A map construction method and system for a driving test subject three terrain surveying and mapping

Abstract

The application discloses a kind of map construction method and system of driving test subject three terrain surveying and mapping, belong to high-precision map construction technical field.The method is collected test route point cloud data by vehicle-mounted laser radar, according to scanning angle resolution parameter and adjacent frame space overlap degree, the effective scanning coverage boundary of each frame is calculated, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point is extracted in boundary, and the road boundary feature point

Description

Technical Field

[0001] This invention relates to the field of high-precision map construction technology, specifically to a map construction method and system for topographic surveying in driving test subject three. Background Technology

[0002] Driving test subject three terrain mapping requires obtaining accurate road geometry and topology information along the test route to support the test evaluation system's real-time judgment of vehicle trajectory. Currently, road map acquisition methods mostly rely on manual measurement or post-processing based on GPS trajectories. Manual measurement is inefficient and prone to human error, making it difficult to guarantee accurate representation of road boundary spatial positions. Satellite trajectory methods can only provide the vehicle's path and cannot directly obtain the complete geometric shape of the road boundary. In lidar-based automatic mapping schemes, the conventional approach is to collect complete point cloud data for the entire road segment and then extract road boundaries using point cloud registration and segmentation algorithms. However, when performing inter-frame matching of point cloud data, such methods suffer from accumulated registration errors due to sparse environments or repetitive road structures, affecting the geometric consistency of the global map. Regarding road boundary feature extraction, existing technologies generally treat the entire frame of point cloud as the processing object, without considering the spatial resolution decay of vehicle-mounted lidar scanning points with increasing distance and the redundancy of information in adjacent frames. Uneven distribution of scanning points makes distant boundary extraction susceptible to noise interference, making it difficult to distinguish between effective scanning areas and invalid stray points. Furthermore, extraction methods based on single geometric or intensity features suffer from insufficient road boundary recognition integrity when road boundary markers are faded, curbs are damaged, or they are occluded, failing to provide reliable spatial constraints for subsequent pose optimization. In the pose estimation stage, conventional methods typically rely on global registration of the original point cloud or graph optimization based on general feature points, failing to fully utilize the stable spatial constraints formed between frames by the strong structural feature of road boundaries. This results in weak constraint relationships during pose solving, and the final constructed road vector map is prone to boundary breaks and topological inconsistencies on long-distance continuous road segments. Summary of the Invention

[0003] This invention provides a map construction method and system for topographic mapping in driving test subject 3. The purpose is to solve the problem of inconsistent global boundary geometry caused by unclear definition of effective coverage area of ​​point cloud and unstable extraction of road boundary features in driving test route map construction, as well as the problem of cumulative error and boundary breakage caused by the failure to effectively introduce road boundary features into pose graph optimization.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a map construction method for terrain surveying in driving test subject three, comprising: responding to a driving start command of the test vehicle on a set route, acquiring multiple frames of continuous point cloud data of the traversed road segment through an onboard LiDAR; calculating the effective scanning coverage boundary corresponding to each frame of point cloud data based on the scanning angular resolution parameters of the onboard LiDAR and the spatial overlap of adjacent frame point cloud data; extracting road boundary feature points from each frame of point cloud data based on the effective scanning coverage boundary, wherein the road boundary feature points include road surface elevation jump points and road surface reflection intensity abrupt change points; and processing the road boundary feature points... As spatial constraint nodes, a spatial relative pose graph is constructed between point cloud data of each frame. The edges of the spatial relative pose graph represent the spatial transformation constraint relationship between two adjacent point cloud data frames. The spatial relative pose graph is solved to obtain the optimal pose estimate of each frame point cloud data in the global coordinate system. Based on the optimal pose estimate, the road boundary feature points in each frame point cloud data are projected onto the global coordinate system to generate a road boundary feature point cloud map. Spatial neighborhood clustering is performed on the feature points in the road boundary feature point cloud map to extract continuous boundary line segments. Based on the spatial extension direction of the boundary line segments, a vector map containing road geometric topology information is generated.

[0005] This method determines the effective scanning coverage boundary by comprehensively utilizing the scanning angular resolution of the lidar and the spatial overlap of adjacent frames. It extracts road boundary feature points only within the effective range, filtering out a large amount of redundant and low-quality point clouds, thus reducing the computational load of subsequent processing. The feature point extraction simultaneously integrates two feature responses: abrupt changes in road elevation and abrupt changes in reflection intensity, demonstrating good recognition capabilities for various road boundary morphologies such as curbs, road markings, and guardrails. Based on the extracted road boundary feature points, a spatial relative pose map containing constraints from adjacent frames and loop closures is constructed. Iteratively solving this pose map effectively eliminates inter-frame cumulative pose drift, making the pose estimation of point clouds in each frame in the global coordinate system more accurate and robust. Based on the optimized pose, the road boundary feature points in the multi-frame point cloud are uniformly projected to the global coordinate system to generate a dense and highly consistent road boundary feature point cloud map. Then, continuous boundary line segments are extracted through spatial neighborhood clustering and line fitting, and continuous vector lines are generated by merging according to the spatial extension direction of the boundary line segments. The resulting vector map fully represents the geometric structure and topological extension relationship of the road, and can be directly used for terrain mapping and scene reconstruction of the driving test subject three test route.

[0006] As a technical solution of the present invention, the effective scanning coverage boundary corresponding to each frame of point cloud data is calculated based on the scanning angular resolution parameters of the vehicle-mounted LiDAR and the spatial overlap of adjacent frame point cloud data. Specifically, this includes: obtaining the horizontal and vertical angular resolutions of the vehicle-mounted LiDAR; determining the spatial azimuth distribution matrix of each scanning point in a single frame of point cloud data based on the horizontal and vertical angular resolutions; performing spatial position overlap calculations on adjacent frame point cloud data to obtain the spatial overlap of the adjacent frame point cloud data; determining the starting and ending azimuth angles of the effective scanning points in a single frame of point cloud data based on the spatial azimuth distribution matrix and the spatial overlap; and defining the effective scanning coverage boundary of the frame of point cloud data using the starting and ending azimuth angles. This adaptively distinguishes reliable scanning areas and low-overlap edge areas in a single frame of point cloud data, ensuring sufficient information redundancy in the point cloud data participating in subsequent feature extraction, which is beneficial for improving the stability and repeatability of road boundary feature extraction.

