A method for automatically generating a high-precision vector map

Abstract

The application discloses a kind of high-precision vector map automatic generation method, mainly according to laser point cloud intensity, geometric structure feature extraction lane marking, through ground filtering, ground, non-ground point cloud is segmented;Adopt block grating to convert three-dimensional point cloud into two-dimensional projection image;Using image filtering, edge extraction operator, dynamic adaptive threshold segmentation, morphological analysis and other processing extraction road marking pixel, based on point cloud image mapping relationship extraction marking point cloud;Using the point cloud contour matching method based on knowledge template library carries out marking classification.Based on deep learning convolutional neural network, road sign detection classification is carried out to road image, and the point cloud region of interest is established by image point cloud fusion;Using clustering screening candidate point cloud, based on principal component analysis carries out local projection, based on knowledge template library point cloud matching extraction traffic sign;Finally, based on GPS data, calibration, parameter fitting, vectorization, data association are carried out, and the construction of high-precision vector map is completed.

Description

TECHNICAL FIELD

[0001] The present application relates to data processing, in particular to a high-precision vector map automatic generation method. BACKGROUND

[0002] With the development of unmanned driving technology, the application of unmanned driving technology is becoming more and more widespread, and the high-precision vector map is defined relative to the ordinary map, which provides higher precision and more dimensional map information, and has become an indispensable part of the unmanned driving technology. Road marking and traffic signs are the most basic and core elements of traffic network. Therefore, the extraction of road vector marking line and signboard is one of the key technologies in the fields of high-precision map drawing, vehicle intelligent driving and road environment three-dimensional modeling. The current research mainly focuses on detecting and tracking road traffic signs from video images, but the road vector marking line and signboard detected from the image mainly records the two-dimensional plane information of the traffic elements rather than the three-dimensional information, and the map constructed based on the video image is a sparse map, and its precision cannot meet the needs of accurate navigation positioning and traffic management of automatic driving. The vehicle-mounted mobile laser measurement system can not only quickly measure the three-dimensional coordinates of the surface points of the target object, but also record the reflection intensity information of the target to the laser, providing a reliable data source for extracting high-precision three-dimensional road marking line information. However, due to the massiveness of the vehicle-mounted laser point cloud data, the unevenness of the spatial distribution and the density, and the complexity of the terrain, the extraction of road information based on laser point cloud and the drawing of high-precision vector map bring certain challenges. For the construction of high-precision map, the point cloud data is used in the prior art to construct high-precision map, but the three-dimensional point cloud processing has low automation degree and depends on manual editing. Therefore, a processing method of fusion extraction can be adopted by using laser point cloud, image and GPS positioning information. SUMMARY

[0003] In order to overcome the above-mentioned defects and shortcomings of the prior art, the purpose of the present application is to provide a high-precision vector map automatic generation method, which improves the automation degree of the vector map.

[0004] The purpose of the present application is achieved by the following technical solutions:

[0005] A high-precision vector map automatic generation method comprises:

[0006] The original road laser point cloud data is processed to separate the ground point cloud and the non-ground point cloud, and then the ground point cloud is fitted to obtain the ground;

[0007] The ground point cloud is subjected to block gridding processing, a pixel point and point cloud index mapping relationship is established and saved, pixel gray values are calculated according to the point cloud intensity in each block grid, the elevation image of the ground point cloud and the density image of the non-ground point cloud, and the pixel gray values are converted into a two-dimensional projection image;

[0008] Lane marking pixel points are extracted from the two-dimensional projection image, and are back-projected to the point cloud space to obtain lane marking point clouds;

[0009] The lane marking point clouds are classified;

[0010] Road image data is acquired, and traffic sign detection and classification neural networks are used to extract traffic signs from the road image data;

[0011] The non-ground road point cloud, camera and radar calibration parameter file and traffic sign are fused to establish a point cloud space of interest;

[0012] The point cloud space of interest is processed to extract candidate traffic sign point clouds, a local projection system is established, and the candidate traffic sign point clouds are projected onto a two-dimensional projection plane to generate a gray image

[0013] According to the generated gray image, a three-dimensional center position of the traffic sign is extracted based on an image matching method of a template knowledge base;

[0014] According to GPS high-precision positioning data and point cloud three-dimensional coordinates, data calibration is performed, then the calibrated lane marking point clouds are subjected to polynomial parameter fitting to vectorize them, then data correlation is performed according to the extracted center position and category information of the traffic sign to export a high-precision vector map for automatic driving.

