Ground penetrating radar positioning method and device based on sequence registration and computer equipment

Abstract

This application relates to a ground-penetrating radar (GPR) positioning method, apparatus, and computing device based on sequence registration. The method utilizes a trained feature extractor to extract features from each frame of the current GPR image sequence and the sample image sequence, respectively obtaining the corresponding current GPR image features and sample image features. A similarity map matrix is ​​constructed based on the normalized relationship between the current GPR image features and the sample image features. By searching this matrix, a coarse matching sequence is obtained in the sample image sequence that coarsely matches the GPR image sequence. Then, a reordering mechanism is applied to the coarse matching sequence based on historical location data to obtain the final matching sequence. Finally, the current location is determined based on the location corresponding to the final matching sequence. This method enables accurate positioning using GPR under various weather conditions.

Description

Technical Field

[0001] This application relates to the field of radar positioning technology, and in particular to a ground-penetrating radar positioning method, apparatus and computing device based on sequence registration. Background Technology

[0002] Ground-penetrating radar (GPR) positioning and navigation technology is an effective solution to overcome the shortcomings of positioning methods based on optical cameras and lidar. This technology is gradually gaining widespread attention. Due to the advantage that underground environments are less prone to change than above-ground environments, this technology holds promise as an effective solution for positioning in various extreme and dynamic environments in the future. Current methods for robot localization using GPR are based on a prior map and employ feature matching for robot relocalization.

[0003] However, this approach still faces many challenges. Firstly, the limited observation range of ground-penetrating radar results in numerous false positive candidate matches in the real-time data collected on the map. Sequence information matching has effectively increased the robot's perception range and improved the robustness of the robot system's localization in past robot localization research. Secondly, changes in groundwater content caused by rain and snow pose a significant challenge to achieving autonomous localization. Summary of the Invention

[0004] Therefore, it is necessary to provide a ground-penetrating radar positioning method, device, and computing equipment based on sequence registration that can achieve stable matching under various weather conditions and adapt to multiple weather data, in order to address the above-mentioned technical problems.

[0005] A ground-penetrating radar (GPR) localization method based on sequence registration, the method comprising:

[0006] Acquire the current echo data obtained by ground-penetrating radar scanning, wherein the current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time;

[0007] Extract sample image sequences related to the current location from the map database, and use a trained feature extractor to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features.

[0008] A similarity map matrix is ​​constructed based on the normalized relationship between each current ground-penetrating image feature and each sample image feature;

[0009] By searching the similarity map matrix, a coarse matching sequence that coarsely matches the ground exploration image sequence is obtained in the sample image sequence;

[0010] The coarse matching sequence is reordered based on historical location data to obtain the final matching sequence, and the current location is determined based on the position corresponding to the final matching sequence.

[0011] In one embodiment, by training a deep Siamese network using template matching, the feature extraction layer in the trained deep neural network is separated and used as the trained feature extractor.

[0012] In one embodiment, training the deep Siamese network includes:

[0013] Acquire multiple template images and multiple search images obtained by ground-penetrating radar scanning the same moving route at different times;

[0014] The template image and the search image corresponding to the same position coordinates are used as a set of training data pairs, and a training dataset containing multiple sets of training data pairs is constructed accordingly.

[0015] The training dataset is input into the deep Siamese network. In the deep Siamese network, the feature extraction layer extracts the feature codes of the template image and the search image in a set of training data pairs respectively. Then, through the cross-correlation layer, the similarity score map of the template image and the search image is output according to the feature codes.

[0016] The logistic loss function is calculated based on the similarity score map and ground truth map predicted by the deep Siamese network. The parameters in the deep Siamese network are adjusted according to the calculation results until the calculation results converge, thus obtaining the trained deep Siamese network.

[0017] In one embodiment, in each set of training data pairs, the center coordinates of the template image and the search image coincide, and the size of the search image is larger than that of the template image.

[0018] In one embodiment, obtaining a coarse matching sequence that coarsely matches the ground-penetrating image sequence in the sample image sequence by searching the similarity map matrix includes:

[0019] Based on the preset coarse matching sequence length and velocity search range, the accumulated similarity of each current ground exploration image in each sample image is calculated in the similarity map matrix, and the best matching sequence is selected as the coarse matching sequence based on the accumulated similarity.

