Ground penetrating radar positioning method and system based on space-time U-shaped network

Abstract

The invention relates to a ground penetrating radar positioning method and system based on a space-time U-shaped network, and the method achieves the synchronous extraction of space-time information through the combination of a U-shaped architecture and a residual dense block, enhances the discrimination feature learning capability, achieves the reservation of features related to target response, and reduces the unnecessary noise features through three-dimensional convolution. According to the scheme, the discriminative features of the ground penetrating radar sequence can be extracted to reduce mismatching and effectively capture space and time information, so that excellent positioning performance is achieved in the ground penetrating radar sequence, and high robustness is shown to weather changes.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle positioning, and relates to a ground penetrating radar positioning method and system based on a spatio-temporal U-shaped network. Background Art

[0002] As an important shallow subsurface geophysical exploration technology, ground penetrating radar (GPR) detects underground targets by emitting electromagnetic waves and has been widely used in fields such as automated underground mapping, underground pipeline detection, and road detection. Localizing ground penetrating radar (LGPR), as a cutting-edge technology in the field of autonomous driving, is a vehicle positioning mode based on a prior map, and its detection depth can reach underground structures. As a vehicle positioning system based on a prior map, LGPR can effectively make up for the deficiencies of lidar, cameras, and GPS / INS sensors in bad weather such as fog, rain, and snow or when the surface environment changes. However, the vertically downward detection characteristic of ground penetrating radar results in a large number of false matching candidate regions in a single-frame image. Most of the existing learning-based ground penetrating radar position recognition methods rely on two-dimensional convolutional neural networks, but they cannot effectively capture key temporal information, which restricts the improvement of the positioning performance of ground penetrating radar. Summary of the Invention

[0003] Aiming at the problems existing in the above traditional technologies, the present invention proposes a ground penetrating radar positioning method based on a spatio-temporal U-shaped network and a ground penetrating radar positioning system based on a spatio-temporal U-shaped network, which can improve the positioning performance of ground penetrating radar.

[0004] To achieve the above object, the embodiments of the present invention adopt the following technical solutions: On the one hand, a ground penetrating radar positioning method based on a spatio-temporal U-shaped network is provided, including the steps of: Obtaining a ground penetrating radar scan sequence diagram; Invoking the constructed spatio-temporal U-shaped network to perform feature encoding on the ground penetrating radar scan sequence diagram; the spatio-temporal U-shaped network includes an encoder, a decoder, and a feature compression block, and the encoder uses three-level three-dimensional convolutional layers for feature extraction to capture the spatio-temporal information of the ground penetrating radar scan sequence diagram; After performing three-dimensional convolutional decoding on the encoded feature map output by the encoder using the decoder of the spatio-temporal U-shaped network, a Sigmoid function layer is used for feature activation; the feature map fed forward from the encoder to the decoder is connected through a residual dense block embedded in the skip connection; After compressing the decoded output feature map output by the decoder using the feature compression block, large-scale feature matching is performed using a Faiss index to obtain the graph feature vector of the ground penetrating radar scan sequence; the graph feature vector is used for ground penetrating radar positioning.

[0005] On the other hand, a ground penetrating radar positioning system based on a spatio-temporal U-shaped network is provided, including: A sequence acquisition module for acquiring the ground penetrating radar scanning sequence diagram; A feature encoding module for calling the constructed spatio-temporal U-shaped network to perform feature encoding on the ground penetrating radar scanning sequence diagram; the spatio-temporal U-shaped network includes an encoder, a decoder, and a feature compression block. The encoder uses three-level three-dimensional convolutional layers for feature extraction to capture the spatio-temporal information of the ground penetrating radar scanning sequence diagram; A feature decoding module for performing three-dimensional convolutional decoding on the encoded feature map output by the encoder using the decoder of the spatio-temporal U-shaped network and then using a Sigmoid function layer for feature activation; the feature map fed forward from the encoder to the decoder is connected through the residual dense block in the embedded skip connection; A compression matching module for compressing the decoded output feature map output by the decoder using the feature compression block and then performing large-scale feature matching using the Faiss index to obtain the graph feature vector of the ground penetrating radar scanning sequence; the graph feature vector is used for ground penetrating radar positioning.