[0007] Preferably, based on the effective scan coverage boundary, road boundary feature points are extracted from the point cloud data of each frame. Specifically, this includes: within the range defined by the effective scan coverage boundary, calculating the spatial distance and elevation difference between each scan point and its adjacent scan points; marking scan points whose spatial distance is less than a preset neighborhood radius and whose elevation difference is greater than a preset jump threshold as road surface elevation jump points; marking scan points whose laser reflection intensity difference between adjacent scan points is greater than a preset intensity jump threshold as road surface reflection intensity jump points; merging the road surface elevation jump points and the road surface reflection intensity jump points to obtain the road boundary feature points; and associating each road boundary feature point with its corresponding scan timestamp and scan azimuth. By using distance constraints and dual determination of elevation and reflection intensity, the road surface area and the boundary area can be effectively distinguished, avoiding mis-extraction caused by local road surface undulations or reflectivity fluctuations. At the same time, time and azimuth information are added to each feature point, facilitating inter-frame data association and pose map construction.

[0008] As another technical solution of the present invention, the road boundary feature points are used as spatial constraint nodes to construct a spatial relative pose graph between point cloud data of each frame. Specifically, this includes: using each frame of point cloud data as a node of the pose graph, and assigning an initial pose estimate to each node; establishing edges between adjacent frames using the spatial distance residual between the road boundary feature points in adjacent frame point cloud data as a constraint condition; establishing loop edges using the spatial distance residual between the road boundary feature points in non-adjacent frame point cloud data as a constraint condition; assigning a weight coefficient to each edge, the weight coefficient being inversely proportional to the magnitude of the spatial distance residual; and combining all the nodes and edges to generate the spatial relative pose graph. Introducing loop edge constraints into the pose graph can significantly reduce the cumulative drift during long-distance travel, while the inverse weight allocation strategy based on residuals allows feature points with higher matching quality to exert a greater influence on the optimization results, improving the robustness of the pose graph solution.

[0009] Preferably, solving the spatial relative pose graph to obtain the optimal pose estimate of each frame of point cloud data in the global coordinate system specifically includes: initializing the global pose variables of each frame of point cloud data and fixing the global pose of the first node as the origin of the global coordinate system; for each edge in the spatial relative pose graph, calculating the deviation between the current global pose transformation matrix of the two nodes connected by the edge and the spatial transformation constraint carried by the edge; linearizing the spatial relative pose graph according to the deviation to construct an incremental equation system of pose variables; iteratively solving the incremental equation system and updating the global pose variables of each node after each iteration; stopping the iteration and outputting the current global pose variables of each node as the optimal pose estimate when the sum of the deviations is less than a preset convergence threshold. Further, in the iterative solution process, the LM algorithm is used to solve the incremental equation system, and the damping factor is updated after each iteration, so that the solution process combines the speed of gradient descent with the high accuracy of the Gauss-Newton method, which is beneficial for obtaining a globally consistent optimal pose configuration in complex road condition pose graphs.

[0010] In a preferred embodiment of the present invention, the road boundary feature points in each frame of point cloud data are projected onto the global coordinate system based on the optimal pose estimation value. Specifically, this includes: obtaining the optimal pose estimation value corresponding to each frame of point cloud data, wherein the optimal pose estimation value includes rotation and translation components; constructing a rotation matrix based on the rotation components and a translation vector based on the translation components; multiplying the local coordinates of each road boundary feature point in a single frame of point cloud data sequentially by the rotation matrix and then adding the translation vector to obtain the global coordinates of the road boundary feature point in the global coordinate system; and performing the local coordinate to global coordinate transformation on all road boundary feature points in all frames of point cloud data to obtain the road boundary feature point cloud map. The projection process after pose map optimization ensures tight stitching of each frame of point cloud data, significantly improving the positional accuracy of road boundary feature points in the global map and providing an accurate spatial reference for subsequent vector boundary generation.

[0011] As a specific solution of the present invention, spatial neighborhood clustering is performed on feature points in the road boundary feature point cloud map to extract continuous boundary line segments. Specifically, this includes: selecting unlabeled feature points as seed points in the road boundary feature point cloud map; establishing a spatial neighborhood search window centered on the seed points; searching for feature points within the search window whose spatial distance to the seed points is less than a preset clustering radius; grouping the found feature points and the seed points into the same feature point cluster; using the newly grouped feature points as new seed points, iteratively performing the spatial neighborhood search until the feature point cluster no longer expands; and performing line fitting on all feature points within the feature point cluster to obtain the boundary line segments. This clustering method can recover boundary line segments with continuous geometric structures from discrete feature points, without relying on prior road models, and is adaptable to road boundaries with different curvatures and alignments.

[0012] Preferably, based on the spatial extension direction of the boundary line segments, a vector map containing road geometric topology information is generated. Specifically, this includes: calculating the coordinates of the two endpoints of each boundary line segment in the global coordinate system and the spatial extension direction vector of the boundary line segment; determining whether adjacent boundary line segments belong to the same road boundary based on the spatial distance between the endpoints of adjacent boundary line segments and the angle between the spatial extension direction vectors of adjacent boundary line segments; merging adjacent boundary line segments belonging to the same road boundary to form continuous road boundary vector lines; and combining all the road boundary vector lines and their spatial extension direction vectors to generate the vector map containing road geometric topology information. Through endpoint proximity and direction consistency checks, the merged vector lines are ensured to be smooth and topologically continuous, ultimately outputting an editable vector map product that can be directly used for driving test scenario analysis.

