A hybrid intelligent driven three-dimensional point cloud change detection method

Abstract

This invention discloses a hybrid intelligent-driven 3D point cloud change detection method, comprising the following steps: dual-temporal point cloud input and preprocessing; establishing intertemporal correspondence between dual-temporal point clouds using nearest neighbor search; multi-scale normal distance geometric prior calculation; texture difference prior calculation; automatic thresholding and normalization processing of the above calculation results; density consistency suppression and prior score construction; construction of a dual-branch encoder-decoder network and prior knowledge injection; change-aware differential attention coupling and acquisition of change-aware enhanced features; decoding output of change detection results. This invention significantly enhances the focus on and representation of truly changed regions without altering the backbone network structure, effectively reducing false detections caused by density differences, occlusion, and noise. It achieves high-precision detection of truly changed regions in dual-temporal urban 3D point clouds and improves the adaptability and robustness of the method under complex occlusion, density inconsistency, and multi-source data conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent interpretation and change detection technology of 3D point clouds, specifically involving a hybrid intelligent-driven method for 3D point cloud change detection. Background Technology

[0002] With the rapid development of spatial information acquisition technologies such as airborne lidar, ground-based laser scanning, UAV photogrammetry, and multi-view 3D reconstruction, acquiring multi-temporal urban 3D point cloud data has become a reality. Dual-temporal or multi-temporal urban point cloud data can intuitively reflect changes such as building additions, demolitions, facade renovations, surface facility updates, and urban environmental evolution. Therefore, it has significant application value in fields such as digital twin cities, urban renewal, disaster monitoring, illegal construction inspections, infrastructure management, and land space governance. The core task of 3D point cloud change detection is to accurately identify areas that have actually changed from point cloud data acquired at different times, and to suppress as many spurious changes as possible caused by noise, occlusion, sampling differences, density inconsistencies, and registration errors. Compared with 2D images, 3D point clouds have characteristics such as irregular distribution, uneven local density, complex sampling perspectives, severe occlusion, and unstable cross-temporal correspondences, making 3D change detection far more complex than 2D change detection.

[0003] Existing methods for detecting changes in 3D point clouds can be broadly categorized into geometric distance-based methods, feature difference-based methods, and deep learning-based methods. Traditional geometric methods typically utilize Euclidean distance, point-to-surface distance, or local normal differences to measure the degree of change, offering some interpretability. However, they are often sensitive to sparse sampling, local occlusion, missing boundaries, and inconsistent density, easily misclassifying non-realistic changes as areas of change. While feature difference-based methods can enhance discriminative capabilities to some extent by utilizing color, intensity, or local structure descriptors, the complex textures and varying material reflectivity in urban scenes mean that relying solely on appearance differences is susceptible to interference from lighting, viewing angle, and sensor variations.

[0004] In recent years, deep learning methods have shown great potential in 3D change detection by learning high-dimensional feature representations between two-temporal point clouds. Typical methods often employ Siamese coding networks, differential modules, attention mechanisms, or feature fusion strategies to jointly model two-temporal point clouds. However, purely data-driven deep networks still have several shortcomings: on the one hand, they lack explicit change cues in the shallow stages, and the network often needs to learn change patterns from complex and redundant raw point features, limiting training efficiency and generalization ability; on the other hand, deep networks are usually sensitive to changes in point cloud density, local missing features, and occlusion, and are prone to misidentifying differences caused by inconsistent sampling as real changes, thus leading to an increased false detection rate.

[0005] In urban scenarios, genuine changes typically manifest as both local geometric structural changes and appearance attribute changes, while pseudo-changes are often accompanied by inconsistent sampling density, unstable neighborhood distribution, and local occlusion. Therefore, if prior knowledge for 3D change detection tasks can be explicitly introduced into the deep network input or shallow feature stage—such as knowledge of geometric changes calculated along surface normals at multiple scales, knowledge of texture changes between corresponding points in two time phases, and confidence knowledge for adaptively suppressing density inconsistencies—it is hoped that the network's ability to focus on genuinely changed regions can be enhanced, and the interference of irrelevant backgrounds and pseudo-changes on differential feature learning can be reduced.

