An underground target detection method based on surface semantic information assistance

Abstract

The present application belongs to the field of geophysical exploration and nondestructive testing technology, and particularly relates to a kind of underground target detection method based on surface semantic information auxiliary, comprising: obtaining the surface laser radar point cloud data and ground penetrating radar data of target detection area;Surface laser radar point cloud data is carried out semantic segmentation, at least one surface target with underground indication significance in target detection area is determined;According to the semantic category of surface target, the corresponding underground area is determined from the preset surface-underground space mapping relationship, and the underground space feature data of target detection area is constructed according to the three-dimensional body data of the underground area;Underground space feature data and ground penetrating radar data are combined to carry out inversion reconstruction, to determine the dielectric constant distribution and / or underground target probability distribution of target detection area.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration and non-destructive testing technology, specifically relating to a method for detecting underground targets based on surface semantic information. Background Technology

[0002] Underground target detection is an important technical means in the construction of urban underground space and the management of municipal facilities. Ground penetrating radar is widely used in the detection of underground pipelines, underground cavities and road defects due to its non-destructive and highly adaptable characteristics.

[0003] However, ground-penetrating radar acquires indirect reflection information of underground structures. Its inversion process requires inferring the distribution of underground targets under limited observation conditions, which naturally leads to problems of non-uniqueness and insufficient stability of the solution, especially in urban scenarios with complex underground environments. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, the purpose of this invention is to propose a method for detecting underground targets based on surface semantic information. By introducing the semantic information of the lidar point cloud on the surface, it is transformed into constraint or auxiliary information in the underground inversion process, thereby improving the detection effect of underground targets.

[0005] To achieve the above and other related objectives, this invention provides a method for detecting underground targets based on surface semantic information, comprising: acquiring surface lidar point cloud data and ground-penetrating radar data of a target detection area; performing semantic segmentation on the surface lidar point cloud data to determine at least one surface target with underground indication significance within the target detection area; determining the corresponding underground region from a preset surface-underground space mapping relationship according to the semantic category of the surface target, and constructing underground spatial feature data of the target detection area based on the three-dimensional volume data of the underground region; and performing inversion reconstruction by combining the underground spatial feature data and the ground-penetrating radar data to determine the underground dielectric constant distribution and / or underground target probability distribution of the target detection area.

[0006] According to a specific embodiment of the present invention, the step of semantic segmentation of the surface lidar point cloud data to determine at least one surface target with underground indication significance within the target detection area includes: classifying the surface lidar point cloud data point by point using a trained three-dimensional point cloud semantic segmentation network, and performing instance clustering on the classified surface lidar point cloud data to obtain at least one surface target.

[0007] According to a specific embodiment of the present invention, the steps of determining the corresponding underground region from a preset surface-underground space mapping relationship based on the semantic category of the surface target, and constructing the underground space feature data of the target detection area based on the three-dimensional volume data of the underground region include: for each surface target, extracting its corresponding semantic category, geometric center coordinates, geometric dimensions, and principal axis direction vector; determining the corresponding underground region from the preset surface-underground space mapping relationship based on the semantic category of the surface target, and constructing a corresponding prior sub-volume based on the confidence level of the underground region and the geometric center coordinates, geometric dimensions, and principal axis direction vector of the surface target; fusing the prior sub-volumes of all surface targets to obtain underground space prior information, which serves as the underground space feature data of the target detection area.

[0008] According to a specific embodiment of the present invention, the prior subbody is constructed according to the following formula: , in, Denotes the i-th prior child body. This represents the baseline confidence level of the underground region corresponding to the i-th surface target. This represents the horizontal distance from the subject to be estimated to the projection axis of the target center on the Earth's surface. This indicates the vertical offset of the subject body relative to the estimated burial depth. This indicates the degree of deviation of the subject from the direction vector of the principal axis. , , These represent the horizontal tolerance parameter, vertical tolerance parameter, and directional tolerance parameter, respectively.

[0009] According to a specific embodiment of the present invention, the step of fusing the prior sub-body of all surface targets to obtain prior information of underground space includes: fusing the prior sub-body of all surface targets using maximum value fusion or weighted superposition fusion to obtain prior information of underground space.

[0010] According to a specific embodiment of the present invention, the step of performing inversion reconstruction by combining the underground space feature data and the ground penetrating radar data to determine the underground dielectric constant distribution and / or underground target probability distribution of the target detection area includes: encoding the underground space feature data according to the data size of the ground penetrating radar data, and concatenating the encoded underground space feature data with the ground penetrating radar data to obtain a corresponding input tensor; inputting the input tensor into a trained inversion network to obtain the underground dielectric constant distribution and / or underground target probability distribution of the target detection area.