[0015] Further, an elevation difference double-threshold ground filtering method is used to separate the original road laser point cloud data to obtain ground point clouds and non-ground road point clouds, then a random sample consensus algorithm is used to fit and refine the ground point clouds to obtain the ground, specifically, all points in the plane with a fitting error less than a set threshold are counted as local points, then the area of the outer convex polygon of the local points is calculated, and the local point with the maximum area is selected as the ground.

[0016] Further, the ground point cloud is subjected to block gridding processing, a pixel point and point cloud index mapping relationship is established and saved, pixel gray values are calculated according to the point cloud intensity in each block grid, the elevation image of the ground point cloud and the density image of the non-ground point cloud, and the pixel gray values are converted into a two-dimensional projection image, specifically:

[0017] First, the ground point cloud is subjected to block gridding processing, and a two-dimensional matrix is used to save the mapping relationship between the pixel points and the point cloud indexes;

[0018] Then, based on the image gray value calculation formula, the gray value is enhanced:

[0019] The elevation image of the ground point cloud and the pixel grayscale value of the density image of the non-ground point are calculated.

[0020] Further, the lane marking pixel points are extracted from the two-dimensional projection image, and are back-projected to the point cloud space to obtain lane marking point clouds, specifically:

[0021] The two-dimensional projection image is subjected to median filtering to remove salt and pepper noise; an edge detection is performed using a canny operator to generate a ground intensity gradient image; then, a density and elevation binary image is generated based on maximum entropy threshold segmentation, and is combined with the ground intensity gradient image to extract road surface pixel points, and then dynamic adaptive threshold segmentation and morphological erosion and expansion are used to finely extract lane marking pixel points;

[0022] Based on the connected component, an optimization analysis is performed, and then marking is performed.

[0023] According to the pixel and point cloud index mapping relationship, the lane marking point clouds are back-projected to the point cloud space to extract the lane marking point clouds.

[0024] Further, the alpha shape rolling algorithm is used to extract the contour information of each lane marking point cloud, and the minimum circumscribed rectangle size of each lane marking point cloud is determined.

[0025] Further, the lane marking point clouds are classified, specifically:

[0026] First, the geometric features of the minimum circumscribed rectangle of the outer contour are used as a coarse classification standard to extract solid and dashed lines and other irregular road markings; then, a point cloud contour matching method based on a template knowledge base is used to classify the irregular road marking point clouds; specifically, candidate point clouds are sequentially subjected to a rotation incremental ICP template library point cloud contour matching, coordinate unification is performed, and then a KNN algorithm is used to calculate the number of nearest neighbor point clouds, and the ratio of the number of point clouds to the total number of template point clouds is used as a classification standard for classification.

[0027] Further, a deep learning convolutional neural network is used to detect and classify traffic signs to obtain a traffic signboard including a current confidence position, and the confidence position includes bounding box information and classification information.

[0028] Further, a candidate traffic sign point cloud is extracted by processing a point cloud region of interest, and a local projection system is established, and the candidate traffic sign point cloud is projected to a two-dimensional projection plane to generate a grayscale image, specifically:

[0029] According to the bounding box and category information, the center pixel coordinates (u, v) of the sign are calculated, and a buffer is increased on the basis of the center pixel to establish a point cloud region of interest. Then, clustering analysis is performed on the point cloud space of interest, and candidate traffic sign point clouds are extracted in detail by using intensity filtering; according to the candidate traffic sign point clouds obtained by clustering, the normal vector obtained by principal component analysis is used as the Z-axis, the vertical direction is used as the X-axis, and the direction perpendicular to the ZOX plane is used as the third axis, so as to establish a local projection system c-xyz with the center of the candidate point cloud block c as the origin, and project the candidate sign point clouds in the region of interest to a two-dimensional plane to generate a gray image.

