A ground penetrating radar inversion voxel sub-graph and ground surface scene integrated registration fusion reconstruction method based on high-precision pose skeleton constraint

Abstract

This invention discloses an integrated registration, fusion, and reconstruction method for ground-penetrating radar inversion voxel sub-maps and surface scenes based on high-precision pose skeleton constraints. Belonging to the field of remote sensing and mapping, the method includes: acquiring a high-precision global pose skeleton through multi-source sensing and SLAM technology, and generating surface sparse geometric voxels and subsurface inversion property sub-maps accordingly; proposing an elastic anchor point mechanism to dynamically attach the subsurface rigid sub-map to pose nodes, forming a master-slave relationship; then, constructing a global octree supporting anisotropic attribute storage, and intelligently fusing in-direction and out-of-direction observation data using scanning direction vectors to eliminate multi-view conflicts and ghosting; finally, when the pose skeleton is optimized due to loop closure detection, updating anchor points drives synchronous correction of the sub-maps to achieve efficient dynamic consistency maintenance, and generating a seamlessly integrated 3D mesh model based on the moving cube algorithm.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing and mapping, and in particular relates to a method for integrated registration, fusion and reconstruction of ground-penetrating radar inverted voxel maps and surface scenes based on high-precision pose skeleton constraints. Background Technology

[0002] With the acceleration of urbanization, the need for precise detection and visual reconstruction of concealed targets such as underground pipelines, cavities, and anomalies is becoming increasingly urgent. Currently, high-precision 3D surface models are typically obtained using lidar simultaneous localization and mapping (SLAM) technology, and ground-penetrating radar (GPR) is used to obtain the distribution of physical properties of underground media. However, these two data sources differ fundamentally in their acquisition principles, data formats, and spatial reference systems. Existing technologies mostly employ separate processing and post-processing stitching to construct an integrated scene.

[0003] However, the lack of an effective rigid correlation mechanism between surface SLAM trajectories and subsurface GPR inversion data leads to severe spatial misalignment due to the accumulation of GPR odometry errors during long-distance exploration. Secondly, when SLAM optimizes the global trajectory through loop closure detection, the statically stored subsurface voxel map requires massive recalculation, making dynamic and consistent updates difficult. Furthermore, when fusing multi-angle, multi-batch GPR scan data for the same area, traditional isotropic averaging methods lose directional sensitivity features, producing ghosting artifacts. Finally, efficient collaborative management and representation of surface geometry and subsurface properties data within a unified data structure remains a gap, hindering cross-modal joint analysis and applications. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for integrated registration, fusion, and reconstruction of ground-penetrating radar inverted voxel images and surface scenes based on high-precision pose skeleton constraints, comprising:

[0005] A multi-source sensing system is constructed to obtain a high-precision global pose skeleton sequence calculated by the lidar SLAM algorithm. Based on this, the lidar point cloud is converted into sparse geometric voxels on the ground surface. At the same time, a deep network is used to invert the ground-penetrating radar data into voxel sub-maps of local rigid underground physical property parameters that characterize the underground dielectric constant.

[0006] Establish a flexible topological association, and attach the underground physical property parameter voxel subgraph as a local rigid subgraph to the corresponding pose node in the high-precision global pose skeleton sequence through virtual anchor points;

[0007] Construct a global octree that supports anisotropic attribute storage, and map the sparse geometric voxels of the ground surface and the local rigid subgraph to the global octree according to the transformation relationship between the pose nodes and the virtual anchor points.

[0008] During the mapping process, conflict detection and processing are performed based on the scanning direction vectors of the sparse geometric voxels on the surface and the local rigid subgraphs, so as to achieve anisotropic fusion or co-directional weighted fusion of multi-view observation data in the global octree.

[0009] When the high-precision global pose skeleton sequence is optimized and updated due to loop closure detection, the virtual anchor point and the attached local rigid subgraph are driven to undergo synchronous rigid transformation based on the updated pose node coordinates in order to maintain the consistency between the surface and underground space.

[0010] Based on the updated global octree, surface geometric isosurfaces and subsurface physical property isosurfaces are extracted to generate an integrated three-dimensional mesh model.

[0011] Optionally, obtaining the high-precision global pose skeleton sequence and sparse surface geometry voxels includes:

[0012] The global pose skeleton sequence is obtained by using a tightly coupled SLAM algorithm of lidar, inertial measurement unit and global positioning system;

[0013] The lidar point cloud is transformed to the world coordinate system based on the global pose skeleton sequence, and voxelization downsampling is performed to generate the sparse geometric voxels of the ground surface containing reflection intensity or occupancy probability attributes.

