Visual SLAM (Simultaneous Localization and Mapping) vehicle positioning method based on high-precision map for GNSS (Global Navigation Satellite System) denial environment

Abstract

The invention discloses a GNSS (Global Navigation Satellite System) denial environment-oriented high-precision map-based visual synchronous localization and mapping SLAM (Simultaneous Localization and Mapping) vehicle localization method, which belongs to the field of vehicle navigation localization and comprises the following steps of: firstly, realizing identification of road lane lines and estimation of vehicle lateral offset distance and angle by adopting a visual lane identification algorithm; and then aiming at the long-time accumulative error problem of the SLAM, constructing a nonlinear optimization model for eliminating the SLAM error by utilizing observability of a road normal error, front and back position constraints constructed by an inertial measurement unit (IMU) and common-view constraints of visual features, so that reliable identification and elimination of the accumulative error are realized by utilizing map matching. According to the method, the vehicle positioning precision is greatly improved, the stable sub-meter positioning level can be kept under the condition of signal loss of a global navigation satellite system (GNSS), and the lane-level path planning and lane-level navigation service capabilities of the vehicle are ensured.

Description

Technical Field

[0001] The present invention relates to a high-precision map-based visual simultaneous localization and mapping (SLAM) vehicle positioning method for a GNSS-denied environment, and belongs to the technical field of vehicle navigation and positioning. Background Art

[0002] Accurate and reliable vehicle positioning is crucial for smart cars. As the most commonly used positioning system, the Global Navigation Satellite System (GNSS) can provide absolute positioning capability without offset in all weather conditions, but its robustness to complex occlusion environments is poor and it is easily affected by multipath effects and signal blocking. GNSS / Inertial Navigation System (INS) combined navigation can overcome the impact of GNSS loss in a short period of time, but the current Inertial Measurement Unit (IMU) devices cannot guarantee high-precision positioning for a long time. In the absence of GNSS, the visual SLAM system can rely on loop detection to eliminate cumulative errors, but in normal outdoor driving behavior, the positioning error often accumulates over time due to the lack of loop detection. Summary of the invention

[0003] In view of the problems existing in the above-mentioned prior art, the present invention constructs an optimized vision / IMU / GNSS / map system framework, integrates high-precision maps (expressed as HD) into the visual inertial odometer for posture correction, and can achieve accurate positioning of vehicles in complex environments over a long period of time and over a wide range. The main technical solutions are as follows:

[0004] (1) Initialize the visual inertial data and calculate the rotation matrix and translation vector from the carrier coordinate system (represented by B-frame) to the transverse Mercator coordinate system (represented by U-frame);

[0005] (2) When the GNSS observation status is poor, the odometer data is associated with the prior map information using the map matching method to obtain the lane-level lateral distance and lane angle; when the GNSS observation status is good, the GNSS data is used to correct the posture;

[0006] (3) Using lane detection, lane change detection, and Hough transform to process the on-board camera data, the distance and angle information of the current vehicle relative to the lane is obtained;

[0007] (4) Construct the residuals of visual reprojection, IMU pre-integration, GNSS, map lateral distance, and map heading angle data items, and add them to the factor graph optimization for multi-information fusion, so as to achieve continuous and stable precise positioning of the odometer;

[0008] (5) Conduct experimental tests on public datasets and real-world scenarios in complex urban environments.

[0009] Furthermore, in step (1), the visual inertial data needs to be integrated, the initialization is completed, and after the GNSS observation data and the map data are unified under the U-frame, the GNSS data is used to obtain the rotation matrix and translation vector from the B-frame to the U-frame.

