Reproducible walking positioning method based on deep learning

Abstract

The application discloses a kind of repeatable walking positioning method based on deep learning, to solve the problem that it is difficult to realize centimeter-level positioning and the same operation line repeatable walking under global satellite navigation system signal shielding or multipath environment, the present application is by: based on reference operation data Construction reference trajectory and generate and pose range associated reference sub-library;Satellite observation is adopted to the current multipath suppression deep learning model output and satellite, frequency point and observation type are used as the granularity of observation credible parameter and observation bias to carry out observation correction;In the tight coupling sliding window factor graph, switch variable is introduced for each satellite observation factor, and measurement covariance is adaptively updated according to observation credible parameter to weaken or close suspected multipath observation;Global descriptor is extracted to the current point cloud, and after candidate sub-graph retrieval in reference sub-library, fine registration is carried out to generate point cloud positioning constraint and covariance;Reference trajectory fitting factor is introduced to constrain lateral error and is optimized with point cloud positioning factor, which realizes the technical effects of robust positioning in shielding section, suppressing multipath error and improving lateral accuracy of repeatable walking.

Description

Technical Field

[0001] This invention relates to the field of high-precision positioning and navigation, and in particular to a repeatable walking positioning method based on deep learning. Background Technology

[0002] With the development of applications such as intelligent agricultural machinery, unmanned delivery vehicles, and park inspection robots, the demand for high-precision positioning of operating equipment in complex environments such as roadside obstructions, tree-lined passages, and urban canyons is becoming increasingly prominent. In particular, in scenarios where repeated operations are required multiple times along the same row or path in the same work area, centimeter-level positioning accuracy and small lateral repeatability error are typically required to be maintained even under long-term operation and environmental interference.

[0003] In existing technologies, high-precision positioning techniques of the Global Navigation Satellite System (GNSS), such as real-time dynamic positioning and precise point positioning, can achieve high accuracy in open environments. However, they are prone to observation anomalies, cycle slips, and decreased fixation rates under signal obstruction and multipath conditions. To improve continuity, engineering practices often involve loosely or tightly coupled fusion of the GNSS and inertial measurement units (IMUs), and employ frameworks such as extended Kalman filtering, sliding window optimization, and factor graphs to enhance robustness. Meanwhile, lidar point cloud localization and mapping technologies are developing rapidly. Subgraph-based map matching and loop closure detection can provide compensation when the GNSS degrades. In recent years, methods utilizing deep learning for satellite observation quality assessment and point cloud global descriptor retrieval have also emerged to enhance usability in complex environments.

[0004] However, the aforementioned existing technologies still have shortcomings:

[0005] 1. For multipath and non-line-of-sight suppression, the availability judgment is mostly at the epoch level or receiver level, or only coarse-grained weighting of observations is applied. It is difficult to refine the results to specific satellites, specific frequencies and specific observation types, and to stably inject the learning results into tightly coupled optimization in the form of bias correction or covariance.

[0006] 2. In tightly coupled sliding window factor plots, when faced with local multipath or occasional abnormal observations, there is a lack of a mechanism to switch and weaken or turn off individual satellite observation factors, which can easily lead to estimation divergence, drift accumulation or positioning jumps.

[0007] 3. Existing point cloud localization or synchronous localization and mapping methods mostly aim for globally consistent localization, and usually do not explicitly model the lateral repetition error relative to the reference trajectory as an optimization constraint; at the same time, the gating and confidence back-injection of mismatches are insufficient in the subgraph retrieval and registration process, which easily leads to repeated walking failures caused by cross-line matching.

[0008] Therefore, a repeatable walking positioning method is needed to address the shortcomings of the existing technology. Summary of the Invention

[0009] One objective of this invention is to propose a deep learning-based repeatable walking localization method. Addressing the challenges of existing technologies such as difficulty in finely suppressing observation anomalies under conditions of global satellite navigation system signal obstruction and multipath propagation, lack of switchable weakening mechanisms for anomaly observations in tightly coupled optimization, and the inability to guarantee lateral repeatability accuracy for the same path due to erroneous point cloud localization matching, the following technical solution is proposed: A reference trajectory is generated based on reference satellite observations, reference inertia, and a reference point cloud, and a reference sub-map library with pose range is established; a multipath suppression deep learning model is used to output observation reliability parameters and observation biases at the granularity of satellite, frequency, and observation type for the current satellite observations, and observation corrections are performed; switch variables are introduced for each satellite observation factor in the tightly coupled sliding window factor graph, and the measurement covariance is adaptively updated based on the reliability parameters; a global descriptor is extracted from the current point cloud, and candidate sub-maps are retrieved and finely registered in the reference sub-map library to generate point cloud localization constraints and covariance; a reference trajectory fitting factor is introduced to constrain lateral errors and jointly optimize the output pose. This invention has the technical effects of improving localization robustness and centimeter-level accuracy in complex environments, reducing the impact of anomaly observations, and significantly improving lateral consistency of repeated walking.

[0010] This invention provides a repeatable walking localization method based on deep learning, comprising:

[0011] S1. Acquire reference sensor data, including reference satellite observation data, reference inertial data, and reference point cloud data. Optimize the reference satellite observation data and reference inertial data in a tightly coupled sliding window factor graph to obtain a reference pose sequence, generate a reference trajectory, and generate a set of reference subgraphs based on the reference point cloud data, establishing a reference subgraph library containing the reference subgraphs and the reference pose range associated with each reference subgraph. S2. Acquire current sensor data, including current satellite observation data, current inertial data, and current point cloud data. S3. Input the current satellite observation data into a multipath suppression deep learning model to obtain a set of observation confidence parameters and a set of observation biases. Correct the current satellite observation data based on the set of observation biases to obtain corrected satellite observation data. S4. Obtain inertial pre-integration constraints based on the current inertial data, construct a tightly coupled sliding window factor graph, and introduce the corrected satellite observation data as satellite observation factors into the tightly coupled sliding window factor graph. Introduce corresponding switch variables for each satellite observation factor, causing the constraint effect to weaken or become ineffective as the switch variables decrease. Update the measurement covariance matrix of the satellite observation factors based on the set of observation confidence parameters. And set prior constraints for the switch variables, construct the inertial pre-integration constraint as an inertial factor, introduce it into the tightly coupled sliding window factor graph and optimize it to obtain the predicted pose sequence and the optimized state; S5, input the current point cloud data into the point cloud feature deep learning model to obtain the current point cloud global descriptor, filter candidate reference subgraphs from the reference subgraph library and search based on the current point cloud global descriptor, register the current point cloud data with the retrieved candidate reference subgraphs to determine the target point cloud registration result, and generate point cloud localization constraints and point cloud localization covariance matrix; S6, in the optimized state, with the current The pose variable corresponding to the epoch is the pose to be estimated. The pose variable is projected onto the reference trajectory to obtain the projection point. The reference trajectory fitting factor is constructed based on the lateral error between the pose variable and the projection point. The point cloud localization constraint is constructed as the point cloud localization factor and its measurement covariance matrix is ​​set according to the point cloud localization covariance matrix. The reference trajectory fitting factor and the point cloud localization factor are introduced into the tightly coupled sliding window factor graph and optimized to obtain the optimized pose of the current epoch. S7. The localization result of the current operation is output based on the optimized pose of the current epoch, including the current position and the lateral error corresponding to the reference trajectory.