[0020] In one embodiment, the accumulated similarity is calculated using the following formula:

[0021]

[0022] In the above formula, T represents the current position index of the ground penetrating radar system, t is the index of the current ground penetrating image sequence acquired in real time, d is the index of the corresponding sample image sequence, and dl is the preset coarse matching sequence length.

[0023] Specifically, the index d of the sample image sequence corresponding to the current ground-penetrating image sequence index t is found based on a preset velocity search range:

[0024] d = M + V(t + dl - T)

[0025] In the above formula, M represents the map tile index used to calculate the cumulative similarity, and V represents the trajectory velocity.

[0026] In one embodiment, the final matching sequence is obtained by reordering the coarse matching sequence based on historical location data:

[0027] The current location is predicted by analyzing the historical locations at multiple consecutive moments.

[0028] The continuity weight is obtained based on the current predicted position, and the accumulated similarity corresponding to each frame of the coarse matching sequence is updated based on the continuity weight.

[0029] The updated accumulated similarities are reordered, and the final matching sequence is obtained based on the sample image corresponding to the highest accumulated similarity.

[0030] In one embodiment, the continuity weight is represented as:

[0031]

[0032] In the above formula, pi represents the index of the i-th coarse matching sequence, px represents the index of the sample image corresponding to the predicted current position in the sample image sequence, and k and σ represent preset parameters.

[0033] A ground-penetrating radar positioning device based on sequence registration, the device comprising:

[0034] The real-time data acquisition module is used to acquire the current echo data obtained by ground-penetrating radar scanning. The current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time.

[0035] The image feature extraction module is used to extract sample image sequences related to the current location from the map database. The trained feature extractor is used to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features.

[0036] The similarity graph matrix construction module is used to construct a similarity graph matrix based on the normalized relationship between each current ground-penetrating image feature and each sample image feature;

[0037] The coarse matching sequence acquisition module is used to obtain a coarse matching sequence that coarsely matches the ground exploration image sequence in the sample image sequence by searching the similarity map matrix;

[0038] The current location positioning module is used to obtain the final matching sequence by reordering the coarse matching sequence based on historical location data, and to locate the current location based on the position corresponding to the final matching sequence.

[0039] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0040] Acquire the current echo data obtained by ground-penetrating radar scanning, wherein the current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time;

[0041] Extract sample image sequences related to the current location from the map database, and use a trained feature extractor to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features.

[0042] A similarity map matrix is ​​constructed based on the normalized relationship between each current ground-penetrating image feature and each sample image feature;

[0043] By searching the similarity map matrix, a coarse matching sequence that coarsely matches the ground exploration image sequence is obtained in the sample image sequence;

[0044] The coarse matching sequence is reordered based on historical location data to obtain the final matching sequence, and the current location is determined based on the position corresponding to the final matching sequence.

[0045] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0046] Acquire the current echo data obtained by ground-penetrating radar scanning, wherein the current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time;

[0047] Extract sample image sequences related to the current location from the map database, and use a trained feature extractor to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features.

[0048] A similarity map matrix is ​​constructed based on the normalized relationship between each current ground-penetrating image feature and each sample image feature;

[0049] By searching the similarity map matrix, a coarse matching sequence that coarsely matches the ground exploration image sequence is obtained in the sample image sequence;

[0050] The coarse matching sequence is reordered based on historical location data to obtain the final matching sequence, and the current location is determined based on the position corresponding to the final matching sequence.

[0051] The aforementioned ground-penetrating radar (GPR) positioning method, device, and computing equipment based on sequence registration utilize a trained feature extractor to extract features from each frame of the current GPR image sequence and the sample image sequence, respectively, to obtain the corresponding current GPR image features and sample image features. A similarity map matrix is ​​constructed based on the normalized relationship between the current GPR image features and the sample image features. By searching this matrix, a coarse matching sequence that coarsely matches the GPR image sequence is obtained in the sample image sequence. Then, a reordering mechanism is applied to the coarse matching sequence based on historical location data to obtain the final matching sequence. The current location is then located based on the location corresponding to the final matching sequence. This method can achieve accurate positioning using GPR under various weather conditions. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating a ground-penetrating radar (GPR) localization method based on sequence registration in one embodiment.