[0010] Furthermore, the map matching method in step (2) constructs a lane-level map matching method based on hidden Markov. The lane data and historical trajectory information perceived from the video image of the camera are used to achieve more accurate matching based on the lane-level map. The lane-level map model is mainly a directed graph composed of nodes, edges and their attributes, where the nodes include the start node, the end node and the set virtual intersection node. The specific steps of constructing the HD map model are as follows:

[0011] (2.1) Draw an HD map based on the lane markings in the orthographic projection map and the laser point cloud map;

[0012] (2.2) Using the automatic generation method of lane segment topological relationship based on R-tree to realize the generation of topological relationship;

[0013] (2.3) Since there is no real road network in the intersection area, there is a lack of real lane information constraints, and vehicles are random and arbitrary when turning. In order to eliminate the positioning offset caused by the lack of constraints in the intersection area, a virtual node is used to represent the center position of the intersection;

[0014] (2.4) Based on the smooth arc trajectory characteristics of the vehicle when passing through the intersection and the map matching information of the vehicle before and after passing the intersection, a multi-lane intersection trajectory mathematical model is established, and the trajectory of the vehicle at the turning point of the intersection is generated using a virtual trajectory generation algorithm;

[0015] (2.5) According to the characteristics of the lane-level road network, relevant methods are used to automatically correct erroneous and suspicious topological information, and automatically generate attributes such as node ID, number of lanes, lane ID, and lane width.

[0016] Furthermore, in step (3), the vehicle camera data is processed using lane detection, lane change detection, and Hough transform to obtain the distance and angle information of the current vehicle relative to the lane, wherein the distance and angle information of the current vehicle relative to the lane is calculated as follows:

[0017] (3.1) Lateral distance: A lane detection method based on a convolutional neural network is used to detect lanes on the road ahead of the vehicle in the camera video image, and each lane line on the road is represented as a point set. In order to obtain accurate and robust lane line positions, the RANSAC method and the use of lane line historical position information can effectively supplement the missing lane lines and prevent the lane lines from interfering with the method due to wear and occlusion. Finally, the pixel distance between the center of the image and the left and right lane lines is calculated using the lane line point at the bottom of the image. When the current lane width is known, the actual distance between the vehicle and the left and right lane lines can be accurately obtained. In addition, the vehicle's lane change behavior can be accurately judged based on the change law of the lateral distance of the vehicle during the lane change process.

[0018] (3.2) Angle: After obtaining the lane using the lane detection method, in order to accurately express the offset and angle between the vehicle and the lane line, the camera image is first transformed into a bird's-eye view through the Hough transform, and the lane detection point set is projected onto the bird's-eye view; in order to prevent some abnormal points from affecting the linear expression of the lane, the RANSAC method is used for linear fitting to obtain a stable lane line and the corresponding angle between the vehicle and the lane line. The linearization formula is as follows:

[0019] f(x)=c0+c1x

[0020] Among them, C0 and c1 are the parameters of the lane line linear fitting equation, and x is the horizontal coordinate in the pixel coordinate system.

[0021] Furthermore, in step (4), the residuals of visual reprojection, IMU pre-integration, GNSS, lateral distance and map heading angle data items are constructed and added to the factor graph optimization for multi-information fusion, thereby achieving continuous and stable precise positioning of the odometer. The back-end factor graph optimization used is based on the sliding window optimizer to fuse the data of each sensor in the factor graph optimization framework. The optimized state variable χ is as follows:

[0022]

[0023] The world coordinate system is represented by W-frame, the camera coordinate system is represented by C-frame, and x k is the IMU state at each time node k, which contains the position of the IMU in the world coordinate system speed Rotation and gyroscope bias and acceleration bias n represents the sliding window size, Represents the transformation matrix between C-frame and B-frame, which contains the position vector between C-frame and B-frame and the rotation vector {γ i ,i=0,…,l} is the inverse depth of the landmark point since the i-th key frame, where l is the number of key frames;

[0024] And minimize the cost function in the form of MAP estimation:

[0025]

[0026] Among them, r p is the prior information residual, H p is the matrix of prior information, r pre is the IMU pre-integrated measurement residual, r C is the reprojection residual of the visual feature points, r G represents the GNSS measurement residual, and They represent the lateral distance residual and heading angle residual obtained by combining with the map, is the pre-integrated residual model at adjacent moments, is the reprojection error model for a single visual feature point, is the measurement error model of single-point GNSS, is the measurement error model between the single-point visual odometer and the HD map, m represents the number of lateral distance and heading angle measurements in the sliding window, Indicates the number of GNSS measurements in the sliding window;

[0027] Furthermore, the residuals of the GNSS, map lateral distance and map heading angle data items are specifically calculated as follows:

[0028] (4.1) GNSS residual:

[0029]

