A 3D point cloud change detection method based on Siamese AdaptConv

Abstract

The application provides a three-dimensional point cloud change detection method based on Siamese AdaptConv, and belongs to the field of intelligent scene reconstruction; firstly, three-dimensional point cloud data of a to-be-detected scene at two different time points is acquired, and an adaptive neighborhood is constructed for each point; then, the pretreated three-dimensional point cloud data is input into a Siamese AdaptConv double-branch network, multi-scale features of each layer are fused respectively, corresponding fusion features are obtained, the change probability of the same position point is further calculated, and a candidate change point set is mapped; DBSCAN clustering is performed, and a candidate region is output; finally, semantic verification is performed on the candidate region, a geographical knowledge graph is pre-constructed, and semantic correlation is calculated; when the semantic correlation exceeds a set threshold value, it is considered that the three-dimensional point cloud change of the candidate region is consistent with the existing entity semantics in the knowledge graph, and the reasonable change is determined; and the application realizes end-to-end multi-sequential feature alignment and change identification.

Description

Technical Field

[0001] This invention belongs to the field of 3D point cloud change detection and intelligent scene reconstruction, specifically involving a 3D point cloud change detection method based on SiameseAdaptConv (a Siamese network architecture that combines the adaptive convolution AdaptConv mechanism). Background Technology

[0002] 3D point cloud change detection is an important component of 3D scene reconstruction and intelligent extraction of aerospace information. It belongs to the interdisciplinary research fields of computer vision, remote sensing and mapping, and geographic information systems. Its main task is to compare point cloud data acquired at different times and identify structural or semantic changes that occur in the scene. It is widely used in scenarios such as urban renewal monitoring, dynamic updates of digital twins, and infrastructure operation and maintenance.

[0003] Existing 3D change detection methods mostly rely on DSM or projected rasterization, which easily leads to a loss of spatial accuracy. Traditional point cloud geometric feature methods, such as C2C and M3C2, are computationally complex and sensitive to noise. Existing deep learning methods suffer from performance degradation in multi-temporal registration accuracy (errors are typically greater than 0.05m), point density difference adaptability (when density differences exceed 2 times), and semantic consistency. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a 3D point cloud change detection method based on Siamese AdaptConv, achieving end-to-end multi-temporal feature alignment and change recognition.

[0005] The specific steps are as follows:

[0006] Step 1: For the scene to be tested, use LiDAR and UAV to move and measure, respectively, to obtain 3D point cloud data at two different time points, and perform preprocessing;

[0007] First-time point cloud dataset Point cloud dataset at the second time step .

[0008] in, , = These represent the three-dimensional coordinate vectors of a single point; , These represent the number of points in the cloud at two different time points.

[0009] Preprocessing refers to performing the following on the point cloud at each time step: coordinate normalization, outlier removal (statistical filtering), downsampling (FPS), and coarse and fine registration (FPFH + RANSAC → ICP); the registration goal is to... and By placing them in the same world coordinate system, local nearest neighbor queries can be compared.

[0010] Step 2: Construct an adaptive neighborhood for each preprocessed point that reflects its local geometric characteristics;

[0011] The adaptive neighborhood construction employs a smooth-transition dynamic neighborhood construction strategy, targeting the center point. The set of points in the local neighborhood The calculation is as follows:

[0012] ;

[0013] In a point cloud, besides the center point Any point other than; Point and points The Euclidean distance between them; It is an adaptive radius that is not fixed, depending on the point. The density of local points in the surrounding area is dynamically calculated. In the point cloud, the distance to the point is... The set of the most recent 50 points.

[0014] Adaptive radius Adjust based on local density smoothing:

[0015] ;

[0016] Where β=0.2 is the maximum adjustment amplitude coefficient; γ=2.0 is the sensitivity of the control adjustment; For global reference density; For point Local point cloud density; The reference neighborhood radius is denoted as .

[0017] Step 3: Transfer the preprocessed 3D point cloud data and Input the Siamese AdaptConv dual-branch network and fuse the multi-scale features of each layer to obtain their respective fused features;

[0018] The specific steps are as follows:

[0019] Step 301: Construct a twin dual-branch feature extraction architecture for the reference branch and the variant branch;

[0020] The reference branch and the variant branch are structurally symmetrical; a weight-sharing mechanism is employed, with both branches sharing the same set of learnable parameters during training and inference. .

[0021] Step 302, process the preprocessed 3D point cloud data and The inputs are fed into the reference branch and the variation branch, respectively, serving as the input layer (i.e., the first branch) of the Siamese AdaptConv two-branch network. =0th floor);

[0022] The initial feature vector of this layer is defined as Its value is determined by the enhanced geometric feature vectors corresponding to the 3D point cloud data at two time points. constitute:

[0023] ;

[0024] Described by linearity; Describing flatness; Describing the scattering degree; It is anisotropic; For feature entropy; This represents the change in surface curvature; It is the average angle between the normal vector of each point in the neighborhood and the average normal vector; is the standard deviation of the angle between the normal vector of each point in the neighborhood and the average normal vector; For points in the neighborhood relative to The standard deviation of the height values; This is the average value; To increase the standard deviation of the strength.

[0025] Reference branch extraction The baseline geometric and semantic features ( Change branch extraction Features in the same spatial location ( ).

[0026] Step 303, using two... The feature vectors of each layer are subjected to adaptive convolution. Layer calculation;

[0027] No. The initial value of the layer is 0;

[0028] The general calculation formula is:

[0029] ;

[0030] in, It is the first Convolutional accumulation and output of layers; It is the number of core points; These are the relative coordinates of the core point; It is a weight generation network, centered on the center point. and neighboring points relative coordinate vector As input, output and core point Corresponding adaptive weights; input vector The composition is as follows: ;

[0031] (t)=max Image function; Scale factor; Indicates the first The layer output, at the point The eigenvector at position n is used as the eigenvector of position n. Layer input;

[0032] Step 304, will the first Layer output The output features of this layer are obtained by transforming the function using a nonlinear activation function.

[0033]

[0034] It is to output linearly Transformation into nonlinear characteristics The activation function is defined as follows: .

[0035] Step 305, will the first Layer output As input to the next layer of convolution, it enables the gradual extraction of multi-level features and the aggregation of spatial information.

