A three-view underground pipeline intelligent detection method based on three-dimensional ground penetrating radar

Abstract

A three-view underground pipeline intelligent detection method based on three-dimensional ground penetrating radar. The traditional method only uses a general model to recognize the radar image of a single B-scan longitudinal profile feature image, and the existing model does not adjust the ground penetrating radar pipeline image features, cannot accurately distinguish the image features of the cavity disease and the pipeline, and has low recognition accuracy. In the present application, the three-view underground pipeline intelligent detection method combines the obtained three-dimensional underground data with the DCO-YOLO model, and obtains the spatial overlap degree of the bounding box in the three-view and the matching pipeline feature information based on the three-dimensional distance intersection ratio algorithm. The spatial overlap degree of the bounding box in the three-view and the matching pipeline feature information are used to complete the rapid recognition of the pipeline features and the three-view matching process.

Description

Technical Field

[0001] This invention specifically relates to an intelligent detection method for underground pipelines based on three-dimensional ground-penetrating radar with three-view diagrams, belonging to the field of underground pipeline detection technology. Background Technology

[0002] As the core infrastructure for maintaining urban water supply, drainage, gas supply, and communication, the structural integrity of urban underground pipelines is directly related to the safe operation of the city. However, due to factors such as aging pipe materials, design defects, complex environmental loads, and insufficient maintenance, pipeline systems are prone to structural damage such as leaks and displacement, which in turn can lead to secondary disasters such as road collapses and urban flooding.

[0003] Traditional pipeline detection methods primarily rely on the principle of electromagnetic induction and are suitable for detecting metallic pipelines. While traditional pipeline-based detection techniques possess some capability, they are inapplicable to non-metallic pipelines, such as plastic and concrete pipes, due to their lack of conductivity, thus increasing the difficulty of detection. Furthermore, traditional pipeline detection suffers from inherent limitations such as poor adaptability to working environments and low detection efficiency, making it difficult to meet the needs of modern large-scale urban pipeline network inspections. With technological advancements, more and more pipelines are being laid using trenchless methods, such as directional drilling and pipe jacking. While this method reduces the damage caused by ground excavation, it also makes it difficult to determine the exact location of pipelines using traditional ground survey methods. In addition, the urban environment contains numerous sources of electromagnetic interference, such as high-voltage power lines and radio towers, which can interfere with pipeline detection signals and affect the accuracy of the results. Ground conditions, such as metal guardrails and traffic conditions, also affect pipeline detection. Metal objects like guardrails can generate electromagnetic interference, while in busy traffic areas, factors such as vehicle vibrations can cause unstable detection signals. In response to this situation, developing efficient pipeline inspection technology is of great engineering significance for assessing the health status of systems and formulating maintenance plans.

[0004] Ground penetrating radar (GPR) technology, with its non-contact electromagnetic detection characteristics, can achieve the organic integration of spatial positioning and structural assessment of underground pipelines by analyzing the reflected signals formed by the difference in the dielectric constant of the strata. However, its single-antenna structure has obvious limitations in the three-dimensional analysis of complex pipeline networks.

[0005] 3D ground-penetrating radar (GPR) simultaneously acquires B-scan (longitudinal profile), C-scan (horizontal slice), and D-scan (transverse profile) scanning data through a multi-channel array antenna system, significantly improving the 3D representation capability of pipeline networks. However, the massive amount of data has led to increasingly prominent problems such as low efficiency and strong subjectivity in manual interpretation. To overcome these limitations, the application of existing deep learning technology in the field of GPR image recognition is gradually emerging. However, existing technologies suffer from low efficiency in 3D data interpretation and insufficient fusion of multi-view features. Traditional methods only use general models to identify single B-scan (longitudinal profile) radar images. Existing models do not make targeted adjustments for the features of GPR pipeline images, making it impossible to accurately distinguish between image features of cavities and pipelines, resulting in low recognition accuracy. Currently, 3D GPR does not fully utilize the three-view features in underground pipeline detection. Relying solely on B-scan images cannot effectively identify the 3D orientation features of underground pipelines, leaving room for improvement in utilization rate, blind spots in the detection range, and affecting detection quality and reliability. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, a three-dimensional ground-penetrating radar-based intelligent detection method for underground pipelines with three-view projection is provided to solve the above problems.