[0030] in, is the translation vector from B-frame to the world coordinate system (expressed by W-frame), Represents the rotation matrix from B-frame to W-frame, is the translation vector from U-frame to W-frame, is the arm value between U-frame and B-frame;

[0031] (4.2) Horizontal distance residual:

[0032] The lateral distance between the vehicle and the lane line obtained by map matching is as follows:

[0033]

[0034] in, is the transformation matrix from W-frame to U-frame, P is the odometer position coordinate, f(·) represents map matching, P m is the point obtained by map matching, d l represents half of the lane width, E(·) represents the Euclidean distance function, and κ is the direction parameter. When the distance between the odometer estimated position and the left lane is less than d l When it is positive, it is greater than d l When is negative;

[0035] The lateral distance between the vehicle and the lane line obtained by visual detection is as follows:

[0036]

[0037] in, is the arm value between the camera coordinate system (C-frame) and B-frame, d w is the lane width, θ v is the visual angle, w l With w r Respectively represent the pixel distance from the camera optical center to the left and right lane boundaries;

[0038] The lateral distance residual is:

[0039]

[0040] (4.3) Angle residual:

[0041]

[0042] Among them, θ o is the heading angle estimated by the odometer, φ is the heading angle between the vehicle and the lane line detected by vision;

[0043] Furthermore, in step (5), experimental tests are carried out on public datasets and real-world scenarios in complex urban environments. In order to fully verify and evaluate the proposed method framework, tests are carried out on public datasets and real-world datasets, respectively, and the experimental results are verified using indicators such as x and y axis errors under U-frame, lateral and longitudinal errors of the vehicle center, maximum error, and absolute trajectory error (ATE). The results of experimental data analysis show that in complex urban environments (even when GNSS signals are denied), the proposed algorithm can achieve real-time, accurate, and robust global positioning, which is of great significance for the research on long-distance, offset-free positioning technology for vehicles that do not rely on GNSS.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This paper proposes a visual / inertial / map SLAM positioning framework based on factor graph optimization. It uses a priori HD maps and rich semantic scene perception of visual cameras to constrain the lateral distance and heading angle of the vehicle, and integrates the HD map into the visual inertial odometer for posture correction. It can achieve accurate positioning of the vehicle in complex environments for a long time and over a large range, and build a real-time and robust posture estimator, so that in complex environments where GNSS is largely missing or even unavailable, the map can be used to provide position and heading error correction. In order to fully evaluate the effectiveness of the proposed system, experiments using public data sets and real-world environments are used to verify the system. The experimental results show that the proposed system can cope with the influence of complex environments and maintain accurate positioning for a long time and over a large range outdoors. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the method provided by the present invention.

[0047] Figure 2 The multi-lane road segment of the present invention uses a road-level map matching graph.

[0048] Figure 3 This is an explanation diagram of virtual nodes at HD map intersections constructed by the present invention.

[0049] Figure 4 It is a comparison result diagram of various algorithm trajectories of the present invention in a rejection environment.

[0050] Figure 5 It is the lane detection image of the present invention, wherein the left image is the original camera image and the right image is the image after lane line detection.

[0051] Figure 6It is an angle calculation diagram after the inverse perspective transformation of the present invention, wherein the left image is a lane line image when the vehicle is parallel to the lane line, and the right image is a lane line image when the vehicle is not parallel to the lane line.

[0052] Figure 7 It is the factor graph framework diagram of the present invention.

[0053] Figure 8 It is a comparison chart of the results of the present invention under the public data set.

[0054] Fig. 9 This is a comparison chart of the results of the present invention in a real scenario. DETAILED DESCRIPTION

[0055] In order to make the purpose and technical solution of the embodiments of the present invention clearer, the present invention is further explained below in conjunction with the accompanying drawings in the embodiments of the present invention:

[0056] like Figure 1 As shown, the present invention discloses a visual SLAM vehicle positioning method based on a high-precision map for a GNSS-denied environment, comprising the following steps:

[0057] Step S1: Integrate the visual inertial data, complete the initialization, and unify the GNSS observation data and map data in the transverse Mercator coordinate system (expressed by U-frame), and then use the GNSS data to obtain the rotation matrix and translation vector from the carrier coordinate system (expressed by B-frame) to the U-frame.