[0006] However, most existing technologies process prior knowledge separately from deep networks, or only use prior knowledge as a post-processing condition. There is a lack of an end-to-end hybrid intelligent framework that unifies and couples "geometric prior-texture prior-density consistency suppression" with a bi-branch deep network and a change-aware differential attention module. Especially in complex urban scenarios, how to establish intertemporal correspondences, construct robust point-level prior change scores, and seamlessly inject them into the input layer, shallow encoding layer, and attention module of a deep network to achieve accurate perception of truly changing areas and effective suppression of pseudo-changes remains a pressing technical problem to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to provide a hybrid intelligent-driven method for detecting changes in 3D point clouds, which can achieve high-precision detection of real change areas in dual-temporal urban 3D point clouds, and improve the adaptability and robustness of the method under complex occlusion, density inconsistency and multi-source data conditions.

[0008] To achieve the above objectives, this invention provides a hybrid intelligent-driven method for detecting changes in 3D point clouds, comprising the following steps: S1. Dual-temporal point cloud input and preprocessing; S2. Nearest neighbor search is used to establish the intertemporal correspondence between point clouds of two time phases; S3, Multi-scale normal distance geometric prior calculation; S4, prior calculation of texture differences; S5. Perform automatic thresholding and normalization on the calculation results of S3 and S4. S6. Density Consistency Suppression and Prior Score Construction; S7. Construct a dual-branch encoder-decoder network and inject prior knowledge; S8. Change-aware differential attention coupling, and acquisition of change-aware enhanced features; S9, Decoding output change detection result.

[0009] As a further aspect of the present invention: Step S1 specifically includes: acquiring first point clouds of the same region at different time phases. Second point cloud The registration, normalization, and sampling were completed to obtain the corresponding set of sampling points. and .

[0010] As a further aspect of the present invention: Step S2 specifically includes: employing K Nearest neighbor search obtains the sample point set respectively s 1 in s 1 point cloud PC Nearest neighbor index set in 2 K 12 and sampling point set s 2 in point cloud PC Nearest neighbor index set in 1 K 21 To establish the intertemporal correspondence between two-temporal point clouds.

[0011] As a further aspect of the present invention: Step S3 specifically includes: ... PC Points in 2 p 2 and its in PC Corresponding points in 1 p 1 * In the radius set At multiple scales, local unit normals are estimated using principal component analysis. n 2 (r) and n 1 (r) ; at scale r The bidirectional projected distance along the normal direction is defined as: ; The geometric difference of the bidirectional normal distance is obtained by averaging the two-way normal distance and taking the maximum response at multiple scales: ; in, Point p 2. Maximum significant displacement response along the local surface normal under multi-scale conditions. Indicates all candidate scales r Below, point p The local surface normal projection displacement response of 2 is taken as the maximum value.

[0012] As a further aspect of the present invention: the texture difference in S4 is defined as follows: ; in, and Representing points respectively p 2 and its corresponding point p 1 *The color, texture, material, or intensity characteristics.

[0013] As a further aspect of the present invention: the threshold in S5 includes a geometric threshold. With texture threshold They are defined as follows: ; in, Geometric difference over all matching point pairs The mean, For texture differences The mean, The theoretical maximum 2 norm for normalized RGB three-channel differences; Normalize geometric and textural differences to : .

[0014] As a further aspect of the present invention: Step S6 specifically includes: constructing a local... The confidence level for density consistency in nearest neighbor statistics is defined as follows: ; in, and They represent the local areas of the corresponding points. The median of nearest neighbor distance, and Indicates the corresponding quantile value , To prevent tiny constants with a denominator of zero, conf is truncated to... ; Gated fusion of geometric and textural differences using confidence scores yields point-level prior change scores. ; Map the prior change score to a prior change mask: After thresholding, channel expansion, or splicing, the Mask constitutes the Prior Change feature.

