A Temporally Correlated Ground Penetrating Radar Image Retrieval and Localization Method

Abstract

This invention relates to a ground-penetrating radar (GPR) image retrieval and localization method based on temporal correlation, comprising: constructing a map database storing map data based on a GPR system; wherein the map data includes GPR images and GPS location tags; constructing a depth descriptor extractor for GPR images, and generating a map feature library based on the depth descriptor extractor and the map database; acquiring real-time GPR images of a target area, and performing feature extraction based on the depth descriptor extractor to obtain real-time image features related to the real-time GPR images; performing a similarity search based on the real-time image features and the map feature library to determine the coarse registration and localization result of the real-time GPR images; expanding the region of the coarse registration and localization result based on the correlation of consecutive frames of GPR images in the map database; and obtaining the fine registration and localization result of the real-time GPR images using a template matching method.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar navigation and positioning technology, and in particular to a ground-penetrating radar image retrieval and positioning method based on time correlation. Background Technology

[0002] Ground penetrating radar (GPR) transmits high-frequency electromagnetic waves into the ground and analyzes the reflected signals to obtain subsurface information. This system is now widely used in underground target detection, road maintenance, and other fields. With the development of autonomous driving, GPR systems offer a new research direction for existing navigation and positioning systems.

[0003] Ground Penetrating Radar (LGPR) is a novel positioning method utilizing ground penetrating radar technology. Unlike traditional optical and lidar-based assisted driving positioning schemes, this method does not rely on surface environmental information but focuses on underground feature data acquired by LGPR, achieving autonomous positioning through registration of this underground data. Compared to surface scenes, underground scenes change less over time, providing more stable and reliable invariant features for positioning. LGPR technology is unaffected by weather or lighting conditions, complementing existing map-based positioning methods and enhancing the resilience of current positioning technologies to common failure modes and environmental changes. Furthermore, with the development of smart cities and the vigorous promotion of new infrastructure, LGPR systems have broad application prospects in industries such as urban rail transit safety, tunnels, coal mines, logistics, and manufacturing.

[0004] Current ground-penetrating radar (GPR) navigation and positioning technologies primarily focus on map-based matching. The aim is to register real-time acquired GPR images with pre-acquired map data to determine the system's current location. Existing GPR positioning technologies utilize cross-correlation-based similarity metrics and heuristic search-based optimization methods to determine location information. However, due to the limited information content and numerous similar structures in GPR images, mismatches can occur during the positioning search, impacting relocalization, navigation, and tracking. Summary of the Invention

[0005] The purpose of this invention is to provide a ground-penetrating radar image retrieval and localization method based on time correlation.

[0006] To achieve the above-mentioned objectives, this invention provides a ground-penetrating radar image retrieval and localization method based on time correlation, comprising the following steps:

[0007] S1. Construct a map database based on a ground-penetrating radar system to store map data; wherein, the map data includes ground-penetrating radar images and GPS location tags;

[0008] S2. Construct a depth descriptor extractor for the ground-penetrating radar image, and generate a map feature library based on the depth descriptor extractor and the map database;

[0009] S3. Acquire real-time ground-penetrating radar images of the target area, and perform feature extraction based on the depth descriptor extractor to obtain real-time image features related to the real-time ground-penetrating radar images;

[0010] S4. Perform a similarity search based on the real-time image features and the map feature library to determine the coarse registration and positioning result of the ground-penetrating radar real-time image;

[0011] S5. Based on the correlation of consecutive frames of ground-penetrating radar images in the map database, the coarse registration and positioning results are expanded regionally, and the fine registration and positioning results of the real-time ground-penetrating radar images are obtained using a template matching method.

[0012] According to one aspect of the present invention, in step S4, in the step of performing a similarity search based on the real-time image features and the map feature library to determine the coarse registration and positioning result of the ground penetrating radar real-time image, a candidate list for the ground penetrating radar image is constructed based on the search results of the real-time image features and the map feature library, and the candidate list is used as the coarse registration and positioning result.

