Research Article

Rapid recognition and localization of virtual assembly components in bridge 3D point clouds based on supervoxel clustering and transformer

School of Computer Science, Semyung University, 65 Semyung-ro, Jecheon-si 27136, Chungcheongbuk-do, Republic of Korea

^* e-mail: 2023624801@semyung.ac.kr

Received: 29 July 2025
Accepted: 12 August 2025

Abstract

Traditional rule-based manual bridge inspection methods often suffer from low efficiency and poor accuracy, making them inadequate for the demands of industrial-scale production. This study aims to achieve rapid recognition and localization of virtual assembly components within bridge 3D point clouds by constructing an intelligent analytical framework that integrates supervoxel clustering with a Transformer architecture. Specifically, an improved supervoxel clustering algorithm is developed, deeply integrating geometric morphology, density distribution, and structural response features to generate multimodal voxel units, thereby enhancing the semantic representation of local features. A graph-based Transformer module is introduced to model spatial relationships and semantic associations among supervoxel nodes through a self-attention mechanism, effectively integrating global contextual information. Additionally, a voxel voting strategy within a pose estimation module is employed to optimize component localization accuracy, forming an end-to-end recognition and localization system. The proposed model demonstrates excellent performance across multiple datasets, including Stanford Large-Scale 3D Indoor Spaces Dataset, ETH Zurich Building Dataset, International Society for Photogrammetry and Remote Sensing Benchmark Dataset, and National Building Museum Point Cloud Dataset. Compared to baseline models, the proposed approach achieves improvements of over 21.5% in semantic segmentation Mean Intersection over Union, instance recognition accuracy, and pose regression precision. In complex multi-box girder bridge scenarios, the recognition accuracy for small-scale connectors improves by up to 37.1%. Computational efficiency increases by more than 18.7%, with inference time reductions of up to 31.5% when processing large-scale data. Overall improvements in bridge component recognition exceed 22.4%, with recognition accuracy for critical connection components increasing by up to 37.4%, and localization accuracy improving by over 26.2%, reaching up to 35.9% for key node localization. The results demonstrate that the proposed model effectively addresses critical challenges in processing bridge point cloud data through multimodal feature fusion and global structural reasoning, significantly enhancing component recognition accuracy and localization precision in complex scenes while maintaining a balance between algorithmic efficiency and model performance. This study provides an efficient solution for the digital delivery and quality control of intelligent bridge construction. By integrating finite element analysis with deep learning, the model enhances semantic understanding of bridge structural functions, contributing significantly to the advancement of intelligent bridge engineering.