[0006] One of the above technical solutions has the following advantages and beneficial effects: The above ground penetrating radar positioning method and system based on the spatio-temporal U-shaped network, by combining the U-shaped architecture with residual dense blocks (RDBs), enhances the discriminative feature learning ability, synchronously extracts spatio-temporal information through three-dimensional convolution, retains the features related to the target response, and reduces the unnecessary noise features. Compared with the traditional method, the above solution can extract the discriminative features of the ground penetrating radar sequence to reduce mis-matching, effectively capture the spatial and temporal information, thereby achieving excellent positioning performance in the ground penetrating radar sequence and showing high robustness to weather changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0008] Figure 1 It is a schematic flowchart of the ground penetrating radar positioning method based on the spatio-temporal U-shaped network in an embodiment; Figure 2 It is a schematic diagram of the network architecture of the spatio-temporal U-shaped network in an embodiment; Figure 3 It is a schematic diagram of the framework of the residual dense block in an embodiment; Figure 4 It is a schematic diagram of the PR curve of the experiment in an embodiment, where Figure 4 (a) is the PR curve on the Run18 (rain) dataset,Figure 4 (b) is the PR curve on the Run90 (snow) dataset; Figure 5 It is a schematic diagram of the results of the ablation experiment in an embodiment; Figure 6 It is a schematic diagram of the module framework of the ground penetrating radar positioning system based on the spatio-temporal U-shaped network in an embodiment. Detailed implementation manners

[0009] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art belonging to the technical field of the present invention. The terms used in the description of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0010] It should be noted that referring to "embodiment" herein means that a specific feature, structure or characteristic described in connection with the embodiment may be included in at least one embodiment of the present invention. Displaying this phrase at various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art can understand that the embodiments described herein can be combined with other embodiments. The term "and / or" used in the description and appended claims of the present invention refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0011] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0012] Since the researchers first proposed the LGPR system architecture, subsequent research has continuously expanded its application scenarios: for example, a GPR positioning framework based on the Monte Carlo method was developed, and a global-local joint positioning strategy in the lunar subsurface environment was also proposed; for another example, in response to the problem of low resolution of GPR images, some literature designed a symmetry-based hyperbolic feature descriptor. However, both traditional signal processing methods and deep learning models are difficult to solve the single-frame mis-matching problem caused by the strong similarity of GPR images.

[0013] Unlike visual / LiDAR positioning that relies on rich surface features, GPR underground features are sparse and difficult to converge. To improve the matching robustness, existing research has also tried to utilize temporal information: for example, the existing method SeqSLAM (Sequential Visual Localization and Mapping) conducts local velocity search through a column-normalized score matrix, the existing methods Network Flow and KNNDTW (K-Nearest Neighbor Dynamic Time Warping) focus on temporal representation learning, and the existing methods SeqVLAD (Sequential Vector Aggregation Descriptor) and SeqOT (Sequential Optimal Transport) respectively adopt sequence compression and attention mechanisms. However, these existing methods fail to achieve joint modeling of spatio-temporal information.

[0014] This specification proposes a Spatio-Temporal U-Net (STU-Net), which synchronously extracts spatial and temporal information from a ground-penetrating radar image sequence through 3D convolution and uses Residual Dense Blocks (RDBs) to achieve multi-scale feature extraction. Experiments on public datasets show that STU-Net outperforms existing methods with significant positioning performance advantages.

[0015] In one embodiment, as Figure 1 shown, a ground-penetrating radar positioning method based on a spatio-temporal U-net may include the following processing steps S10 to S16: S10, obtain a ground-penetrating radar scan sequence map; S12, call the constructed spatio-temporal U-net to perform feature encoding on the ground-penetrating radar scan sequence map; the spatio-temporal U-net includes an encoder, a decoder, and a feature compression block. The encoder uses three-level 3D convolutional layers for feature extraction to capture the spatio-temporal information of the ground-penetrating radar scan sequence map; S14, after using the decoder of the spatio-temporal U-net to perform 3D convolutional decoding on the encoded feature map output by the encoder, use the Sigmoid function layer for feature activation; the feature map fed forward from the encoder to the decoder is connected through the residual dense block in the embedded skip connection; S16, after using the feature compression block to compress the decoded output feature map output by the decoder, perform large-scale feature matching using the Faiss index to obtain the graph feature vector of the ground-penetrating radar scan sequence; the graph feature vector is used for ground-penetrating radar positioning.

