CN114659514B - A LiDAR-IMU-GNSS fusion positioning method based on voxel-based precise registration - Google Patents

Abstract

The present invention discloses a LiDAR‑IMU‑GNSS fusion positioning method based on voxelized precise registration. Firstly, a voxelized point cloud downsampling method based on curvature segmentation is proposed, which performs coarse classification through curvature threshold, and uses hash mapping function to perform point cloud voxel downsampling, thereby retaining the spatial characteristic distribution attributes of the source point cloud to a greater extent. Secondly, a point cloud registration model based on the nearest neighbor of a point and a neighborhood point set is constructed, and an iteration termination threshold is set to reduce the probability of occurrence of the local optimal solution problem, thereby increasing the registration time of a single-frame point cloud by an order of magnitude. Finally, a LiDAR‑IMU‑GNSS fusion positioning model based on optimization is constructed, and the local pose estimation value is globally corrected using confidence-weighted GNSS observations, which can achieve more continuous and accurate pose estimation and map reconstruction on average in complex urban environments than similar advanced methods.

Description

A LiDAR-IMU-GNSS fusion positioning method based on voxel-based precise registration

Technical Field

The present invention belongs to the field of multi-sensor real-time positioning and mapping, and specifically relates to a LiDAR-IMU-GNSS fusion positioning method based on voxelized precise registration.

Background technique

For any autonomous robot system, accurate and robust mobile carrier navigation and positioning technology is one of the most basic technologies. For the positioning problem of mobile observation platforms. In recent years, multi-source fusion positioning technology based on SLAM (Simultaneous Localization and Mapping) has received widespread attention from relevant companies and researchers. The multi-source fusion positioning scheme based on SLAM can be divided into two types according to the main sensors: vision-based SLAM and lidar-based SLAM. Due to the superiority of the sensor, the lidar-based SLAM scheme can collect spatial fingerprint information more frequently and more accurately, and is more suitable for the real-time high-precision positioning requirements of vehicle-mounted platforms. Therefore, the lidar-based SLAM scheme represented by LiDAR-Inertial Odometry (LIO) is widely used to obtain three-dimensional geographic information in complex environments and for carrier positioning and map reconstruction.

The point cloud registration of LiDAR is a key step in the pose estimation of mobile carriers, which strictly affects the pose estimation and map reconstruction results of the positioning method. Common point cloud filter registration methods include Normal Distributions Transform (NDT), Iterative Closest Point (ICP) and various improved methods. The core of the NDT method is to use the probability density function of the source point cloud and the target point cloud as the objective function, and use the nonlinear optimization method to minimize the probability density function between the two to obtain the optimal solution. Although it has better real-time performance, it requires multiple points to construct the covariance matrix and has low robustness in sparse point cloud areas. As another method of point cloud registration, the ICP method has higher positioning accuracy, but it must re-search the nearest neighbor points and obtain the transformation matrix in each iteration process. Its high computational cost makes it difficult to adapt to the vehicle platform with limited computing resources. In summary, on the basis of compressing the computational cost, a high-precision real-time point cloud registration method suitable for vehicle platforms still needs to be studied.

In addition, as a local sensor, there is a cumulative offset between the local map and the global map when LIO estimates the pose of the current frame, which limits the positioning accuracy of the LIO positioning and mapping solution in large outdoor environments. Global observation information from the Global Navigation Satellite System (GNSS) is needed to provide a reliable global constraint correction for the LIO system. Traditional fusion methods often introduce three-dimensional GNSS measurement factors to assist LIO in the optimization framework, but the measurement information of a single key frame is relatively redundant, and the reliability of the GNSS factors added when driving to the GNSS multipath area is poor. In order to reduce data redundancy, existing methods propose to screen GNSS corner point factors to constrain local poses, but they do not consider the situation where the number of corner points on straight sections is insufficient, the measurement information utilization rate is low, and its application in large and complex outdoor environments is limited.

From the above analysis, it can be seen that there is an urgent need to propose a LIO-GNSS fusion solution suitable for vehicle-mounted platforms. The main technical points include: (1) Realizing real-time and high-precision point cloud registration on the basis of compressing the computational cost. (2) Using GNSS to perform global cumulative error correction on LIO on the basis of making full use of GNSS measurement information.

