A global visual prior map construction method based on vector road network prior non-rigid correction

Abstract

The application discloses a kind of global visual prior map construction methods based on vector road network prior non-rigid correction, belong to the cross field of computer vision, mobile robot navigation and automatic driving technology, for global visual prior map construction, including obtaining original image and constructing original visual map, obtaining the vector road network data corresponding to the trajectory coverage area of original visual map, extracting control point and establishing matching relationship;Non-rigid change model is established based on matching relationship, and corrected position is obtained by combining orthogonal projection method, the corrected heading angle is obtained according to the displacement direction change before and after adjacent map frame correction, and finally global visual prior map is generated.The application introduces vector road network prior and combines non-rigid correction, eliminates the cumulative drift of visual odometry, improves the global accuracy of prior map under the premise of keeping visual feature mode consistent, and reduces the construction and application cost.

Description

Technical Field

[0001] This invention discloses a global visual prior map construction method based on vector road network prior non-rigid correction, which belongs to the interdisciplinary field of computer vision, mobile robot navigation and autonomous driving technology. Background Technology

[0002] Visual SLAM and visual inertial odometry have become important localization and mapping methods for mobile robots, autonomous vehicles, and intelligent sensing systems due to their low sensor cost, flexible deployment, and strong environmental adaptability. These methods typically estimate the motion state of the vehicle through camera observations and inertial measurement information, and then construct an environmental map based on this. However, visual SLAM or visual inertial odometry inherently relies primarily on relative observations, and their mapping results often lack a stable and low-cost global absolute reference. As running time and distance increase, cumulative drift easily occurs, leading to significant deviations between the trajectory and map and the actual geographic space.

[0003] To improve positioning accuracy and global consistency, existing technologies typically employ two approaches. The first approach introduces global sensors such as GNSS to provide absolute position constraints. However, this approach is prone to signal weakening, multipath effects, and intermittent loss of lock-on in scenarios such as tunnels, urban canyons, and densely obstructed areas, making long-term stable operation difficult. The second approach pre-builds a high-precision prior map and performs map matching and positioning during the online phase. This approach usually relies on LiDAR, high-precision map acquisition equipment, or large-scale Structure-from-Motion reconstruction, and suffers from high mapping costs, complex data acquisition and maintenance, and difficulty in low-cost, large-scale deployment.

[0004] In contrast, raw visual maps directly generated by visual SLAM or visual inertial odometry are naturally modally consistent with the visual sensors used in the online localization phase, making them more suitable as visual prior maps. However, these raw visual maps suffer from cumulative drift, and their errors typically manifest as non-rigid deformations with uneven spatial distribution. These include not only large-scale positional shifts but also local geometric distortions and attitude deviations. Traditional rigid body registration or single global similarity transformations are insufficient to effectively correct these non-rigid errors, making it difficult to directly convert raw visual maps into high-quality prior maps suitable for global localization.

[0005] Meanwhile, publicly available vector road network data in road scenarios has advantages such as low cost, wide coverage, and clear geometric structure, which can provide effective global geometric priors for vehicle trajectories. Therefore, using vector road network priors to perform non-rigid corrections on the original visual map generated by the vision system, thereby constructing a low-cost, globally consistent visual prior map suitable for subsequent localization and fusion, has become an important solution for prior map construction. Summary of the Invention

[0006] The purpose of this invention is to provide a global visual prior map construction method based on vector road network prior non-rigid correction, so as to solve the problems of modal inconsistency and high construction cost between prior maps and online visual positioning systems in the prior art.

