基于几何结构优化和熵注意力机制的点云变化检测方法

doi:10.11947/j.AGCS.2026.20250415

摘要/Abstract

摘要：

三维变化检测旨在从同一场景不同时相的点云中获取地物变化信息，对理解地表动态变化过程至关重要。然而，现有基于深度学习的点云变化检测方法常因忽视局部几何细节和全局上下文特征建模不足，导致在点云密度不均、存在孔洞及视角重叠的复杂城市场景中性能受限。本文提出一种基于几何结构优化和熵注意力机制的点云变化检测方法。在编码器阶段组合利用局部复杂度分析模块和分层采样模块，通过融合点云曲率、密度、信息熵三维特征，共同量化点云邻域的空间分布。并以此作为下采样先验知识计算采样概率，通过自适应的分层采样方法在缩小点云规模的同时优化宏观几何结构，兼顾了点云全局覆盖和重点区域增强。改进了传统双路径注意力机制，将信息熵作为先验信息，通过对注意力权重施加熵偏置，动态引导空间与特征通道注意力聚焦于结构复杂的高信息量区域，增强了全局上下文特征表征。选取7种主流基线网络，在SLPCCD街景三维变化检测数据集和Urb3DCD模拟机载激光点云变化检测数据集上进行模型对比试验，结果表明，本文方法在定量和定性分析方面均优于其他主流网络，在SLPCCD数据集上的平均交并比比基线网络3DCDNet提升超17.01%，比先前最优模型Ms-DANet提升了3.35%。本文方法能够有效改善因点云密度不均、点云孔洞、视角重叠造成的错检及漏检问题。

关键词: 三维点云, 变化检测, 几何结构优化, 熵注意力, 双路径注意力机制

Abstract:

The aim of three-dimensional change detection technology is to identify temporal feature alterations from multi-temporal point clouds within identical geographical scenes, which plays a crucial role in understanding surface dynamics. Conversely, contemporary deep-learning-based approaches always underperform in complex urban settings characterized by non-uniform point density, occlusions, and multi-viewpoint overlaps, which arises from insufficient incorporation of local geometric intricacies and global contextual dependencies. Therefore, we propose GS-EANet (geometric structure refinement and entropy-attention network) which effectively models scene-aware geometric complexity for robust 3D change detection. During the encoding stage, GS-EANet combines local complexity analysis module and stratified sampling module, which uses three key features including point curvature, density, and information entropy, to describe the spatial arrangement of points in a neighborhood. This multi-feature fusion serves as prior knowledge for calculating sampling probabilities. The stratified sampling module based on sampling probability notably optimizes geometric structures while down-sampling, thereby achieving an optimal balance between global scene coverage and emphasis on critical regions. Furthermore, we advance conventional dual-path attention mechanisms by incorporating information-entropy, which dynamically steers spatial and channel attention toward structurally complex regions, substantially enhancing global contextual representation. Comprehensive evaluations on the SLPCCD and Urb3DCD datasets demonstrate that GS-EANet surpasses seven state-of-the-art baseline methods in both quantitative and qualitative assessments. Notably, on the SLPCCD dataset, GS-EANet achieves a mean intersection over union (mIoU) of 17.01% higher than 3DCDNet. Moreover, GS-EANet outperforms the previous best-performing model, Ms-DANet, by 3.35%. The proposed approach effectively mitigates false or missed detections caused by density variations, occlusions, and viewpoint overlaps, establishing a new state-of-the-art in 3D change detection for complex urban environment.

Key words: 3D point cloud, change detection, geometric structure optimization, entropy attention, dual-path attention mechanism

中图分类号:

P237

朱涵, 戴晨光, 张振超, 刘宣广, 卢金浩. 基于几何结构优化和熵注意力机制的点云变化检测方法[J]. 测绘学报, 2026, 55(8): 1439-1451.