[0013] Step S5, which involves expanding the region of the coarse registration and positioning result based on the correlation of consecutive frames of ground-penetrating radar images in the map database, and obtaining the fine registration and positioning result using template matching, includes:

[0014] S51. Using the correlation of consecutive frames of ground-penetrating radar images in the map database, the candidate list is expanded to obtain an expanded candidate list;

[0015] S52. Perform image template matching using the real-time ground-penetrating radar image and the candidate extended list, and output the fine registration and positioning result of the real-time ground-penetrating radar image based on the image template matching result.

[0016] According to one aspect of the invention, step S2, the step of constructing a depth descriptor extractor for the ground-penetrating radar image, includes:

[0017] S21. Construct the extractor network architecture for the depth descriptor extractor;

[0018] S22. The extractor network architecture is trained using discrete slice images with a Triplanar structure to obtain the depth descriptor extractor; wherein, the discrete slice images are obtained based on gridded sampling of ground penetrating radar images in the map database.

[0019] According to one aspect of the present invention, in step S21, the step of constructing the extractor network architecture for the deep descriptor extractor, wherein the extractor network architecture comprises: an input layer, a first convolutional layer connected to the input layer, a first wavelet pooling layer connected to the first convolutional layer, a first ReLU activation function layer connected to the first wavelet pooling layer, a second convolutional layer connected to the first ReLU activation function layer, a second wavelet pooling layer connected to the second convolutional layer, a second ReLU activation function layer connected to the second wavelet pooling layer, a third convolutional layer connected to the second ReLU activation function layer, a fourth convolutional layer connected to the third convolutional layer, a fifth convolutional layer connected to the fourth convolutional layer, a third wavelet pooling layer connected to the fifth convolutional layer, a first fully connected layer connected to the third wavelet pooling layer, a dropout layer connected to the first fully connected layer, and a second fully connected layer connected to the dropout layer;

[0020] The input layer is used to receive discrete slice images using a Triplanar structure;

[0021] The second fully connected layer is used to output map feature descriptors corresponding to the discrete slice image.

[0022] According to one aspect of the present invention, step S2, the step of generating a map feature library based on the depth descriptor extractor and the map database, includes:

[0023] S23. Based on the ground-penetrating radar images in the map database, perform gridded sampling to obtain image volume data, and convert the image volume data into discrete slice images using a Triplanar structure;

[0024] S24. The discrete slice images are sequentially input into the depth descriptor extractor for feature extraction to obtain map feature descriptors corresponding to the discrete slice images, and the map feature descriptors are summarized to generate the map feature library.

[0025] According to one aspect of the present invention, step S3, which involves acquiring a real-time ground-penetrating radar image of the target area and performing feature extraction based on the depth descriptor extractor to obtain real-time image features related to the real-time ground-penetrating radar image, includes:

[0026] S31. Acquire real-time ground-penetrating radar images of the target area, perform gridded sampling on the real-time ground-penetrating radar images to obtain real-time volumetric data of the images, and convert the real-time volumetric data of the images into real-time discrete slice images using a Triplanar structure;

[0027] S32. Input the real-time discrete slice image into the depth descriptor extractor for feature extraction, and obtain the real-time feature descriptor corresponding to the real-time discrete slice image to characterize the real-time image features.

[0028] According to one aspect of the present invention, step S4, which involves constructing a candidate list for the ground-penetrating radar image based on the real-time image features and the search results of the map feature library, and using the candidate list as the coarse registration and positioning result, includes:

[0029] S41. Construct an objective function for similarity search between the real-time image features and the map feature library, wherein the objective function is expressed as:

[0030]

[0031] in, f represents the real-time feature descriptor, and f represents the map feature descriptor in the map feature library;

[0032] S42. Calculate the similarity between the real-time feature descriptor and the map feature descriptor based on the objective function, and sort the similarity from high to low;

[0033] S43. Based on the order of the similarity, obtain the image index of the discrete slice image corresponding to the first N1 similarities to construct the candidate list.

[0034] According to one aspect of the present invention, in step S51, the candidate list is expanded by utilizing the correlation of consecutive frames of ground-penetrating radar images in the map database to obtain an expanded candidate list. Based on the order of the consecutive frames of ground-penetrating radar images, the image indexes of the candidate list are sequentially expanded at a preset sampling interval N2 to generate the expanded candidate list.

