Deep learning-based surveying and mapping point cloud data intelligent correction method and system

Abstract

The invention discloses a deep learning-based surveying and mapping point cloud data intelligent correction method and system, and relates to the technical field of data processing, and the method comprises the steps: obtaining original point cloud data of surveying and mapping equipment, carrying out the denoising and adaptive density normalization processing, constructing standardized point cloud data, and extracting a multi-scale feature set; inputting a deep learning model to perform feature learning, extracting different scale spatial features containing global surveying and mapping errors and local detail distortion in parallel, performing interactive fusion, and decoupling and identifying parameterized global errors and local distortion displacement fields; and fusing scene physical constraints, carrying out iterative adaptive optimization until a convergence condition is met, and outputting final corrected point cloud data. The technical problem of poor geometric fidelity and reliability of the point cloud data caused by difficulty in cooperative processing of global system errors and local complex distortion with huge scale difference in the prior art is solved, and the technical effect of improving the geometric fidelity and reliability of massive point cloud data is achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for intelligent correction of surveying point cloud data based on deep learning. Background Technology

[0002] During the acquisition of point cloud data, due to the complex operating environment and the physical mechanism of the equipment itself, the acquired raw point cloud contains a variety of systematic and random errors and distortions, which seriously restrict the high-level application value of the data. Among them, the errors are mainly divided into parameterized global errors and unstructured local distortions. Global errors usually originate from sensor intrinsic parameter calibration deviations, multi-source data registration misalignments, inaccurate coordinate system transformation models, etc., which manifest as systematic deviations in rotation, translation, and scaling of the point cloud as a whole. Local distortions are caused by dynamic environmental interference (such as moving objects, atmospheric refraction), platform vibration, intensity attenuation and incident angle effects caused by changes in scanning angle and distance, as well as multiple echoes and occlusions from complex surfaces (such as vegetation, grid structures), etc., which manifest as stretching, compression, distortion, or blurring of details in the local structure of the point cloud, with irregular shapes and uneven spatial distribution. Traditional point cloud correction and optimization mainly rely on geometric constraints, statistical filtering, or error equations based on physical models. For example, optimization is performed after registration using the Iterative Closest Point (ICP) algorithm or absolute accuracy correction is performed using ground control points. These methods have many limitations. For instance, they are difficult to effectively model and correct highly nonlinear, spatially varying, and complex local distortions. They also heavily rely on manual intervention and experience-based settings, resulting in low levels of automation and intelligence, and low efficiency when processing massive point clouds.

[0003] Therefore, current technologies face the challenge of coordinating global system errors with significant scale differences and complex local distortions, resulting in poor geometric fidelity and reliability of point cloud data. Summary of the Invention

[0004] This application provides a deep learning-based intelligent correction method and system for surveying point cloud data, which solves the technical problem in the prior art of the difficulty in coordinating the processing of global system errors with huge scale differences and local complex distortions, resulting in poor geometric fidelity and reliability of point cloud data, and achieves the technical effect of improving the geometric fidelity and reliability of massive point cloud data.

[0005] This application provides a deep learning-based intelligent correction method for surveying point cloud data. The method includes: acquiring raw point cloud data from a surveying device; performing denoising and semantically guided adaptive density normalization to construct standardized point cloud data; performing multi-scale partitioning based on the semantic labels of the standardized point cloud data and local geometric complexity; extracting a multi-scale feature set; inputting the multi-scale feature set into a deep learning model for feature learning; extracting spatial features of different scales containing global surveying errors and local detail distortions in parallel; interactively fusing the spatial features of different scales; decoupling and identifying the parameterized global error and local distortion displacement field in the point cloud data; generating a point-by-point correction uncertainty quantization value; using the parameterized global error, local distortion displacement field, and uncertainty quantization value as initial values, incorporating scene physical constraints, performing iterative adaptive optimization until convergence conditions are met, and outputting the final corrected point cloud data.

[0006] In a possible implementation, the intelligent correction method for surveying point cloud data based on deep learning includes: establishing a correction tracking record set by tracking error statistics, abnormal area markings, and comparison data before and after correction based on correction process data; performing error diagnosis and correction localization based on the correction tracking record set; and constructing a traceable error diagnosis report with visual charts and complete processing metadata.

[0007] In a possible implementation, the standardized point cloud data is constructed by: calculating the average k-nearest neighbor distance of each point through statistical filtering and removing outliers whose distance exceeds a preset threshold; performing semantic segmentation on the denoised point cloud to distinguish different categories, including at least ground, buildings, and vegetation, and obtaining a semantic label for each point; mapping different voxel sizes according to the semantic labels, adaptively downsampling each voxel size, and constructing the density-normalized standardized point cloud data.

[0008] In a possible implementation, multi-scale partitioning is performed based on the semantic labels of the standardized point cloud data combined with local geometric complexity, and a multi-scale feature set is extracted. This includes: calculating the eigenvalues ​​of the covariance matrix of the local neighborhood for each point in the standardized point cloud data, and determining the curvature, linearity, flatness, and dispersion of the point based on the eigenvalues ​​as a measure of local geometric complexity; constructing an adaptive octree for spatial partitioning based on the semantic labels and the local geometric complexity to obtain a multi-level structure containing coarse, medium, and fine scales; and extracting features at each scale based on the point cloud blocks corresponding to the multi-level structure to obtain the multi-scale feature set. Among these, the coarse-scale partitioning prioritizes the coverage of the ground semantic region to capture global mapping errors; the fine-scale partitioning targets high-curvature regions to preserve local detail features.

[0009] In a possible implementation, the multi-scale feature set is input into a deep learning model for feature learning, and spatial features at different scales containing global mapping errors and local detail distortions are extracted in parallel. This includes: the deep learning model comprising a fully connected self-attention mechanism network and a graph convolutional network, which are configured in parallel; the multi-scale feature set is input into the parallel self-attention mechanism network and graph convolutional network for feature learning, wherein the self-attention mechanism network learns the long-range dependencies and global context information of the point cloud to extract global spatial structure features; and the graph convolutional network learns the local neighborhood geometric structure and detail features of the point cloud based on k-neighborhood or a fixed radius, with each point aggregating feature information from its neighbors to obtain local detail geometric features.

