A teaching-repeating positioning method based on millimeter wave radar matching network

Abstract

The present application relates to the field of SLAM and deep learning, and proposes a teaching-repeating positioning method based on a millimeter wave radar matching network, designs a millimeter wave radar matching network, and constructs a geographic positioning system based on the network, mainly uses the millimeter wave radar with global positioning data to collect signal image data in the target path to construct a teaching sequence, then, the relative pose transformation calculation is carried out through the registration result of the repeating sequence and the teaching sequence, once the teaching sequence is established, the subsequent area does not need to rely on the GPS device to obtain the global positioning data again, and only needs to use the relative relationship between the current data and the prior data to realize accurate positioning. The present application solves the problem that the existing positioning scheme cannot accurately position under the condition that GPS is bad, and improves the strong robustness under various weather conditions, and can also realize all-weather high-precision positioning for a long time in the absence of GPS signals.

Description

Technical Field

[0001] This invention relates to the fields of SLAM and deep learning, specifically to a teach-repeated localization method based on millimeter-wave radar matching networks. Background Technology

[0002] SLAM (Simultaneous Localization and Mapping) and deep learning are currently hot research topics in the field of robot navigation and localization. The goal of SLAM is to solve the problems of environmental perception and autonomous navigation by simultaneously performing map building and robot localization. It typically utilizes sensor data such as cameras and LiDAR for real-time map building and pose estimation. However, traditional SLAM methods may have limitations in certain situations, such as in environments with missing or unreliable GPS signals.

[0003] Accurate and long-term positioning is one of the key technologies for autonomous driving. The advent of GPS has not completely solved the need for precise positioning; more stable and accurate positioning over longer periods is still required. On the other hand, although odometry based on various sensors can provide relatively accurate pose estimations, they are inevitably affected by cumulative drift and require global information for constraint. Therefore, vehicle positioning technology typically relies on Global Navigation Satellite Systems (GNSS) to estimate absolute pose. However, in complex environments, commercial GPS signals are often unusable or have low accuracy due to obstruction or interference. Summary of the Invention

[0004] To address the aforementioned issues, this invention designs a millimeter-wave radar matching network based on synchronous mapping and positioning techniques and deep learning methods, and constructs a geolocation system based on this network. This solves the problem that existing positioning schemes cannot accurately locate under GPS failure conditions, and improves robustness under various weather conditions. After the system is initialized, it can maintain high positioning accuracy for a long time without relying on GPS signals.

[0005] To achieve the above objectives, the present invention provides a teaching-repeated localization method based on a millimeter-wave radar matching network, comprising the following steps:

[0006] S1. Collect millimeter-wave signal image data with satellite positioning data within the target path using millimeter-wave radar, and construct a teaching sequence;

[0007] S2. Construct a millimeter-wave radar matching network;

[0008] S3. For the first frame of millimeter-wave radar data input in the repeating sequence, the millimeter-wave radar matching network is used to retrieve the teaching frame with the highest matching degree within the target path, and the relative pose transformation between the input frame and the teaching frame is calculated according to the matching relationship to obtain the starting point initialization positioning.

[0009] S4. For subsequent radar millimeter-wave data input by the repeating sequence, the millimeter-wave matching network is used as an odometry to obtain feature point matching pairs between consecutive frames of the repeating sequence, and the pose transformation between frames is calculated to obtain the relative pose between frames and the estimated prior global position.

[0010] S5. Based on the prior global position, find the teaching frames within the specified range, use the millimeter-wave radar matching network for positioning and registration, obtain the feature point matching pairs between the teaching frames within the range and the current input frame, and calculate the pose transformation between frames to obtain the relative pose between frames.

[0011] S6. Combining the odometry estimation results from step S4 with the positioning and registration results from step S5, the pose graph is optimized to obtain the optimized pose transformation results.

