Ground Penetrating Radar Data Key Point Detection and Target Localization Method

Abstract

A key point detection and target localization method based on ground-penetrating radar (GPR) data is proposed. First, a global attention mechanism is introduced into the YOLOv8 framework to improve the performance of deep neural networks by reducing information loss and amplifying global interactive representations. Second, a multi-head self-attention mechanism is introduced, combining spatial feature enhancement subnets and deep recurrent networks, to achieve high-resolution mapping of underground structures. Finally, generalized intersection-union (GUU) loss and full intersection-union (GUU) loss are introduced into the loss function to optimize detection accuracy under both non-overlapping and overlapping conditions by considering the distance to the outer region of the bounding box, the center point of the target detection box, and the aspect ratio. This method first performs GPR target detection based on the YOLOv8-GSI network to obtain the regions where candidate targets are located. Then, some training weights from the first stage are shared and passed to the second stage, which trains the second-stage YOLOv8-GSI network based on these weights to obtain accurate key point detection from the candidate target features, thereby achieving automated localization of underground targets.

Description

Technical Field

[0001] This invention relates to the field of key point detection and target localization technology in ground penetrating radar images, specifically to a method for key point detection and target localization in ground penetrating radar data. Background Technology

[0002] Ground-penetrating radar (GPR) is a non-destructive testing and assessment technique widely used in the inspection of civil infrastructure. Applications include mapping underground assets, locating and inspecting rebar for corrosion, assessing concrete bridge deck degradation, detecting sinkholes, and controlling construction quality. Impact GPR (hereinafter referred to as GPR) emits short electromagnetic (EM) pulses into the ground, propagating in a cone shape, and receives signals reflected from objects with different EM properties than the underground medium. These reflected signals form a GPR scan, a primary source for data visualization, analysis, and interpretation.

[0003] Traditional GPR (Geometric Pixel Array) technology has achieved significant results in the field of underground target detection. However, the detection and localization of underground targets remains a challenging task. The introduction of deep learning technology has brought new possibilities to GPR image processing. In particular, the application of two-stage and single-stage detection methods has greatly improved the accuracy and efficiency of detection. The two-stage detection method achieves high-precision target detection by generating candidate regions and performing accurate classification and bounding box regression on these regions. Although it shows high accuracy in some applications, this method is slow, has a long training time, and insufficient information fusion between stages, which easily leads to false alarms and affects the overall performance of the system. In addition, although the single-stage method has advantages in speed and real-time performance, it still has shortcomings in directly and accurately locating underground targets. Existing single-stage methods can quickly locate the approximate location of the target, but lack precise control, and its performance is not ideal, especially in the detection of underground structures in complex or noisy backgrounds.

[0004] Therefore, although current deep learning methods have made some progress in underground target detection, how to further improve detection accuracy and real-time response capabilities in complex environments remains a pressing technical problem to be solved in this field. A new method or improvement to existing technologies is needed to achieve more accurate and efficient identification and localization of complex underground structures.

[0005] Technical comparison with patent CN118015457A "Ground Penetrating Radar Underground Cavity Image Detection Method Based on Improved YOLOv8 Algorithm".

[0006] Patent CN118015457A is mainly used for underground cavity image detection, while this patent is for general underground target detection. The dataset is mainly generated from bridge metal and simulation software GprMax.

[0007] The main objective of patent CN118015457A is target detection of underground targets, while this patent adds key point detection on the basis of target detection, which can better locate the specific location and shape of underground targets.

[0008] Based on the YOLOv8 network model, patent CN118015457A rewrites ordinary convolution into local convolution. This patent introduces multiple attention mechanisms and adds generalized intersection loss on the union, which can better extract deep semantic features in GPR images. Summary of the Invention