[0013] This invention also provides a map building system for terrain surveying in driving test subject three, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the map building method for terrain surveying in driving test subject three described above. This system can be embedded in an onboard computing platform and generate high-precision road vector maps in real time or near real time during the driving of the test vehicle, providing reliable map data support for the automatic evaluation and scenario reproduction of driving test subject three.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By combining the scanning angular resolution parameters of the vehicle-mounted LiDAR with the spatial overlap of adjacent frame point cloud data, the effective scanning coverage boundary of each frame point cloud is determined. The spatial azimuth distribution matrix of scanning points established by the horizontal and vertical angular resolutions can accurately describe the density distribution of scanning lines within a single frame point cloud. Then, by using the spatial overlap of adjacent frames, duplicate scanning areas and sparse edge areas caused by vehicle movement are eliminated, ensuring that the extracted road boundary feature points are concentrated within an effective scanning range with stable coverage quality and controllable information redundancy. Within this effective boundary range, road surface elevation jump points are extracted simultaneously using spatial distance and elevation changes, and road surface reflection intensity abrupt change points are extracted using laser reflection intensity changes. These two types of feature points are merged as road boundary feature points. This dual-feature joint extraction method can reliably obtain the road boundary position in different scenarios, such as sharp curb edges or insufficient visual contrast of road markings. It compensates for the incomplete extraction under conditions of boundary damage or occlusion when relying solely on single geometric or intensity features. The output road boundary feature points are more stable, providing evenly distributed and structurally continuous spatial constraint nodes for subsequent pose map construction. By constructing a spatial relative pose map using road boundary feature points in the point clouds of each frame as spatial constraint nodes, edge constraints are established using the spatial distance residuals between road boundary feature points in adjacent and non-adjacent frames, and weight coefficients inversely proportional to the distance residuals are assigned, allowing the road boundary, a strongly structural feature, to directly participate in the global pose optimization process. Utilizing the continuous spatial extension of road boundary feature points, clear directional and distance constraints can be provided for inter-frame transformation relationships. Compared to methods relying on scattered point registration, this enhances the geometric meaning of constraint edges in the pose map, resulting in more stable convergence directions when solving incremental equations. After iteratively solving for the optimal pose estimates of each frame in the global coordinate system, the road boundary feature points of each frame are uniformly projected to generate a geometrically aligned road boundary feature point cloud map in the global coordinate system. Then, spatial neighborhood clustering and line fitting are performed on the feature points to extract continuous boundary line segments. Finally, line segments belonging to the same road boundary are merged based on the endpoint spatial distance and extension direction vector to form continuous road boundary vector lines. This process effectively suppresses boundary misalignment and breakage caused by local pose errors in long-distance sequences, and the generated vector map maintains the consistency of road geometry and topology. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the map construction method for topographic surveying in the driving test subject 3; Figure 2 This is a flowchart of the construction and optimization of spatial relative pose graph based on road boundary feature matching; Figure 3 This is a flowchart of the process for generating a point cloud map of road boundary features; Figure 4 This is a schematic diagram illustrating the determination of elevation jumps and reflection intensity abrupt changes at feature points along the road boundary; Figure 5 This is a comparison chart of the residual distribution before and after optimization of the spatial relative pose graph; Figure 6 It is a global coordinate distribution map of feature points at the road boundary; Figure 7 This is a schematic diagram of the clustering of feature points at the road boundary and the fitting of boundary line segments. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1This invention provides a map construction method for terrain surveying in driving test subject three, comprising: responding to a driving start command of the test vehicle on a set route, acquiring multiple frames of continuous point cloud data of the traversed road segment using an onboard LiDAR; calculating the effective scanning coverage boundary corresponding to each frame of point cloud data based on the scanning angular resolution parameters of the onboard LiDAR and the spatial overlap of adjacent frame point cloud data; extracting road boundary feature points from each frame of point cloud data based on the effective scanning coverage boundary, the road boundary feature points including road surface elevation jump points and road surface reflection intensity abrupt change points; and using the road boundary feature points as spatial constraint nodes. A spatial relative pose graph is constructed among the point cloud data of each frame. The edges of the spatial relative pose graph represent the spatial transformation constraints between two adjacent point cloud data frames. The spatial relative pose graph is solved to obtain the optimal pose estimate of each frame of point cloud data in the global coordinate system. Based on the optimal pose estimate, the road boundary feature points in each frame of point cloud data are projected onto the global coordinate system to generate a road boundary feature point cloud map. Spatial neighborhood clustering is performed on the feature points in the road boundary feature point cloud map to extract continuous boundary line segments. Based on the spatial extension direction of the boundary line segments, a vector map containing road geometric topology information is generated.

[0019] In practical implementation, the horizontal and vertical angular resolutions of the vehicle-mounted LiDAR are obtained. The horizontal and vertical angular resolutions of the vehicle-mounted LiDAR are determined by its factory specifications. Indicates the angular interval between adjacent horizontal scan lines, and the vertical angular resolution. This represents the angular interval between adjacent vertical scan layers. Based on the horizontal and vertical angular resolutions, the spatial azimuth distribution matrix of each scan point in a single frame of point cloud data is determined. The vehicle-mounted LiDAR emits laser pulses in fixed angular increments within a single frame scan cycle, with each scan point corresponding to a set of horizontal and vertical azimuth angles. For any scan point in a single frame of point cloud data, the scan point index is used. Identifier, scan point Corresponding horizontal azimuth and vertical azimuth Calculated using the following formula: ; in, The horizontal starting azimuth angle for a single frame scan. This is the vertical starting azimuth angle for a single frame scan. For horizontal angular resolution, For vertical angular resolution, For scan points The sequence number in the horizontal scan sequence For scan points The sequence number in the vertical scan layer. All scan points. These are combined to form a spatial azimuth distribution matrix.

[0020] Spatial overlap calculation is performed on adjacent frame point cloud data to obtain the spatial overlap of adjacent frame point cloud data. Using the initial pose provided by the vehicle positioning system, the point cloud data of two adjacent frames are transformed into the same local coordinate system. By dividing the three-dimensional space into a grid, the number of scan points falling into the same grid in the two frame point cloud data is counted. The ratio of the number of overlapping scan points in the same grid to the total number of scan points is used as the spatial overlap.

[0021] Based on the spatial azimuth distribution matrix and spatial overlap, the starting and ending azimuth angles of the effective scan points in a single frame of point cloud data are determined. Based on the calculated spatial overlap, the 3D spatial overlap region between two frames of point clouds is identified. In the spatial azimuth distribution matrix, the horizontal and vertical azimuth ranges corresponding to the 3D spatial overlap region are determined. The minimum value of the horizontal azimuth range is determined as the starting azimuth angle of the effective scan point, and the maximum value is determined as the ending azimuth angle. The effective scan coverage boundary of the single frame of point cloud data is defined by the starting and ending azimuth angles, and subsequent feature extraction is performed only at the scan points defined by the effective scan coverage boundary.