[0058] Step S2: When the GNSS observation status is poor, the lane data and historical trajectory information perceived from the camera's video image are used to associate the odometer data with the HD map information using a lane-level map matching algorithm based on the lane-level map, such as Figure 2 As shown in the figure, the lane-level lateral distance and lane angle are obtained. When the GNSS observation status is good, the GNSS data is used to correct the posture; among them, the map matching method is a lane-level map matching method based on hidden Markov. The lane-level map model is mainly a directed graph composed of nodes, edges and their attributes. The nodes include the starting node, the ending node and the set virtual intersection node. The edge includes the lane centerline and the extended edge in the intersection area. The definition of edge attributes can increase the prior information of the road network in addition to the topological relationship. In addition to the basic information such as the starting node id and the ending node id, its attributes also increase the definition of lane-level road attributes such as the lane id of the current edge, the number of road lanes and the lane width. The specific steps for building the HD map model are as follows:

[0059] S2.1 Draw an HD map based on the lane markings in the orthographic projection map and the laser point cloud map;

[0060] S2.2 uses the lane segment topology automatic generation method based on R-tree to realize the generation of topological relationship;

[0061] S2.3 Since there is no real road network in the intersection area, there is a lack of real lane information constraints, and vehicles are random and arbitrary when turning. In order to eliminate the positioning offset caused by the lack of constraints in the intersection area, a virtual node is used to represent the center position of the intersection, such as Figure 3 As shown;

[0062] S2.4 Based on the smooth arc trajectory characteristics of vehicles passing through intersections and the map matching information of vehicles before and after passing through intersections, a multi-lane intersection trajectory mathematical model is established, and the trajectory of vehicles at the turning point of intersections is generated using a virtual trajectory generation algorithm, such as Figure 4 As shown;

[0063] According to the characteristics of the lane-level road network, S2.5 uses relevant methods to automatically correct erroneous and suspicious topological information, and automatically generates attributes such as node ID, number of lanes, lane ID, and lane width.

[0064] Step S3: The vehicle-mounted camera data is processed using lane detection, lane change detection, and Hough transform to obtain the distance and angle information of the current vehicle relative to the lane, wherein the distance and angle information of the current vehicle relative to the lane is calculated as follows:

[0065] S3.1 Lateral distance: The lane detection method based on convolutional neural network is used to detect the lane of the road ahead of the vehicle in the camera video image, and each lane line in the road is represented as a point set, such as Figure 5 As shown in the figure; in order to obtain accurate and robust lane line positions, the use of the RANSAC method and the lane line historical position information can effectively supplement the missing lane lines and prevent the lane lines from interfering with the method due to wear and occlusion; finally, the pixel distance between the image center and the left and right lane lines is calculated using the lane line point at the bottom of the image. When the current lane width is known, the actual distance between the vehicle and the left and right lane lines can be accurately obtained; in addition, the vehicle's lane change behavior can be accurately judged based on the change law of the lateral distance of the vehicle during the lane change process;

[0066] S3.2 Angle: After obtaining the lane using the lane detection method, in order to accurately express the offset and angle between the vehicle and the lane line, the camera image is first transformed through Hough transform to obtain a bird's-eye view, and the lane detection point set is projected onto the bird's-eye view, such as Figure 6 As shown; in order to prevent some abnormal points from affecting the linear expression of the lane, the RANSAC method is used to process and perform linear fitting to obtain stable lane lines and the corresponding vehicle and lane line angles. The linearization formula is as follows:

[0067] f(x)=c0+c1x

[0068] Among them, c0 and c1 are the parameters of the lane line linear fitting equation, and x is the horizontal coordinate in the pixel coordinate system.

