A method and system for identifying underground targets

Abstract

The embodiment of the present application discloses a method and system for identifying underground targets, which extracts feature maps from a ground penetrating radar three-dimensional data set C-scan and a two-dimensional data set Top-scan, respectively, C-scan 3D feature map and ROI focus feature map, and fuses the two types of feature maps, the fused feature map can focus on those underground disease targets which cannot be distinguished in C-scan but have obvious features in Top-scan, in addition, the underground disease usually does not have a fixed shape when reflected in the Top-scan plane view, therefore, the fusion of the ROI focus feature map and the C-scan 3D feature map can mask the interference caused by the noise extracted from the C-scan in the fused feature map, and enhance the disease features. The present application can accurately identify the type and position of the underground target.

Description

Technical Field

[0001] This invention relates to ground-penetrating radar technology for detecting and identifying underground targets, and more particularly to a method and system for identifying underground targets. Background Technology

[0002] Airport runways, as the area for aircraft takeoff and landing, are essential infrastructure for ensuring safe aircraft operations. Due to environmental factors and repeated landings, runway structures inevitably deteriorate over time, leading to various underground defects such as delamination, settlement, and cracks. As these underground defects develop, the stress state of the airport runway surface changes, seriously threatening aircraft operational safety. Early detection of underground defects is necessary to ensure aircraft safety and reduce maintenance costs.

[0003] Ground-penetrating radar (GPR) is one of the most important tools in non-destructive testing (NDT). The interpretation of GPR data primarily relies on experts in the field. However, human interpretation is highly subjective, time-consuming, and costly. Therefore, it is necessary to develop automated methods for detecting underground targets using GPR data analysis.

[0004] Existing detection methods often rely solely on analyzing B-scan images to identify underground targets. Since voids, subsidence, and cracks exhibit similar reflection patterns in B-scan images, these methods may have some effectiveness in identifying underground targets such as pipelines, but they are prone to misidentifying voids, subsidence, and cracks. Furthermore, these detection methods can only classify underground targets, not pinpoint their location. Summary of the Invention

[0005] The main objective of this invention is to propose an underground target identification method and system, which aims to solve the aforementioned problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides a method for identifying underground targets, the method comprising:

[0007] Receives C-scan data from ground-penetrating radar of the underground area to be detected;

[0008] Based on the C-scan, a C-scan 3D feature map is obtained;

[0009] Receive multiple top-scan profiles, each of which corresponds to 2D data of the C-scan at different depth planes;

[0010] Based on the multiple profile top-scans, multiple top-scan feature maps and multiple 2D candidate boxes corresponding to the multiple top-scan feature maps are obtained respectively;

[0011] The multiple 2D candidate boxes are mapped to the corresponding Top-scan feature maps, and the values ​​of the parts of the Top-scan feature maps that do not correspond to the 2D candidate boxes are set to 0, thereby obtaining multiple ROI focusing feature maps;

[0012] The C-scan 3D feature map and the multiple ROI focused feature maps are fused to obtain a fused feature map;

[0013] Based on the fused feature map, the category and location of underground targets in the underground area to be detected are determined.

[0014] In some embodiments, the step of obtaining the C-scan 3D feature map based on the C-scan includes:

[0015] The C-scan is divided into multiple equally spaced voxels;

[0016] The multiple voxels are sequentially processed through 3D convolutional layers and a 3D dense network to obtain feature map x. l ;

[0017] Based on the feature map x l The channel attention map Atte is obtained. rc ;

[0018] Atte rc =σ(F c (AvgPool(x l ))+F c (MaxPool(x l )));

[0019] Where σ is the sigmoid activation function, F c It is a fully connected layer, AvgPool is average pooling, and MaxPool is max pooling;

[0020] The channel attention map Atte rc and the feature map x l Element-wise multiplication is performed to obtain the C-scan 3D feature map.

[0021] In some embodiments, the C-scan includes a series of B-scan images obtained by multiple parallel scans of the underground area to be detected by a multi-channel ground-penetrating radar;

[0022] The voxel contains the B-scan image obtained in a single scan of the multi-channel ground-penetrating radar.

[0023] In some embodiments, the 3D convolutional layer consists of 1x, 2x, 4x, and 8x downsampling convolutional layers, and the 3D dense network consists of three dense layers.

