A Density‐Adaptive, Cloud‐Oriented Framework for Fine‐Grained Object Recognition in Large‐Scale Multimodal Urban LiDAR
Transactions in GIS · 2026-11 · 30 (7)
DOI: 10.1111/tgis.70442 ↗Kurz gefasst
Fachartikel in der Zeitschrift Transactions in GIS, online erschienen am 06.11.2026. Verfasst von Rytis Maskeliūnas und 3 weiteren Autor:innen. Die Arbeit nennt 67 Literaturangaben.
Abstract (Originalsprache)
National‐scale urban LiDAR datasets challenge conventional monolithic processing workflows because of their large volume, multimodal heterogeneity, and spatially non‐uniform point density, making scalable architectures that jointly exploit geometric, spectral, and geospatial contextual information essential for fine‐grained recognition. This study develops and evaluates a density‐adaptive, parameter‐efficient, cloud‐oriented extension for scalable tile‐based fine‐grained object recognition in large‐scale multimodal urban LiDAR data. The proposed coarse‐to‐fine pipeline integrates aerial LiDAR‐derived layers, RGB orthophotos, and OpenStreetMap priors into geo‐consistent 512 × 512 raster tiles with 3‐, 6‐, or 9‐channel inputs. RandLA‐Net and SegFormer provide coarse semantic segmentation, while a lightweight 3.8‐million‐parameter YOLOv11‐based detector identifies benches, poles, and traffic signs. To address non‐uniform LiDAR sampling, a per‐pixel density map is used for preprocessing control and adaptive selection of the DBSCAN neighborhood radius, followed by RANSAC‐based geometric filtering. The deployment design further organizes stateless tile‐processing jobs within a fault‐tolerant, horizontally scalable cloud architecture. Evaluation on more than 3.1 billion LiDAR points from Estonian urban areas showed that the full 9‐channel configuration improved scene mIoU by 10.7 percentage points and micro‐object recall by 16.4 percentage points compared with RGB‐only input. Under a unified comparison protocol, the proposed detector achieved 79.1% micro‐object F1 and 86.9% scene mIoU, outperforming the 35‐million‐parameter BEVFusion baseline, which achieved 76.3% and 86.1%, respectively. Density‐stratified experiments further demonstrated stable performance across varying point‐density conditions, while analytical projections based on the measured single‐node runtime indicated favorable horizontal scaling under the stated orchestration and I/O‐overhead assumptions. These findings demonstrate that controlled multimodal fusion, density‐adaptive refinement, and coarse‐to‐fine processing provide an accurate and computationally efficient foundation for national‐scale urban LiDAR analytics, supporting scalable smart‐city mapping, infrastructure monitoring, and urban planning applications.
Zitieren
APA
Maskeliūnas, R., Maqsood, S., Vaškevičius, M., & Gelšvartas, J. (2026). A Density‐Adaptive, Cloud‐Oriented Framework for Fine‐Grained Object Recognition in Large‐Scale Multimodal Urban LiDAR. Transactions in GIS, 30(7). https://doi.org/10.1111/tgis.70442
Vancouver
Maskeliūnas R, Maqsood S, Vaškevičius M, Gelšvartas J. A Density‐Adaptive, Cloud‐Oriented Framework for Fine‐Grained Object Recognition in Large‐Scale Multimodal Urban LiDAR. Transactions in GIS. 2026;30(7). doi:10.1111/tgis.70442
BibTeX
@article{Maskelinas2026,
author = {Maskeliūnas, Rytis and Maqsood, Sarmad and Vaškevičius, Mantas and Gelšvartas, Julius},
title = {A Density‐Adaptive, Cloud‐Oriented Framework for Fine‐Grained Object Recognition in Large‐Scale Multimodal Urban LiDAR},
journal = {Transactions in GIS},
year = {2026},
volume = {30},
number = {7},
doi = {10.1111/tgis.70442}
}
EUVIMED