[0011] According to a specific embodiment of the present invention, the inversion network adopts a three-dimensional fully convolutional encoder-decoder network based on multi-scale feature aggregation, comprising: an input layer, an encoding path, a bottleneck layer, a decoding path, skip connections, and an output layer; wherein, the input layer is used to receive an input tensor; the encoding path is used to extract shallow local features and deep abstract features from the input tensor step by step, and generate corresponding feature maps; the decoding path is used to restore the spatial resolution of the feature maps output by the encoding path step by step, and reconstruct the distribution of underground medium parameters based on them; the bottleneck layer is used to compress the feature maps generated by the encoding path and pass them to the decoding path; the skip connections are used to pass the high-resolution detail features extracted by the encoding path to the decoding path; the output layer is used to generate the corresponding underground dielectric constant distribution and / or underground target probability distribution based on the output of the decoding path.

[0012] According to a specific embodiment of the present invention, at least one level of the encoding path is provided with a priori modulation unit, and the priori modulation unit enhances the features according to the following formula: ,and , in, Represents the coding features of layer I. This represents the encoded features enhanced by prior modulation. This represents the features in the underground space feature data. express The corresponding modulation weighting diagram, This represents the Sigmoid activation function.

[0013] According to a specific embodiment of the present invention, the training steps of the inversion network include: performing reverse optimization on the pre-constructed inversion network using a joint loss function to obtain a trained inversion network; wherein, the joint loss function includes at least a regression loss term, a segmentation loss term, and a boundary gradient loss term, and the regression loss term is used to constrain the difference between the underground dielectric constant distribution predicted by the inversion network and the true label, the segmentation loss term is used to constrain the consistency between the underground target probability distribution predicted by the inversion network and the true label, and the boundary gradient loss term is used to enhance the clarity of the underground target boundary.

[0014] According to a specific embodiment of the present invention, inversion reconstruction is performed by combining the underground space feature data and the ground penetrating radar data to determine the underground dielectric constant distribution and / or underground target probability distribution of the target detection area, including: constructing a corresponding area mask based on the underground space feature data, and performing inversion reconstruction by combining the underground space feature data, the area mask, and the ground penetrating radar data to determine the underground dielectric constant distribution and / or underground target probability distribution of the target detection area.

[0015] The beneficial effects of the present invention are as follows: By introducing prior information on underground space constructed based on surface semantic information, the present invention can constrain or guide the possible distribution areas, structural boundaries and medium parameter distribution of underground targets, effectively reduce the solution space of the ground penetrating radar inversion problem, reduce the uncertainty in the inversion process, and reduce the pseudo-anomaly response caused by complex underground environment.

[0016] Meanwhile, this invention employs an inversion network structure based on multi-scale feature aggregation. By progressively expanding the receptive field through cascaded convolution and fusing spatial features of different scales, it can simultaneously characterize small-scale underground anomalies and large-scale continuous structures, thereby improving the resolution and accuracy of underground target inversion imaging. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0018] Figure 1 This is a flowchart illustrating a method for detecting underground targets based on surface semantic information, provided in one embodiment of the present invention. Figure 2 This is another flowchart illustrating a method for detecting underground targets based on surface semantic information, provided in one embodiment of the present invention. Figure 3 This is a schematic diagram of the inversion network training process provided in one embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an underground target detection system based on surface semantic information assistance provided in one embodiment of the present invention; Figure 5 This is a structural block diagram of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0020] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0021] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, publicly known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0022] Example 1 Please see Figure 1 , 2 The method for detecting underground targets based on surface semantic information, as shown, includes: Step S100: Obtain surface lidar point cloud data and ground-penetrating radar data of the target detection area.

[0023] Specifically, in practical applications, a mobile data acquisition platform (such as a vehicle-mounted mobile measurement system, a hand-pushed radar detection vehicle, or a drone or robot carrying a detection payload) can be used to simultaneously scan the target detection area. This mobile data acquisition platform integrates a surface lidar, a ground-penetrating radar, and a combined navigation module. The surface lidar is used to acquire the surface semantic information (surface lidar point cloud data) of the target detection area, the ground-penetrating radar is used to acquire ground-penetrating radar data (such as two-dimensional B-scan profile data, three-dimensional array radar C-scan data) of the target detection area, and the combined navigation module includes GNSS (satellite navigation), an IMU (inertial measurement unit), and a pose calculation unit, used to record the trajectory and attitude information of the data acquisition platform during its movement.