[0024] In some embodiments, the step of obtaining multiple Top-scan feature maps and multiple 2D candidate boxes corresponding to the multiple Top-scan feature maps based on the multiple profile Top-scan includes:

[0025] The multiple top-scan profile maps are processed sequentially through 2D convolutional layers and 2D dense networks to obtain multiple top-scan feature maps;

[0026] The multiple Top-scan feature maps are processed by object classification and 2D bounding box regression to obtain 2D candidate boxes;

[0027] The 2D bounding box regression has a loss function L tr satisfy:

[0028]

[0029] Among them, t i =[x,y,w,h] T , represents the labeled 2D border, [x,y] T Let [w, h] be the center coordinates of the 2D bounding box. T The width and height of the 2D border. This is for predicting the bounding box.

[0030] In some embodiments, the step of fusing the C-scan 3D feature map and the plurality of ROI focusing feature maps to obtain a fused feature map includes:

[0031] Through interpolation and transformation, the multiple ROI focused feature maps and the C-scan 3D feature map are made to have the same size in the planar dimension of the ROI focused feature map;

[0032] The C-scan 3D feature map and the multiple ROI focused feature maps are added element-wise to obtain a fused feature map.

[0033] In some embodiments, the step of determining the category and location of underground targets in the underground area to be detected based on the fused feature map includes:

[0034] The fused feature map is processed using a deconvolutional sub-network to separate the foreground pixels as initial 3D candidate boxes.

[0035] The initial 3D candidate boxes are mapped to the corresponding positions in the fused feature map, and ROI-pooling is performed.

[0036] The output of the ROI-pooling operation is processed by object classification and bounding box regression to obtain the category and bounding box of the underground target. The location of the underground target is determined based on the bounding box of the underground target.

[0037] In some embodiments, the foreground pixel separation process has a loss function L f (p t )satisfy:

[0038] L f (p t )=-α t (1-p t ) γ logp t ,

[0039]

[0040] Where, α t γ represents the weighting factor, p is the focus parameter, and p is the probability that a pixel belongs to the foreground pixel.

[0041] In some embodiments, the object classification and bounding box regression processing has a loss function L r Satisfy: L r =α∑L c +β∑L h ;

[0042] in,

[0043]

[0044] Where, n c y represents the total number of categories. i and Let b represent the label value and predicted value of a point belonging to the i-th category, respectively. i Let be a vector representing the labeled 3D bounding box. Let δ be a vector representing the predicted 3D bounding box, where δ is a parameter and α and β are balancing weight factors.

[0045] Furthermore, to achieve the above objectives, the present invention also proposes an underground target recognition system, which includes a 3DCNN module, a 2D CNN module, and a multi-view fusion network module;

[0046] The 3D CNN module is used to receive the ground-penetrating radar C-scan of the underground area to be detected and output the C-scan 3D feature map;

[0047] The 2D CNN module is used to receive multiple top-scan profiles, each corresponding to 2D data of the C-scan at different depth planes; based on the multiple top-scan profiles, multiple top-scan feature maps and multiple 2D candidate boxes corresponding to each top-scan feature map are obtained; the multiple 2D candidate boxes are mapped to the corresponding top-scan feature maps, and the values ​​of the parts of the top-scan feature maps that do not correspond to the 2D candidate boxes are set to 0, thus obtaining multiple ROI focusing feature maps;

[0048] The multi-view fusion module is used to fuse the C-scan 3D feature map and the multiple ROI focused feature maps to obtain a fused feature map; based on the fused feature map, the category and location of underground targets in the underground area to be detected are determined.

[0049] The proposed method and system for underground target identification involves receiving three-dimensional volumetric data (C-scan) of the underground area to be detected and obtaining 2D profile maps (TOP-scan) at different depths based on the C-scan. The C-scan is then processed using a 3D neural network module to obtain C-scan 3D feature maps. Multiple Top-scan feature maps and corresponding 2D candidate boxes are obtained based on these multiple profile maps. The values ​​of non-corresponding 2D candidate boxes in the multiple Top-scan feature maps are set to 0, resulting in multiple Region of Interest (ROI) focused feature maps. Since some underground defects exhibit significant features in the TOP-scan, these significant features can be preserved by obtaining the ROI focused feature maps. However, these defects are not clearly distinguishable in the C-scan, resulting in corresponding noise in the C-scan 3D feature maps. Fusing the ROI focused feature maps with the C-scan 3D feature maps can mask the interference from the noise extracted from the C-scan in the fused feature map and enhance the defect features, thereby accurately identifying the type of underground target in the fused feature map.