[0015] As a further aspect of the present invention: Step S7 specifically includes: constructing a dual-branch encoder-decoder network with shared weights, and... and The data is input into two encoding branches respectively; at the same time, the Mask or Prior Change is concatenated with the original attributes including point cloud coordinates, color, and intensity, or the feature map is reweighted point by point in the shallow encoding feature stage, thereby completing the injection of prior knowledge at the input level and the shallow feature level.

[0016] As a further aspect of the present invention: Step S8 specifically includes: setting a change-aware differential attention module in one or more layers of the encoder, and the joint feature formed by the first temporal point cloud feature and the second temporal point cloud feature after corresponding association is represented as a dual-temporal point feature. Features of dual time phase points Perform linear projection Furthermore, a priori change mask is used to modulate the attention map element-wise to obtain change-aware enhanced features: ; in, CSA(⋅) represents element-wise multiplication; L(⋅) represents the output of the change-aware differential attention module; SoftMax(⋅) represents the normalization function, used to convert the modulated features into attention weights.

[0017] As a further aspect of the present invention, step S9 specifically includes: inputting the multi-level features enhanced by prior knowledge and reweighted by differential attention into the decoder, restoring the point-level spatial resolution through layer-by-layer upsampling and skip connections, and outputting the change labels, change probability maps, or change region segmentation results of the first temporal point cloud and / or the second temporal point cloud.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a geometry-texture-density joint prior knowledge system for 3D point cloud change detection tasks. The geometry prior uses multi-scale normal distance to represent the true geometric changes along the local surface normal, which is more suitable for sparse sampling, incomplete boundary and local occlusion scenarios than simple Euclidean distance. The texture prior can effectively supplement the appearance difference information of two phases. The density consistency suppression mechanism can adaptively suppress the pseudo changes caused by inconsistent sampling density, neighborhood distribution differences and local missing data, thereby improving the stability and interpretability of detection at the change prior level.

[0019] 2. This invention does not use prior knowledge as an independent post-processing condition. Instead, it seamlessly injects prior change scores, prior change masks, and prior change features into a dual-branch encoder-decoder deep network, achieving unified coupling at the input level, shallow feature level, and attention level. This forms a truly hybrid intelligent end-to-end change detection framework that retains the physical interpretability of explicit priors while leveraging the adaptive learning capabilities of deep networks for complex high-dimensional features.

[0020] 3. This invention introduces a change-aware differential attention module in one or more layers of the encoder. By modulating the attention map, differential response, or similarity mapping through a prior change mask, the network prioritizes potential change regions and suppresses interference from unchanged backgrounds during the feature extraction stage. This effectively alleviates the problems of low proportion of change regions, large background redundancy, and strong error propagation in complex urban scenes.

[0021] 4. Without changing the overall structure of the backbone network, this invention can significantly enhance the network's attention to and representation of real-world changing regions. It features modularity, lightweight design, and easy reusability, and can be easily embedded into various backbone networks and training processes, demonstrating significant engineering value.

[0022] 5. This invention has strong adaptability and robustness to urban 3D point cloud data from different sources, at different scales, with different sampling densities, and under different occlusion conditions. It is applicable to airborne laser point clouds, ground laser point clouds, UAV photogrammetric point clouds, multi-view reconstructed point clouds, and their combined data. In complex urban scenarios, it can significantly reduce the false detection rate and improve the accuracy and generalization ability of change detection results.

[0023] 6. The output of this invention includes not only point-level change labels and change probability maps, but also prior change masks, change confidence levels, and statistical results of change regions, which are beneficial for subsequent change interpretation, result quality control, and application decisions. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention; Figure 2 This is a rendering of the invention applied to a city-level 3D point cloud change detection dataset. Detailed Implementation

[0025] The present invention will be further illustrated by the following examples.