[0014] Optionally, establishing flexible topological associations specifically includes:

[0015] The voxel subgraphs of the underground physical property parameters generated by continuous inversion are encapsulated into local rigid subgraphs;

[0016] Based on the time synchronization relationship, the pose node corresponding to each of the local rigid subgraphs is determined in the global pose skeleton sequence;

[0017] A virtual anchor point is created on the pose node, and the local transformation relationship of the local rigid subgraph relative to the virtual anchor point is calculated to complete the mounting.

[0018] Optionally, constructing a global octree that supports anisotropic attribute storage and performing conflict detection and handling includes:

[0019] Construct a global octree, whose leaf nodes adopt a dynamic data structure to store one or more attribute state packets. Each attribute state packet contains at least an attribute value, a scan direction vector, and a confidence weight.

[0020] When mapping voxels to the global octree, if the target leaf node already contains voxels, the similarity between the scan direction vectors of the voxel to be stored and the already stored voxels is calculated.

[0021] If the similarity is greater than a preset threshold, the attribute values ​​of the two are weighted and fused, and the confidence weight is updated.

[0022] If the similarity is less than or equal to a preset threshold, a new independent attribute state package is added to the target leaf node to store the voxel to be stored.

[0023] Optionally, driving the virtual anchor point and the local rigid subgraph to undergo synchronous rigid transformation includes:

[0024] Establish constraint edges between the pose node and the virtual anchor point in the SLAM factor graph;

[0025] When the coordinates of the pose node are updated due to global optimization, the transformation matrix of the virtual anchor point is recalculated based on the updated coordinates;

[0026] Using the updated transformation matrix, a one-time rigid transformation is performed on the coordinates of all voxels within the local rigid subgraph to which the virtual anchor point is attached, in order to correct their positions in the global octree.

[0027] Optionally, generating an integrated 3D mesh model includes:

[0028] Traverse the global octree to identify leaf nodes that contain valid table attributes or subsurface properties;

[0029] The moving cube algorithm is used to extract the occupied isosurface from the sparse geometric voxels of the surface and the dielectric constant isosurface from the voxel subgraph of the underground physical property parameters.

[0030] The extracted isosurfaces are connected and synthesized to form a seamlessly integrated three-dimensional mesh model of the surface and subsurface.

[0031] Optionally, the data structure of the attribute status package further includes a scan batch identifier and a timestamp;

[0032] Before performing the weighted fusion, it is also determined whether the scan batch identifiers of the voxels to be stored and the voxels already stored are the same.

[0033] If the batch identifiers are different and the attribute values ​​differ by more than the preset change threshold, it is determined that the scene's physical state has changed, and the voxel data with the newer timestamp will be used first.

[0034] Optionally, using a deep network to invert ground-penetrating radar data into voxel submaps of subsurface physical parameters includes:

[0035] Preprocessing of time-domain profile data acquired by ground-penetrating radar;

[0036] The preprocessed data is input into a deep neural network with an encoder-decoder structure, and the network directly outputs a three-dimensional voxel grid representing the dielectric constant distribution of underground space.

[0037] Invalid voxels are removed from the output 3D voxel mesh, and the resulting mesh is encapsulated as the local rigid subgraph.

[0038] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0039] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0040] Compared with the prior art, the present invention has the following advantages and technical effects:

[0041] This invention achieves centimeter-level precision registration of surface and subsurface spaces with strong constraints, fundamentally eliminating the problem of spatiotemporal reference fragmentation caused by heterogeneous data. The proposed anisotropic octree fusion mechanism intelligently distinguishes and fuses in-direction and out-of-direction scan data, effectively eliminating multi-view scanning conflicts and "ghosting" phenomena, and fully preserving the directional characteristics of subsurface targets. The innovative "elastic anchor point" dynamic maintenance mechanism updates the sparse anchor point transformation matrix only during loop closure optimization, driving overall rigid correction of the associated subgraph, avoiding point-by-point recalculation of massive voxels, and significantly improving the efficiency of large-scale scene dynamic updates. Finally, a unified data representation system supporting the storage of surface geometry and subsurface properties is established, providing a standardized data foundation for cross-modal intelligent diagnostics. Attached Figure Description

[0042] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0043] Figure 1 This is an overall process flow diagram of an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the structure of the noise reduction network in an embodiment of the present invention, wherein (a) is the overall network structure of the noise reducer, and (b) is the detailed structure of the feature learning module;

[0045] Figure 3 This is a schematic diagram of the architecture of the inversion network and the multi-scale feature aggregation module (MSFA) according to an embodiment of the present invention;

[0046] Figure 4This is a schematic diagram illustrating the principle of unified fusion of surface point cloud voxelization and subsurface inversion voxel sub-graph in a global spatial hash table according to an embodiment of the present invention.