[0010] In a possible implementation, the deep learning model further includes a mesoscale region encoder, wherein the mesoscale region encoder is a neural network based on a local self-attention mechanism. Through the local self-attention mechanism, it calculates the association weights between each point in the point cloud block and other points in its spatial neighborhood, extracts the features of local regional structural consistency and mid-frequency continuous distortion, and outputs mesoscale spatial features.

[0011] In possible implementations, spatial features at different scales are interactively fused to decouple the parameterized global error and local distortion displacement field in the point cloud data, generating a point-by-point correction uncertainty quantification value. This includes: performing top-down feature propagation on multi-scale spatial features based on a scale-hierarchical structure, upsampling coarse-scale features and concatenating them with fine-scale features from the previous layer to enhance the contextual information of fine-scale features; performing bottom-up feature aggregation, pooling fine-scale features and fusing them with coarse-scale features from the previous layer to enhance the detail information of coarse-scale features; aggregating features from different scales based on the target point, and calculating the fusion weights of multi-scale features based on the semantic label and geometric complexity context of the target point; using the fusion weights to perform weighted summation of multi-scale features to generate global interactive fusion features and local interactive fusion features for the target point; decoupling the parameterized global error describing the overall coordinate system deviation and the local distortion displacement field describing the detail deformation based on the global and local interactive fusion features, and generating a point-by-point correction uncertainty quantification value based on the variance estimate of the predicted points in the global and local features.

[0012] In a possible implementation, the parameterized global error, local distortion displacement field, and uncertainty quantization value are used as initial values, integrated with scene physical constraints, and iterative adaptive optimization is performed until the convergence condition is met, outputting the final corrected point cloud data. This includes: applying the parameterized global error to the overall point cloud to obtain a first intermediate point cloud; superimposing the local distortion displacement field onto the first intermediate point cloud to obtain a second intermediate point cloud, and retaining the uncertainty quantization value as the confidence map for this round of optimization; based on the optimization confidence map, reducing the learning rate for high-confidence regions and using them as anchor points, increasing the learning rate for low-confidence regions and applying additional smoothing or physical constraints, and simultaneously optimizing according to the current round of optimization. The confidence level is dynamically adjusted to optimize the neighborhood weights, allowing the anchor point constraints to propagate to surrounding low-confidence regions. The second intermediate point cloud is projected onto the scene's physical constraint space to obtain a third intermediate point cloud that satisfies both constraints and semantic consistency constraints. The residual between the current iteration and the previous iteration's second intermediate point cloud is calculated in the pre-projection space, and the uncertainty update value of the third intermediate point cloud is calculated. When the residual is less than a threshold and the average value of the uncertainty quantization update value is less than the target value, the convergence condition is deemed met, and the third intermediate point cloud is output as the final corrected point cloud data. When the convergence condition is not met, the third intermediate point cloud and the updated uncertainty quantization value are used as the initial values ​​for a new round, and the iteration continues.

[0013] In possible implementations, the scene physical constraints include: rigid constraints, planar constraints, smooth constraints, and topological constraints; wherein, the rigid constraints ensure that points on the same semantically labeled object maintain a constant relative distance after correction; the planar constraints ensure that points identified as planes conform to planar requirements after correction; the smooth constraints force a gradual change in the correction vectors of adjacent points in natural terrain areas; and the topological constraints ensure that linear elements with fixed topological connection structures maintain connectivity and continuity after correction.

[0014] This application also provides a deep learning-based intelligent correction system for surveying point cloud data. The system includes: a point cloud data standardization module, used to acquire the original point cloud data from surveying equipment, perform denoising and semantically guided adaptive density normalization processing to construct standardized point cloud data, and perform multi-scale partitioning based on the semantic labels of the standardized point cloud data combined with local geometric complexity to extract a multi-scale feature set; a spatial feature extraction module, used to input the multi-scale feature set into a deep learning model for feature learning, extract spatial features of different scales containing global surveying errors and local detail distortions in parallel, and interactively fuse spatial features of different scales to decouple and identify parameterized global errors and local distortion displacement fields in the point cloud data, generating point-by-point correction uncertainty quantification values; and a corrected point cloud data output module, used to take the parameterized global errors, local distortion displacement fields, and uncertainty quantification values ​​as initial values, incorporate scene physical constraints, perform iterative adaptive optimization until convergence conditions are met, and output the final corrected point cloud data.

[0015] This application proposes a deep learning-based intelligent correction method and system for surveying point cloud data. The method acquires raw point cloud data from surveying equipment, performs denoising and adaptive density normalization, constructs standardized point cloud data, and extracts multi-scale feature sets. These features are then input into a deep learning model for feature learning, and in parallel, spatial features at different scales containing global surveying errors and local detail distortions are extracted and interactively fused to decouple and identify parameterized global errors and local distortion displacement fields. Scene physical constraints are incorporated, and iterative adaptive optimization is performed until convergence conditions are met, outputting the final corrected point cloud data. This method solves the technical problem in existing technologies where it is difficult to collaboratively process global system errors with large scale differences and complex local distortions, leading to poor geometric fidelity and reliability of point cloud data. It achieves the technical effect of improving the geometric fidelity and reliability of massive point cloud data. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a schematic diagram of a method for intelligent correction of surveying point cloud data based on deep learning, provided in an embodiment of this application.

[0018] Figure 2This is a schematic diagram of the structure of a deep learning-based intelligent correction system for surveying point cloud data, provided in an embodiment of this application.