Han Zhu, Chenguang Dai, Zhenchao Zhang, Xuanguang Liu, Jinhao Lu. Point cloud change detection with geometric structure refinement and entropy-attention mechanism[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(8): 1439-1451.

导出引用管理器 EndNote|Reference Manager|ProCite|BibTeX|RefWorks

链接本文: http://xb.chinasmp.com/CN/10.11947/j.AGCS.2026.20250415

http://xb.chinasmp.com/CN/Y2026/V55/I8/1439

图/表 13

图1

图2

图3

图4

图5

表1

表2

图6

表3

表4

图7

图8

表5

参考文献 44

[1]	李朋飞, 李豆, 胡晋飞, 等. 机载LiDAR监测黄土高原土壤侵蚀的能力评估[J]. 测绘学报, 2023, 52(8): 1342-1354. DOI: . doi: 10.11947/j.AGCS.2023.20210698
	Li Pengfei, Li Dou, Hu Jinfei, et al. Assessing the ability of airborne LiDAR to monitor soil erosion on the Chinese Loess Plateau[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(8): 1342-1354. DOI: . doi: 10.11947/j.AGCS.2023.20210698
[2]	Khan S H, He Xuming, Porikli F, et al. Forest change detection in incomplete satellite images with deep neural networks[J]. IEEE Transactions on Geoscience and Remote Sensing, 2017, 55(9): 5407-5423.
[3]	杨安秀. 机载LiDAR测深点云滤波与分类方法研究[J]. 测绘学报, 2023, 52(7): 1234. DOI: . doi: 10.11947/j.AGCS.2023.20220015
	Yang Anxiu. Research on the point cloud filtering and classification for airborne LiDAR bathymetry[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(7): 1234. DOI: . doi: 10.11947/j.AGCS.2023.20220015
[4]	Wang Xiaochen, Ma Hongchao, Zhang Liang, et al. Point cloud classification and accuracy analysis based on feature fusion[J]. Journal of Geodesy and Geoinformation Science, 2021, 4(3): 38-48.
[5]	Chen Hao, Shi Zhenwei. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection[J]. Remote Sensing, 2020, 12(10): 1662.
[6]	Lu Miao, Chen Jun, Tang Huajun, et al. Land cover change detection by integrating object-based data blending model of Landsat and MODIS[J]. Remote Sensing of Environment, 2016, 184: 374-386.
[7]	Yang Juntao, Kang Zhizhong, Yang Zhou. A robust model fitting-based method for transmission line extraction from airborne LiDAR point cloud data[J]. Journal of Geodesy and Geoinformation Science, 2021, 4(3): 60-71.
[8]	杨必胜, 韩旭, 董震. 适用于城市场景大规模点云语义标识的深度学习网络[J]. 测绘学报, 2021, 50(8): 1059-1067. DOI: . doi: 10.11947/j.AGCS.2021.20210093
	Yang Bisheng, Han Xu, Dong Zhen. A deep learning network for semantic labeling of large-scale urban point clouds[J]. Acta Geodaetica et Cartographica Sinica, 2021, 50(8): 1059-1067. DOI: . doi: 10.11947/j.AGCS.2021.20210093
[9]	Yan L, Dai J, Tan J, et al. Global fine registration of point cloud in LiDAR SLAM based on pose graph[J]. Journal of Geodesy and Geoinformation Science, 2020, 3(2): 26-35.
[10]	Wang Zhixue, Zhang Yu, Luo Lin, et al. TransCD: scene change detection via transformer-based architecture[J]. Optics Express, 2021, 29(25): 41409.
[11]	Alcantarilla P F, Stent S, Ros G, et al. Street-view change detection with deconvolutional networks[J]. Autonomous Robots, 2018, 42(7): 1301-1322.
[12]	Li Bin, Zhao Xiaofa, Chen Yong, et al. Three dimensional laser point cloud slicing method for calculating irregular volume[J]. Journal of Geodesy and Geoinformation Science, 2019, 2(4): 31-43.