[0007] A method for intelligent detection of underground pipelines based on three-dimensional ground-penetrating radar (GPR) and three-view underground pipelines is proposed. The method combines the acquired underground three-dimensional data with the DCO-YOLO model and then uses the three-dimensional distance intersection-union algorithm to obtain the spatial overlap of the annotation boxes in the three-view and the matching pipeline feature information. The method uses the spatial overlap of the annotation boxes in the three-view and the matching pipeline feature information to complete the rapid identification of pipeline features and the three-view matching process.

[0008] As a preferred approach: Three-dimensional ground-penetrating radar (GPR) is used to acquire underground three-dimensional data of actual roads. After acquiring the underground three-dimensional data, a preprocessing process is performed, including inverse selective discrete Fourier transform, background removal filtering, progressive low-pass filtering, and gain adjustment. The preprocessing process also includes homogenization adjustment of the three-dimensional radar data. The homogenization adjustment process involves multi-view decomposition of the radar data to obtain a three-view image set: B-scan longitudinal profile feature image, C-scan horizontal slice feature image, and D-scan transverse profile feature image. The image parameters of the three-view image set are a depth of 3m, a span of 30m, and a vertical span of 1.32m. The three-view image set is then normalized by adjusting the image size, optimizing the contrast, removing irrelevant patterns, and removing text annotations to form homogenized data, which is used as input image data.

[0009] As a preferred approach: After preprocessing the underground 3D data, the DCO-YOLO model that is used in conjunction with the underground 3D data is adapted. The adaptation process involves improving the DCO-YOLO model to form an initial model that is compatible with the preprocessed homogenized data. The model is then processed according to the spatial location of the underground pipelines corresponding to the actual roads. Specifically, based on the direction of the underground pipelines, the initial model is trained by dividing the dataset into different label categories. The pipeline burial depth is calculated based on the pipeline B-scan longitudinal profile feature image, and the pipeline burial depth is determined to be the distance from the pipeline B-scan longitudinal profile feature image to the ground.

[0010] As a preferred approach: the input image data undergoes a restoration process before calculating 3D-DIoU. The restoration process is as follows:

[0011] Based on the positions of the three views in the input image data, a finite three-dimensional space R with length L, width W, and height H is constructed. The relative positions of the predicted bounding boxes in their respective views are preserved. Each bounding box contains three dimensions (x, y, z) in the three-dimensional space R. When the input image data is the x-data of the pipeline feature bounding box 1-B on the B-Scan view, the corresponding transformation formula is:

[0012]

[0013] In the above formula, x B1 and x B2 These represent the coordinates of the left and right edges of the 1-B annotation box projected onto the spatial area R; X B1 and X B2 These represent the coordinates of the left and right edges of the 1-B bounding box in the input image; main_view(X,Y,width,height) represents storing the coordinates and dimensions of the top left vertex of the B-Scan.

[0014] As a preferred solution: Each bounding box is mapped to space R to ensure that it is supplemented with data of one dimension. The supplementation process is as follows: When the bounding box is a pipe B-scan longitudinal profile feature image bounding box, the supplementation process is that the pipe B-scan longitudinal profile feature image bounding box contains spatial information of pipe width x and height z, but lacks spatial information of length y. The corresponding y information of D-Scan is supplemented, and so on. That is, the z information of B-scan longitudinal profile feature image is given to C-scan horizontal slice feature image, and the x information of C-scan horizontal slice feature image is given to D-scan transverse profile feature image.

[0015] As a preferred approach, the process of obtaining the spatial overlap of the annotation boxes in the three views and matching the pipeline feature information based on the three-dimensional distance intersection-union algorithm is as follows:

[0016] The matching process of the three views is judged by calculating and comparing 3D-DIoU. 3D-DIoU reflects the spatial correlation of the annotation boxes of the three views of the pipeline feature after being mapped to the three-dimensional space R. The formula for calculating 3D-DIoU is:

[0017]

[0018] In the above formula, 3D-DIoU is the intersection volume ratio and union volume of the two 3D boxes; d is the Euclidean distance between the center points of the two 3D boxes; and c is the diagonal length of the smallest circumscribed cube covering the two 3D boxes.

[0019] Calculate the 3D-DIoU value of each bounding box mapped to the 3D space. Compare the 3D-DIoU value of each bounding box mapped to the 3D space with its corresponding set threshold. When the threshold conditions are met, it indicates that the three bounding boxes are the B-scan longitudinal profile feature image, C-scan horizontal slice feature image, and D-scan transverse profile feature image corresponding to the same pipe, that is, the three view boxes are in a successful matching state.