[0012] Preferably, step S1 specifically includes the following steps:

[0013] S11. Use a vehicle equipped with millimeter-wave radar and satellite positioning system to traverse the target area, collect millimeter-wave images and satellite positioning data within the area, and record timestamps.

[0014] S12. Use the nearest timestamp principle to retrieve the satellite positioning data corresponding to each frame of millimeter wave, and obtain the world coordinate system of the data acquisition location corresponding to all feasible areas in the teaching sequence.

[0015] Preferably, step S2 specifically includes the following steps:

[0016] S21. Use a local feature (CNN) convolutional neural network with feature pyramid network (FPN) to downsample the two input frames of images. After convolution and pooling operations, obtain the coarse-grained 1 / 8 feature map and the fine-grained 1 / 2 feature map corresponding to the two input frames.

[0017] S22. Flatten the coarse-grained 1 / 8 feature map into a one-dimensional vector and add position encoding, input it into the local feature transformer module, and output the reconstructed features.

[0018] S23. Generate a confidence matrix of features based on the confidence threshold and the nearest neighbor criterion, and obtain coarse-grained matching results based on the confidence scores.

[0019] S24. For each selected coarse-grained match, a local window is extracted from the fine-grained feature map, and the match is refined to the sub-pixel level within this window as the final matching result.

[0020] S25. Train the network structure.

[0021] Preferably, the local feature transformer module has multiple self-attention layers and cross-attention layers.

[0022] Preferably, step S3 specifically includes the following steps:

[0023] S31. Traverse all teaching frames, use the millimeter-wave radar matching network to perform matching, and select the teaching frame with the highest matching degree as the starting frame based on the output confidence.

[0024] S32. Filter out valid matching point pairs based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit.

[0025] S33. Use RANSAC (Random Consensus Sampling) to calculate the relative rotation and translation relationships between the matching points in the remaining point pairs;

[0026] S34. Obtain the initial positioning of the starting point based on the rotation and translation relationship of the positioning step S33 in the teaching frame.

[0027] Preferably, step S33 specifically involves: first, randomly selecting a set of minimum samples from the remaining valid matching point pairs; calculating the relative rotation and translation transformation based on the initial sample set using the least squares method; for the remaining matching point pairs, calculating the error between them and the estimated transformation using Euclidean distance; determining the set of interior points consistent with the estimated transformation according to a pre-set threshold; if the current number of interior points is greater than the previous maximum number of interior points, updating the maximum number of interior points and saving the corresponding estimated transformation; if not, repeating the above steps until the stopping condition is met or the upper limit of the number of iterations is reached; finally, using the final set of interior points, re-estimating the rotation and translation transformation.

[0028] Preferably, step S4 specifically includes the following steps:

[0029] S41. Input the current frame and the previous frame millimeter-wave image in the repeating sequence, and use the millimeter-wave radar matching network to perform matching.

[0030] S42. Filter out valid matching point pairs based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit.

[0031] S43. Use RANSAC (Random Consistent Sampling) to calculate the relative rotation and translation relationships between the matching points in the remaining point pairs;

[0032] S44. Based on the positioning of the previous frame and the relative rotation and translation relationship of step S43, obtain the prior positioning estimate of the starting point in the current frame.

[0033] Preferably, step S5 specifically includes the following steps:

[0034] S51. Based on the prior localization estimate of the current frame, given the retrieval range, obtain all teaching frames within the range and their global localization in the teaching sequence;

[0035] S52. Input the millimeter-wave images of the current frame and the teaching frame in the repeating sequence, and perform matching using the millimeter-wave radar matching network;

[0036] S53. Filter out valid matching point pairs based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit.

[0037] S54. Use RANSAC (Random Consensus Sampling) to calculate the relative rotation and translation relationships between the matching points in the remaining point pairs;

[0038] S55. Obtain the relative positioning relationship between the current frame and all teaching frames.