[0036] Step 306: The features extracted from each level of the Siamese AdaptConv dual-branch network are weighted and fused using the multi-scale aggregation module MS-AGG to obtain two corresponding fused features:

[0037] ;

[0038] in, For the first The s-th scale of each layer; each layer has One scale; The number of layers in the AdaptConv adaptive graph convolutional network; Let be the attention weight for the s-th scale.

[0039] Step 4: Utilize point cloud data and The fusion features are used to calculate the probability of change at the same location. ;

[0040] The specific steps are as follows:

[0041] Step 401, two point cloud data and The set of fused feature vectors:

[0042]

[0043] in, and They are the same spatial location Multi-scale aggregated features at two time points.

[0044] Step 402, calculate the same spatial point The magnitude of structural or semantic changes between two time points, i.e., the point-level feature distance metric:

[0045]

[0046] Step 403, measure the distance. Input discriminator Mapped to the probability of change via the Sigmoid function :

[0047]

[0048] In the formula, It is the Sigmoid activation function. This represents the confidence level that point y belongs to the category of change. It is the input feature distance, which is the degree of difference in point cloud features at two time points calculated by the Siamese network. It is a learnable bias parameter.

[0049] Step 5: Utilize the probability of change Train the Siamese AdaptConv dual-branch network;

[0050] Weighted binary cross-entropy loss is used during training:

[0051]

[0052] in ∈{0,1} is the truth label; Category weights; It is a weighted binary cross-entropy loss; the base of the log function is e.

[0053] Then, by adding a local consistency regularization term R to the loss function, the final total loss function L is obtained as follows:

[0054]

[0055] in The regularization term weight coefficient;

[0056]

[0057] Step 6: Using confidence thresholds The probability of change Mapped to a set of candidate change points It performs DBSCAN clustering, removes clusters with area / voxel volume less than the threshold, removes geometrically unreasonable clusters through structured consistency check, and outputs the final candidate region;

[0058] Candidate regions are either polygonal or point cloud masks representing changing regions.

[0059] The mapping formula is:

[0060] ;

[0061] in, The i-th candidate region is composed of spatially continuous points of change.

[0062] Step 7: Perform semantic verification on the candidate regions, mapping each candidate region to a semantic query vector:

[0063] Candidate region The semantic query vector for the mapping is: ; in, The land cover categories identified by the SiameseAdaptConv dual-branch network; The geographic coordinates of the region's center point; For time sequence labels.

[0064] Step 8: Pre-construct a geographic knowledge graph, retrieve the set of semantic entity nodes corresponding to each semantic query vector, and use the TransE embedding model to calculate semantic relevance;

[0065] Semantic query vector The corresponding formula for calculating semantic relevance is:

[0066]

[0067] in, for The corresponding set of semantic entity nodes, This indicates an embedded mapping function. This is a similarity metric.

[0068] Step 9: Determine whether the semantic relevance exceeds the set threshold. If it does, the change in the 3D point cloud of the candidate region is considered to be consistent with the semantics of existing entities in the knowledge graph and is judged as a reasonable change. Otherwise, the 3D point cloud of the candidate region is marked as an abnormal change. The abnormal change result will be fed back to the main detection module for secondary correction to achieve closed-loop optimization.

[0069] The advantages of this invention are:

[0070] (1) Algorithm level: This invention proposes a Siamese AdaptConv network with shared weights to realize automatic identification and semantic verification of geometric structure changes in multi-temporal point clouds of urban buildings, which significantly improves the accuracy of change detection.

[0071] (2) Structural level: This invention constructs a unified "local-global" discrimination model through multi-scale feature aggregation and spatial consistency constraint mechanism, which significantly reduces isolated false detections.

[0072] (3) System level: The present invention designs an end-to-end detection process and can be extended to a cloud-edge collaborative architecture to achieve minute-level response and lightweight deployment.

[0073] (4) In terms of algorithm structure, this invention achieves a breakthrough in the transformation from two-dimensional image to direct learning of three-dimensional original point cloud. In terms of processing flow, it achieves an upgrade from local feature extraction to global consistent detection. It has significant advantages such as high precision, high robustness and real-time deployability, and has good engineering application and industrial promotion value. Attached Figure Description

[0074] Figure 1 This is a flowchart of an end-to-end system for three-dimensional change detection based on Siamese AdaptConv, according to the present invention.

[0075] Figure 2 This is a schematic diagram illustrating the principle of integrating Siamese AdaptConv network feature extraction and change probability calculation in this invention.

[0076] Figure 3 This is a schematic diagram illustrating the point cloud change detection effect at time T1 based on Siamese AdaptConv according to the present invention;

[0077] Figure 4 This is a schematic diagram illustrating the point cloud change detection effect at time T2 based on Siamese AdaptConv according to the present invention. Detailed Implementation

[0078] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some, not all, embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0079] This invention proposes a 3D point cloud change detection method based on Siamese AdaptConv, which is used to realize the automatic identification and semantic verification of geometric structure changes in multi-temporal point cloud scenes. It is applicable to applications such as 3D change monitoring of urban buildings, smart city renewal and infrastructure operation and maintenance.

[0080] The proposed 3D point cloud change detection method based on Siamese AdaptConv consists of four core modules: data preprocessing, feature extraction, change detection, and semantic verification. The overall process is an end-to-end processing from the original point cloud input to the output of the changed region.

[0081] like Figure 1 As shown, the specific steps are as follows:

[0082] Step 1: For the scene to be tested, use LiDAR and UAV to move and measure, respectively, to obtain 3D point cloud data at two different time points, and perform preprocessing;

[0083] First-time point cloud dataset Point cloud dataset at the second time step .

[0084] in, , = These represent the three-dimensional coordinate vectors of a single point; , These represent the number of points in the cloud at two different time points.

[0085] Preprocessing refers to performing the following on the point cloud at each time step: coordinate normalization, outlier removal (statistical filtering), downsampling (FPS), and coarse and fine registration (FPFH + RANSAC → ICP); the registration goal is to... and By placing them in the same world coordinate system, local nearest neighbor queries can be compared.

[0086] Step 2: Construct an adaptive neighborhood for each preprocessed point that reflects its local geometric characteristics;

[0087] To achieve multi-scale context awareness and address uneven point cloud density, the adaptive neighborhood construction employs a smooth-transition dynamic neighborhood construction strategy, targeting the center point. The set of points in the local neighborhood The calculation is as follows, taking into account both scale consistency and point stability:

[0088]

[0089] It revolves around the center point The set of points in the local neighborhood of ; In a point cloud, besides the center point Any point other than; Point and points The Euclidean distance between them; It is a non-fixed adaptive radius, depending on the point. The density of local points in the surrounding area is dynamically calculated. In the point cloud, the distance to the point is... The set of the most recent 50 points.

