Crowdsourcing data-based lane signal map construction method and system

Abstract

The invention discloses a lane signal map construction method and system based on crowdsourcing data, and belongs to the technical field of vehicle localization, and the method comprises the steps: employing a Bayesian state estimation algorithm to estimate the obtained crowdsourcing data, and obtaining a vehicle motion track, inertial navigation error model parameters and magnetometer error model parameters; carrying out feature representation on the magnetic field sequence based on the vehicle motion trails by adopting a deep representation learning method to obtain magnetic field sequence features of key frames, and carrying out key frame association on the magnetic field sequence features by utilizing a loopback detection method to obtain loopback constraints among different trails; and constructing a three-dimensional trajectory inequality constraint based on key frame association, and constructing a global trajectory optimization constraint and obtaining a lane signal map in combination with inertial navigation error model parameters, magnetometer error model parameters and loopback constraints among different trajectories. According to the invention, the method achieves the construction of the lane-level road feature map based on crowdsourcing data, and remarkably reduces the cost of a lane-level positioning scheme based on road feature matching.

Description

Technical Field

[0001] This invention belongs to the field of vehicle positioning technology, specifically relating to a method and system for constructing lane signal maps based on crowdsourced data. Background Technology

[0002] The demand for lane-level positioning and navigation services for consumer-grade devices is growing rapidly. In open outdoor areas, Global Navigation Satellite System (GNSS) can provide relatively accurate and reliable positioning, and consumer-grade devices can achieve lane-level positioning by receiving correction information from base stations. However, in challenging environments such as urban canyons and tunnels, GNSS-based positioning technology cannot guarantee positioning accuracy. Existing solutions achieve full-area lane-level positioning by fusing GNSS and inertial navigation information with lane-level magnetic field maps. However, the construction and maintenance of large-scale lane-level road feature maps is costly. To achieve large-scale mapping, lane-level road feature map construction based on crowdsourced data sources is a feasible solution.

[0003] However, building location maps based on crowdsourced data sources is a challenging problem. Existing crowdsourced mapping solutions mainly focus on building indoor WiFi maps and indoor magnetic field maps. Details are as follows:

[0004] (1) Crowdsourced construction of indoor WiFi map.

[0005] The main steps in crowdsourcing the construction of indoor WiFi maps are trajectory recovery, trajectory association, and trajectory optimization. Since WiFi signals themselves carry unique identifiers for base stations, trajectory association based on WiFi signals is relatively simple. Therefore, crowdsourcing the construction of WiFi maps is also relatively straightforward. However, outdoor areas typically lack WiFi signal coverage, making this method unsuitable.

[0006] (2) Crowdsourced construction of indoor magnetic field map.

[0007] The main steps of indoor magnetic field map crowdsourcing construction technology are also trajectory recovery, trajectory association, and trajectory optimization. Indoor magnetic field map crowdsourcing construction can leverage the geometric constraints formed by the intersections between indoor trajectories, and indoor magnetic fields change relatively quickly and have a higher distribution. However, how to address the problems of low signal strength and weak geometric constraints between trajectories on lanes remains to be studied.

[0008] In conclusion, how to build lane signal maps based on crowdsourced data remains an unsolved problem. Summary of the Invention

[0009] This invention addresses the problems of low signal strength and weak geometric constraints between trajectories in existing technologies by providing a method for constructing lane signal maps based on crowdsourced data. It employs a loop closure confirmation process using magnetic field sequence representation learning based on contrastive learning, data augmentation, and adaptive clustering with KD-tree + k-NN candidate retrieval. Furthermore, it solves the problems of heading drift in single sub-trajectories, unobservable sensor bias in crowdsourced data, and inconsistent trajectory shapes by integrating initial value estimation using error-state Kalman filtering (ESKF) from GNSS, Micro-Electro-Mechanical Systems Inertial Measurement Unit (MEMS-IMU), and magnetometers, combined with graph optimization global VDR solution and global average magnetic field constraints through zero-velocity and non-holonomic constraints.

[0010] According to one aspect of the present invention, a method for constructing a lane signal map based on crowdsourced data is provided, comprising:

[0011] The Bayesian state estimation algorithm is used to estimate the acquired crowdsourced data to obtain the vehicle trajectory, inertial navigation error model parameters, and magnetometer error model parameters.

[0012] A deep representation learning method is used to represent the magnetic field sequence based on the vehicle motion trajectory to obtain the magnetic field sequence features of key frames. A loop closure detection method is used to associate the magnetic field sequence features of key frames.