[0069] Step S4: construct the residuals of visual reprojection, IMU pre-integration, GNSS, map lateral distance and map heading angle data items, and add them to the factor graph optimization for multi-information fusion, so as to achieve continuous and stable precise positioning of the odometer. The back-end factor graph optimization used is based on the sliding window optimizer to fuse the data of each sensor in the factor graph optimization framework. The optimized state variable χ is as follows:

[0070]

[0071] The world coordinate system is represented by W-frame, the camera coordinate system is represented by C-frame, and x k is the IMU state at each time node k, which contains the position of the IMU in the world coordinate system speed Rotation and gyroscope bias and acceleration bias n represents the sliding window size, Represents the transformation matrix between C-frame and B-frame, which contains the position vector between C-frame and B-frame and the rotation vector {γ i ,i=0,…,l} is the inverse depth of the landmark point since the i-th key frame, where l is the number of key frames;

[0072] And minimize the cost function in the form of MAP estimation:

[0073]

[0074] Among them, r p is the prior information residual, H p is the matrix of prior information, r pre is the IMU pre-integrated measurement residual, r C is the reprojection residual of the visual feature points, r G represents the GNSS measurement residual, and They represent the lateral distance residual and heading angle residual obtained by combining with the map, is the pre-integrated residual model at adjacent moments, is the reprojection error model for a single visual feature point, is the measurement error model of single-point GNSS, is the measurement error model between the single-point visual odometer and the HD map, m represents the number of lateral distance and heading angle measurements in the sliding window, Indicates the number of GNSS measurements in the sliding window.

[0075] like Figure 7 As shown, the residuals of GNSS, lateral distance and map heading angle data items are calculated as follows:

[0076] S4.1 GNSS residuals:

[0077]

[0078] in, is the translation vector from B-frame to W-frame, Represents the rotation matrix from B-frame to U-frame, is the translation vector from U-frame to W-frame, is the arm value between U-frame and B-frame;

[0079] S4.2 Map lateral distance residual:

[0080] The lateral distance between the vehicle and the lane line obtained by map matching is as follows:

[0081]

[0082] in, is the transformation matrix from W-frame to U-frame, P is the odometer position coordinate, f(·) represents map matching, P m is the point obtained by map matching, d l represents half of the lane width, E(·) represents the Euclidean distance function, and κ is the direction parameter. When the distance between the odometer estimated position and the left lane is less than d l When it is positive, it is greater than d l When is negative;

[0083] The lateral distance between the vehicle and the lane line obtained by visual detection is as follows:

[0084]

[0085] in, is the arm value between C-frame and B-frame, d w is the lane width, θ v is the visual angle, w l With w r Respectively represent the pixel distance from the camera optical center to the left and right lane boundaries;

[0086] The lateral distance residual is:

[0087]

[0088] S4.3 Angle residual: After obtaining the accurate lane lines on the left and right sides of the vehicle through the lane detection algorithm, the angle is calculated according to the lane line expression obtained by the linear fitting formula. The angle value θ v The specific calculation is:

[0089]

[0090] Among them, c0 and c1 are the parameters of the lane line linear fitting equation respectively;

[0091] Based on the accurate lane lines on the map obtained by map matching, the lane line vector V in the U-frame coordinate system l =(e x ,e y ) and the x-axis vector V U =(u x ,u y ) is as follows:

[0092]

[0093] Adjust it to the range of [-π,π], then:

[0094]

[0095] Therefore, the actual angle of the vehicle θ = θ m +θ v , then the vehicle's attitude Euler angle can be expressed as: A U =(α,β,θ). By U Convert to direction cosine matrix R U , and convert it to the unified W-frame coordinate system to get R W , and its calculation formula is as follows:

[0096]

[0097] According to the Rodriguez formula, the direction cosine matrix R U Convert to Euler angles The heading angle is φ;

[0098] The vehicle attitude estimated by the odometer is expressed in quaternion as:

[0099] q W =q w +q x i+q y j+q z k

[0100] Among them, i, j, k are the imaginary parts of the quaternion, q x ,q y ,q z are the imaginary coefficients of the quaternion, q w is the real part of the quaternion;

[0101] The formula to convert it to Euler angles is:

[0102]

[0103] Therefore, the angle residual term can be:

[0104]

[0105] Among them, θ o is the heading angle estimated by the odometer, φ is the heading angle between the vehicle and the lane line detected by vision;