[0090] Neighborhood searches are all accelerated using KD-Tree.

[0091] To avoid drastic changes in the number of neighborhood points, an adaptive radius is used. Adjust based on local density smoothing:

[0092]

[0093] Where β=0.2 is the maximum adjustment amplitude coefficient, which means the maximum adjustment radius is 1.2r; γ=2.0 is the sensitivity of the control adjustment; For global reference density; For point Local point cloud density; The baseline neighborhood radius, defined as 0.1m, is used to construct local neighborhoods; this formula ensures a continuous and smooth transition of the radius as density changes.

[0094] Density calculation: Define query point Local point cloud density and global reference density as follows:

[0095] = =

[0096] in, Take the median density of all point clouds in the scene; Indicated by Centered on, with radius as The number of points; This represents the total number of points in the scene. The total volume of the space.

[0097] Step 3: Transfer the preprocessed 3D point cloud data and Input the Siamese AdaptConv dual-branch network and fuse the multi-scale features of each layer to obtain their respective fused features;

[0098] like Figure 2 As shown, the Siamese AdaptConv dual-branch network achieves feature evolution from local to global by stacking multiple layers of AdaptConv operators:

[0099] Low-level geometric feature extraction: refers to the KPConv-L1 layer and related input layers, mainly utilizing enhanced geometric feature vectors. Extract the basic microscopic geometric properties of point clouds, such as normal vectors, curvature, and linearity;

[0100] Mid-level structural feature extraction: refers to the output of KPConv-L2 layer convolution, which captures the local combined structure and morphological features of point cloud through a weight generation network based on kernel point convolution mechanism;

[0101] Global semantic features: These refer to features at the KPConv-L3 layer and deeper. As the receptive field gradually expands, global information covering the macro-semantics and spatial distribution patterns of the scene is extracted.

[0102] The specific steps are as follows:

[0103] Step 301: Construct a twin dual-branch feature extraction architecture for the reference branch and the variant branch;

[0104] The reference branch and the variant branch are structurally symmetrical; a weight-sharing mechanism is employed, with both branches sharing the same set of learnable parameters during training and inference. .

[0105] Siamese shared weight strategy: To ensure the alignment of two temporal branches in the feature space, the Siamese network uses a parameter sharing strategy, namely: ;in This is the set of learnable parameters for the AdaptConv network.

[0106] By sharing weights, the comparability and alignment of two time-series features can be guaranteed.

[0107] Step 302, process the preprocessed 3D point cloud data and The inputs are fed into the reference branch and the variation branch, respectively, serving as the input layer (i.e., the first branch) of the Siamese AdaptConv two-branch network. =0th floor);

[0108] The initial feature vector of this layer is defined as Its value is determined by the enhanced geometric feature vectors corresponding to the 3D point cloud data at two time points. Its core function is to drive AdaptConv: it acts as the AdaptConv weight generation network. One of the key inputs. Indicates the first Layered mesh at points The eigenvector at that location; It is the first The output features of the layer;

[0109] Enhanced geometric eigenvectors Specifically as follows:

[0110]

[0111] Described by linearity; Describing flatness; Describing the scattering degree; It is anisotropic; For feature entropy; This represents the change in surface curvature; It is the average angle between the normal vector of each point in the neighborhood and the average normal vector; is the standard deviation of the angle between the normal vector of each point in the neighborhood and the average normal vector; For points in the neighborhood relative to The standard deviation of the height values; This is the average value; To increase the standard deviation of the strength.

[0112] Eigenvalue analysis of the covariance matrix of a neighborhood point set is one of the most effective methods for capturing local geometric shapes; the 3x3 covariance matrix Σ of the neighborhood points is calculated and eigenvalues ​​are obtained through eigenvalue decomposition. Then, the following set of shape descriptors is calculated:

[0113] Linearity (L): (Describe the degree to which it resembles a linear structure);

[0114] Flatness (P): (Describe the degree to which it resembles a planar structure);

[0115] Scattering degree (S): (Describes the degree of discreteness of the distribution);

[0116] Anisotropy (A): ;

[0117] Feature entropy (E): ;

[0118] in It is used to describe the complexity of local shapes.

[0119] Surface variation (C): This provides a robust curvature estimate.

[0120] Then, the existing features are expanded to provide more detailed statistical information:

[0121] Normal vector consistency (μ): Calculates the average angle between the normal vector of each point in the neighborhood and the average normal vector. and standard deviation , describes the smoothness of the surface.

[0122] Height difference (h): Calculates the distance between points in the neighborhood and the height difference. Standard deviation of height values ​​( (), replacing the simple range to describe local fluctuations.

[0123] Reflection intensity (I): excluding the average value In addition, increase the standard deviation of strength To capture the material variations on the surface of the same ground feature.

[0124] For each query point, a neighborhood that reflects its local geometric characteristics is constructed, and a set of information-rich primary features are calculated within the constructed neighborhood. These features will directly participate in the dynamic generation of AdaptConv convolutional kernel weights, providing powerful local geometric prior knowledge for the AdaptConv weight generation network.

[0125] Reference branch extraction The baseline geometric and semantic features ( Change branch extraction Features in the same spatial location ( ).

[0126] Step 303, using two... The feature vectors of each layer are subjected to adaptive convolution. Layer calculation;

[0127] No. The initial value of the layer is 0;

[0128] The general formula for calculating a basic convolutional layer is:

[0129]

[0130] in, It is the first Convolutional accumulation and output of layers; It is the number of core points; It is the first The relative offset of each kernel point in Euclidean space; dynamic weights It is obtained by normalizing the score output by a lightweight weighted generative network (MLP) using Softmax, with the center point as the center point. and neighboring points relative coordinate vector Or use richer local features as input, and output the kernel point. The corresponding adaptive weights; that is = Input vector The composition is as follows: ;

[0131] (t)=max Image function; Scale factor; Indicates the first The layer output, at the point The eigenvector at position n is used as the eigenvector of position n. Layer input; It provides a macroscopic geometric blueprint for a local area; Provided points Relative to the center The precise relative position;

[0132] Weight generation network Learning to integrate this information, for each core point The network predicts the most suitable weights. This means that in planar regions, the network can learn to generate smooth convolutional kernels to extract planar features; in areas such as building edges or roof corners, the network generates directional convolutional kernels to highlight their unique geometric characteristics. This mechanism makes the feature extraction process itself highly adaptive to the local geometry.