[0013] Based on the keyframe association results, a three-dimensional trajectory inequality constraint is constructed. Combined with the inertial navigation error model parameters, magnetometer error model parameters, and lapsing constraints between keyframes, a global trajectory optimization constraint is constructed. The lane signal map is then obtained based on the global trajectory optimization constraint.

[0014] Furthermore, a Bayesian state estimation algorithm is used to estimate the acquired crowdsourced data, including:

[0015] Initial value estimation of vehicle trajectory is performed using the Bayesian state estimation algorithm;

[0016] Based on the initial value estimation results of the vehicle's motion trajectory, zero-speed constraints, zero-angle increment constraints, and non-holonomic motion constraints are constructed, and the graph optimization problem is defined together with pre-integration constraints and global average magnetic field constraints.

[0017] The graph optimization problem is solved to obtain the vehicle trajectory, inertial navigation error model parameters, and magnetometer error model parameters.

[0018] Furthermore, the loop closure detection method is used to associate keyframes with magnetic field sequence features, including:

[0019] The magnetic field sequence features of keyframes are normalized and a KD-tree index is constructed.

[0020] When new candidate keyframes are generated, k-NN query is used to return several candidate keyframes;

[0021] The spatiotemporal proximity exclusion rule is used to exclude candidate keyframes that are within a set range of the current keyframe in terms of space or time.

[0022] Threshold filtering is performed on the remaining candidate keyframes to filter out candidate keyframes whose scores are lower than a preset threshold with the current keyframe, and these candidate keyframes are then associated with the current keyframe.

[0023] Furthermore, based on the keyframe association results, three-dimensional trajectory inequality constraints are constructed, including:

[0024]

[0025] In the formula, k represents the trajectory. and These represent the three-dimensional translation and heading angles, respectively. Represents the Cauchy kernel function and cost function. This represents the three-dimensional spatial distance constraint between two keyframes of the loop closure. and Indicates the angle threshold. Indicates belonging to Summation of all constraints.

[0026] Furthermore, global trajectory optimization constraints are constructed, including:

[0027]

[0028] In the formula, k represents the trajectory. These represent the IMU's position, velocity, and attitude in the navigation coordinate system, respectively. These represent the accelerometer zero bias, gyroscope zero bias, and magnetometer zero bias of the current sub-trajectory, respectively. This represents the magnetic field vector in the navigation coordinate system. Constraints were defined within each sub-trajectory. This indicates the global average magnetic field constraint. This indicates the closure constraint between keyframes. This indicates the alignment point magnetic field vector constraint. This represents the sum of all constraints that are part of a loop constraint. This represents the sum of all constraints belonging to the alignment point magnetic field vector constraint. This represents the sum of all constraints that belong to the global average magnetic field constraint.

[0029] Furthermore, a lane signal map is obtained based on global trajectory optimization constraints, including:

[0030] Lane magnetic sequences are extracted based on global trajectory optimization constraints to generate a lane signal map. The lane signal map includes the magnetic field sequence and magnetic field spatial distribution information corresponding to the lane, which is used for lane-level positioning and positioning enhancement.

[0031] Furthermore, the use of deep representation learning methods to feature-represent the magnetic field sequence based on vehicle motion trajectory also includes:

[0032] The residuals of the magnetometer error model, absolute heading error, short-term magnetic disturbances, and observation noise are simulated based on the magnetic field sequence, and the simulation results are used to augment the magnetic field sequence.

[0033] According to one aspect of the present invention, a lane signal map construction system based on crowdsourced data is provided, comprising:

[0034] The Bayesian estimation module is used to estimate the acquired crowdsourced data using the Bayesian state estimation algorithm to obtain the vehicle trajectory, inertial navigation error model parameters, and magnetometer error model parameters.

[0035] The keyframe association module is used to perform feature representation on the magnetic field sequence based on the vehicle motion trajectory using a deep representation learning method to obtain the magnetic field sequence features of the keyframes, and to perform keyframe association on the magnetic field sequence features of the keyframes using a loop closure detection method.

[0036] The lane signal map module is used to construct three-dimensional trajectory inequality constraints based on keyframe association results. It combines inertial navigation error model parameters, magnetometer error model parameters, and lapsing constraints between keyframes to construct global trajectory optimization constraints, and obtains the lane signal map based on the global trajectory optimization constraints.

[0037] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the lane signal map construction method based on crowdsourced data.