[0024] It is understandable that during data acquisition, surface lidar point cloud data and ground-penetrating radar data of the target detection area will be acquired simultaneously. For the acquired ground-penetrating radar data, to ensure the effectiveness of the subsequent inversion and reconstruction input, it can be preprocessed, such as zero-time correction, direct-wave removal, background removal, bandpass filtering, and other preprocessing methods. Zero-time correction aligns the echo position at the air-surface interface to the zero point of the time axis; direct-wave removal and background removal weaken the masking of shallow subsurface signals by strong surface reflections and system background responses; and bandpass filtering suppresses high-frequency noise and low-frequency drift. Therefore, one or more preprocessing methods can be used to enhance, denoise, adjust, and convert the acquired surface lidar point cloud data and / or ground-penetrating radar data according to actual needs. This is not limited to the aforementioned preprocessing methods, and no excessive restrictions are imposed. Modifications and refinements made by those skilled in the art to the embodiments of this invention without departing from the spirit of this invention still fall within the scope of the invention application.

[0025] Furthermore, after preprocessing the surface lidar point cloud data and / or ground-penetrating radar data, unified coordinate registration can be performed on the surface lidar point cloud data and ground-penetrating radar data based on the trajectory and attitude information recorded by the integrated navigation module. Specifically, rigid body transformation relationships between the lidar coordinate system, the ground-penetrating radar coordinate system, and the platform coordinate system can be constructed using the extrinsic parameter calibration results, and both types of data can be uniformly transformed into the same absolute geographic coordinate system or local engineering coordinate system. Correspondingly, through the above spatiotemporal registration process, a spatial correspondence between the surface lidar point cloud data and the underground ground-penetrating radar data can be established, thereby enabling the planar position of the surface semantic target to match the corresponding area in the underground data, providing a foundation for subsequent underground spatial feature data and inversion reconstruction.

[0026] It should be added that the surface lidar point cloud data and ground penetrating radar data can also be first registered in coordinates, and then the surface lidar point cloud data and / or ground penetrating radar data can be preprocessed separately. There is no limitation on this. Modifications and refinements made by those skilled in the art to the embodiments of the present invention without departing from the spirit of the present invention still fall within the scope of the invention application patent of the present invention.

[0027] Step S200: Semantic segmentation is performed on the surface lidar point cloud data to determine at least one surface target with underground indication significance within the target detection area.

[0028] Specifically, a trained 3D point cloud semantic segmentation network can be used to classify the original point cloud, i.e., the aforementioned surface lidar point cloud data, point by point, and then perform instance clustering on the classified point cloud data to extract independent surface target instances. These surface targets include, but are not limited to, manhole covers, storm drain grates, fire hydrants, linear road facilities, road edge structures, and road surface cracks.

[0029] Furthermore, for each identified surface target, its corresponding semantic category, 3D spatial location, geometric dimensions, and directional attributes can be extracted. The 3D spatial location of a surface target can be determined by its geometric center, boundary point cloud, or 3D bounding box, while its geometric dimensions can be represented by the bounding box dimensions in three spatial directions. Simultaneously, for surface targets with obvious directionality, their principal axis direction vectors can be extracted to describe their extension direction in surface space. Therefore, the geometric center coordinates, geometric dimensions, and principal axis direction vectors of each surface target can be extracted.

[0030] Step S300: Based on the semantic category of the surface target, determine the corresponding underground region from the preset surface-underground space mapping relationship, and construct the underground space feature data of the target detection area based on the three-dimensional volume data of the underground region.

[0031] After extracting the geometric center coordinates, geometric dimensions, and principal axis direction vector of the surface target, underground space feature data can be constructed based on the preset surface-underground space mapping relationship, such as prior information of underground space that characterizes the distribution probability or geometric range of underground structures.

[0032] In this embodiment, prior information about underground space is specifically used as underground space feature data to characterize at least one of the following: the area where an underground target may exist, its probability of existence, its direction of extension, and background medium parameters. Furthermore, the prior information about underground space is preferably three-dimensional volume data corresponding to the underground area to be inverted, whose spatial coordinate system is aligned with the ground-penetrating radar data, and whose grid size is consistent with the pre-constructed inversion network, or is consistent with the inversion network after resampling.