[0009] To address the issues of accuracy and efficiency in target detection and localization during Ground Penetrating Radar (GPR) data processing, this invention proposes a keypoint detection and target localization method for GPR data. Traditional GPR image processing methods suffer from significant data interference, low target recognition accuracy, and inaccurate keypoint localization, which is particularly pronounced in complex underground environments. Therefore, the goal of this invention is to improve the detection accuracy of underground structural targets and the localization accuracy of keypoints, while reducing background noise interference and improving overall processing speed and efficiency. This invention constructs a two-stage detection model, based on a reused single-stage detection model—the YOLOv8-GSI network—with the same network framework. The first-stage model trains the YOLOv8-GSI network to achieve GPR target detection, and shares some of the trained weights with the second-stage model. The second-stage model uses a keypoint detection dataset to train the YOLOv8-GSI network to achieve keypoint detection in cropped images. Finally, through mapping relationships, the positions of keypoints in the cropped image relative to the original image are obtained, realizing keypoint detection and target localization in GPR images.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] The method for key point detection and target localization using ground-penetrating radar data includes the following steps:

[0012] S1: Experimental data acquisition;

[0013] The dataset includes measured data and simulation data. The measured data was obtained from ground-penetrating radar of civilian bridge facilities, and the simulation data was generated by GprMax 3D modeling to simulate the conductivity and space occupancy of various media.

[0014] S2: Data preprocessing;

[0015] Direct wave removal and 2D Kirchhoff Migration migration were performed on the simulation dataset to reduce interference and improve the spatial positioning accuracy of underground structures;

[0016] S3: YOLOv8-GSI network construction;

[0017] The YOLOv8 framework integrates global attention and multi-head self-attention mechanisms, and introduces GIoU and CIoU losses to improve object detection accuracy and keypoint detection performance. The global attention mechanism enhances the performance of deep neural networks by reducing information loss and amplifying global interactive representations. The multi-head self-attention mechanism processes input features in parallel through multiple attention heads and combines spatial features to enhance subnets, achieving high-resolution mapping of underground structures. GIoU and CIoU losses optimize detection accuracy under non-overlapping and overlapping conditions by increasing consideration of the outer region of the bounding box, the distance to the center point of the object detection box, and the aspect ratio.

[0018] S4: Target recognition;

[0019] The YOLOv8-GSI network was trained with bounding boxes to build the first model. The accurate target recognition in this stage provides the necessary foundation for subsequent key point detection.

[0020] S5: Key point annotation;

[0021] Using the first-stage model, the original dataset is cropped to obtain a sub-image containing only a single target. By labeling the vertex and two sets of horizontally symmetrical vertically evenly distributed key points, a total of 5 key points are obtained to obtain a key point detection dataset. This enables accurate fitting of the single target contour and determination of the target's spatial location.

[0022] S6: Keypoint detection;

[0023] Keypoint detection is performed on the small images obtained from the first stage of cropping. First, some weights of the first model are transferred to the second model to shorten the training time of the second model. Second, the second model is trained based on the keypoint annotation dataset. Finally, the trained second model is used to perform model inference on the cropped images output from the first stage.

[0024] S7: Keypoint Mapping;

[0025] The key point locations identified in the cropped image are mapped back to the original GPR image. In this process, the positions of the key points need to be accurately transferred from the cropped image to the original image to ensure that the relative position of each point remains unchanged. For each target, the highest point in the middle of its key points is selected for highlighting. The vertex coordinates determine the spatial location of the underground target, providing an important visual reference for subsequent data analysis and applications.

[0026] Furthermore, step S3 includes the following steps:

[0027] S31: Add a global attention mechanism;

[0028] The introduction of a global attention mechanism aims to improve the performance of deep neural networks by reducing information loss and enhancing global interactive representations. This mechanism employs 3D permutation and multilayer perceptrons for channel attention, combined with a convolutional spatial attention submodule to enhance the interactive effect of global features. 3D permutation can preserve information in different dimensions, and the multilayer perceptron further processes this information, amplifying the interdependence between channels. Meanwhile, the convolutional spatial attention submodule aggregates spatial information through convolution operations, enhancing the interactive representation of global features.

[0029] S32: Add a self-attention mechanism;

[0030] The multi-head self-attention mechanism achieves high-resolution mapping and detail capture by processing input features in parallel. It utilizes multiple attention heads to process input features in parallel, with each attention head independently focusing on different parts of the input features, thereby capturing more detailed information. Combined with spatial feature enhancement subnets and deep recurrent networks, the multi-head self-attention mechanism further improves the resolution of feature maps in the spatial dimension, ensuring high-resolution mapping of underground structures.