[0038] According to one aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the lane signal map construction method based on crowdsourced data.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. This invention solves the problems of heading drift of single sub-trajectory, unobservable sensor zero bias in crowdsourced data, and inconsistent trajectory shape by integrating the initial ESKF estimation of GNSS, MEMS-IMU and magnetometer, combined with graph optimization global VDR solution and global average magnetic field constraint of zero velocity constraint and non-holonomic constraint. It realizes smooth three-dimensional recovery of crowdsourced trajectory and estimation of magnetometer zero bias, and provides a globally consistent and usable trajectory basis and error model for subsequent lane-level feature alignment and mapping.

[0041] 2. This invention solves the problem of incorrect matching caused by noise, installation posture and short-term disturbances in magnetic field sequences under crowdsourced collection conditions by using magnetic field sequence representation learning based on contrastive learning, data augmentation, and loop closure confirmation process of adaptive clustering and KD tree + k-NN candidate retrieval. It realizes robust similarity measurement and accurate loop closure keyframe recognition of lane-level magnetic field sequences, and achieves reliable lane-level trajectory association on low-quality crowdsourced data. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 The present invention provides a flowchart of a lane signal map construction method based on crowdsourced data.

[0044] Figure 2 This is a comparative learning diagram provided by the present invention.

[0045] Figure 3 This is a schematic diagram of data augmentation in comparative learning provided by the present invention. Detailed Implementation

[0046] It should be noted that the VDR of this invention refers to the vehicle trajectory obtained by multi-source fusion methods such as GNSS / vision / vehicle mileage / inertial navigation, and the specific schemes and sensors used are not limited.

[0047] 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.

[0048] like Figure 1 As shown, this invention proposes a method for constructing a lane signal map based on crowdsourced data, including: S1: Obtaining a relatively accurate vehicle motion trajectory. A preset Bayesian state estimation algorithm is used to fuse position information provided by GNSS, vehicle kinematic constraints, motion dynamic information provided by inertial navigation, and magnetic field information provided by a magnetometer to estimate the vehicle motion trajectory (vehicle position, speed, attitude), inertial navigation (INS) error model parameters, and magnetometer error model parameters. S2: Utilizing the similarity of magnetic field features at the same location, a magnetic field sequence similarity measurement method based on comparative learning is used to study the lane-level keyframe association of crowdsourced magnetic field sequences, thereby achieving a precise association method for magnetic field sequence features. An adaptive clustering method (k-NN query) is used in the association process. S3: Based on global graph optimization, relative pose information within the trajectory and loop closure constraints between trajectories are fused to construct a lane signal map. Here, crowdsourced data refers to multi-sensor data collected by the user's smart device (such as a mobile phone) without the user's awareness. Crowdsourced data can include collected GNSS, INS, magnetometer, barometer, and other data. Sensors such as GNSS, inertial navigation, and magnetometers include accelerometers, gyroscopes, magnetometers, odometry, and GNSS; these sensors are integrated into a smartphone, which is fixed to the vehicle via a phone holder. Vehicle kinematic constraints include zero-velocity constraints, zero-angular increment constraints, and non-holonomic constraints. The inertial navigation system provides motion dynamics information including relative position, attitude, and velocity between keyframes, while the magnetometer provides magnetic field information including absolute and relative magnetic fields, as well as a global magnetic field vector.

[0049] Specifically, the embodiments of the present invention provide the following details for S1:

[0050] S1.1 Initial value estimation of vehicle trajectory based on filtering

[0051] This invention presents a post-processing algorithm for vehicle motion trajectory recovery. Based on initial trajectory parameter values ​​provided by Kalman filtering, zero-speed constraints, zero-angle increment constraints, non-holonomic motion constraints, and global average magnetic field constraints are constructed to achieve a globally optimal VDR solution based on graph optimization. Error-state Kalman filtering (a Bayesian state estimation algorithm) is used for vehicle pose estimation. The system state of the filter is as follows:

[0052] (1)

[0053] In the formula, , , These represent the IMU's position, velocity, and attitude in the navigation coordinate system, respectively. , The parameters for the inertial navigation error model are denoted by and represent the zero biases of the accelerometer and gyroscope, respectively, and are modeled as a first-order Gaussian Markov process. The state transition of the filter uses a simplified INS mechanical arrangement, specifically as follows:

[0054] (2)

[0055] In the formula, and These represent the measured values ​​of acceleration and angular velocity, respectively. Represents gravitational acceleration. Represents an exponential mapping over the special orthogonal group SO(3); Indicates time difference, Indicates the posture at the next moment. Indicates the velocity at the next moment.