[0033] Furthermore, the surface-to-subsurface spatial mapping relationship includes: point mapping, for example, when the surface target is a manhole cover, it is inferred in a probabilistic manner that there may be a manhole passage or cavity at the corresponding underground location, and a corresponding confidence weight is assigned; linear mapping, for example, when the surface target is a street lamp pole, fire hydrant, or linear road feature, based on the category of the target and the distribution pattern of similar facilities in the surrounding area, it is inferred in a statistical probability form that there may be a pipeline structure extending in a specific direction at the corresponding underground location, and a corresponding confidence weight is assigned; regional mapping, for example, when the surface target is a specific road surface material, the corresponding underground region is mapped to have a specific background value of medium parameters, and a corresponding confidence weight is assigned, and so on.

[0034] Specifically, for each surface target According to its corresponding semantic category The corresponding underground region is determined from the preset surface-subsurface spatial mapping relationship, and the confidence level of the underground region and the geometric center coordinates of the surface target are used as the basis for the determination. Geometric dimensions The prior subbody is constructed from the principal axis direction vector. Finally, the prior sub-objects of all surface targets are fused to obtain the prior information of underground space.

[0035] The prior sub-body can be constructed using a probability decay method, as follows: , This represents the baseline confidence level of the underground region corresponding to the i-th surface target. This represents the spatial coordinates of a specific element to be estimated within the underground region. This represents the horizontal distance from the subject to be estimated to the projection axis of the target center on the Earth's surface. This indicates the vertical offset of the subject body relative to the estimated burial depth. This indicates the degree of deviation of the subject from the direction vector of the principal axis. , , These represent the horizontal tolerance parameter, vertical tolerance parameter, and directional tolerance parameter, respectively.

[0036] Understandably, for surface targets that do not have a clear direction, the direction deviation term can be used. The corresponding exponential decay factor is set to 1; for scenarios lacking reliable prior knowledge of burial depth, the vertical offset term can be... The corresponding exponential decay factor is set to 1, so that only horizontal position constraints are used, or horizontal position constraints and directional constraints are used together to construct the prior information of underground space.

[0037] Finally, maximum value fusion or weighted superposition fusion can be used to fuse the prior sub-volumes of all surface targets to generate prior information of underground space.

[0038] For example, using the maximum value fusion method, the corresponding prior information of underground space can be represented as: , Where N represents the total number of surface targets.

[0039] Alternatively, using a weighted overlay and fusion method, the corresponding prior information of underground space can be represented as: , in, Indicates the first The fusion weights of the prior sub-sub ... Preferably, It can be determined based on the reliability of surface target categories, the credibility of mapping relationships, target integrity, historical maintenance information, or failure probability factors, and can meet the following requirements. .

[0040] It should be added that, considering the situation of abandoned facilities, relocated facilities, or ineffective surface markers in old urban areas, the confidence level in the above formula can be adjusted. The confidence level is adjusted, and the adjusted confidence level can be expressed as: , in, This represents the abandonment probability factor or failure probability factor corresponding to the i-th surface target, in order to reduce the misleading effect of failed surface targets on the subsurface inversion process.

[0041] Step S400: Combine the underground space feature data with the ground penetrating radar data to perform inversion reconstruction, so as to determine the underground dielectric constant distribution and / or underground target probability distribution of the target detection area.

[0042] After constructing the prior information of underground space, it can be incorporated into the ground-penetrating radar data inversion and reconstruction process. To this end, this embodiment also constructs a semantically guided underground target inversion network, employing a three-dimensional fully convolutional encoder-decoder network based on multi-scale feature aggregation, including an input layer, encoding path, bottleneck layer, decoding path, skip connections, and output layer.

[0043] The input layer is configured as a multi-channel layer to receive ground-penetrating radar (GPR) data and prior information about the subsurface space. The prior information and GPR data are concatenated along the channel dimension to form a multi-channel input tensor for the inversion network. This design allows the inversion network to automatically suppress noise responses in non-semantically related regions by utilizing the spatial location weights implied by surface semantics during the initial feature extraction stage. The encoding path extracts shallow local features and deep abstract features from the input tensor step-by-step, generating corresponding feature maps. The decoding path restores the spatial resolution of the feature maps output by the encoding path step-by-step and reconstructs the distribution of subsurface medium parameters based on them. Accordingly, the encoding path includes several cascaded downsampling modules for extracting multi-scale abstract features, and the decoding path includes corresponding upsampling modules for restoring spatial resolution. Preferably, a multi-scale feature aggregation unit can be introduced into the network's convolutional modules. This involves combining multiple convolutional layers with different dilation rates or kernel sizes to expand the receptive field, adapting to the detection needs of targets at different burial depths and sizes. The bottleneck layer compresses the feature map generated by the encoding path and passes it to the decoding path. Skip connections pass high-resolution detail features extracted by the encoding path to the decoding path, thereby improving the ability to recover the boundaries and subtle anomalies of underground targets. The output layer maps high-dimensional features back to the physical parameter space through convolutional layers, and outputs three-dimensional volume data representing the distribution of physical properties (such as relative permittivity) of the underground medium using activation functions, thus achieving accurate imaging of underground targets.