[0047] Figure 5 This is a schematic diagram illustrating the integrated 3D reconstruction effect of surface and underground scenes according to an embodiment of the present invention. Detailed Implementation

[0048] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0050] Example 1

[0051] like Figure 1 As shown, this embodiment provides a method for integrated registration, fusion, and reconstruction of ground-penetrating radar inverted voxel images and surface scenes based on high-precision pose skeleton constraints, including:

[0052] A multi-source sensing system is constructed to obtain a high-precision global pose skeleton sequence calculated by the lidar SLAM algorithm. Based on this, the lidar point cloud is converted into sparse geometric voxels on the ground surface. At the same time, a deep network is used to invert the ground-penetrating radar data into voxel sub-maps of local rigid underground physical property parameters that characterize the underground dielectric constant.

[0053] Establish a flexible topological association, and attach the underground physical property parameter voxel subgraph as a local rigid subgraph to the corresponding pose node in the high-precision global pose skeleton sequence through virtual anchor points;

[0054] Construct a global octree that supports anisotropic attribute storage, and map the sparse geometric voxels of the ground surface and the local rigid subgraph to the global octree according to the transformation relationship between the pose nodes and the virtual anchor points.

[0055] During the mapping process, conflict detection and processing are performed based on the scanning direction vectors of the sparse geometric voxels on the surface and the local rigid subgraphs, so as to achieve anisotropic fusion or co-directional weighted fusion of multi-view observation data in the global octree.

[0056] When the high-precision global pose skeleton sequence is optimized and updated due to loop closure detection, the virtual anchor point and the attached local rigid subgraph are driven to undergo synchronous rigid transformation based on the updated pose node coordinates in order to maintain the consistency between the surface and underground space.

[0057] Based on the updated global octree, surface geometric isosurfaces and subsurface physical property isosurfaces are extracted to generate an integrated three-dimensional mesh model.

[0058] The first step is the construction and calibration of a multi-source sensing system.

[0059] A vehicle-mounted joint data acquisition system was built, and the sensor suite included: a 16-line lidar (LiDAR), a high-precision inertial measurement unit (IMU), a GPS positioning module, and an array ground-penetrating radar (GPR).

[0060] Establish a unified coordinate system: Define the world coordinate system determined by the initial GPS frame. LiDAR coordinate system with the center of the lidar as the origin GPR coordinate system with the center of the ground penetrating radar as the origin .

[0061] Before the data acquisition task begins, the rigid body transformation extrinsic parameter matrices between each sensor are calculated through calibration experiments. (From ground-penetrating radar to lidar) and (IMU to lidar).

[0062] Control the vehicle to travel along the area to be measured (such as urban roads) and simultaneously collect surface laser point cloud data, GPS / IMU navigation data and underground GPRC-scan volume data.

[0063] The second step is to calculate the high-precision pose skeleton of the ground surface.

[0064] Relative pose constraints are obtained through scan-matching of LiDAR point clouds. These constraints are then combined with IMU pre-integration constraints and GPS absolute position constraints to construct a factor graph optimization model. The model is then used to calculate the vehicle's position at each time step in the world coordinate system. High-precision six-DOF pose This forms a globally consistent "pose skeleton" sequence, which serves as a unified spatiotemporal reference for subsequent registration of surface and subsurface data.

[0065] The third step is the generation of sparse geometric elements on the ground surface.

[0066] Each frame of the LiDAR point cloud is determined based on the current pose. Transform to world coordinates. Set the resolution of the global octree (e.g., ...). For each valid laser reflection point, it is mapped as a "sparse geometric voxel".

[0067] Voxel attribute assignment: The reflection intensity of a point is used as the voxel value, and the voxel is marked as "occupied". Only voxels in non-empty regions are stored to achieve sparse representation.

[0068] The fourth step is the inversion of the underground rigid voxel graph.