[0019] Figure labeling: Point cloud data standardization module 10, spatial feature extraction module 20, and corrected point cloud data output module 30. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0021] This application provides an intelligent correction method for surveying point cloud data based on deep learning, such as... Figure 1 As shown, the method includes: Step S100: Obtain the original point cloud data from the surveying equipment, perform denoising and semantically guided adaptive density normalization processing to construct standardized point cloud data, and perform multi-scale partitioning based on the semantic labels of the standardized point cloud data and local geometric complexity to extract multi-scale feature sets.

[0022] Preferably, raw point cloud data from surveying equipment such as LiDAR, oblique photogrammetry, and 3D laser scanners are acquired. Statistical filtering is used to denoise the data, removing noise points that are not on the actual ground surface or ground features, thus improving the signal-to-noise ratio. Next, semantically guided adaptive density normalization is performed, including semantic segmentation of each denoised point cloud and assigning semantic labels such as "ground," "building facade," "roof," "vegetation," and "road." Then, adaptive downsampling is employed, setting different target densities based on the semantic labels. For example, smaller voxel sizes are used for downsampling of ground and roads to retain high accuracy and density, while larger voxel sizes are used for vegetation and irregular objects for sparser sampling. Within the same semantic category, the sampling rate is further adjusted by combining geometric information such as local curvature, resulting in slightly sparser sampling in flat areas and slightly denser sampling at edges and corners. This generates point clouds with relatively uniform density but reasonable density differences between different ground feature categories, thus obtaining standardized point cloud data.

[0023] Furthermore, step S100 also includes: calculating the average k-nearest neighbor distance of each point through statistical filtering, and removing outliers whose distance exceeds a preset threshold; performing semantic segmentation on the denoised point cloud to distinguish different categories including at least ground, buildings, and vegetation, and obtaining semantic labels for each point; mapping different voxel sizes according to the semantic labels, adaptively downsampling each voxel size, and constructing the density-normalized standardized point cloud data.

[0024] Preferably, the original point cloud data is denoised using statistical filtering. Specifically, the average distance between each point in the point cloud and its k nearest neighbors is calculated. The global mean and standard deviation of the average distance for all points in the entire point cloud are calculated. Assuming that the distance distribution of the entire point cloud follows a Gaussian distribution, points whose average distance exceeds the mean by several times (e.g., 3 times) the standard deviation are considered isolated and unreasonable outliers, such as splash points or isolated points caused by transient sensor malfunctions, and are removed, thus eliminating outliers caused by sensor errors, splashes, or airborne particles. Then, a pre-trained deep learning semantic segmentation model is used to perform semantic segmentation on the denoised point cloud. This involves analyzing the geometric features (such as height, normal vector, curvature, etc.) and contextual relationships of each point and its surrounding points, predicting and assigning semantic labels to each point, including at least "ground," "building," and "vegetation." Ground serves as a reference for geometric correction; buildings have regular and rigid geometric features that require high-precision preservation; vegetation surfaces are complex and irregular and may exhibit dynamic swaying, requiring different data processing strategies than rigid objects. Then, based on the semantic labels, different voxels are mapped to points of different categories. Size refers to setting different target processing precisions. For ground / road mapping, the voxel size is smaller, such as 0.1 meters; for building facade / roof mapping, the voxel size is medium, such as 0.2 meters; and for vegetation mapping, the voxel size is larger, such as 0.5 meters, to significantly reduce the amount of data. Then, adaptive downsampling is performed on each voxel size. This includes dividing the space into multiple voxels for all points belonging to the same semantic category using the voxel size set for that category. Within each voxel, a representative point (the centroid or center point of all points) is used to replace all the original points within that voxel, and finally, density-normalized standardized point cloud data is obtained.

[0025] Furthermore, step S100 also includes: calculating the eigenvalues ​​of the covariance matrix of the local neighborhood for each point in the standardized point cloud data; determining the curvature, linearity, flatness, and dispersion of the point based on the eigenvalues ​​as a measure of local geometric complexity; constructing an adaptive octree for spatial partitioning based on the semantic labels and the local geometric complexity to obtain a multi-level structure containing coarse, medium, and fine scales; extracting features at each scale based on the point cloud blocks corresponding to the multi-level structure to obtain the multi-scale feature set; wherein, the coarse-scale partitioning prioritizes covering the ground semantic region to capture global mapping errors; and the fine-scale partitioning targets high-curvature regions to preserve local detail features.

[0026] Preferably, for each point in the standardized point cloud data, its k nearest neighbors are determined, and the covariance matrix of the local neighborhood is calculated to describe the distribution of multiple neighboring points in three spatial directions. Then, the covariance matrix is ​​decomposed into eigenvalues ​​to determine the eigenvalues. The magnitude of these eigenvalues ​​directly reflects the geometric type of the local surface. Based on the eigenvalues, the curvature, linearity, flatness, and dispersion of the point are calculated. The larger the curvature / dispersion value, the more diffuse the point is in a certain direction, the more curved or discrete the surface is, and the more complex the geometry, such as the edge of a leaf or the texture of a sculpture. The larger the linearity value, the more it resembles a line, such as a power line or a railing. The larger the flatness value, the more it resembles a plane, such as a wall or the ground. A set of quantitative indicators is calculated for each point to constitute a measure of local geometric complexity. For example, a point located on a flat ground has high flatness and low curvature, while a point located at the corner of a building's eaves has high curvature and low flatness.

[0027] Preferably, based on scene content and geometric features, combined with semantic tags and local geometric complexity, an adaptive octree is constructed for spatial partitioning. An octree is a spatial data structure used to recursively subdivide a cubic space into eight sub-cubes. Specifically, coarse-scale partitioning prioritizes the ground, generating larger point cloud blocks for semantic regions labeled "ground," with the ground serving as an overall reference benchmark to help capture large-scale global errors such as overall tilt and settlement. Fine-scale partitioning targets high-curvature areas, continuously subdividing curved building corners, sculptural decorations, and vegetation edges with high local geometric complexity until very small point cloud blocks are generated to preserve and highlight local details. Meso-scale partitioning lies between the two, used to handle regional structures such as building facades and regular roofs. The final result is a multi-layered structure containing coarse, meso, and fine scales—an octree with uneven spatial partitioning granularity. The root node, intermediate nodes, and leaf nodes naturally correspond to the coarse, meso, and fine scales, respectively, and each node contains a point cloud block.