[0016] The above-mentioned ground penetrating radar positioning method based on the spatio-temporal U-shaped network enhances the discriminative feature learning ability by combining the U-shaped architecture with residual dense blocks (RDBs), synchronously extracts spatio-temporal information through 3D convolution, retains the features related to the target response, and reduces the unnecessary noise features. Compared with the traditional methods, the above solution can extract the discriminative features of the ground penetrating radar sequence to reduce the mis-matching, effectively capture the spatial and temporal information, so as to achieve excellent positioning performance in the ground penetrating radar sequence and show high robustness to weather changes.

[0017] Specifically, first, given the observation sequence Q l = s l , s l+1 ,…, s l+p T and the prior map M = s 1, s 2,…, s n T ( p is the number of observation frames, n is the number of prior frames), the goal of LGPR is to find the optimal matching position of the observation sequence in the prior map M. To achieve the precise positioning of the observation sequence, first, the feature representation is extracted from the input sequence, and then it is used for matching and positioning. This mode has achieved remarkable success in both visual positioning and lidar-based positioning. In this embodiment, a neural network is constructed to extract the feature representation from each frame of the observation sequence Q l and the prior map M, and the feature representations { q k} and { m k} are obtained respectively. Then, the goal is achieved by finding the optimal matching position of the query descriptor m k in the feature representation { q k}.

[0018] Network architecture part: Intuitively, both the temporal information and the spatial information are crucial for LGPR. To capture the temporal information in the ground penetrating radar (GPR) scan sequence (figure), the proposed spatio-temporal U-shaped network STU-Net first stacks the consecutive frames, and then uses the 3D convolutional layer to extract the spatio-temporal information simultaneously. To model the spatio-temporal correlations at different scales, the proposed spatio-temporal U-shaped network STU-Net adopts the U-shaped structure, and its overall architecture is as Figure 2 shown. STU-Net adopts a three-level architecture of encoder-decoder-feature compression block. ​​

[0019] It can be understood that the encoder has a three-level structure. At the first level of the encoder, the input ground penetrating radar (GPR) sequence (B-scan sequence) is first input into a three-level three-dimensional convolutional layer (3×3×3 kernel) for feature extraction.

[0020] In one embodiment, each three-dimensional convolutional layer of the encoder includes a basic block composed of a double convolutional layer and a ReLU activation function layer, and the convolutional layer and the ReLU activation function layer are alternately connected in series. The first two three-dimensional convolutional layers of the encoder also respectively include a downsampling layer with a stride of 2.

[0021] Furthermore, each three-dimensional convolutional layer includes a basic block (i.e., double convolutional layer + ReLU activation), and the previous three-dimensional convolutional layer also includes a downsampling layer with a stride of 2. Then, a basic block is used to extract deep features, and then the obtained deep features are downsampled with a stride of 2. The feature extraction at the first level can be expressed as: F stage1 =conv 3×3×3 (Blcok(Down( Q l ))) where conv 3×3×3 ()is a three-dimensional convolutional layer, Blcok()is a basic block, and Down()is a convolutional downsampling layer. According to the pattern of the first level, basic blocks and convolutional downsampling layers are used to extract features at a coarser scale. Inside each basic block, two 3×3×3 convolutional layers are used for feature extraction, and the ReLU activation function is used to activate features after each convolutional layer: F out =conv 3×3×3 (ReLU(conv( F in ))) where F out represents the output of the basic block, and F in represents the input of the basic block.

[0022] Thanks to the spatio-temporal modeling ability of the three-dimensional convolutional layer, the encoder can capture spatio-temporal information simultaneously. After the encoder finishes processing, the obtained encoded feature map will be input into the decoder for progressive recovery. Specifically, the encoded feature map obtained from the encoder is first upsampled by a factor of 3, and then input into a basic block and a 3×3×3 convolutional layer. The decoding process of the encoded feature map at the first level can be expressed as: F stage1 =conv 3×3×3 (Blcok(Up( Fstage3 ))) Among them, Up() represents the upsampling operation. Then, this decoding structure is repeated up to three levels (i.e., F stage3 ) and input it into a Sigmoid function layer to generate the final decoded output feature map, and input it into the feature compression block for feature compression to obtain the feature vector of the input ground penetrating radar sequence.