[0022] Within the effective scan coverage boundary, the spatial distance and elevation difference between each scan point and its adjacent scan points are calculated point by point. For each scan point within the effective scan coverage boundary, another scan point geometrically adjacent to the scan point on the same scan line is selected as the adjacent scan point. The Euclidean distance between the coordinates of the two scan points is calculated as the spatial distance, and the absolute value of the difference between the elevation coordinates of the two scan points is calculated as the elevation difference. Both the spatial distance and the elevation difference are expressed in meters.

[0023] Scanning points whose spatial distance is less than a preset neighborhood radius and whose elevation difference is greater than a preset jump threshold are marked as road elevation jump points. The preset neighborhood radius is set at 0.3 meters, based on the fact that the horizontal point spacing of the vehicle-mounted LiDAR at a distance of 10 meters from the target is approximately 0.3 meters; the preset jump threshold is set at 0.1 meters, based on the fact that the typical lower limit of the elevation difference between the road curb and the road surface is 0.1 meters. Scanning points that simultaneously meet the conditions of a spatial distance less than 0.3 meters and an elevation difference greater than 0.1 meters are marked as road elevation jump points.

[0024] Scanning points where the difference in laser reflection intensity between adjacent scanning points exceeds a preset intensity mutation threshold are marked as points where the road surface reflection intensity changes abruptly. The preset intensity mutation threshold is set to 20, based on the assumption that changes in laser reflection intensity exceeding 20 occur when the road surface or curbstone material changes. Scanning points where the absolute value of the difference in laser reflection intensity between two adjacent scanning points exceeds 20 are marked as points where the road surface reflection intensity changes abruptly.

[0025] By merging the points where the road surface elevation changes abruptly and the points where the road surface reflection intensity changes abruptly, the road boundary feature points are obtained. The set of scan points marked as points where the road surface elevation changes abruptly and the set of scan points marked as points where the road surface reflection intensity changes abruptly are then combined, and duplicate scan points are removed to form the set of road boundary feature points.

[0026] For each feature point in the set of road boundary feature points, associate the corresponding scan timestamp and scan azimuth angle. Parse the acquisition time of each scan point from the data packet of the vehicle-mounted LiDAR as the scan timestamp; extract the horizontal and vertical azimuth angles of the corresponding scan point from the spatial azimuth angle distribution matrix as the scan azimuth angle. Store the scan timestamp and scan azimuth angle in the attribute information of the corresponding road boundary feature point.

[0027] See Figure 4 The figure illustrates the relationship between elevation differences and spatial distances after classifying scan points according to the road boundary feature point extraction method in this embodiment. The horizontal axis represents the spatial distance between a scan point and its adjacent scan points, in meters, ranging from 0 to 0.9 meters; the vertical axis represents the elevation difference of the corresponding scan point, in meters, ranging from 0 to 0.27 meters. The figure distinguishes three types of scan points using different symbols: blue dots represent non-feature points, red triangles represent points of abrupt changes in road surface elevation, and green squares represent points of abrupt changes in road surface reflection intensity.

[0028] In the diagram, the preset neighborhood radius, represented by the orange dashed line, is 0.3 meters, and this boundary is used to limit the upper limit of the spatial distance. The preset jump threshold, represented by the purple dotted line, is 0.1 meters, and is used to limit the lower limit of the elevation difference. The red triangular points are all distributed in the upper right region where the spatial distance is less than 0.3 meters and the elevation difference is greater than 0.1 meters. This clearly demonstrates the judgment rule in this embodiment that scan points that meet the condition of "spatial distance less than the preset neighborhood radius (0.3 meters) and elevation difference greater than the preset jump threshold (0.1 meters)" are marked as road surface elevation jump points. The green squares are more dispersed, and their elevation differences are mostly less than 0.1 meters, but the spatial distance can be greater than 0.3 meters. This matches the characteristics of road surface reflection intensity abrupt change points, indicating that the judgment basis mainly relies on the change in laser reflection intensity rather than the elevation difference.

[0029] The blue dots, as non-feature points, are concentrated in areas where the spatial distance is less than 0.3 meters and the elevation difference is less than 0.1 meters. This indicates that these points neither meet the threshold condition for elevation jump points nor have significant abrupt changes in reflection intensity, which conforms to the definition of non-feature points in this embodiment.

[0030] In specific implementation, please refer to Figure 2Each frame of point cloud data is used as a node in a spatial relative pose map, and an initial pose estimate is assigned to each node. The point cloud data is acquired by an onboard LiDAR, with each frame corresponding to one node. The total number of nodes is denoted as […]. , No. The initial pose estimate for each node is expressed as follows: ,in It is a three-dimensional translation vector. It is a quaternion representing rotation. The value range is 1 to The initial pose estimate is derived from pose data provided by the vehicle positioning system, which includes position coordinates and attitude quaternions.

[0031] Using the spatial distance residual between road boundary feature points in adjacent frame point cloud data as a constraint, edges between adjacent frames are established. For two consecutive frames of point cloud data with continuous timestamps, road boundary feature point sets are extracted respectively. The feature points of the two frames are transformed to a unified coordinate system using the initial pose estimation values. Nearest neighbor matching is performed on the feature points of the two frames to form feature point matching pairs. For each pair of matched feature points, the Euclidean distance between the feature points is calculated, and the Euclidean distance is used as the spatial distance residual. The edges between adjacent frames carry the spatial transformation constraints calculated from the feature point matching pairs. The spatial transformation constraints are expressed in the form of relative pose transformations, which are obtained by minimizing the sum of the Euclidean distances of all matched feature point pairs.

[0032] Using the spatial distance residual between road boundary feature points in non-adjacent frame point cloud data as a constraint, loop closure edges are established. Non-adjacent frame point cloud data refers to two frames of point cloud data whose timestamps are not adjacent but whose spatial locations are close. Loop closures are detected using an appearance-based location re-identification method. A loop closure is considered detected when the cosine similarity of the feature point descriptors of the two frames exceeds a preset similarity threshold. The preset similarity threshold is set to 0.85, based on the fact that the accuracy of feature point matching reaches over 95% when the cosine similarity of the feature descriptors is greater than 0.85. For the two frames of point cloud data where loop closures are detected, feature point matching is performed again, and the spatial distance residual is calculated. Loop closure edges are then established based on the spatial distance residual.

[0033] A weight coefficient is assigned to each edge in the spatial relative pose graph, and the weight coefficient is inversely proportional to the magnitude of the spatial distance residual. For any edge in the spatial relative pose graph, the spatial distance residuals of all matching feature point pairs corresponding to that edge are obtained, the root mean square value of the spatial distance residuals is calculated, and the reciprocal of the root mean square value is used as the weight coefficient of that edge. The larger the weight coefficient, the higher the reliability of the spatial transformation constraint.