[0026] like Figure 1 As shown, a hybrid intelligent-driven method for detecting changes in 3D point clouds includes the following steps: S1. Dual-temporal point cloud input and preprocessing; Specifically, acquiring the first point cloud of the same region at different time phases. Second point cloud The first and second point clouds can be derived from airborne laser point clouds, ground-based laser point clouds, UAV photogrammetric point clouds, multi-view reconstructed point clouds, or combinations thereof. The acquired dual-temporal point clouds undergo coordinate system optimization, coarse registration, fine registration, outlier removal, scale normalization, and point sampling to obtain a set of sampled points for subsequent calculations. and .

[0027] Preferably, the sampling method can be random sampling, voxel downsampling, farthest point sampling, or a combination thereof; preprocessing may also include color normalization, intensity normalization, and initial normal estimation to improve the stability of cross-period feature matching.

[0028] S2. Nearest neighbor search is used to establish the intertemporal correspondence between point clouds of two time phases; Specifically, adopt Nearest neighbor search obtains the sample point set respectively In point cloud Nearest neighbor index set in and sampling point set In point cloud Nearest neighbor index set in To establish the intertemporal correspondence between two-temporal point clouds.

[0029] Preferably, the nearest neighbor in another time phase can be used as the candidate corresponding point; alternatively, it can be based on... Spatial consistency or bidirectional consistency of nearest neighbor sets further filters for more stable corresponding point pairs.

[0030] This step provides a cross-period point-pair basis for constructing confidence scores for geometric differences, texture differences, and density consistency. Compared to directly calculating differences across the global point cloud, this step limits the calculation of prior knowledge to a locally reasonable matching range, thereby reducing error propagation caused by large-scale non-correspondence relationships.

[0031] S3, Multi-scale normal distance geometric prior calculation; Specifically, for PC Points in 2 p 2 and its in PC Corresponding points in 1 p 1 * In the radius set At multiple scales, local unit normals are estimated using principal component analysis. n 2 (r) and n 1 (r) ; at scale r The bidirectional projected distance along the normal direction is defined as: ; The geometric difference of the bidirectional normal distance is obtained by averaging the two-way normal distance and taking the maximum response at multiple scales: ; in, Point p 2. Maximum significant displacement response along the local surface normal under multi-scale conditions. Indicates all candidate scales r Below, pointp The local surface normal projection displacement response of 2 is taken as the maximum value. Using this value instead of the simple point-to-point Euclidean distance can reduce misjudgments under different sampling densities, local missing and occlusion conditions, and enhance the ability of geometric change priors to express the real structural changes.

[0032] Preferred, radius set Multiple neighborhood scales, ranging from small to large, can be set according to the scene scale; for building boundaries, roof edges, or facade transition areas, the multi-scale normal distance can simultaneously maintain local sensitivity and global robustness.

[0033] S4, prior calculation of texture differences; Specifically, texture difference is defined as: ; in, and Representing points respectively p 2 and its corresponding point p 1 * The color, texture, material, or intensity characteristics.

[0034] Preferably, the color feature can be RGB, HSV, or grayscale; the intensity feature can be laser reflection intensity; and the material feature can be a high-dimensional descriptor obtained through an external encoder or point cloud attribute statistics. To reduce the influence of local noise, the average or weighted average of the texture differences between nearest neighbors across periods can also be taken to improve the smoothness and stability of the texture prior.

[0035] S5. Perform automatic thresholding and normalization on the calculation results of S3 and S4. Considering the differences in dual-temporal point clouds across various scenarios, sensors, and sampling conditions, a fixed threshold is insufficient to adapt to diverse urban environments simultaneously. Therefore, this invention adaptively determines the geometric and texture thresholds using the statistical mean of all matched point pairs.

[0036] Specifically, thresholds include geometric thresholds. With texture threshold They are defined as follows: ; in, Geometric difference over all matching point pairs The mean, For texture differences The mean, The theoretical maximum 2 norm for normalized RGB three-channel differences; Normalize geometric and textural differences to : .

[0037] Through the above processing, geometric and texture responses from different sources and with different dimensions can be mapped to a unified numerical range, providing a foundation for subsequent gating fusion.