[0050] Furthermore, TOP-scan is obtained by slicing C-scan at different depth planes. Therefore, the C-scan 3D feature map based on C-scan and the ROI focusing feature map based on TOP-scan can be effectively fused together to form a 3D fused feature map. Then, the specific location of the underground target can be obtained based on the 3D fused feature map, which solves the problem that it is difficult to determine the location of the underground target in the depth direction based solely on the Top-scan feature map. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] The methods, systems, and / or procedures shown in the accompanying drawings will be further described with reference to exemplary embodiments. These exemplary embodiments will be described in detail with reference to the drawings. These exemplary embodiments are non-limiting exemplary embodiments, wherein reference numerals in the various views of the drawings represent similar mechanisms.

[0053] Figure 1 This application includes a ground-penetrating radar (GPR) antenna and 3D GPR data layout diagram for some embodiments.

[0054] Figure 2 This application includes some embodiments of the detection robot and GPR data collection process diagrams.

[0055] Figure 3 This is a flowchart illustrating some embodiments of the underground target identification method described in this application.

[0056] Figure 4 This is a diagram illustrating the architecture of an underground target recognition system (MV-GPRNet) based on some embodiments of this application.

[0057] Figure 5 This is a comparison chart of the detection results of MV-GPRNet in this application with four other detection methods.

[0058] Figure 6 This is a comparison chart showing the detection results in MV-GPRNet of this application with and without the dense layer module and radar channel attention module configured.

[0059] Figure 7 This is a comparison chart of detection results for the dense layer module and radar channel attention module in MV-GPRNet of this application with different configurations.

[0060] Figure 8 This is a comparison chart showing the detection results in MV-GPRNet of this application with and without the use of Top-scan. Detailed Implementation

[0061] To better understand the above technical solutions, the technical solutions of this application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of this application and the specific features in the embodiments are detailed descriptions of the technical solutions of this application, rather than limitations on the technical solutions of this application. In the absence of conflict, the embodiments of this application and the technical features in the embodiments can be combined with each other.

[0062] In the detailed description below, numerous specific details are illustrated with examples to provide a comprehensive understanding of the relevant guidance. However, it will be apparent to those skilled in the art that this application can be practiced without these details. In other instances, well-known methods, procedures, systems, components, and / or circuits have been described at a relatively high level without detail to avoid unnecessarily obscuring aspects of this application.

[0063] These and other characteristics, the functions disclosed in the current application, the methods of execution, the functions of related elements in the structure, the combination of components, and the economic efficiency of production, may become more apparent in consideration of the following description with reference to the accompanying drawings, all of which form part of this application. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this application. It should be understood that these drawings are not drawn to scale.

[0064] This application uses flowcharts to illustrate the execution process performed by a system according to embodiments of this application. It should be clearly understood that the execution processes in the flowcharts may not be executed sequentially. Instead, these execution processes may be executed in reverse order or simultaneously. Additionally, at least one other execution process may be added to the flowchart. One or more execution processes may be deleted from the flowchart.

[0065] Please refer to Figure 1 and Figure 2 This application utilizes a detection robot to collect information from the underground area to be inspected. The detection robot is equipped with a multi-channel ground-penetrating radar (GPR). The detection robot moves along a preset trajectory within the corresponding ground area and completes the inspection work, such as... Figure 2 As shown in the upper left corner, the detection robot collects information in the xy-plane where the ground is located. During the data collection task, the robot moves from one end to the other along the positive / negative y-direction, completing the collection of a straight line segment. Then, the robot moves a certain distance in the x-direction and begins moving along the negative / positive y-direction, repeating the above process to complete the data collection. Thus, the entire trajectory has multiple parallel line segments. For example... Figure 1As shown, the multi-channel GPR antenna array is fixed perpendicular to the robot's movement method, meaning the antenna array is perpendicular to the y-direction and the antennas are equidistantly distributed along the array. During data acquisition, the GPR antenna transmitter of each channel radiates electromagnetic waves to the ground, and the receiver collects the signals reflected from underground objects or geological interfaces, generating a one-dimensional waveform, i.e., an A-scan. As the robot moves along the y-direction, each GPR antenna generates a set of A-scans at fixed intervals, forming a set of two-dimensional data, i.e., a B-scan. It can be seen that the B-scan represents a vertical slice of the ground. Combined with... Figure 2 When GPR moves along multiple parallel line segments, it can collect multiple parallel B-scans, thus forming a three-dimensional dataset, namely C-scan.