[0034] The spatial relative pose graph is generated by combining all nodes and edges from the spatial relative pose graph. The spatial relative pose graph is stored using a graph data structure. Nodes store the index and pose variables of the frame point cloud data, and edges store the indexes of the two connected nodes, spatial transformation constraints, and weight coefficients.

[0035] Initialize the global pose variables for each frame of point cloud data, and fix the global pose of the first node to the origin of the global coordinate system. The global pose variables include rotation and translation components. The rotation components are represented as quaternions, and the translation components are represented as three-dimensional vectors. The global pose variable of the first node is set to the rotation component. Translation component In subsequent optimization processes, the global pose variable value of the first node remains unchanged.

[0036] For each edge in the spatial relative pose graph, calculate the deviation between the current global pose transformation matrix of the two nodes connected by the edge and the spatial transformation constraint carried by the edge. Connecting nodes and nodes , record nodes The current global pose variable is ,node The current global pose variable is ,side The spatial transformation constraints carried are Then the deviation The calculation formula is: ; in, and All The rigid body transformation matrix consists of rotation matrices and translation vectors; It is the observation relative pose transformation matrix; operators This represents finding the inverse of a matrix. Indicates from The logarithmic mapping from a rigid body transformation matrix to a six-dimensional Lie algebra vector, where the six-dimensional Lie algebra vector is... ,in For three-dimensional translation components, For three-dimensional rotational components; operators This indicates the conversion of the Lie algebra matrix form into a six-dimensional vector form. It is a six-dimensional deviation vector, with the first three dimensions corresponding to translational deviation and the last three dimensions corresponding to rotational deviation.

[0037] The spatial relative pose graph is linearized based on the deviation, constructing a system of incremental equations for the pose variables. After calculating the deviation for all edges, a first-order Taylor expansion is used to linearize the deviation, expressing the deviation for each edge as a linear function of the increment of the node pose variables. Combining the linearized equations for all edges yields a system of linear equations for the increments of all node pose variables, denoted as the increment vector. Then the incremental equation system is ,in It is the overall Jacobian matrix formed by concatenating the Jacobian matrices of all edges row by row. It is a diagonal weight matrix composed of the weight coefficients of each edge. It is the overall deviation vector formed by concatenating the deviations of all edges row by row. The incremental equations are stored and solved in sparse matrix form.

[0038] The incremental equations are solved iteratively, and the global pose variables of each node are updated after each iteration. The incremental vector is obtained by solving the incremental equations using the sparse Choliski decomposition method. Increment vector Each element in the array corresponds to the pose update of a node. For the ... For each node, extract the corresponding six-dimensional update value from the increment vector. The six-dimensional update quantity is converted into a pose update matrix through exponential mapping. Then update the node's global pose variable: ,in The pose matrix before the update. This is the updated pose matrix. This process is iterative.

[0039] When the sum of the deviations is less than a preset convergence threshold, the iteration stops and the global pose variables of each node are output as the optimal pose estimate. The sum of the deviations is defined as the deviation vector of all edges. The sum of norms. The preset convergence threshold is set to... The setting is based on the premise that geometric errors caused by changes in pose estimation are visually indistinguishable when they are less than 1 mm. When the total deviation is less than... When the iteration terminates, the rotation quaternion and translation vector in the current global pose variables of each node are output as the optimal pose estimate.

[0040] See Figure 5In the figure, the horizontal axis represents the spatial distance residual in meters, and the vertical axis represents the frequency within the corresponding residual interval. In the legend, the blue bars represent the distribution of the spatial distance residual before optimization, and the red bars represent the distribution of the spatial distance residual after optimization. The residual distribution before optimization exhibits a right-skewed long-tailed shape, with most residuals concentrated in the 0 to 1.5 meter range, and a small number exceeding 2 meters, indicating significant errors and instability in the initial pose estimation. The residuals after optimization are significantly clustered in the minimum interval, mainly distributed between 0 and 0.1 meters, and the frequency is significantly increased, showing that the optimization process effectively reduced the spatial distance residual and improved the accuracy and consistency of pose estimation. The peak height of the optimized distribution is much higher than that before optimization, indicating that the spatial distance error of most matching feature point pairs is compressed to the millimeter level.

[0041] In specific implementation, please refer to Figure 3 The optimal pose estimate for each frame of point cloud data is obtained. This optimal pose estimate includes rotation and translation components. The optimal pose estimate originates from the output of the spatial relative pose graph solution step, and each frame of point cloud data corresponds to a set of optimal pose estimates. For the ... In frame point cloud data, the rotation component in the optimal pose estimate is represented in quaternion form. ,in Let be the real part of the quaternion. , , These are the three imaginary coefficients of the quaternion. The translation component in the optimal pose estimate is represented in three-dimensional vector form as follows: ,in , , These are the translation vectors in the global coordinate system. axis, axis, The component along the axial direction.

[0042] Construct the rotation matrix based on the rotation components, and construct the translation vector based on the translation components. The construction of the rotation matrix follows the transformation relationship from quaternions to rotation matrices. For the ... Frame point cloud data, composed of quaternions The calculated rotation matrix It is Orthogonal matrix, rotation matrix The values ​​of each element are determined by the quaternion components. A translation vector is a three-dimensional vector. No additional construction is required.

[0043] The local coordinates of each road boundary feature point in a single frame of point cloud data are multiplied by a rotation matrix and then added to a translation vector to obtain the global coordinates of the road boundary feature points in the global coordinate system. For the first... The first frame of point cloud data The local coordinates of the road boundary feature points in the local coordinate system of the vehicle-mounted LiDAR are as follows: The global coordinates of the feature points corresponding to the road boundary. Calculated using the following formula: ; in, Indicates the first The first frame of point cloud data The three-dimensional coordinate vectors of each road boundary feature point in the local coordinate system of the vehicle-mounted LiDAR. Indicates by the first The rotation matrix is ​​constructed from the rotation components of the optimal pose estimate corresponding to the frame point cloud data. Indicates by the first The translation vector is constructed from the translation components of the optimal pose estimate corresponding to the frame point cloud data. This represents the three-dimensional coordinate vector of the feature point of the road boundary in the global coordinate system after coordinate transformation.