[0012] Optionally, S1 includes:

[0013] Acquire reference satellite observation data, reference inertial data, and reference point cloud data during the reference operation, and synchronize the reference satellite observation data, reference inertial data, and reference point cloud data in time to obtain synchronized reference data;

[0014] Based on the synchronous reference data, a tightly coupled sliding window factor map is constructed. The reference satellite observation data is constructed as a satellite observation factor and introduced into the tightly coupled sliding window factor map. The reference inertial data is constructed as an inertial factor and introduced into the tightly coupled sliding window factor map. The tightly coupled sliding window factor map is optimized to obtain a reference pose sequence.

[0015] Based on the reference pose sequence, reference keyframes are selected in the reference point cloud data, and a set of reference sub-graphs is generated based on the reference keyframes. When generating each reference sub-graph, the reference point cloud corresponding to the reference keyframe is fused with the reference point clouds corresponding to other reference keyframes that meet the preset neighborhood conditions to obtain the reference sub-graph corresponding to the reference keyframe.

[0016] Generate a reference trajectory based on the reference pose sequence;

[0017] For each reference subgraph in the set of reference subgraphs, a reference pose range is determined, wherein the reference pose range includes the start pose index and the end pose index in the reference pose sequence;

[0018] The reference subgraph set and the reference pose range corresponding to each reference subgraph are written into the reference subgraph library to obtain the reference subgraph library.

[0019] Optionally, S2 includes:

[0020] Acquire the current satellite observation data, current inertial data, and current point cloud data during the current operation, and synchronize the current satellite observation data, current inertial data, and current point cloud data in time to obtain synchronized current data;

[0021] The current satellite observation data in the synchronized current data is organized into a satellite observation set according to epochs, wherein each satellite observation includes a satellite identifier, a frequency identifier, an observation type identifier, and an observation value;

[0022] Output the current satellite observation data, the current inertial data, and the current point cloud data.

[0023] Optionally, S3 includes:

[0024] The current satellite observation data is input into the trained multipath suppression deep learning model to obtain an observation confidence parameter set and an observation bias set. The observation confidence parameter set refers to the observation weight and noise scaling coefficient corresponding to each satellite observation output in the current satellite observation data, and the observation bias set refers to the observation bias corresponding to each satellite observation output in the current satellite observation data.

[0025] The observation bias set is matched one-to-one with each satellite observation in the current satellite observation data, and the observation values ​​of each satellite observation are corrected to obtain corrected satellite observation data.

[0026] Optionally, S4 includes:

[0027] Based on the current inertial data, inertial pre-integration processing is performed to obtain inertial pre-integration constraints; a tightly coupled sliding window factor map is constructed using the pose variables and inertial bias variables within the sliding window as variables to be estimated; each satellite observation in the corrected satellite observation data is constructed as a satellite observation factor and introduced into the tightly coupled sliding window factor map, wherein a switch variable corresponding to the satellite observation is introduced for each satellite observation factor, and the switch variable is a continuous variable with a value range between zero and one;

[0028] The residual term of the satellite observation factor is scaled according to the switch variable, so that when the switch variable approaches zero, the constraint effect of the satellite observation factor on optimization is weakened or invalidated. Based on the observation weights and noise scaling coefficients corresponding to the satellite observations in the set of observation confidence parameters, the measurement covariance matrix of the satellite observation factor is updated, and a priori constraint is set for the switch variable to approach one. The inertial pre-integration constraint is constructed as an inertial factor and introduced into the tightly coupled sliding window factor graph. The tightly coupled sliding window factor graph is optimized to obtain the predicted pose sequence within the sliding window, and the optimized state of the predicted pose sequence and the tightly coupled sliding window factor graph is output.

[0029] Optionally, S5 includes:

[0030] The current point cloud data is input into a trained point cloud feature deep learning model to obtain a global descriptor for the current point cloud. Based on the current epoch predicted pose in the predicted pose sequence, a corresponding reference pose index is determined on the reference trajectory. Reference sub-graphs whose reference pose range contains the reference pose index are selected from the reference sub-graph library to obtain a retrieval subset. Similarity retrieval is performed on the current point cloud global descriptor in the retrieval subset to obtain a set of candidate reference sub-graphs with a preset number of K. Using the current epoch predicted pose in the predicted pose sequence as the initial value, the current point cloud data is matched with each candidate reference sub-graph in the candidate reference sub-graph set using deep feature-guided fine point cloud registration to obtain a set of point cloud registration results. Registration residuals and matching confidence are calculated for each point cloud registration result. The target point cloud registration result is determined based on the set of point cloud registration results, and point cloud localization constraints and point cloud localization covariance matrices are generated based on the target point cloud registration result.

[0031] Furthermore, determining the target point cloud registration result includes: under the condition that the matching confidence of each point cloud registration result is greater than a preset confidence threshold, selecting the point cloud registration result with the smallest registration residual as the target point cloud registration result;

[0032] Furthermore, the reference pose range includes a start pose index and an end pose index, and the filtering of reference subgraphs includes: retaining reference subgraphs that satisfy the condition that the start pose index is less than or equal to the reference pose index and the reference pose index is less than or equal to the end pose index, as the retrieval subset.

[0033] Optionally, S6 includes:

[0034] Using the pose variable corresponding to the current epoch in the tightly coupled sliding window factor graph as the pose to be estimated, the position corresponding to the pose variable is projected onto the reference trajectory to obtain the projection point, and the lateral error is calculated based on the pose variable and the projection point to construct the reference trajectory fitting factor; the point cloud localization constraint is constructed as the point cloud localization factor, and the measurement covariance matrix of the point cloud localization factor is set according to the point cloud localization covariance matrix; the reference trajectory fitting factor and the point cloud localization factor are introduced into the optimization state of the tightly coupled sliding window factor graph, and the updated tightly coupled sliding window factor graph is optimized to obtain the optimized pose of the current epoch.

[0035] Optionally, the S7 includes:

[0036] Based on the current epoch, the pose is optimized and the sliding window is updated so that the variables to be estimated in the tightly coupled sliding window factor graph include pose variables and inertial bias variables within a preset time length.

[0037] Based on the optimized state of the tightly coupled sliding window factor map corresponding to the updated sliding window, the positioning result of the current operation is output, including the current position and the lateral error corresponding to the reference trajectory.

[0038] Optionally, the training methods for the multipath suppression deep learning model and the point cloud feature deep learning model are as follows:

[0039] Acquire training sensor data, which includes training satellite observation data output by the global navigation satellite system receiver, training inertial data output by the inertial measurement unit, and training point cloud data output by the lidar. Synchronize the training sensor data in time to obtain synchronized training data.

[0040] Based on the training inertial data in the synchronous training data, inertial pre-integration is performed, and combined with the high-precision carrier phase solution results of the global satellite navigation system in an open environment, the synchronous training data is fused and solved to obtain the reference pose sequence.

[0041] Based on the reference pose sequence and combined with the satellite ephemeris, the theoretical satellite observation values ​​for each epoch are calculated; the theoretical satellite observation values ​​are then differentially analyzed with the training satellite observation data in the synchronous training data to obtain the observation residual for each satellite observation.

[0042] A multipath suppression training sample set is constructed with satellite observations as the granularity. Each multipath suppression training sample includes at least a satellite identifier, frequency identifier, observation type identifier, observation value, and observation residual corresponding to the satellite observation. It further includes observation quality features and motion state features extracted from the synchronous training data. The multipath suppression deep learning model is trained with the multipath suppression training sample set as input and the observation residual as supervision information. The multipath suppression deep learning model outputs observation weights, noise scaling coefficients, and observation biases for each satellite observation.