[0038] S6. Density Consistency Suppression and Prior Score Construction; Specifically, to suppress spurious changes caused by inconsistent sampling density, occlusion, and local sparsity, a local-based... The confidence level for density consistency in nearest neighbor statistics is defined as follows: ; in, and They represent the local areas of the corresponding points. The median of nearest neighbor distance, and Indicates the corresponding quantile value , To prevent tiny constants with a denominator of zero, specifically, The value ranges from 0.5 to 5, preferably from 1 to 2; The value ranges from 0.5 to 5, preferably from 1 to 2; The range of values ​​is Preferred and truncate conf to ; Gated fusion of geometric and textural differences using confidence scores yields point-level prior change scores. ; Map the prior change score to a prior change mask: Mask can be used directly as a point-level mask, or it can be thresholded, channel expanded, linearly transformed, or concatenated with the original point attributes to form a prior change feature.

[0039] The above processing enables the geometric differences, texture differences, and density consistency confidence to form a unified explicit knowledge representation.

[0040] S7. Construct a dual-branch encoder-decoder network and inject prior knowledge; Specifically, a dual-branch encoder-decoder network with shared weights is constructed, and... and The inputs are respectively fed into two encoding branches; the dual-branch encoder-decoder network can adopt a twin structure, with parameters shared between branches to ensure that the features of both time phases are in a unified feature space. In this invention, prior knowledge is not used as an independent post-processing condition, but is directly and seamlessly injected into the deep network. Specific methods include: 1) Concatenate Prior Change with the original attributes such as point cloud coordinates, color, and intensity as the network input; 2) In the shallow encoding stage, Mask is used to reweight the point feature map point by point; 3) Embed the Prior Change as an additional channel in the encoder feature map; 3) Use prior change masks to modulate the similarity or difference mapping between bi-temporal features.

[0041] In this way, the network can obtain explicit change cues at the input level and shallow feature level, avoiding blindly searching for change patterns from raw point features by relying entirely on data-driven methods.

[0042] S8. Change-aware differential attention coupling, and acquisition of change-aware enhanced features; Specifically, a change-aware differential attention module is set in one or more layers of the encoder. The joint feature formed by the corresponding association between the first-phase point cloud features and the second-phase point cloud features is represented as a dual-phase point feature. Features of dual time phase points Perform linear projection Furthermore, a priori change mask is used to modulate the attention map element-wise to obtain change-aware enhanced features: ; in, The expression represents element-wise multiplication, and the Mask is broadcast across the channel dimension. CSA(⋅) represents the output of the change-aware differential attention module; L(⋅) represents the linear projection function; and SoftMax(⋅) represents the normalization function, used to convert the modulated features into attention weights. The role of the change-aware differential attention module is to enable the network to prioritize potentially changing regions during the encoding stage, assigning higher weights to changing regions and lower weights to unchanged backgrounds and pseudo-changing regions, thereby improving the effectiveness of differential feature extraction.

[0043] Preferably, the change-aware differential attention module can be set at one level of the encoder or at multiple levels to jointly model shallow local detail changes and deep global structural changes. The change-aware differential attention module can be combined with a self-attention structure or work in conjunction with a differential module, residual module, or feature fusion module.

[0044] S9, Decoding output change detection result; Specifically, the multi-level features enhanced by prior knowledge and reweighted by differential attention are input into the decoder. The point-level spatial resolution is restored through layer-by-layer upsampling and skip connections, and the change labels, change probability maps or change region segmentation results of the first phase point cloud and / or the second phase point cloud are output.

[0045] The output also includes a point-level change probability map, a prior change mask, change confidence assessment, and statistical results of the change region. The change confidence is jointly characterized by the prior change score, network output probability, and their consistency. The false detection risk in local areas is assessed using density consistency confidence and residual statistics. The statistical estimates of the fluctuation range of the false detection rate, residual statistics, or detection performance indicators are termed uncertainty, which is estimated using Monte Carlo sampling, bootstrap methods, or multiple random subsampling methods to quantify the model's robustness in complex urban scenes, occluded areas, and inconsistent density conditions.

[0046] Figure 2 As shown in Table 1, the effect diagrams and statistical indicators of the method of the present invention on a city-level 3D point cloud change detection dataset are presented respectively.