[0056] Under GNSS availability, the positioning results from GNSS point positioning are used to correct the INS status. Considering the lever arm effect, the positional transformation relationship between the GNSS antenna phase center and the IMU measurement center is as follows:

[0057] (3)

[0058] In the formula, Indicates the position of the GNSS antenna phase center. Indicates the location of the IMU measurement center. This represents the coordinate scale transformation matrix (radius of curvature matrix), used to establish a linear approximation between the geodetic coordinate representation and the linear displacement in the local navigation coordinate system. Indicates a gesture; This represents the projection of the arm vector pointing from the IMU measurement center to the GNSS antenna phase center onto the vehicle system. Under GNSS rejection conditions, vehicle trajectory estimation mainly relies on autonomous estimation by the MEMS-IMU and vehicle motion constraints.

[0059] Using MEMS-IMU to estimate vehicle trajectory mainly relies on two constraints: 1) zero-speed constraint and zero-angle increment constraint in a stationary state; 2) nonholonomic constraint (NHC) during vehicle movement.

[0060] 1) Zero velocity constraint and zero angle increment constraint

[0061] The system determines whether the vehicle is stationary by using raw IMU measurements. If the vehicle is identified as stationary, the system state can be corrected using zero-speed and zero-angle incremental constraints.

[0062] Zero-speed constraints utilize the information that the vehicle's velocity is 0 and its coordinates remain unchanged when stationary. However, since the vehicle is not a rigid body, it is practically difficult to guarantee that its instantaneous velocity and angular velocity are both zero. To improve stability, position invariance is used as a constraint. The observation equations are as follows:

[0063] (4)

[0064] In the formula, This represents the displacement vector between two time points. and These represent the start and end times of the zero-velocity state, respectively. and These represent the positions of the carrier in the navigation coordinate system at two different times.

[0065] The zero-angle increment constrained observation equation is defined as:

[0066] (5)

[0067] In the formula, This indicates the relative posture at two moments. and These represent the rotations from the carrier coordinate system to the navigation coordinate system at two different moments; express The transpose of .

[0068] 2) Incompleteness constraints

[0069] Non-integrity constraints utilize the property that a vehicle cannot move laterally. Specifically, during travel, the lateral and vertical velocities at the center point of the rear wheels are zero. The observation equation is:

[0070] (6)

[0071] In the formula, and Rotation and translation between the vehicle system and the load system can be estimated online with GNSS assistance. This represents the projection of the vehicle's velocity relative to the navigation coordinate system at time t onto the vehicle frame. This represents the angular velocity measurement value of the gyroscope.

[0072] S1.2 Trajectory Recovery with Global Magnetic Field Vector Enhancement

[0073] Because the vehicle-mounted mobile phone is stably installed, motion constraints can be fully utilized, ensuring minimal divergence of the individual trajectory relative to the heading. In vehicle trajectory recovery, the main purpose of using the global magnetic field vector is to ensure that the headings of all trajectories are roughly aligned, and that the magnetometer zero bias (magnetometer error model parameters) is relatively consistent. The corresponding graph optimization problem is defined as:

[0074] (7)

[0075] In the formula, These represent the accelerometer zero bias, gyroscope zero bias, and magnetometer zero bias of the current sub-trajectory, respectively. Due to the finite length of each sub-trajectory, the zero bias modeling of each sensor has been simplified to a certain extent, directly treating the sensor zero bias over a short period of time as a constant value. This represents the magnetic field vector in the navigation coordinate system, used to unify the heading of all sub-trajectories. This indicates the global average magnetic field constraint. This represents the sum of all constraints belonging to this type. Graph optimization problems, also known as nonlinear least squares problems, represent complex estimation problems as graph structures and solve for the optimal state by minimizing the overall error.

[0076] By setting the eastward component of the magnetic field to follow a zero-mean Gaussian distribution, the aim of this method is to ensure the observability of the global heading and global magnetic field vector. The graph optimization cost function used for each sub-trajectory is defined as follows:

[0077] (8)

[0078] In the formula, k represents the trajectory. This indicates pre-integral constraints between keyframes. This represents the sum of all constraints belonging to this type. The pre-integration model can be reduced to integrating IMU measurements in a specific coordinate system to obtain a time-position interval. ,speed and posture The relative changes are then used to constrain the relative position, attitude, and velocity between keyframes; This represents the sum of all constraints that are part of the pre-integral constraints. This represents the sum of all constraints that are non-integrity constraints. This represents the sum of all constraints that fall under the categories of zero velocity constraints and zero angle increment constraints. and These represent the sum of all constraints belonging to absolute and relative magnetic field constraints, respectively.