[0031] S33: Add generalized intersection loss;

[0032] By introducing the generalized intersection-union loss into the loss function, the minimum closed rectangle of the two axis-aligned rectangular bounding boxes is calculated and the difference between this closed rectangle and the joint region of the two bounding boxes is measured, thereby optimizing the bounding box regression.

[0033] S34: Add full intersection-union loss;

[0034] By introducing a complete intersection-union loss into the loss function, and by considering three geometric factors—overlap area, normalized center point distance, and aspect ratio—a significant improvement in the accuracy of bounding box regression is achieved.

[0035] Furthermore, step S4 includes the following steps:

[0036] S41: Model training;

[0037] Targets were labeled on GPR images, and the YOLOv8-GSI network was used to train the first model, which focuses on improving target recognition performance.

[0038] First, the targets in the GPR images need to be labeled to accurately mark the location of each target and build a target detection dataset. Then, the YOLOv8-GSI model is trained using the target detection dataset. During the training process, the key point detection part of the model is ignored, and the focus is on improving its performance in target recognition. Through supervised learning methods, the model is trained to accurately identify and locate various underground structure targets in GPR images.

[0039] S42: Model Inference;

[0040] Inference is performed on new GPR images using a first-stage YOLOv8-GSI model that has been trained.

[0041] First, the GPR image to be processed is input into the trained model. The model identifies and labels the target region in the image, quickly locating the position of the underground structure target. Then, the original image is cropped according to the output target coordinates to obtain a cropped target image of a single target.

[0042] S43: Image cropping;

[0043] Based on the location of the detection bounding box in the model, the GPR image is precisely cropped to prepare the input image for the key point recognition stage, ensuring the accuracy of the cropping process.

[0044] Furthermore, step S6 includes the following steps:

[0045] S61: Model Training

[0046] The goal of this stage is to train the model to identify and locate key points in the cropped image. Using the cropped image obtained from the first stage, the key points are accurately labeled, including the vertices of each target and other important feature points. These labeled images are then used as input to train the second YOLOv8-GSI model. In this process, the model learns how to identify the exact location of key points from the cropped image, thus achieving the localization of key point locations for a single target.

[0047] S62: Model Inference;

[0048] The trained second model is used to process the image cropped in the first stage. In this stage, the cropped image is input into the model, which is responsible for identifying and marking the location of key points in the image. This step is a critical link in the key point detection process because it directly affects the accuracy of key point mapping. The model's inference results provide detailed information about the location of each target key point, laying the foundation for subsequent mapping steps.

[0049] The benefits and indicators achieved by this invention.

[0050] 1. Improved target detection accuracy: By combining global attention mechanism and self-attention mechanism, and using GIoU loss, the accuracy of target detection and the accuracy of key point localization are significantly improved, with a target detection accuracy of 93.1%.

[0051] 2. Achieve high-precision target positioning: Through key point detection, accurate positioning of underground targets is achieved, with key point recognition mAP50 reaching 82.9%, ensuring accurate positioning.

[0052] 2. Reduce noise interference: Optimized data preprocessing steps effectively remove background noise and direct wave interference, clearly revealing the underground structure.

[0053] 3. Improved processing speed: The high efficiency of the new model reduces processing time, making rapid identification of underground structures and key point detection possible.

[0054] 4. Wide applicability: This method is not only applicable to the detection of civil facilities such as bridges, but can also be extended to other fields that require high-precision underground detection. Attached Figure Description

[0055] Figure 1 This is the overall process for key point detection and target localization methods using ground-penetrating radar data;

[0056] Figure 2 This is a diagram of the YOLOv8-GSI network architecture;

[0057] Figure 3 This is a diagram of the global attention mechanism network structure;

[0058] Figure 4 This is a diagram of the self-attention mechanism network structure;

[0059] Figure 5 This is an image showing the target detection results;

[0060] Figure 6 This is a diagram illustrating the key points.