[0020] The beneficial effects of this invention are as follows:

[0021] This invention combines acquired underground 3D data with a DCO-YOLO model and uses a 3D distance intersection-union algorithm to derive the spatial overlap of bounding boxes in the three-view drawings and match pipeline feature information. This forms a detailed and complete processing procedure for underground 3D data. Utilizing the spatial overlap of bounding boxes and matching pipeline feature information in the three-view drawings, it achieves rapid pipeline feature identification and three-view matching, accurately distinguishing between cavities and pipeline image features with high accuracy. This invention's prediction and feature matching of 3D radar data can accurately and efficiently detect the distribution and operational status of underground pipelines, thus providing a basis for health assessment and maintenance planning. The specific advantages of this invention are:

[0022] First, the model in this application improves upon YOLOv11 by replacing the original uu.Upsample module with Dysample, thereby enhancing the model's sensitivity to edge features in ground-penetrating radar images. It also introduces CGLU to improve the C2PSA module, enhancing the local details of pipe features in ground-penetrating radar images. Furthermore, it introduces Outlook Attention to improve the C3K2 module, making the model pay more attention to the linear changes around the pipe waveform and suppressing background noise unrelated to the central features, ultimately achieving the goal of improving detection accuracy.

[0023] Second: This application proposes a concept of mapping the three-view features of B-scan longitudinal profile feature image, C-scan horizontal slice feature image, and D-scan transverse profile feature image to three-dimensional space. In particular, the introduction of 3D-DIoU to estimate the interrelationship of the spatial positions of each view enables accurate matching of the features of each view. By integrating the three-view features to determine the pipeline direction, the detection results are more accurate. Attached Figure Description

[0024] Figure 1 This is a flowchart of a three-view intelligent detection method for underground pipelines based on three-dimensional ground-penetrating radar.

[0025] Figure 2 This is a schematic diagram of the DCO-YOLO underground pipeline three-view intelligent recognition algorithm structure;

[0026] Figure 3 This is a schematic diagram of the Dysample structure;

[0027] Figure 4 Here are schematic diagrams of the GLU and CGLU structures;

[0028] Figure 5 This is a schematic diagram of the OutlookAttention structure;

[0029] Figure 6 A schematic diagram of 3D-DIoU;

[0030] Figure 7 This is a visual comparison diagram of the DCO-YOLO underground pipeline identification network of the present invention with other methods;

[0031] Figure 8 This is a visualization diagram illustrating the pipeline features and spatial location prediction of the present invention. Detailed Implementation

[0032] The following specific examples illustrate the implementation of the present invention. 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.

[0033] Specific implementation method one: Combining Figures 1 to 8This embodiment describes the intelligent detection method for underground pipelines using three-view drawings. It combines the acquired underground 3D data with a DCO-YOLO model and then uses a 3D distance intersection-union algorithm to determine the spatial overlap of the bounding boxes in the three-view drawings and to match pipeline feature information. This spatial overlap and matching information are then used to quickly identify pipeline features and perform the three-view matching process. The 3D distance intersection-union algorithm used in this embodiment is the 3D-DIoU algorithm.

[0034] Combination Figure 1 As shown, the specific operation process of this invention is as follows:

[0035] First, the data comes from the underground three-dimensional data below the actual road. The collected three-dimensional ground-penetrating radar data is preprocessed, specifically including background removal filtering, progressive low-pass filtering, gain adjustment, and other necessary preprocessing processes.

[0036] Secondly, based on the pipeline direction characteristics, the preprocessed data is labeled and divided into training and validation sets. The core C3k2, C2PSA and Upsample modules in the YOLOv11 framework are improved, and the improved model is used for training and prediction.

[0037] Then, the two-dimensional view of the predicted three-dimensional ground-penetrating radar image is restored to three-dimensional space, and the matching of the three-view features of the pipeline is realized based on 3D-DIoU;

[0038] Finally, the predicted matching pipe features and other pipe features are drawn using different colored bounding boxes. The corresponding pipe type and burial depth are output.

[0039] This invention collects underground data using three-dimensional ground-penetrating radar, preprocesses it to generate a set of B-Scan, C-Scan, and D-Scan three-view images, uses the DCO-YOLO model for feature recognition, and maps the three-view recognition boxes to three-dimensional space based on the three-dimensional distance intersection-union algorithm. It calculates the spatial overlap to match pipeline features, thus realizing the process of pipeline orientation classification and burial depth calculation.