[0035] According to one aspect of the present invention, step S52, which involves performing image template matching using the real-time ground-penetrating radar image and the candidate extended list, and outputting the fine registration and positioning result of the real-time ground-penetrating radar image based on the image template matching result, includes:

[0036] S521. Based on the image index contained in the candidate extended list, obtain discrete slice images of the ground penetrating radar image and construct a candidate extended image set;

[0037] S522. Based on the real-time discrete slice images of the ground penetrating radar real-time images, and using a similarity metric based on grayscale information to perform image template matching with the candidate extended image set, and sorting according to the image template matching results, the image index with the highest similarity is obtained, thereby determining the fine registration and positioning results of the ground penetrating radar real-time images.

[0038] According to one aspect of the present invention, in step S522, the step of performing image template matching with the candidate extended image set using a similarity metric based on grayscale information, a similarity metric based on grayscale information is constructed using a normalized cross-correlation matching method, which is expressed as:

[0039]

[0040] Where m represents the number of pixels on the horizontal axis of the image, n represents the number of pixels on the vertical axis of the image, i represents the horizontal coordinate position of the pixel, j represents the vertical coordinate position of the pixel, and I1(i,j) is the gray value of the pixel at position (i,j) in the discrete slice image of the ground penetrating radar image in the map database. Let I2(i,j) be the pixel mean of a discrete slice of ground-penetrating radar (GPR) images in the map database, and let I2(i,j) be the gray value of the pixel at position (i,j) in the real-time discrete slice of a GPR image acquired by the GPR. The pixel mean of a real-time discrete slice image of a ground-penetrating radar (GPR) image acquired by GPR.

[0041] According to one aspect of the present invention, to address the problem of weak matching ability of ground-penetrating radar (GPR) images in the matching and positioning task, the present invention introduces a Triplanar structure image group to increase the amount of registration information from the perspective of information content; to address the problem of differences in GPR images of the same road segment at different time phases, a new GPR image feature extraction network is designed by utilizing the robustness of GPR image low-frequency data to solve the GPR image registration task at different time phases; a GPR map database and a GPR map feature library are constructed respectively, and the data collected in real time by the motion platform is compared with the two map databases to achieve coarse-to-fine registration and positioning, reducing the occurrence of mismatches and providing an important guarantee for achieving safe and reliable GPR navigation and positioning.

[0042] According to one aspect of the present invention, the present invention can achieve image registration at different time phases, and can achieve robust positioning, especially under severe weather conditions such as rain and snow. Attached Figure Description

[0043] Figure 1 This is a schematic diagram illustrating the steps of a ground-penetrating radar image retrieval and positioning method according to an embodiment of the present invention;

[0044] Figure 2 This is a flowchart schematically illustrating a ground-penetrating radar image retrieval and positioning method according to an embodiment of the present invention;

[0045] Figure 3 This is a schematic representation of the frequency analysis of ground-penetrating radar images under different weather conditions;

[0046] Figure 4 This is a schematic diagram illustrating the network architecture of the extractor in a ground-penetrating radar image retrieval and localization method according to an embodiment of the present invention.

[0047] Figure 5 This is a schematic flowchart illustrating the image creation process of a Triplanar structure in a ground-penetrating radar image retrieval and positioning method according to an embodiment of the present invention.

[0048] Figure 6 This is a flowchart illustrating a ground-penetrating radar image retrieval and positioning method according to an embodiment of the present invention, in which real-time map features are searched for similarity with a map feature database. Detailed Implementation

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0050] In describing embodiments of the present invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer" express orientations or positional relationships based on the orientations or positional relationships shown in the relevant drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limitations on the present invention.

[0051] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The embodiments cannot be described in detail here, but the embodiments of the present invention are not limited to the following embodiments.

