A method for point cloud completion by airborne depth sounding radar

Abstract

This invention discloses a point cloud completion method for airborne depth sounding radar, belonging to the field of lidar data processing technology. It is used for point cloud completion in airborne depth sounding radar, including acquiring raw point cloud data, performing tile division and coordinate transformation, extracting ground points and constructing a terrain-normalized point cloud; calculating local point clusters based on the neighborhood density of the raw points, obtaining spatial location and attribute information of the point clusters, constructing a surrogate point set, and constructing an input vector and initial feature vector based on the surrogate points; inputting the surrogate point set into a neural network encoder to generate fine-grained completion points and their confidence levels, calculating the fusion weights of the fine-grained completion points, removing redundant points to restore the original coordinates, and outputting the complete point cloud completion result. This invention achieves highly robust completion of complex terrain and uneven density regions in point clouds through a dynamic completion strategy guided by adaptive voxel clustering, terrain normalization, and missing data scoring, significantly improving the geometric consistency and detail restoration accuracy of the completed points.

Description

Technical Field

[0001] This invention discloses a method for point cloud completion of airborne depth sounding radar, belonging to the field of lidar data processing technology. Background Technology

[0002] Airborne lidar point clouds are characterized by large scenes, realistic coordinates, significant terrain undulations, uneven point density, multiple echoes, and rich intensity attributes. Their completion tasks require the reasonable restoration of missing areas while maintaining consistency with realistic terrain, elevation benchmarks, and spatial coordinates. Existing point cloud completion methods are mainly divided into traditional geometric completion methods and deep learning-based completion methods. Traditional geometric completion methods detect missing boundaries and fill gaps using interpolation, surface fitting, triangulation, or implicit surface reconstruction. While effective on simple point clouds, they are prone to surface collapse and distortion for large-area missing areas or complex surfaces. Furthermore, they rely on the assumption of local geometric continuity, making it difficult to handle issues such as mixed features, large-scale tile formation, and cross-tile consistency in airborne point clouds.

[0003] In recent years, completion methods based on encoder-decoder, generative networks, or Transformers have enhanced feature representation through attention mechanisms, achieving end-to-end learning from incomplete point clouds to complete point clouds. However, existing methods are mainly designed for general 3D object point clouds, without considering the characteristics of airborne point clouds such as true projected coordinates, terrain undulations, flight strip differences, multi-echo attributes, and extremely uneven density, and also lack specialized processing for cross-tile continuity in large scenes. Summary of the Invention

[0004] The purpose of this invention is to provide a method for point cloud completion of airborne depth sounding radar, in order to solve the problem in the prior art that it is difficult to directly use an object-level point cloud completion model for efficient completion of airborne point clouds under conditions such as real coordinate system, complex terrain, uneven point density at different elevations, flight zone differences, echo attributes, local occlusion, tiled storage, and large-scale computing.

[0005] A method for completing point clouds of an airborne depth sounding radar includes: S1. Obtain the original point cloud data, divide the original point cloud data into several tiles, convert the original point coordinates into tile coordinates, extract the ground points of the tiles and construct a ground elevation model, convert the elevation of the original points into the height relative to the ground based on the ground elevation model, and construct a terrain-normalized point cloud. S2. Calculate the neighborhood density of the original points and perform adaptive voxel clustering to obtain local point clusters; calculate the spatial location information of the surrogate points through a learnable weighted average and extract the attribute information of the local point clusters to form a surrogate point set; construct the input vector based on the spatial location information and attribute information of the surrogate points, and construct the initial feature vector of the surrogate points through multilayer perceptron mapping, learnable embedding table and position encoding. S3. Input the surrogate point set into the neural network encoder, update the surrogate point feature vector through cross-attention calculation in the encoder and construct the fused feature set; generate query points based on the missing score of each surrogate point, perform cross-attention calculation on the query points and the fused feature set, obtain fine-grained completion points through coordinate prediction and residual refinement, and predict the confidence of fine-grained completion points. S4. Calculate the fusion weight based on the confidence of the fine completion points, the center area weight, the density consistency weight, and the adjacent consistency weight. Remove redundant points by setting the target neighborhood density and confidence filtering threshold, restore the remaining completion points to the original coordinate system, and output the complete point cloud completion result.

[0006] S1 includes, S1.1, acquiring raw point cloud data, based on the raw point cloud data. Construct the original point set: ; ; ; In the formula, , , for Spatial location information, for Attribute information, The original reflection intensity, This is the current echo number. The scanning angle; right Establish metadata : ; In the formula, As a coordinate reference system, A unique identifier for point cloud files. Assign a number to the corresponding flight strip. For the time of collection, Spatial range; The original point cloud data is divided into several tiles, and the origin of each tile is selected. , , , and for The three-dimensional coordinates, based on Will Converting the 3D coordinates to tile coordinates yields the tile points. , , , and for The three-dimensional coordinates are used to store tile metadata. : ; In the formula, Number the tiles. This is a list of IDs for adjacent tiles.