Summary of the invention

To solve the above problems, the present invention discloses a LiDAR-IMU-GNSS fusion positioning method based on voxelized precise registration, which achieves higher positioning accuracy than other similar methods in actual urban environments, and can provide accurate and continuous pose estimation results for vehicle-mounted platforms during driving in complex urban environments.

To achieve the above object, the technical solution of the present invention is as follows:

A LiDAR-IMU-GNSS fusion positioning method based on voxel-based precise registration includes the following steps:

(1) A voxel-based downsampling method based on curvature segmentation is adopted: given a set of original point cloud sequences collected by LiDAR, all points in the original point cloud sequence are traversed and coarse clustering is performed using the breadth-first method;

Specifically, the geometric angle threshold based on Euclidean distance is used to finely segment point cloud clusters with similar depths. Assume that the laser radar scanning center is O, and the two similar edge points p _a and p _b in the point cloud cluster have depths d _a and d _b (d _a > d _b ), respectively. Assume that the number of point clouds in the point cloud cluster where point p _i is located is M, then the roughness of point cloud p _i is:

Set the roughness threshold to Traverse M and put/> The points are classified as edge feature points, and the /> The points are classified as plane feature point sets, and downsampling operations are performed separately. Then, the hash function mapping value is used instead of random sampling to quickly downsample the point cloud clusters after the curvature rough classification; let the coordinates of a feature point in a set of point cloud sequences in the voxel space be p(x,y,z). Let the voxel grid size be r, then the dimension of the voxel grid in the x direction is _Dx = ( _xmax - _xmin )/r, and the x-direction index of p in the voxel grid is _hx = ( _xxmin )/r, and the same applies to the y and z directions.

After obtaining the three-dimensional index of the feature point in the voxel space, the hash function is used to quickly sort the feature point index and map it to N containers (N=80). The hash mapping function is:

hash( _hx , _hy , _hz )=( _hx + _hy · _Dx + _hz · _Dx · _Dy )%N (2)

To avoid hash conflicts, the conflict detection condition is set as:

hash(h _x ,hy _y ,h _z )＝hash(h' _x ,h' _y ,h' _z )(h _x ≠h' _x |hy ≠h' _y |h _{z ≠} h' _z ) (3)

Once a hash conflict is detected, the index value in the current container is output and the container is cleared, and the new index value is placed in the container.

(2) Point cloud registration based on smooth voxelization

The method of solving T based on the distribution of feature points within voxels is used to transform the problem of constructing the nearest neighbor model of a single point pair using a dendrogram into the nearest neighbor model of a point and a set of neighborhood points. First, the two sets of point cloud sequences are approximated as Gaussian distributions, that is, where i∈(1,n),/> and/> are the covariance matrices of two sets of point cloud sequences respectively; let the neighborhood point set of a _i be/> Among them, λ is the neighborhood judgment threshold, from which the rigid body transformation error is calculated/> for:

This step realizes the smoothing of all neighboring point clouds in the neighborhood of _ai ; thus, the maximum likelihood estimate of the inter-frame pose transformation matrix T can be inferred as follows, where _Ni is the number of point clouds in the neighborhood of _ai .

In addition to smoothing all neighboring point clouds in the neighborhood of _ai , in order to avoid falling into the local optimal blind area after multiple iterations, the iteration termination threshold ε is set as follows:

|RMSE _k+1 -RMSE _k |＞ε (6)

Among them, RMSE _k+1 and RMSE _k are the root mean square errors of the first k+1 iterations and the first k iterations, respectively. The iteration is completed when the absolute value of the change in the root mean square error |RMSE _k+1 -RMSE _k |≤ε or the maximum number of iterations is reached.

(3) Adopt a method suitable for the unification of the LiDAR-IMU-GNSS multi-sensor pose space;

set up is the external parameter conversion matrix from IMU to GNSS,/> is the external parameter conversion matrix from laser radar to IMU; since the hardware is fixed to the mobile carrier, both are constant after calibration; introduce spatial conversion parameters/> (/> Include translation parameters and rotation parameters/> ) associates two pose spaces; the problem of unifying the multi-source pose spaces at time t is expressed as:

In the formula, The initial value is set to the unit matrix. Each time the GNSS factor is added to solve the global optimal solution, the next moment will be updated. Value, correct the accumulated offset between the local coordinate system and the global coordinate system;

(4) Global map construction

Using graph optimization methods, we regard the construction of the global map as a maximum a posteriori estimation problem and perform nonlinear optimization on the state vector within the sliding window. The global optimization function is constructed as follows:

Where, ρ is the GNSS confidence; For the current global point cloud/> Local point cloud derived from local matching between frames/> The pose conversion matrix between them; the specific meaning of each sensor cost function in the formula is as follows;

① LiDAR-Inertial Odometry Factor: Based on the local pose estimation result of LiDAR-Inertial Odometry, the carrier position in the local coordinate system at time t can be obtained. and rotation/> From this, the local residual factor can be constructed as follows:

②GNSS factor: Assume that the time interval between two frames of GNSS observations is Δt, and the time alignment is achieved with the LIO pose estimation value by interpolation. Now, the GNSS measurement value in the ENU coordinate system is given as LIO pose observations in the global coordinate system Then the GNSS residual factor is expressed as follows:

When the carrier moves to an area with good GNSS signals, in order to fully and reliably utilize GNSS observations, a GNSS factor is added with GNSS confidence as a weight. The GNSS confidence is determined by the number of visible and effective GNSS satellites.

③ Voxelized loop closure factor: Considering the possible overlap of the driving area of the mobile carrier, it is necessary to add a loop detection link to establish the loop constraints between non-adjacent frames; the loop closure factor can be constructed according to formula (5) as follows:

Beneficial effects of the present invention:

The present invention proposes a LiDAR-IMU-GNSS fusion positioning method based on voxelized precise registration. First, a voxelized point cloud downsampling method based on curvature segmentation is proposed. The curvature threshold is used for rough classification, and the point cloud voxel downsampling is performed using a hash mapping function, which retains the spatial feature distribution properties of the source point cloud to a greater extent. Secondly, a point cloud registration model based on the nearest neighbor of a point and a neighborhood point set is constructed, and an iteration termination threshold is set to reduce the probability of the occurrence of the local optimal solution problem. The registration time of a single frame point cloud is increased by an order of magnitude. Finally, a fusion positioning model based on graph optimization is constructed, and the local pose of the LIO is globally corrected using GNSS observations based on confidence weighting. The experimental results show that in a large-scale measured outdoor environment, the root mean square error of the absolute positioning trajectory error of this method is an average of 2.317m, which is higher than the accuracy of similar methods. The experimental results fully prove that the method of the present invention can achieve more continuous and more accurate pose estimation and map reconstruction than similar advanced methods in a complex urban environment.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a schematic diagram of the unified multi-sensor pose space;

Figure 2 is a box plot comparing the absolute trajectory errors of the positioning of the present system and similar methods;

Figure 3 is a diagram showing the mapping effect of this system.

Detailed ways

The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

As shown in the figure, the LiDAR-IMU-GNSS fusion positioning method based on voxel-based precise registration described in the present invention is specifically as follows:

(1) Voxel-based downsampling based on curvature segmentation

The present invention proposes a voxelized downsampling method based on curvature segmentation. Given a set of original point cloud sequences collected by a laser radar, all points in the original point cloud sequence are traversed and coarse clustering is performed using the breadth-first method. Furthermore, a geometric angle threshold based on Euclidean distance is used to finely segment point cloud clusters with similar depths. Assume that the laser radar scanning center is O, and the two close edge points p _a and p _b in the point cloud cluster have depths d _a and d _b (d _a > d _b ), respectively. Assume that the number of point clouds in the point cloud cluster where point p _i is located is M, then the roughness of point cloud p _i is:

Set the roughness threshold to Traverse M and put/> The points are classified as edge feature points, and the /> The points are classified as plane feature point sets and down-sampled respectively. The present invention uses hash function mapping values instead of random sampling to quickly downsample the point cloud clusters after rough curvature classification. Suppose the coordinates of a feature point in a set of point cloud sequences in the voxel space are p(x, y, z). Suppose the voxel grid size is r, then the dimension of the voxel grid in the x direction is D _x =(x _max -x _min )/r, and the x-direction index of p in the voxel grid is h _x =(xx _min )/r, and the same applies to the y and z directions.

After obtaining the three-dimensional index of the feature point in the voxel space, the hash function is used to quickly sort the feature point index and map it to N containers (N=80). The hash mapping function is:

hash( _hx , _hy , _hz )=( _hx + _hy · _Dx + _hz · _Dx · _Dy )%N (2)

To avoid hash conflicts, the conflict detection condition is set as:

hash(h _x ,hy _y ,h _z )＝hash(h' _x ,h' _y ,h' _z )(h _x ≠h' _x |hy ≠h' _y |h _{z ≠} h' _z ) (3)

Once a hash conflict is detected, the index value in the current container is output and the container is cleared, and the new index value is placed in the container.