[13]	Menderes A, Erener A, Sarp G. Automatic detection of damaged buildings after earthquake hazard by using remote sensing and information technologies[J]. Procedia Earth and Planetary Science, 2015, 15: 257-262.
[14]	Zhu Zan, Gan Shu, Wang Jianqi, et al. A novel airborne 3D laser point cloud hole repair algorithm considering topographic features[J]. Journal of Geodesy and Geoinformation Science, 2020, 3(3): 29-38.
[15]	Xiao Wen, Cao Hui, Tang Miao, et al. 3D urban object change detection from aerial and terrestrial point clouds: a review[J]. International Journal of Applied Earth Observation and Geoinformation, 2023, 118: 103258.
[16]	刘宣广, 李玉洁, 张振超, 等. CVS-Net：基于时空关系建模与边缘信息强化的遥感影像语义变化检测方法[J]. 地球信息科学学报, 2025, 27(5): 1144-1162.
	Liu Xuanguang, Li Yujie, Zhang Zhenchao, et al. CVS-net: semantic change detection in remote sensing images using spatial-temporal modeling and edge enhancement[J]. Journal of Geo-information Science, 2025, 27(5): 1144-1162.
[17]	刘宣广, 李蒙蒙, 汪小钦, 等. 基于面向对象孪生神经网络的高分辨率遥感影像建筑物变化检测[J]. 遥感学报, 2024, 28(2): 437-454.
	Liu Xuanguang, Li Mengmeng, Wang Xiaoqin, et al. Use of object-based siamese neural network to build change detection from very high resolution remote-sensing images[J]. Journal of Remote Sensing, 2024, 28(2): 437-454.
[18]	Zhang L, Zhang Y, Chen Z, et al. Splitting and merging based multi-model fitting for point cloud segmentation[J]. Journal of Geodesy and Geoinformation Science, 2019, 2(2): 78-89.
[19]	Sun Yifan, Dai Chenguang, Li Wenke, et al. Supervised contrastive learning for indoor point cloud over segmentation[J]. IEEE Transactions on Multimedia, 2025, 27: 9189-9201.
[20]	Kharroubi A, Poux F, Ballouch Z, et al. Three dimensional change detection using point clouds: a review[J]. Geomatics, 2022, 2(4): 457-485.
[21]	Williams R D. DEMs of difference[J]. Geomorphological Techniques, 2012, 2(3): 1-17.
[22]	Bailey G, Li Yingkui, McKinney N, et al. Las2DoD: change detection based on digital elevation models derived from dense point clouds with spatially varied uncertainty[J]. Remote Sensing, 2022, 14(7): 1537.
[23]	Scott C P, Beckley M, Phan M, et al. Statewide USGS 3DEP LiDAR topographic differencing applied to Indiana, USA[J]. Remote Sensing, 2022, 14(4): 847.
[24]	Girardeau-Montaut D, Roux M, Raphaël M, et al. Change detection on points cloud data acquired with a ground laser scanner[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2005, 36: 30-35.
[25]	Lague D, Brodu N, Leroux J. Accurate 3D comparison of complex topography with terrestrial laser scanner: application to the Rangitikei canyon (N-Z)[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2013, 82: 10-26.
[26]	Stilla U, Xu Y. Change detection of urban objects using 3D point clouds: a review[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2023, 197: 228-255.
[27]	Wang Zhixue, Zhang Yu, Luo Lin, et al. An end-to-end point-based method and a new dataset for street-level point cloud change detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5703015.