[0007] S1 includes S1.2, which uses cloth simulation filtering to extract the set of ground points of the tiles. ,based on The ground point in the middle is Add ground point markers; Using ground point sets Construct an irregular triangular network ground model to obtain the ground elevation function. Set neighborhood distance ,according to Fit the neighborhood point set to the plane to obtain Interpolated elevation obtained by interpolation through irregular triangular mesh And perform terrain normalization: ; In the formula, for The original elevation, for Relative to the height of the ground; Constructing terrain normalization points : .

[0008] S2 includes S2.1, point-to-point. In radius Internal calculation of neighborhood density ,according to Determine the current tile voxel size and based on Calculate voxel index Points with the same or adjacent voxel indices are aggregated into local point clusters. , For indexes of local point clusters; right Computational Agent Center : ; ; In the formula, As weight, For local edge intensity, For point quality scoring, For the Sigmoid function, , and These are learnable scalar parameters; right Calculate the covariance matrix ,right Perform eigenvalue decomposition and construct based on eigenvalues linearity Flatness Dispersion and roughness ; Constructing proxy point attribute features : ; In the formula, for The mean height of the interior point relative to the ground. for The maximum height of the interior point relative to the ground. for The variance of the internal height relative to the ground. for The mean value of the inward reflection intensity. for The variance of the inward reflection intensity; Agent point is represented as , All agent points within a tile constitute a set. , , It refers to the number of agent points.

[0009] S2 includes, S2.2, for each Construct the input vector : ; In the formula, for The slope at the center of mass, for Curvature at the center of mass; Will Input to a multilayer perceptron, and obtain Comprehensive attribute embedding Then, the density is graded. : ; In the formula, For low-level thresholds, High-level threshold; Mapping discrete levels to vectors using learnable tables , build initial feature vector : ; In the formula, For position encoding vectors, Density level.

[0010] S3 includes, S3.1, and Input a sparse Transformer encoder, for Establish local neighborhood : ; In the formula, yes Recent One agency point, for The index; Calculate attention: ; In the formula, For the index of the network layer, For the first In-layer proxy point The updated feature vector, For query vector, For key vectors, For value vectors, For query vector dimensions, For geometric position offset, This is the normalization function.

[0011] S3 includes S3.2, constructing tiles. Adjacent tile set , Construct a set of feature vectors for all proxy points in adjacent tiles. , ,from Memory tiles obtained from the pool : ; In the formula, for The Middle The updated feature vector of the surrogate encoder The number of adjacent tiles. For pooling operations, For max pooling along the feature dimension, For average pooling, Pooling for bird's-eye view; Cross-attention is applied to the current tile and the memory tile to obtain new features after fusion of memories. : ; In the formula, This is a cross-attention operation; according to Constructing a fusion feature set , .

[0012] S3 includes S3.3, calculating the missing score for each agent point. : ; In the formula, , , , and These are learnable weight coefficients. This is a density missing term. It is a marginal mutation term. For scan angle or flight strip anomalies, For echo anomalies, For the tile boundary uncertainty term; Calculate the number of query points generated by each proxy point. : ; In the formula, Minimum number of query points generated per proxy point The maximum number of query points that can be generated per proxy point To round down; Construct a query point set based on the initial position and characteristics of each query point. : ; In the formula, To query the serial number, For the first The first agent point generated the first One query point; and Perform cross-attention, and obtain the coarse-completed full point offset from the coordinate prediction head output. : ; In the formula, For the index of the query point, For the first The initial three-dimensional position of each query point Multilayer perceptron for coordinate prediction; right Extract surrogate point features and fuse them to obtain local context. Calculate prediction residuals to obtain refined completion points. : ; In the formula, For residual prediction, multilayer perceptron, For splicing operations, For the first The three-dimensional position of each agent point The neighborhood search radius; For each Predict the confidence level of fine-grained completion points : ; In the formula, To predict the confidence level of a multilayer perceptron. For the first Local point density of agent points.

[0013] S4 includes, S4.1, for tiles Fine-grained completion points Calculate the fusion weights : ; In the formula, For the network's predicted output, Weight for the central area For density consistency, For adjacent consistency; Set cluster radius threshold Distance less than Establish fusion clusters from candidate points: ; In the formula, The center of the cluster; Fusion point for: ; Fusion confidence for: ; Perform deduplication and filtering, including setting confidence filtering thresholds. and target neighborhood density ,calculate neighborhood density ,like Continue with confidence filtering, if ,delete ; Output the deduplicated and filtered merged point cloud.