[0040] This invention has been experimentally verified. Experiments show that, in the specific implementation process, the detection accuracy P, recall rate R, and average precision of this invention reach 96.2%, 93.3%, and 96.7% respectively in 100 kilometers of measured data. The data information is significantly better than that of traditional models, and it can provide an efficient and high-precision solution for intelligent detection of underground pipelines.

[0041] Specific Implementation Method Two: This implementation method is a further limitation of Specific Implementation Method One. The intelligent detection method for underground pipelines in three-view drawings in this implementation method is specifically divided into the following steps:

[0042] Step 1: Acquire 3D underground data of the actual road using 3D ground-penetrating radar. Preprocess the data across the three dimensions to generate a three-view image set. Specifically, preprocess the raw data acquired by the frequency-stepping 3D ground-penetrating radar, including inverse discrete Fourier transform, background removal filtering, progressive low-pass filtering, and gain adjustment.

[0043] Specifically, considering the frequency domain signal characteristics of frequency-stepping three-dimensional radar, the frequency domain signal is converted into a time domain signal through inverse discrete Fourier transform to optimize the time domain resolution. The formula is as follows:

[0044]

[0045] In the above formula, S(f k ) represents the k-th frequency component, N represents the number of frequency sampling points, and s(t) represents the time-domain signal.

[0046] Background removal filtering eliminates interference signals caused by antenna direct-coupled waves, ground reflections, and system noise, highlighting effective reflection characteristics. Progressive low-pass filtering suppresses high-frequency electronic noise and low-frequency ground reflection interference, eliminating periodic noise and smoothing random noise. Gain adjustment compensates for signal attenuation effects, enhancing the visibility of deep, weak targets and ensuring the comparability of reflected energy from different media in the image.

[0047] Step 2: Label the preprocessed data and divide it into training and validation sets, specifically:

[0048] Based on the pipe orientation, the dataset is divided into different annotation categories. For vertical pipes, the B-scan longitudinal profile feature image is 1-B, the C-scan horizontal slice feature image is 1-C, and the D-scan transverse profile feature image is 1-D; for horizontally inclined pipes, the B-scan longitudinal profile feature image is 1-B, the C-scan horizontal slice feature image is 2-C, and the D-scan transverse profile feature image is 2-D; for depth-oriented inclined pipes, the B-scan longitudinal profile feature image is 1-B, the C-scan horizontal slice feature image is 3-C, and the D-scan transverse profile feature image is 3-D.

[0049] Since the hyperbolic features of the corresponding B-Scan are not significantly different regardless of the direction of the pipeline, the longitudinal profile feature images of the B-Scan for the three types of pipelines are all 1-B.

[0050] LabelImg software was used to annotate multi-view GPR images, extracting pipeline geometric features and spatial attributes. Annotation quality was validated through pre-training to optimize feature recognition accuracy. The dataset was divided into training and validation sets in an 8:2 ratio, and all annotation files were uniformly converted to .txt format for easier training of the subsequent object recognition model.

[0051] Step 3: Combining Figure 2 As shown, training and prediction are based on the DCO-YOLO model. The core Upsample, C2PSA, and C3k2 modules in the original YOLOv11 framework are improved by introducing DySample dynamic upsampling, CGLU local feature awareness, and OutlookAttention cross-dimensional correlation mechanism. This significantly enhances the representation ability of small-scale target edge features, making the model pay more attention to the local details and edge features of the pipeline waveform. The improved DCO-YOLO model is used for training and prediction, specifically in the following parts:

[0052] Part One: Combination Figure 3 As shown, the DySample dynamic upsampling module is used to replace the traditional Upsample upsampling, and content-aware feature reconstruction is achieved through two-stage offset generation and pixel rearrangement. For the input low-resolution feature map X∈R... H×W×C DySample first generates the position offset of each target sampling point directly through a lightweight linear layer. Where s is the upsampling factor. The dimension of the offset is 2s. 2 A dynamic adjustment mechanism is introduced, generating a dynamic range factor O through a linear layer and a sigmoid function. The formula for O is:

[0053] O=0.5sigmoid(linear1(X)Linear2(X)) (1)

[0054] In the above formula, the sigmoid function, which is σ in the figure, can map any real number input to the range of (0,1) and is used to control the range of sampling points.