[0007] A method for constructing a global visual prior map based on vector road network prior non-rigid correction includes: S1. Acquire the original images, and use SLAM to perform pose estimation and map construction on the original image sequence to obtain the original visual map. S2. Obtain vector road network data corresponding to the area covered by the original visual map trajectory, and represent the vector road network data as several polyline segments, nodes and node connection structures; S3. From the original visual map trajectory and vector road network, control points are extracted based on curvature scale space. A matching cost is constructed using a matching method based on shape context descriptors. The initial matching result is calculated using the optimal allocation algorithm. Matching control point pairs are eliminated using consistency checks to obtain the control point matching relationship between the original visual map trajectory and the vector road network. S4. Based on the control point matching relationship between the original visual map trajectory and the vector road network, establish a thin plate spline non-rigid transformation model from the original visual map trajectory to the vector road network prior, and use the transformation model to perform non-rigid correction on the position of each map frame in the original visual map to obtain the position of the map frame after preliminary correction. S5. Project the pre-corrected map frame position orthogonally onto the corresponding or nearest vector road network segment in the map plane to obtain the corrected position; S6. Calculate the heading correction amount based on the change in displacement direction before and after correction of adjacent map frames, update the heading angle in the attitude of the map frame based on the heading correction amount, and obtain the corrected heading angle. S7. Generate a global visual prior map using the map frame pose after position correction and heading angle correction and the corresponding visual feature description information.

[0008] S1 includes acquiring raw images and constructing raw visual maps: ; ; In the formula, The original visual map, For the first One map frame, This represents the total number of map frames. For map frame location, This refers to the pose of the map frame. This refers to visual feature description information associated with the map frame; It should include at least normalized feature coordinates, feature descriptors, pixel coordinates, and corresponding landmark information.

[0009] S3 includes S3.1, extracting control points from the original visual map trajectory and vector road network based on curvature scale space, establishing a matching cost matrix using a shape context descriptor-based matching method, calculating the initial matching results using the Hungarian algorithm, and filtering control point pairs using the RANSAC method to obtain the control point matching relationship between the original visual map trajectory and the vector road network; the original visual map trajectory is a spatial geometric path formed by connecting the positions of each map frame in chronological order.

[0010] S3 includes S3.2, which introduces rotation invariance and scale invariance during the construction of control point descriptors; Rotation invariance includes: taking the current control point as the center, calculating the relative position vectors from the current control point to other sampling points on the same curve, forming a set of position vectors; performing principal component analysis on the set of position vectors, and extracting the direction of the first principal component as the reference direction when constructing the descriptor; Scale invariance includes normalizing the relative distances between control points.

[0011] S4 includes, S4.1, where the control points on the original visual map trajectory are... The corresponding vector road network control points are , Then for any original map frame location The corresponding non-rigid transformation of the thin plate spline is: ; ; ; In the formula, To map the target location, For mapping the target location, the first Each coordinate component , , , For affine terms, For the first The control point at the ... The weighting coefficients of each coordinate component. For radial basis functions, To control the total number of point pairs, For the index number of the control point pair, For the Euclidean norm, , , for The three-dimensional coordinates.

[0012] S4 includes S4.2, constructing the kernel matrix and affine term matrix based on all control point pairs, and simultaneously solving the control point interpolation constraints and smoothing constraints to form a system of linear equations, and obtaining... , , , and ; Will Input the transformation model one by one to obtain the map frame positions after preliminary correction.

[0013] S5 includes searching for the corresponding or nearest road vector line segment in the horizontal plane for the initially corrected map frame position, and orthogonally projecting the initially corrected map frame position onto the corresponding or nearest road vector line segment.

[0014] S6 includes setting the displacement vector between adjacent map frames before correction as... The displacement vector between adjacent map frames after correction is Then the heading correction amount for: ; In the formula, The cross product symbol is used. The dot product symbol; Corrected heading angle for: ; In the formula, This is the original heading angle.

[0015] S7 includes a global visual prior map, comprising the corrected position, corrected pose, feature-normalized coordinates, and feature descriptors for each map frame.

[0016] Compared with existing technologies, this invention has the following advantages: This invention utilizes raw visual maps generated by visual SLAM or visual inertial odometry, and introduces publicly available vector road network data as a global geometric prior. It can complete the construction of a global visual prior map without relying on high-cost LiDAR, high-precision map acquisition equipment, or large-scale 3D reconstruction systems, offering advantages such as low cost and easy deployment. Through a coarse-to-fine correction strategy combining control point matching, thin-plate spline non-rigid transformation, local projection refinement, and heading correction, it can simultaneously suppress large-scale positional deviations, small-scale local geometric distortions, and attitude deviations, making it more suitable for handling the non-uniform cumulative drift formed by visual maps during long-distance travel. The obtained global visual prior map directly inherits the visual feature representation of the original visual map, and is naturally compatible with the feature extraction, matching and relocalization processes used in the online visual localization stage, which is conducive to improving the usability and effectiveness of the prior map in subsequent localization and fusion. Based on the relative relationship between the map frame pose and its local visual description information, only global non-rigid correction of the map frame pose is needed to achieve consistent updates of related visual information, avoiding the high complexity problem caused by point-by-point registration of all map points or dense point clouds. This invention is particularly suitable for the construction of global visual prior maps in urban roads, park roads, open roads and vehicle trajectory scenarios, and has good engineering application value. Attached Figure Description