[0066] Figure 3 This is a flowchart illustrating an underground target identification method according to some embodiments of this application. The method is used to identify... Figure 1 and 2 The data collected by the robot is processed, which may include the following steps S100-S700. Figure 4 The diagram illustrates the architecture of an underground target recognition system (MV-GPRNet) according to some embodiments of this application. MV-GPRNet includes (a) a B-scan-based 3D convolutional neural network module, (b) a top-scan-based 2D CNN module, and (c) a multi-view fusion module. MV-GPRNet can at least perform... Figure 3 The method for identifying underground targets is shown. Based on steps S100-S700, it will combine... Figure 4 Some optional embodiments are described below. It should be noted that these embodiments should be understood as examples and not as essential technical features for implementing this solution.

[0067] Step S100: Receive the ground-penetrating radar C-scan of the underground area to be detected.

[0068] Step S200: Obtain the C-scan 3D feature map based on the C-scan.

[0069] In this embodiment, the detection robot collects information along a predetermined trajectory, thereby obtaining a series of parallel GPR B-scan images. All B-scan images are combined to form a three-dimensional dataset C-scan. After receiving the C-scan, the MV-GPRNet system converts it into a C-scan 3D feature map (C-scan 3DFeaturemap) through a B-scan-based 3D CNN module.

[0070] Step S300: Receive multiple top-scan profile views.

[0071] Step S400: Based on the multiple top-scan profiles, obtain multiple top-scan feature maps and multiple 2D candidate boxes corresponding to the multiple top-scan feature maps respectively.

[0072] Step S500: Map the multiple 2D candidate boxes to the corresponding Top-scan feature maps, and set the values ​​of the parts of the Top-scan feature maps that do not correspond to the 2D candidate boxes to 0, thereby obtaining multiple ROI focusing feature maps.

[0073] Throughout the entire inspection process of the inspection robot, the data recorded by all GPR antennas at a certain underground depth forms another two-dimensional view, called a top-scan, which is an image of a horizontal cross-section at a certain underground depth. It can be understood that each top-scan corresponds to the 2D data of the C-scan at a specific depth plane, and all top-scans correspond to the 2D data of the C-scan at different depth planes.

[0074] The inventors of this invention discovered that some underground disease targets exhibit strong identifying features in the Top-scan but weak features in the B-scan. Therefore, the inventors processed the multi-view data—C-scan and Top-scan—separately so that their feature maps could be fused later.

[0075] Since the continuity between adjacent top-scans is not as obvious as that between adjacent B-scans, the MV-GPRNet system uses a 2D CNN module to extract facet-by-face feature maps from the top-scans.

[0076] After receiving multiple top-scan profiles, the MV-GPRNet system converts them into multiple top-scan feature maps and multiple 2D candidate boxes corresponding to each top-scan feature map using a top-scan-based 2D CNN module. The multiple 2D candidate boxes are then mapped to their corresponding top-scan feature maps, and the values ​​of the parts of the top-scan feature maps that do not correspond to the 2D candidate boxes are set to 0, resulting in multiple ROI-focused feature maps.

[0077] Step S600: The C-scan 3D feature map and the multiple ROI focused feature maps are fused to obtain a fused feature map.

[0078] Step S700: Based on the fused feature map, determine the category and location of the underground target in the underground area to be detected.

[0079] In this embodiment, the MV-GPRNet system receives the output of the 3D CNN module (C-scan 3D feature map) and the output of the 2D CNN module (multiple ROI focused feature maps) through the multi-view fusion module, and fuses the two to obtain a fused feature map. Based on the depth of the corresponding top-scan, the ROI focused feature maps can be easily merged into the C-scan 3D feature map without needing to calculate the spatial location corresponding to the C-scan based on the ROI location.

[0080] The multi-view fusion module processes the fused feature map to determine the category and location of underground targets in the underground area to be detected. In this embodiment, the types of underground targets include, but are not limited to: voids, subsidence, cracks, and pipes.