[0028] Preferably, feature extraction at each scale is performed on the point cloud blocks corresponding to the multi-level structure to obtain feature vectors representing the collective features of the point cloud within the block. Specifically, for each node or point cloud block in the octree, all points within the block are used as input, and a shared, lightweight feature extraction network, such as a convolutional network, is used for processing to output a fixed-length feature vector, which is used to encode the overall geometric and semantic attributes of the point cloud block. Finally, a multi-scale feature set is determined. Among them, the feature vectors of coarse-scale nodes carry large-scale, global contextual information; the feature vectors of fine-scale nodes carry small-scale, highly detailed local information; and the feature vectors of meso-scale nodes carry information about regional consistency.

[0029] Step S200: Input the multi-scale feature set into the deep learning model for feature learning, extract spatial features of different scales containing global mapping errors and local detail distortions in parallel, and perform interactive fusion of spatial features of different scales to decouple and identify parameterized global errors and local distortion displacement fields in point cloud data, and generate point-by-point correction uncertainty quantification values.

[0030] Step S200 further includes the following: the deep learning model includes a fully connected self-attention mechanism network and a graph convolutional network, which are set in parallel. The multi-scale feature set is input into the parallel self-attention mechanism network and the graph convolutional network for feature learning. Specifically, the self-attention mechanism network learns the long-range dependencies and global context information of the point cloud to extract global spatial structure features; the graph convolutional network learns the local neighborhood geometric structure and detailed features of the point cloud based on k-neighborhood or a fixed radius. Each point aggregates feature information from its neighboring points to obtain local detailed geometric features.

[0031] Preferably, a deep learning model is used to learn features from a multi-scale feature set, and global mapping errors and local detail distortions at different spatial scales are extracted in parallel. Global mapping errors are manifested as systematic deviations of the entire point cloud or large areas, such as overall rotation, translation, scaling, or slight bending. Local detail distortions are manifested as small-scale, irregular twisting, stretching, or compression in the point cloud, such as blurred building edges due to vibration or uneven walls due to incident angle effects. The deep learning model includes three well-defined neural network paths for processing features at different scales. Specifically, the first neural network path is a fully connected self-attention mechanism network, such as the encoder-decoder structure of Transformer, which receives features from coarse-scale and some meso-scale feature sets, extracts global spatial structure features, and allows any two points or feature blocks to interact directly. By calculating the correlation between all features, the contextual relationship of the entire scene is understood, and the systematic offset or deformation pattern of the entire point cloud as a whole is captured through long-range dependency modeling. The second neural network pathway is a graph convolutional network, which mainly receives features from fine-scale and some mesoscale feature sets to extract local detailed geometric features. The graph convolutional network treats the point cloud as a graph, where points are nodes and the connections between points and their neighbors (k-nearest neighbors or within a fixed radius) are edges. By repeatedly aggregating and transmitting information within the local neighborhood, each point obtains features from its neighbors, updates its own features, and then transmits them to the next neighbor, thereby characterizing the subtle geometric changes and topological relationships of the local surface, identifying regions where the point cloud arrangement exhibits unnatural distortions, expansions, or contractions, and ultimately generating a local distortion displacement field that describes fine deformations.

[0032] Furthermore, step S200 also includes the deep learning model further including: a mesoscale region encoder, wherein the mesoscale region encoder is a neural network based on a local self-attention mechanism, which calculates the association weights between each point in the point cloud block and other points in its spatial neighborhood through the local self-attention mechanism, extracts the features of local regional structural consistency and mid-frequency continuous distortion, and outputs mesoscale spatial features.

[0033] Preferably, the third neural network pathway is a mesoscale region encoder, which is a neural network based on a local self-attention mechanism to extract mesoscale regional features. Specifically, for each query point in a mesoscale point cloud block, the attention calculation is restricted to a local neighborhood centered on that point with a fixed spatial radius or a fixed number of nearest neighbors. By calculating the feature correlation between the query point and all key points in the local neighborhood, attention weights are dynamically generated. The important structural information in the neighborhood is weighted and aggregated to extract the features of local regional structural consistency and mid-frequency continuous distortion, and outputs mesoscale spatial features.

[0034] Furthermore, step S200 also includes: performing top-down feature propagation on multi-scale spatial features based on a scale-hierarchical structure, upsampling coarse-scale features and concatenating them with fine-scale features from the previous layer to enhance the contextual information of fine-scale features; performing bottom-up feature aggregation, pooling fine-scale features and fusing them with coarse-scale features from the previous layer to enhance the detail information of coarse-scale features; aggregating features from different scales based on the target point, and calculating the fusion weights of multi-scale features based on the semantic label and geometric complexity context of the target point; using the fusion weights to perform weighted summation of multi-scale features to generate global interactive fusion features and local interactive fusion features for the target point; decoupling the parameterized global error describing the overall coordinate system deviation and the local distortion displacement field describing the detail deformation based on the global and local interactive fusion features, and generating a point-by-point correction uncertainty quantification value based on the variance estimation of the predictions in the global and local features.

[0035] Preferably, spatial features at different scales are interactively fused, that is, top-down feature propagation is performed on multi-scale spatial features. Specifically, coarse-scale features are upsampled through nearest neighbor interpolation, transposed convolution, etc., to the same spatial resolution as the fine-scale feature map of the next layer. Then, the upsampled coarse-scale features are concatenated with the original fine-scale features in the channel dimension to enhance the contextual information of the fine-scale features. Fine-scale features are downsampled through pooling or convolution with stride to the same spatial resolution as the coarse-scale feature map of the previous layer. Then, the pooled fine-scale features are fused with the original coarse-scale features element-wise or attention-weighted to enhance the detail information of the coarse-scale features. For example, after fusing a large amount of local geometric information of fine-scale points into global ground features, it is possible to more keenly perceive whether there are subtle and continuous overall undulations over a large area.