[0133] This design enables the method to have stronger discriminative power and adaptability when faced with complex geometric changes in urban environments, laying a solid data foundation for subsequent high-precision change detection.

[0134] Step 304, will the first Layer output The output features of this layer are obtained by transforming the function using a nonlinear activation function.

[0135]

[0136] It is the linear output of the AdaptConv convolution. Transformation into nonlinear characteristics The activation function is defined as follows: .

[0137] Step 305, will the first Layer output As input to the next layer of convolution, it enables the gradual extraction of multi-level features and the aggregation of spatial information.

[0138] ;

[0139] Step 306: The features extracted from each level of the Siamese AdaptConv dual-branch network are weighted and fused using the multi-scale aggregation module MS-AGG to obtain two corresponding fused features:

[0140] To further integrate feature representations at different scales, a multi-scale feature aggregation module (MS-AGG) is proposed. Let the feature sets from different layers be... The fusion feature is then defined as:

[0141] ;

[0142] in, For the first The s-th scale of each layer; each layer has One scale; The number of layers in the AdaptConv adaptive graph convolutional network; The attention weights at the s-th scale are obtained by Softmax normalization:

[0143] =

[0144] The above fusion feature definition and While the mathematical expressions remain consistent in form, the core of this invention lies in its use of a weight-sharing Siamese architecture. This architecture processes dual-temporal features through completely consistent aggregation logic and attention weights, aiming to eliminate model-level systematic errors and ensure high alignment of the feature space. Although the formula logic is the same, the feature variables... Point cloud data at different times and The process drives generation; when features are added, removed, or deformed at the same spatial location, the input feature values ​​change accordingly, causing significant differences in the numerical distribution of the final output fused feature vector. This "logical isomorphism, numerical heterogeneity" design enables subsequent step 402 to calculate feature distances. Accurately quantifying the magnitude of changes in a scenario is a necessary prerequisite.

[0145] This module is based on Softmax normalized weight calculation and assigns weight coefficients to features from different scales to maintain discrimination consistency within the range of local point density variation (0.5-2 times the global average density).

[0146] Step 4: Utilize point cloud data and The fusion features are used to calculate the probability of change at the same location. ;

[0147] The specific steps are as follows:

[0148] Step 401, two point cloud data and The set of fused feature vectors:

[0149]

[0150] in, and They are the same spatial location Multi-scale aggregated features at two time points.

[0151] Step 402, calculate the same spatial point The magnitude of structural or semantic changes between two time points, i.e., the point-level feature distance metric:

[0152]

[0153] Step 403, measure the distance. Input discriminator Mapped to the probability of change via the Sigmoid function :

[0154]

[0155] In the formula, It is the Sigmoid activation function. This represents the confidence level that point y belongs to the category of change. It is the input feature distance, which is the degree of difference in point cloud features at two time points calculated by the Siamese network. It is a learnable bias parameter.

[0156] Step 5: Utilize the probability of change Train the Siamese AdaptConv dual-branch network;

[0157] Weighted binary cross-entropy loss is used during training:

[0158]

[0159] in ∈{0,1} is the truth label; Category weights; It is a weighted binary cross-entropy loss; the base of the log function is e.

[0160] Then, to further enhance the spatial consistency of the predictions, a local consistency regularization term R is added to the loss function to constrain the Euclidean distance between the predicted probabilities of adjacent points to be less than 0.1, thereby improving the spatial continuity of the predictions; the final total loss function L is obtained as follows:

[0161]

[0162] in is the weight coefficient of the regularization term; R is the consistency regularization term, used to encourage the continuity of spatial neighborhood predictions.

[0163]

[0164] This regularization term, jointly optimized with the main loss, helps suppress isolated false detections. This joint optimization mechanism can simultaneously improve the accuracy and spatial coherence of change detection.

[0165] Step 6: Using confidence thresholds The probability of change Mapped to a set of candidate change points Then, DBSCAN clustering is performed on the candidate point set (parameters...). (minPts), remove clusters with area / voxel volume less than the threshold, remove geometrically unreasonable clusters through structured consistency verification (e.g., fitting plane / volume difference), and output the final candidate region;

[0166] Candidate regions are either polygonal or point cloud masks representing changing regions.

[0167] The mapping formula is:

[0168]

[0169] in, The i-th candidate region is composed of spatially continuous points of change.

[0170] The specific implementation logic is as follows:

[0171] In this embodiment, a clustering algorithm is used to spatially aggregate the set of change points C. The core implementation logic is as follows:

[0172] To extract the point cloud indices whose change probability is greater than a threshold τ, the following statement is used:

[0173] candidate_indices = np.where(change_probability > tau)

[0174] To perform DBSCAN clustering, aggregate scattered points into entity clusters, use the following statement:

[0175] clusters = DBSCAN(eps=0.5, min_samples=10).fit(points[candidate_indices])

[0176] Traverse the clusters and extract candidate regions using bounding boxes. The statement is as follows:

[0177] for cluster_id in np.unique(clusters.labels_): if cluster_id != -1:

[0178] Rectangle(xy=(min_x, min_y), width=dx, height=dy)

[0179] Step 7: Perform semantic verification on the candidate regions, mapping each candidate region to a semantic query vector;

[0180] Candidate region The semantic query vector for the mapping is: ; in, The land cover categories identified by the SiameseAdaptConv dual-branch network; The geographic coordinates of the region's center point; For time-series labels (indicating when the change occurred).

[0181] Step 8: Preconstruct a geographic knowledge graph G=(E,R), retrieve the set of semantic entity nodes corresponding to each semantic query vector, and calculate the semantic relevance using the TransE embedding model:

[0182] Semantic query vector The corresponding formula for calculating semantic relevance is:

[0183]

[0184] in, for The corresponding set of semantic entity nodes, This indicates an embedded mapping function. This is a similarity metric.

[0185] Step 9: Determine whether the semantic relevance exceeds the set threshold. If it does, the change in the 3D point cloud of the candidate region is considered to be consistent with the semantics of existing entities in the knowledge graph and is judged as a reasonable change. Otherwise, the 3D point cloud of the candidate region is marked as an abnormal change. The result of the abnormal change will be fed back to the main detection module for secondary correction, thereby achieving closed-loop optimization.