[0061] Figure 7 This is a diagram showing the model recognition results. Detailed Implementation

[0062] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0063] This method develops a keypoint detection and target localization method for GPR data based on a dual single-stage depth model. The method utilizes a dual reuse of the single-stage keypoint detection network—the YOLOv8-GSI network—to construct a two-stage detection model. In the first stage, the YOLOv8-GSI network is trained to identify targets in radar images, and some of the trained weights are transferred to the second-stage model. The second stage uses the keypoint-annotated dataset to further train the YOLOv8-GSI network, focusing on keypoint detection in cropped images. By establishing a mapping relationship, this model can accurately determine the specific location of keypoints in the cropped image relative to the original image, thereby achieving accurate keypoint detection and target localization in ground-penetrating radar images. The flowchart is shown below. Figure 1 As shown.

[0064] S1: Experimental Data Acquisition

[0065] This technique involves two types of datasets: real-world datasets and simulation datasets.

[0066] The measured dataset consists of ground-penetrating radar images collected from civilian bridge infrastructure. These images reflect actual usage conditions and are directly used for model training and validation, ensuring the technology is adapted to real-world application environments.

[0067] Simulation Dataset: A 3D ground-penetrating radar model was created using GprMax software to simulate the actual environment. Simulation parameters included the electrical conductivity and spatial distribution of different media. For example, in the simulation, a hexahedral region with coordinates (0,0,0) to (2.5,0.45,0.0025) was defined to represent concrete, while a hexahedron with coordinates (0,0.45,0) to (2.5,0.5,0.0025) represented air.

[0068] S2: Data Preprocessing

[0069] In the data preprocessing stage, direct wave removal is first performed to eliminate interference from the underground structure images. Subsequently, the 2D Kirchhoff Migration algorithm is applied to process the simulation dataset to accurately extract the spatial location and detailed information of the underground structures. These steps are crucial for the accuracy and efficiency of subsequent analysis.

[0070] S3: YOLOv8-GSI Network Construction

[0071] This invention relates to a single-stage deep learning model, namely the YOLOv8-GSI network. Based on the YOLOv8 architecture, this invention first introduces a global attention mechanism, which effectively improves the overall performance of the deep neural network by reducing information loss and expanding the representation of global interactions. Secondly, by combining a spatial feature-enhanced subnetwork with a deep recurrent network, a multi-head self-attention mechanism is introduced to achieve high-resolution mapping of underground structures. Finally, this invention introduces Generalized Intersection over Union (GIoU) ​​and Complete Intersection over Union (CIoU) loss functions. This improvement optimizes detection accuracy in both non-overlapping and overlapping cases by considering the distance between the outer region of the bounding box, the center point of the target detection box, and their aspect ratio, thereby improving the model's localization accuracy and stability in various detection scenarios. The YOLOv8-GSI network structure diagram is shown below. Figure 2 As shown.

[0072] S31: Global Attention Mechanism

[0073] The CSPlayer_2Conv module in the P5 layer of CSPDarkNet-53 is replaced with a Global Attention Mechanism (GAM). The introduction of GAM aims to enhance the interaction between channel and spatial information, reduce information loss, and thus improve the global representation capability of features. CSPDarkNet-53 extracts features from each channel and scale in GPR images. In the P5 layer, the replaced GAM module can effectively dynamically reweight the input features to highlight more important features and suppress secondary information, thereby achieving more accurate image recognition results. Specifically, by using a multilayer perceptron (MLP) for feature reorganization in the channel attention module and applying convolutional layers to directly weight spatial information in the spatial attention module, the model's ability to resolve complex geological structures is effectively improved. The global attention mechanism network structure is as follows: Figure 3 As shown.

[0074] S32: Self-attention mechanism:

[0075] The second ConvModule in the detection head is replaced with a self-attention mechanism. By processing global information across the entire input feature map, the self-attention mechanism can more comprehensively capture the widely distributed semantic information in the image, which is crucial for tasks such as object detection and keypoint detection. Unlike traditional fixed convolutional kernels, self-attention layers can dynamically adjust their response based on the input data, enhancing or suppressing specific features, thus providing greater flexibility and robustness. The self-attention mechanism network structure is as follows: Figure 4 As shown.