[0014] S4 includes S4.2, which involves fusing points in the deduplicated and filtered point cloud. , Restored to Plane coordinates and absolute elevation : ; ; ; In the formula, , for Two-dimensional coordinates, for The relative height to the ground; based on Output the data structure of the final completion point : ; In the formula, for The intensity of reflection, for echo number, for Confidence level, for The source file identifier; All the tiles After merging, output the complete point cloud completion result.

[0015] Compared with existing technologies, this invention has the following advantages: It converts absolute elevation to relative ground height through terrain normalization, eliminating the interference of terrain undulations on the geometric features of point clouds, allowing subsequent processing to focus on the true height of ground features; through density-adaptive voxel clustering and learnable weighted surrogate point representation, it effectively addresses the problems of uneven density and variable ground feature scale in airborne point clouds, improving feature representation capabilities in complex scenes; through dynamic generation of query point numbers guided by missing score, it allocates more queries to severely missing areas to strengthen completion, and reduces computational redundancy in sufficiently observed areas, achieving a balance between efficiency and accuracy; through residual refinement modules and cross-tile attention mechanisms, it ensures that completed points fit local surface details and maintain geometric consistency between tiles; through multi-factor fusion weights (confidence, center area, density consistency, adjacency consistency) and density-constrained filtering, it effectively removes redundant and low-quality points, ultimately outputting a completed point cloud strictly aligned with the original coordinate system. Therefore, compared with the prior art, the present invention significantly improves the robustness of the completion capability for complex terrain, uneven density and large scene cross-tile area in airborne depth sounding radar point cloud, while maintaining the integrity of ground structure and consistency of spatial coordinates. Attached Figure Description

[0016] Figure 1 This is a flowchart of the technology of this invention; Figure 2 This is a system architecture diagram of the present invention; Figure 3 This is a point cloud image that is missing in an embodiment of the present invention; Figure 4 This is a point cloud image after point cloud completion processing according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] A method for completing point clouds of an airborne depth sounding radar includes: S1. Obtain the original point cloud data, divide the original point cloud data into several tiles, convert the original point coordinates into tile coordinates, extract the ground points of the tiles and construct a ground elevation model, convert the elevation of the original points into the height relative to the ground based on the ground elevation model, and construct a terrain-normalized point cloud. S2. Calculate the neighborhood density of the original points and perform adaptive voxel clustering to obtain local point clusters; calculate the spatial location information of the surrogate points through a learnable weighted average and extract the attribute information of the local point clusters to form a surrogate point set; construct the input vector based on the spatial location information and attribute information of the surrogate points, and construct the initial feature vector of the surrogate points through multilayer perceptron mapping, learnable embedding table and position encoding. S3. Input the surrogate point set into the neural network encoder, update the surrogate point feature vector through cross-attention calculation in the encoder and construct the fused feature set; generate query points based on the missing score of each surrogate point, perform cross-attention calculation on the query points and the fused feature set, obtain fine-grained completion points through coordinate prediction and residual refinement, and predict the confidence of fine-grained completion points. S4. Calculate the fusion weight based on the confidence of the fine completion points, the center area weight, the density consistency weight, and the adjacent consistency weight. Remove redundant points by setting the target neighborhood density and confidence filtering threshold, restore the remaining completion points to the original coordinate system, and output the complete point cloud completion result.

[0019] S1 includes, S1.1, acquiring raw point cloud data, based on the raw point cloud data. Construct the original point set: ; ; ; In the formula, , , for Spatial location information, for Attribute information, The original reflection intensity, This is the current echo number. The scanning angle; right Establish metadata : ; In the formula, As a coordinate reference system, A unique identifier for point cloud files. Assign a number to the corresponding flight strip. For the time of collection, Spatial range; The original point cloud data is divided into several tiles, and the origin of each tile is selected. , , , and for The three-dimensional coordinates, based on Will Converting the 3D coordinates to tile coordinates yields the tile points. , , , and for The three-dimensional coordinates are used to store tile metadata. : ; In the formula, Number the tiles. This is a list of IDs for adjacent tiles.

[0020] S1 includes S1.2, which uses cloth simulation filtering to extract the set of ground points of the tiles. ,based on The ground point in the middle is Add ground point markers; Using ground point sets Construct an irregular triangular network ground model to obtain the ground elevation function. Set neighborhood distance ,according to Fit the neighborhood point set to the plane to obtain Interpolated elevation obtained by interpolation through irregular triangular mesh And perform terrain normalization: ; In the formula, for The original elevation, for Relative to the height of the ground; Constructing terrain normalization points : .