(2) Point cloud registration based on smooth voxelization

This paper proposes a method to solve T based on the distribution of feature points within voxels, which transforms the problem of constructing a nearest neighbor model for a single point pair using a dendrogram into constructing a nearest neighbor model for a point and a set of neighboring points. First, the two sets of point cloud sequences are approximated as Gaussian distributions, that is, where i∈(1,n),/> and/> are the covariance matrices of two sets of point cloud sequences. Let the neighborhood point set of a _i be/> Among them, λ is the neighborhood judgment threshold, which is used to calculate the rigid body transformation error for:

This step realizes the smoothing of all neighboring point clouds in the neighborhood of _ai . From this, the maximum likelihood estimate of the inter-frame pose transformation matrix T can be inferred as follows, where _Ni is the number of point clouds in the neighborhood of _ai .

In addition to smoothing all neighboring point clouds in the neighborhood of _ai , in order to avoid falling into the local optimal blind area after multiple iterations, the iteration termination threshold ε is further set as follows:

|RMSE _k+1 -RMSE _k |＞ε (6)

Where RMSE _k+1 and RMSE _k are the root mean square errors of the first k+1 iterations and the first k iterations, respectively. The iteration is completed when the absolute value of the change in the root mean square error |RMSE _k+1 -RMSE _k |≤ε or the maximum number of iterations is reached.

(3) Multi-sensor pose space unification

The present invention proposes a method suitable for unifying the position and posture space of multiple sensors such as LiDAR, IMU and GNSS. is the external parameter conversion matrix from IMU to GNSS,/> is the external parameter conversion matrix from the laser radar to the IMU. Since the hardware is fixed to the mobile carrier, both are constant after calibration. Introducing the spatial conversion parameter/> (including translation parameters/> and rotation parameters ) associates two pose spaces. The problem of unifying multi-source pose spaces at time t can be expressed as:

In the formula, The initial value is set to the unit matrix. Each time the GNSS factor is added to solve the global optimal solution, the next moment will be updated. Value, correct the accumulated offset between the local coordinate system and the global coordinate system.

(4) Global map construction

The present invention uses graph optimization methods to regard the construction of a global map as a maximum a posteriori estimation problem and performs nonlinear optimization on the state vector within the sliding window. The global optimization function is constructed as follows:

Where ρ is the GNSS confidence. For the current global point cloud/> Local point cloud derived from local matching between frames/> The specific meaning of each sensor cost function in the formula is as follows.

① LiDAR-Inertial Odometry Factor: Based on the local pose estimation result of LiDAR-Inertial Odometry, the carrier position in the local coordinate system at time t can be obtained. and rotation/> From this, the local residual factor can be constructed as follows:

②GNSS factor: Assume that the time interval between two frames of GNSS observations is Δt, and the LIO pose estimation value is interpolated to achieve time alignment. Now given the GNSS measurement value in the ENU coordinate system LIO pose observations in the global coordinate system Then the GNSS residual factor is expressed as follows:

When the carrier moves to an area with good GNSS signals, in order to fully and reliably utilize GNSS observations, a GNSS factor is added with GNSS confidence as a weight. The GNSS confidence is determined by the number of visible and effective GNSS satellites.

③ Voxelized loop closure factor: Considering the possible overlap of the driving area of the mobile carrier, it is necessary to add a loop detection link to establish the loop constraints between non-adjacent frames. According to formula (5), the loop closure factor can be constructed as follows:

The positioning accuracy of the technical solution of the present invention is quantitatively evaluated based on public data set experiments. The positioning accuracy of the fusion method is evaluated using the KITTI_RAW data set and compared with other similar advanced methods. This paper selects four different outdoor scenes to verify the performance of the fusion method, including urban environments, open areas, highways, and forest roads. The voxel grid size of the fusion method is set to 0.3×0.2×0.3, the maximum iteration threshold is set to 30, and the iteration termination tolerance threshold is set to 1e-8, so as to ensure that the number of feature point clouds participating in the matching in the sparse area of the outdoor point cloud is stable while meeting the real-time requirements. The evaluation strategy is to compare the true value of the trajectory with the pose estimation results output by different methods, and obtain the root mean square error (APE_RMSE) value of the absolute trajectory error.