[0044] In one specific embodiment, the inversion network adopts a four-level encoding-decoding structure, and the number of feature channels at each level can be set to 32, 64, 128, and 256 respectively, while the number of channels in the bottleneck layer is set to 512.

[0045] The encoding path comprises four cascaded encoding blocks. Each block contains a Multi-Scale Feature Aggregation Module (MSFA) and a downsampling layer. The MSFA includes multiple parallel convolutional branches, each employing 3D convolutions with different dilation rates to extract subsurface anomaly responses at different spatial scales. The outputs of these branches are then concatenated along the channel dimension and fused through convolutional layers to form a multi-scale representation that simultaneously incorporates small-scale local anomaly features and large-scale continuous structural features. For example, the MSFA consists of three consecutive 3D convolutional layers, each with a 3×3×3 kernel and a 1×1×1 stride. The first layer extracts local micro-features, the second layer expands the receptive field based on the first layer's results, and the third layer further expands it. These three layers' feature maps are concatenated along the channel dimension to form a fused multi-scale feature representation. This design allows the inversion network to simultaneously capture features of both micro-cracks (small-scale) and large cavities (large-scale). Meanwhile, the downsampling layer, which follows the multi-scale feature aggregation module, can use a max pooling layer with a stride of 2×2×2 to compress the amount of data and extract significant features.

[0046] Furthermore, to ensure that prior information about underground space can continue to play a role within the inversion network, a priori modulation unit is set at at least one level of the coding path, and the priori modulation unit can enhance the features according to the following formula: ,and , in, Represents the coding features of layer I. This represents the encoded features enhanced by prior modulation. This represents the features obtained by downsampling prior information about underground space. express The corresponding modulation weighting diagram, This represents the Sigmoid activation function.

[0047] The above enhancement methods can suppress invalid features or clutter features that deviate significantly from the prior knowledge of underground space.

[0048] The decoding path is symmetrical to the encoding path and contains four decoding blocks. Each decoding block contains an upsampling layer and a multi-scale feature aggregation module, and the structure of the multi-scale feature aggregation module is consistent with that of the multi-scale feature aggregation module in the encoding path. It is used to reconstruct spatial details from the upsampled features. The upsampling layer can be a transposed convolutional layer with a stride of 2×2×2 to increase the resolution of the feature map.

[0049] In addition, the high-resolution multi-scale feature maps output by each coding block in the coding path are directly transmitted to the corresponding level of the decoding block through skip connections, and concatenated with the upsampled feature maps to supplement the high-frequency spatial information (such as the sharpness of pipeline edges) lost during the pooling process.

[0050] The output layer can adopt a dual-output head structure, where the first output head is used to output the relative permittivity distribution of the underground medium. The second output head is used to output the probability distribution of underground targets. The dual-output head structure can recover the parameter distribution of the underground continuous medium on the one hand, and simultaneously obtain the probabilistic representation of the underground target area on the other hand, which facilitates subsequent target identification, boundary extraction and geometric parameter analysis.

[0051] Based on the aforementioned inversion network, prior information about underground space can be encoded into a three-dimensional data volume with the same size as ground-penetrating radar data, and then stitched together in the channel dimension to form the input tensor of the inversion network.

[0052] It should be added here that, in obtaining prior information about underground space It can then be based on a preset threshold. Generate a key inversion area mask from the prior information of this underground space. Preferably, when When, the corresponding voxel is marked as the key inversion region; when At that time, the corresponding voxels are marked as non-key inversion regions. Among them, the threshold... It can be customized based on target category, scene complexity, and mapping reliability.

[0053] Correspondingly, prior information about underground space and / or masking of key inversion areas can be encoded into a three-dimensional data volume with the same size as the ground-penetrating radar data, thereby enabling integration with the ground-penetrating radar data. By concatenating the components, we obtain the corresponding input tensor. .

