimitrios},
title = {Established Deep Learning Object Detectors Remain Adequate for Road Infrastructure Monitoring},
journal = {Technical Annals},
year = {2026},
volume = {2},
number = {3},
doi = {10.12681/ta.45596}
}
EUVIMED