[0021] S2 includes S2.1, point-to-point. In radius Internal calculation of neighborhood density ,according to Determine the current tile voxel size and based on Calculate voxel index Points with the same or adjacent voxel indices are aggregated into local point clusters. , For indexes of local point clusters; right Computational Agent Center : ; ; In the formula, As weight, For local edge intensity, For point quality scoring, For the Sigmoid function, , and These are learnable scalar parameters; right Calculate the covariance matrix ,right Perform eigenvalue decomposition and construct based on eigenvalues linearity Flatness Dispersion and roughness ; Constructing proxy point attribute features : ; In the formula, for The mean height of the interior point relative to the ground. for The maximum height of the interior point relative to the ground. for The variance of the internal height relative to the ground. for The mean value of the inward reflection intensity. for The variance of the inward reflection intensity; Agent point is represented as , All agent points within a tile constitute a set. , , It refers to the number of agent points.

[0022] S2 includes, S2.2, for each Construct the input vector : ; In the formula, for The slope at the center of mass, for Curvature at the center of mass; Will Input to a multilayer perceptron, and obtain Comprehensive attribute embedding Then, the density is graded. : ; In the formula, For low-level thresholds, High-level threshold; Mapping discrete levels to vectors using learnable tables , build initial feature vector : ; In the formula, For position encoding vectors, Density level.

[0023] S3 includes, S3.1, and Input a sparse Transformer encoder, for Establish local neighborhood : ; In the formula, yes Recent One agency point, for The index; Calculate attention: ; In the formula, For the index of the network layer, For the first In-layer proxy point The updated feature vector, For query vector, For key vectors, For value vectors, For query vector dimensions, For geometric position offset, This is the normalization function.

[0024] S3 includes S3.2, constructing tiles. Adjacent tile set , Construct a set of feature vectors for all proxy points in adjacent tiles. , ,from Memory tiles obtained from the pool : ; In the formula, for The Middle The updated feature vector of the surrogate encoder The number of adjacent tiles. For pooling operations, For max pooling along the feature dimension, For average pooling, Pooling for bird's-eye view; Cross-attention is applied to the current tile and the memory tile to obtain new features after fusion of memories. : ; In the formula, This is a cross-attention operation; according to Constructing a fusion feature set , .

[0025] S3 includes S3.3, calculating the missing score for each agent point. : ; In the formula, , , , and These are learnable weight coefficients. This is a density missing term. It is a marginal mutation term. For scan angle or flight strip anomalies, For echo anomalies, For the tile boundary uncertainty term; Calculate the number of query points generated by each proxy point. : ; In the formula, Minimum number of query points generated per proxy point The maximum number of query points that can be generated per proxy point To round down; Construct a query point set based on the initial position and characteristics of each query point. : ; In the formula, To query the serial number, For the first The first agent point generated the first One query point; and Perform cross-attention, and obtain the coarse-completed full point offset from the coordinate prediction head output. : ; In the formula, For the index of the query point, For the first The initial three-dimensional position of each query point Multilayer perceptron for coordinate prediction; right Extract surrogate point features and fuse them to obtain local context. Calculate prediction residuals to obtain refined completion points. : ; In the formula, For residual prediction, multilayer perceptron, For splicing operations, For the first The three-dimensional position of each agent point The neighborhood search radius; For each Predict the confidence level of fine-grained completion points : ; In the formula, To predict the confidence level of a multilayer perceptron. For the first Local point density of agent points.

[0026] S4 includes, S4.1, for tiles Fine-grained completion points Calculate the fusion weights : ; In the formula, For the network's predicted output, Weight for the central area For density consistency, For adjacent consistency; Set cluster radius threshold Distance less than Establish fusion clusters from candidate points: ; In the formula, The center of the cluster; Fusion point for: ; Fusion confidence for: ; Perform deduplication and filtering, including setting confidence filtering thresholds. and target neighborhood density ,calculate neighborhood density ,like Continue with confidence filtering, if ,delete ; Output the deduplicated and filtered merged point cloud.

[0027] S4 includes S4.2, which involves fusing points in the deduplicated and filtered point cloud. , Restored to Plane coordinates and absolute elevation : ; ; ; In the formula, , for Two-dimensional coordinates, for The relative height to the ground; based on Output the data structure of the final completion point : ; In the formula, for The intensity of reflection, for echo number, for Confidence level, for The source file identifier; All the tiles After merging, output the complete point cloud completion result.

[0028] The first step of this invention is airborne point cloud data acquisition and attribute organization, including acquiring the original airborne radar point cloud of the area to be completed and organizing the point coordinates and airborne attributes into network input data. The original point cloud data is read from the point cloud file. During reading, attribute fields such as 3D coordinates, reflection intensity, echo sequence number, total echo count, scan angle, classification label, and flight strip number are extracted for each point. After reading, all points are saved according to a unified data structure, with each point corresponding to a set of coordinate information and a set of attribute information. The data reading process does not only extract 3D coordinates but also retains attributes such as intensity, echo, flight strip, and scan angle. This allows the subsequent model to distinguish whether the shortage of points is due to actual holes or insufficient sampling caused by scan angle, vegetation obstruction, flight strip edges, or low reflectivity. Metadata records are established for the input point cloud. Metadata includes point cloud file name, coordinate reference system, spatial range, number of points, and acquisition time. Independent metadata is established for each file, and a global index is created to record the spatial position of each file in the overall survey area. All subsequent tiles, completion points, and fusion results are associated with this metadata.