[0044] A local-to-global coordinate transformation is performed on all road boundary feature points in all frame point cloud data to obtain a road boundary feature point cloud map. All frame point cloud data are traversed, and for each road boundary feature point in each frame, the global coordinates are calculated according to the above coordinate transformation method. All calculated global coordinate points are collected into a 3D point cloud dataset, which is the road boundary feature point cloud map. Each point in the road boundary feature point cloud map contains global 3D coordinate information, as well as scan timestamp and scan azimuth attributes inherited from the original road boundary feature points.

[0045] See Figure 6 The figure illustrates the two-dimensional distribution of road boundary feature points processed according to the method described in this embodiment. The horizontal axis represents the distance along the X-axis in the global coordinate system, in meters, ranging from 0 meters to approximately 50 meters; the vertical axis represents the distance along the Y-axis in the global coordinate system, in meters, ranging from -5 meters to 5 meters. The road boundary feature points are marked as dots in the legend, indicating that the coordinate system transformation was completed by sequentially multiplying the local coordinates in each frame of point cloud data by the corresponding rotation matrix and adding a translation vector, thus obtaining the three-dimensional positions of all road boundary feature points in the global coordinate system.

[0046] The two clearly parallel and evenly distributed point cloud stripes in the figure obviously correspond to the left and right boundaries of the road, located in fixed areas of approximately Y=1.5 meters and Y=-2 meters respectively. The point cloud is continuously distributed along the X-axis without obvious breaks, indicating that the point cloud data of each frame has been accurately aligned according to the optimal pose estimation value, ensuring the spatial continuity and integrity of the boundary feature points. The uniform density and stable distribution of the point cloud demonstrate the effectiveness and accuracy of the optimal pose estimation value obtained by solving the spatial relative pose map in this embodiment, ensuring the high-quality construction of the feature point cloud map.

[0047] In practice, unlabeled feature points are selected as seed points in the road boundary feature point cloud map, and a spatial neighborhood search window is established centered on the seed points. Each feature point in the road boundary feature point cloud map maintains a labeling state, with the initial value of the labeling state being unlabeled. All feature points in the road boundary feature point cloud map are traversed, and when a traversed feature point is found to be unlabeled, that feature point is selected as the seed point. The spatial neighborhood search window is a three-dimensional spherical spatial region centered on the three-dimensional coordinates of the seed point, and the radius of the three-dimensional spherical spatial region is a preset clustering radius.

[0048] Within the search window, feature points whose spatial distance from the seed point is less than the preset clustering radius are searched, and these feature points are grouped into the same feature point cluster as the seed point. The preset clustering radius is set to 0.5 meters, based on the fact that road widths are generally 3.5 to 3.75 meters, and feature points along road boundaries should be continuously distributed within a 0.5-meter range. Using the seed point's 3D coordinates as the center, the Euclidean distance between the 3D coordinates of all unlabeled feature points within the search window and the seed point's 3D coordinates is calculated. Unlabeled feature points with an Euclidean distance less than 0.5 meters are all grouped into the current feature point cluster, and the labeling status of these newly grouped feature points is updated to labeled.

[0049] Using newly added feature points as seed points, iteratively perform spatial neighborhood search until the feature point cluster stops expanding. For each newly added feature point in the current feature point cluster, use that feature point as a new seed point. Establish a new spatial neighborhood search window centered on the 3D coordinates of the new seed point. Within the new spatial neighborhood search window, search for unlabeled feature points whose spatial distance from the new seed point is less than the preset clustering radius. Add the found unlabeled feature points to the current feature point cluster and update their labeling status. Repeat the above seed point selection and neighborhood search process. When no new feature points are added to the feature point cluster in a single search round, the current feature point cluster stops expanding.

[0050] Line fitting is performed on all feature points within the feature point cluster to obtain the boundary line segment. The least squares method is used to fit a three-dimensional line to all feature points within the cluster; the resulting spatial line is the boundary line segment. A covariance matrix is ​​constructed for the three-dimensional coordinates of all feature points within the cluster. The eigenvector corresponding to the largest eigenvalue of the covariance matrix is ​​calculated; this eigenvector determines the spatial extension direction of the boundary line segment. The mean of the three-dimensional coordinates of all feature points within the cluster is calculated; this mean determines the spatial location points traversed by the boundary line segment.

[0051] Calculate the coordinates of the two endpoints of each boundary segment in the global coordinate system and the spatial extension direction vector of the boundary segment. For a boundary segment, project all feature points within the feature point cluster onto a straight line along the spatial extension direction vector. Find the two feature points corresponding to the maximum and minimum projected coordinates. Use the 3D coordinates of the feature point corresponding to the maximum projected coordinate as the coordinates of the first endpoint of the boundary segment, and the 3D coordinates of the feature point corresponding to the minimum projected coordinate as the coordinates of the second endpoint of the boundary segment. The spatial extension direction vector is the direction vector determined by the eigenvector corresponding to the largest eigenvalue of the covariance matrix during the straight line fitting process. The spatial extension direction vector is a unit vector, representing the spatial orientation of the boundary segment in the global coordinate system.

[0052] The determination of whether adjacent boundary segments belong to the same road boundary is based on the spatial distance between their endpoints and the angle between their spatial extension direction vectors. The Euclidean distance between the first or second endpoint of the first boundary segment and the first or second endpoint of the second boundary segment is calculated, and the minimum Euclidean distance among the four endpoint pairs is taken as the spatial distance between the endpoints. The angle between the spatial extension direction vectors of the first and second boundary segments is calculated using the inverse cosine of the dot product of the two spatial extension direction vectors, with a value ranging from 0 to 90 degrees. The preset distance threshold for the spatial distance between endpoints is set to 1.0 meter, based on the assumption that the spacing at discontinuities in road boundaries typically does not exceed 1.0 meter. The preset angle threshold for the angle between spatial extension direction vectors is set to 15 degrees, based on the assumption that the extension direction of the same road boundary will not exhibit abrupt changes exceeding 15 degrees. When the spatial distance between endpoints is less than 1.0 meter and the angle between spatial extension direction vectors is less than 15 degrees, adjacent boundary segments are determined to belong to the same road boundary.