[0053] Figure 2 This is a flowchart illustrating a method for training a deep Siamese network in one embodiment;

[0054] Figure 3 This is a schematic diagram of the structure of a deep Siamese network based on U-Net in one embodiment;

[0055] Figure 4 This is a schematic diagram showing the feature comparison of the current ground-penetrating image before and after passing through the feature extractor in one embodiment.

[0056] Figure 5 This is a schematic diagram illustrating coarse sequence matching using a limited speed range in one embodiment;

[0057] Figure 6 This is a flowchart illustrating the ground-penetrating radar sequence matching and localization framework using this method in one embodiment;

[0058] Figure 7 A detailed schematic diagram illustrating the positioning results using this method in rainy or snowy weather;

[0059] Figure 8 This is a structural block diagram of a ground-penetrating radar positioning device based on sequence registration in one embodiment;

[0060] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] To address the issues of numerous false positive candidate matches and varying underground scene conditions in existing ground-penetrating radar (GPR) positioning technologies, in one embodiment, such as... Figure 1 As shown, a ground-penetrating radar (GPR) localization method based on sequence registration is provided, including the following steps:

[0063] Step S100: Obtain the current echo data obtained by ground-penetrating radar scanning. The current echo data is a current ground-penetrating image sequence that includes multiple frames of images sorted by time.

[0064] Step S110: Extract sample image sequences related to the current location from the map database, and use the trained feature extractor to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features respectively.

[0065] Step S120: Construct a similarity map matrix based on the normalized relationship between each current ground-penetrating image feature and each sample image feature;

[0066] Step S130: By searching the similarity map matrix, a coarse matching sequence that coarsely matches the ground exploration image sequence is obtained in the sample image sequence;

[0067] Step S140: The coarse matching sequence is reordered based on historical location data to obtain the final matching sequence, and the current location is located based on the position corresponding to the final matching sequence.

[0068] Because ground-penetrating radar (GPR) is typically deployed close to the ground, the area of ​​ground imagery it acquires is generally small. Therefore, when locating a position based on real-time ground imagery, the current image is usually compared one by one with images in an existing map database to obtain a matching image. The location information in the matching image is then used to pinpoint the current location and enable navigation. However, due to the limited observation range of GPR and the scarcity of location information, a large number of false positive matches occur during the matching process, resulting in inaccurate positioning. Furthermore, weather conditions can also contribute to positioning inaccuracies.

[0069] In this embodiment, to overcome the above two problems, a method based on inter-sequence registration rather than single-frame image registration is provided, and a method for extracting stable features of ground-penetrating images obtained under multiple weather conditions is also designed.

[0070] Specifically, in step S100, the ground-penetrating radar scans the ground in real time to obtain a sequence of multiple consecutive ground-penetrating images within a preset time period. The current ground-penetrating image sequence is matched in a map database to locate the current position based on the specific geographic information in the map database.

[0071] In step S110, the multiple sample images in the map library are also images obtained by ground-penetrating radar (GPR) observation of the ground, but their acquisition time is much earlier than the current time. That is to say, the sample images in the map library are images of ground observation using GPR on different roads in advance, and each frame image has specific geographical location information. Therefore, before performing sequence matching, multiple sample images of the same road related to the current location can be extracted from the map library.

[0072] In this embodiment, by training the deep Siamese network using template matching, the feature extraction layer in the trained deep neural network is separated and used as a trained feature extractor.

[0073] Specifically, the existing deep twin network framework is used to train the ground penetrating radar image template matching. The trained template matching network is then split, and the feature extraction layer can be extracted to provide robust multi-weather features for subsequent similarity map construction.

[0074] In this embodiment, as Figure 2 As shown, a method for training a deep Siamese network is also provided, the specific steps of which include:

[0075] Step S200: Obtain multiple template images and multiple search images obtained by ground penetrating radar scanning the same moving route at different times;

[0076] Step S210: Take the template image and the search image corresponding to the same position coordinate as a set of training data pairs, and construct a training dataset containing multiple sets of training data pairs accordingly.

[0077] Step S220: Input the training dataset into the deep Siamese network. In the deep Siamese network, the feature extraction layer extracts the feature codes of the template image and the search image in a set of training data pairs respectively. Then, through the cross-correlation layer, the similarity score map of the template image and the search image is output according to the feature codes.