[0055] Pixel Shuffle redistributes channel-dimensional information into the spatial dimension, thus achieving upsampling. The offset O is superimposed on the standard upsampling grid G ​​to form a dynamic sampling grid S. The original grid G ​​represents the fixed sampling position of the bilinear interpolation. Subsequently, the grid_sample function is used to resample the input feature X according to the dynamic grid S, obtaining the upsampled feature X′.

[0056] X' = ​​grid_sample(X,S) (2)

[0057] Part Two: Combination Figure 4The illustrated structure introduces a Convolutionally Gated Linear Unit (CGLU) into the C2PSA module. It extracts neighborhood features through deep convolution to generate position-related channel weights, including bilinear transformation branches: a value branch where the input undergoes a linear transformation to generate main features, preserving global information; and a gate branch where the input undergoes a linear transformation before being activated by a Gaussian Error Linear Unit (GELU). However, the weights of its gate branch originate from the input itself, resulting in a limited receptive field. The improved CGLU (ConvolutionalGLU) builds upon the GLU by adding a 3×3 deep convolution before the GLU's gate branch, utilizing neighboring pixel features to generate locally perceptual gated signals. Deep convolution extracts features around the token, assigning unique channel attention weights to each position, thus addressing the coarse-grained problem of SE (Search Engine).

[0058] The output of the gated branch (local features) is multiplied element-wise with the output of the value branch (global features) to generate fine-grained channel attention weights, as shown in the following formula.

[0059] ConvGLU(X)=Linear1(X)⊙GELU(DepthwiseC onv(Linear2(X)))(3)

[0060] In the above formula, DepthwiseCONV (DWConv) is a 3×3 depthwise convolution, and ⊙ represents element-wise multiplication. CGLU not only focuses on the global features of the input image but also observes the local features around each pixel, dynamically adjusting the channel weights at different locations. CGLU introduces a 3×3 depthwise convolution in its gated branch to dynamically extract neighborhood features and generate location-related channel weights. In the time-frequency images of radar signals, edge and phase transition features in different regions can be enhanced through this module.

[0061] Part Three: Combination Figure 5 As shown, the OutlookAttention mechanism is embedded in the C3K2 module, which aggregates multi-granularity features based on local window attention, and processes the input feature map X∈R. H×W×C Values ​​V and attention weights A are generated through linear projection, and a local window (K×K) for each center token is extracted using an unfold operation, resulting in... The central token is generated through a linear layer. After reshaping, the data are normalized using the Softmax function. Local features are then aggregated using matrix multiplication.

[0062]

[0063] In the above formula, Y Δi,jRepresenting the aggregated local features, Matmul is a matrix multiplication function that sums the attention weights and values ​​in a weighted manner.

[0064] The aggregation result of the local window is restored to the global feature map by the Fold operation, as shown in the following formula.

[0065]

[0066] In the above formula, the aggregated result Y with input (i,j) on the right is... Δi,j The final output is obtained by summing the aggregated results of all local windows covering positions (i,j).

[0067] The detection performance of the model is evaluated using accuracy, recall, and mean precision.

[0068] Step 4: Implement pipeline feature recognition processing based on the DCO-YOLO model. Specifically, use the DCO-YOLO trained in Step 3 to accurately identify targets in the three-dimensional ground-penetrating radar three-view images of underground pipelines.

[0069] Step 5: Reconstruct the three 2D views of the identified 3D ground-penetrating radar image into 3D space, and match the features of the three views of the pipeline based on 3D-DIoU, specifically:

[0070] Based on the positions of the three views in the input image (x-coordinate of the top-left vertex, y-coordinate of the top-left corner, width, and height), a finite three-dimensional space R with length L, width W, and height H is constructed. The relative positions of the predicted bounding boxes in their respective views are preserved. Each bounding box contains three dimensions of data (x, y, z) in the three-dimensional space R. Taking the x-data of the pipeline feature bounding box 1-B on the B-Scan view as an example, the conversion formula is:

[0071]

[0072] In the above formula, x B1 ,x B2 , representing the coordinates of the left and right edges of the 1-B annotation box projected onto the space R. B1 X B2 These represent the coordinates of the left and right edges of the 1-B bounding box in the input image, respectively. main_view(X,Y,width,height) stores the coordinates and dimensions of the top left vertex of the B-Scan.

[0073] The identification box 1-B in the main view contains spatial information about the pipe width (x) and height (z), but lacks spatial information about the length (y). This information can be supplemented from the y information of nD. Similarly, the z information of 1-B can be supplemented to nC, and the x information of nC can be supplemented to nD.