[0017] Figure 1 This is a flowchart of the technology of this invention; Figure 2 This is a schematic diagram of the extraction and matching of control points from the original visual map trajectory and the vector road network. Figure 1 ; Figure 3 This is a schematic diagram of the extraction and matching of control points from the original visual map trajectory and the vector road network. Figure 2 . Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] A method for constructing a global visual prior map based on vector road network prior non-rigid correction includes: S1. Acquire the original images, and use SLAM to perform pose estimation and map construction on the original image sequence to obtain the original visual map. S2. Obtain vector road network data corresponding to the area covered by the original visual map trajectory, and represent the vector road network data as several polyline segments, nodes and node connection structures; S3. From the original visual map trajectory and vector road network, control points are extracted based on curvature scale space. A matching cost is constructed using a matching method based on shape context descriptors. The initial matching result is calculated using the optimal allocation algorithm. Matching control point pairs are eliminated using consistency checks to obtain the control point matching relationship between the original visual map trajectory and the vector road network. S4. Based on the control point matching relationship between the original visual map trajectory and the vector road network, establish a thin plate spline non-rigid transformation model from the original visual map trajectory to the vector road network prior, and use the transformation model to perform non-rigid correction on the position of each map frame in the original visual map to obtain the position of the map frame after preliminary correction. S5. Project the pre-corrected map frame position orthogonally onto the corresponding or nearest vector road network segment in the map plane to obtain the corrected position; S6. Calculate the heading correction amount based on the change in displacement direction before and after correction of adjacent map frames, update the heading angle in the attitude of the map frame based on the heading correction amount, and obtain the corrected heading angle. S7. Generate a global visual prior map using the map frame pose after position correction and heading angle correction and the corresponding visual feature description information.

[0020] S1 includes acquiring raw images and constructing raw visual maps: ; ; In the formula, The original visual map, For the first One map frame, This represents the total number of map frames. For map frame location, This refers to the pose of the map frame. This refers to visual feature description information associated with the map frame; It should include at least normalized feature coordinates, feature descriptors, pixel coordinates, and corresponding landmark information.

[0021] S3 includes S3.1, extracting control points from the original visual map trajectory and vector road network based on curvature scale space, establishing a matching cost matrix using a shape context descriptor-based matching method, calculating the initial matching results using the Hungarian algorithm, and filtering control point pairs using the RANSAC method to obtain the control point matching relationship between the original visual map trajectory and the vector road network; the original visual map trajectory is a spatial geometric path formed by connecting the positions of each map frame in chronological order.

[0022] S3 includes S3.2, which introduces rotation invariance and scale invariance during the construction of control point descriptors; Rotation invariance includes: taking the current control point as the center, calculating the relative position vectors from the current control point to other sampling points on the same curve, forming a set of position vectors; performing principal component analysis on the set of position vectors, and extracting the direction of the first principal component as the reference direction when constructing the descriptor; Scale invariance includes normalizing the relative distances between control points.

[0023] S4 includes, S4.1, where the control points on the original visual map trajectory are... The corresponding vector road network control points are , Then for any original map frame location The corresponding non-rigid transformation of the thin plate spline is: ; ; ; In the formula, To map the target location, For mapping the target location, the first Each coordinate component , , , For affine terms, For the first The control point at the ... The weighting coefficients of each coordinate component. For radial basis functions, To control the total number of point pairs, For the index number of the control point pair, For the Euclidean norm, , , for The three-dimensional coordinates.