The test results of the data set are shown in the following table:

Table 1 Absolute trajectory error (ATE_RMSE(m)) of each method in the KITTI dataset

As can be seen from the above table, thanks to the precise point cloud registration and the multi-sensor weighted nonlinear optimization framework that introduces GNSS global constraints, the average root mean square error of the absolute trajectory error of this method is about 2.317m, which is significantly lower than the positioning errors of the above three similar advanced methods, and the positioning accuracy is better than the above three advanced methods. From the experimental results, it can be seen that the global optimization link in the method of the present invention has well completed the role of suppressing local cumulative drift, making the pose estimation results closer to the global optimal solution. And because of the inherent advantage of local sensors that are not affected by the signal refraction and reflection environment, it makes up for the positioning anomalies caused by the multipath effect of traditional GNSS positioning in urban environments, and ensures the complete reliability of the fusion system positioning.

The test results show that the LiDAR-IMU-GNSS fusion positioning method based on voxelized precise registration proposed in the present invention achieves higher positioning accuracy than other similar methods in actual urban environments, and can provide accurate and continuous pose estimation results for vehicle-mounted platforms during driving in complex urban environments.

Claims (5)

1. A LiDAR-IMU-GNSS fusion positioning method based on voxel precise registration comprises the following steps:

(1) The voxel downsampling method based on curvature segmentation is adopted: giving an original point cloud sequence acquired by a group of laser radars, traversing all points in the original point cloud sequence, and performing coarse clustering by using a breadth-first method;

Specifically, the point cloud clusters with similar depths are finely segmented by utilizing a geometric angle threshold value based on Euclidean distance; let the laser radar scan center be O, two similar edge points p _a、p_b in the point cloud cluster, the depth is d _a、d_b,d_a＞d_b respectively, the point cloud number in the point cloud cluster where the set point p _i is located is M, and then the p _i point cloud roughness is:

Setting the roughness threshold as Traversing M, will/>Points of (1) are classified as edge feature point sets, will/>The points of the curve are classified into a plane characteristic point set, downsampling operation is respectively carried out, and then a hash function mapping value is used for replacing random sampling to rapidly downsample the point cloud cluster after the curvature coarse classification; setting the coordinates of a certain characteristic point in a group of point cloud sequences in a voxel space as p (x, y, z); setting the size of the voxel grid as r, wherein the dimension of the voxel grid in the x direction is D _x＝(x_max-x_min)/r, and the x direction index of p in the voxel grid is h _x＝(x-x_min)/r, and the y and z directions are the same;

after three-dimensional indexes of the feature points in the voxel space are obtained, the feature point indexes are rapidly ordered by adopting a hash function, the feature point indexes are mapped into N containers, N=80, and the hash mapping function is as follows:

hash(h_x,h_y,h_z)＝(h_x+h_y·D_x+h_z·D_x·D_y)％N (2)

in order to avoid hash collision, the collision detection conditions are set as follows:

hash(h_x,h_y,h_z)＝hash(h'_x,h'_y,h'_z) (h_x≠h'_x|h_y≠h'_y|h_z≠h'_z) (3)

Outputting the index value in the current container and emptying the container once the hash collision is detected, and placing a new index value into the container;

(2) Point cloud registration based on smoothing voxel processing

Adopting a method for solving T based on characteristic point distribution in voxels, and converting a problem of constructing a single-point nearest neighbor model by using a dendrogram into a nearest neighbor model of a construction point and a neighborhood point set; first, two sets of point cloud sequences are approximated as gaussian distributions, i.e Wherein i is E (1, n),/>And/>Respectively two groups of point cloud sequence covariance matrixes; let the neighborhood point set of a _i be/>Wherein λ is a neighborhood decision threshold, thereby calculating rigid body transformation error/>The method comprises the following steps:

The step realizes the smooth processing of all adjacent point clouds in the a _i neighborhood; the maximum likelihood estimation of the pose transformation matrix T of the frame can be estimated as follows, wherein N _i is the number of point clouds in a _i neighborhood;

in addition to smoothing all neighboring point clouds in the a _i neighborhood, to avoid trapping in a locally optimal blind zone after multiple iterations, an iteration termination threshold epsilon is set as follows:

|RMSE_k+1-RMSE_k|＞ε (6)