[0106] Step S5: Use the public dataset KAIST Complex Urban Dataset to evaluate the proposed method. The sequence urban 39 with a trajectory length of 10678 meters is selected for actual evaluation, as shown in Figure 8 As shown in the figure, the proposed algorithm is compared with the trajectory of VINS-Mono and ORB-SLAM3. It can be seen that the patented algorithm has only a small drift, while VINS-Mono and ORB-SLAM3 have large offset errors, and their errors gradually accumulate and become larger in the absence of effective absolute position constraints. In addition, a real-world dataset is used for verification. The dataset is actually collected and processed in a certain urban area in Nanjing. The collection trajectory distance is 5.78 kilometers. Fig. 9 The trajectory of each algorithm is displayed on a real map scene. Compared with ORB-SLAM3, VINS-Mono has a significantly smaller positioning error in the initial stage due to better heading estimation, while the positioning error in the later stage becomes larger due to the deviation of heading estimation. In general, the trajectories of the two algorithms have a large trajectory deviation compared to the ground truth, and have lost the function of vehicle positioning. In contrast, the algorithm proposed in this patent obtains a trajectory with very small error due to the correction of heading angle and lateral distance errors.

Claims

1. A high-precision map-based visual SLAM vehicle positioning method for GNSS-denied environments, characterized in that: The steps include: (1) Initialize the visual inertial data and calculate the rotation matrix and translation vector from the carrier coordinate system to the transverse Mercator coordinate system, where the carrier coordinate system is represented by B-frame and the transverse Mercator coordinate system is represented by U-frame; (2) When the GNSS observation status is poor, the odometer data is associated with the prior high-precision map HD information using the map matching method to obtain the lane-level lateral distance and lane angle; when the GNSS observation status is good, the GNSS data is used to correct the posture; (3) Using lane detection, lane change detection, and Hough transform to process the vehicle camera data, the lateral distance and angle information of the current vehicle relative to the lane is obtained; (4) Construct the residuals of visual reprojection, IMU pre-integration, GNSS, map lateral distance, and map heading angle data items, and add them to the factor graph optimization for multi-information fusion, so as to achieve continuous and stable precise positioning of the odometer; (5) Conduct experimental tests on public datasets and real-world scenarios in complex urban environments.

2. The method for visual SLAM vehicle positioning based on high-precision maps for GNSS-denied environments according to claim 1, characterized in that: In the step (1), the visual inertial data needs to be integrated and initialized, and after the GNSS observation data and the map data are unified under the U-frame, the rotation matrix and translation vector from the B-frame to the U-frame are obtained using the GNSS data.

3. The method for visual SLAM vehicle positioning based on high-precision maps for GNSS-denied environments according to claim 1, characterized in that: The map matching method in step (2) constructs a lane-level map matching method based on hidden Markov. The lane data and historical trajectory information perceived from the video image of the camera are used to achieve more accurate matching based on the lane-level map. The lane-level map model is a directed graph composed of nodes, edges and their attributes, wherein the nodes include a start node, an end node and a set virtual intersection node. The specific steps of constructing the HD map model are as follows: (2.1) Draw an HD map based on the lane markings in the orthographic projection map and the laser point cloud map; (2.2) Using the automatic generation method of lane segment topological relationship based on R-tree to realize the generation of topological relationship; (2.3) Since there is no real road network in the intersection area, there is a lack of real lane information constraints, and vehicles are random and arbitrary when turning. In order to eliminate the positioning offset caused by the lack of constraints in the intersection area, a virtual node is used to represent the center position of the intersection; (2.4) Based on the smooth arc trajectory characteristics of the vehicle when passing through the intersection and the map matching information of the vehicle before and after passing the intersection, a multi-lane intersection trajectory mathematical model is established, and the trajectory of the vehicle at the turning point of the intersection is generated using a virtual trajectory generation algorithm; (2.5) According to the characteristics of the lane-level road network, relevant methods are used to automatically correct erroneous and suspicious topological information, and automatically generate node ID, lane number, lane ID, and lane width attributes.