[0043] Based on the reference pose sequence, keyframe selection and subgraph generation are performed on the training point cloud data in the synchronous training data to obtain a training subgraph set. In the training subgraph set, positive sample pairs are constructed according to the condition that the relative poses between keyframe poses meet a preset threshold, and negative sample pairs are constructed according to the condition that the relative poses between keyframe poses do not meet the preset threshold, to obtain a point cloud retrieval training sample set.

[0044] Using the point cloud retrieval training sample set as input, a point cloud feature deep learning model is trained using metric learning loss. This model outputs a point cloud global descriptor for the input point cloud, and ensures that the similarity of the point cloud global descriptor to positive sample pairs is greater than its similarity to negative sample pairs.

[0045] The beneficial effects of this invention are:

[0046] 1. By using a multipath suppression deep learning model to output observation weights, noise scaling factors, and observation biases at the granularity of satellite, frequency, and observation type, and performing bias correction and covariance adaptive update on satellite observations, suspected multipath and non-line-of-sight observations are finely weakened, thereby reducing the impact of observation anomalies such as pseudorange and carrier phase on tightly coupled solution, and improving the fixation rate and centimeter-level positioning stability in occluded scenarios.

[0047] 2. In the tightly coupled sliding window factor graph, introduce corresponding switch variables for each satellite observation factor and set prior constraints so that they can be automatically weakened or invalidated when the observation is abnormal, thereby avoiding optimization divergence and trajectory jump caused by a sudden change in a single satellite or a single type of observation, and improving the robustness and continuity of the positioning results.

[0048] 3. Based on the pose range filtering and Top-K candidate sub-graph retrieval of the reference sub-graph library, the point cloud localization covariance is output after fine registration for point cloud factor weighting. At the same time, the reference trajectory fitting factor is introduced to constrain the lateral error, thereby reducing the risk of cross-line positioning caused by mismatch and significantly improving the lateral accuracy and consistency of repeated walking in the same work row. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a flowchart of a repeatable walking localization method based on deep learning proposed in this invention. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0052] refer to Figure 1 A deep learning-based repeatable walking localization method includes:

[0053] S1. Acquire reference sensor data, including reference satellite observation data, reference inertial data, and reference point cloud data. Optimize the reference satellite observation data and reference inertial data in a tightly coupled sliding window factor graph to obtain a reference pose sequence, generate a reference trajectory, and generate a set of reference subgraphs based on the reference point cloud data, establishing a reference subgraph library containing the reference subgraphs and the reference pose range associated with each reference subgraph. S2. Acquire current sensor data, including current satellite observation data, current inertial data, and current point cloud data. S3. Input the current satellite observation data into a multipath suppression deep learning model to obtain a set of observation confidence parameters and a set of observation biases. Correct the current satellite observation data based on the set of observation biases to obtain corrected satellite observation data. S4. Obtain inertial pre-integration constraints based on the current inertial data, construct a tightly coupled sliding window factor graph, and introduce the corrected satellite observation data as satellite observation factors into the tightly coupled sliding window factor graph. Introduce corresponding switch variables for each satellite observation factor, causing the constraint effect to weaken or become ineffective as the switch variables decrease. Update the measurement covariance matrix of the satellite observation factors based on the set of observation confidence parameters. And set prior constraints for the switch variables, construct the inertial pre-integration constraint as an inertial factor, introduce it into the tightly coupled sliding window factor graph and optimize it to obtain the predicted pose sequence and the optimized state; S5, input the current point cloud data into the point cloud feature deep learning model to obtain the current point cloud global descriptor, filter candidate reference subgraphs from the reference subgraph library and search based on the current point cloud global descriptor, register the current point cloud data with the retrieved candidate reference subgraphs to determine the target point cloud registration result, and generate point cloud localization constraints and point cloud localization covariance matrix; S6, in the optimized state, with the current The pose variable corresponding to the epoch is the pose to be estimated. The pose variable is projected onto the reference trajectory to obtain the projection point. The reference trajectory fitting factor is constructed based on the lateral error between the pose variable and the projection point. The point cloud localization constraint is constructed as the point cloud localization factor and its measurement covariance matrix is ​​set according to the point cloud localization covariance matrix. The reference trajectory fitting factor and the point cloud localization factor are introduced into the tightly coupled sliding window factor graph and optimized to obtain the optimized pose of the current epoch. S7. The localization result of the current operation is output based on the optimized pose of the current epoch, including the current position and the lateral error corresponding to the reference trajectory.

[0054] In this specific embodiment, S1 includes:

[0055] The reference satellite observation data, reference inertial data, and reference point cloud data are acquired during the reference operation. The three types of data are unified under the same time reference to complete time synchronization. The frame timestamp of the reference point cloud data is used as the epoch time axis. The reference satellite observation data is aligned to the corresponding point cloud frame timestamp according to the most recent epoch and the satellite identifier, frequency identifier, observation type identifier, and observation value are retained. The reference inertial data is divided into two adjacent point cloud frames according to the timestamp and arranged in the original sampling order for subsequent inertial pre-integration.

[0056] A tightly coupled sliding window factor graph is constructed based on synchronous reference data, with the sliding window length set to... The epoch, the first in the window The state to be estimated at each epoch is defined as follows: ,in It includes a reference pose variable and a reference inertia bias variable. The reference pose variable includes the reference position. Reference speed With reference attitude quaternion The reference inertial bias variable includes the reference accelerometer zero bias. Zero bias with reference gyroscope And in the factor graph, adjacent epochs Construct an inertia factor for the epoch. Construct satellite observation factors;

[0057] The optimization objective of the factor graph, expressed in weighted nonlinear least squares form, is:

[0058] ;

[0059] in This indicates the starting epoch index of the current sliding window. This indicates the number of epochs contained in the sliding window. Represents the epoch The state to be estimated This indicates that the reference inertial data is in the epoch. To the epoch The inertial pre-integral residual obtained by integrating between them, This represents the covariance matrix corresponding to the inertial pre-integration residual. Represents the epoch The number of satellite observations involved in tight coupling Represents the epoch The The observation residuals corresponding to each satellite observation, This represents the measurement covariance matrix corresponding to the satellite observation residuals. Describe the Markov norm and satisfy ,in Represents the residual vector. Represents the weight matrix, symbol Indicates transpose;

[0060] The objective function is solved using Gauss-Newton iteration. Each time the sliding window is updated, five iterations are performed, using the optimization result of the previous window as the initial value for the current window. This yields a time-ordered sequence of reference poses, and the reference position for each epoch in the sequence is determined. A reference trajectory is formed by connecting the indices;

[0061] Reference keyframes are selected from the reference point cloud data based on the reference pose sequence. The epoch corresponding to the first point cloud frame is set as the reference keyframe. Subsequently, when the change in the reference position corresponding to an adjacent reference keyframe reaches 1 m or the change in the reference attitude heading angle reaches a certain value, the reference keyframe is selected. When the current point cloud frame corresponds to an epoch, it is determined as a new reference keyframe and its pose index in the reference pose sequence is recorded.

[0062] A set of reference sub-graphs is generated with each reference keyframe as the center. The point cloud corresponding to the reference keyframe and the point clouds corresponding to the five reference keyframes before and after it are fused to form a reference sub-graph. During fusion, the point cloud of each frame is transformed to the coordinate system of the central reference keyframe according to its corresponding reference pose and voxel filtering downsampling is performed. The voxel side length is set to 0.2m to ensure that the point cloud density of the sub-graph is consistent and easy to store.

[0063] For each reference subgraph, a reference pose range is determined. The reference pose range consists of a start pose index and an end pose index. The start pose index is set to 0 when it is less than 5 and the center reference keyframe pose index. The end pose index is set to 5 and the center reference keyframe pose index is added to. When the end pose index exceeds the end index of the reference pose sequence, it is truncated to the end index.