[0079] The pose and velocity updates defined in the pre-integration can be expressed as:

[0080] (9)

[0081] By calculating pre-integral observations in advance , and The cost function between keyframes t and t+s constructed based on INS mechanical orchestration can be expressed as:

[0082] (10)

[0083] Furthermore, the pre-integration constraint here requires perturbation analysis of the above equation regarding the Jacobian matrix of zero bias of each sensor. For simplification, only the basic principle of pre-integration is given here.

[0084] This represents a non-holonomic constraint. In graph optimization algorithms, using the relative pose and velocity obtained through pre-integration, short-term relative lateral and longitudinal displacement constraints within the vehicle system can be constructed, specifically defined as follows:

[0085] (11)

[0086] (12)

[0087] In the formula, and Both refer to the installation relationship of the carrier relative to the vehicle. and This represents the attitude and velocity at the i-th epoch after keyframe t. The coefficient matrix represents the three-dimensional velocity that is mapped to the first two dimensions.

[0088] The zero-velocity constraint and zero-angle increment constraint are defined as shown above. and These represent absolute and relative magnetic field constraints, respectively. The absolute magnetic field vector constraint is related to the carrier attitude, magnetometer bias, and global magnetic field vector. To reduce the influence of local magnetic field disturbances, the short-term average measurement value of the magnetometer is used to form the constraint, specifically defined as follows:

[0089] (13)

[0090] (14)

[0091] (15)

[0092] In the formula, This represents the magnetic field vector of the n-system. This represents the short-term average of the magnetometer measurements at time i. This represents the short-term average of the magnetometer measurements at time j. This indicates that the magnetometer has zero bias. This indicates the number of magnetometer measurements acquired within a short period of time. This indicates the value measured by the magnetometer.

[0093] Based on the initial values ​​provided by the aforementioned filter, trajectory recovery based on global magnetic field vector enhancement can yield a smooth 3D vehicle motion trajectory. This optimization algorithm uses the constraint that the average magnetic field measured on each sub-trajectory is consistent with the global average magnetic field, ensuring that the headings of each trajectory are roughly aligned. The absolute and relative magnetic field constraints included in the graph optimization cost function guarantee zero-bias convergence of the system magnetometer, thus ensuring that the estimated magnetic field sequences of each trajectory can be used for keyframe association.

[0094] Specifically, the embodiments of the present invention provide the following details for S2:

[0095] S2.1 Characteristic Representation of Magnetic Field Sequences

[0096] Correct correlation between magnetic field sequences is a prerequisite for generating high-quality lane magnetic field maps. However, the magnetic field sequence is the result of a magnetometer sampling a continuous magnetic field in three-dimensional space. The sampled magnetic field sequence is affected by the accuracy of the relative position of the trajectory, the accuracy of the absolute heading, the accuracy of the absolute velocity, and the accuracy of the magnetometer error model parameters. Since crowdsourced data comes from mass-market smartphones, the data collection conditions are completely uncontrollable, making it difficult to impose high requirements on the quality and accuracy of the data collection. Therefore, to adapt to the low reliability of crowdsourced data, a deep representation learning method is used to train a neural network model to learn a robust and accurate magnetic field sequence feature. Here, deep representation learning is a representation learning method within deep learning.

[0097] The learning method can be expressed as learning a function:

[0098] (16)

[0099] In the formula, x represents the original data input, and y represents the feature vector obtained from the original input. The space in which they reside is called the feature space. The goal of representation learning is to make a certain distance metric (e.g., cosine distance, Euclidean distance, Manhattan distance, etc.) in the y space optimally represent the similarity between the original data, that is, the samples satisfy the following condition:

[0100] (17)

[0101] In the formula, The similarity function between two original inputs is difficult to find directly because the magnetic field sequence is affected by many factors. The distance function in the feature space is usually defined as the cosine function. The goal of this study is to find a mapping... In the mapped target space, the distance metric function Functions that correspond to the similarity of magnetic field sequences in the original space .

[0102] To achieve this goal, a comparative learning approach is adopted. For example... Figure 2 As shown, for any triplet ( Indicates the anchor point; This indicates a positive sample, belonging to the same class as the anchor point; (This indicates a negative sample, which is not of the same class as the anchor point), satisfying that the distance from the anchor point to the positive sample and the difference between the distance from the anchor point to the negative sample are greater than a threshold.