[0081] Through steps S100-S700 above, the underground target identification method provided in this embodiment extracts feature maps from the three-dimensional dataset C-scan and the two-dimensional dataset Top-scan respectively: C-scan 3D feature map and ROI focused feature map, and then fuses the two types of feature maps. The fused feature map can focus on underground disease targets that cannot be distinguished in C-scan but have obvious features in Top-scan. In addition, underground diseases usually do not have a fixed shape when reflected in the Top-scan planar view. Therefore, fusing the ROI focused feature map with the C-scan 3D feature map can mask the interference caused by noise extracted from C-scan in the fused feature map and enhance the disease features, thereby accurately identifying the type of underground target in the fused feature map. The fused feature map is a three-dimensional dataset, so the specific location of the underground target can be directly obtained in the output result.

[0082] In some embodiments, step S200 includes:

[0083] Step S210: Divide the C-scan into multiple equally spaced voxels.

[0084] In step S220, the multiple voxels are sequentially processed through 3D convolutional layers and a 3D dense network to obtain feature map x. l .

[0085] Step S230, based on the feature map x l The channel attention map Atte is obtained. rc .

[0086] Step S240, the channel attention map Atte rc and the feature map x l Element-wise multiplication is performed to obtain the C-scan 3D feature map.

[0087] In this embodiment, the 3D CNN module has 3D convolutional layers (3D Conv), a 3D dense network, and a radar-channel attention module.

[0088] In this embodiment, the C-scan is first divided into equally spaced voxels υ, each voxel having a spatial resolution of υx×υy×υz. υx, υy, and υz represent the length, height, and width of each voxel, respectively. Typically, υx is set to the number of channels in the GPR array, ensuring that the data in a voxel comes from only one parallel scan. That is, a voxel contains all the B-scan image data obtained in one parallel scan, thus avoiding data alignment errors after multiple scans. Furthermore, the 3D convolutional layers can consist of 1x, 2x, 4x, and 8x downsampling convolutional layers. A series of 3D convolutional layers sequentially transforms the voxels into feature volumes with 1x, 2x, 4x, and 8x downsampling specifications. These feature volumes can be considered as a series of voxel-by-voxel feature vectors.

[0089] To enhance the utilization of voxel-by-voxel features, the 3D dense network in the 3D CNN module consists of three dense layers, and a non-linear transformation H is performed in each 3D feature extraction process. l The l-th dense layer receives voxel-by-voxel feature maps, x0, ..., x1, generated by all preceding dense layers 0, ..., l-1. l-1 , as input, and obtain feature map x l :

[0090] x l =H l ([x0, x1, ..., x l-1 ])

[0091] Where, [x0, x1, ..., x l-1 ] represents the voxel-wise feature map, x0, ..., x l-1 , concatenation.

[0092] The radar channel attention module uses the feature map x obtained from the last dense layer. l As input. In the radar channel attention module, x l The feature tensor X is obtained through average pooling operation. avg x l X is obtained through max pooling operation max Then X avg and X max Enter a shared network. The shared network consists of a fully connected layer F with a hidden layer. c Composition. To reduce parameter overhead, the hidden activation size is set to C / r×1×1, where r is the reduction ratio. X avg and X maxThe output after passing through the shared network is then processed by element-wise addition and an activation function with weights ranging from [0,1] to obtain the radar channel attention map Atte. rc Its dimensions are C×1×1:

[0093] Atte rc =σ(F c (AvgPool(x l ))+F c (MaxPool(x l )))

[0094] Where σ is the sigmoid activation function, F c It is a fully connected layer, AvgPool is average pooling, and MaxPool is max pooling.

[0095] Finally, the model will use the channel attention map Atte rc and feature map x l Element-wise multiplication yields the C-scan3D feature map A. rc (X l ):

[0096]

[0097] in, This indicates element-wise multiplication.

[0098] This embodiment enhances the effective domain feature analysis of a specific radar channel by introducing a radar channel attention module.

[0099] In this embodiment, multiple voxels first pass through the dense layer and then through the radar channel attention module, meaning the dense layer is in front of the radar channel attention module. It is understood that in some embodiments, the dense layer may be after the radar channel attention module, meaning multiple voxels pass through the radar channel attention module first and then through the dense layer; in other embodiments, the dense layer and the radar channel attention module are connected in parallel, meaning multiple voxels pass through the dense layer and the radar channel attention module respectively.