[0053] Adjacent boundary segments belonging to the same road boundary are merged to form a continuous road boundary vector line. For two adjacent boundary segments identified as belonging to the same road boundary, the three-dimensional coordinates of all feature points in the two adjacent boundary segments are calculated. After merging these three-dimensional coordinates, straight line fitting is performed again to obtain the merged road boundary vector line. The road boundary vector line is represented in the form of a three-dimensional spatial line segment, storing the three-dimensional coordinates of the starting point, the three-dimensional coordinates of the ending point, and the spatial extension direction vector of the road boundary vector line.

[0054] By combining all road boundary vector lines and their spatial extension direction vectors, a vector map containing road geometric topology information is generated. The vector map contains the geometric data of all road boundary vector lines. The geometric data of each road boundary vector line includes the 3D coordinates of its starting point, the 3D coordinates of its ending point, its spatial extension direction vector, and the identifier of the road boundary it belongs to. All road boundary vector lines of the same road boundary share the same road boundary identifier, while road boundary vector lines of different road boundaries have different road boundary identifiers. The connectivity between road boundaries in the vector map is implicitly expressed through the spatial proximity of the endpoint coordinates of the road boundary vector lines.

[0055] See Figure 7 The figure shows the feature point clusters and their corresponding boundary line segment fitting results in the feature point cloud map of the road boundary described in this embodiment. The coordinate axes represent the spatial positions in the global coordinate system, with the horizontal axis being the X-coordinate, ranging from 0 to 40 meters, and the vertical axis being the Y-coordinate, ranging from -4 to 4 meters. The figure identifies four different feature point clusters with scattered dots of different colors and shapes: feature point cluster 1 (blue dots), feature point cluster 2 (red triangles), feature point cluster 3 (green squares), and feature point cluster 4 (orange rhombuses). The feature points in each cluster exhibit a clear hierarchical distribution along the Y-axis. Feature point clusters 1 and 2 are concentrated around the Y-axis at approximately 2 meters, while feature point clusters 3 and 4 are concentrated around the Y-axis at approximately -2 meters, demonstrating the spatial distribution characteristics of different road boundaries.

[0056] The feature points within each feature point cluster are fitted with a straight line in three-dimensional space using black lines to form boundary segments. The fitted boundary segments extend primarily horizontally. From the X-axis coordinates, the X-coordinate range for feature point clusters 1 and 3 is approximately 5 to 15 meters, while that for feature point clusters 2 and 4 is approximately 22 to 32 meters, demonstrating the spatial continuity and segmentation of the road boundary segments. The spatial extension direction of the boundary segments is consistent with the overall distribution direction of the feature point clusters, meeting the technical requirement of obtaining boundary segments through least squares fitting of feature point clusters in this embodiment.

[0057] The figure shows clear spatial intervals between different feature point clusters, and feature points within the same cluster are closely clustered, with spatial distances all less than the preset clustering radius of 0.5 meters, demonstrating the effectiveness of spatial neighborhood clustering. The spatial distances and directional angles between the endpoints of the boundary line segments do not show any abnormalities, meeting the preset threshold for merging adjacent boundary line segments in this embodiment.

[0058] In practical implementation, the Levenberg-Marquardt (LM) algorithm is used to solve the incremental equations during the iterative solution process, and the damping factor is updated after each iteration. The LM algorithm, also known as the Levenberg-Marquardt algorithm, constructs a damped incremental equation system in each iteration. The damped incremental equation system has the following form: ; in, It is the overall Jacobian matrix formed by concatenating the Jacobian matrices of all edges row by row. It is a diagonal weight matrix composed of the weight coefficients of each edge. It is the overall deviation vector formed by concatenating the deviations of all edges row by row. For incremental vectors, Represents the overall Jacobian matrix transpose, Represented by matrix A diagonal matrix composed of the diagonal elements. Damping factor It is a positive real scalar. Damping factor The initial value is set to The rationale for this setting is that the LM algorithm requires a large damping factor in the initial stage to ensure the convergence stability of the iterative solution. It is the balance point between the Gauss-Newton method and the gradient descent method.

[0059] The damping factor is updated based on the changing trend of the current total deviation. The current total deviation is defined as the deviation vector of all edges in the current iteration step. The sum of norms. In each iteration, the current damping factor is used. Solving the damped incremental equations yields the current incremental vector. Using the current increment vector Temporarily update the global pose variables of all nodes and calculate the sum of the deviations after the temporary update. Compare the sum of the deviations after the temporary update with the sum of the deviations before the update.

[0060] The damping factor is decreased when the total sum of deviations decreases. A decrease in the total sum of deviations, meaning the sum of deviations after the temporary update is less than the sum of deviations before the update, indicates that the linearization approximation of the current iteration step is relatively accurate. Decreasing the damping factor makes the algorithm closer to the Gauss-Newton method. The damping factor is decreased by adjusting the current damping factor... Divide by the damping factor adjustment coefficient ,Right now Damping factor adjustment coefficient The value of is fixed at 10, which is based on the fact that the scaling factor of the damping factor is usually set to 10 in the standard practice of the LM algorithm.

[0061] The damping factor is increased when the total deviation increases or remains constant. An increase in the total deviation means the updated deviation is greater than the original deviation; a constant deviation means the updated deviation equals the original deviation. This indicates that the linearization approximation of the current iteration step is inaccurate, and increasing the damping factor makes the algorithm closer to gradient descent. The damping factor is increased by adjusting the current damping factor... Multiply by the damping factor adjustment coefficient ,Right now .

[0062] After updating the damping factor, the damped incremental equations are solved again using the updated damping factor to obtain a new incremental vector. The global pose variables of all nodes are then formally updated based on this new incremental vector. After the formal update is complete, the sum of the updated deviations is recorded, and the next iteration begins. Each iteration repeats the above damping factor update process until the sum of the deviations is less than a preset convergence threshold.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A map construction method for topographic surveying in driving test subject three, characterized in that, include: In response to the driving start command of the test vehicle on the set route, the vehicle-mounted lidar collects multiple frames of continuous point cloud data of the road section it passes through. Based on the scanning angular resolution parameters of the vehicle-mounted lidar and the spatial overlap of adjacent frame point cloud data, the effective scanning coverage boundary corresponding to each frame point cloud data is calculated. Based on the effective scan coverage boundary, road boundary feature points are extracted from the point cloud data of each frame. The road boundary feature points include road surface elevation jump points and road surface reflection intensity abrupt change points. Using the road boundary feature points as spatial constraint nodes, a spatial relative pose graph is constructed between point cloud data of each frame. The edges of the spatial relative pose graph represent the spatial transformation constraint relationship between two adjacent point cloud data frames. Solve the spatial relative pose diagram to obtain the optimal pose estimate of each frame of point cloud data in the global coordinate system; Based on the optimal pose estimation value, the road boundary feature points in each frame of point cloud data are projected onto the global coordinate system to generate a road boundary feature point cloud map. Spatial neighborhood clustering is performed on the feature points in the road boundary feature point cloud map to extract continuous boundary line segments, and a vector map containing road geometric topology information is generated based on the spatial extension direction of the boundary line segments.