[0047] Table 1: Statistical Indicators of a City-Level 3D Point Cloud Change Detection Dataset

[0048] Combination Figure 2 As shown in Table 1, the method of this invention can extract the main change regions relatively completely, and the prediction results are consistent with the true labels, with relatively few false detections in the background region. The overall accuracy (OA), recall (Recall), and mean intersection-union (mIoU) of the method of this invention are 96.95%, 93.96%, and 79.04%, respectively, all of which are superior to the comparative methods such as Siamese kernel convolution (SKP), encoder fusion Siamese kernel convolution (Encoder Fusion SKP), and knowledge-guided change detection (PGN3DCD). Compared with the basic SKP method, OA, recall, and mIoU are improved by 5.17, 8.32, and 15.32 percentage points, respectively; compared with the PGN3DCD method, OA and mIoU are improved by 0.71 and 1.90 percentage points, respectively.

[0049] From the perspective of Intersection over Union (IoU), the method of this invention achieves an IoU of 96.01% for unchanged categories and 72.17% for newly added categories, both showing good performance. The IoU for removed categories is 62.75%, slightly lower than PGN3DCD, but significantly higher than SKP and Encoder Fusion SKP. These results demonstrate that this invention can effectively suppress false background detections and has a strong ability to identify newly added changed targets.

[0050] The method of this invention performs stably on real urban point cloud data, indicating that geometric, texture and density priors can provide effective guidance for the network on regions of change and reduce the impact of spurious changes caused by uneven point cloud sampling, occlusion and reconstruction errors.

[0051] In this invention, prior knowledge allows for adaptive construction based on sampling density, occlusion degree, sensor type, and appearance differences of urban point clouds in different time periods. Specifically, the geometric prior uses multi-scale normal distance to represent local structural changes, the texture prior originates from the differences in color, grayscale, material, or intensity between corresponding points in the two time periods, and the density consistency suppression term is used to adaptively suppress spurious changes caused by inconsistent sampling density, local sparsity, and occlusion. The intertemporal correspondence is established through K-nearest neighbor search between point clouds in the two time periods, and the generated point-level prior change score serves as explicit guiding information for subsequent deep networks.

[0052] The prior knowledge computation framework is centered on intertemporal correspondence. After obtaining the corresponding point pairs in two time periods, it calculates geometric difference, texture difference, and density consistency confidence. Geometric difference is obtained by estimating the local surface normals at multiple neighborhood scales and calculating the projected distance between intertemporal point pairs along the normal direction. Texture difference is obtained by the feature difference between corresponding points. Density consistency confidence is constructed based on local K-nearest neighbor distance statistics, median distance, and quantile distance. Subsequently, geometric difference and texture difference are normalized and gated weighted fusion is performed using density consistency confidence to generate point-level prior change scores. When the prior change scores meet the preset mapping method, a prior change mask or prior change feature is generated and injected into the deep network to participate in end-to-end training and inference.

[0053] In practical implementation, this invention allows for parameterized configuration of urban 3D point cloud data from various sources. Urban 3D point cloud data includes airborne laser point clouds, ground-based laser point clouds, UAV photogrammetric point clouds, multi-view reconstructed point clouds, and combinations thereof. For different data sources and application scenarios, the K-nearest neighbor parameter, neighborhood radius set, automatic thresholding strategy, density consistency suppression parameter, prior injection level, and network training settings can be adjusted according to scene scale, point cloud density, occlusion level, and task requirements. Therefore, while maintaining the basic structure of the backbone network, the applicability of this invention in different backbone networks, different training strategies, and different urban scene point cloud data can be improved.

Claims

1. A hybrid intelligent-driven method for detecting changes in three-dimensional point clouds, characterized in that, Includes the following steps: S1. Dual-temporal point cloud input and preprocessing; S2. Nearest neighbor search is used to establish the intertemporal correspondence between point clouds of two time phases; S3, Multi-scale normal distance geometric prior calculation; S4, prior calculation of texture differences; S5. Perform automatic thresholding and normalization on the calculation results of S3 and S4. S6. Density Consistency Suppression and Prior Score Construction; S7. Construct a dual-branch encoder-decoder network and inject prior knowledge; S8. Change-aware differential attention coupling, and acquisition of change-aware enhanced features; S9, Decoding output change detection result.