[0052] Combination Figure 1 and Figure 2 As shown, according to one embodiment of the present invention, a ground-penetrating radar image retrieval and localization method based on time correlation includes the following steps:

[0053] S1. Construct a map database based on the ground-penetrating radar system to store map data; wherein, the map data includes ground-penetrating radar images and GPS location tags;

[0054] S2. Construct a depth descriptor extractor for ground-penetrating radar images, and generate a map feature library based on the depth descriptor extractor and the map database;

[0055] S3. Acquire real-time ground-penetrating radar images of the target area and extract features based on the depth descriptor extractor to obtain real-time image features of the ground-penetrating radar images;

[0056] S4. Perform a similarity search based on real-time image features and map feature database to determine the coarse registration and positioning results of the ground penetrating radar real-time image;

[0057] S5. Based on the correlation of consecutive frames of ground-penetrating radar images in the map database, the coarse registration and positioning results are expanded regionally, and the fine registration and positioning results of real-time ground-penetrating radar images are obtained by using template matching.

[0058] According to one embodiment of the present invention, in step S1, the step of constructing a map database for storing map data based on a ground-penetrating radar system includes: a ground-penetrating radar, a GPS positioning device, and a data processing device; wherein, the ground-penetrating radar acquires underground detection data, the GPS positioning device acquires GPS location tags that match the detection data, and the data processing device generates a ground-penetrating radar image based on the detection data, associates the GPS location tags with the generated ground-penetrating radar image, and obtains a grid map; thus, the map data stored in the map database includes the ground-penetrating radar image and the GPS location tags.

[0059] According to one embodiment of the present invention, step S2, the step of constructing a depth descriptor extractor for ground-penetrating radar images, includes:

[0060] S21. Construct an extractor network architecture for the deep descriptor extractor; wherein the extractor network architecture is built based on a deep learning network. Because the amount of image information obtained from real-time acquisition of multi-channel ground-penetrating radar images is relatively small, differences will exist between acquired ground-penetrating radar images due to changes in time, weather, season, and other factors. While the overall contours and shapes of the low-level features in the images remain largely unchanged, the detailed edges exhibit varying degrees of change and blurring. See [link to relevant documentation]. Figure 3 As shown. Therefore, in constructing the extractor network architecture, a wavelet pooling layer is further introduced to focus on robust low-frequency information in the image during feature extraction, thus reducing the impact of clutter interference during matching. For details, see... Figure 4As shown, the extractor network architecture includes: an input layer, a first convolutional layer connected to the input layer, a first wavelet pooling layer connected to the first convolutional layer, a first ReLU activation function layer connected to the first wavelet pooling layer, a second convolutional layer connected to the first ReLU activation function layer, a second wavelet pooling layer connected to the second convolutional layer, a second ReLU activation function layer connected to the second wavelet pooling layer, a third convolutional layer connected to the second ReLU activation function layer, a fourth convolutional layer connected to the third convolutional layer, a fifth convolutional layer connected to the fourth convolutional layer, a third wavelet pooling layer connected to the fifth convolutional layer, a first fully connected layer connected to the third wavelet pooling layer, a dropout layer connected to the first fully connected layer, and a second fully connected layer connected to the dropout layer. In this embodiment, the second fully connected layer serves as the core of the entire extractor network architecture, used to output map feature descriptors corresponding to discrete slice images. Furthermore, to improve the robustness of the feature extraction process, discrete slice images with a Triplanar structure are used as input, fully utilizing the three-dimensional information provided by the ground-penetrating radar system to increase the amount of information in the image used for matching and localization. Therefore, in this embodiment, the input layer is used to receive discrete slice images with a Triplanar structure.

[0061] S22. The depth descriptor extractor is obtained by training the extractor network architecture with discrete slice images using a Triplanar structure; wherein, the discrete slice images are obtained by gridded sampling of ground-penetrating radar images in a map database, for details see Figure 5 As shown, the ground-penetrating radar images with a three-dimensional structure based on the map database are sampled in a grid to obtain image volume data of 20*20*175. Then, discrete slice images are constructed based on the image volume data. The size of the image volume data can be set according to actual needs.