[0074] Furthermore, the process of determining the matching of three views by calculating and comparing 3D-DIoU includes:

[0075] Combination Figure 6 As shown, 3D-DIoU reflects the spatial correlation of 1-B, nC, nD after mapping to three-dimensional space R. The formula for calculating 3D-DIoU is as follows:

[0076]

[0077] In the above formula, 3D IoU is the ratio of the intersection volume to the union volume of the two 3D bounding boxes, d is the Euclidean distance between the center points of the two 3D bounding boxes, and c is the diagonal length of the smallest circumscribed cube covering the two 3D bounding boxes.

[0078] Calculate the 3D-DIoU values ​​mapped from 1-B to nC and 1-B to nD in three-dimensional space, and compare them with the set threshold of 0.4. When the threshold conditions are met, it means that the three recognition boxes are the B-scan longitudinal profile feature image, C-scan horizontal slice feature image, and D-scan transverse profile feature image corresponding to the same pipe, that is, the three view boxes are matched.

[0079] Step Six: Visualize the detection results. When a set of matching pipe features is identified in the input image, the burial depth is calculated based on the position of the 1-B bounding box. The burial depth is represented by the actual distance from the upper edge of the 1-B bounding box to the ground. The predicted set of matching pipe features and the remaining pipe features are drawn with different colored bounding boxes. The corresponding pipe type and burial depth are output according to the category of the matching pipe features.

[0080] Specific Implementation Method 3: This implementation method is a further limitation of Specific Implementation Method 1 or 2. In this implementation method, three-dimensional underground data of actual roads are collected by three-dimensional ground penetrating radar. After acquiring the underground three-dimensional data, a preprocessing process is performed. The preprocessing process includes inverse selective discrete Fourier transform, background removal filtering, progressive low-pass filtering, and gain adjustment. The preprocessing process also includes homogenization adjustment of the three-dimensional radar data. The homogenization adjustment process is to obtain a three-view image set by multi-view decomposition of the radar data, namely B-scan longitudinal profile feature image, C-scan horizontal slice feature image, and D-scan transverse profile feature image. The image parameters of the three-view image set are depth 3m, span 30m, and vertical span direction 1.32m. The three-view image set is then processed by standardization to adjust the image size, optimize contrast, remove irrelevant patterns, and remove text annotations to form homogenized data, which is used as input image data.

[0081] Specific Implementation Method Four: This implementation method is a further limitation of Specific Implementation Methods One, Two, or Three. In this implementation method, after preprocessing the underground 3D data, the DCO-YOLO model that is matched with the underground 3D data is adapted. The adaptation process involves improving the DCO-YOLO model to form an initial model that is adapted to the preprocessed homogenized data. The model is processed according to the spatial location of the underground pipeline corresponding to the actual road. Specifically, based on the underground pipeline route, the initial model is trained by dividing the dataset into different label categories. The pipeline burial depth is calculated based on the pipeline B-scan longitudinal profile feature image, and the pipeline burial depth is the distance from the pipeline B-scan longitudinal profile feature image to the ground.

[0082] Specific Implementation Method 5: This implementation method is a further limitation of Specific Implementation Methods 1, 2, 3 or 4. In this implementation method, the input image data is restored before 3D-DIoU is calculated. The restoration process is to restore the three two-dimensional views of the input image data to a three-dimensional space. The process is to construct a finite three-dimensional space R with length L, width W and height H based on the position of the three views in the input image, keep the relative position of the predicted recognition box in the corresponding view unchanged, and add information of one dimension to convert it into a three-dimensional annotation box in space R.

[0083] The specific restoration process is as follows:

[0084] Based on the positions of the three views in the input image data, a finite three-dimensional space R with length L, width W, and height H is constructed. The relative positions of the predicted bounding boxes in their respective views are preserved. Each bounding box contains three dimensions (x, y, z) in the three-dimensional space R. When the input image data is the x-data of the pipeline feature bounding box 1-B on the B-Scan view, the corresponding transformation formula is:

[0085]

[0086] In the above formula, x B1 and x B2 These represent the coordinates of the left and right edges of the 1-B annotation box projected onto the spatial area R; X B1 and X B2 These represent the coordinates of the left and right edges of the 1-B bounding box in the input image; main_view(X,Y,width,height) represents storing the coordinates and dimensions of the top left vertex of the B-Scan.