[0039] Preferably, step S6 specifically includes the following steps:

[0040] S61. Using the prior localization estimation of the odometry in step S4 and the relative localization relationship between the current frame and the teaching frame in S5, construct a pose graph optimization problem, and use the relative pose of the consecutive frames of the repeating sequence and the relative pose of the teaching-repeating sequence to calculate the edge error between vertices in the pose graph.

[0041] S62. The pose graph is optimized by using prior edges and integrating the results of posterior teaching-repeated localization.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This invention provides a teach-repeat localization method based on millimeter-wave radar matching networks, which combines millimeter-wave radar and deep learning technology, enabling long-term stable and high-precision localization in all weather conditions and at all times without the need for GPS signals. Attached Figure Description

[0044] Figure 1 This is a structural diagram of the method of the present invention;

[0045] Figure 2 This is a flowchart of the present invention;

[0046] Figure 3 This is a schematic diagram of the millimeter-wave image input for this invention;

[0047] Figure 4 This is a network structure diagram of the present invention;

[0048] Figure 5 This is a matching effect diagram output by the matching network of the present invention;

[0049] Figure 6 This is a schematic diagram of the pose graph structure of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] This invention aims to address the problem in existing technologies where commercial GPS signals are often unusable or have low accuracy in complex environments due to obstruction or interference. A teaching-repeating localization method based on prior maps provides a feasible solution, overcoming the limitations of GPS signals and offering higher-precision six-DOF attitude estimation. For long-term positioning needs in specific environments or along specific routes, such as indoors or on repetitive transportation routes, the teaching-repeating method can be employed. This method first uses a millimeter-wave radar with global positioning data to acquire signal image data within the target path to construct a teaching sequence. Then, relative pose transformation is calculated using the registration results of the repeating sequence and the teaching sequence. Once the teaching sequence is established, subsequent entry into the area does not require relying on GPS equipment to obtain global positioning data; accurate positioning can be achieved simply by utilizing the relative relationship between the current data and prior data. This method can achieve stable, all-weather, high-precision positioning over long periods without GPS signals.

[0052] The present invention proposes a teach-repeated localization method based on millimeter-wave radar matching networks, comprising the following steps:

[0053] S1. Collect millimeter-wave signal image data with satellite positioning data within the target path using millimeter-wave radar, and construct a teaching sequence;

[0054] S2. Construct a millimeter-wave radar matching network;

[0055] S3. For the first frame of millimeter-wave radar data input in the repeating sequence, the millimeter-wave radar matching network is used to retrieve the teaching frame with the highest matching degree within the target path, and the relative pose transformation between the input frame and the teaching frame is calculated according to the matching relationship to obtain the starting point initialization positioning.

[0056] S4. For subsequent radar millimeter-wave data input by the repeating sequence, the millimeter-wave matching network is used as an odometry to obtain feature point matching pairs between consecutive frames of the repeating sequence, and the pose transformation between frames is calculated to obtain the relative pose between frames and the estimated prior global position.

[0057] S5. Based on the prior global position, find the teaching frames within the specified range, use the millimeter-wave radar matching network for positioning and registration, obtain the feature point matching pairs between the teaching frames within the range and the current input frame, and calculate the pose transformation between frames to obtain the relative pose between frames.

[0058] S6. Combining the odometry estimation results from step S4 with the positioning and registration results from step S5, the pose graph is optimized to obtain the optimized pose transformation results.

[0059] Millimeter-wave radar is an emerging sensor in the field of SLAM (Simultaneous Localization and Mapping). It can provide high-precision distance and velocity information. Compared to other sensors such as cameras or lidar, millimeter-wave radar has better penetration and detection capabilities for targets in the environment, maintaining stable performance even in harsh weather conditions such as rain, snow, and fog. This makes millimeter-wave radar a powerful tool for acquiring data within a target path and constructing teaching sequences. Furthermore, the operating principle of millimeter-wave radar gives it excellent spatiotemporal consistency. It can provide a continuous data stream with high temporal stability and consistency. This makes millimeter-wave radar highly reliable for use as odometry on subsequent data from repeated input sequences, enabling feature point matching and pose transformation calculation between consecutive frames, further improving the accuracy and stability of localization. Therefore, in teach-and-repeat localization tasks, millimeter-wave radar has unique advantages over existing pre-built map solutions such as lidar.