4. The method of visual SLAM vehicle positioning based on high-precision maps for GNSS-denied environments according to claim 1, characterized in that: In step (3), lane detection, lane change detection and Hough transform are used to process the vehicle camera data to obtain the distance and angle information of the current vehicle relative to the lane, wherein the current vehicle angle information is calculated as follows: After obtaining the lane using the lane detection method, in order to accurately express the offset and angle between the vehicle and the lane line, the camera image is first transformed through the Hough transform to obtain a bird's-eye view, and the lane detection point set is projected onto the bird's-eye view; in order to prevent some abnormal points from affecting the linear expression of the lane, the RANSAC method is used for linear fitting to obtain a stable lane line and the corresponding angle between the vehicle and the lane line. The linearization formula is as follows: f(x)=c0+c1x Among them, c0 and c1 are the parameters of the lane line linear fitting equation, and x is the horizontal coordinate in the pixel coordinate system.

5. The method of visual SLAM vehicle positioning based on high-precision maps for GNSS-denied environments according to claim 1, characterized in that: In step (4), the residuals of visual reprojection, IMU pre-integration, GNSS, map lateral distance and map heading angle data items are constructed and added to the factor graph optimization for multi-information fusion, so as to achieve continuous and stable precise positioning of the odometer. The back-end factor graph optimization used is based on the sliding window optimizer to fuse the data of each sensor in the factor graph optimization framework. The optimized state variable χ is as follows: The world coordinate system is represented by W-frame, the camera coordinate system is represented by C-frame, and x k is the IMU state at each time node k, which contains the position of the IMU in the world coordinate system speed Rotation Vector and gyroscope bias and acceleration bias n represents the sliding window size, Represents the transformation matrix between C-frame and B-frame, which contains the position vector between C-frame and B-frame and the rotation vector {γ i ,i=0,…,l} is the inverse depth of the landmark point since the i-th key frame, where l is the number of key frames; The cost function is minimized in the form of MAP estimation, and its formula is as follows: Among them, r p is the prior information residual, H p is the matrix of prior information, r pre is the IMU pre-integrated measurement residual, r C is the reprojection residual of the visual feature points, r G represents the GNSS measurement residual, and They represent the lateral distance residual and heading angle residual obtained by combining with the map, is the pre-integrated residual model at adjacent moments, is the reprojection error model for a single visual feature point, is the measurement error model of single-point GNSS, is the measurement error model between the single-point visual odometer and the HD map, m represents the number of lateral distance and heading angle measurements in the sliding window, Indicates the number of GNSS measurements in the sliding window.

6. The method of visual SLAM vehicle positioning based on high-precision maps for GNSS-denied environments according to claim 5, characterized in that: The residual of the map heading angle data item is specifically calculated as follows: After obtaining the accurate lane lines on the left and right sides of the vehicle through the lane detection algorithm, the angle is calculated according to the lane line expression obtained by the linear fitting formula. The angle value θ v The specific calculation is: Among them, c0 and c1 are the parameters of the lane line linear fitting equation respectively; Based on the accurate lane lines on the map obtained by map matching, the lane line vector V in the U-frame coordinate system l =(e x ,e y ) and the x-axis vector V U =(u x ,u y ) is as follows: Adjust it to the range of [-π,π], then: Therefore, the actual angle of the vehicle θ = θ m +θ v , then the vehicle's attitude Euler angle can be expressed as: A U =(α,β,θ); by A U Convert to direction cosine matrix R U , and convert it to the unified W-frame coordinate system to get R W , and its calculation formula is as follows: According to the Rodriguez formula, the direction cosine matrix R U Convert to Euler angles The heading angle is φ; The vehicle attitude estimated by the odometer is expressed in quaternion as: q W =q w +q x i+q y j+q z k Among them, i, j, k are the imaginary parts of the quaternion, q x ,q y ,q z are the imaginary coefficients of the quaternion, q w is the real part of the quaternion; The formula to convert it to Euler angles is: Therefore, the angle residual term is: Among them, θ o is the heading angle estimated by the odometer, and φ is the heading angle between the vehicle and the lane line detected by vision.

7. The method of visual SLAM vehicle positioning based on high-precision maps for GNSS-denied environments according to claim 1, characterized in that: In the step (5), experimental tests are conducted on public datasets and real-world scenarios in complex urban environments. In order to fully verify and evaluate the proposed method framework, tests are conducted on public datasets and real-world datasets, and the experimental results are analyzed using the x and y axis errors under U-frame, the lateral and longitudinal errors of the vehicle center, the maximum error, and the absolute trajectory error ATE indicators.

Images