[0023] In one embodiment, when the decoder upsamples the encoded feature map, bilinear interpolation with a scaling factor of 2 is used for upsampling.

[0024] It should be noted that bilinear interpolation upsampling operation with a scaling factor of 2 is used here to ensure the decoding quality.

[0025] After being processed by the decoder, the obtained decoded output feature map will be input into the feature compression block. In particular, three groups of 3×7×3 convolutions are used to compress the feature map to obtain the feature vector of the input sequence, and Faiss index (fc) is used for large-scale feature matching.

[0026] Furthermore, during the process of the feature compression block compressing the decoded output feature map output by the decoder, two convolutional layers are used to compress the features, where each convolutional layer consists of three groups of 3×7×3 convolutions.

[0027] Specifically, at the last stage of the ground penetrating radar spatio-temporal U-shaped network (GPR STU-Net), two convolutional layers (each convolutional layer consists of three groups of 3×7×3 convolutions) are used to compress the features. Then, Faiss index is used to perform similarity matching on a large number of features.

[0028] In the spatio-temporal U-shaped network STU-Net, the features of the encoder and the decoder are connected through the residual dense blocks (RDBs) in the embedded skip connection to avoid the problem of loss of target response information caused by the downsampling operation in the encoder. Taking the low-level feature map such as the B-scan sequence of the ground penetrating radar as an example, the position target information is easily submerged by noise, and directly connecting the features of the encoder and the decoder may introduce unnecessary noise information into the decoder. Therefore, residual dense blocks are used to connect the feature maps fed forward from the encoder to the decoder, and retain the features related to the target response, reducing the unnecessary noise features. The framework of the residual dense block is as Figure 3 shown, which includes three dense connection layers, a local feature fusion layer and a residual learning operation.

[0029] In one embodiment, the spatio-temporal U-shaped network is optimized through the triplet loss function.

[0030] It can be understood that regarding the loss function: a triplet loss function is adopted to optimize the spatio-temporal U-shaped network STU-Net. Specifically, for each query descriptor q k , the set of positive descriptors q pos = q p} and the set of negative descriptors q neg = q n} are used to calculate the triplet loss: .

[0031] In the triplet loss function , is the margin used to distinguish positive and negative samples, α = 0.5 is the boundary threshold, d ( ) is the square of the Euclidean distance, and the geographical distance threshold for positive / negative samples is set to 3 meters. N pos and N neg are respectively q pos and q neg The number of. The closest reference image to each query in the embedding space is found as the final positive sample, and samples with an actual geographical distance exceeding 3 meters are regarded as negative samples.

[0032] In some embodiments, experimental verification is also carried out: Dataset and settings: Use the GROUNDED dataset: containing multi-environment (highway / urban area, etc.) and multi-weather (sunny / rainy / snowy) data collected by 11-channel GPR. The collected data is annotated with the true positioning information of the real-time kinematic global positioning system (RTK-GPS) to provide ground truth positioning. This dataset covers seven different road environments (highway, city center, and urban area), three weather conditions (sunny, rainy, and snowy), and has multi-lane mapping challenges.

[0033] Experimental division: The training set is Route4-Run46 of the GROUNDED dataset. Validation / test maps: Route5-Run56 (sunny, Sunny). The query sets are Run90 (rainy, Rainy), Run57 (sunny), and Run18 (snowy, Snowy) of the GROUNDED dataset.

[0034] Table 1

[0035] Comparison experiment results: The comparison of the positioning performance of the GROUNDED dataset (the bold ones are the optimal results) is shown in Table 1. The method of the present invention (i.e., STU-Net (ours)) significantly outperforms the existing method LGPRNet in the recall@1 metric, with a recall@1 metric of 0.74 compared to only 0.48 for LGPRNet. This is because the spatio-temporal U-shaped network STU-Net of the present invention makes full use of 3D convolution and the U-shaped structure to better capture the spatio-temporal information in the ground penetrating radar (GPR) sequence. In addition, the method of the present invention achieves a high accuracy in practical applications, Figure 4 The precision-recall curve (PR curve) further verifies its robustness in rainy and snowy weather. Among them, Figure 4 (a) is the PR curve on the Run18 (rain) dataset, Figure 4 and (b) is the PR curve on the Run90 (snow) dataset.