[0078] Step S230: Calculate the logstic loss function based on the similarity score map and ground truth map predicted by the deep Siamese network. Adjust the parameters in the deep Siamese network according to the calculation results until the calculation results converge, and then obtain the trained deep Siamese network.

[0079] In step S200, ground-penetrating radar scanning data is received and processed in real time, and the ground-penetrating radar data of the same road segment under different weather conditions and different lane positions is synchronously located using RTK GPS and ground-penetrating radar timestamps.

[0080] In step S210, in each training data pair, the center coordinates of the template image and the search image coincide, and the size of the search image is larger than that of the template image.

[0081] Specifically, in the actual construction of template and search images, a 50×50 template image T and a 150×150 search image S can be extracted from the same radar observation image, using different locations as center points. A training dataset of 4000 pairs is constructed using random sampling and the above strategy.

[0082] Next, an end-to-end multi-weather template matching network MWS-Net based on deep learning is constructed. MWS-Net is based on the Siamese network framework and includes a feature extraction layer and a cross-correlation layer. The feature extraction layer uses the U-Net network architecture.

[0083] In step S220, the template image and the search image are encoded by the feature extraction layer and then output as a similarity score map through the cross-correlation layer.

[0084] Next, in step S230, the logstic loss function is calculated based on the similarity score map predicted by the deep Siamese network and the ground truth map. The logstic loss function is expressed as:

[0085] l(y,v)=log(1+exp(-yv)) (1)

[0086] In formula (1), y represents the ground truth binary image corresponding to the image pair, and v represents the similarity score image predicted by the network. The loss of the similarity score image is defined as the mean loss between each pixel:

[0087]

[0088] If the pixel coordinates of the similarity score map are within a radius R of the center (considering the network's stride k), then they are considered positive examples, i.e., their label y is:

[0089]

[0090] In this embodiment, U-Net is used as the backbone network of the deep twin network, such as Figure 3 As shown, the network was trained on the large public ground-penetrating radar dataset GROUNDED using the Adam stochastic gradient descent optimizer, with a batch size of 32 and a weight decay of 0.0001. The entire network was trained for 5 epochs.

[0091] In one embodiment, training of the network is implemented using PyTorch on a server with an RTX 3090 GPU and an Intel Core i9-9900X CPU.

[0092] Next, in the trained deep Siamese network, the convolutional layers are extracted as feature extractors to extract features from each frame of the current ground exploration image sequence and the sample image sequence, respectively, to obtain the corresponding current ground exploration image features and sample image features.

[0093] like Figure 4 As shown, this is a schematic diagram comparing the features of the current ground-penetrating image before and after passing through the feature extractor. In this diagram, 4(a) is the original feature map of the current ground-penetrating image, and 4(b) is the feature map of the current ground-penetrating image after being extracted by the feature extractor.

[0094] In step S120, a similarity map matrix is ​​constructed based on the normalized relationship between each current ground-penetrating image feature q and each sample image feature m. The normalized cross-correlation is calculated as follows:

[0095]

[0096] In formula (4), S(i,j) specifically represents the normalized cross-correlation degree between the current ground-penetrating image of frame i and the sample image of frame j, which is the similarity. The size of the similarity map matrix is ​​i*j.

[0097] In step S130, by searching the similarity map matrix, a coarse matching sequence that coarsely matches the ground exploration image sequence is obtained in the sample image sequence. This includes: calculating the accumulated similarity of each current ground exploration image in each sample image in the similarity map matrix according to the preset coarse matching sequence length and velocity search range, and selecting the best matching sequence as the coarse matching sequence based on the accumulated similarity.

[0098] In one embodiment, the speed search range is set to [0.8, 1.2].

[0099] In this embodiment, the accumulated similarity is calculated using the following formula:

[0100]

[0101] In formula (5), T represents the current position index of the ground penetrating radar system, t is the index of the current ground penetrating image sequence acquired in real time, d represents the index of the corresponding sample image sequence, and dl is the preset coarse matching sequence length.

[0102] Because the ground-penetrating radar moves at different speeds when acquiring the current ground-penetrating image sequence and the sample image sequence, the relationship between indices t and d is not one-to-one. Therefore, by setting different speed search ranges, the index d of the sample image sequence corresponding to index t of the current ground-penetrating image sequence is found:

[0103] d=M+V(t+dl-T) (6)

[0104] In formula (6), M represents the map tile index used to calculate the cumulative similarity, and V represents the trajectory velocity.