[0060] The following is a detailed explanation of each step.

[0061] Step S1: Collect millimeter-wave signal image data with satellite positioning data within the target path using millimeter-wave radar, and construct a teaching sequence.

[0062] Specifically, it mainly includes the following steps:

[0063] S11. Using a vehicle equipped with millimeter-wave radar and a satellite positioning system, traverse the target area, such as a cyclical work route or a limited indoor scene, to collect millimeter-wave images and satellite positioning data within feasible areas that may be repeatedly visited. The millimeter-wave radar data consists of single-channel grayscale PNG images with equal height and width, while the satellite positioning data includes latitude and longitude estimates with error information and UTM coordinates. Collect millimeter-wave images and satellite positioning data within the area and record timestamps.

[0064] S12. Use the nearest timestamp principle to retrieve the satellite positioning data corresponding to each frame of millimeter wave, and obtain the world coordinate system of the data acquisition location corresponding to all feasible areas in the teaching sequence.

[0065] Step S2: Construct a millimeter-wave radar matching network.

[0066] Specifically, it mainly includes the following steps:

[0067] S21, The two received millimeter-wave radar images are denoted as I. A I B A local feature convolutional neural network (CNN) with a feature pyramid network (FPN) is used to downsample a pair of input images. After convolution and pooling operations, coarse-grained 1 / 8 feature maps corresponding to the two input images are obtained. and 1 / 2 fine-grained feature map

[0068] S22. Flatten the coarse-grained feature map into a one-dimensional vector and add positional encoding so that the flattened encoding has global positional information. Input the vector into the local feature transformer module, which has N... c Each layer employs self-attention and cross-attention layers; these attention mechanisms effectively capture key features in an image and reconstruct those features. The final output is the reconstructed feature. In this approach, the self-attention layer receives features from the same feature map, while the cross-attention layer receives inputs from two different feature maps. The attention layer input consists of an index vector Q, a key vector K, and a value vector V. Q obtains information from V, and the attention weights corresponding to V are calculated based on the dot product of Q and K. Intuitively, the attention operation selects relevant information by measuring the similarity between indices and keys. The output is a sum of value vectors weighted by similarity scores. As a result, if the similarity is high, relevant information is extracted from the value vectors.

[0069] S23. Generate the confidence matrix of the features based on the confidence threshold and the nearest neighbor criterion:

[0070]

[0071] P c (i,j)=softmax(S(i,·)) j ·softmax(S(·,j)) i

[0072] By setting an appropriate confidence threshold, based on the confidence matrix P c Select those with a confidence level higher than the threshold θ c The matching is performed, and the nearest neighbor (MNN) criterion is further enforced to filter out possible outliers in the coarse matching. Finally, the candidate results M of the coarse-grained matching are obtained. c .

[0073] S24. For each selected coarse-grained match, extract a w×w local window from the fine-grained feature map, refine it to the sub-pixel level within this window, and use this as the final matching result. For each coarse match... The present invention first positions its location Positioning in fine feature mapping and Then, crop two sets of local windows of size w×w. Use a local feature transformer module to transform the features within each window by N. f This generates two transformed local feature maps. and respectively and Centered on. Then... The center vector and All vectors in the dataset are correlated to generate a heatmap, representing... Each pixel in the neighborhood and The matching probability is obtained by calculating the expected value of the probability distribution in I. B Final position with sub-pixel precision Collect all matches To produce the final fine-grained match M f .