Wherein, RMSE _k+1 and RMSE _k are root mean square errors of the previous k+1 iterations and the previous k iterations respectively, and when the absolute value of the variation of the root mean square errors is |rmse _k+1-RMSE_k |ε or reaches the maximum number of iterations, the iterations are completed;

(3) The method suitable for unifying the pose space of the laser radar-IMU-GNSS multi-sensor is adopted;

Is provided with For IMU to GNSS extrinsic conversion matrix,/>An external parameter conversion matrix for the laser radar reaching the IMU; because the hardware is fixedly connected with the mobile carrier, the two are fixed values after calibration; introducing spatial conversion parameters/>Associating two pose spaces; /(I)Including translation parameters/>And rotation parameter/>The problem of multi-source pose space unification at the time t is expressed as follows:

In the method, in the process of the invention, Setting the initial value as a unit array, and updating the/>, at the next moment after adding GNSS factors to solve the global optimal solution each timeA value correcting an accumulated offset between the local coordinate system and the global coordinate system;

(4) Global map construction

The method comprises the steps that a map optimization method is utilized, a global map is constructed as a maximum posterior estimation problem, and nonlinear optimization is carried out on state vectors in a sliding window; the global optimization function is constructed as follows:

Wherein ρ is GNSS confidence; For the current global point cloud/> Local point cloud derived after local matching with inter-frameA pose conversion matrix between the two; the specific meaning of each sensor cost function in the formula is as follows;

① Lidar-inertial odometer factor: the carrier position under the local coordinate system at the moment t can be obtained according to the local pose estimation result of the laser radar-inertial odometer And rotation/>The local residual factor can thus be constructed as follows:

② GNSS factors: setting the time interval of two frames of GNSS observation values as delta t, and realizing time alignment with LIO pose estimation values in an interpolation mode; GNSS measurements in an ENU coordinate system are now given Pose observations of LIO in global coordinate System/>The GNSS residual factor is expressed as follows:

When the carrier moves to a GNSS signal good area, GNSS factors are added by taking GNSS confidence as a weight in order to fully and reliably utilize GNSS observation values, wherein the GNSS confidence is determined by the number of visible effective satellites of the GNSS;

③ Voxelized loop-back factor: considering the possible coincidence of the running areas of the mobile carrier, a loop detection link needs to be added to establish loop constraint existing between non-adjacent frames; the loop back factor can be constructed according to equation (5) as follows:

2. The LiDAR-IMU-GNSS fusion positioning method based on voxel precise registration according to claim 1, wherein the point cloud voxel downsampling method based on curvature segmentation in the step (1) is characterized in that firstly coarse classification is carried out through a curvature threshold value, then the point cloud voxel downsampling is carried out by utilizing a Hash mapping function instead of a random sampling consistency method, and the spatial feature distribution attribute of source point clouds is reserved to a greater extent.

3. The LiDAR-IMU-GNSS fusion positioning method based on voxel-based fine registration of claim 1, wherein the point cloud registration based on smooth voxel-based processing in the step (2) performs smooth processing on the point cloud sequence aiming at single-to-multiple distribution in the point cloud sequence, a point cloud registration model based on the nearest neighborhood of the point and neighborhood point set is constructed, and an iteration termination threshold is set to reduce the occurrence probability of a local optimal solution problem.

4. The LiDAR-IMU-GNSS fusion positioning method based on voxel-based fine registration according to claim 1, wherein in the multi-sensor pose space unification method in the step (3), a pose estimation result of a laser radar and an IMU in a local map and a GNSS measurement value in a global map are spatially unified by a lever arm compensation method, and a pose conversion matrix at the next moment is updated after a GNSS factor is added to solve a global optimal solution each time, so that the accumulated offset between a local coordinate system and the global coordinate system is corrected.

5. The LiDAR-IMU-GNSS fusion positioning method based on voxel-based fine registration of claim 1, wherein in the step (4), a global map construction framework based on graph optimization is adopted, and weighted GNSS global position observations and voxel loop factors are used as global constraints to inhibit the accumulated drift of a local sensor; and optimizing and correcting the historical track by utilizing an optimized smooth voxel point cloud registration method through registration between the point cloud of the prior local map and the current global point cloud, so that the pose estimated value is converged to the global optimal result.