A Hybrid Intelligence Driven Method for 3D Point Cloud Change Detection

Abstract

Multi-temporal three-dimensional (3D) point cloud change detection is essential for updating reality-based 3D models and urban spatial databases, but remains affected by uneven sampling, occlusion, registration residuals, and local reconstruction errors. This study proposes a hybrid-intelligence-driven method that integrates interpretable surveying-domain measurements with learned point-cloud features. Cross-temporal correspondences are first established to construct three complementary cues: multiscale bidirectional normal-projection responses for geometric dis-placement, attribute differences for auxiliary texture evidence, and density-consistency confidence for reliability modulation. After normalization, these cues are fused into a continuous joint prior and injected into an inherited dual-branch Siamese KPConv network through the change-aware difference attention module. The methodological contribution therefore lies in reformulating and reliability-modulating the prior change evidence rather than introducing a new backbone or attention architecture. Experiments were conducted on the real-world HKCD dataset and the simulation-based Urb3DCD-V2 dataset. The proposed method achieved an mAcc of 96.95%, a recall of 93.96%, and an mIoU of 77.17% on HKCD, and corresponding values of 94.70%, 90.80%, and 88.17% on Urb3DCD-V2. Ablation experiments confirmed that geometric displacement provides the primary evidence, texture information offers complementary discrimination, and density confidence improves overall reliability under the evaluated conditions. However, density modulation slightly reduced the IoU of the removed and demolition classes, revealing a trade-off between suppressing responses associated with inconsistent sampling and preserving genuine object-removal evidence. Overall, the proposed prior representation achieves competitive point-level change-detection performance and provides an interpretable approach for integrating surveying knowledge with deep feature learning in urban 3D model updating.