[0103] (18)

[0104] The network is trained using the classic triplet loss function, which can be expressed as:

[0105] (19)

[0106] In the formula, α represents the hyperparameter, used to force the minimum distance difference between positive and negative samples in the feature space. To train the network model using the cost function defined above, the triples described in the previous section need to be generated from the original dataset. The specific process is as follows... Figure 3 As shown, two samples are randomly sampled from the dataset, and these two samples are negative samples of each other. Simultaneously, positive samples are generated by performing data augmentation operations on the anchor points. Data augmentation operations can be understood as actively transforming the original data, causing changes while preserving its essential information. For example, in image data augmentation, typical operations include warping, affine transformation, and adding noise.

[0107] S2.2 Data Augmentation Methods for Magnetic Field Sequence Features

[0108] To find data augmentation methods for magnetic field sequence features, it is necessary to analyze the factors that influence these features. Empirically, the main influencing factors include the relative pose accuracy of the trajectory, the accuracy of the absolute heading, the estimation error of the magnetometer noise model, and short-term magnetic field disturbances (such as changes in the phone's pose relative to the vehicle body and the proximity of external ferromagnetic objects). By simulating the magnetometer error model residuals, absolute heading errors, short-term magnetic disturbances, and observation noise on the original magnetic field sequence, diverse and robust training samples are generated to improve the robustness of representation learning / matching to noise and disturbances.

[0109] The magnetic field observation sequence is a function of time, defined as:

[0110] (20)

[0111] In the formula, These represent the magnetometer's proportional factor error and zero bias error, respectively. This indicates the rotation and translation of the sensor relative to the navigation coordinate system at the corresponding moment; This represents the magnetic field vector corresponding to a geographical location, and this vector can be assumed to be a fixed value; This is the observation noise of the magnetometer, which can be assumed to follow a zero-mean Gaussian distribution, and the standard deviation of the distribution is known; This refers to short-term magnetic field disturbances, which have random characteristics and are usually unpredictable and cannot be compensated for using traditional methods.

[0112] Among these factors, the magnetometer error model parameters can be estimated using online calibration, but the accuracy of online calibration is limited. For absolute position and attitude, the horizontal and longitudinal positions are relatively reliable, and the relative horizontal position and relative heading accuracy are also high, so their errors can be ignored in practical use. The magnitude of the absolute heading error needs to be considered during training. The standard deviation of the magnetometer's observation noise is small and is usually ignored in use. Short-term magnetic field interference is difficult to detect and eliminate simply, and is one of the main challenges in magnetic field sequence matching. Therefore, data augmentation strategies for magnetic field sequence observations mainly need to consider the residuals of the magnetometer error model parameters, heading errors, and the impact of short-term magnetic field interference.

[0113] S2.3 Correlation of Magnetic Field Sequence Characteristics

[0114] Based on the magnetic field sequence features extracted insensitive to specific interferences through the above steps, adaptive clustering and loop closure detection methods are used for processing. A KD-tree is then employed for rapid magnetic loop keyframe identification within the distance metric space. Specifically, all extracted magnetic field sequence feature vectors are first normalized, and the magnetic field features of all keyframes are used to construct a KD-tree index. When new candidate keyframes are generated, a k-NN query is performed using the KD-tree to quickly return several approximate matching points in Euclidean space. These initially obtained coarse candidate magnetic loop keyframes are then excluded using an exclusion rule, removing those that are too close spatially or temporally. The remaining candidate keyframes are then threshold-screened; when the best score is below a preset threshold, it is confirmed as a magnetic loop keyframe.

[0115] Specifically, the embodiments of the present invention provide the following details for S3:

[0116] S3.1 Trajectory Optimization with Three-Dimensional Trajectory Inequality Constraints

[0117] This optimization process is based on the keyframe loop closure information obtained in S2 and the absolute heading information provided by the global magnetic field vector, and uses inequality constraints to achieve relative position constraints between trajectories. In order to make full use of the high-precision vehicle motion trajectory obtained in step S1, in the trajectory optimization with inequality constraints, the poses of all keyframes in the same sub-trajectory are represented by the pose of the first keyframe, while avoiding erroneous loop closure results from destroying the relative shape of the trajectory.

[0118] In summary, the three-dimensional trajectory inequality-constrained trajectory optimization problem is defined as follows:

[0119] (twenty one)

[0120] In the formula, k represents the trajectory. and These represent three-dimensional translation and heading angle, respectively, and their physical meaning is the change in the three-dimensional position and heading of the corresponding trajectory in the navigation coordinate system;

[0121] (twenty two)

[0122] In the formula, This represents the independent variable of the function. Represents the Cauchy kernel function and cost function. The three-dimensional spatial distance constraint between two keyframes of a loop closure is expressed in the following form:

[0123] (twenty three)

[0124] In the formula, This represents the position of the i-th keyframe of trajectory k, as indicated by the starting frame of the current trajectory. This represents the three-dimensional rotation matrix corresponding to the angular heading transformation. , This represents the 3D position of two different keyframes. An angle threshold is used empirically. and Take ±20 degrees respectively.