[0076] Multi-head self-attention mechanisms process input features in parallel using multiple independent attention heads, allowing each head to focus on capturing detailed information in a specific region. This mechanism significantly improves the network's ability to capture details when processing high-resolution images. Combined with a spatial feature enhancement subnetwork, this technique finely improves the spatial resolution of feature maps through convolutional layers while preserving important local features. Furthermore, deep recurrent networks deepen the understanding and representation of complex patterns by recursively passing information between multiple layers. The combined application of these techniques ensures high-resolution mapping of underground structures, improving performance and adaptability in complex scenes.

[0077] S33, Loss Function Design

[0078] Generalized intersection loss (GIou) and complete intersection-union ratio loss (CIoU) are used to optimize the detection accuracy of bounding boxes. In GPR images, underground targets are unevenly distributed, and there are various problems such as complex overlap, occlusion, and different scales.

[0079] GIoU loss measures the intersection ratio between the predicted and ground truth bounding boxes and addresses the gradient vanishing problem of traditional IoU when the bounding boxes do not overlap by introducing a minimum closed rectangle. This is particularly important for multi-target detection in ground-penetrating radar images, especially when targets partially overlap. GIoU loss considers the complete geometric relationship between bounding boxes, making it more effective in handling targets of different sizes and partial overlap, thus significantly improving the localization accuracy of bounding boxes for complex underground structures. The GIoU calculation formula is shown in equations (1, 2).

[0080] Where A and B are the predicted bounding box and the ground truth bounding box, respectively, and C is the smallest closed rectangle containing A and B. This represents the region in C excluding the union of A and B. This approach encourages predicted boxes to converge towards the ground truth boxes by reducing the area within C not covered by either bounding box, effectively updating gradients even in the absence of overlap.

[0081] The Complete Intersection Ratio (CIoU) loss, based on IoU, adds considerations of center point distance and aspect ratio consistency, making it more suitable for handling situations with large differences in target shape and size. CIoU further optimizes the accurate matching of bounding boxes by penalizing the distance deviation between the center points of the predicted box and the ground truth box, as well as the inconsistency in aspect ratio. This is especially important when the targets do not overlap at all or have significant shape distortion, such as when detecting multi-layered or intersecting underground pipelines, CIoU can effectively improve the accuracy and stability of detection. The CIoU calculation formula is shown in formula (3):

[0082] It is the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box. It is the length of the diagonal of the smallest enclosed rectangle. The value is calculated by the difference in aspect ratio between the predicted and ground truth bounding boxes, while α is a weighting factor that is dynamically adjusted based on the overlap area. This composite penalty term not only promotes the spatial overlap between the predicted and ground truth bounding boxes but also fosters their consistency in shape and position, thereby providing more accurate and stable target detection performance in complex ground-penetrating radar imagery scenarios.

[0083] S4: Target Recognition

[0084] S41: Model Training

[0085] In this phase, targets in the ground-penetrating radar (GPR) images are first labeled. The labeling process involves selecting bounding boxes for each potential target region and labeling these regions with information including the target's category and location, thus constructing a target detection dataset. This labeled data will be used to train the YOLOv8-GSI model.

[0086] In the first stage of training, the model learns how to extract features from labeled data to accurately identify and locate targets in unlabeled GPR images. The model training employs an iterative optimization method, adjusting model parameters through repeated forward and backward propagation until the desired recognition accuracy and robustness are achieved. During training, special attention is paid to ignoring keypoint detection to ensure the model focuses on target region recognition. The resulting first-stage model is called the first model.

[0087] S42: Model Reasoning

[0088] Once model training is complete, the actual GPR image is input into the pre-trained first model for inference. The model analyzes the input GPR image, identifies the target region in the image, and generates target detection boxes.

[0089] During inference, the model calculates a confidence score for each detection box. A higher confidence score indicates a greater likelihood that the region contains the target. Detection boxes with higher confidence scores from the inference results are retained for subsequent processing. The detection results are as follows: Figure 5 As shown, the model's detection performance under different environments and conditions is demonstrated, verifying the model's accuracy and reliability in practical applications.

[0090] S43: Image Cropping

[0091] Based on the detection bounding boxes generated in the previous step, the GPR image is precisely cropped to extract a small area where a single target exists. The purpose of cropping is to extract the image region within the detection bounding boxes and generate keypoint detection.