2. The hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 1, characterized in that, The specific steps of S1 include: acquiring the first point cloud of the same region at different time phases. Second point cloud The registration, normalization, and sampling were completed to obtain the corresponding set of sampling points. and .

3. The hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 2, characterized in that, The specific steps of S2 include: using... K Nearest neighbor search obtains the sample point set respectively s 1 in s 1 point cloud PC Nearest neighbor index set in 2 K 12 and sampling point set s 2 in point cloud PC Nearest neighbor index set in 1 K 21 To establish the intertemporal correspondence between two-temporal point clouds.

4. The hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 3, characterized in that, The specific steps in S3 include: […]. PC Points in 2 p 2 and its in PC Corresponding points in 1 p 1 * In the radius set At multiple scales, local unit normals are estimated using principal component analysis. n 2 (r) and n 1 (r) ; at scale r The bidirectional projected distance along the normal direction is defined as: ； The geometric difference of the bidirectional normal distance is obtained by averaging the two-way normal distance and taking the maximum response at multiple scales: ； in, Point p 2. Maximum significant displacement response along the local surface normal under multi-scale conditions. Indicates all candidate scales r Below, point p The local surface normal projection displacement response of 2 is taken as the maximum value.

5. The hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 4, characterized in that, In S4, texture differences are defined as follows: ； in, and Representing points respectively p 2 and its corresponding point p 1 * The color, texture, material, or intensity characteristics.

6. The hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 5, characterized in that, Thresholds in S5 include geometric thresholds. With texture threshold They are defined as follows: ； in, Geometric difference over all matching point pairs The mean, For texture differences The mean, The theoretical maximum 2 norm for normalized RGB three-channel differences; Normalize geometric and textural differences to : 。 7. The hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 6, characterized in that, The specific steps of S6 include: building a local... The confidence level for density consistency in nearest neighbor statistics is defined as follows: ； in, and They represent the local areas of the corresponding points respectively. The median of nearest neighbor distances. and Indicates the corresponding quantile value , To prevent tiny constants with a denominator of zero, conf is truncated to... ; Gated fusion of geometric and textural differences using confidence scores yields point-level prior change scores. ； Map the prior change score to a prior change mask: After thresholding, channel expansion, or splicing, the Mask constitutes the Prior Change feature.

8. The hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 7, characterized in that, The specific steps of S7 include: constructing a dual-branch encoder-decoder network with shared weights, and... and The data is input into two encoding branches respectively; at the same time, the Mask or Prior Change is concatenated with the original attributes including point cloud coordinates, color, and intensity, or the feature map is reweighted point by point in the shallow encoding feature stage, thereby completing the injection of prior knowledge at the input level and the shallow feature level.

9. A hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 8, characterized in that, The specific steps of S8 include: setting up change-aware differential attention modules in one or more layers of the encoder; and representing the joint features formed by the corresponding association between the first-phase point cloud features and the second-phase point cloud features as dual-phase point features. Features of dual time phase points Perform linear projection Furthermore, a priori change mask is used to modulate the attention map element-wise to obtain change-aware enhanced features: ； in, CSA(⋅) represents element-wise multiplication; L(⋅) represents the output of the change-aware differential attention module; SoftMax(⋅) represents the normalization function, used to convert the modulated features into attention weights.

10. A hybrid intelligent-driven three-dimensional point cloud change detection method according to claim 9, characterized in that, The specific steps of S9 include: inputting the multi-level features enhanced by prior knowledge and reweighted by differential attention into the decoder, restoring the point-level spatial resolution through layer-by-layer upsampling and skip connections, and outputting the change labels, change probability maps or change region segmentation results of the first phase point cloud and / or the second phase point cloud.

Images