[0024] S4 includes S4.2, constructing the kernel matrix and affine term matrix based on all control point pairs, and simultaneously solving the control point interpolation constraints and smoothing constraints to form a system of linear equations, and obtaining... , , , and ; Will Input the transformation model one by one to obtain the map frame positions after preliminary correction.

[0025] S5 includes searching for the corresponding or nearest road vector line segment in the horizontal plane for the initially corrected map frame position, and orthogonally projecting the initially corrected map frame position onto the corresponding or nearest road vector line segment.

[0026] S6 includes setting the displacement vector between adjacent map frames before correction as... The displacement vector between adjacent map frames after correction is Then the heading correction amount for: ; In the formula, The cross product symbol is used. The dot product symbol; Corrected heading angle for: ; In the formula, This is the original heading angle.

[0027] S7 includes a global visual prior map, comprising the corrected position, corrected pose, feature-normalized coordinates, and feature descriptors for each map frame.

[0028] The following description, in conjunction with the accompanying drawings, further illustrates the invention. This invention performs monocular-inertial SLAM on camera images and an inertial navigation unit to obtain an original visual map and an original map trajectory; it then processes the 2D road network and elevation data to obtain a 3D road network; it performs control point detection and matching on the original visual map, the original map trajectory, and the 3D road network, including CSS-based control point extraction and Shape Context-based control point matching, to obtain control point correspondences; subsequently, it performs non-rigid correction, including global distortion correction based on TPS transform, local distortion correction based on orthogonal projection, and attitude correction based on adjacent map frame vectors; finally, it obtains a corrected visual map and a corrected map trajectory based on map correction parameters.

[0029] The following is a further explanation with reference to an embodiment. This embodiment provides a method for constructing a global visual prior map based on non-rigid correction of a vector road network prior. This method is mainly aimed at the task of constructing a visual prior map for vehicles in road scenes. It uses visual SLAM or visual inertial odometry to generate an original visual map, and combines it with road vector network priors to perform non-rigid correction on the original visual map, finally constructing a global visual prior map. In this embodiment, a vehicle-mounted monocular camera and an inertial measurement unit are used to collect road scene data, and a visual inertial odometry is used to generate an original visual map; then, road vector network data corresponding to the trajectory coverage area is acquired; further, through control point extraction and matching, thin plate spline non-rigid correction, local orthogonal projection refinement, and heading correction, a global visual prior map is generated.

[0030] The first step is to construct the original visual map. First, a camera mounted on a mobile platform continuously acquires images of the road scene, and simultaneously acquires angular velocity and acceleration data output by the inertial measurement unit. The image sequence and inertial measurement data are then input into the visual inertial odometry algorithm to complete the relative motion estimation between adjacent frames and map construction, resulting in the original visual map arranged in chronological order.

[0031] In this embodiment, the original visual map consists of multiple map frames, each including at least its position, pose, and associated visual feature description information. To balance subsequent map application capabilities and storage efficiency, it is preferable to save the pose, normalized feature coordinates, and feature descriptor of each map frame; when 3D visualization or offline analysis is required, corresponding 3D landmark information can also be additionally saved. The original visual map can be represented as a set of map frames as described above, where each map frame contains its pose and associated local visual feature information.

[0032] It should be noted that the visual feature information is defined relative to the local coordinate system of the corresponding map frame. Therefore, in the subsequent correction process, it is not necessary to perform global registration point by point for all visual feature points. It is only necessary to correct the pose of each map frame to achieve the overall consistency update of the associated visual information.

[0033] The second step is the prior acquisition and preprocessing of the vector road network. After the original visual map is constructed, the road vector network data corresponding to the area covered by the original visual map is acquired. The vector road network data can come from public map platforms, existing road vector layers, or pre-built road topology databases, such as information on road centerlines, road polylines, and intersection topological relationships.

[0034] To facilitate subsequent control point extraction and matching, this embodiment preprocesses the vector road network data. Specifically, the vector road network is represented as several polyline segments, nodes, and their connections, and the road centerlines are sampled and smoothed to enable geometric comparison with the original visual map trajectory under a unified planar representation. Here, the vector road network does not participate in the visual front-end pose estimation but serves as an external geometric prior after the original visual map is constructed, participating in constraints only during the map correction stage.