[0064] The point cloud data, center reference keyframe pose index, and corresponding reference pose range of each reference subgraph are written into the reference subgraph library. The reference subgraph library is stored in sequential file format and a unique subgraph identifier is recorded for each reference subgraph to support subsequent filtering by pose range and reading by subgraph identifier.

[0065] In this specific embodiment, S2 includes:

[0066] The system acquires current satellite observation data, current inertial data, and current point cloud data during the current operation period, and unifies the three types of data under the same time reference to achieve time synchronization. Specifically, the point cloud timestamp of each frame of the current point cloud data is defined as an epoch timestamp and used as the synchronization reference. The current satellite observation data is output by the global satellite navigation system receiver and includes the satellite identifier, frequency identifier, observation type identifier, and observation value corresponding to each observation. The current inertial data is output by the inertial measurement unit and includes discrete sampling sequences of angular velocity and specific force.

[0067] The current satellite observation data is assigned to the nearest epoch timestamp according to the timestamp. When the time difference between a satellite observation timestamp and its nearest epoch timestamp is greater than 0.05 s, the satellite observation is discarded to avoid cross-epoch aliasing. The retained satellite observations are organized into a satellite observation set according to epoch and written into a memory queue for subsequent steps to read epoch by epoch.

[0068] The current inertial data is divided into two adjacent epoch timestamps and stored in the original sampling order, so that continuous inertial samples within the time period can be directly read between any adjacent epochs for subsequent inertial pre-integration. If the inertial sampling sequence has reversed or duplicate timestamps, the abnormal samples are directly deleted and the deleted sequence is used as the valid current inertial data.

[0069] The synchronized current satellite observation data will be in epochs. The organization is as follows:

[0070] ;

[0071] in Represents the epoch The collection of satellite observations This indicates the epoch index, which is consistent with the current point cloud frame number. Represents the epoch The internal sequence number, Represents the epoch The number of satellite observations retained within the country, Indicates the first The satellite identifier corresponding to each observation. Indicates the first The frequency point identifier corresponding to each observation. Indicates the first The observation type identifier corresponding to each observation. Indicates the first The observed values ​​corresponding to each observation;

[0072] After completing the above synchronization and organization, the current satellite observation data, the current inertial data, and the current point cloud data are output. The current satellite observation data is output in the form of a set of satellite observations per epoch, the current inertial data is output in the form of an inertial sampling sequence between adjacent epochs, and the current point cloud data is output in the form of point cloud frames per epoch.

[0073] In this specific embodiment, S3 includes:

[0074] For the epoch Satellite Observations Collection Perform multipath suppression processing one by one, among which This indicates the epoch index, which is consistent with the current point cloud frame number. Represents the epoch The internal sequence number, Represents the epoch Number of internal satellite observations Indicates satellite identifier, Indicates frequency point identifier, Indicates the observation type identifier. Represents the observed value;

[0075] Each satellite observation in the current satellite observation data is expanded into an observation record that includes an observation quality field, which includes the carrier-to-noise ratio. Lock duration and satellite elevation angle ,in by The unit is output by the receiver. Output by the receiver in seconds (s). The result, in degrees, is calculated by the receiver based on the ephemeris and the approximate position and is output along with the observation record.

[0076] Numerical features are normalized according to a fixed rule to form numerical sub-vectors of the model input. The normalization rule is to normalize the observed values... Divide by To suppress dimensional differences, the carrier-to-noise ratio Linear normalization is performed using a center value of 40 and a scale of 10 to normalize the elevation angle. Divide by 90 for linear normalization to lock the duration. Divide by 60 to perform linear normalization, and then concatenate the four normalized values ​​in a fixed order;

[0077] satellite identifier Frequency point identification With observation type identifier Each is mapped to an integer index and fed into three sets of learnable embedding tables to obtain three embedding vectors, where the satellite embedding table has the following parameters: And the satellite index range is The frequency point embedding table parameters are And the frequency index range is The observation type embedding table parameter is And the observation type index range is The three embedded vectors are then concatenated with the numerical sub-vectors in a fixed order to form the input feature vector for each satellite observation.

[0078] The input feature vector is used as input to train a multipath suppression deep learning model, which is a feedforward multilayer perceptron and its model parameters are denoted as follows. The multilayer perceptron consists of three fully connected hidden layers with widths of 64, 32, and 16 respectively, and uses the ReLU activation function. The output layer has a three-head structure and outputs the observation weights. Noise scaling factor With observation bias ,in Constrained by the Sigmoid function , Constrained by the Sigmoid function and affine transformation pass Function and scaling coefficient constraints and with the observed values Maintain the same physical dimensions;

[0079] Era All observations within Construct a set of observational reliability parameters and Both constitute the observation bias set. Index and The observation records are stored one-to-one to ensure that they can be retrieved later according to the observation granularity.

[0080] Based on the observation bias set, the current satellite observation data is corrected, and corrected satellite observation data is generated. The observation correction satisfies the following:

[0081] ;

[0082] in Represents the epoch No. Corrected observations from satellite observations, This represents the corresponding original observation value. This represents the output of the multipath suppression deep learning model and the observation bias corresponding to the satellite observation. Indicates epoch index, The sequence number is indicated, and the corrected satellite observation records are displayed using the original satellite identifier. Frequency point identification With observation type identifier Keep it unchanged and only use replace To generate corrected satellite observation data.

[0083] In this specific embodiment, S4 includes:

[0084] Read the current inertial data and correct the satellite observation data between adjacent epochs, and divide the current inertial data into intervals based on the epoch timestamp. Inertial sampling sequences within the range are used to construct inertial pre-integral constraints, where epoch indexes are used. Consistent with the current point cloud frame number, the inertial sampling sequence includes angular velocity measurements. Compared with specific force measurement And the mean value integral method is used in the interval The pre-integral quantity is obtained by accumulating in chronological order. and and the corresponding pre-integral covariance matrix ,in Indicates from the epoch To the epoch The pre-integral displacement increment, Indicates the pre-integral velocity increment. This represents the pre-integral attitude increment quaternion. This represents the pre-integral uncertainty obtained from the propagation of inertial noise density and zero-biased random walk parameters, where the noise density is taken as the accelerometer noise density. Gyroscope noise density Accelerometer zero-bias random walk gyroscope zero-bias random walk ;

[0085] With sliding window length Construct a tightly coupled sliding window factor graph for each epoch, and then... The state to be estimated is denoted as And from the current position Current speed Current attitude quaternion Current accelerometer zero bias With the current gyroscope zero bias The structure consists of a superscript cur indicating the current job, for each pair of adjacent epochs. Introduce an inertia factor into the factor diagram and use a pre-integral quantity. With the pre-integral covariance matrix Forming inertial pre-integral constraints;

[0086] For the epoch Each corrected satellite observation record Construct satellite observation factors, among which Indicates satellite identifier, Indicates frequency point identifier, Indicates the observation type identifier. This indicates the correction of the observation values, and the position of the satellite in the Earth-fixed coordinate system at that epoch is calculated from the ephemeris. This is used to construct a tightly coupled observation model, so that the satellite observation residuals are corrected from the current state. Predicted observations and The difference is obtained;

[0087] Introduce a corresponding switch variable for each satellite observation factor. ,in For the range of values ​​within A continuous scalar and an interval projection is performed after each nonlinear iteration update to make Stay Within, the residual vector of the satellite observation factors is multiplied as a whole. To achieve When the factor approaches 0, its constraint effect weakens or becomes ineffective. When the factor approaches 1, it maintains normal constraint strength;