[0069] The collected GPR data is divided into time windows and input into a pre-trained deep learning network.

[0070] like Figure 2 As shown, the noise reduction network is a noise suppression structure based on a three-dimensional convolutional neural network. Its input is noisy ground-penetrating radar C-scan data, and its output is noise-suppressed C-scan data. Figure 2 As shown in (a), the network comprises three main components: an initial feature extraction module, a feature learning module, and a reconstruction module. The initial feature extraction module consists of a three-dimensional convolutional layer with a kernel size of 3×3×3, a stride of 1×1×1, and C1 output channels. It is used to extract initial feature maps from the input data and performs non-linear activation using the ReLU activation function. The feature learning module consists of m cascaded feature learning units, each with the structure shown below. Figure 2 As shown in (b), it includes two residual blocks and one feature attention block. Each residual block consists of two 3D convolutional layers with the same number of channels, and identity mapping is achieved through cross-layer connections to avoid gradient vanishing or exploding problems. The feature attention block first calculates channel statistics through global average pooling, then generates channel attention vectors through two fully connected layers and the Sigmoid function, and multiplies these vectors with the feature map using channel weighting to enhance the response of important features. The reconstruction module consists of a single-channel 3D convolutional layer, which sums the deep features processed by the feature learning module with the initial features, and then performs convolution and activation operations to finally reconstruct the denoised C-scan data.

[0071] like Figure 3As shown, the inversion network is an encoder-decoder network based on a 3D U-Net architecture. Its input is denoised C-scan data, and its output is a 3D dielectric constant distribution map of the target region. The network consists of an encoding path, a decoding path, and skip connections. The encoding path contains n encoding blocks, each containing a multi-scale feature aggregation module (MSFA) and a max-pooling layer with a stride of 2×2×2. The MSFA consists of three consecutive 3D convolutional layers, each with a kernel size of 3×3×3 and a stride of 1×1×1. Through progressively expanding receptive fields, features at different scales are extracted. These three feature maps at different scales are concatenated along the channel dimension to form a fused multi-scale feature representation. The decoding path is symmetrical to the encoding path, containing n decoding blocks. Each decoding block includes a transposed convolutional layer with a stride of 2×2×2 for upsampling, and the same MSFA as in the encoding block, used to progressively reconstruct spatial details from the compressed features. The multi-scale feature maps output by each coding block in the encoding path are concatenated with the feature maps at a resolution level in the middle of the decoding path through skip connections to compensate for the high-frequency information lost during pooling. A single-channel 3D convolutional layer is set at the end of the network, combined with a linear activation function, to finally output a 3D dielectric constant distribution map.

[0072] Subgraph encapsulation: The network outputs a fixed-size 3D tensor (e.g., A voxel represents the dielectric constant distribution of a local physical space underground. This three-dimensional tensor is defined as a "local rigid subgraph," and the current scan batch identifier (BatchID) is recorded in the subgraph metadata to distinguish different acquisition jobs or reciprocating scan sequences.

[0073] The fifth step is to establish topological associations based on elastic anchor points.

[0074] like Figure 4 As shown, unlike traditional methods that directly write voxels into absolute coordinates, this step establishes an elastic topology of "surface pose nodes - underground rigid subgraph":

[0075] Synchronize based on timestamp, find the first A sub-map of underground rigidity Corresponding surface pose skeleton node .

[0076] At pose node Create a "Virtual Anchor" on it.

[0077] Computational subgraph Local transformation matrix relative to the anchor point The matrix is ​​then stored in a fixed position. At this point, the subgraph is attached to the skeleton, and its absolute position changes dynamically as the skeleton nodes move.

[0078] The sixth step is to construct an anisotropic octree and merge the conflicts.

[0079] A unified global octree is constructed to manage surface and subsurface voxels. To resolve conflicts in observation data from different perspectives at the same spatial location, the octree employs a storage mechanism that supports multi-state coexistence.

[0080] Leaf node data structure definition:

[0081] Each leaf node of the global octree is configured as a dynamic container (such as a linked list or dynamic array) to store one or more attribute state packets. Each attribute state packet contains the following core information fields:

[0082] Physical attribute value field: used to store surface reflectivity values ​​or underground dielectric constant values;

[0083] Confidence weight field: Used to record the confidence level of the current attribute value, usually related to the sensor measurement distance or the probability output of the inversion network;

[0084] Scan direction vector field: Used to store the unit direction vector of the sensor's line of sight in the world coordinate system when the data was acquired, representing the observation angle;

[0085] The scan batch identifier field (BatchID) is used to store the acquisition task number or scan sequence number to which the current data belongs, and is used to distinguish observation data from different time periods or different work batches.