[28]	Hu Qingyong, Yang Bo, Xie Linhai, et al. RandLA-net: efficient semantic segmentation of large-scale point clouds[C]//Proceedings of 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020: 11105-11114.
[29]	Han Ming, Sha Jianjun, Wang Yanheng, et al. PBFormer: point and bi-spatiotemporal transformer for pointwise change detection of 3D urban point clouds[J]. Remote Sensing, 2023, 15(9): 2314.
[30]	Zhan Wenxiao, Cheng Ruozhen, Chen Jing. PGN3DCD: prior-knowledge-guided network for urban 3D point cloud change detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5705415.
[31]	Lu Jinhao, Dai Chenguang, Zhang Zhenchao, et al. Ms-DANet: multiscale difference-aware network for 3D point cloud change detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 5702113.
[32]	Shen Zhao, Jia Xin, Zhang Jinglei. Rotation-invariant convolution with point sort and curvature radius for point cloud classification and segmentation[J]. IEEE Access, 2025, 13: 10432-10446.
[33]	Li Yaqian, Zhang Ze, Li Haibin, et al. DPPCN: density and position-based point convolution network for point cloud segmentation[J]. Pattern Analysis and Applications, 2025, 28(2): 69.
[34]	Jiang G, Yan W Y, Lichti D D. A maximum entropy-based optimal neighbor selection for multispectral airborne LiDAR point cloud classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5705018.
[35]	An Yi, Wang Jian, He Lijun, et al. Learning-based sampling method for point cloud segmentation[J]. IEEE Sensors Journal, 2024, 24(15): 24140-24151.
[36]	赵夫群, 李映萱. 基于点重要性判断的点云简化[J]. 计算机系统应用, 2023, 32(9): 197-202.
	Zhao Fuqun, Li Yingxuan. Point cloud simplification based on point importance judgment[J]. Computer Systems & Applications, 2023, 32(9): 197-202.
[37]	Jaderberg M, Simonyan K, Zisserman A, et al. Spatial transformer networks[C]//Proceedings of 2015 Advances in Neural Information Processing Systems. 2015: 2017-2025.
[38]	Hu J, Shen L, Albanie S, et al. Squeeze-and-excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(8): 2011-2023.
[39]	De Gélis I, Lefèvre S, Corpetti T. Change detection in urban point clouds: an experimental comparison with simulated 3D datasets[J]. Remote Sensing, 2021, 13(13): 2629.
[40]	Charles R Q, Hao S, Mo K C, et al. PointNet: deep learning on point sets for 3D classification and segmentation[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 77-85.
[41]	Kipf T, Welling M. Semi-supervised classification with graph convolutional networks[J]. Neural Processing Letters, 2021, 54: 2629-2644.
[42]	Qi C R, Yi L, Su H, et al. PointNet++: deep hierarchical feature learning on point sets in a metric space[C]//Proceedings of 2017 Advances in Neural Information Processing Systems. 2017: 5099-5108.
[43]	Ma X, Qin C, You H X, et al. Rethinking network design and local geometry in point cloud: a simple residual MLP framework[C]//Proceedings of 2022 International Conference on Learning Representations. 2022.
[44]	Tran T H G, Ressl C, Pfeifer N. Integrated change detection and classification in urban areas based on airborne laser scanning point clouds[J]. Sensors, 2018, 18(2): 448.