[0088] Based on the observation weights corresponding one-to-one with the observation in the set of observation confidence parameters. With noise scaling factor Update the measurement covariance matrix of the satellite observation factors, wherein the update satisfies:

[0089] ;

[0090] in Represents the epoch No. The updated measurement covariance matrix of the satellite observation factors. This represents the noise scaling factor and its value range is... Represents the observation weights and their value range is... This represents the lower limit constant for the weights and has a value of 0.05 to avoid a zero denominator. Indicator and observation type identifier The corresponding benchmark measurement covariance matrix is ​​determined by a fixed rule. ,in Indicates the baseline standard deviation and when When it is a pseudo-range ,when When the carrier phase is ,when For Doppler , where I represents the identity matrix consistent with the residual dimension;

[0091] For each switch variable A priori constraint is set to approach 1, and this prior constraint is represented in the factor graph as a function of the residuals. A prior factor is constructed with a prior standard deviation of 0.05 to keep the switch on under normal observation conditions and allow it to be downgraded when observations are abnormal;

[0092] After constructing the inertia factor, satellite observation factor, and switch prior factor, Gauss-Newton iterative optimization is performed on the tightly coupled sliding window factor map, with 5 iterations performed for each epoch update. The optimization result of the previous epoch is used as the initial value for the current iteration, and historical states exceeding the window are marginalized to maintain a constant window length, thereby outputting the predicted pose sequence within the sliding window. Simultaneously, it outputs the optimized state of the tightly coupled sliding window factor graph, wherein the optimized state includes each state to be estimated within the window. Current estimated values, each switch variable The current estimates, linearization points of each factor, and marginalized prior factors are used in subsequent steps to continue incremental updates within the same optimization framework.

[0093] In this specific embodiment, S5 includes:

[0094] Read the current epoch prediction pose and record it as ,in This indicates the epoch index, which is consistent with the current point cloud frame number. Indicates the predicted position at the current epoch. This represents the quaternion for the current epoch's predicted pose, and simultaneously reads and records the current point cloud data as... ;

[0095] right A deterministic preprocessing procedure is executed, which includes transforming the point cloud to a map coordinate system consistent with the reference trajectory using fixed extrinsic parameters, removing points with a distance greater than 60 m from the sensor origin, removing points with elevation coordinates less than -2 m or greater than 2 m, and downsampling using voxel filtering with a voxel side length of 0.2 m, thereby obtaining the current point cloud for retrieval and registration. ;

[0096] Will The point cloud feature deep learning model trained is used to generate a global descriptor for the current point cloud. This point cloud feature deep learning model is denoted as... And the parameters are Its structure consists of a point-level feature extraction backbone network and a global aggregation layer. The point-level feature extraction backbone network uses a shared multilayer perceptron to encode the 3D coordinates of each point and sequentially sets four linear layers with widths of [missing information]. Furthermore, each linear layer is followed by a ReLU activation function and batch normalization. Then, the point-level features are input into the NetVLAD global aggregation layer, and the number of cluster centers is set to [value missing]. The aggregation yields a fixed-length vector, which is then passed through a linear dimensionality reduction layer to output a dimension of [value missing]. global descriptor and perform Normalization is performed to obtain the current global descriptor of the point cloud. And guarantee ;

[0097] A reference pose index corresponding to the current epoch is determined on the reference trajectory, which is generated from the reference pose sequence obtained in step S1 and based on the reference position sequence. It means that, among them Indicates the reference pose index is Reference position, The end index of the reference pose sequence is represented by calculation. With all The reference pose index is obtained by taking the index corresponding to the minimum Euclidean distance. ;

[0098] Reference subgraphs are filtered by pose range in the reference subgraph library to form a retrieval subset. The reference subgraph library contains a set of reference subgraphs. and the reference pose range associated with each reference subgraph ,in The subgraph is identified as Reference subplot point cloud, Indicates the total number of reference subgraphs. Indicates the first The initial pose index of each reference subgraph Indicates the first The termination pose index of each reference subgraph is selected, and the filtering rule is to retain only those that satisfy the condition. Reference subgraphs and use them to form retrieval subsets ;

[0099] Retrieving subsets Perform a similarity search within the function to obtain a preset number of results. The candidate reference subgraph set, the reference subgraph library pre-stores the global descriptor of each reference subgraph. And satisfy During retrieval, for each calculate and The cosine similarity is calculated and sorted from highest to lowest, with the top results taken. The candidate reference subgraphs are composed of several reference subgraphs. ;

[0100] Using the predicted pose of the current epoch as the initial value, perform deep feature-guided fine registration of the point cloud for each candidate reference sub-image. Specifically, this involves... With candidate reference subgraph Use respectively The backbone network outputs point-level feature vectors for each point and establishes point-to-point correspondences through the nearest neighbors in the feature space. Subsequently, weighted singular value decomposition is used to solve the rigid body transformation, and ICP fine registration is performed for up to 30 iterations with the point-to-plane residual as the objective to obtain the registration result. ,in This represents a rigid body transformation that transforms the current point cloud from the current point cloud coordinate system to the reference subgraph coordinate system;

[0101] Calculate the registration residual for each registration result. Match confidence ,in Defined as the root mean square error of the distances from all interior points to the plane after fine registration convergence, with units of meters. Defined as the proportion of interior points in a feature correspondence where the distance from a point to a plane is less than 0.3 m out of the total number of corresponding points, and the value range is [value missing]. ;

[0102] Determine the target point cloud registration results and set the confidence threshold. And in satisfying Selecting registration residuals from candidate reference subplots The smallest one corresponds to The target point cloud registration result is recorded as its reference sub-image index. ;

[0103] Point cloud localization constraints are generated based on the target point cloud registration results, and the point cloud localization covariance matrix is ​​output. The point cloud localization constraints are derived from the target registration results. With reference subgraph The poses of the sub-maps recorded in the reference sub-map library are used to jointly determine the current pose measurement in the map coordinate system, which is then used to construct the point cloud localization factor. At the same time, the point cloud localization covariance matrix is ​​generated according to the following formula. To achieve registration quality reinjection:

[0104] ;

[0105] in Indicates the current epoch. The point cloud localization covariance matrix is matrix, This represents the registration residual of the target point cloud registration result. This represents the matching confidence score of the target point cloud registration result. This represents the lower confidence limit constant and has a value of 0.1 to avoid a denominator of zero. Represent the point cloud localization reference covariance matrix and take a fixed value. The reference variances are the corresponding position three-axis and attitude three-axis, respectively, with the attitude variance in units of . and the point cloud positioning constraints are combined with Output.

[0106] In this specific embodiment, S6 includes:

[0107] Read the optimization state of the tightly coupled sliding window factor graph, and select the current epoch from the optimization state. The corresponding pose variable is used as the pose to be estimated, which is determined by the current position. With the current attitude quaternion Composition, in which This indicates the epoch index, which is consistent with the current point cloud frame number. This represents a three-dimensional position vector in the map coordinate system. Represents the orientation of a unit quaternion in the map coordinate system;

[0108] Read the reference trajectory and use the reference position sequence It means that, among them Indicates the reference pose index is Reference position, Indicates the end index of the reference pose sequence;

[0109] Will Project onto the reference trajectory to obtain the projection point The projection process is performed on the polyline segment of the reference trajectory and indexed by the reference pose. As a local search center, within the index range Inner traversal of adjacent reference position pairs And calculate for each broken line segment The closest point to the given line segment is determined as the point with the smallest distance. The trajectory tangential unit vector is determined by the direction vector of the corresponding polyline segment. ,in The direction of travel is represented by the normalized direction vector of the broken line segment, taking only its component in the horizontal plane.