[0086] Timestamp field: Used to record the specific time when the data was collected.

[0087] Mapping and collision detection process:

[0088] Based on the pose calculated in step 5, the surface or subsurface voxels are mapped to their specific spatial locations in the global octree. If historical data already exists in the leaf nodes of the octree to which the new voxel is to be stored, the following conflict resolution logic is executed:

[0089] Directional similarity calculation: Extract the scanning direction vector of the voxel to be stored and the scanning direction vector of each stored attribute state packet in the leaf node container, and calculate the directional cosine similarity between them respectively.

[0090] Same-direction fusion update: If the calculated maximum similarity is higher than the preset fusion threshold (e.g., 0.9), the current data is determined to be a same-direction repeated scan. At this time, the corresponding stored attribute state package is selected, and the attribute values ​​of the voxel to be stored are weighted and averaged with the stored attribute values, and the confidence weight is accumulated to improve the accuracy and smoothness of the data.

[0091] Independent storage in opposite directions: If the calculated maximum similarity is lower than or equal to the fusion threshold, the current data is determined to be a multi-view cross-scan (e.g., pipeline detection at a crossroads). In this case, no modifications are made to the original data. Instead, an independent attribute status package is added to the container of the leaf node, which fully preserves the attribute values, scanning direction vectors, and scanning batch identifiers of the voxels to be stored.

[0092] Batch conflict handling: If a scan is determined to be in the same direction, but the scan batch identifier of the voxel to be stored is different from the already stored data, and the attribute values ​​differ significantly (exceeding the preset change threshold), then it is determined that the environment in that area has undergone physical changes (such as road excavation and repair). In this case, the scan batch data with the newer timestamp is retained first, overwriting the older batch data.

[0093] Step 7: Dynamic consistency maintenance based on loop closure detection.

[0094] When the surface SLAM system detects historical trajectory drift through loop closure detection and triggers global factor graph optimization:

[0095] Obtain the optimized new pose skeleton sequence.

[0096] Update the absolute transformation matrix of all affected virtual anchor points in the world coordinate system.

[0097] Using the updated matrix, drive the rigid transformation of the mounted underground rigid subgraph. Complexity).

[0098] After the subgraph position is updated, only the index pointer of the affected region in the octree needs to be updated. There is no need to recalculate the attribute values ​​in a massive number of voxels, which achieves millisecond-level dynamic map correction.

[0099] Step 8: Integrated visual reconstruction.

[0100] Traverse the final globally anisotropic octree:

[0101] Surface reconstruction: Extract states with reflectivity greater than a threshold from each leaf node and generate a surface geometric mesh using the Moving Cubes algorithm.

[0102] Underground reconstruction: Extract the state of dielectric_constant anomalies (such as voids or pipelines) from each leaf node. For nodes with multiple "coexisting" states, dynamically select the best display data or merge the display based on the user's perspective.

[0103] like Figure 5 As shown, rendering involves importing the surface and underground mesh models into the rendering engine to generate a unified 3D scene with clear surface textures, accurate underground pipeline routing, and no stitching gaps or ghosting.

[0104] Example 2

[0105] This embodiment provides a method for integrated registration, fusion, and reconstruction of ground-penetrating radar inverted voxel images and surface scenes based on high-precision pose skeleton constraints, including:

[0106] S1: Construct a multi-source sensing system to acquire a high-precision pose skeleton sequence of the Earth's surface, sparse geometric voxels of the Earth's surface, and voxel sub-maps of underground physical property parameters; wherein, the pose skeleton sequence is obtained by a tightly coupled lidar-SLAM algorithm; the voxel sub-maps of underground physical property parameters are obtained by inversion of ground-penetrating radar data through a deep learning network and are used to characterize the dielectric constant of the underground medium.

[0107] S2: Establish an elastic topological association between "surface pose node - underground rigid subgraph"; encapsulate the continuously inverted underground physical property parameters voxels into a local rigid subgraph, and attach the local rigid subgraph to the corresponding surface high-precision pose skeleton node through virtual anchor points;

[0108] S3: Construct a global octree that supports anisotropic attribute storage, and map the sparse surface geometric voxels and underground physical property parameter voxels to the global octree according to the pose information after mounting.