[0100] In some embodiments, step S400 includes:

[0101] In step S410, the multiple top-scan profiles are processed sequentially through 2D convolutional layers and 2D dense networks to obtain multiple top-scan feature maps.

[0102] Step S420: The multiple Top-scan feature maps are processed by object classification and 2D bounding box regression to obtain 2D candidate boxes.

[0103] In this embodiment, the 2D CNN module has 2D convolutional layers (2D Conv) and a 2D dense network, and the 2D dense network has several 2D dense layers. Preferably, the 2D dense network has 3 2D dense layers.

[0104] The 2D CNN module generates 2D candidate boxes from a series of 2D prior boxes, each 2D bounding box being parameterized as [x, y, w, h]. T Where, [x,y] T Let [w, h] be the center coordinates of the 2D bounding box. T The width and height of the 2D border.

[0105] The loss function L of 2D bounding box regression tr satisfy:

[0106]

[0107] Among them, t i =[x,y,w,h] T , which are the labeled 2D borders. This is for predicting the bounding box.

[0108] In some embodiments, step S600 includes:

[0109] Step S610: Through interpolation and transformation processing, the multiple ROI focused feature maps and the C-scan 3D feature map are made to have the same size in the planar dimension of the ROI focused feature map.

[0110] Step S620: The C-scan 3D feature map and the multiple ROI focused feature maps are added element by element to obtain a fused feature map.

[0111] In this embodiment, when fusing the C-scan 3D feature map and multiple ROI focused feature maps, it is necessary to ensure that the size of the two-dimensional ROI focused feature map is the same as the size of the slice feature map in the horizontal direction of the three-dimensional C-scan 3D feature map. That is, the ROI focused feature map and the C-scan 3D feature map have the same size in the planar dimension of the ROI focused feature map. This process is achieved through interpolation transformation.

[0112] In some embodiments, step S700 includes:

[0113] Step S710: The fused feature map is processed using a deconvolutional sub-network to separate the foreground pixels as initial 3D candidate boxes.

[0114] Step S720: Map the initial 3D candidate box to the corresponding position in the fused feature map and perform ROI-pooling operation.

[0115] In step S730, the output of the ROI-pooling operation is processed by object classification and bounding box regression to obtain the category and bounding box of the underground target, and the location of the underground target is determined based on the bounding box of the underground target.

[0116] Since underground defects are naturally separated and do not overlap, this embodiment uses a bottom-up approach to generate 3D candidate boxes. Pixels within the 3D boxes are considered foreground pixels. The multi-view fusion module uses a 3D deconvolution sub-network to separate the foreground pixels in each voxel, and the 3D bounding boxes generated from these foreground pixels are the initial 3D candidate boxes. The 3D deconvolution sub-network upsamples the fusion features to achieve dense prediction. The ground-truth separation mask is naturally provided by the 3D ground-truth boxes. Since the number of foreground elements is much smaller than the number of background elements, this embodiment uses a loss function L. f To address the class imbalance problem:

[0117] L f (p t )=-α t (1-p t ) γ logp t ,

[0118]

[0119] Where, α t γ represents the weighting factor, p is the focus parameter, and p is the probability that a pixel belongs to the foreground pixel.

[0120] Next, this embodiment obtains feature maps of fixed shapes through ROI-pooling operations. These fixed-shape feature maps are used to regress 3D bounding boxes. The loss function L in object classification and bounding box regression processing... r satisfy:

[0121] L r =α∑L c +β∑L h ;

[0122] in,

[0123]

[0124] Where, n c y represents the total number of categories.i and Let b represent the label value and predicted value of a point belonging to the i-th category, respectively. i Let be a vector representing the labeled 3D bounding box. Let be a vector representing the predicted 3D bounding box, where δ is a parameter, and α and β are balancing weight factors. Further, the vector of the 3D bounding box b is [x, y, z, l, w, h]. T Where [x,y,z] T The coordinates represent the top-left corner of the bounding box; l, w, and h represent the length, width, and height, respectively. The resulting 3D bounding box position represents the location of the underground target.

[0125] The underground target identification method of the present invention has the following experimental results.

