As dynamic virtual replicas of the physical world, digital twins require highly timely updates of the reality-captured 3D models that constitute their foundation. To sustain high-fidelity synchronization with physical entities, efficient and accurate detection of real-world changes is essential. This paper addresses a core challenge in digital twin applications: adaptive change detection and intelligent analysis of multi-temporal reality-captured 3D models-traditional methods struggle with low efficiency, high false rates, and lack of change semantics when handling large-scale, noisy, non-rigid real-scene 3D data.This paper proposes ACD-IA, a novel integrated deep learning framework that: it adaptively extracts point cloud geometric/texture features via multi-scale fusion (suppressing illumination/data noise), uses an attention-based Siamese network to measure cross-temporal feature differences and localize changes (coarse-to-fine), and integrates lightweight semantic segmentation to identify change types (e.g., new buildings, vegetation growth). Experiments on a self-built multi-temporal UAV/LiDAR dataset (urban scenes) and the public SHREC 2023 benchmark show ACD-IA outperforms mainstream methods in precision, recall, F1-score, and provides rich semantics, supporting digital twin dynamic updates and decision-making.
Rong Chen (Thu,) studied this question.