[0035] The third step involves extracting and matching control points between the original visual map trajectory and the vector road network. To establish the geometric correspondence between the original visual map trajectory and the vector road network, control points need to be extracted from both, forming a control point correspondence. The control point extraction and matching process is illustrated below. Figure 2 and Figure 3 As shown.

[0036] In this embodiment, geometrically significant points, such as corner points, intersection points, curvature abrupt change points, and key points of polylines, are preferably extracted from the original visual map trajectory and vector road network curves as control points. Since the original trajectory and road vector data primarily provide geometric contour information and lack strong visual features such as image texture, they are not suitable for directly using image feature point detection and matching methods. This embodiment preferably employs a control point detection method based on curvature scale space to extract control points from the trajectory curves and road network curves.

[0037] Specifically, the original trajectory curve and road polyline are smoothed to reduce the impact of noise and local small disturbances on curvature calculation. Then, the curve curvature is calculated at multiple scales, and significant curvature locations are detected as candidate control points. This enhances the stability of control points at different curve scales, ensuring that control points mainly fall at geometrically stable locations such as intersections and turns.

[0038] After obtaining candidate control points, to establish the correspondence between the control points of the original visual map trajectory and the control points of the vector road network, this embodiment preferably adopts a control point matching method based on shape context. Specifically, taking each candidate control point as a reference point, its relative spatial distribution with respect to other sampling points in the same curve is statistically analyzed to construct a local geometric descriptor for the reference point; then, a matching cost matrix is ​​established based on the distance between different control point descriptors; and finally, an optimal allocation algorithm is used to obtain the initial matching result. Preferably, the optimal allocation algorithm is the Hungarian algorithm.

[0039] To improve the robustness of control point matching, after obtaining the initial matching results, a random sampling consistency test method is further used to filter the correspondences of control points, removing mismatched control point pairs that are obviously inconsistent or do not meet geometric constraints. Preferably, the consistency test method is RANSAC.

[0040] Furthermore, to improve the matching stability of the original trajectory and the vector road network under different coordinate systems and scales, this embodiment introduces rotational invariance and scale invariance processing during the construction of control point descriptors. For rotational invariance, instead of using the tangent direction that relies solely on local geometry as the reference direction, the first principal direction is extracted using principal component analysis based on the set of relative position vectors formed from the reference control point to the other sampling points, and this is used as the reference direction for descriptor construction. For scale invariance, the relative distances between control points are normalized to ensure that the control point descriptors obtained at different scales remain comparable.

[0041] When the original visual map trajectory is long, the cumulative drift is large, and the automatic matching results are unstable, a small number of control point correspondences can be manually supplemented or corrected based on the automatic matching to enhance the reliability of subsequent non-rigid correction.

[0042] The fourth step is thin-plate spline non-rigid correction. After obtaining the correspondence between the control points of the original visual map trajectory and the control points of the vector road network, a thin-plate spline non-rigid transformation model is constructed from the original trajectory to the vector road network. This model performs large-scale non-rigid correction on the map frame positions in the original visual map, such as... Figure 3 As shown.

[0043] In this embodiment, the control points on the original visual map trajectory are set as follows: The corresponding vector road network control points are ,in For any map frame location to be corrected The thin-plate spline transformation function described in the aforementioned invention is used to map its position, thereby obtaining the preliminarily corrected map frame position. The thin-plate spline transformation model consists of affine terms and radial basis function terms, wherein the affine terms describe the overall translation, rotation, and scaling trends, and the radial basis function terms describe the local smooth non-rigid deformation induced by control points.

[0044] Furthermore, to solve the thin-plate spline transformation model, a kernel matrix and an affine term matrix can be constructed based on all control point pairs. A system of linear equations is then formed by simultaneously applying control point interpolation constraints and smoothing constraints to obtain the weight coefficients and affine term coefficients for each control point. After obtaining the transformation model parameters, the positions of each map frame in the original visual map are input into the model one by one to obtain the preliminary corrected position results.