[0029] The second step of this invention involves CRS management and local coordinate translation. Airborne point clouds typically use true projected coordinates, and position encoding is dominated by absolute coordinates, affecting the model's learning of local geometry. Local coordinate translation is required before inputting the data into the network. The large scene point cloud to be processed is divided into spatial processing regions, and a local origin is selected for each region. The global coordinates of all points within this region are subtracted from the local origin coordinates to obtain the local coordinates. The local origin is saved separately and does not participate in network feature learning. Each tile saves its own local origin, tile number, file number, and coordinate offset information. Subsequent model inputs only use local coordinates, and the output results are then restored to global coordinates based on the saved local origins.

[0030] The third step of this invention is ground extraction and terrain normalization, which includes first extracting ground points from the local point cloud. A cloth simulation filter is used to extract ground points. After extraction, ground points and non-ground points are saved separately, and a ground point identifier is added to each point. ; ; In the formula, The total number of ground points, For ground point indexing, , , , For the first The three-dimensional coordinates of a ground point For the first Attribute information of each ground point; A ground elevation model is constructed using extracted ground points, forming an irregular triangular mesh ground model. For sparse areas of ground points, interpolation is performed using neighboring points to complete the model. For areas with abrupt terrain changes, local topographical variations are preserved without excessive smoothing. After construction, the ground elevation model is bound and saved to the current tile or processing area; local plane fitting is used to obtain the final elevation model. At point Nearby ground neighborhood points are selected, and the fit is performed: ; but: ; Airborne point clouds are situated in realistic terrain environments, where the absolute elevations vary significantly across mountainous areas, slopes, and built-up areas. Normalization using ordinary unit spheres would distort these true height relationships. This patented solution employs terrain normalization, converting absolute elevations into relative ground heights (HAG).

[0031] The tile division process of this invention includes setting the tile side length to be... The overlap width is 0, and the sliding window step size is 0. , No. The spatial range of each tile is: ; ; The tile point set is: ; Each tile preserves: ; Set a mask in the center area of ​​the tile for: ; In the formula, express Distance to the edge of the tile.

[0032] The fourth step of this invention involves density estimation and adaptive voxel partitioning, including dynamically adjusting the sampling scale based on local density and performing spatial neighborhood statistics on points within each tile. Voxels are used to count the number of points near each location. The statistical results are saved as a point density field. Local density is calculated separately for different types of regions. If a point belongs to multiple overlapping tiles, its local density record in each tile is calculated separately, and the density results are subsequently used in surrogate point generation and query completion generation. Point In radius Internal count of neighboring points: ; ; according to Determine the voxel size of the current region : ; In the formula, This is the scaling factor. To prevent division by zero of extremely small constants, For the smallest voxel size, For the maximum voxel size, This is a truncation function; Calculate voxel index : ; Points with the same or adjacent voxel indices are aggregated into local point clusters. : ; In the formula, It is an index for a local point cluster.

[0033] The fifth step of this invention is the generation of density-adaptive surrogate points, which involves clustering each local point into a single point cluster. Aggregates data into surrogate points, which are compressed representations of local geometry, density, terrain, and airborne attributes. A surrogate point is generated for each effective voxel or local point cluster. The location of the surrogate point is determined by the weighted center of all points within that cluster. The attributes of the surrogate point consist of height statistics, density statistics, intensity statistics, echo statistics, and geometric morphology statistics within the cluster. Different features are recorded for surrogate points in different regions, ultimately compressing a large number of original points within a tile into a smaller set of surrogate points. Information from each surrogate point is concatenated to form surrogate point features, which are then converted into fixed-length high-dimensional feature vectors using a multilayer perceptron. Attributes from different sources are encoded separately and then merged. The merged surrogate point features serve as input to the subsequent Transformer encoder; for point clusters... Calculate the covariance matrix: ; In the formula, It is the transpose symbol; right Eigenvalue decomposition yields: ; Linearity calculation: ; Flatness calculation: ; Scattering calculation: ; Roughness calculation: ; calculate echo entropy: ; In the formula, It is the first echo-like The proportion in; Based on the above information, construct the attribute features of the proxy points and output the set of proxy points.