[0109] S4: Execute conflict detection based on scan direction vector and scan batch during the mapping process: When a new voxel is mapped to an already occupied leaf node in the octree, calculate the similarity between the scan direction vectors of the new and old voxels, and perform in-direction weighted fusion or out-of-direction independent storage based on the similarity.

[0110] S5: Perform dynamic consistency maintenance based on elastic anchor points: When the high-precision pose skeleton of the ground surface triggers global optimization due to closed-loop detection, update the transformation matrix of the virtual anchor points and drive the mounted underground local rigid sub-graph to perform rigid transformation synchronously in the global coordinate system.

[0111] S6: Traverse the updated global octree and use the Moving Cubes algorithm to extract the occupancy isosurface of sparse surface geometric voxels and the medium property isosurface of underground physical property parameter voxels respectively, to generate an integrated three-dimensional mesh model that seamlessly connects surface texture and underground physical properties.

[0112] Furthermore, the conflict detection and processing in step S4 specifically includes:

[0113] S41: Extract the scan direction vectors of the voxels to be stored and the voxels already stored in the leaf nodes of the octagonal tree.

[0114] S42: Calculate the cosine similarity between two scanning direction vectors;

[0115] S43: If the cosine similarity is greater than the preset fusion threshold, it is determined to be a repeated scan in the same direction. The physical property parameter values ​​of the two voxels are weighted and averaged, and the confidence weight of the node is updated.

[0116] S44: If the cosine similarity is less than or equal to the fusion threshold, it is determined to be a multi-view cross-scan. A new state entry is added to the data list of the current octagonal leaf node, and the attribute value and scanning direction vector of the voxel to be stored are stored independently to preserve the anisotropic features of the spatial location.

[0117] Furthermore, the leaf nodes of the global octree adopt a dynamic linked list structure, and each node contains one or more attribute state packets. The data structure of each attribute state packet includes at least: dielectric constant attribute value, surface reflectivity attribute value, scan direction vector, confidence weight, and timestamp.

[0118] Furthermore, step S5 specifically includes:

[0119] S51: Establish constraint edges between pose nodes and virtual anchor points in the SLAM factor graph, wherein the virtual anchor points store the relative transformation relationship between the local rigid subgraph and the origin of the global coordinate system;

[0120] S52: When the SLAM backend optimizes and updates the coordinates of the pose nodes, the relative positions of the voxels inside the local rigid subgraph remain unchanged, and the model transformation matrix of the virtual anchor point is recalculated only based on the updated pose node coordinates.

[0121] S53: Using the updated model transformation matrix, perform a one-time rigid rotation and translation on the coordinates of all voxels within the local rigid subgraph to complete their position correction in the global octree.

[0122] Furthermore, the step of obtaining the voxel sub-map of underground physical property parameters specifically includes:

[0123] After preprocessing the time-domain B-scan profile data sequence collected by ground penetrating radar, it is input into a pre-trained deep learning inversion network.

[0124] Using the encoder-decoder structure of the deep learning network, the two-dimensional temporal radar profile is mapped into a three-dimensional spatial domain dielectric constant voxel grid, and invalid voxels with air values ​​are removed to generate the local rigid subgraph.

[0125] Furthermore, the step of obtaining the sparse geometric voxels of the land surface includes:

[0126] The point cloud data collected by the lidar is stitched together into a local point cloud map according to the pose skeleton sequence.

[0127] The local point cloud map is downsampled using voxelization, voxels in non-idle areas are retained, and voxel values ​​are set as lidar reflection intensity or occupancy probability to generate the sparse geometric voxels of the ground surface.