[0186] Through the aforementioned semantic enhancement mechanism, after completing the geometric layer change detection, the model further realizes semantic layer verification and logical consistency judgment, constructing a three-stage reasoning closed loop of "geometric detection → semantic verification → logical judgment".

[0187] The specific verification logic is implemented as follows:

[0188] The semantic relevance score si (based on the TransE embedding model) is calculated using the following statement:

[0189] semantic_similarity = compute_transe_similarity(query_vector, entity_vectors)

[0190] Decision-making process:

[0191] if semantic_similarity > threshold: # Determines the change as reasonable;

[0192] label_type = "Reasonable Change" #at Figure 4 The solid line box indicates the middle part; else: # if it is determined to be an abnormal change (such as: an isolated object on the water), it is fed back to the main module for secondary correction;

[0193] label_type = "Abnormal Change" #at Figure 4 The middle part is indicated by a black dashed box;

[0194] feedback_to_main_module( # Triggering closed-loop optimization mechanism;

[0195] In engineering deployment, a resource scheduling module based on task load thresholds automatically transfers tasks to the cloud for execution when the CPU utilization of edge nodes exceeds 80%, ensuring that the processing latency does not exceed 100ms.

[0196] Example:

[0197] This invention employs deep learning network architecture optimization, geographic entity semantic modeling, and cloud-edge collaborative computing to construct an end-to-end intelligent processing pipeline. This transforms traditional 3D change detection, which relies on manual threshold settings, into an automated and precise semantic understanding process. The method achieves decimeter-level reconstruction accuracy, minute-level response speed, and a false detection rate of less than 5%, demonstrating a breakthrough in existing technological bottlenecks and holding significant importance for the digital governance of smart cities and the development of the aerospace information industry.

[0198] To verify the effectiveness of this invention, the following hardware and software platform was used in the experiment:

[0199] The edge node module utilizes an NVIDIA Jetson AGX Xavier with 32GB of memory for real-time inference and data preprocessing. The cloud platform module employs an AWS EC2 P3 (8 × V100 GPUs) or a local cluster, running the PyTorch framework (version ≥ 1.9). The development environment uses Ubuntu 20.04, Python 3.8, CUDA 11.3, Open3D, Faiss, and scikit-learn. Data sources include airborne LiDAR (wavelength 1064 nm, ≥ 10 points / m²), oblique photography (0.05 m resolution), and Sentinel-2 multispectral imagery (10 m). The experimental area is 100 m × 100 m, with a point cloud sampling interval of 0.05 m and a density of approximately 2000 points / m². This environment satisfies both high-precision change detection requirements and allows for verification of algorithm robustness under multi-source heterogeneous conditions.

[0200] The specific operating steps are as follows:

[0201] Step 1: Multi-source data acquisition and preprocessing pipeline;

[0202] This invention employs a multi-source data fusion scheme to balance spatial coverage with the accuracy of ground feature details.

[0203] Data sources include: airborne LiDAR point cloud: point density ≥ 10 points / m², flight altitude 300–500 m; UAV oblique photogrammetry data: resolution 0.05 m, forward overlap 80%, lateral overlap 70%; high-resolution satellite imagery: resolution 0.65 m.

[0204] The multi-source data is first uniformly formatted and initially aligned through an adaptive quality control mechanism, and then preprocessed with registration and denoising to form high-quality fused point cloud data, providing stable input for subsequent Siamese AdaptConv feature extraction.

[0205] The distribution of the raw point cloud data at the first time step (T1) is as follows: Figure 3 As shown, it serves as a benchmark reference for change detection; Figure 3 This shows a top-down view of the point cloud at time T1 (an earlier time point). Figure 3 The map clearly marks "Building 1" (located in the area of ​​coordinates X: 95-110m, Y: 50-65m) and "Building 2" (located in the area of ​​coordinates X: 105-115m, Y: 40-50m), with the rest being open space.

[0206] 1) Spatiotemporal registration and fusion;

[0207] To ensure the consistency of data from different sensors in time and space, an improved ICP (Iterative ClosestPoint) algorithm is used for high-precision alignment of multi-source point clouds.

[0208] The registration process is divided into two stages: coarse registration and fine registration.

[0209] Coarse registration stage: Initial transformation is estimated using FPFH (Fast Point Feature Histogram) features combined with the RANSAC algorithm;

[0210] Fine-tuning stage: A KD-tree neighborhood acceleration mechanism is introduced, with an upper limit of 200 iterations and a convergence threshold of 0.001 m. The optimization objective is to minimize the energy function: ;

[0211] Where R is the rotation matrix and t is the translation vector. Here, λ represents the source and target points, respectively, and 0.1 is the regularization coefficient. This design suppresses overfitting through orthogonal constraints, ensuring high-precision spatial alignment of multi-source point clouds.

[0212] Furthermore, the time synchronization mechanism employs cubic spline interpolation to fill in missing timestamps, ensuring time alignment of multiple time-series data. After optimization, the registration error is controlled within 0.05m, and the fusion integrity is improved to 95%, approximately 60% higher than traditional methods.

[0213] 2) Multi-stage filtering and quality control;

[0214] After geometric alignment is completed, a hierarchical filtering strategy is adopted to suppress the influence of noise and outliers.

[0215] First, statistical filtering (with the standard deviation multiple threshold set to 2.0) is used to remove outliers;

[0216] Then, an adaptive Gaussian filter is applied for smoothing, with the filter kernel size dynamically adjusted according to the local point density.

[0217] ;

[0218] in For local point density, Let x be the adaptive scaling parameter at point x. The initial global scale is 0.8 meters. For reference point density, The density adjustment factor is set to 0.75.

[0219] The result is a noise removal rate of 95%, an effective information retention rate of over 90%, and a significant improvement in detail fidelity at the boundaries of complex building complexes.

[0220] Through the aforementioned multi-source data fusion and filtering mechanisms, the resulting point cloud possesses both local geometric accuracy and overall spatial continuity. This fused point cloud serves as the input to the subsequent Siamese AdaptConv model, providing a high-quality data foundation for feature extraction, change detection, and semantic verification modules.