[0036] Preferably, the enhanced features of the target point at all scales after bidirectional interaction are aggregated, and a weight prediction network is established based on a multilayer perceptron. This weight prediction network is called to calculate the fusion weights of multi-scale features using the semantic label and local geometric complexity of the target point as key inputs. If the target point is a ground point located in a flat area, the coarse-scale and meso-scale features are more important for its correction and have higher feature fusion weights. If the target point is a building edge point, the feature fusion weights of fine-scale and meso-scale should be higher. Then, the features at different scales are weighted and summed using the determined feature fusion weights to finally generate the global interactive fusion features and local interactive fusion features of the target point.

[0037] Preferably, the global interactive fusion features and local interactive fusion features are input into the regression module, respectively. For each point, a parameterized global error describing the overall coordinate system deviation and a local distortion displacement field describing the detailed deformation are output. The parameterized global error may be a 6-DOF rigid body transformation or a more complex low-order polynomial surface. The local distortion displacement field indicates the amount of displacement in the X, Y, and Z directions required to eliminate local detailed deformation at that point. Based on the variance estimates predicted in the global and local features, the regression module simultaneously outputs the mean and variance of the predicted values. The larger the variance, the more uncertain the correction suggestion for that point is, which may be because the point is located in a feature ambiguity region, such as dense vegetation or where features of different scales contradict each other. Finally, a point-by-point correction uncertainty quantification value is generated to intuitively represent the reliability of the suggested correction for that point.

[0038] Step S300: The parameterized global error, local distortion displacement field, and uncertainty quantization value are used as initial values, integrated with scene physical constraints, and iterative adaptive optimization is performed until the convergence condition is met, and the final corrected point cloud data is output.

[0039] Step S300 further includes: applying the parameterized global error to the overall point cloud to obtain a first intermediate point cloud; superimposing the local distortion displacement field onto the first intermediate point cloud to obtain a second intermediate point cloud, and retaining the uncertainty quantization value as the confidence map for this round of optimization; based on the optimization confidence map, reducing the learning rate for high-confidence regions and using them as anchor points, increasing the learning rate for low-confidence regions and applying additional smoothing or physical constraints, and simultaneously dynamically adjusting the neighborhood optimization weights according to the confidence of this round of optimization, so that the anchor point constraints propagate to the surrounding low-confidence regions; and then using the second... The intermediate point cloud is projected onto the scene's physical constraint space to obtain a third intermediate point cloud that satisfies both the constraints and semantic consistency constraints. The residual between the current iteration and the second intermediate point cloud from the previous iteration is calculated in the space before projection, and the uncertainty update value of the third intermediate point cloud is calculated. When the residual is less than a threshold and the average value of the uncertainty quantization update value is less than the target value, the convergence condition is determined to be met, and the third intermediate point cloud is output as the final corrected point cloud data. When the convergence condition is not met, the third intermediate point cloud and the updated uncertainty quantization value are used as the initial values ​​for a new round, and the iteration continues.

[0040] Preferably, rigid body transformation or low-order polynomial surfaces are applied to the overall point cloud to initially correct the overall coordinate system deviation and determine the first intermediate point cloud. The predicted point-by-point displacement vectors are then superimposed onto each point of the first intermediate point cloud to initially repair local detail deformations, obtaining the second intermediate point cloud. Simultaneously, the uncertainty quantification value of each point is used as the confidence map for this round of optimization. Based on the optimization confidence map, the learning rate is reduced in high-confidence regions (i.e., small-step fine-tuning) to avoid destroying the good results already obtained, and these points are treated as relatively reliable fixed reference points. The learning rate is increased in low-confidence regions (i.e., large-step adjustments) to seek a better solution. It also applies additional smoothing or physical constraints, such as forcing the displacement changes of low-confidence regions to smoothly transition with their high-confidence neighborhoods, or forcing low-confidence points belonging to the same object to satisfy rigid non-deformation constraints; simultaneously, it dynamically adjusts the neighborhood optimization weights based on the current round of optimization confidence, that is, the optimization objective function includes a smoothing term based on dynamic weighting of confidence. High-confidence anchor points use this smoothing term to strongly propagate their positional constraints to the surrounding low-confidence regions, guiding the points in these regions to move to more reasonable positions. Reliable anchor points are used to constrain and correct unreliable regions, realizing intelligent information propagation and correction in the optimization process.

[0041] Preferably, the second intermediate point cloud is projected onto the scene's physical constraint space. These physical constraints include rigid constraints, planar constraints, smoothness constraints, and topological constraints, resulting in a third intermediate point cloud that satisfies both constraints and semantic consistency constraints, ensuring physical plausibility (e.g., all walls are flat, and all floors are continuous and smooth). Then, the positional differences between the second intermediate point cloud at the start of the current iteration and the third intermediate point cloud after physical projection are compared. The residual in the space before projection is calculated to measure the impact of physical constraints on the current solution. Based on the results and residuals from this optimization round, the uncertainty of each point's position is reassessed. Qualitatively, the updated uncertainty quantification value is obtained. When the residual is less than the threshold and the average value of the updated uncertainty quantification value is less than the target value, it indicates that the physical constraints no longer require large adjustments to the point cloud, and the position of the entire point cloud is already very certain. It is determined that the convergence condition is met, and the iteration stops and the final corrected point cloud data is output. If any convergence condition is not met, the third intermediate point cloud obtained in this round of optimization and the updated uncertainty are used as the input for the next round of iteration, and adaptive optimization iteration continues, thereby ensuring that the output final corrected point cloud data is optimal and physically reliable, meeting the strict requirements of the surveying and mapping field for high-precision and high-fidelity data.