[0126] The specific experimental procedure is as follows:

[0127] All modules of MV-GPRNet were trained together. The entire network was trained for 60 epochs with a batch size of 1 and a learning rate of 0.00001. Gradient clipping was used to avoid gradient explosion, and a momentum coefficient of 0.9 was used to accelerate gradient descent. Each GPR B-scan was reformatted to a size of 448 pixels × 448 pixels, corresponding to a 7m × 7m concrete segment. When training the 3D GPR data separation, α... t =0.25, γ=2; initially, α=1, β=1, δ=1. To quantify the performance of different methods, the experiment used three metrics: precision, recall, and F1-measure. These three metrics are determined based on the intersection-over-union (IoU) threshold, which represents the overlap between the predicted bounding box and the original label box. If the IoU value is greater than the preset threshold T... IoU This predicted bounding box is considered the true value (TP). In the experiment, T... IoU The value is 0.5.

[0128] Airport Data Acquisition: Experiments were conducted at a real airport, with the dataset collected by a detection robot equipped with a GPR (Geometry Probe). The robot was fitted with an equidistant 14-channel Raptor. TM GPR, antenna frequency 900MHz. GPR set to range-triggered mode. Ground penetrating radar coverage width 0.7m. (Example) Figure 2As shown, to conduct defect inspections, the robot navigates within a predefined survey area on the airport runway to collect images and GPR data. The robot first moves from a starting point and then follows a linear path for each scan. Once the robot completes the current scan, it moves to the next scan until the entire area to be inspected is completely covered. During scanning, the robot uses a 4G / 5G connection to transmit GPR data to a nearby data analysis center, where the collected data is then automatically analyzed offline.

[0129] In this experiment, each disease was randomly assigned to the training and testing data in a 7:3 ratio. Real-world ground surface irradiation (GPR) data were collected from three international airports in China. To obtain B-scan and Top-scan data, two datasets with different views extracted from the GPR data were created, named AUD-B and AUD-T. The AUD-B dataset contains 5300 B-scan images, and the AUD-T dataset contains 4526 Top-scan images. The total inspection area of ​​the three airports was 21083 square meters, with a ground depth of 1.53 meters. The horizontal sampling rate of the GPR was 390 A-scans per meter, and the vertical sampling rate was 1024 A-scans per meter. Two human experts independently labeled four typical underground diseases / objects, including voids, subsidence, cracks, and pipes.

[0130] The experimental results are shown in Figure 5-8 .from Figure 5 As can be seen, the F1-measure results of the MV-GPRNet of this invention have a significant advantage over the other four existing methods. Figure 6 The "×" indicates that the corresponding module is missing, and the "√" indicates that the corresponding module is present. It can be seen that the detection effect will be significantly reduced when either or both of the dense layer module and the radar channel attention module are missing. Figure 7 In the diagram, Dense+Attention indicates that the dense layer is placed before the radar channel attention module, while Attention+Dense indicates that the radar channel attention module is placed before the dense layer. It can be seen that the detection performance of a sequential arrangement of the dense layer and the radar channel attention module is better than that of a side-by-side arrangement. From... Figure 8 As can be seen, for the four types of underground targets, the F1-measure results after adding depth dimension features (i.e., adding Top-scan) are all better than the F1-measure results of extracting features only from B-scan.

[0131] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0132] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0133] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0134] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for identifying underground targets, the method comprising: Receives C-scan data from ground-penetrating radar of the underground area to be detected; Based on the C-scan, a C-scan 3D feature map is obtained; Receive multiple top-scan profiles, each of which corresponds to 2D data of the C-scan at different depth planes; Based on the multiple profile top-scans, multiple top-scan feature maps and multiple 2D candidate boxes corresponding to the multiple top-scan feature maps are obtained respectively; The multiple 2D candidate boxes are mapped to the corresponding Top-scan feature maps, and the values ​​of the parts of the Top-scan feature maps that do not correspond to the 2D candidate boxes are set to 0, thereby obtaining multiple ROI focusing feature maps; The C-scan 3D feature map and the multiple ROI focused feature maps are fused to obtain a fused feature map; Based on the fused feature map, the category and location of underground targets in the underground area to be detected are determined; The step of obtaining the C-scan 3D feature map based on the C-scan includes: The C-scan is divided into multiple equally spaced voxels; The multiple voxels are sequentially processed through 3D convolutional layers and 3D dense networks to obtain feature maps. x l ; Based on the feature map x l Obtain the channel attention map Atte rc ; Atte rc =σ( F c ( AvgPool ( x l ))+ F c ( MaxPool ( x l ))); Where σ is the sigmoid activation function, F c It is a fully connected layer. AvgPool For average pooling, MaxPool For max pooling; The channel attention map Atte rc and the feature map x l Element-wise multiplication is performed to obtain the C-scan 3D feature map.