[0034] The sixth step of this invention is terrain-density joint feature embedding, which includes converting the location and attributes of surrogate points into high-dimensional feature vectors that can be processed by a neural network. The seventh step is sparse Transformer encoding, which includes modeling the spatial relationships between surrogate points and inputting the surrogate points into a sparse Transformer encoder. Before encoding, a local neighborhood index is established for each surrogate point. During encoding, each surrogate point only performs attention calculations with surrogate points within its local neighborhood. For distant regions, only a small number of representative surrogate points are selected to participate in relationship modeling. The encoder can use local window attention, graph attention, or hierarchical sparse attention. After each layer of encoding, the surrogate point features are fused with the spatial structure information of the surrounding neighborhood; the geometric position bias is: ; In the formula, For the first The spatial location vector of each proxy point For the first The relative ground height of each agency point For the first Point density of agent points.

[0035] The eighth step of this invention is cross-tile memory fusion, which includes extracting summary features from neighboring tiles of the current tile to form cross-tile memory tokens (proxy points). Neighboring tile summaries can be obtained by pooling adjacent tile proxy points. The proxy points encoded in the current tile interact with these memory tokens. After the interaction, proxy points near the current tile boundary obtain the context information of adjacent tiles. The fused proxy point features enter the query completion generation module. The cross-tile memory token fusion layer is located after local sparse Transformer encoding and before query point generation. It does not directly generate points but supplements the context outside the current tile boundary. This module can reduce boundary misalignment, duplicate point supplementation, and structural discontinuity problems caused by "each tile supplementing its own context."

[0036] The ninth step of this invention is the calculation of missing scores and adaptive query generation. Instead of uniformly generating completion points across all regions, it determines where missing data is more likely and assigns more queries to those regions; the density of missing data is: ; The marginal mutation term is: ; If a certain area has low density, strong HAG mutation, and obvious echo anomalies, then Larger values ​​generate more queries; if a region is well-observed and has flat terrain, then... Smaller size, generates fewer queries; The initial position of each query is: ; in Generated from learnable offsets: ; in It is the first Learnable encoding of a query; Query feature initialization: ; The final query set is: .

[0037] The tenth step of this invention involves coarse completion decoding, residual refinement decoding, and confidence completion. This includes generating first-stage completion points based on the encoded surrogate point token and completion query. The coarse completion points only provide approximate locations, while the residual refinement module further generates local details. The residual refinement module ensures that the completion points fit the local surface, preventing coarse points from being too sparse or uneven. In airborne point cloud completion, not all missing regions have definite ground truth values, so each completion point needs to output a confidence score for subsequent fusion.

[0038] The eleventh step of this invention is the fusion and deduplication of confidence scores in overlapping regions, including fusion weight calculation, spatial clustering fusion, and deduplication and filtering, wherein the weight of the central region is defined as: ; in It is the distance from the point to the edge of the tile; Density consistency weight: ; Adjacent tile consistency weight: .

[0039] The twelfth step of this invention is to restore and output the global coordinates, and finally restore the local coordinates and HAG height to the true CRS coordinates, and finally output the complete large-scale airborne point cloud completion result.

[0040] The following description, in conjunction with the accompanying drawings, further illustrates the method flow of this invention. Figure 1 As shown, the original point cloud of the airborne LiDAR is acquired, then normalized, and after normalization, overlapping tiles are divided, followed by completion processing, including terrain-density feature encoding, adaptive query generation, sparse Transformer completion encoding, and density adaptive proxy generation. After completion processing, coarse completion and residual refinement are performed, followed by overlapping area fusion and global back-scraping, and finally large-scale point cloud completion output.

[0041] The present invention also includes an airborne depth sounding radar point cloud completion system, such as Figure 2 As shown, the system includes a point cloud data storage unit, a preprocessing unit, a tile organization unit, a surrogate point generation unit, a completion network inference unit, and a post-processing and output unit. The preprocessing unit includes a coordinate transformation module, a ground extraction module, and attribute standardization. The tile organization unit includes a sliding window block segmentation module, an overlap region marking module, and an adjacency index. The surrogate point generation unit includes a density estimation module, an adaptive voxel module, and a token aggregation module. The completion network inference unit includes a sparse encoder, a query generator, and a decoder refiner. The post-processing and output unit includes a confidence fusion module, a deduplication filtering module, and point cloud reconstruction.