[0074] S25. Train the above network structure using supervised learning with different millimeter-wave radar images having known translations and rotations. Set up two stages of supervision: coarse-grained and fine-grained. The loss function for coarse-grained learning is the confidence matrix P returned by the optimal transport layer or the double softmax operator. c Negative log-likelihood loss;

[0075]

[0076] The fine-grained loss function is L2 loss, for each query point. This invention calculates the total variance of the corresponding heatmap. To measure its uncertainty;

[0077]

[0078] The final loss consists of coarse-level and fine-level losses: L = L c +L f .

[0079] Step S3: For the first frame of millimeter-wave radar data input in the repeating sequence, the millimeter-wave radar matching network is used to retrieve the teaching frame with the highest matching degree within the target path, and the relative pose transformation between the input frame and the teaching frame is calculated according to the matching relationship to obtain the initial positioning of the starting point.

[0080] Specifically, it mainly includes the following steps:

[0081] S31. Traverse all teaching frames and use the millimeter-wave radar matching network constructed in step S2 to perform matching. Match the first frame of the repeating sequence with the teaching sequence frame. The matching network outputs the matched feature point pairs and their confidence scores. By accumulating the confidence scores of all point pairs, select the teaching frame with the highest matching score.

[0082] S32. Valid matching point pairs are filtered out based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit. By filtering the matching points, invalid matches in the no-signal part and erroneous matches with too large differences in millimeter wave echoes can be removed, thereby improving the matching effect.

[0083] S33. Calculate the relative rotation and translation relationship between matching points using RANSAC (Random Consistent Sampling) on ​​the remaining point pairs. First, randomly select a minimum sample from the remaining valid matching point pairs. Calculate the relative rotation and translation transformation based on the initial sample set using the least squares method. For the remaining matching point pairs, calculate the error between them and the estimated transformation using Euclidean distance. Determine the set of inliers consistent with the estimated transformation according to a pre-set threshold. If the current number of inliers is greater than the previous maximum number of inliers, update the maximum number of inliers and save the corresponding estimated transformation. If not, repeat the above steps until the stopping condition is met or the maximum number of iterations is reached. Finally, re-estimate the rotation and translation transformation using the final set of inliers.

[0084] S34. Obtain the initial positioning of the starting point based on the rotation and translation relationship of the positioning step S33 in the teaching frame.

[0085] Step S4: For subsequent radar millimeter-wave data input by the repeating sequence, the millimeter-wave matching network is used as an odometry to obtain feature point matching pairs between consecutive frames of the repeating sequence, and the pose transformation between frames is calculated to obtain the relative pose between frames and the estimated prior global position.

[0086] Specifically, it mainly includes the following steps:

[0087] S41. Input the current frame and the previous frame millimeter-wave image in the repeating sequence, and use the millimeter-wave radar matching network constructed in step S2 to perform matching;

[0088] S42. Filter out valid matching point pairs based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit.

[0089] S43. Use RANSAC (Random Consistent Sampling) to calculate the relative rotation and translation relationships between the matching points in the remaining point pairs;

[0090] S44. Based on the positioning of the previous frame and the relative rotation and translation relationship of step S43, obtain the prior positioning estimate of the starting point in the current frame.

[0091] Step S5: Based on the prior global position, find the teaching frames within the specified range, use the millimeter-wave radar matching network for positioning and registration, obtain the feature point matching pairs between the teaching frames within the range and the current input frame, and calculate the pose transformation between frames to obtain the relative pose between frames.

[0092] Specifically, it mainly includes the following steps:

[0093] S51. Based on the prior localization estimate of the current frame, given the retrieval range, obtain all teaching frames within the range and their global localization in the teaching sequence;

[0094] S52. Input the millimeter-wave images of the current frame and the teaching frame in the repeating sequence, and perform matching using the millimeter-wave radar matching network constructed in step S2;

[0095] S53. Filter out valid matching point pairs based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit.