[0062] According to one embodiment of the present invention, step S2, the step of generating a map feature library based on a depth descriptor extractor and a map database, includes:

[0063] S23. Based on the ground-penetrating radar images in the map database, perform grid sampling to obtain image volume data, and convert the image volume data into discrete slice images using the Triplanar structure. The discrete slice image is defined as S = Trip(V), where V represents the 20*20*175 image volume data obtained by grid sampling mentioned in S22;

[0064] S24. The discrete slice images are sequentially input into the depth descriptor extractor for feature extraction, obtaining map feature descriptors corresponding to the discrete slice images, and the map feature descriptors are summarized to generate a map feature library. In this embodiment, the map feature descriptor is represented as:

[0065] f = ψ(S,θ)

[0066] Where ψ() represents the feature descriptor extraction function learned by the network, and θ represents the parameters that can be updated and learned in the network.

[0067] According to one embodiment of the present invention, step S3, which involves acquiring a real-time ground-penetrating radar image of the target area and performing feature extraction based on a depth descriptor extractor to obtain real-time image features related to the real-time ground-penetrating radar image, includes:

[0068] S31. Acquire real-time ground-penetrating radar images of the target area, perform grid sampling on the real-time ground-penetrating radar images to obtain real-time volumetric data of the image volume, and convert the real-time volumetric data of the image volume into real-time discrete slice images using a Triplanar structure; in this embodiment, the construction process of the real-time discrete slice images is consistent with the aforementioned construction process of the discrete slice images, and will not be repeated here.

[0069] S32. Input the real-time discrete slice images into the depth descriptor extractor sequentially to extract features and obtain real-time feature descriptors corresponding to the real-time discrete slice images to characterize the real-time image features.

[0070] Combination Figure 1 , Figure 2 and Figure 6 As shown, according to one embodiment of the present invention, in step S4, the step of performing a similarity search based on real-time image features and a map feature library to determine the coarse registration and positioning result of the ground penetrating radar real-time image involves constructing a candidate list of ground penetrating radar images based on the search results of the real-time image features and the map feature library, and using the candidate list as the coarse registration and positioning result; it includes:

[0071] S41. Construct an objective function for real-time similarity search between image features and map feature databases, wherein the objective function can be constructed based on the l2 norm, and is expressed as:

[0072]

[0073] in, represents the real-time feature descriptor, and f represents the map feature descriptor in the map feature library;

[0074] S42. Calculate the similarity between real-time feature descriptors and map feature descriptors based on the objective function, and sort them from high to low similarity.

[0075] S43. Based on the similarity ranking, obtain the image indices of the discrete slice images corresponding to the top N1 similarity scores to construct a candidate list, where the candidate list is represented as:

[0076]

[0077] Where l1, ..., l N1 The image index representing a discrete slice image.

[0078] Combination Figure 1 , Figure 2 and Figure 6 As shown, according to one embodiment of the present invention, step S5, which involves expanding the region of the coarse registration and positioning result based on the correlation of consecutive frames of ground-penetrating radar images in the map database, and obtaining the fine registration and positioning result using a template matching method, includes:

[0079] S51. The candidate list is expanded using the correlation of consecutive frames of ground-penetrating radar images in the map database to obtain an expanded candidate list. In this embodiment, based on the order of consecutive frames of ground-penetrating radar images, the image indices of the candidate list are sequentially expanded at a preset sampling interval N2 to generate the expanded candidate list. The expanded candidate list is represented as follows:

[0080]

[0081] S52. Perform image template matching using real-time ground-penetrating radar images and a candidate extended list, and output fine registration and positioning results of the real-time ground-penetrating radar images based on the image template matching results, including:

[0082] S521. Based on the image indices contained in the candidate extended list, obtain discrete slice images of the ground penetrating radar image and construct a candidate extended image set; wherein, the candidate extended image set is represented as:

[0083]

[0084] in, Representation and image index The corresponding discrete slice image;

[0085] S522. Based on real-time discrete slice images of ground-penetrating radar (GPR) images, image template matching is performed between the GPR images and a candidate extended image set using a similarity metric based on grayscale information. Based on the image template matching results, the discrete slice images corresponding to the candidate extended image set are reordered according to similarity from high to low to obtain the image index with the highest similarity, thereby determining the fine-grained registration and positioning result of the GPR real-time image. Furthermore, based on the reordered image index with the highest similarity, the GPS location label of the corresponding GPR image can be obtained to complete the output of the positioning result.