[0042] The parameters in this embodiment of the invention are only set for this point cloud map. When different point cloud maps are input, parameter adjustments are required based on the characteristics of the point cloud map. Partial radar pose data includes time, longitude, latitude, elevation, pitch angle, roll angle, and heading angle. Point cloud completion is performed using the scan results of a certain area, with 402,440 points and a planar range of approximately... The average rectangular density is approximately 71 points per square meter. For this point cloud file, tiles are divided according to the XY plane, with no separate segmentation in the Z direction. The segmentation parameters include the tile scale. Overlap width sliding step size The division is along the X and Y planes, with 8 tiles in 2 columns and 4 rows; tile information is shown in Table 1: Table 1. Tile Information ; During point cloud completion, a maximum of 256 surrogate points are generated within each tile. For any surrogate point, its 32 nearest surrogate points within the same tile are searched based on local coordinate distance, forming its local neighborhood, with a size of 32. If the total number of surrogate points within a tile is less than 32, then all surrogate points within that tile except for itself are used as its local neighborhood. Each tile has a maximum of 256 surrogate points, and each surrogate point's 32 nearest surrogate points are used as its local neighborhood. The number of query points generated for each proxy point is set to a range of 0 to 2. =0, =2. When the point cloud corresponding to a certain proxy point is complete and the density is normal, no query point is generated; when the area has low density, occlusion, or missing edges, 1 to 2 query points are generated. A single tile can retain a maximum of 256 proxy points, so each tile can generate a maximum of 512 query points.

[0043] For each coarse completion point, the features of surrogate points within a 3m range are searched and pooled. When there are insufficient surrogate points in the neighborhood, the 8 nearest surrogate points are added; when there are too many surrogate points in the neighborhood, a maximum of 32 nearest surrogate points are retained. Cluster radius threshold. The threshold is set to 0.2m. This means that within an overlapping tile region, if the spatial distance between two or more candidate completion points is less than 0.2m, these points are considered to correspond to the same local spatial location, are assigned to the same fusion cluster, and the fusion point location and fusion confidence are calculated based on the fusion weights of each point. (Confidence filtering threshold) The value is set to 0.6. In this embodiment of the model, the average planar density of the original point cloud is approximately 71 points / m², and the expected average planar density of the point cloud after completion is approximately 180 points / m². Therefore, the target neighborhood density is set to 180 points / m². The completion result is as follows: Figure 3 and Figure 4 As shown, Figure 3 The ground points are relatively sparse, and ground objects are blurry. After completing the point cloud using the method of this invention, Figure 4 Both the ground and ground objects have been effectively completed, and the shapes of the ground objects are more prominent and obvious.

[0044] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for completing point clouds of an airborne depth sounding radar, characterized in that, include: S1. Obtain the original point cloud data, divide the original point cloud data into several tiles, convert the original point coordinates into tile coordinates, extract the ground points of the tiles and construct a ground elevation model, convert the elevation of the original points into the height relative to the ground based on the ground elevation model, and construct a terrain-normalized point cloud. S2. Calculate the neighborhood density of the original points and perform adaptive voxel clustering to obtain local point clusters; calculate the spatial location information of the surrogate points through a learnable weighted average and extract the attribute information of the local point clusters to form a surrogate point set; construct the input vector based on the spatial location information and attribute information of the surrogate points, and construct the initial feature vector of the surrogate points through multilayer perceptron mapping, learnable embedding table and position encoding. S3. Input the surrogate point set into the neural network encoder, calculate and update the surrogate point feature vectors through cross attention in the encoder, and construct a fused feature set; Query points are generated based on the missing scores of each proxy point. The query points are then cross-attentioned with the fused feature set. Fine-grained completion points are obtained through coordinate prediction and residual refinement, and the confidence of the fine-grained completion points is predicted. S4. Calculate the fusion weight based on the confidence, center area weight, density consistency weight, and adjacent consistency weight of the fine completion points. Remove redundant points by setting the target neighborhood density and confidence filtering threshold, restore the remaining completion points to the original coordinate system, and output the complete point cloud completion result. S3 includes, S3.1, and Input a sparse Transformer encoder, for Establish local neighborhood : ； In the formula, yes Recent One agency point, for The index; Calculate attention: ； In the formula, For the index of the network layer, For the first In-layer proxy point The updated feature vector, For query vector, For key vectors, For value vectors, For query vector dimensions, For geometric position offset, This is the normalization function; S3 includes S3.2, constructing tiles. Adjacent tile set , Construct a set of feature vectors for all proxy points in adjacent tiles. , ,from Memory tiles obtained from the pool : ； In the formula, for The Middle The updated feature vector of the surrogate encoder The number of adjacent tiles. For pooling operations, For max pooling along the feature dimension, For average pooling, Pooling for bird's-eye view; Cross-attention is applied to the current tile and the memory tile to obtain new features after fusion of memories. : ； In the formula, This is a cross-attention operation; according to Constructing a fusion feature set , ; S3 includes S3.3, calculating the missing score for each agent point. : ； In the formula, , , , and These are learnable weight coefficients. This is a density missing term. It is a marginal mutation term. For scan angle or flight strip anomalies, For echo anomalies, For the tile boundary uncertainty term; Calculate the number of query points generated by each proxy point. : ； In the formula, Minimum number of query points generated per proxy point The maximum number of query points that can be generated per proxy point To round down; Construct a query point set based on the initial position and characteristics of each query point. : ； In the formula, To query the serial number, For the first The first agent point generated the first One query point; and Perform cross-attention, and obtain the coarse-completed full point offset from the coordinate prediction head output. : ； In the formula, For the index of the query point, For the first The initial three-dimensional position of each query point Multilayer perceptron for coordinate prediction; right Extract surrogate point features and fuse them to obtain local context. Calculate prediction residuals to obtain refined completion points. : ； In the formula, For residual prediction, multilayer perceptron, For splicing operations, For the first The three-dimensional position of each agent point The neighborhood search radius; For each Predicting the confidence level of fine-grained completion points : ； In the formula, To predict the confidence level of a multilayer perceptron. For the first Local point density of agent points.