2. The map construction method for topographic surveying in driving test subject three according to claim 1, characterized in that, Based on the scanning angular resolution parameters of the vehicle-mounted lidar and the spatial overlap of adjacent frame point cloud data, the effective scanning coverage boundary corresponding to each frame of point cloud data is calculated, specifically including: Obtain the horizontal and vertical angular resolutions of the vehicle-mounted LiDAR; Based on the horizontal angular resolution and the vertical angular resolution, determine the spatial azimuth distribution matrix of each scan point in a single frame of point cloud data; Spatial overlap calculation is performed on adjacent frame point cloud data to obtain the spatial overlap of the adjacent frame point cloud data; Based on the spatial azimuth distribution matrix and the spatial overlap, determine the starting and ending azimuth angles of the effective scan points in a single frame of point cloud data; The effective scanning coverage boundary of the point cloud data in this frame is defined by the starting azimuth angle and the ending azimuth angle.

3. The map construction method for topographic surveying in driving test subject three according to claim 2, characterized in that, Based on the effective scan coverage boundary, road boundary feature points are extracted from the point cloud data of each frame, specifically including: Within the effective scan coverage boundary, the spatial distance and elevation difference between each scan point and its adjacent scan points are calculated point by point; Scanning points whose spatial distance is less than a preset neighborhood radius and whose elevation difference is greater than a preset jump threshold are marked as road surface elevation jump points. Scanning points where the difference in laser reflection intensity between adjacent scanning points is greater than a preset intensity change threshold are marked as road surface reflection intensity change points. By merging the road surface elevation jump points and the road surface reflection intensity abrupt change points, the road boundary feature points are obtained; Associate each road boundary feature point with its corresponding scan timestamp and scan azimuth angle.

4. The map construction method for topographic surveying in driving test subject three according to claim 1, characterized in that, Using the road boundary feature points as spatial constraint nodes, a spatial relative pose map is constructed between point cloud data of each frame, specifically including: Each frame of point cloud data is used as a node in the pose graph, and an initial pose estimate is assigned to each node. Using the spatial distance residual between the road boundary feature points in adjacent frame point cloud data as a constraint, an edge is established between adjacent frames; Using the spatial distance residuals between the road boundary feature points in non-adjacent frame point cloud data as constraints, loop edges are established. Each edge is assigned a weight coefficient, which is inversely proportional to the magnitude of the spatial distance residual. By combining all the nodes and edges, the spatial relative pose graph is generated.

5. The map construction method for topographic surveying in driving test subject three according to claim 4, characterized in that, Solving the spatial relative pose diagram yields the optimal pose estimate of each frame of point cloud data in the global coordinate system, specifically including: Initialize the global pose variables of the point cloud data in each frame, and fix the global pose of the first node to the origin of the global coordinate system; For each edge in the spatial relative pose graph, calculate the deviation between the current global pose transformation matrix of the two nodes connected by the edge and the spatial transformation constraint carried by the edge; The spatial relative pose diagram is linearized based on the deviation to construct an incremental set of equations for the pose variables. The incremental equations are solved iteratively, and the global pose variables of each node are updated after each iteration. When the sum of the deviations is less than the preset convergence threshold, the iteration stops and the global pose variables of each node are output as the optimal pose estimate.

6. The map construction method for topographic surveying in driving test subject three according to claim 5, characterized in that, Based on the optimal pose estimation value, the road boundary feature points in each frame of point cloud data are projected onto the global coordinate system, specifically including: Obtain the optimal pose estimation value corresponding to each frame of point cloud data, wherein the optimal pose estimation value includes rotation components and translation components; Construct a rotation matrix based on the rotation components and a translation vector based on the translation components; The local coordinates of each road boundary feature point in a single frame of point cloud data are multiplied by the rotation matrix and then added to the translation vector to obtain the global coordinates of the road boundary feature point in the global coordinate system. Perform the local coordinate to global coordinate transformation on all road boundary feature points in all frame point cloud data to obtain the road boundary feature point cloud map.

7. The map construction method for topographic surveying in driving test subject three according to claim 1, characterized in that, Spatial neighborhood clustering is performed on the feature points in the road boundary feature point cloud map to extract continuous boundary line segments, specifically including: In the road boundary feature point cloud map, unmarked feature points are selected as seed points, and a spatial neighborhood search window is established with the seed points as the center. Within the search window, find feature points whose spatial distance from the seed point is less than a preset clustering radius, and group the found feature points and the seed point into the same feature point cluster. Using the newly incorporated feature points as new seed points, the spatial neighborhood search is iteratively performed until the feature point cluster no longer expands. The boundary line segment is obtained by performing straight line fitting on all feature points within the feature point cluster.

8. A map construction method for topographic surveying in driving test subject three according to claim 7, characterized in that, Based on the spatial extension direction of the boundary line segments, a vector map containing road geometric topology information is generated, specifically including: Calculate the coordinates of the two endpoints of each boundary line segment in the global coordinate system and the spatial extension direction vector of the boundary line segment; Based on the spatial distance between the endpoints of adjacent boundary segments and the angle between the spatial extension direction vectors of adjacent boundary segments, determine whether adjacent boundary segments belong to the same road boundary. Adjacent boundary segments belonging to the same road boundary are merged to form a continuous road boundary vector line; By combining all the road boundary vector lines and their spatial extension direction vectors, a vector map containing road geometric topology information is generated.

9. A map construction method for topographic surveying in driving test subject three according to claim 5, characterized in that, In the iterative solution process, the LM algorithm is used to solve the incremental equation system, and the damping factor is updated after each iteration.

10. A map building system for topographic surveying in driving test subject three, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the map construction method for terrain surveying in driving test subject three as described in any one of claims 1 to 9.

Images