[0125] The solution to the above three-dimensional trajectory inequality-constrained trajectory optimization problem is used to update the pose and velocity of each keyframe within each sub-trajectory. The update formula is as follows:

[0126] (twenty four)

[0127] S3.2 Global Trajectory Optimization

[0128] Global trajectory optimization, based on the sub-trajectory internal constraints provided by the IMU and the loop closure constraints provided by the loop closure detection results, solves for the optimal trajectory estimate to obtain the final vehicle motion trajectory. The global optimization problem is defined as:

[0129] (25)

[0130] In the formula, The constraints within each sub-trajectory are defined, including: pre-integral constraints on the relative pose of the keyframe, zero-velocity and zero-angular-velocity change constraints based on the zero-velocity state, non-holonomic constraints based on vehicle motion characteristics, and absolute and relative magnetic field constraints based on magnetometer observations. It is a global average magnetic field constraint used to ensure that the global heading and global magnetic field are considerable; This represents the sum of all constraints that are part of a loop constraint. This represents the sum of all constraints belonging to the alignment point magnetic field vector constraint. This represents the sum of all constraints that belong to the global average magnetic field constraint.

[0131] The closure constraint between keyframes is defined as follows:

[0132] (26)

[0133] In the formula, This represents the pose and rotation corresponding to the i-th keyframe of the k-th trajectory. and This refers to the alignment point in two keyframes that passes through the same physical location. This property of passing through the same physical location is obtained based on fine alignment of the magnetic field sequence.

[0134] The alignment point magnetic field vector constraint (the constraint that the magnetic field vectors at the magnetic loop keyframes are equal) is defined as follows:

[0135] (27)

[0136] In the formula, and This indicates the relative pose of the alignment point to the keyframe. and This indicates that the magnetometers of the two sub-trajectories have zero bias. and This indicates the alignment point corresponds to the magnetometer measurement value, which is obtained by linear interpolation of the original magnetometer measurement value on the time axis.

[0137] The implementation of the various embodiments of the present invention is based on programmed processing by a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide a lane signal map construction system based on crowdsourced data. This system is used to execute a lane signal map construction method based on crowdsourced data from the above method embodiments.

[0138] The system includes: a Bayesian estimation module, used to estimate the acquired crowdsourced data using a Bayesian state estimation algorithm to obtain vehicle trajectory, inertial navigation error model parameters, and magnetometer error model parameters; a keyframe association module, used to perform feature representation of the magnetic field sequence based on the vehicle trajectory using a deep representation learning method to obtain the magnetic field sequence features of keyframes, and to perform keyframe association using a loop closure detection method; and a lane signal map module, used to construct three-dimensional trajectory inequality constraints based on the keyframe association results, combine the inertial navigation error model parameters, magnetometer error model parameters, and loop closure constraints between keyframes to construct global trajectory optimization constraints, and obtain a lane signal map based on the global trajectory optimization constraints.

[0139] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides an electronic device, including a memory and a processor. The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize a lane signal map construction method based on crowdsourced data as proposed in the above embodiments.

[0140] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, this program solves the problem of erroneous matching caused by noise, installation posture, and short-term disturbances affecting magnetic field sequences under crowdsourced collection conditions. It achieves robust similarity measurement and accurate loop closure keyframe identification for lane-level magnetic field sequences, enabling reliable lane-level trajectory association on low-quality crowdsourced data.

[0141] Finally, it should be noted that the above specific embodiments are merely representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and many variations are possible. Any simple modifications, equivalent changes, and alterations made to the above specific embodiments based on the technical essence of the present invention should be considered within the protection scope of the present invention.

Claims

1. A method for constructing lane signal maps based on crowdsourced data, characterized in that, include: The Bayesian state estimation algorithm is used to estimate the acquired crowdsourced data to obtain the vehicle trajectory, inertial navigation error model parameters, and magnetometer error model parameters. A deep representation learning method is used to represent the magnetic field sequence based on the vehicle motion trajectory to obtain the magnetic field sequence features of key frames. A loop closure detection method is used to associate the magnetic field sequence features of key frames. Based on the keyframe association results, a three-dimensional trajectory inequality constraint is constructed. Combined with the inertial navigation error model parameters, magnetometer error model parameters, and lapsing constraints between keyframes, a global trajectory optimization constraint is constructed. The lane signal map is then obtained based on the global trajectory optimization constraint.