[0110] based on unit vector of gravity axis Construct the horizontal normal unit vector ,in Depend on and The cross product is obtained and normalized to ensure Located on the horizontal plane and with Orthogonal;

[0111] The lateral error is calculated based on the estimated pose and the projection point, and a reference trajectory fitting factor is constructed. The lateral error is defined as follows:

[0112] ;

[0113] in Indicates the current epoch. The lateral error scalar Indicates the current epoch. The corresponding horizontal normal unit vector, symbol Indicates transpose. Indicates the current epoch. The position vector to be estimated, Indicates the current epoch. The position vector of the projection point on the reference trajectory;

[0114] When writing the reference trajectory fitting factor into the factor plot, the residual is used. Construct a one-dimensional observation constraint and set the measurement covariance for this factor. And take To constrain lateral repeatability error to the centimeter level;

[0115] Read the point cloud localization constraints and represent them as point cloud pose measurements in the map coordinate system. and point cloud localization covariance matrix ,in This indicates the position measurement obtained from point cloud registration. This represents the attitude measurement obtained from point cloud registration. Indicates the corresponding Measurement covariance matrix;

[0116] Point cloud localization constraints are constructed as point cloud localization factors and introduced into the optimization state of the factor graph, wherein the residuals of the point cloud localization factors are composed of position errors. Composed of attitude error, the attitude error is formed by and The relative rotation is converted into a three-dimensional small-angle vector through a logarithmic mapping to ensure that the residual is additive in Euclidean space, and according to... Set the measurement covariance matrix of the point cloud localization factor to achieve the back injection of registration confidence onto the factor weights;

[0117] After simultaneously adding the reference trajectory fitting factor and the point cloud localization factor to the same tightly coupled sliding window factor map, Gauss-Newton iterative optimization is performed, and 5 iterations are executed during the current epoch update. The initial value of the iteration is taken from the optimization state. and The current estimate is used to obtain the current epoch optimized pose, and the optimization result is used to update the corresponding state in the optimization state. and .

[0118] In this specific embodiment, S7 includes:

[0119] Read the current epoch optimized pose and record it as ,in This indicates the epoch index, which is consistent with the current point cloud frame number. Indicates the current position in the map coordinate system. This represents the current attitude quaternion in the map coordinate system, and simultaneously reads the lateral error calculated in step S6 from the reference trajectory fitting factor. As the lateral error output corresponding to the reference trajectory;

[0120] based on( The optimized state of the tightly coupled sliding window factor graph is updated using a sliding window, ensuring that the variables to be estimated only include pose and inertial bias variables within a preset time length, where the preset time length is [value missing]. And record the epoch timestamp as And it is given by the timestamp of the current point cloud frame, through the window's starting epoch index. Determine the coverage area of ​​the current sliding window, the satisfy ,in Indicates the starting epoch index of the window. Indicates the candidate epoch index. Represents the epoch timestamp, Indicates the current epoch. timestamp, Indicates the preset time length;

[0121] All indices less than The historical state is removed from the variables to be estimated in the factor graph and marginalized to generate marginalized prior factors. The marginalization process linearizes the inertial factor, satellite observation factor, switch prior factor, point cloud positioning factor and reference trajectory fitting factor connected to the removed state at the current linearization point. After eliminating the removed state through Schur complement, prior information related only to the retained state is obtained. This prior information is written into the updated tightly coupled sliding window factor graph in the form of prior factors to ensure that the historical information continues to constrain subsequent epochs.

[0122] After marginalization is completed, the current epoch will be... The optimization result is used as the corresponding state within the window. Write the latest estimate into the optimization state and set the next epoch. The initial value is written into the optimization state for subsequent incremental optimization in step S4, where the initial value for the next epoch is... Combination The current inertial data in the interval is obtained by inertial pre-integration propagation and used as the initial value for newly added nodes in the tightly coupled sliding window factor graph;

[0123] The localization result of the current job is output based on the optimized state of the tightly coupled sliding window factor graph corresponding to the updated sliding window. The localization result includes the current epoch timestamp. Current location Current posture and the lateral error corresponding to the reference trajectory .

[0124] In this specific embodiment, training sensor data is acquired and a reproducible experimental dataset is formed. The training sensor data is collected by the same operating platform continuously traveling in open and occluded environments and includes training satellite observation data output by the global satellite navigation system receiver, training inertial data output by the inertial measurement unit, and training point cloud data output by the lidar. The timestamp of each frame of the training point cloud data is used as an epoch timestamp to perform time synchronization on the three types of data. The training satellite observation data is aligned according to the nearest epoch and retains quality fields such as satellite identifier, frequency identifier, observation type identifier, observation value, carrier-to-noise ratio, and lock duration. The training inertial data is divided into continuous sampling sequences according to adjacent epochs for inertial pre-integration.

[0125] A reference pose sequence for supervised learning is generated based on synchronous training data. Specifically, inertial pre-integration is performed on the training inertial data in the same manner as in step S4, and an offline tightly coupled smooth optimization problem is constructed. Simultaneously, in open environment segments, centimeter-level high-precision poses obtained by carrier phase fixing are used as external strong constraints to introduce optimization to anchor global drift. The reference pose sequence contains the reference position for each epoch. , reference attitude quaternion Reference speed and reference inertial bias and And with epoch index One-to-one correspondence;

[0126] Based on the aforementioned reference pose sequence and combined with satellite ephemeris, theoretical satellite observation values ​​for each epoch are calculated. These theoretical satellite observation values ​​are constructed using standard geometric distance, Earth rotation correction, tropospheric and ionospheric delay corrections, and receiver and satellite clock bias corrections. The observation residuals are obtained by subtracting the theoretical satellite observation values ​​from the training satellite observation data in the synchronous training data according to satellite identifier, frequency identifier, and observation type identifier. ,in Represents the epoch The The observation residuals of each training satellite observation are matched one-to-one with the subsequent multipath suppression training samples;

[0127] A multipath suppression training sample set is constructed using satellite observations as the granularity, and each multipath suppression training sample contains a satellite identifier. Frequency point identification Observation type identifier Observed values Carrier-to-noise ratio Lock duration Satellite elevation angle and supervisory information The model input features are generated according to the consistent normalization and embedding rules in step S3, where the satellite embedding table parameters are... And the satellite index range is The frequency point embedding table parameters are And the frequency index range is The observation type embedding table parameter is And the observation type index range is The embedded vectors and numerical features are concatenated and then input into a multipath suppression deep learning model to train and obtain model parameters. The multipath suppression deep learning model is consistent with step S3 and is a multilayer perceptron with three fully connected hidden layers. The widths of the hidden layers are 64, 32, and 16 respectively, and the ReLU activation function is used. The output layer has a three-head structure and outputs the observation weights. Noise scaling factor With observation bias ,in By using the Sigmoid constraint Constrained by Sigmoid and affine transformation pass With scaling factor constraints ;

[0128] In the construction of training supervision, observation residuals are used as the basis. Simultaneously monitor the observation bias, observation weights, and noise scaling factor, where the observation bias monitoring target is taken as... Observation weight supervision target selection The noise scaling factor is used to supervise the target. And respectively trained using Huber regression loss. Training using binary cross-entropy loss Training using Huber regression loss The transition parameter of Huber loss is taken Furthermore, the residual unit remains consistent with the corresponding output;

[0129] Based on the reference pose sequence, keyframe selection and training sub-graph generation are performed on the training point cloud data in the synchronous training data. The keyframe selection rules are the same as in step S1, and the keyframes are selected when the translation of adjacent keyframes is greater than or equal to 1 m or the change in heading angle is greater than or equal to 1 m. As a triggering condition, the training sub-image fuses the point clouds of the five keyframes before and after each keyframe as the center, and performs coordinate transformation and voxel downsampling with a side length of 0.2 m in the coordinate system of the center keyframe.