[0042] Furthermore, step S300 also includes the following: the scene physical constraints include: rigid constraints, planar constraints, smooth constraints, and topological constraints; wherein, the rigid constraints ensure that points on the same semantically labeled object maintain a constant relative distance after correction; the planar constraints ensure that points identified as planes meet planar requirements after correction; the smooth constraints force the correction vectors of adjacent points to change smoothly in natural terrain areas; and the topological constraints ensure that linear elements with fixed topological connection structures maintain connection relationships and continuity after correction.

[0043] Preferably, the scene physical constraints include rigid constraints, planar constraints, smoothing constraints, and topological constraints. Among them, rigid constraints require that points on the same semantically labeled object maintain a constant relative distance after correction, meaning that the object can move or rotate as a whole, but cannot be stretched, compressed, or twisted. Planar constraints require that points identified as planes (such as building exterior walls, interior ceilings, and road surfaces) meet the planar requirements after correction, which is used to strongly correct unrealistic distortions such as unevenness and bulges in wall point clouds caused by sensor errors, occlusion, etc. Smoothing constraints force the correction vector changes of adjacent points to be gradual in natural terrain areas, preventing local noise from being misjudged as real terrain undulations and over-corrected during the optimization process, resulting in unrealistic "waves" or "spikes" on the ground. Topological constraints require that linear elements with fixed topological connection structures maintain connection relationships and continuity after correction, that is, for linear infrastructure networks, such as power lines, pipelines, railings, and road markings, their physical connection relationships and linear continuity must be maintained after correction, and there should be no breaks, misalignments, or non-physical intersections, ensuring the correctness of the functional logic of key engineering elements.

[0044] In the above text, refer to Figure 1 This paper describes in detail an intelligent correction method for surveying point cloud data based on deep learning according to an embodiment of the present invention. Next, we will refer to... Figure 2 This invention describes an intelligent correction system for surveying point cloud data based on deep learning, according to an embodiment of the present invention.

[0045] According to an embodiment of the present invention, a deep learning-based intelligent correction system for surveying point cloud data is used to solve the technical problem in the prior art of struggling to coordinate global system errors with huge scale differences and complex local distortions, resulting in poor geometric fidelity and reliability of point cloud data. This system achieves the technical effect of improving the geometric fidelity and reliability of massive point cloud data. Figure 2 As shown, a deep learning-based intelligent correction system for surveying point cloud data includes: a point cloud data standardization module 10, a spatial feature extraction module 20, and a corrected point cloud data output module 30.

[0046] The point cloud data standardization module 10 is used to acquire the original point cloud data of the surveying equipment, perform denoising and semantically guided adaptive density normalization processing to construct standardized point cloud data, and perform multi-scale partitioning based on the semantic labels of the standardized point cloud data and local geometric complexity to extract multi-scale feature sets; the spatial feature extraction module 20 is used to input the multi-scale feature sets into a deep learning model for feature learning, extract spatial features of different scales containing global surveying errors and local detail distortions in parallel, and perform interactive fusion of spatial features of different scales to decouple and identify parameterized global errors and local distortion displacement fields in the point cloud data, and generate point-by-point correction uncertainty quantization values; the corrected point cloud data output module 30 is used to take the parameterized global errors, local distortion displacement fields and uncertainty quantization values ​​as initial values, incorporate scene physical constraints, perform iterative adaptive optimization until the convergence condition is met, and output the final corrected point cloud data.

[0047] The following section describes in detail the specific configuration of a deep learning-based intelligent correction system for surveying point cloud data. It also includes: establishing a correction tracking record set based on correction process data, tracking error statistics, anomaly area marking, and comparison data before and after correction; performing error diagnosis and correction localization based on the correction tracking record set; and constructing a traceable error diagnosis report with visual charts and complete metadata processing.

[0048] The specific configuration of the point cloud data standardization module 10 will be described in detail below. The point cloud data standardization module 10 further includes: calculating the average k-nearest neighbor distance for each point through statistical filtering, and removing outliers whose distance exceeds a preset threshold; performing semantic segmentation on the denoised point cloud to distinguish different categories, including at least ground, buildings, and vegetation, and obtaining a semantic label for each point; mapping different voxel sizes according to the semantic labels, adaptively downsampling each voxel size, and constructing the density-normalized standardized point cloud data.

[0049] The specific configuration of the point cloud data standardization module 10 will be described in detail below. The point cloud data standardization module 10 further includes: calculating the eigenvalues ​​of the covariance matrix of the local neighborhood for each point in the standardized point cloud data; determining the curvature, linearity, flatness, and dispersion of the point based on the eigenvalues ​​as a measure of local geometric complexity; constructing an adaptive octree for spatial partitioning based on semantic labels and the local geometric complexity to obtain a multi-level structure containing coarse, medium, and fine scales; extracting features at each scale based on the point cloud blocks corresponding to the multi-level structure to obtain the multi-scale feature set; wherein, the coarse-scale partitioning prioritizes covering the ground semantic region to capture global mapping errors; the fine-scale partitioning targets high-curvature regions to preserve local detail features.

[0050] The specific configuration of the spatial feature extraction module 20 will be described in detail below. The spatial feature extraction module 20 further includes: the deep learning model comprising a fully connected self-attention mechanism network and a graph convolutional network, wherein the self-attention mechanism network and the graph convolutional network are configured in parallel; the multi-scale feature set is input into the parallel self-attention mechanism network and the graph convolutional network respectively for feature learning; wherein, the self-attention mechanism network learns the long-range dependencies and global context information of the point cloud to extract global spatial structure features; and the graph convolutional network learns the local neighborhood geometric structure and detailed features of the point cloud based on k-neighborhood or a fixed radius, with each point aggregating feature information from its neighboring points to obtain local detailed geometric features.