[0221] Step 2: Point cloud feature extraction and multi-feature fusion;

[0222] This invention employs a five-layer AdaptConv dynamic convolutional network structure (kernel number (K=15), influence radius) in the point cloud feature extraction stage. Its convolution kernel is defined in continuous Euclidean space and can be directly applied to sparse point sets to achieve continuous convolution operations on point-level features.

[0223] In this invention, the dynamics of the convolution kernel are mainly reflected in the adaptive change of the kernel weights with the local geometric features of the input point cloud.

[0224] Let the local center point be y, and its neighborhood points be x∈N(y). The convolution kernel contains K discrete kernel points, whose spatial locations are denoted as yk. For each pair of neighborhood relations (x, y), this invention generates a weighted network. Adaptively generate dynamic weights corresponding to each kernel point: .

[0225] in, ; This represents the local geometric properties calculated by the present invention. This represents the features of the previous convolutional layer. This design allows the convolutional kernel to no longer have fixed parameters, but instead generate different weights for different scenes, regions, and sampling densities, achieving an adaptive expression of the structural shape.

[0226] The convolution output is transformed by the non-linear activation function ReLU: ;

[0227] By progressively expanding the receptive field through hierarchical downsampling (at ratios of 1 / 4, 1 / 8, 1 / 16, 1 / 32, and 1 / 64), the network can simultaneously capture local geometric details and global structural context. Compared to traditional voxel convolution, AdaptConv's kernel points are defined in a continuous space, avoiding the spatial information loss caused by voxelization, and improving mIoU to 85.3%.

[0228] Weight generation network A typical implementation is a two-layer fully connected MLP, whose input is a vector:

[0229] ;

[0230] Kernel weight normalization is achieved using Softmax:

[0231] ;

[0232] in, Normalized weight representing the k-th kernel point: This weight is a probability value calculated using the Softmax function, ranging from 0 to 1, representing the contribution of the kernel point to feature extraction in the geometric relationship between the current local region point y and its neighboring point x; This represents the exponential operation on the raw scores: its function is to convert the raw real scores output by the weight generation network into positive numbers, and to amplify the differences in scores between different kernel points through the characteristics of the exponential function, providing a non-negative input basis for subsequent normalization; The summation term represents the total index score of all kernel points: this summation term serves as the normalization denominator, ensuring that the weight of each kernel point is normalized to a distribution with a total probability of 1, thereby constraining the total amount of attention resources of the network to remain constant when processing each local region; K represents the total number of kernel points in the convolution kernel. The raw score represents the output of the weight generation network.

[0233] To enhance feature discriminative power, this invention integrates geometric and semantic multimodal features: ;in Indicates coordinate characteristics, For normal vectors or curvature features, For color (RGB), The intensity is the reflection intensity.

[0234] It should be noted that the multimodal attention weighting mechanism of this invention is different from that in the aforementioned AdaptConv. The generated kernel weights are not duplicated, and their levels of action are different: It focuses on predicting local neighborhood weights within the convolutional kernel, while the attention mechanism is used to redistribute the channel-level importance of the fused multimodal features, further improving the discriminative ability from the feature semantic level.

[0235] Introducing an attention weighting mechanism to achieve adaptive adjustment of feature importance: ;in It is a multilayer perceptron. This indicates weighted by elements.

[0236] This mechanism dynamically adjusts the fusion weights based on feature importance, enabling the network to focus on key areas of change in complex urban environments. Experiments show that the attention fusion mechanism improves feature discrimination by approximately 30%, significantly enhancing the detection performance of small target changes.

[0237] Step 3: Change detection initialization;

[0238] In the change detection phase, this invention constructs a Siamese AdaptConv dual-branch structure, with each branch containing 5 convolutional layers and sharing a parameter set. (Right now = This ensures consistency in the feature space of the two temporal point clouds. After multi-scale aggregation, two temporal point-level feature sets are obtained:

[0239] ;

[0240] To quantify the difference between the clouds at two time points in the feature space, the average feature distance is defined:

[0241] ;

[0242] A larger distance indicates a significant geometric or semantic change in the local structure or surface. To achieve scale adaptation and context fusion, this invention combines a Feature Pyramid Network (FPN) structure for multi-scale aggregation and introduces deformable convolutional offsets in AdaptConv to enhance deformation robustness.

[0243] Finally, the discriminator maps the feature differences to the probability space of change: ;

[0244] in The Sigmoid activation function outputs... This represents the confidence level that point y belongs to the category of change. When p(y) > (threshold) When the value is 0.7, it is considered a significant change point, thus realizing the detection of coarse-grained change areas.

[0245] Experimental results show that the change recall rate reaches 92% and the false alarm rate is reduced to 8%, providing reliable candidate regions for subsequent fine semantic verification.

[0246] The specific detection results of this method for newly constructed building areas and demolished building areas are as follows: Figure 4 As shown, Figure 4 Is with Figure 3 The corresponding point cloud top-down view and change detection results at time T2 (the later time point). By comparing with time T1, it can be clearly seen that: 1) "Building 2" in the lower right area has been demolished (marked with a dashed box, with no point cloud inside the box); 2) "New building" appeared in the upper middle area (approximately X: 98-103m, Y: 60-65m), and the "new area" is marked with a solid box. "Building 1" in the upper left to middle area remains unchanged. Figure 4 This effectively validates the method's ability to identify changes in buildings, such as new construction and demolition.

[0247] Step 4: Knowledge graph construction and semantic enhancement verification;

[0248] To further verify the semantic validity of the detected change regions, this invention introduces a geographic knowledge graph enhancement module to perform semantic consistency reasoning based on the geometric detection results. The system constructs a geographic knowledge graph based on Neo4j. , containing approximately Each entity node is associated with one of 20 semantic relationships. The TransE embedding model is used to embed each entity node... Represented as a 256-dimensional semantic vector .

[0249] For the detected change area Define semantic query vectors:

[0250] in For land cover category labels (output by AdaptConv). For geographical coordinates, This is a marker for time-series changes.

[0251] Calculate its semantic similarity with the corresponding entity in the knowledge graph. If If the change does not conform to the geographic semantic constraints, it is considered an abnormal change; otherwise, it is considered a reasonable change. Abnormal results will be fed back to the main detection module for secondary correction, realizing a closed-loop reasoning process of "geometric detection - semantic verification - logical judgment".