2. The method as described in claim 1, characterized in that, The C-scan includes a series of B-scan images obtained by multiple parallel scans of the underground area to be detected by a multi-channel ground-penetrating radar. The voxel contains the B-scan image obtained in a single scan of the multi-channel ground-penetrating radar.

3. The method as described in claim 1, characterized in that, The 3D convolutional layer consists of 1x, 2x, 4x, and 8x downsampling convolutional layers, and the 3D dense network consists of 3 dense layers.

4. The method as described in claim 1, characterized in that, The step of obtaining multiple Top-scan feature maps and multiple 2D candidate boxes corresponding to the multiple Top-scan feature maps based on the multiple profile maps includes: The multiple top-scan profile maps are processed sequentially through 2D convolutional layers and 2D dense networks to obtain multiple top-scan feature maps; The multiple Top-scan feature maps are processed by object classification and 2D bounding box regression to obtain 2D candidate boxes; The loss function of the 2D bounding box regression is... satisfy: in, , which are the labeled 2D borders. The center coordinates of the 2D border The width and height of the 2D border. This is for predicting the bounding box.

5. The method as described in claim 1, characterized in that, The step of fusing the C-scan 3D feature map and the multiple ROI focused feature maps to obtain a fused feature map includes: Through interpolation and transformation, the multiple ROI focused feature maps and the C-scan 3D feature map are made to have the same size in the planar dimension of the ROI focused feature map; The C-scan 3D feature map and the multiple ROI focused feature maps are added element-wise to obtain a fused feature map.

6. The method as described in claim 1, characterized in that, The step of determining the category and location of underground targets in the underground area to be detected based on the fused feature map includes: The fused feature map is processed using a deconvolutional sub-network to separate the foreground pixels as initial 3D candidate boxes. The initial 3D candidate boxes are mapped to the corresponding positions in the fused feature map, and ROI-pooling is performed. The output of the ROI-pooling operation is processed by object classification and bounding box regression to obtain the category and bounding box of the underground target. The location of the underground target is determined based on the bounding box of the underground target.

7. The method as described in claim 6, characterized in that, The foreground pixel separation process has a loss function satisfy: ， in, γ represents the weighting factor, p is the focus parameter, and p is the probability that a pixel belongs to the foreground pixel.

8. The method as described in claim 6, characterized in that, The object classification and bounding box regression processing have a loss function satisfy: ; in, The total number of categories, and Let these represent the label value and predicted value of a point belonging to the i-th category, respectively. Let be a vector representing the labeled 3D bounding box. Let be a vector representing the predicted 3D bounding box. For parameters, and To balance the weighting factors.

9. An underground target recognition system, the system comprising a 3D CNN module, a 2D CNN module, and a multi-view fusion network module; The 3D CNN module is used to receive the ground-penetrating radar C-scan of the underground area to be detected and output the C-scan 3D feature map; The 2D CNN module is used to receive multiple top-scan profiles, each corresponding to 2D data of the C-scan at different depth planes; based on the multiple top-scan profiles, multiple top-scan feature maps and multiple 2D candidate boxes corresponding to each top-scan feature map are obtained; the multiple 2D candidate boxes are mapped to the corresponding top-scan feature maps, and the values ​​of the parts of the top-scan feature maps that do not correspond to the 2D candidate boxes are set to 0, thus obtaining multiple ROI focusing feature maps; The multi-view fusion module is used to fuse the C-scan 3D feature map and the multiple ROI focused feature maps to obtain a fused feature map; based on the fused feature map, the category and location of underground targets in the underground area to be detected are determined; in, The 3D CNN module is further used to divide the C-scan into multiple equally spaced voxels; the multiple voxels are sequentially processed by 3D convolutional layers and a 3D dense network to obtain feature maps. x l Based on the feature map x l Obtain the channel attention map Atte rc ; Atte rc =σ( F c ( AvgPool ( x l ))+ F c ( MaxPool ( x l )));where σ is the sigmoid activation function, F c It is a fully connected layer. AvgPool For average pooling, MaxPool For max pooling; the channel attention map Atte rc and the feature map x l Element-wise multiplication is performed to obtain the C-scan 3D feature map.

Images