[0092] The cropping process segments the original image based on the size and position of the detection boxes, retaining only the content within the boxes. The cropped image is more focused, reducing interference from irrelevant information and improving the accuracy of keypoint detection. The precisely cropped image will be used in the next stage of keypoint detection, ensuring data quality and processing effectiveness at each step.

[0093] S5: Key Point Annotation

[0094] Precise keypoint annotation is performed using the cropped image generated in the first stage. By annotating the vertices and two sets of evenly distributed vertical keypoints with horizontal symmetry in the cropped image, precise localization of individual target locations can be achieved. The selection of these keypoints is based on typical features in GPR images, such as the vertices and edge points of hyperbolas, resulting in a keypoint detection dataset.

[0095] During keypoint annotation, salient feature points in the image need to be manually marked to ensure accuracy and consistency. The annotated image will be used for the next stage of model training, providing foundational data for keypoint detection. A schematic diagram of keypoint annotation is shown below. Figure 6 As shown, this illustrates the distribution of key points in the image and the annotation method.

[0096] S6: Keypoint Detection

[0097] S61: Model Training

[0098] The keypoint detection phase primarily involves specific training of the second-stage YOLOv8-GSI network. The training data consists of cropped images generated in the first stage, where keypoints have been precisely labeled, including the vertices of the target and other important features. Based on the keypoint detection dataset, the YOLOv8-GSI model is trained to identify and locate keypoints in the cropped images.

[0099] The model training employs an optimization method similar to that used in object detection, iteratively updating model parameters to allow the model to gradually learn and accurately predict keypoint locations. During training, the model needs to consider both the location and structural information of keypoints to ensure accurate identification of all important keypoints during detection.

[0100] S62: Model Reasoning

[0101] The trained model will then be used to process the cropped images generated in the first stage. During the inference stage, the cropped images are input into the model, which is responsible for identifying and labeling the locations of keypoints in the images. These keypoint locations are output as image coordinates, providing the basis for keypoint mapping in the next stage.

[0102] The inference results include the specific coordinates and confidence score of each keypoint. The model then selects the most likely keypoint locations based on the confidence scores. The inference results represent the keypoint locations in the cropped image coordinate system.

[0103] S7: Keypoint Mapping

[0104] In this step, based on the mapping relationship of the two-dimensional spatial coordinate system, the positions of key points are transferred from the cropped image to the original image, ensuring that the relative position of each point remains unchanged. The mapping process converts the coordinates of key points in the cropped image to the coordinate system of the original image based on the position information of the detection box.

[0105] Specifically, for each target, the highest point in the middle of its key points is highlighted to ensure that the vertex coordinates accurately reflect the spatial location of the underground target. The mapped result is as follows. Figure 7 As shown, the precise location and distribution of key points in the original image are demonstrated, verifying the accuracy and consistency of the mapping process.

[0106] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for ground penetrating radar data key point detection and target positioning, comprising the following steps, characterized in that: S1: Experimental data collection; The data set includes measured data and simulation data. The measured data is obtained by using ground penetrating radar on civil bridge facilities. The simulation data is generated by GprMax three-dimensional modeling, simulating the conductivity and space occupancy of various media; S2: Data preprocessing; The simulation data set is subjected to direct wave removal and 2D Kirchhoff migration to reduce interference and improve the spatial positioning accuracy of underground structures; S3: YOLOv8-GSI network construction; Integrate global attention mechanism and multi-head self-attention mechanism in YOLOv8 framework, and introduce GIoU loss and CIoU loss to improve target detection accuracy and key point detection performance. Global attention mechanism reduces information loss and puts global interaction representation, enhancing the performance of deep neural network; Multi-head self-attention mechanism processes input features in parallel through multiple attention heads, and combines spatial feature enhancement subnetwork to realize high-resolution mapping of underground structures; GIoU loss and CIoU loss optimize the detection accuracy under non-overlapping and overlapping problems by considering the external area of the bounding box, the distance between the target detection frame center point and the length-width ratio; S4: Target identification; Train the detection frame of YOLOv8-GSI network to build the first model. Through accurate target identification in this stage, the necessary foundation is provided for subsequent key point detection; S5: Key point labeling; Use the first stage model to crop the original data set to obtain sub-images containing only a single target. By labeling the top point and the two groups of vertically uniform key points symmetrically, a total of 5 key points are obtained. The key point detection data set is obtained, the single target contour is accurately fitted, and the spatial position of the target is determined; S6: Key point detection; Firstly, part of the weights of the first model is migrated to the second model to shorten the training time of the second stage model; Secondly, based on the key point labeling data set, the second model is trained; finally, the trained second model is used to perform model inference on the cropped images output by the first stage; S7: Key point mapping; Map the identified key point positions in the cropped images back to the original GPR images. In this process, the positions of the key points need to be accurately transferred from the cropped images to the original images, ensuring that the relative positions of each point remain unchanged. For each target, the highest point among the key points is selected for highlighted labeling, and the top point coordinates determine the spatial position of the underground target, providing an important visual reference for subsequent data analysis and application.