[0045] In this embodiment, thin-plate spline transformation is mainly used to correct large-scale non-uniform positional deviations caused by long-distance cumulative drift from visual SLAM or visual inertial odometry, such as overall length deviations between adjacent turning intersections, overall positional misalignments, and slow bending deformations. Since visual feature information is locally defined relative to each map frame, only the position of each map frame needs to be corrected to achieve an overall geometric update of the associated map information, without the need for direct point-by-point non-rigid registration of all feature points.

[0046] The fifth step is local orthogonal projection refinement. Thin-plate spline transformation can significantly reduce the cumulative drift error of the original visual map at large scales. However, since control points are usually concentrated in geometrically significant areas such as corners and intersections, residual local lateral deviations may still exist in long, straight road sections between adjacent control points. To further improve the consistency between the corrected trajectory and road geometry, this embodiment performs local orthogonal projection refinement on the map frame positions based on the non-rigid thin-plate spline correction.

[0047] Specifically, for each map frame position after thin-plate spline transformation, its corresponding or nearest road vector line segment is searched in the horizontal plane, and the map frame position is orthogonally projected onto the road line segment to obtain a further refined map frame position. This step effectively eliminates the remaining lateral offset of the map frame position relative to the road centerline, making the corrected trajectory more closely fit the road network geometry.

[0048] In this embodiment, orthogonal projection is preferably performed only in planar coordinates, while elevation coordinates can retain the values ​​provided by the original visual map or inertial estimation, thereby avoiding the introduction of additional errors when road elevation is inaccurate or road network elevation information is missing.

[0049] The sixth step is heading correction. After the map frame position undergoes non-rigid correction using thin-plate splines and local orthogonal projection, the heading angle in the map frame attitude still retains the deviation from the original visual map. Therefore, this embodiment further corrects the heading angle in the map frame attitude based on the changes in the displacement direction of adjacent map frames before and after correction.

[0050] Specifically, two adjacent map frames are taken, and displacement vectors are constructed between the two frames before correction and between the two frames after correction. Since the motion direction between the two frames before and after correction can be approximately considered to satisfy the planar rigidity direction change relationship within a small local range, the heading correction amount can be obtained based on the angle between these two displacement vectors. The heading correction amount can be calculated according to the heading correction relationship described in the aforementioned invention.

[0051] After obtaining the heading correction for each map frame, this correction is superimposed on the heading angle of the original map frame to obtain the corrected heading angle. For visual inertial systems, since pitch and roll angles are usually constrained by the direction of gravity and are not directly affected by non-rigid corrections of planar position, this embodiment preferably updates only the heading angle while keeping the pitch and roll angles unchanged. The corrected map frame attitude can continue to be stored in quaternion form, or it can be converted to Euler angles or rotation matrix form as needed.

[0052] The seventh step is to generate a global visual prior map. After completing the map frame position correction and pose correction, the corrected map frame position, the corrected map frame pose, and the visual feature description information associated with each map frame are saved in a unified manner to generate a global visual prior map.

[0053] In this embodiment, the generated global visual prior map includes at least: the corrected position, corrected pose, normalized feature coordinates, and feature descriptors of each map frame. If necessary, auxiliary information such as pixel coordinate information, partial visible landmark information, and map frame timestamps can also be added.

[0054] Since the global visual prior map directly inherits the visual feature representation of the original visual map, it exhibits good modal consistency with the online visual localization, visual relocalization, and prior map-assisted fusion localization processes. In other words, the subsequent online localization stage can still employ visual feature extraction, matching, and relocalization methods compatible with the original visual SLAM or visual-inertial systems, without the need to construct additional complex cross-modal feature mapping relationships.