模型	IoU			OA	mIoU
模型	移除	增加	不变	OA	mIoU
PointNet^[40]*	23.47	26.84	89.36	89.16	46.56
GCN^[41]*	28.23	35.85	88.61	89.80	50.90
PointNet++^[42]*	45.36	57.21	95.31	96.03	65.96
PointMLP^[43]*	46.83	40.11	90.35	92.56	59.10
3DCDNet^[27-28]	52.82	58.23	93.87	93.51	68.31
PGN3DCD^[30]	74.58	75.25	96.67	96.97	82.17
Ms-DANet^[31]	73.54	75.50	96.88	97.14	81.97
GS-EANet	79.84	78.14	97.98	97.56	85.32

模型	FLOPs/G	Params/M	推理速度/（帧/s）	mIoU/（%）
3DCDNet^[27-28]	88.20	6.68	53.29	68.31
Ms-DANet^[31]	106.31	6.37	44.73	81.97
GS-EANet	93.41	6.45	48.80	85.32

网络	LCA	SSM	DPA	EG-DPA	IoU			OA	mIoU
网络	LCA	SSM	DPA	EG-DPA	移除	增加	不变	OA	mIoU
3DCDNet^[27-28]	×	×	×	×	52.82	58.23	93.87	93.51	68.31
GS-EANet-1	√	×	×	×	55.98	60.21	94.30	94.03	70.16
GS-EANet-2	×	×	√	×	58.69	62.06	93.99	93.74	71.58
GS-EANet-3	×	×	×	√	63.96	64.87	95.53	94.79	74.79
GS-EANet-4	√	√	×	×	60.32	62.98	95.02	95.04	72.77
GS-EANet-5	√	√	√	×	74.08	75.14	97.11	96.82	82.11
GS-EANet	√	√	×	√	79.84	78.14	97.98	97.56	85.32

数据集	网络	IoU			OA	mIoU
数据集	网络	移除	增加	不变	OA	mIoU
子数据集1-b ALS Low Res	3DCDNet^[27-28]	37.62	45.84	98.63	96.21	60.70
	Ms-DANet^[31]	53.10	62.45	98.55	98.57	71.37
	GS-EANet	56.12	65.09	98.64	98.37	73.28
子数据集1-a ALS Low Res	3DCDNet^[27-28]	48.31	62.17	98.03	96.58	69.50
	Ms-DANet^[31]	63.86	78.67	99.04	99.06	80.52
	GS-EANet	66.21	78.22	99.65	98.89	81.36
子数据集1-c ALS Low Res	3DCDNet^[27-28]	55.03	62.31	99.08	98.14	72.14
	Ms-DANet^[31]	66.85	83.49	99.20	99.22	83.18
	GS-EANet	70.66	86.21	99.05	98.98	85.31
子数据集2 ALS High Res	3DCDNet^[27-28]	55.63	70.59	99.03	98.67	75.08
	Ms-DANet^[31]	67.74	76.59	99.06	99.08	81.13
	GS-EANet	70.36	80.21	99.23	99.21	83.27
子数据集3 ALS High Noise	3DCDNet^[27-28]	40.31	58.69	99.02	98.32	65.80
	Ms-DANet^[31]	55.27	72.15	98.92	98.93	75.45
	GS-EANet	65.04	69.89	99.23	99.05	78.05
子数据集4 Photogrammetry	3DCDNet^[27-28]	51.23	65.21	98.06	98.25	71.74
	Ms-DANet^[31]	61.42	76.95	99.15	99.17	79.17
	GS-EANet	67.32	78.21	99.06	99.12	81.53
子数据集5 Multi-Sensor	3DCDNet^[27-28]	45.63	62.58	98.12	98.62	68.78
	Ms-DANet^[31]	54.84	71.96	98.68	98.71	75.16
	GS-EANet	59.98	76.36	99.06	99.12	78.47

数据集	Ms-DANet的mIoU/（%）	GS-EANet		跨区域迁移试验后的mIoU/（%）	跨区域迁移精度变化情况/（%）
数据集	Ms-DANet的mIoU/（%）	最优权重系数组合（ω_C，ω_H，ω_ρ）	GS-EANet最优mIoU/（%）	跨区域迁移试验后的mIoU/（%）	跨区域迁移精度变化情况/（%）
Urb3DCD 1-a	71.37	（0.4，0.4，0.2）	73.28	73.06	↓0.22
Urb3DCD 1-b	80.52	（0.4，0.4，0.2）	81.36	81.22	↓0.14
Urb3DCD 1-c	83.18	（0.4，0.4，0.2）	85.31	85.08	↓0.23
Urb3DCD 2	81.13	（0.3，0.5，0.2）	83.27	82.89	↓0.38
Urb3DCD 3	75.45	（0.2，0.5，0.3）	78.05	77.53	↓0.52
Urb3DCD 4	79.17	（0.3，0.3，0.4）	81.53	81.39	↓0.14
Urb3DCD 5	75.16	（0.3，0.2，0.5）	78.47	78.15	↓0.33