2. The ground penetrating radar data key point detection and target positioning method of claim 1, wherein: The step S3 comprises the following steps: S31: Add global attention mechanism; The global attention mechanism is introduced to improve the performance of deep neural networks by reducing information loss and enhancing global interaction representation, which adopts 3D permutation and multilayer perceptron for channel attention, combined with a convolutional spatial attention submodule to improve the global feature interaction effect. The 3D permutation can preserve information in different dimensions, and the multilayer perceptron further processes these information to amplify the interdependence between channels. The convolutional spatial attention submodule aggregates spatial information through convolution operations to enhance the global feature interaction representation. S32: Add self-attention mechanism; The multi-head self-attention mechanism achieves high-resolution mapping and detail capture by processing input features in parallel. Each attention head independently focuses on different parts of the input features, capturing more detailed information. Combined with the spatial feature enhancement subnet and the deep recursive network, the multi-head self-attention mechanism further improves the resolution of the feature map in the spatial dimension, ensuring high-resolution mapping of underground structures. S33: Add generalized intersection loss; The generalized intersection loss is introduced in the loss function, which calculates the minimum enclosing rectangle of two axis-aligned rectangular bounding boxes and measures the difference between this enclosing rectangle and the joint area of the two bounding boxes, achieving optimization of bounding box regression. S34: Add complete intersection over union loss; The complete intersection over union loss is introduced in the loss function, which considers three geometric factors: overlap area, normalized center point distance, and aspect ratio, significantly improving the accuracy of bounding box regression. 3.The ground penetrating radar data key point detection and target positioning method of claim 1, wherein: The step S4 includes the following steps: S41: Model training; Target annotation is performed on the GPR image, and the YOLOv8-GSI network is used for training to build the first model, which focuses on improving target recognition performance. First, the targets in the GPR image need to be annotated, and the positions of each target are accurately annotated to build a target detection dataset. Then, the YOLOv8-GSI model is trained using the target detection dataset. During training, the keypoint detection part of the model is specifically ignored, and the performance in target recognition is focused on. Through supervised learning, the model is trained to accurately recognize and locate various underground structure targets in the GPR image. S42: Model inference; The trained first-stage YOLOv8-GSI model is used to infer new GPR images. First, input the GPR image to be processed into the trained model, and the model identifies and annotates the target area in the image to quickly locate the position of the underground structure target. Then, according to the output target coordinates, the original image is cropped to obtain a single target cropped target image. S43: Image cropping; According to the model detection frame position, the GPR image is accurately cropped to prepare the input image for the keypoint recognition stage, ensuring the accuracy of the cropping process. 4.The ground penetrating radar data key point detection and target positioning method of claim 1, wherein: The step S6 includes the following steps: S61: Model training The goal of this stage is to train the model to recognize and locate key points in the cropped images. Using the cropped images processed in the first stage, the model is trained to accurately label key points, including the vertices of each target and other important feature points. These labeled images are then used as input to train the second YOLOv8-GSI model. During this process, the model learns how to identify the exact locations of key points from the cropped images, enabling the localization of key points for a single target. S62: Model inference; The trained second model is used to process the images cropped in the first stage. In this stage, the cropped images are input into the model, which is responsible for identifying and marking the locations of key points in the images. This step is a crucial part of the key point detection process, as it directly affects the accuracy of key point mapping. The inference results of the model provide detailed information about the locations of key points for each target, laying the foundation for the subsequent mapping step.

Images