[0087] Specific Implementation Method Six: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, or Five. In this implementation method, each recognition box is ensured to be supplemented with data of one dimension in the space R. The supplementation process is as follows: When the recognition box is a pipe B-scan longitudinal profile feature image recognition box, the supplementation process is that the pipe B-scan longitudinal profile feature image recognition box contains spatial information of pipe width x and height z, but lacks spatial information of length y. The corresponding y information of D-Scan is supplemented, and so on. That is, the z information of B-scan longitudinal profile feature image is given to C-scan horizontal slice feature image, and the x information of C-scan horizontal slice feature image is given to D-scan transverse profile feature image.

[0088] Specific Implementation Method Seven: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, or Six. In this implementation method, the process of obtaining the spatial overlap of the annotation boxes in the three views and matching the pipeline feature information based on the three-dimensional distance intersection-union algorithm is as follows:

[0089] The matching process of the three views is judged by calculating and comparing 3D-DIoU. 3D-DIoU reflects the spatial correlation of the annotation boxes of the three views of the pipeline feature after being mapped to the three-dimensional space R. The formula for calculating 3D-DIoU is:

[0090]

[0091] In the above formula, 3D-DIoU is the intersection volume ratio and union volume of the two 3D boxes; d is the Euclidean distance between the center points of the two 3D boxes; and c is the diagonal length of the smallest circumscribed cube covering the two 3D boxes.

[0092] Calculate the 3D-DIoU value of each bounding box mapped to the 3D space. Compare the 3D-DIoU value of each bounding box mapped to the 3D space with its corresponding set threshold. When the threshold conditions are met, it indicates that the three bounding boxes are the B-scan longitudinal profile feature image, C-scan horizontal slice feature image, and D-scan transverse profile feature image corresponding to the same pipe, that is, the three view boxes are in a successful matching state.

[0093] When this invention is deployed for the detection of underground pipelines beneath roads, a frequency-stepping three-dimensional ground-penetrating radar is used to detect underground pipelines over 100km of urban roads, covering typical distribution areas of the urban underground pipe network. Preprocessing is performed according to the above steps, followed by model experiments and comparative analysis.

[0094] Combination Figure 7 A comparative analysis of the recognition performance was conducted with the model test results in Table 1, specifically as follows:

[0095] Table 1. Results of Model Comparison Experiments

[0096]

[0097] According to the data in the table above, the DCO-YOLO model used in this invention has significant advantages over other deep learning models in the experiment. It improves the accuracy by 2%, recall by 2.1%, and mean precision by 0.9 compared to the original YOLOv11 model, thus verifying the effectiveness of the multi-dimensional feature enhancement strategy.

[0098] Combination Figure 8 As shown, the visualization results of pipeline feature and spatial location prediction in this invention can accurately identify the pipeline direction and burial depth, thus verifying the feasibility of this invention.

[0099] Based on extensive practical experience and professional knowledge in this field, this invention designs an intelligent detection method for underground pipelines using three-dimensional ground-penetrating radar (GPR) based on three-view calculations. Addressing key technical challenges in existing GPR data processing, such as insufficient spatiotemporal feature fusion, inadequate utilization of multi-view correlation, and low accuracy in complex scene recognition, this method innovatively constructs a dual-strategy collaborative detection framework. Through bidirectional optimization of three-dimensional data multi-view feature fusion and an improved DCO-YOLO target detection model, accurate identification and spatial positioning of underground pipelines are achieved.