[0105] In this embodiment, the process of performing coarse matching on the similarity map matrix using a limited speed range is as follows: Figure 5 As shown.

[0106] In this embodiment, the matching of the data collected at the current moment through speed search is a coarse match. Then, a reordering mechanism is used to rearrange the candidate matching (coarse match) items to achieve a fine match.

[0107] In step S140, the coarse matching sequence is reordered based on historical location data to obtain the final matching sequence: the current location is predicted based on the historical locations at multiple consecutive times to obtain the current predicted location; the continuity weight is obtained based on the current predicted location; the accumulated similarity corresponding to each frame of the coarse matching sequence is updated based on the continuity weight; finally, the updated accumulated similarity is reordered, and the final matching sequence is obtained based on the sample image corresponding to the highest accumulated similarity.

[0108] In this embodiment, historical location data refers to multiple consecutive locations estimated before the current time, which are obtained by removing sample images that are far from the historical locations in the coarse matching sequence.

[0109] Specifically, assuming the ground-penetrating radar moves at a constant speed within the selected sequence length, that is, the historical trajectory over a period of time can be estimated using a simple constant speed motion model to estimate the displacement change Δpx and predict the current position px:

[0110]

[0111] Then, based on the estimated current position, a continuity weight is designed to reassign the accumulated similarity corresponding to the sample images in the coarse matching sequence. The continuity weight is expressed as:

[0112]

[0113] In formula (8), pi represents the index of the i-th coarse matching sequence, and px represents the index of the sample image corresponding to the predicted current position in the sample image sequence. The closer to the current position, the higher the weight; the farther away from the current position, the lower the weight. k and σ represent preset parameters.

[0114] In one embodiment, k = 15 and σ = 20.

[0115] Next, the sample images in the coarse matching sequence are reordered based on the reassigned accumulated similarity. The position with the highest accumulated similarity is the best matching point. The modified accumulated similarity... The calculation is as follows:

[0116]

[0117] Finally, select the accumulated similarity. The value with the highest value is the final sequence localization result.

[0118] In this embodiment, as Figure 6 As shown, a framework flowchart for ground-penetrating radar sequence matching and localization using this method is also provided.

[0119] In this paper, experimental simulations are also performed based on the proposed method, and the method is tested using evaluation schemes in the field of visual location recognition. For each tracking sequence, it is aligned with map library data via RTK GPS.

[0120] First, ablation experiments validated that sequence registration improves performance. Our method was validated on the publicly available GROUNDED dataset and our own collected dataset under different weather conditions and multi-lane experimental challenges using the recall metric Recall@K (K=1,5,15). The results are shown in Table 1.

[0121] Table 1 compares the performance of different solutions under various challenging scenarios.

[0122]

[0123] The baseline algorithm uses unprocessed raw GPR data, while the comparison algorithms employ advanced hand-crafted descriptors gprHOG and HOVPS. The baseline for the sequence registration algorithm is seqSLAM, a commonly used visual sequence registration scheme. In the experiments, the sequence length dl was set to 10. By using the sequence registration scheme, the matching performance of all features was improved by approximately 20%. The MWS-seq method designed in this paper achieved the best results in almost all experimental settings, effectively solving the problem of excessive false positives.

[0124] In addition, to analyze the effects of different sequence lengths and reordering mechanisms in more detail, this invention further conducted ablation experiments on the GROUNDED dataset and its own dataset, and the results are shown in Table 2.

[0125] Table 1 Ablation experiments with sequence length and reordering

[0126]

[0127] It can be observed that the localization performance of different schemes is significantly improved with the increase of sequence length. When the sequence length exceeds 10, even the original data without any feature processing can yield good localization results. The reordering mechanism based on motion continuity in this method also effectively improves performance. Specifically, the improvement in reordering performance is positively correlated with the accuracy of past historical trajectories; the more accurate the past position estimation, the more significant the performance improvement from reordering.

[0128] like Figure 7 The diagram shows a detailed illustration of the positioning results using this method in rainy or snowy weather.