[0130] Construct a point cloud retrieval training sample set from the training subgraph set. Positive sample pairs are defined as having a relative translation of less than or equal to 2 m from the center keyframes of the two subgraphs and a relative heading angle less than or equal to... It is determined that negative sample pairs are formed by relative translation greater than or equal to 10 m or relative heading angle greater than or equal to 10 m. Determine and form triplet groups based on anchor point subgraphs, positive subgraphs, and negative subgraphs;

[0131] Using a point cloud feature deep learning model Output the triplet subgraphs with a dimension of 256 and... Normalized global descriptors and trained to obtain model parameters The point cloud feature deep learning model is consistent with step S5 and consists of a shared multilayer perceptron point-level backbone network and a NetVLAD aggregation layer. The width of the linear layers in the point-level backbone network is as follows: Furthermore, each layer is followed by batch normalization and ReLU activation function, and the number of NetVLAD cluster centers is taken as... The output layer linearly reduces the aggregated vector to 256 dimensions and normalizes it;

[0132] The two types of models are trained using the same training schedule and jointly trained using a total loss function, which is:

[0133] ;

[0134] in Indicates the total loss. This represents the supervised loss of a multi-path suppression deep learning model. This represents the learning loss metric of a deep learning model for point cloud features. Represents the weighting coefficients of the two types of loss and takes ;

[0135] Training uses the Adam optimizer and sets the learning rate. Weight decay The training rounds are 50. For multipath suppression, the training batch size is 1024 satellite observation samples, shuffled according to epoch. For point cloud retrieval, the training batch size is 32 sets of triples, and all subgraph triples are traversed in each round. The training is then fixed. and And used for online inference in steps S3 and S5 respectively.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0137] This invention addresses the problems of observational anomalies caused by signal obstruction and multipath propagation in global navigation satellite systems (GNSS), as well as the difficulty in achieving stable convergence of lateral repetition errors when performing multiple operations on the same work row. It integrates satellite observations, inertial pre-integration, and point cloud positioning constraints into a tightly coupled sliding window factor graph for joint optimization. On one hand, a multipath suppression deep learning model outputs observation reliability parameters and observation biases for each satellite observation, correcting the biases of the observations. Simultaneously, it adaptively updates the measurement covariance of the satellite observation factors based on the reliability parameters, quantitatively weakening abnormal observations before they enter the optimizer. On the other hand, it uses a global point cloud descriptor to search for candidate subgraphs in a reference subgraph library and performs precise registration to generate point cloud positioning constraints. The registration covariance is then used for point cloud factor weighting, providing reliable geometric positioning information even when satellite observations degrade. Furthermore, the current pose is projected onto a reference trajectory, constructing a reference trajectory fitting factor for lateral errors. This ensures that the optimization objective directly includes constraints on the lateral consistency of repeated walks, ultimately achieving robust centimeter-level positioning in obstructed sections and stable suppression of lateral errors for repetitive operations.

[0138] In terms of algorithm structure, this invention addresses the aforementioned technical problems with structured improvements for the optimizer: First, it refines the deep learning output from the traditional overall availability judgment to the granularity of satellite, frequency point, and observation type, and injects it into the tightly coupled factor graph in the form of observation bias and covariance updates, enabling multipath errors to affect residual weights and uncertainty propagation in an interpretable and computable manner; Second, it introduces switch variables and sets prior constraints for each satellite observation factor, allowing suspected abnormal observations to be automatically weakened or invalidated during the optimization process, avoiding solution divergence and trajectory jumps caused by local multipaths or occasional anomalies; Third, it adopts pose range filtering and candidate subgraph gating from the reference subgraph library, significantly reducing the mismatch range of point cloud retrieval, and uses covariance back-injection of registration confidence information for factor weighting, further suppressing repeated walking failures caused by cross-line matching, thereby better achieving the centimeter-level positioning and repeatable walking technology effects required in this case.

Claims

1. A repeatable walking localization method based on deep learning, characterized in that, include: S1. Acquire reference sensor data, including reference satellite observation data, reference inertial data, and reference point cloud data. Optimize the reference satellite observation data and reference inertial data in a tightly coupled sliding window factor graph to obtain a reference pose sequence, generate a reference trajectory, and generate a set of reference subgraphs based on the reference point cloud data, establishing a reference subgraph library containing the reference subgraphs and the reference pose range associated with each reference subgraph. S2. Acquire current sensor data, including current satellite observation data, current inertial data, and current point cloud data. S3. Input the current satellite observation data into a multipath suppression deep learning model to obtain a set of observation confidence parameters and an observation bias set. Correct the current satellite observation data based on the observation bias set to obtain corrected satellite observation data. S4. Based on the current inertial data, obtain the inertial pre-integration constraints, construct a tightly coupled sliding window factor graph, and introduce the corrected satellite observation data as satellite observation factors into the tightly coupled sliding window factor graph. Introduce corresponding switch variables for each satellite observation factor, so that the constraint effect weakens or becomes ineffective as the switch variables are applied. Update the measurement covariance matrix of the satellite observation factors according to the set of reliable observation parameters, and set prior constraints for the switch variables. Construct the inertial pre-integration constraints as inertial factors, introduce them into the tightly coupled sliding window factor graph, and optimize them to obtain the predicted pose sequence and optimized state. S5. Input the current point cloud data into the point cloud feature deep learning model to obtain the current point cloud global descriptor, and filter from the reference sub-graph library. Select candidate reference subgraphs and perform retrieval based on the current point cloud global descriptor. Register the current point cloud data with the retrieved candidate reference subgraphs to determine the target point cloud registration result, and generate point cloud localization constraints and point cloud localization covariance matrix; S6, in the optimization state, take the pose variable corresponding to the current epoch as the pose to be estimated, project the pose variable onto the reference trajectory to obtain the projection point, construct the reference trajectory fitting factor based on the lateral error between the pose variable and the projection point, construct the point cloud localization constraints as point cloud localization factors and set its measurement covariance matrix according to the point cloud localization covariance matrix, introduce the reference trajectory fitting factor and point cloud localization factor into the tightly coupled sliding window factor graph and optimize it to obtain the optimized pose of the current epoch; S7. Optimize pose based on the current epoch and output the positioning result of the current operation, including the current position and the lateral error corresponding to the reference trajectory.

2. The deep learning-based repeatable walking localization method according to claim 1, characterized in that, S1 includes: Acquire reference satellite observation data, reference inertial data, and reference point cloud data during the reference operation, and synchronize the reference satellite observation data, reference inertial data, and reference point cloud data in time to obtain synchronized reference data; Based on the synchronous reference data, a tightly coupled sliding window factor map is constructed. The reference satellite observation data is constructed as a satellite observation factor and introduced into the tightly coupled sliding window factor map. The reference inertial data is constructed as an inertial factor and introduced into the tightly coupled sliding window factor map. The tightly coupled sliding window factor map is optimized to obtain a reference pose sequence. Based on the reference pose sequence, reference keyframes are selected in the reference point cloud data, and a set of reference sub-graphs is generated based on the reference keyframes. When generating each reference sub-graph, the reference point cloud corresponding to the reference keyframe is fused with the reference point clouds corresponding to other reference keyframes that meet the preset neighborhood conditions to obtain the reference sub-graph corresponding to the reference keyframe. Generate a reference trajectory based on the reference pose sequence; For each reference subgraph in the set of reference subgraphs, a reference pose range is determined, wherein the reference pose range includes the start pose index and the end pose index in the reference pose sequence; The reference subgraph set and the reference pose range corresponding to each reference subgraph are written into the reference subgraph library to obtain the reference subgraph library.