[0252] This invention implements a multi-stage change detection and verification mechanism from the geometric layer to the semantic layer, constructing a structurally consistent, logically closed-loop, and highly interpretable 3D point cloud change detection framework; furthermore, it introduces contrastive loss during training to enhance the separability of positive and negative samples. Let: ;

[0253] in This is the interval threshold; y is the total number of training samples. 1 represents a positive sample and 0 represents a negative sample; d(y) is the feature distance between samples; this design enhances the robustness of the discriminator.

[0254] Step 5: Multimodal fusion and refined change detection;

[0255] Based on preliminary detection and semantic verification, this invention further proposes a multimodal fusion and refined change detection module, which achieves high-precision change recognition of 3D point cloud scenes by combining geometric features and semantic information, dynamic threshold optimization and a dual-branch Siamese AdaptConv network.

[0256] (1) Adaptive threshold setting and multi-level filtering;

[0257] To overcome the high misclassification rate of the fixed threshold method in complex terrain, an improved Otsu dynamic thresholding algorithm is proposed, which comprehensively considers the variance of geometric difference and the variance of semantic consistency. The threshold is calculated as follows: ;

[0258] in: The optimal dynamic threshold; =0.5, which is the weighting coefficient; Between-class variance representing geometric differences; This represents the variance of the semantic feature distribution.

[0259] This method can adaptively adjust the threshold according to the scene complexity, maintaining stable performance in areas with high-rise buildings and road intersections. After thresholding the initial change probability (p(y)), a three-level filtering mechanism is executed:

[0260] Geometric consistency check: Based on shape, volume, and neighborhood geometric constraints, isolated noise points are eliminated;

[0261] Spatiotemporal continuity verification: Calculate the time difference between adjacent frames Consistency of spatial trajectories of internal change points;

[0262] Semantic rationality verification: combining knowledge graph constraints Verify the consistency between the change category and the geographic entity logic.

[0263] Finally, the DBSCAN clustering algorithm was used to remove residual noise, with the clustering parameters set to [value missing]. , Experimental results show that the module has a false alarm rate of less than 5% and a small target recall rate of 85%.

[0264] (2) Siamese AdaptConv refined detection network;

[0265] Based on the filtered candidate point set, the Siamese AdaptConv fine detection network is used for point-level difference reinforcement learning.

[0266] The network consists of a dual-branch encoder and a U-Net decoder. Each branch contains five layers of AdaptConv modules (number of kernel points (K=15), influence radius...). .

[0267] The formula for calculating the distance of point-level features is:

[0268] in For the first point cloud One point, These are the fused feature vectors of the two temporal point clouds, respectively.

[0269] To highlight areas of significant change, an attention-weighted mechanism is introduced: ;in The attention weight matrix is ​​normalized by the Softmax function and the importance of different regions is adaptively adjusted.

[0270] The final model outputs a probability map of change, which, after Sigmoid activation, yields the confidence level of the change. ;

[0271] The loss function uses weighted multi-class cross-entropy: Where: C=3, which is the total number of categories (newly built, demolished, unchanged); The weights for the i-th class are (new = 2.0, demolish = 1.5, unchanged = 1.0). For real labels, The model predicts probabilities; the base of the log function is e.

[0272] This design achieves progressive optimization of multi-scale features through geometric-semantic dual-feature fusion and attention enhancement mechanisms.

[0273] Experimental results show that the overall accuracy is 92.3%, the F1-score is 95.8%, and the inference speed is 40% faster than PointNet++. Figure 4 The test results shown demonstrate that this invention can accurately identify the construction and demolition of buildings.

[0274] Step 6: Knowledge graph-enhanced reasoning;

[0275] Based on the detection results in step 5, in order to further verify whether the detected building changes comply with urban semantic planning constraints, this invention designs a knowledge graph-enhanced reasoning module.

[0276] The system constructs a semantic relational database of urban buildings ({G}=(E,R)) based on the TransE embedding model (256 dimensions), and represents the embedding features of nodes and relations in vector form.

[0277] The similarity between a query point and a candidate semantic entity is defined as follows:

[0278] Where: q is the query feature corresponding to the detection point; e is the neighborhood or candidate entity in the knowledge graph; r is the semantic relation vector. When a change occurs, the system determines that it conforms to geographic semantic constraints; otherwise, it is marked as an abnormal change and feedback is provided for correction. This module can quickly identify whether newly added or demolished buildings conform to urban planning logic, improving semantic reasoning accuracy to 93.7% and reducing the false detection rate to 4.1%.

[0279] Step 7: Cloud-edge collaborative computing power scheduling and real-time deployment;

[0280] To balance real-time performance and computational efficiency, this invention proposes a hierarchical cloud-edge collaborative computing power scheduling mechanism to achieve integrated deployment of data processing, model inference, and resource optimization.

[0281] The system adopts a two-tier architecture:

[0282] Edge nodes: responsible for processing latency-sensitive real-time tasks (response time <100 ms), typically using NVIDIA Jetson AGX hardware with 32 GB of memory;

[0283] Cloud cluster: responsible for model training, batch inference and validation, deployed on AWS EC2 P3 instance (8×V100GPU).

[0284] To achieve dynamic load balancing, a scheduling optimization model based on reinforcement learning is constructed, and its reward function is defined as:

[0285] - latency

[0286] in: The immediate reward for taking action a in state s; System throughput (MB / s); latency is the task delay (ms). Calculate the cost per unit time (USD / h). Parameter settings: =0.6, =0.3, =0.1.

[0287] When CPU utilization exceeds 80%, the system automatically triggers capacity expansion to achieve elastic scaling of tasks. Combined with edge caching and data sharding strategies, asynchronous loading and parallel computing of data streams can be achieved.

[0288] This invention uses the Siamese AdaptConv deep network as its core, integrating knowledge graph reasoning and cloud-edge collaborative computing to construct an end-to-end intelligent processing framework for 3D point cloud change detection. This method achieves dual-temporal point cloud feature extraction and difference discrimination through multi-scale convolution and attention mechanisms, and introduces semantic constraints and computational power collaborative optimization to achieve a leap from geometric layer change detection to semantic-level intelligent recognition. The system is scalable and engineering-deployable, suitable for 3D change scenarios of urban buildings.