[0054] Furthermore, based on the acquired prior information about underground space, the ground-penetrating radar data can be pre-enhanced using a spatial weighting method to obtain weighted ground-penetrating radar data. , is represented as:

[0055] in, This represents element-wise multiplication. This represents the prior gain coefficient.

[0056] The above methods can enhance the effective response in key areas and suppress clutter responses in non-key areas.

[0057] Accordingly, the input tensor can be represented as .

[0058] Finally, the input tensor The input is fed into the trained inversion network to obtain the underground dielectric constant distribution and / or underground target probability distribution of the target detection area.

[0059] Understandably, to simultaneously ensure the accuracy of underground dielectric constant inversion, the clarity of underground target boundaries, and consistency with prior knowledge of underground space, a joint loss function can be used to optimize the inversion network during the model training phase. Specifically, the joint loss function can be expressed as: , in, , , These are the weighting coefficients for each loss term. This represents the regression loss term, used to constrain the difference between the underground dielectric constant distribution predicted by the inversion network and the true label. This represents the segmentation loss term, used to constrain the consistency between the underground target probability distribution predicted by the inversion network and the true labels. This represents the boundary gradient loss term, used to enhance the clarity of underground target boundaries.

[0060] Furthermore, considering that the mask coding of key inversion areas is performed together with the prior information of underground space, the corresponding joint loss function can also be expressed as: , in, This represents the prior consistency loss term, used to constrain the output of the inversion network to maintain a soft consistency relationship with the mask of the key inversion region, so that the network can utilize semantic priors without being completely limited by erroneous priors. express The weighting coefficients.

[0061] In one specific embodiment, the prior consistency loss term can be expressed as: , in, This represents the prior confidence weight corresponding to the position of the i-th voxel. This represents the value of the mask for the key inversion region at the i-th voxel. This loss term guides the network to prioritize the prior high-confidence region while preserving the network's ability to correct prior biases based on the actual radar response.

[0062] In addition, for training optimization, the AdamW optimizer can be used, and the learning rate can be updated using a cosine annealing strategy or a piecewise decay strategy.

[0063] Furthermore, to improve the relevance and stability of ground-penetrating radar inversion under different surface semantic conditions, multiple inversion networks can be constructed. For example, network A for shaft-type targets, network B for pipeline-type targets, and network C for non-salient semantic indication areas.

[0064] Network A is preferably used for surface target identification results that include manhole covers, inspection manhole covers, valve manhole covers, fire hydrant bases, or other semantic target areas associated with underground vertical structures; Network B is preferably used for surface target identification results that include storm drain grates, linear road facilities, curb drainage structures, or other semantic target areas associated with underground linear pipeline structures; Network C is preferably used for background areas where no clear underground indicator semantic targets have been identified, or for areas where semantic indicator information is weak and the type of underground structure is unclear.

[0065] Correspondingly, networks A, B, and C use the same inversion network architecture, but their training sample composition, network parameters, and loss function weights will be adjusted accordingly. Specifically, for example... Figure 3 As shown, Network A is trained on the localized concentrated distribution characteristics of shaft-type targets to enhance its ability to recover vertical targets, nodal targets, and locally strong anomaly structures; Network B is trained on the continuous extension characteristics of pipeline-type targets to enhance its ability to model the continuity, orientation, and depth variations of underground linear targets; Network C is trained on samples from ordinary background areas to handle conventional underground structure inversion tasks that lack explicit surface semantic guidance. Therefore, in practical applications, the appropriate inversion network can be selected based on the surface semantic labels of the currently acquired ground-penetrating radar data.

[0066] After completing the semantically guided inversion reconstruction and obtaining the underground dielectric parameter distribution and underground target probability distribution results of the target detection area, the underground anomaly area can be extracted, parameters analyzed, and results output based on the inversion results.

[0067] Specifically, the probability distribution of underground targets is first segmented using a threshold to extract candidate underground anomaly regions, and then connected component analysis is used to obtain mutually independent underground target regions. If necessary, morphological filtering, void filling, and noise removal can be further performed to improve the stability of target region extraction.

[0068] Secondly, for each extracted underground target area, its target parameters are further calculated. These target parameters include one or more of the following: target center location, burial depth range, geometric dimensions, extension direction, and target category. The target center location can be determined based on the geometric center or weighted center of the underground target area, the burial depth range can be determined based on the distribution of the underground target area in the depth direction, the geometric dimensions can be determined based on the bounding box size or equivalent size of the underground target area, and the extension direction can be determined based on the principal axis direction of the underground target area.