[0129] In the aforementioned ground-penetrating radar (GPR) localization method based on sequence registration, to overcome the shortcomings of existing technologies in multi-weather matching localization, a deep learning-based matching method is provided. This method utilizes supervised learning to design a multi-weather GPR image template matching task, and uses the feature extraction structure of the network to extract stable features for multiple weather conditions, and uses real-time data q i With map data m j After passing through a feature extraction network and calculating the similarity using normalized cross-correlation coefficients, the similarity score S between the i-th frame of real-time data and the j-th frame of map data is obtained. i,jThe similarity matrix M is obtained by calculating the similarity between real-time data and map data one by one. Since the data collected at different times may have different collection speeds, to adapt sequence matching to different collection speeds, a speed search range can be set. Corresponding sequences are collected and matched in the form of different 'rays' on the similarity matrix. The speed range limitation not only adapts to different collection speeds but also limits the matching search area, reducing search complexity. Considering the continuity of sequence positioning, this method also designs a re-ranking mechanism to fine-tune and re-rank candidate matches. Based on past positioning results, a simple assumption is made that the motion in a sequence is uniform. Therefore, a simple uniform motion model can be used to give the pose estimate of the current moment based on past positioning results, and corresponding weights are designed for positions closer to the current pose estimate for re-ranking. In this method, a trainable deep learning process is designed. By establishing a more easily trained task, stable and unchanging underground features under various weather conditions are effectively obtained. This method effectively solves the problem of excessive false positives in the matching process caused by sensor limitations, and designs a sequence matching-based framework, laying the foundation for ground-penetrating radar-based positioning and navigation tasks.

[0130] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0131] In one embodiment, such as Figure 8 As shown, a ground-penetrating radar positioning device based on sequence registration is provided, including: a real-time data acquisition module 300, an image feature extraction module 310, a similarity map matrix construction module 320, a coarse matching sequence acquisition module 330, and a current location positioning module 340, wherein:

[0132] The real-time data acquisition module 300 is used to acquire the current echo data obtained by ground-penetrating radar scanning, wherein the current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time.

[0133] The image feature extraction module 310 is used to extract sample image sequences related to the current location from the map library, and to use a trained feature extractor to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features respectively.

[0134] The similarity graph matrix construction module 320 is used to construct a similarity graph matrix based on the normalized relationship between each current ground exploration image feature and each sample image feature;

[0135] The coarse matching sequence acquisition module 330 is used to obtain a coarse matching sequence that coarsely matches the ground exploration image sequence in the sample image sequence by searching the similarity map matrix;

[0136] The current location positioning module 340 is used to obtain the final matching sequence by reordering the coarse matching sequence based on historical location data, and to locate the current location based on the position corresponding to the final matching sequence.

[0137] Specific limitations regarding the sequence registration-based ground-penetrating radar (GPR) positioning device can be found in the limitations of the sequence registration-based GPR positioning method described above, and will not be repeated here. Each module in the aforementioned sequence registration-based GPR positioning device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0138] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores map database data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a ground-penetrating radar (GPR) positioning method based on sequence registration.

[0139] Those skilled in the art will understand that Figure 9The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0140] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0141] Acquire the current echo data obtained by ground-penetrating radar scanning, wherein the current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time;

[0142] Extract sample image sequences related to the current location from the map database, and use a trained feature extractor to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features.

[0143] A similarity map matrix is ​​constructed based on the normalized relationship between each current ground-penetrating image feature and each sample image feature;

[0144] By searching the similarity map matrix, a coarse matching sequence that coarsely matches the ground exploration image sequence is obtained in the sample image sequence;

[0145] The coarse matching sequence is reordered based on historical location data to obtain the final matching sequence, and the current location is determined based on the position corresponding to the final matching sequence.

[0146] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0147] Acquire the current echo data obtained by ground-penetrating radar scanning, wherein the current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time;

[0148] Extract sample image sequences related to the current location from the map database, and use a trained feature extractor to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features.

[0149] A similarity map matrix is ​​constructed based on the normalized relationship between each current ground-penetrating image feature and each sample image feature;

[0150] By searching the similarity map matrix, a coarse matching sequence that coarsely matches the ground exploration image sequence is obtained in the sample image sequence;

[0151] The coarse matching sequence is reordered based on historical location data to obtain the final matching sequence, and the current location is determined based on the position corresponding to the final matching sequence.