[0096] S54. Use RANSAC (Random Consistent Sampling) to calculate the relative rotation and translation relationships between the matching points in the remaining point pairs;

[0097] S55. Obtain the relative positioning relationship between the current frame and all teaching frames.

[0098] Step S6: Combine the odometry estimation results from step S4 with the positioning and registration results from step S5 to optimize the pose graph and obtain the optimized pose transformation results.

[0099] Specifically, it mainly includes the following steps:

[0100] S61. Using the prior localization estimation of the odometry in step S4 and the relative localization relationship between the current frame and the teaching frame in S5, construct a pose graph optimization problem, and use the relative pose of the consecutive frames of the repeating sequence and the relative pose of the teaching-repeating sequence to calculate the edge error between vertices in the pose graph.

[0101] S62. The pose graph is optimized by using prior edges and integrating the results of posterior teaching-repeated localization.

[0102] This invention uses a pose graph optimization method to fuse odometry results and localization results based on teach-and-repeat. The odometry output pose is denoted as... The registration result output by the relocation registration is denoted as This invention uses odometer output and relocation registration results The problem of constructing a pose graph optimization problem to obtain more accurate poses. The odometer error edge is defined as follows:

[0103]

[0104] To integrate the global positioning obtained from registration, an additional positioning error edge is added in this invention:

[0105]

[0106] Where t and R are the translation and rotation components of T, and R|z represents the z-axis rotation component of the rotation matrix R, i.e., taking the yaw angle.

[0107] The application of millimeter-wave radar in the teach-repeat localization method fully leverages its advantages such as high precision, strong robustness, and spatiotemporal consistency. It plays a crucial role in addressing the problem of long-term stable high-precision positioning in the absence of GPS and provides a feasible technical solution for achieving all-weather, all-time high-precision positioning. Based on deep learning methods, features that are difficult to extract from millimeter-wave data can be more easily associated and registered, thereby completing the high-precision positioning task.

[0108] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A teach-repeated localization method based on millimeter-wave radar matching networks, characterized in that, Includes the following steps: S1. Collect millimeter-wave signal image data with satellite positioning data within the target path using millimeter-wave radar, and construct a teaching sequence; S2. Construct a millimeter-wave radar matching network; S3. For the first frame of millimeter-wave radar data input in the repeating sequence, the millimeter-wave radar matching network is used to retrieve the teaching frame with the highest matching degree within the target path, and the relative pose transformation between the input frame and the teaching frame is calculated according to the matching relationship to obtain the starting point initialization positioning. S4. For subsequent radar millimeter-wave data input by the repeating sequence, the millimeter-wave radar matching network is used as an odometry to obtain feature point matching pairs between consecutive frames of the repeating sequence, and the pose transformation between frames is calculated to obtain the relative pose between frames and the estimated prior global position. S5. Based on the prior global position, find the teaching frames within the specified range, use the millimeter-wave radar matching network for positioning and registration, obtain the feature point matching pairs between the teaching frames within the range and the current input frame, and calculate the pose transformation between frames to obtain the relative pose between frames. S6. Combining the odometry estimation result from step S4 with the positioning and registration result from step S5, the pose graph is optimized to obtain the optimized pose transformation result. Step S2 specifically includes the following steps: S21. Use a Local Feature Convolutional Neural Network (CNN) with Feature Pyramid Network (FPN) to downsample the two input frames of images. After convolution and pooling operations, obtain the coarse-grained 1 / 8 feature map and the fine-grained 1 / 2 feature map corresponding to the two input frames. S22. Flatten the coarse-grained 1 / 8 feature map into a one-dimensional vector and add position encoding, input it into the local feature transformer module, and output the reconstructed features. S23. Generate a confidence matrix of features based on the confidence threshold and the nearest neighbor criterion, and obtain coarse-grained matching results based on the confidence scores. S24. For each selected coarse-grained match, a local window is extracted from the fine-grained feature map, and the match is refined to the sub-pixel level within this window as the final matching result. S25. Train the network structure.