[0069] Simultaneously, consistency verification can be performed between the inversion results and prior information on underground space. When the inversion results and prior information on underground space have high consistency in spatial location, extension direction, or target category, the credibility of the extracted underground target area can be improved. When there is a significant deviation between the two, the credibility of the extracted underground target area can be reduced, or it can be marked as a target to be verified.

[0070] In addition, based on the probability response intensity, spatial continuity, boundary clarity, and consistency with prior information about underground space, the underground target areas can be classified into credibility levels or risk levels, and a list of underground target areas can be generated.

[0071] In practical applications, the results of underground dielectric parameter distribution, underground target area extraction, and underground target area list can be visualized and output. The output formats include one or more of the following: two-dimensional profile map, depth slice map, three-dimensional target distribution map, and detection report. These can then be sent to municipal maintenance platforms, underground disease monitoring platforms, or geographic information system platforms for subsequent analysis, verification, and engineering applications.

[0072] It should be noted that the steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they contain the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0073] Example 2 Please see Figure 4 As shown, this application also provides a method for detecting underground targets based on surface semantic information, including: The data acquisition module 10 is used to acquire surface lidar point cloud data and ground penetrating radar data of the target detection area.

[0074] The data processing module 20 is used to perform semantic segmentation on the surface lidar point cloud data to determine at least one surface target with underground indication significance within the target detection area.

[0075] The surface-to-subsurface association module 30 is used to determine the corresponding subsurface area from the preset surface-to-subsurface spatial mapping relationship according to the semantic category of the surface target, and to construct the subsurface spatial feature data of the target detection area based on the three-dimensional volume data of the subsurface area.

[0076] The inversion and reconstruction module 40 is used to combine the underground space feature data with the ground penetrating radar data to perform inversion and reconstruction, so as to determine the underground dielectric constant distribution and / or underground target probability distribution of the target detection area.

[0077] It should be noted that the underground target detection system based on surface semantic information assisted by the above embodiments and the underground target detection method based on surface semantic information assisted by the above embodiment 1 belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the underground target detection method based on surface semantic information assisted by the above embodiment 1 can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above, and no limitation is imposed here.

[0078] Example 3 Please see Figure 5 As shown, this application also provides an electronic device, including a memory 2, a processor 1, and a program stored in the memory and executable on the processor, wherein the processor executes the steps of any of the methods described above.

[0079] The memory includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory, magnetic storage, disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory can be an external storage device of the electronic device, such as a plug-in portable hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, Flash Card, etc. Furthermore, the memory can include both internal and external storage units of the electronic device. The memory can be used not only to store application software and various types of data installed on the electronic device, but also to temporarily store data that has been output or will be output.

[0080] In some embodiments, the processor may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory and calls data stored in the memory to perform various functions and process data of the electronic device. The processor executes the operating system and various installed application programs of the electronic device. The processor executes the application programs to implement the steps in the above method embodiments.

[0081] For example, the program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of program instruction segments capable of performing a specific function, which describe the execution process of the program in the electronic device.

[0082] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute some of the functions of the various embodiments of the present invention.

[0083] In summary, this invention comprehensively utilizes surface lidar point cloud data and underground ground-penetrating radar data to establish a correlation between surface semantic information and underground structural responses, thereby achieving collaborative detection and joint inversion of multi-source heterogeneous information and improving the reliability and accuracy of integrated surface and underground detection.

[0084] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for detecting underground targets based on surface semantic information, characterized in that, include: Acquire surface lidar point cloud data and ground-penetrating radar data of the target detection area; Semantic segmentation is performed on the surface lidar point cloud data to identify at least one surface target with underground indication significance within the target detection area; Based on the semantic category of the surface target, the corresponding underground region is determined from a preset surface-to-underground space mapping relationship, and the underground space feature data of the target detection area is constructed based on the three-dimensional volume data of the underground region. The steps include: for each surface target, extracting its corresponding semantic category, geometric center coordinates, geometric dimensions, and principal axis direction vector; determining the corresponding underground region from the preset surface-to-underground space mapping relationship based on the semantic category of the surface target, and constructing a corresponding prior sub-volume based on the confidence level of the underground region and the geometric center coordinates, geometric dimensions, and principal axis direction vector of the surface target; fusing the prior sub-volumes of all surface targets to obtain underground space prior information, which serves as the underground space feature data of the target detection area; wherein, the prior sub-volume is constructed according to the following formula: ， In the formula, Denotes the i-th prior child body. This represents the baseline confidence level of the underground region corresponding to the i-th surface target. This represents the horizontal distance from the subject to be estimated to the projection axis of the target center on the Earth's surface. This indicates the vertical offset of the subject body relative to the estimated burial depth. This indicates the degree of deviation of the subject from the direction vector of the principal axis. , , Let represent the horizontal tolerance parameter, vertical tolerance parameter, and directional tolerance parameter, respectively, and let the geometric dimensions of the surface target be . At that time, based on the length of the surface target ,Width Determine the horizontal tolerance parameter and orientation tolerance parameters According to the height of surface targets Determine the vertical tolerance parameters; By combining the underground space feature data with the ground penetrating radar data, an inversion and reconstruction is performed to determine the underground dielectric constant distribution and / or underground target probability distribution in the target detection area.

