Browse by author
Lookup NU author(s): Dr Husnain SheraziORCiD
Full text for this publication is not currently held within this repository. Alternative links are provided below where available.
Copyright © 2026. Published by Elsevier Ltd. Accurate three-dimensional (3D) object understanding using artificial intelligence is critical for engineering applications such as environmental monitoring, smart urban planning, and autonomous navigation. Despite recent progress in deep learning-based point cloud analysis, existing classification and segmentation methods often emphasize local geometry or rely on rigid voxelization, resulting in limited long-range dependency modeling, structural information loss, and reduced robustness under sparse and noisy light detection and ranging (LiDAR) sensing conditions. These issues constrain the reliability of current 3D perception systems in complex environments. To address these challenges, this paper proposes a Voxel Graph-Adaptive Transformer (VGaT) framework that integrates voxel-based representation, graph-based spatial reasoning, and adaptive transformer learning within a unified deep learning architecture. VGaT jointly captures fine-grained local geometry and global semantic context through a Voxel Graph Feature Aggregator (VGFA) and a Graph-Adaptive Transformer (GaT) with parallel channel and spatial attention, enabling efficient and content-adaptive information propagation. In addition, a local–global feature aggregation strategy is introduced to improve point-wise segmentation robustness under varying density and occlusion conditions. Extensive experiments on benchmark datasets, including ModelNet40, ScanObjectNN, the Stanford Large-Scale Three-Dimensional Indoor Spaces (S3DIS) dataset, and the SemanticKITTI dataset derived from the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) benchmark, demonstrate consistent performance improvements over existing methods. The proposed framework achieves 94.58% accuracy on ModelNet40 and mean intersection over union (IoU) scores of 89.47%, 86.50%, and 64.70% on the remaining datasets, confirming its effectiveness for reliable and context-aware 3D perception in real-world engineering applications.
Author(s): Mushtaq H, Deng X, Ullah I, Ali M, Raza Sherazi HH
Publication type: Article
Publication status: Published
Journal: Engineering Applications of Artificial Intelligence
Year: 2026
Volume: 181
Print publication date: 01/10/2026
Online publication date: 13/06/2026
Acceptance date: 01/06/2026
ISSN (print): 0952-1976
ISSN (electronic): 1873-6769
Publisher: Elsevier Ltd
URL: https://doi.org/10.1016/j.engappai.2026.115309
DOI: 10.1016/j.engappai.2026.115309
Altmetrics provided by Altmetric