2. The teaching-repeated localization method based on millimeter-wave radar matching networks according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Use a vehicle equipped with millimeter-wave radar and satellite positioning system to traverse the target area, collect millimeter-wave images and satellite positioning data within the area, and record timestamps. S12. Use the nearest timestamp principle to retrieve the satellite positioning data corresponding to each frame of millimeter wave, and obtain the world coordinate system of the data acquisition location corresponding to all feasible areas in the teaching sequence.

3. The teach-repeated localization method based on millimeter-wave radar matching network according to claim 1, characterized in that, The local feature transformer module has multiple self-attention layers and cross-attention layers.

4. The teach-repeat localization method based on millimeter-wave radar matching network according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Traverse all teaching frames, use the millimeter-wave radar matching network to perform matching, and select the teaching frame with the highest matching degree as the starting frame based on the output confidence. S32. Filter out valid matching point pairs based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit. S33. Use RANSAC (Random Consistent Sampling) to calculate the relative rotation and translation relationship between the matching points in the remaining point pairs; S34. Obtain the initial positioning of the starting point based on the rotation and translation relationship of the positioning step S33 in the teaching frame.

5. The teach-repeated localization method based on a millimeter-wave radar matching network according to claim 4, characterized in that, Step S33 specifically involves: First, randomly selecting a set of minimum samples from the remaining valid matching point pairs; calculating the relative rotation and translation transformation based on the initial sample set using the least squares method; for the remaining matching point pairs, calculating the error between them and the estimated transformation using Euclidean distance; determining the set of interior points consistent with the estimated transformation according to a pre-set threshold; if the current number of interior points is greater than the previous maximum number of interior points, updating the maximum number of interior points and saving the corresponding estimated transformation; if not, repeating the above steps until the stopping condition is met or the upper limit of the number of iterations is reached; finally, using the final set of interior points, re-estimating the rotation and translation transformation.

6. The teach-repeat localization method based on millimeter-wave radar matching network according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Input the current frame and the previous frame millimeter-wave image in the repeating sequence, and use the millimeter-wave radar matching network to perform matching. S42. Filter out valid matching point pairs based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit. S43. Use RANSAC (Random Consistent Sampling) to calculate the relative rotation and translation relationships between the matching points in the remaining point pairs; S44. Based on the positioning of the previous frame and the relative rotation and translation relationship of step S43, obtain the prior positioning estimate of the starting point in the current frame.

7. The teach-repeat localization method based on millimeter-wave radar matching network according to claim 1, characterized in that, Step S5 specifically includes the following steps: S51. Based on the prior localization estimate of the current frame, given the retrieval range, obtain all teaching frames within the range and their global localization in the teaching sequence; S52. Input the millimeter-wave images of the current frame and the teaching frame in the repeating sequence, and perform matching using the millimeter-wave radar matching network; S53. Filter out valid matching point pairs based on whether the pixels corresponding to the matching points on the radar image all exceed the lower intensity limit and whether the difference between the two pixels is less than the upper difference limit. S54. Use RANSAC (Random Consistent Sampling) to calculate the relative rotation and translation relationship between the matching points in the remaining point pairs; S55. Obtain the relative positioning relationship between the current frame and all teaching frames.

8. The teach-repeated localization method based on millimeter-wave radar matching network according to claim 1, characterized in that, Step S6 specifically includes the following steps: S61. Using the prior localization estimation of the odometry in step S4 and the relative localization relationship between the current frame and the teaching frame in S5, construct a pose graph optimization problem, and use the relative pose of the consecutive frames of the repeating sequence and the relative pose of the teaching-repeating sequence to calculate the edge error between vertices in the pose graph. S62. The pose graph is optimized by using prior edges and integrating the results of posterior teaching-repeated localization.

Images