[0086] According to one embodiment of the present invention, in step S522, the step of matching image templates with the candidate extended image set using a similarity metric based on grayscale information, employs a normalized cross-correlation (NCC) method to construct a similarity metric based on grayscale information. Specifically, by performing a similarity metric on discrete slice images of ground-penetrating radar images in a map database and real-time discrete slice images of real-time ground-penetrating radar images acquired in real-time, the similarity metric is expressed as:

[0087]

[0088] Where m represents the number of pixels on the horizontal axis of the image, n represents the number of pixels on the vertical axis of the image, i represents the horizontal coordinate position of the pixel, j represents the vertical coordinate position of the pixel, and I1(i,j) is the gray value of the pixel at position (i,j) in the discrete slice image of the ground penetrating radar image in the map database. Let I2(i,j) be the pixel mean of a discrete slice of ground-penetrating radar (GPR) images in the map database, and let I2(i,j) be the gray value of the pixel at position (i,j) in the real-time discrete slice of a GPR image acquired by the GPR. The pixel mean of a real-time discrete slice image of a ground-penetrating radar (GPR) image acquired by GPR.

[0089] In this embodiment, the similarity measure ρ ranges from [-1, 1], and a higher value indicates a higher similarity between the two images.

[0090] The above description is merely an example of a specific solution of the present invention. For any devices and structures not described in detail herein, it should be understood that they are implemented using common devices and methods already available in the art.

[0091] The above description is merely one embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A ground-penetrating radar image retrieval and localization method based on temporal correlation, characterized in that, Includes the following steps: S1. Construct a map database based on a ground-penetrating radar system to store map data; wherein, the map data includes ground-penetrating radar images and GPS location tags; S2. Construct a depth descriptor extractor for the ground-penetrating radar image, and generate a map feature library based on the depth descriptor extractor and the map database; wherein, the step of constructing the depth descriptor extractor for the ground-penetrating radar image includes: S21. Construct an extractor network architecture for the deep descriptor extractor; wherein the extractor network architecture includes: an input layer, a first convolutional layer connected to the input layer, a first wavelet pooling layer connected to the first convolutional layer, a first ReLU activation function layer connected to the first wavelet pooling layer, a second convolutional layer connected to the first ReLU activation function layer, a second wavelet pooling layer connected to the second convolutional layer, a second ReLU activation function layer connected to the second wavelet pooling layer, a third convolutional layer connected to the second ReLU activation function layer, a fourth convolutional layer connected to the third convolutional layer, a fifth convolutional layer connected to the fourth convolutional layer, a third wavelet pooling layer connected to the fifth convolutional layer, a first fully connected layer connected to the third wavelet pooling layer, a dropout layer connected to the first fully connected layer, and a second fully connected layer connected to the dropout layer; The input layer is used to receive discrete slice images using a Triplanar structure; The second fully connected layer is used to output map feature descriptors corresponding to the discrete slice image; S22. The extractor network architecture is trained using discrete slice images with a Triplanar structure to obtain the depth descriptor extractor; wherein, the discrete slice images are obtained based on gridded sampling of ground-penetrating radar images in the map database; S3. Acquire real-time ground-penetrating radar images of the target area, and perform feature extraction based on the depth descriptor extractor to obtain real-time image features related to the real-time ground-penetrating radar images; S4. Perform a similarity search based on the real-time image features and the map feature library to determine the coarse registration and positioning result of the ground penetrating radar real-time image; wherein, a candidate list for the ground penetrating radar image is constructed based on the search results of the real-time image features and the map feature library, and the candidate list is used as the coarse registration and positioning result; S5. Based on the correlation of consecutive frames of ground-penetrating radar images in the map database, the coarse registration and positioning results are expanded regionally, and a fine registration and positioning result of the real-time ground-penetrating radar image is obtained using a template matching method; wherein, the method includes: S51. Expand the candidate list by utilizing the correlation of consecutive ground-penetrating radar image frames in the map database to obtain an expanded candidate list; wherein, based on the order of the consecutive ground-penetrating radar image frames, at a preset sampling interval... N 2. The image indices of the candidate list are sequentially expanded to generate the expanded candidate list; S52. Perform image template matching using the real-time ground-penetrating radar image and the candidate extended list, and output the fine registration and positioning result of the real-time ground-penetrating radar image based on the image template matching result.