2. The method for detecting underground targets based on surface semantic information as described in claim 1, characterized in that, The steps of semantically segmenting the surface lidar point cloud data to determine at least one surface target with underground indication significance within the target detection area include: The trained 3D point cloud semantic segmentation network is used to classify the surface lidar point cloud data point by point, and the classified surface lidar point cloud data is clustered into instances to obtain at least one surface target.

3. The method for detecting underground targets based on surface semantic information as described in claim 1, the step of fusing prior sub-entities of all surface targets to obtain prior information about underground space includes: The prior sub-volumes of all surface targets are fused using maximum value fusion or weighted superposition fusion to obtain prior information of underground space.

4. The method for detecting underground targets based on surface semantic information as described in claim 1, characterized in that, The steps of combining the underground space feature data with the ground-penetrating radar data to perform inversion and reconstruction to determine the underground dielectric constant distribution and / or underground target probability distribution in the target detection area include: The underground space feature data is encoded according to the data size of the ground penetrating radar data, and the encoded underground space feature data is concatenated with the ground penetrating radar data to obtain the corresponding input tensor; The input tensor is input into the trained inversion network to obtain the underground dielectric constant distribution and / or underground target probability distribution of the target detection area.

5. The method for detecting underground targets based on surface semantic information according to claim 4, characterized in that, The inversion network adopts a three-dimensional fully convolutional encoder-decoder network based on multi-scale feature aggregation, including: an input layer, an encoding path, a bottleneck layer, a decoding path, skip connections, and an output layer; The input layer is used to receive input tensors; The encoding path is used to extract shallow local features and deep abstract features from the input tensor step by step, and generate corresponding feature maps; The decoding path is used to recover the spatial resolution of the feature map output by the encoding path step by step, and to reconstruct the distribution of underground medium parameters based on it; The bottleneck layer is used to compress the feature map generated by the encoding path and pass it to the decoding path; The skip connection is used to pass the high-resolution detail features extracted from the encoding path to the decoding path; The output layer is used to generate the corresponding underground dielectric constant distribution and / or underground target probability distribution based on the output of the decoding path.

6. The method for detecting underground targets based on surface semantic information according to claim 5, characterized in that, At least one level of the encoding path is provided with a priori modulation unit, and the priori modulation unit enhances the features according to the following formula: ,and , in, Represents the coding features of layer I. This represents the encoded features enhanced by prior modulation. This represents the features in the underground space feature data. express The corresponding modulation weighting diagram, This represents the Sigmoid activation function.

7. The method for detecting underground targets based on surface semantic information according to claim 5, characterized in that, The training steps for the inversion network include: The pre-constructed inversion network is back-optimized using the joint loss function to obtain a trained inversion network. The joint loss function includes at least a regression loss term, a segmentation loss term, and a boundary gradient loss term. The regression loss term is used to constrain the difference between the underground dielectric constant distribution predicted by the inversion network and the true label. The segmentation loss term is used to constrain the consistency between the underground target probability distribution predicted by the inversion network and the true label. The boundary gradient loss term is used to enhance the clarity of the underground target boundary.

8. The method for detecting underground targets based on surface semantic information according to claim 1, characterized in that, By combining the underground space feature data with the ground-penetrating radar data for inversion and reconstruction, the underground dielectric constant distribution and / or underground target probability distribution of the target detection area are determined, including: Based on the underground space feature data, a corresponding regional mask is constructed, and inversion reconstruction is performed by combining the underground space feature data, the regional mask, and the ground penetrating radar data to determine the underground dielectric constant distribution and / or underground target probability distribution of the target detection area.

Images