[0036] Ablation experiment: Specifically, first, the influence of different sequence lengths on the final accuracy is studied. The method of the present invention is evaluated on ground penetrating radar (GPR) sequences with lengths ranging from 40 to 100, and the performance of the existing methods SeqOT and LGPRNet is shown as a baseline. The average recall@1 is calculated on Run18 and Run90 of the GROUNDED dataset to evaluate the influence of sequence length on performance, and the results are as Figure 5 shown. Figure 5 It shows that when the sequence length increases from 40 to 100, the average recall@1 of the spatio-temporal U-shaped network STU-Net continuously outperforms the baseline methods, verifying the effectiveness of the temporal modeling of the present invention.

[0037] The spatio-temporal U-shaped network (STU-Net) proposed in this paper realizes the spatio-temporal joint modeling of ground penetrating radar sequences through 3D convolution and the U-shaped architecture, achieving significant performance improvement on public datasets and showing strong robustness under multiple weather conditions.

[0038] It should be understood that although Figure 1 the steps in Figure 1 are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover

[0039] In one embodiment, asFigure 6 As shown in Figure 6 , a ground penetrating radar positioning system 100 based on a spatio-temporal U-shaped network is also provided, which includes a sequence acquisition module 11, a feature encoding module 13, a feature decoding module 15, and a compression matching module 17. Among them, the sequence acquisition module 11 is used to acquire the ground penetrating radar scanning sequence diagram. The feature encoding module 13 is used to call the constructed spatio-temporal U-shaped network to perform feature encoding on the ground penetrating radar scanning sequence diagram; the spatio-temporal U-shaped network includes an encoder, a decoder, and a feature compression block. The encoder uses three-level three-dimensional convolutional layers for feature extraction to capture the spatio-temporal information of the ground penetrating radar scanning sequence diagram. The feature decoding module 15 is used to perform three-dimensional convolutional decoding on the encoded feature map output by the encoder using the decoder of the spatio-temporal U-shaped network, and then use the Sigmoid function layer for feature activation; the feature map fed forward from the encoder to the decoder is connected through the residual dense block embedded in the skip connection. The compression matching module 17 is used to compress the decoded output feature map output by the decoder using the feature compression block, and then perform large-scale feature matching using the Faiss index to obtain the graph feature vector of the ground penetrating radar scanning sequence; the graph feature vector is used for ground penetrating radar positioning.

[0040] The above-mentioned ground penetrating radar positioning system 100 based on a spatio-temporal U-shaped network combines the U-shaped architecture with residual dense blocks (RDBs) to enhance the discriminative feature learning ability, synchronously extracts spatio-temporal information through three-dimensional convolution, retains the features related to the target response, and reduces unnecessary noise features. Compared with traditional methods, the above scheme can extract the discriminative features of the ground penetrating radar sequence to reduce mis-matching, effectively capture spatial and temporal information, thereby achieving excellent positioning performance in the ground penetrating radar sequence and showing high robustness to weather changes.

[0041] In one embodiment, each level of the three-dimensional convolutional layer of the encoder includes a basic block composed of a double convolutional layer and a ReLU activation function layer, and the convolutional layer and the ReLU activation function layer are alternately connected in series. The first two levels of the three-dimensional convolutional layer of the encoder also respectively include a downsampling layer with a stride of 2.

[0042] In one embodiment, when the decoder performs upsampling on the encoded feature map, bilinear interpolation with a scaling factor of 2 is used for upsampling.

[0043] In one embodiment, during the process of the feature compression block compressing the decoded output feature map output by the decoder, two convolutional layers are used for compressing features, where each convolutional layer consists of three groups of 3×7×3 convolutions.

[0044] In one embodiment, the spatio-temporal U-shaped network is optimized through a triplet loss function.

[0045] For the specific limitations of the above-mentioned ground penetrating radar positioning system 100 based on the spatio-temporal U-shaped network, reference may be made to the corresponding limitations of the various embodiments of the ground penetrating radar positioning method based on the spatio-temporal U-shaped network in the foregoing text, which will not be elaborated herein.

[0046] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0047] The above embodiments only represent several implementation manners of the present invention, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the protection scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, which all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the appended claims.