Claims

1. A three-dimensional point cloud change detection method based on Siamese AdaptConv, characterized in that, The specific steps are as follows: Step 1: For the scene to be tested, use LiDAR and UAV to move and measure, respectively, to acquire 3D point cloud data at two different time points. and And perform preprocessing; Step 2: Construct an adaptive neighborhood for each preprocessed point that reflects its local geometric characteristics; The adaptive neighborhood construction employs a smooth-transition dynamic neighborhood construction strategy, targeting the center point. The set of points in the local neighborhood The calculation is as follows: ; In a point cloud, besides the center point Any point other than; Point and points The Euclidean distance between them; It is an adaptive radius that is not fixed, depending on the point. The density of local points in the surrounding area is dynamically calculated. In the point cloud, the distance to the point is... The set of the most recent 50 points; Step 3: Transfer the preprocessed 3D point cloud data and Input the Siamese AdaptConv dual-branch network and fuse the multi-scale features of each layer to obtain their respective fused features; Step 4: Utilize point cloud data and The fusion features are used to calculate the probability of change at the same location. ; Step 5: Utilize the probability of change Train the Siamese AdaptConv dual-branch network; Step 6: Using confidence thresholds The probability of change Mapped to a set of candidate change points It performs DBSCAN clustering, removes clusters with area / voxel volume less than the threshold, removes geometrically unreasonable clusters through structured consistency check, and outputs the final candidate region; Step 7: Perform semantic verification on the candidate regions, mapping each candidate region to a semantic query vector; Step 8: Pre-construct a geographic knowledge graph, retrieve the set of semantic entity nodes corresponding to each semantic query vector, and use the TransE embedding model to calculate semantic relevance; Step 9: Determine whether the semantic relevance exceeds the set threshold. If it does, the change in the 3D point cloud of the candidate region is considered to be consistent with the semantics of existing entities in the knowledge graph and is judged as a reasonable change. Otherwise, the 3D point cloud of the candidate region is marked as an abnormal change. The abnormal change result is fed back to the main detection module for secondary correction to achieve closed-loop optimization.

2. The method as described in claim 1, characterized in that, In step one, the first time-instance point cloud dataset Point cloud dataset at the second time step ; in, , = These represent the three-dimensional coordinate vectors of a single point; , These represent the number of points in the point cloud at the two time points, respectively. Preprocessing refers to performing the following on the point cloud at each time step: coordinate normalization, removal of isolated points, downsampling, coarse registration, and fine registration.

3. The method as described in claim 1, characterized in that, In step two, the adaptive radius Adjust based on local density smoothing: ; Where β is the maximum adjustment amplitude coefficient; To control the sensitivity of the adjustment; For global reference density; For point Local point cloud density; The reference neighborhood radius is denoted as .

4. The method as described in claim 2, characterized in that, The specific steps of step 3 are as follows: Step 301: Construct a twin dual-branch feature extraction architecture for the reference branch and the variant branch; The reference branch and the variant branch are structurally symmetrical; a weight-sharing mechanism is employed, with both branches sharing the same set of learnable parameters during training and inference. ; Step 302, process the preprocessed 3D point cloud data and The inputs are fed into the reference branch and the variation branch, respectively, as the input layer of the Siamese AdaptConv dual-branch network; Input layer, i.e., the first =0 layer, its initial feature vector is defined as The value is 3D point cloud data. and Enhanced geometric eigenvectors Reference branch extraction The baseline geometric and semantic features ( Change branch extraction Features in the same spatial location ( ); Step 303, using two... The feature vectors of each layer are subjected to adaptive convolution. Layer calculation; No. The initial value of the layer is 0; The general calculation formula is: ; in, It is the first Layer convolution plus output; It is the number of core points; These are the relative coordinates of the core point; It is a weight generation network, centered on the center point. and neighboring points relative coordinate vector As input, output and core point Corresponding adaptive weights; input vector The composition is as follows: ； (t)=max Image function; Scale factor; Indicates the first The layer output, at the point The eigenvector at position n is used as the eigenvector of position n. Layer input; Step 304, will the first Layer output The output features of this layer are obtained by transforming the function using a nonlinear activation function. ; It is to output linearly Transformation into nonlinear characteristics The activation function is defined as follows: ; Step 305, will the first Layer output As input to the next layer of convolution, it enables the gradual extraction of multi-level features and the aggregation of spatial information; Step 306: The features extracted from each level of the Siamese AdaptConv dual-branch network are weighted and fused using the multi-scale aggregation module MS-AGG to obtain two corresponding fused features: ; ； in, For the first The s-th scale of each layer; each layer has One scale; The number of layers in the AdaptConv adaptive graph convolutional network; Let be the attention weight for the s-th scale.

5. The method as described in claim 4, characterized in that, In step 302, the geometric feature vector is enhanced. : ; Described by linearity; Describing flatness; Describing the scattering degree; It is anisotropic; For feature entropy; This represents the change in surface curvature; It is the average angle between the normal vector of each point in the neighborhood and the average normal vector; is the standard deviation of the angle between the normal vector of each point in the neighborhood and the average normal vector; For points in the neighborhood relative to The standard deviation of the height values; This is the average value; To increase the standard deviation of the strength.

6. The method as described in claim 1, characterized in that, The specific steps of step 4 are as follows: Step 401, two point cloud data and The set of fused feature vectors: ； in, and They are the same spatial location Multi-scale aggregated features at two time points; Step 402, calculate the same spatial point The magnitude of structural or semantic changes between two time points, i.e., the point-level feature distance metric: ; Step 403, measure the distance. Input discriminator Mapped to the probability of change via the Sigmoid function : ; In the formula, It is the Sigmoid activation function. It is the input feature distance. It is a learnable bias parameter.

7. The method as described in claim 6, characterized in that, In step 5, a weighted binary cross-entropy loss is used during training: ; in ∈{0,1} is the truth label; Category weights; It is a weighted binary cross-entropy loss; Then, a local consistency regularization term R is added to the loss function: ; The final total loss function L is obtained as follows: ;in This is the weight coefficient for the regularization term.

8. The method as described in claim 6, characterized in that, In step 6, the mapping formula is: ； in, The i-th candidate region is composed of spatially continuous points of change.

9. The method as described in claim 1, characterized in that, In step 7, the candidate region The semantic query vector for the mapping is: ; in, The land cover categories identified by the Siamese AdaptConv dual-branch network; The geographic coordinates of the region's center point; For time sequence labels.

10. The method as described in claim 1, characterized in that, In step 8, the semantic query vector The corresponding formula for calculating semantic relevance is: ; in, for The corresponding set of semantic entity nodes, This indicates an embedded mapping function. This is a similarity metric.

Images