[0055] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a global visual prior map based on vector road network prior non-rigid correction, characterized in that, include: S1. Acquire the original images, and use SLAM to perform pose estimation and map construction on the original image sequence to obtain the original visual map. S2. Obtain vector road network data corresponding to the area covered by the original visual map trajectory, and represent the vector road network data as several polyline segments, nodes and node connection structures; S3. From the original visual map trajectory and vector road network, control points are extracted based on curvature scale space. A matching cost is constructed using a matching method based on shape context descriptors. The initial matching result is calculated using the optimal allocation algorithm. Matching control point pairs are eliminated using consistency checks to obtain the control point matching relationship between the original visual map trajectory and the vector road network. S4. Based on the control point matching relationship between the original visual map trajectory and the vector road network, establish a thin plate spline non-rigid transformation model from the original visual map trajectory to the vector road network prior, and use the transformation model to perform non-rigid correction on the position of each map frame in the original visual map to obtain the position of the map frame after preliminary correction. S5. Project the pre-corrected map frame position orthogonally onto the corresponding or nearest vector road network segment in the map plane to obtain the corrected position; S6. Calculate the heading correction amount based on the change in displacement direction before and after correction of adjacent map frames, update the heading angle in the attitude of the map frame based on the heading correction amount, and obtain the corrected heading angle. S7. Generate a global visual prior map using the map frame pose after position correction and heading angle correction and the corresponding visual feature description information.

2. The global visual prior map construction method based on vector road network prior non-rigid correction according to claim 1, characterized in that, S1 includes acquiring raw images and constructing raw visual maps: ； ； In the formula, The original visual map, For the first One map frame, This represents the total number of map frames. For map frame location, This refers to the pose of the map frame. This refers to visual feature description information associated with the map frame; It should include at least normalized feature coordinates, feature descriptors, pixel coordinates, and corresponding landmark information.

3. The global visual prior map construction method based on vector road network prior non-rigid correction according to claim 2, characterized in that, S3 includes S3.1, extracting control points from the original visual map trajectory and vector road network based on curvature scale space, establishing a matching cost matrix using a matching method based on shape context descriptors, calculating the initial matching results using the Hungarian algorithm, and filtering control point pairs using the RANSAC method to obtain the control point matching relationship between the original visual map trajectory and the vector road network; The original visual map trajectory is a spatial geometric path formed by connecting the positions of each map frame in chronological order.

4. The global visual prior map construction method based on vector road network prior non-rigid correction according to claim 3, characterized in that, S3 includes S3.2, which introduces rotation invariance and scale invariance during the construction of control point descriptors; Rotation invariance includes: taking the current control point as the center, calculating the relative position vectors from the current control point to other sampling points on the same curve, forming a set of position vectors; performing principal component analysis on the set of position vectors, and extracting the direction of the first principal component as the reference direction when constructing the descriptor; Scale invariance includes normalizing the relative distances between control points.

5. The global visual prior map construction method based on vector road network prior non-rigid correction according to claim 4, characterized in that, S4 includes, S4.1, where the control points on the original visual map trajectory are... The corresponding vector road network control points are , Then for any original map frame location The corresponding non-rigid transformation of the thin plate spline is: ； ； ； In the formula, To map the target location, For mapping the target location, the first Each coordinate component , , , For affine terms, For the first The control point at the ... The weighting coefficients of each coordinate component. For radial basis functions, To control the total number of point pairs, For the index number of the control point pair, For the Euclidean norm, , , for The three-dimensional coordinates.

6. The global visual prior map construction method based on vector road network prior non-rigid correction according to claim 5, characterized in that, S4 includes S4.2, constructing the kernel matrix and affine term matrix based on all control point pairs, and simultaneously solving the control point interpolation constraints and smoothing constraints to form a system of linear equations, and obtaining... , , , and ; Will Input the transformation model one by one to obtain the map frame positions after preliminary correction.

7. The global visual prior map construction method based on vector road network prior non-rigid correction according to claim 6, characterized in that, S5 includes searching for the corresponding or nearest road vector line segment in the horizontal plane for the initially corrected map frame position, and orthogonally projecting the initially corrected map frame position onto the corresponding or nearest road vector line segment.

8. The global visual prior map construction method based on vector road network prior non-rigid correction according to claim 7, characterized in that, S6 includes setting the displacement vector between adjacent map frames before correction as... The displacement vector between adjacent map frames after correction is Then the heading correction amount for: ； In the formula, The cross product symbol is used. The dot product symbol; Corrected heading angle for: ； In the formula, This is the original heading angle.

9. The global visual prior map construction method based on vector road network prior non-rigid correction according to claim 8, characterized in that, S7 includes a global visual prior map, comprising the corrected position, corrected pose, feature-normalized coordinates, and feature descriptors for each map frame.

Images