[0128] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0129] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0130] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for integrated registration, fusion, and reconstruction of ground-penetrating radar inverted voxel images and surface scenes based on high-precision pose skeleton constraints, characterized in that, include: A multi-source sensing system is constructed to obtain a high-precision global pose skeleton sequence calculated by the lidar SLAM algorithm. Based on this, the lidar point cloud is converted into sparse geometric voxels on the ground surface. At the same time, a deep network is used to invert the ground-penetrating radar data into voxel sub-maps of local rigid underground physical property parameters that characterize the underground dielectric constant. Establish a flexible topological association, and attach the underground physical property parameter voxel subgraph as a local rigid subgraph to the corresponding pose node in the high-precision global pose skeleton sequence through virtual anchor points; Construct a global octree that supports anisotropic attribute storage, and map the sparse geometric voxels of the ground surface and the local rigid subgraph to the global octree according to the transformation relationship between the pose nodes and the virtual anchor points. During the mapping process, conflict detection and processing are performed based on the scanning direction vectors of the sparse geometric voxels on the surface and the local rigid subgraphs, so as to achieve anisotropic fusion or co-directional weighted fusion of multi-view observation data in the global octree. When the high-precision global pose skeleton sequence is optimized and updated due to loop closure detection, the virtual anchor point and the attached local rigid subgraph are driven to undergo synchronous rigid transformation based on the updated pose node coordinates in order to maintain the consistency between the surface and underground space. Based on the updated global octree, surface geometric isosurfaces and subsurface physical property isosurfaces are extracted to generate an integrated three-dimensional mesh model.

2. The method according to claim 1, characterized in that, Obtaining the high-precision global pose skeleton sequence and sparse surface geometry voxels includes: The global pose skeleton sequence is obtained by using a tightly coupled SLAM algorithm of lidar, inertial measurement unit and global positioning system; The lidar point cloud is transformed to the world coordinate system based on the global pose skeleton sequence, and voxelization downsampling is performed to generate the sparse geometric voxels of the ground surface containing reflection intensity or occupancy probability attributes.

3. The method according to claim 1, characterized in that, Establishing flexible topology associations specifically includes: The voxel subgraphs of the underground physical property parameters generated by continuous inversion are encapsulated into local rigid subgraphs; Based on the time synchronization relationship, the pose node corresponding to each of the local rigid subgraphs is determined in the global pose skeleton sequence; A virtual anchor point is created on the pose node, and the local transformation relationship of the local rigid subgraph relative to the virtual anchor point is calculated to complete the mounting.

4. The method according to claim 1, characterized in that, Constructing a global octree that supports anisotropic attribute storage and performing conflict detection and handling includes: Construct a global octree, whose leaf nodes adopt a dynamic data structure to store one or more attribute state packets. Each attribute state packet contains at least an attribute value, a scan direction vector, and a confidence weight. When mapping voxels to the global octree, if the target leaf node already contains voxels, the similarity between the scan direction vectors of the voxel to be stored and the already stored voxels is calculated. If the similarity is greater than a preset threshold, the attribute values ​​of the two are weighted and fused, and the confidence weight is updated. If the similarity is less than or equal to a preset threshold, a new independent attribute state package is added to the target leaf node to store the voxel to be stored.

5. The method according to claim 1, characterized in that, Driving the virtual anchor point and the local rigid subgraph to perform synchronous rigid transformation includes: Establish constraint edges between the pose node and the virtual anchor point in the SLAM factor graph; When the coordinates of the pose node are updated due to global optimization, the transformation matrix of the virtual anchor point is recalculated based on the updated coordinates; Using the updated transformation matrix, a one-time rigid transformation is performed on the coordinates of all voxels within the local rigid subgraph to which the virtual anchor point is attached, in order to correct their positions in the global octree.

6. The method according to claim 1, characterized in that, Generating an integrated 3D mesh model includes: Traverse the global octree to identify leaf nodes that contain valid table attributes or subsurface properties; The moving cube algorithm is used to extract the occupied isosurface from the sparse geometric voxels of the surface and the dielectric constant isosurface from the voxel subgraph of the underground physical property parameters. The extracted isosurfaces are connected and synthesized to form a seamlessly integrated three-dimensional mesh model of the surface and subsurface.

7. The method according to claim 1, characterized in that, The data structure of the attribute status package also includes a scan batch identifier and a timestamp; Before performing the weighted fusion, it is also determined whether the scan batch identifiers of the voxels to be stored and the voxels already stored are the same. If the batch identifiers are different and the attribute values ​​differ by more than the preset change threshold, it is determined that the scene's physical state has changed, and the voxel data with the newer timestamp will be used first.

8. The method according to claim 1, characterized in that, Using deep networks to invert ground-penetrating radar data into voxel submaps of subsurface physical parameters includes: Preprocessing of time-domain profile data acquired by ground-penetrating radar; The preprocessed data is input into a deep neural network with an encoder-decoder structure, and the network directly outputs a three-dimensional voxel grid representing the dielectric constant distribution of underground space. Invalid voxels are removed from the output 3D voxel mesh, and the resulting mesh is encapsulated as the local rigid subgraph.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images