2. The method for completing point clouds of airborne depth sounding radar according to claim 1, characterized in that, S1 includes, S1.1, acquiring raw point cloud data, based on the raw point cloud data. Construct the original point set: ； ； ； In the formula, , , for Spatial location information, for Attribute information, The original reflection intensity, This is the current echo sequence number. The scanning angle; right Establish metadata : ； In the formula, As a coordinate reference system, A unique identifier for point cloud files. Assign a number to the corresponding flight strip. For the time of collection, Spatial range; The original point cloud data is divided into several tiles, and the origin of each tile is selected. , , , and for The three-dimensional coordinates, based on Will Converting the 3D coordinates to tile coordinates yields the tile points. , , , and for The three-dimensional coordinates are used to store tile metadata. : ； In the formula, Number the tiles. This is a list of IDs for adjacent tiles.

3. The method for completing point clouds of airborne depth sounding radar according to claim 2, characterized in that, S1 includes S1.2, which uses cloth simulation filtering to extract the set of ground points of the tiles. ,based on The ground point in the middle is Add ground point markers; Using ground point sets Construct an irregular triangular network ground model to obtain the ground elevation function. Set neighborhood distance ,according to Fit the neighborhood point set to the plane to obtain Interpolated elevation obtained by interpolation through irregular triangular mesh And perform terrain normalization: ； In the formula, for The original elevation, for Relative to the height of the ground; Constructing terrain normalization points : 。 4. The method for completing point clouds of airborne depth sounding radar according to claim 3, characterized in that, S2 includes S2.1, point-to-point. In radius Internal calculation of neighborhood density ,according to Determine the current tile voxel size and based on Calculate voxel index Points with the same or adjacent voxel indices are aggregated into local point clusters. , For indexes of local point clusters; right Computational Agent Center : ； ； In the formula, As weight, For local edge intensity, For point quality scoring, For the Sigmoid function, , and These are learnable scalar parameters; right Calculate the covariance matrix ,right Perform eigenvalue decomposition and construct based on eigenvalues linearity Flatness Dispersion and roughness ; Constructing proxy point attribute features : ； In the formula, for The mean height of the interior point relative to the ground. for The maximum height of the interior point relative to the ground. for The variance of the internal height relative to the ground. for The mean value of the inward reflection intensity. for The variance of the inward reflection intensity; Agent point is represented as , All agent points within a tile constitute a set. , , It refers to the number of agent points.

5. The method for completing point clouds of airborne depth sounding radar according to claim 4, characterized in that, S2 includes, S2.2, for each Construct the input vector : ； In the formula, for The slope at the center of mass, for Curvature at the center of mass; Will Input to a multilayer perceptron, and obtain Comprehensive attribute embedding Then, the density is graded. : ； In the formula, For low-level thresholds, High-level threshold; Mapping discrete levels to vectors using learnable tables , build initial feature vector : ； In the formula, For position encoding vectors, Density level.

6. The method for completing point clouds of airborne depth sounding radar according to claim 5, characterized in that, S4 includes, S4.1, for tiles Fine-grained completion points Calculate the fusion weights : ； In the formula, For the network's predicted output, Weight for the central area For density consistency, For adjacent consistency; Set cluster radius threshold Distance less than Establish fusion clusters from candidate points: ； In the formula, The center of the cluster; Fusion point for: ； Fusion confidence for: ； Perform deduplication and filtering, including setting confidence filtering thresholds. and target neighborhood density ,calculate neighborhood density ,like Continue with confidence filtering, if ,delete ; Output the deduplicated and filtered merged point cloud.

7. The method for completing point clouds of airborne depth sounding radar according to claim 6, characterized in that, S4 includes S4.2, which involves fusing points in the deduplicated and filtered point cloud. , Restored to Plane coordinates and absolute elevation : ； ； ； In the formula, , for Two-dimensional coordinates, for The relative height to the ground; based on Output the data structure of the final completion point : ； In the formula, for The intensity of reflection, for echo number, for Confidence level, for The source file identifier; All the tiles After merging, output the complete point cloud completion result.

Images