Claims

1. A method for intelligent detection of underground pipelines based on three-dimensional ground-penetrating radar (GPR) with three-view projections, characterized in that: The intelligent detection method for underground pipelines using three-view drawings combines the acquired underground three-dimensional data with the DCO-YOLO model and then uses the three-dimensional distance intersection-union algorithm to obtain the spatial overlap of the annotation boxes in the three-view drawings and the matching pipeline feature information. The method then uses the spatial overlap of the annotation boxes in the three-view drawings and the matching pipeline feature information to complete the rapid identification of pipeline features and the three-view matching process. Underground data is collected by three-dimensional ground-penetrating radar, and B-Scan, C-Scan, and D-Scan three-view image sets are generated after preprocessing. The DCO-YOLO model is used for feature recognition, and the three-view recognition boxes are mapped to three-dimensional space based on the three-dimensional distance intersection-union algorithm. The spatial overlap is calculated to match pipeline features, realizing the pipeline orientation classification and burial depth calculation process. The DCO-YOLO model is obtained by replacing the original YOLOv11 upsampling with the DySample dynamic upsampling module, introducing the convolutional gated linear unit CGLU in the C2PSA module, and embedding the OutlookAttention mechanism in the C3K2 module. The input image data is uniformized data formed by standardizing the three-view image set, adjusting the image size, optimizing the contrast, removing irrelevant patterns and text annotations; The input image data undergoes a restoration process before calculating 3D-DIoU. The restoration process is as follows: Based on the positions of the three views in the input image data, a finite three-dimensional space R with length L, width W, and height H is constructed, preserving the relative positions of the predicted bounding boxes in their respective views. Each bounding box contains three dimensions of data (x, y, z) in the three-dimensional space R. When the input image data is the x data of the pipeline feature bounding box 1-B on the B-Scan view, the corresponding transformation formula is: ， ， In the above formula, and These represent the coordinates of the left and right edges of the 1-B annotation box, projected onto space R. and These represent the coordinates of the left and right edges of the 1-B bounding box in the input image; main_view(X, Y, width, height) stores the coordinates and dimensions of the top left vertex of the B-Scan. Each bounding box is mapped to space R to ensure that it is supplemented with data of one dimension. The supplementation process is as follows: when the bounding box is a pipe B-scan longitudinal profile feature image bounding box, the supplementation process is that the pipe B-scan longitudinal profile feature image bounding box contains spatial information of pipe width x and height z, but lacks spatial information of length y. The corresponding y information of D-Scan is supplemented, and so on. That is, the z information of B-scan longitudinal profile feature image is given to C-scan horizontal slice feature image, and the x information of C-scan horizontal slice feature image is given to D-scan transverse profile feature image. The process of obtaining the spatial overlap of the annotation boxes in the three views and matching the pipeline feature information based on the three-dimensional distance intersection-union algorithm is as follows: The matching process of the three views is judged by calculating and comparing 3D-DIoU. 3D-DIoU reflects the spatial correlation of the annotation boxes of the three view features of the pipeline after being mapped to the three-dimensional space R. The formula for calculating 3D-DIoU is: In the above formula, 3D IoU is the ratio of the intersection volume to the union volume of the two 3D bounding boxes; d is the Euclidean distance between the center points of the two 3D bounding boxes; and c is the diagonal length of the smallest circumscribed cube covering the two 3D bounding boxes. Calculate the 3D-DIoU value mapped to the three-dimensional space for each bounding box. Compare the 3D-DIoU value mapped to the three-dimensional space for each bounding box with its corresponding set threshold. When the threshold conditions are met, it indicates that the three bounding boxes are the B-scan longitudinal profile feature image, C-scan horizontal slice feature image, and D-scan transverse profile feature image corresponding to the same pipe, that is, the three view boxes are in a successful matching state.

2. The intelligent detection method for underground pipelines based on three-dimensional ground-penetrating radar (GPR) with three-view projection as described in claim 1, characterized in that: Three-dimensional underground data of actual roads are collected by three-dimensional ground-penetrating radar. After obtaining the underground three-dimensional data, a preprocessing process is carried out, which includes inverse selective discrete Fourier transform, background removal filtering, progressive low-pass filtering, and gain adjustment. The preprocessing process also includes homogenization adjustment of the three-dimensional radar data. The homogenization adjustment process is to obtain a three-view image set by multi-view decomposition of radar data, namely B-scan longitudinal profile feature image, C-scan horizontal slice feature image, and D-scan transverse profile feature image. The image parameters of the three-view image set are depth 3m, span 30m, and vertical span direction 1.32m.

3. The intelligent detection method for underground pipelines based on three-dimensional ground-penetrating radar (GPR) with three-view projection as described in claim 2, characterized in that: After preprocessing the underground 3D data, the DCO-YOLO model associated with the underground 3D data is adapted. The adaptation process involves improving the DCO-YOLO model to form an initial model that is compatible with the preprocessed homogenized data. The model is then processed based on the spatial location of the underground pipelines corresponding to the actual roads. Specifically, based on the underground pipeline route, the initial model is trained by dividing the dataset into different label categories. The pipeline burial depth is calculated based on the pipeline B-scan longitudinal profile feature image, and the pipeline burial depth is determined as the distance from the pipeline B-scan longitudinal profile feature image to the ground.

Images