[0051] The specific configuration of the spatial feature extraction module 20 will be described in detail below. The spatial feature extraction module 20 further includes: the deep learning model further includes: a mesoscale region encoder, wherein the mesoscale region encoder is a neural network based on a local self-attention mechanism. Through the local self-attention mechanism, it calculates the association weights between each point within a point cloud block and other points in its spatial neighborhood, extracts features of local regional structural consistency and mid-frequency continuous distortion, and outputs mesoscale spatial features.

[0052] The specific configuration of the spatial feature extraction module 20 will be described in detail below. The spatial feature extraction module 20 further includes: performing top-down feature propagation on multi-scale spatial features based on a scale-hierarchical structure, upsampling coarse-scale features and concatenating them with fine-scale features from the previous layer to enhance the contextual information of fine-scale features; performing bottom-up feature aggregation, pooling fine-scale features and fusing them with coarse-scale features from the previous layer to enhance the detail information of coarse-scale features; aggregating features from different scales based on the target point, and calculating the fusion weights of multi-scale features based on the semantic label and geometric complexity context of the target point; using the fusion weights to perform weighted summation of multi-scale features, generating global interactive fusion features and local interactive fusion features for the target point; decoupling the parameterized global error describing the overall coordinate system deviation and the local distortion displacement field describing the detailed deformation based on the global and local interactive fusion features, and generating a point-by-point correction uncertainty quantification value based on the variance estimation of the predicted points in the global and local features.

[0053] The specific configuration of the corrected point cloud data output module 30 will be described in detail below. The corrected point cloud data output module 30 further includes: applying the parameterized global error to the overall point cloud to obtain a first intermediate point cloud; superimposing the local distortion displacement field onto the first intermediate point cloud to obtain a second intermediate point cloud, and retaining the uncertainty quantization value as the confidence map for this round of optimization; based on the optimization confidence map, reducing the learning rate for high-confidence regions and using them as anchor points, increasing the learning rate for low-confidence regions and applying additional smoothing or physical constraints, and simultaneously dynamically adjusting the neighborhood optimization weights according to the current round of optimization confidence, so that the anchor point constraints propagate to the surrounding low-confidence regions; and outputting the corrected point cloud data. The second intermediate point cloud is projected onto the scene's physical constraint space to obtain a third intermediate point cloud that satisfies the constraints and semantic consistency constraints. The residual between the current iteration and the second intermediate point cloud in the previous iteration is calculated in the space before projection, and the uncertainty update value of the third intermediate point cloud is calculated. When the residual is less than a threshold and the average value of the uncertainty quantization update value is less than the target value, the convergence condition is determined to be met, and the third intermediate point cloud is output as the final corrected point cloud data. When the convergence condition is not met, the third intermediate point cloud and the updated uncertainty quantization value are used as the initial value for a new round, and the iteration continues.

[0054] The specific configuration of the corrected point cloud data output module 30 will be described in detail below. The corrected point cloud data output module 30 further includes: the scene physical constraints include: rigid constraints, planar constraints, smoothing constraints, and topological constraints; wherein, the rigid constraints ensure that points on the same semantically labeled object maintain a constant relative distance after correction; the planar constraints ensure that points identified as planes conform to planar requirements after correction; the smoothing constraints force a gradual change in the correction vectors of adjacent points in natural terrain areas; and the topological constraints ensure that linear elements with fixed topological connections maintain connectivity and continuity after correction.

[0055] The intelligent correction system for surveying point cloud data based on deep learning provided in this embodiment of the invention can execute the intelligent correction method for surveying point cloud data based on deep learning provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent correction of surveying point cloud data based on deep learning, characterized in that, include: The original point cloud data of the surveying equipment is acquired, and noise reduction and semantically guided adaptive density normalization are performed to construct standardized point cloud data. Based on the semantic labels of the standardized point cloud data and the local geometric complexity, multi-scale division is performed, and multi-scale feature sets are extracted. The multi-scale feature set is input into a deep learning model for feature learning, and different scale spatial features containing global mapping errors and local detail distortions are extracted in parallel. The different scale spatial features are interactively fused to decouple and identify parameterized global errors and local distortion displacement fields in point cloud data, and generate point-by-point correction uncertainty quantification values. The parameterized global error, local distortion displacement field, and uncertainty quantization value are used as initial values, and the scene physical constraints are incorporated to perform iterative adaptive optimization until the convergence condition is met, and the final corrected point cloud data is output.

2. The intelligent correction method for surveying point cloud data based on deep learning according to claim 1, characterized in that, Also includes: Based on the calibration process data, track error statistics, abnormal area markings, and comparison data before and after calibration to establish a calibration tracking record set; Error diagnosis and correction are performed based on the correction tracking record set, and a traceable error diagnosis report with visual charts and complete processing metadata is constructed.

3. The intelligent correction method for surveying point cloud data based on deep learning according to claim 1, characterized in that, Constructing standardized point cloud data includes: By using statistical filtering, the average k-nearest neighbor distance for each point is calculated, and outliers whose distance exceeds a preset threshold are removed. Semantic segmentation is performed on the denoised point cloud to distinguish different categories, including at least ground, buildings, and vegetation, and to obtain the semantic label of each point. Based on the semantic labels, different voxel sizes are mapped, and adaptive downsampling is performed on each voxel size to construct the density-normalized standardized point cloud data.

4. The intelligent correction method for surveying point cloud data based on deep learning according to claim 3, characterized in that, Based on the semantic labels of the standardized point cloud data and combined with local geometric complexity, multi-scale partitioning is performed, and a multi-scale feature set is extracted, including: For each point in the standardized point cloud data, calculate the eigenvalues ​​of the covariance matrix of the local neighborhood, and determine the curvature, linearity, flatness, and dispersion of the point based on the eigenvalues, which serve as a measure of local geometric complexity. Based on the semantic tags and the local geometric complexity, an adaptive octree is constructed to partition the space, resulting in a multi-level structure containing coarse-scale, medium-scale, and fine-scale elements. Based on the point cloud blocks corresponding to the multi-level structure, features at each scale are extracted to obtain the multi-scale feature set. Among them, coarse-scale segmentation prioritizes the coverage of ground semantic regions to capture global mapping errors; fine-scale segmentation targets high-curvature regions to preserve local detailed features.