[0152] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0153] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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, they should be considered to be within the scope of this specification.

[0154] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A sequence registration based ground penetrating radar positioning method, characterized in that, The method includes: Acquire the current echo data obtained by ground-penetrating radar scanning, wherein the current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time; Extract sample image sequences related to the current location from the map database, and use a trained feature extractor to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features. A similarity map matrix is ​​constructed based on the normalized relationship between each current ground-penetrating image feature and each sample image feature; By searching the similarity map matrix, a coarse matching sequence that coarsely matches the ground-penetrating image sequence is obtained in the sample image sequence. Specifically, based on a preset coarse matching sequence length and velocity search range, the accumulated similarity of each current ground-penetrating image in each sample image is calculated in the similarity map matrix, and the best matching sequence is selected as the coarse matching sequence based on the accumulated similarity. The accumulated similarity is calculated using the following formula: In the above formula, T represents the current position index of the ground penetrating radar system, t is the index of the current ground penetrating image sequence acquired in real time, d is the index of the corresponding sample image sequence, and dl is the preset coarse matching sequence length. Specifically, the index d of the sample image sequence corresponding to the current ground-penetrating image sequence index t is found based on a preset velocity search range: In the above formula, M represents the map tile index used to calculate the cumulative similarity, and V represents the trajectory velocity; The coarse matching sequence is reordered based on historical location data to obtain the final matching sequence. The current location is then determined based on the position corresponding to the final matching sequence. Specifically, the current location is predicted based on historical locations from multiple consecutive moments to obtain the predicted current location. A continuity weight is obtained based on the predicted current location, and the accumulated similarity of each frame in the coarse matching sequence is updated based on this continuity weight. The updated accumulated similarities are then reordered, and the final matching sequence is obtained based on the sample image corresponding to the highest accumulated similarity. The continuity weight is expressed as follows: In the above formula, denotes an index of the i-th rough matching sequence, denotes an index of a sample image corresponding to the predicted current position in the sample image sequence, and denotes a preset parameter.

2. The ground penetrating radar positioning method of claim 1, wherein, By training a deep Siamese network using template matching, the feature extraction layer in the trained deep neural network is separated and used as the trained feature extractor.

3. The ground penetrating radar positioning method of claim 2, wherein, Training the deep Siamese network includes: Acquire multiple template images and multiple search images obtained by ground-penetrating radar scanning the same moving route at different times; The template image and the search image corresponding to the same position coordinates are used as a set of training data pairs, and a training dataset containing multiple sets of training data pairs is constructed accordingly. The training dataset is input into the deep Siamese network. In the deep Siamese network, the feature extraction layer extracts the feature codes of the template image and the search image in a set of training data pairs respectively. Then, through the cross-correlation layer, the similarity score map of the template image and the search image is output according to the feature codes. The logistic loss function is calculated based on the similarity score map and ground truth map predicted by the deep Siamese network. The parameters in the deep Siamese network are adjusted according to the calculation results until the calculation results converge, thus obtaining the trained deep Siamese network.

4. The ground penetrating radar positioning method of claim 3, wherein, In each of the training data pairs, the center coordinates of the template image and the search image coincide, and the size of the search image is larger than that of the template image.

5. A sequence registration based ground penetrating radar positioning apparatus, characterized by, The device implements the ground-penetrating radar positioning method based on sequence registration as described in any one of claims 1-4, including: The real-time data acquisition module is used to acquire the current echo data obtained by ground-penetrating radar scanning. The current echo data is a current ground-penetrating image sequence including multiple frames of images sorted by time. The image feature extraction module is used to extract sample image sequences related to the current location from the map database. The trained feature extractor is used to extract features from each frame of the current ground exploration image sequence and the sample image sequence to obtain the corresponding current ground exploration image features and sample image features. The similarity graph matrix construction module is used to construct a similarity graph matrix based on the normalized relationship between each current ground-penetrating image feature and each sample image feature; The coarse matching sequence acquisition module is used to obtain a coarse matching sequence that coarsely matches the ground exploration image sequence in the sample image sequence by searching the similarity map matrix; The current location positioning module is used to obtain the final matching sequence by reordering the coarse matching sequence based on historical location data, and to locate the current location based on the position corresponding to the final matching sequence. 6.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-5 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

Images