3. The repeatable walking localization method based on deep learning according to claim 1, characterized in that, S2 include: Acquire the current satellite observation data, current inertial data, and current point cloud data during the current operation, and synchronize the current satellite observation data, current inertial data, and current point cloud data in time to obtain synchronized current data; The current satellite observation data in the synchronized current data is organized into a satellite observation set according to epochs, wherein each satellite observation includes a satellite identifier, a frequency identifier, an observation type identifier, and an observation value; Output the current satellite observation data, the current inertial data, and the current point cloud data.

4. The deep learning-based repeatable walking localization method according to claim 1, characterized in that, S3 includes: The current satellite observation data is input into the trained multipath suppression deep learning model to obtain an observation confidence parameter set and an observation bias set. The observation confidence parameter set refers to the observation weight and noise scaling coefficient corresponding to each satellite observation output in the current satellite observation data, and the observation bias set refers to the observation bias corresponding to each satellite observation output in the current satellite observation data. The observation bias set is matched one-to-one with each satellite observation in the current satellite observation data, and the observation values ​​of each satellite observation are corrected to obtain corrected satellite observation data.

5. The deep learning-based repeatable walking localization method according to claim 1, characterized in that, S4 includes: Based on the current inertial data, inertial pre-integration processing is performed to obtain inertial pre-integration constraints; a tightly coupled sliding window factor map is constructed using the pose variables and inertial bias variables within the sliding window as variables to be estimated; each satellite observation in the corrected satellite observation data is constructed as a satellite observation factor and introduced into the tightly coupled sliding window factor map, wherein a switch variable corresponding to the satellite observation is introduced for each satellite observation factor, and the switch variable is a continuous variable with a value range between zero and one; The residual term of the satellite observation factor is scaled according to the switch variable, so that when the switch variable approaches zero, the constraint effect of the satellite observation factor on optimization is weakened or invalidated. Based on the observation weights and noise scaling coefficients corresponding to the satellite observations in the set of observation confidence parameters, the measurement covariance matrix of the satellite observation factor is updated, and a priori constraint is set for the switch variable to approach one. The inertial pre-integration constraint is constructed as an inertial factor and introduced into the tightly coupled sliding window factor graph. The tightly coupled sliding window factor graph is optimized to obtain the predicted pose sequence within the sliding window, and the optimized state of the predicted pose sequence and the tightly coupled sliding window factor graph is output.

6. The repeatable walking localization method based on deep learning according to claim 1, characterized in that, S5 include: The current point cloud data is input into a trained point cloud feature deep learning model to obtain a global descriptor for the current point cloud. Based on the current epoch predicted pose in the predicted pose sequence, a corresponding reference pose index is determined on the reference trajectory. Reference sub-graphs whose reference pose range contains the reference pose index are selected from the reference sub-graph library to obtain a retrieval subset. Similarity retrieval is performed on the current point cloud global descriptor in the retrieval subset to obtain a set of candidate reference sub-graphs with a preset number of K. Using the current epoch predicted pose in the predicted pose sequence as the initial value, the current point cloud data is matched with each candidate reference sub-graph in the candidate reference sub-graph set using deep feature-guided fine point cloud registration to obtain a set of point cloud registration results. Registration residuals and matching confidence scores are calculated for each point cloud registration result. The target point cloud registration result is determined based on the set of point cloud registration results, and point cloud localization constraints and point cloud localization covariance matrices are generated based on the target point cloud registration result.

7. The repeatable walking localization method based on deep learning according to claim 1, characterized in that, S6 include: Using the pose variable corresponding to the current epoch in the tightly coupled sliding window factor graph as the pose to be estimated, the position corresponding to the pose variable is projected onto the reference trajectory to obtain the projection point, and the lateral error is calculated based on the pose variable and the projection point to construct the reference trajectory fitting factor; the point cloud localization constraint is constructed as the point cloud localization factor, and the measurement covariance matrix of the point cloud localization factor is set according to the point cloud localization covariance matrix; the reference trajectory fitting factor and the point cloud localization factor are introduced into the optimization state of the tightly coupled sliding window factor graph, and the updated tightly coupled sliding window factor graph is optimized to obtain the optimized pose of the current epoch.

8. The deep learning-based repeatable walking localization method according to claim 1, characterized in that, S7 includes: Based on the current epoch, the pose is optimized and the sliding window is updated so that the variables to be estimated in the tightly coupled sliding window factor graph include pose variables and inertial bias variables within a preset time length. Based on the optimized state of the tightly coupled sliding window factor map corresponding to the updated sliding window, the positioning result of the current operation is output, including the current position and the lateral error corresponding to the reference trajectory.

9. The deep learning-based repeatable walking localization method according to claim 1, characterized in that, The training methods for the multipath suppression deep learning model and the point cloud feature deep learning model are as follows: Acquire training sensor data, which includes training satellite observation data output by the global navigation satellite system receiver, training inertial data output by the inertial measurement unit, and training point cloud data output by the lidar. Synchronize the training sensor data in time to obtain synchronized training data. Based on the training inertial data in the synchronous training data, inertial pre-integration is performed, and combined with the high-precision carrier phase solution results of the global satellite navigation system in an open environment, the synchronous training data is fused and solved to obtain the reference pose sequence. Based on the reference pose sequence and combined with the satellite ephemeris, the theoretical satellite observation values ​​for each epoch are calculated; the theoretical satellite observation values ​​are then differentially analyzed with the training satellite observation data in the synchronous training data to obtain the observation residual for each satellite observation. A multipath suppression training sample set is constructed with satellite observations as the granularity. Each multipath suppression training sample includes at least a satellite identifier, frequency identifier, observation type identifier, observation value, and observation residual corresponding to the satellite observation. It further includes observation quality features and motion state features extracted from the synchronous training data. The multipath suppression deep learning model is trained with the multipath suppression training sample set as input and the observation residual as supervision information. The multipath suppression deep learning model outputs observation weights, noise scaling coefficients, and observation biases for each satellite observation. Based on the reference pose sequence, keyframe selection and subgraph generation are performed on the training point cloud data in the synchronous training data to obtain a training subgraph set. In the training subgraph set, positive sample pairs are constructed according to the condition that the relative poses between keyframe poses meet a preset threshold, and negative sample pairs are constructed according to the condition that the relative poses between keyframe poses do not meet the preset threshold, to obtain a point cloud retrieval training sample set. Using the point cloud retrieval training sample set as input, a point cloud feature deep learning model is trained using metric learning loss. This model outputs a point cloud global descriptor for the input point cloud, and ensures that the similarity of the point cloud global descriptor to positive sample pairs is greater than its similarity to negative sample pairs.

10. A repeatable walking localization method based on deep learning according to claim 6, characterized in that, The determination of the target point cloud registration result includes: under the condition that the matching confidence of each point cloud registration result is greater than the preset confidence threshold, selecting the point cloud registration result with the smallest registration residual as the target point cloud registration result.

11. The deep learning-based repeatable walking localization method according to claim 6, characterized in that, The reference pose range includes a start pose index and an end pose index. The filtering of reference subgraphs includes retaining reference subgraphs that satisfy the condition that the start pose index is less than or equal to the reference pose index and the reference pose index is less than or equal to the end pose index, as the retrieval subset.

Images