2. The method for constructing a lane signal map based on crowdsourced data according to claim 1, characterized in that, The Bayesian state estimation algorithm is used to estimate the acquired crowdsourced data, including: Initial value estimation of vehicle trajectory is performed using the Bayesian state estimation algorithm; Based on the initial value estimation results of the vehicle's motion trajectory, zero-speed constraints, zero-angle increment constraints, and non-holonomic motion constraints are constructed, and the graph optimization problem is defined together with pre-integration constraints and global average magnetic field constraints. The graph optimization problem is solved to obtain the vehicle trajectory, inertial navigation error model parameters, and magnetometer error model parameters.

3. The method for constructing a lane signal map based on crowdsourced data according to claim 1, characterized in that, Keyframe association is performed using the loop closure detection method to analyze the magnetic field sequence features of keyframes, including: The magnetic field sequence features of keyframes are normalized and a KD-tree index is constructed. When new candidate keyframes are generated, k-NN query is used to return several candidate keyframes; The spatiotemporal proximity exclusion rule is used to exclude candidate keyframes that are within a set range of the current keyframe in terms of space or time. Threshold filtering is performed on the remaining candidate keyframes to filter out candidate keyframes whose scores are lower than a preset threshold with the current keyframe, and these candidate keyframes are then associated with the current keyframe.

4. The method for constructing a lane signal map based on crowdsourced data according to claim 1, characterized in that, Three-dimensional trajectory inequality constraints are constructed based on keyframe association results, including: , In the formula, k represents the trajectory. and These represent the three-dimensional translation and heading angles, respectively. Represents the Cauchy kernel function and cost function. This represents the three-dimensional spatial distance constraint between two keyframes of the loop closure. and Indicates the angle threshold. Indicates belonging to Summation of all constraints.

5. The method for constructing a lane signal map based on crowdsourced data according to claim 1, characterized in that, Construct global trajectory optimization constraints, including: , In the formula, k represents the trajectory. These represent the IMU's position, velocity, and attitude in the navigation coordinate system, respectively. These represent the accelerometer zero bias, gyroscope zero bias, and magnetometer zero bias of the current sub-trajectory, respectively. This represents the magnetic field vector in the navigation coordinate system. Constraints were defined within each sub-trajectory. This indicates the global average magnetic field constraint. This indicates the closure constraint between keyframes. This indicates the alignment point magnetic field vector constraint. This represents the sum of all constraints that are part of a loop constraint. This represents the sum of all constraints belonging to the alignment point magnetic field vector constraint. This represents the sum of all constraints that belong to the global average magnetic field constraint.

6. The method for constructing a lane signal map based on crowdsourced data according to claim 1, characterized in that, The lane signal map is obtained based on global trajectory optimization constraints, including: Lane magnetic sequences are extracted based on global trajectory optimization constraints to generate a lane signal map. The lane signal map includes the magnetic field sequence and magnetic field spatial distribution information corresponding to the lane, which is used for lane-level positioning and positioning enhancement.

7. The method for constructing a lane signal map based on crowdsourced data according to claim 1, characterized in that, The method employs deep representation learning to represent the features of magnetic field sequences based on vehicle motion trajectories, and also includes: The residuals of the magnetometer error model, absolute heading error, short-term magnetic disturbances, and observation noise are simulated based on the magnetic field sequence, and the simulation results are used to augment the magnetic field sequence.

8. A lane signal map construction system based on crowdsourced data, characterized in that, include: The Bayesian estimation module is used to estimate the acquired crowdsourced data using the Bayesian state estimation algorithm to obtain the vehicle trajectory, inertial navigation error model parameters, and magnetometer error model parameters. The keyframe association module is used to perform feature representation on the magnetic field sequence based on the vehicle motion trajectory using a deep representation learning method to obtain the magnetic field sequence features of the keyframes, and to perform keyframe association on the magnetic field sequence features of the keyframes using a loop closure detection method. The lane signal map module is used to construct three-dimensional trajectory inequality constraints based on keyframe association results. It combines inertial navigation error model parameters, magnetometer error model parameters, and lapsing constraints between keyframes to construct global trajectory optimization constraints, and obtains the lane signal map based on the global trajectory optimization constraints.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the lane signal map construction method based on crowdsourced data as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the lane signal map construction method based on crowdsourced data as described in any one of claims 1 to 7.

Images