5. The intelligent correction method for surveying point cloud data based on deep learning according to claim 1, characterized in that, The multi-scale feature set is input into a deep learning model for feature learning, and spatial features at different scales containing global mapping errors and local detail distortions are extracted in parallel, including: The deep learning model includes a fully connected self-attention mechanism network and a graph convolutional network, which are configured in parallel. The multi-scale feature set is input into the parallel self-attention mechanism network and graph convolutional network for feature learning. Specifically, the self-attention mechanism network learns the long-range dependencies and global context information of the point cloud to extract global spatial structure features. The graph convolutional network learns the local neighborhood geometric structure and detailed features of the point cloud based on k-neighborhood or a fixed radius. For each point, feature information is aggregated from neighboring points to obtain local detailed geometric features.

6. The intelligent correction method for surveying point cloud data based on deep learning according to claim 5, characterized in that, The deep learning model further includes a mesoscale region encoder, wherein the mesoscale region encoder is a neural network based on a local self-attention mechanism. Through the local self-attention mechanism, it calculates the association weights between each point in the point cloud block and other points in its spatial neighborhood, extracts the features of local regional structural consistency and mid-frequency continuous distortion, and outputs mesoscale spatial features.

7. The intelligent correction method for surveying point cloud data based on deep learning according to claim 6, characterized in that, By interactively fusing spatial features at different scales, the parameterized global error and local distortion displacement field in point cloud data are decoupled and identified, generating point-by-point correction uncertainty quantification values, including: Based on the scale-hierarchical structure, top-down feature propagation is performed on multi-scale spatial features, upsampling coarse-scale features and concatenating them with fine-scale features from the previous layer to enhance the contextual information of fine-scale features. Perform bottom-up feature aggregation, pooling fine-scale features and fusing them with coarse-scale features from the previous layer to enhance the detail information of coarse-scale features; Based on the target point, features from different scales are aggregated, and the fusion weights of multi-scale features are calculated based on the semantic labels and geometric complexity context of the target point. By using fusion weights, multi-scale feature weighted summation is performed to generate global interactive fusion features and local interactive fusion features of the target point, respectively. Based on the global interactive fusion features and the local interactive fusion features, the parameterized global error describing the overall coordinate system deviation and the local distortion displacement field describing the detailed deformation are decoupled and obtained respectively. Based on the variance estimation of the predicted points in the global features and local features, the point-by-point correction uncertainty quantification value is generated.

8. The intelligent correction method for surveying point cloud data based on deep learning according to claim 1, characterized in that, The parameterized global error, local distortion displacement field, and uncertainty quantization value are used as initial values, integrated with scene physical constraints, and iterative adaptive optimization is performed until the convergence condition is met, outputting the final corrected point cloud data, including: The parameterized global error is applied to the overall point cloud to obtain the first intermediate point cloud; the local distortion displacement field is superimposed on the first intermediate point cloud to obtain the second intermediate point cloud, and the uncertainty quantization value is retained as the confidence map for this round of optimization. Based on the optimized confidence map, the learning rate is reduced for high-confidence regions and used as anchor points, while the learning rate is increased for low-confidence regions and additional smoothing or physical constraints are applied. Simultaneously, the neighborhood optimization weights are dynamically adjusted according to the current round of optimization confidence, so that the anchor point constraints propagate to the surrounding low-confidence regions. The second intermediate point cloud is projected onto the scene physical constraint space to obtain a third intermediate point cloud that satisfies the constraints and semantic consistency constraints. Calculate the residual between the current iteration and the previous iteration of the second intermediate point cloud in the pre-projection space, and calculate the uncertainty update value of the third intermediate point cloud; When the residual is less than the threshold and the average value of the uncertainty quantization update value is less than the target value, the convergence condition is determined to be met, and the third intermediate point cloud is output as the final corrected point cloud data. If the convergence condition is not met, the third intermediate point cloud and the updated uncertainty quantization value are used as the initial values ​​for a new round of iteration.

9. The intelligent correction method for surveying point cloud data based on deep learning according to claim 8, characterized in that, The physical constraints of the scene include: rigid constraints, planar constraints, smooth constraints, and topological constraints. The rigid constraint condition is that points on the same semantically labeled object maintain a constant relative distance after correction; The planar constraint condition is that points identified as planes meet the planar requirements after correction; The smoothing constraint condition forces the correction vector of adjacent points to change smoothly in natural terrain areas. The topological constraint condition is to maintain the connection relationship and continuity of linear elements with a fixed topological connection structure after correction.

10. A deep learning-based intelligent correction system for surveying point cloud data, characterized in that, The system is used to implement the intelligent correction method for surveying point cloud data based on deep learning as described in any one of claims 1 to 9, and the system comprises: The point cloud data standardization module is used to acquire the raw point cloud data of the surveying equipment, perform noise reduction and semantically guided adaptive density normalization processing, construct standardized point cloud data, and perform multi-scale partitioning based on the semantic labels of the standardized point cloud data and local geometric complexity, and extract multi-scale feature sets. The spatial feature extraction module is used to input the multi-scale feature set into the deep learning model for feature learning, extract spatial features of different scales containing global mapping errors and local detail distortions in parallel, and interactively fuse spatial features of different scales to decouple and identify parameterized global errors and local distortion displacement fields in point cloud data, and generate point-by-point correction uncertainty quantification values. The corrected point cloud data output module is used to take the parameterized global error, local distortion displacement field and uncertainty quantization value as initial values, integrate them with scene physical constraints, perform iterative adaptive optimization until the convergence condition is met, and output the final corrected point cloud data.

Images