Claims

1. A ground penetrating radar positioning method based on a spatio-temporal U-shaped network, characterized in that, Including the steps: Obtain the ground penetrating radar (GPR) scan sequence diagram; Call the constructed spatio-temporal U-shaped network to perform feature encoding on the GPR scan sequence diagram; the spatio-temporal U-shaped network includes an encoder, a decoder, and a feature compression block. The encoder uses three levels of 3D convolutional layers for feature extraction to capture the spatio-temporal information of the GPR scan sequence diagram; After using the decoder of the spatio-temporal U-shaped network to perform 3D convolutional decoding on the encoded feature map output by the encoder, use the Sigmoid function layer for feature activation; the feature map fed forward from the encoder to the decoder is connected through the residual dense block in the embedded skip connection; After using the feature compression block to compress the decoded output feature map output by the decoder, perform large-scale feature matching using the Faiss index to obtain the graph feature vector of the GPR scan sequence; the graph feature vector is used for GPR positioning.

2. The ground penetrating radar positioning method based on the spatio-temporal U-shaped network according to claim 1, wherein Each level of the 3D convolutional layer of the encoder includes a basic block composed of a double convolutional layer and a ReLU activation function layer, and the convolutional layer and the ReLU activation function layer are alternately connected in series. The first two levels of the 3D convolutional layer of the encoder also respectively include a downsampling layer with a stride of 2.

3. The ground penetrating radar positioning method based on the spatio-temporal U-shaped network according to claim 1 or 2, characterized in that, When the decoder performs upsampling on the encoded feature map, it uses bilinear interpolation with a scaling factor of 2 for upsampling.

4. The ground penetrating radar positioning method based on the spatio-temporal U-shaped network according to claim 3, wherein During the process of the feature compression block compressing the decoded output feature map output by the decoder, two convolutional layers are used for compressing features, where each convolutional layer consists of three groups of 3×7×3 convolutions.

5. The ground penetrating radar positioning method based on the spatio-temporal U-shaped network according to claim 3, characterized in that, The spatio-temporal U-shaped network is optimized through the triplet loss function.

6. A ground penetrating radar positioning system based on a spatio-temporal U-shaped network, characterized in that, Including: A sequence acquisition module for obtaining the GPR scan sequence diagram; A feature encoding module for calling the constructed spatio-temporal U-shaped network to perform feature encoding on the GPR scan sequence diagram; the spatio-temporal U-shaped network includes an encoder, a decoder, and a feature compression block. The encoder uses three levels of 3D convolutional layers for feature extraction to capture the spatio-temporal information of the GPR scan sequence diagram; A feature decoding module for using the decoder of the spatio-temporal U-shaped network to perform 3D convolutional decoding on the encoded feature map output by the encoder and then using the Sigmoid function layer for feature activation; the feature map fed forward from the encoder to the decoder is connected through the residual dense block in the embedded skip connection; A compression and matching module for using the feature compression block to compress the decoded output feature map output by the decoder and then performing large-scale feature matching using the Faiss index to obtain the graph feature vector of the GPR scan sequence; the graph feature vector is used for GPR positioning.

7. The ground penetrating radar positioning system based on the spatio-temporal U-shaped network according to claim 6, characterized in that, Each level of the 3D convolutional layer of the encoder includes a basic block composed of a double convolutional layer and a ReLU activation function layer, and the convolutional layer and the ReLU activation function layer are alternately connected in series. The first two levels of the 3D convolutional layer of the encoder also respectively include a downsampling layer with a stride of 2.

8. The ground penetrating radar positioning system based on the spatio-temporal U-shaped network according to claim 6 or 7, characterized in that, When the decoder performs upsampling on the encoded feature map, it uses bilinear interpolation with a scaling factor of 2 for upsampling.

9. The ground penetrating radar positioning system based on the spatio-temporal U-shaped network according to claim 8, wherein, During the process of feature map compression of the decoded output feature map output by the decoder by the feature compression block, two convolutional layers are used to compress features, where each convolutional layer consists of three groups of 3×7×3 convolutions.

10. The ground penetrating radar positioning system based on the spatio-temporal U-shaped network according to claim 8, wherein, The spatio-temporal U-shaped network is optimized through a triplet loss function.

Images