2. The ground-penetrating radar image retrieval and positioning method according to claim 1, characterized in that, Step S2, the step of generating a map feature library based on the depth descriptor extractor and the map database, includes: S23. Based on the ground-penetrating radar images in the map database, perform gridded sampling to obtain image volume data, and convert the image volume data into discrete slice images using a Triplanar structure; S24. The discrete slice images are sequentially input into the depth descriptor extractor for feature extraction to obtain map feature descriptors corresponding to the discrete slice images, and the map feature descriptors are summarized to generate the map feature library.

3. The ground-penetrating radar image retrieval and positioning method according to claim 2, characterized in that, Step S3, which involves acquiring a real-time ground-penetrating radar image of the target area and performing feature extraction based on the depth descriptor extractor to obtain real-time image features of the ground-penetrating radar image, includes: S31. Acquire real-time ground-penetrating radar images of the target area, perform gridded sampling on the real-time ground-penetrating radar images to obtain real-time volumetric data of the images, and convert the real-time volumetric data of the images into real-time discrete slice images using a Triplanar structure; S32. Input the real-time discrete slice image into the depth descriptor extractor for feature extraction, and obtain the real-time feature descriptor corresponding to the real-time discrete slice image to characterize the real-time image features.

4. The ground-penetrating radar image retrieval and positioning method according to claim 3, characterized in that, Step S4, which involves constructing a candidate list for the ground-penetrating radar image based on the real-time image features and the search results from the map feature library, and using the candidate list as the coarse registration and positioning result, includes: S41. Construct an objective function for similarity search between the real-time image features and the map feature library, wherein the objective function is expressed as: in, This refers to the real-time feature descriptor. f This represents a map feature descriptor in the map feature library; S42. Calculate the similarity between the real-time feature descriptor and the map feature descriptor based on the objective function, and sort the similarity from high to low; S43. Based on the ranking order of the similarity, obtain the order of the previous... N The image index of the discrete slice image corresponding to the similarity is used to construct the candidate list.

5. The ground-penetrating radar image retrieval and positioning method according to claim 4, characterized in that, Step S52, which involves performing image template matching using the real-time ground-penetrating radar image and the candidate extended list, and outputting the fine registration and positioning result of the real-time ground-penetrating radar image based on the image template matching result, includes: S521. Based on the image index contained in the candidate extended list, obtain discrete slice images of the ground penetrating radar image and construct a candidate extended image set; S522. Based on the real-time discrete slice images of the ground penetrating radar real-time images, and using a similarity metric based on grayscale information to perform image template matching with the candidate extended image set, and sorting according to the image template matching results, the image index with the highest similarity is obtained, thereby determining the fine registration and positioning results of the ground penetrating radar real-time images.

6. The ground-penetrating radar image retrieval and positioning method according to claim 5, characterized in that, In step S522, the step of matching image templates with the candidate extended image set using a similarity metric based on grayscale information, a similarity metric based on grayscale information is constructed using a normalized cross-correlation matching method, which is expressed as: in, m This represents the number of pixels along the horizontal axis of the image. n This represents the number of pixels along the vertical axis of the image. i Indicates the horizontal coordinate position of the pixel. j Indicates the vertical coordinate position of the pixel. The first discrete slice of ground-penetrating radar image in the map database The grayscale value of the location pixel. This represents the pixel mean of a discrete slice of ground-penetrating radar images in a map database. The first discrete slice of a real-time ground-penetrating radar image acquired by the ground-penetrating radar. The grayscale value of the location pixel. The pixel mean of a real-time discrete slice image of a ground-penetrating radar (GPR) image acquired by GPR.

Images