[0030] Further, according to the generated gray image, the three-dimensional center position of the traffic sign is extracted based on the image matching method of the template knowledge base, specifically:

[0031] Firstly, image enhancement is realized by using the expansion operation, a gradient image is obtained by using the sobel edge detection algorithm, and a constructed planar template is used as a convolution kernel to perform convolution operation with the gradient image. When the template just covers the sign, the intensity value is maximum, and the corresponding center is the accurate center position of the traffic sign.

[0032] Further, the traffic sign detection and classification neural network is a deep learning convolutional neural network YOLOv5.

[0033] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0034] The method based on gray scale enhancement adopted converts the intensity information of the point cloud into gray scale information, better realizes the conversion of the three-dimensional point cloud into a two-dimensional projection image, realizes dimension reduction compared with directly processing the three-dimensional point cloud, and is conducive to improving the processing speed, and the lane marking in the gray image is obvious, which is conducive to the subsequent image processing effect;

[0035] Then, the canny edge detection operator and the dynamic adaptive threshold segmentation can identify the possible lane marking as much as possible, are robust to noise points, have good processing effect, and have low missing and false extraction rates

[0036] The lane marking classification based on the template knowledge base has good classification effect, and basically all the extracted lane markings can be correctly classified;

[0037] After the extracted lane marking is calibrated based on high-precision GPS data, the lane marking is vectorized into a map according to the automatic driving navigation map, and the effect is good. DETAILED DESCRIPTION

[0038] Figure 1 is a workflow diagram of the present application;

[0039] Figure 2 is a schematic diagram of the elevation difference double-threshold ground filtering of the present application;

[0040] Figure 3 is a template knowledge base point cloud matching effect drawing of the present application;

[0041] Figure 4 is a specific implementation flowchart in the embodiment of the present application. DETAILED DESCRIPTION

[0042] The present application will be further described in conjunction with embodiments, but the embodiments of the present application are not limited thereto.

[0043] As Figures 1-4 shown, a high-precision vector map automatic generation method mainly based on high-precision positioning and traffic element recognition includes the following steps:

[0044] S1 processes original road laser point cloud data to separate ground point cloud and non-ground road point cloud, and then fits the ground point cloud to obtain the ground;

[0045] Specifically, it includes:

[0046] S1.1 first counts the minimum circumscribed bounding box size of the original road laser point cloud data, the maximum and minimum coordinate values, and then performs gridding processing based on the bounding box; the resolution-grid size is set to 1.5m;

[0047] S1.2 performs high-threshold double-threshold ground filtering segmentation based on the elevation information of the point cloud in the grid: the elevation difference double thresholds are the minimum elevation of the point cloud in the grid and the minimum elevation of the grid and the 3*3 neighborhood grid and The minimum elevation difference in the grid, the two thresholds are obtained by experience 0.3-0.5, and each point p in the grid is classified as a roughly determined ground point cloud G or a non-ground road point cloud NG through the two thresholds, as shown below: k

[0048]

[0049] S1.3 uses a random sample consensus algorithm to refine the fitted ground plane point cloud to obtain a fitted ground plane, counts all points in the plane with a fitting error less than a set threshold as local points, and then calculates the area of the outer convex polygon of the local points; repeat the step, and select the local point area maximum fitting plane as the ground.

[0050] S2 performs block gridding processing on the ground point cloud, establishes and saves the pixel and point cloud index mapping relationship, calculates the pixel gray value according to the point cloud intensity in each block grid, the elevation image of the ground point cloud and the density image of the non-ground point cloud, and converts it into a two-dimensional projection image;

[0051] Specifically, ​

[0052] S2.1 Firstly, the ground point cloud is divided into blocks with a size of 200x200 m and gridded with a size of 0.05x0.05 m. The number of rows and columns of the gridded grid is set as the image resolution, i.e. each grid is taken as a pixel point, and a two-dimensional matrix is used to save the mapping relationship between the pixel point and the point cloud index;

[0053] S2.2 The gray value of each pixel of the high-intensity image is enhanced according to the intensity value of each point cloud in the grid: specifically, the gray value of the image pixel is determined according to the proportion of the grid intensity mean value to the maximum intensity of the whole block point cloud. The formula is as follows:

[0054]

[0055] In the formula, G i，j is the gray value corresponding to the grid (i, j), av_I i，j is the intensity mean value of all points in the grid (i, j), and the maximum intensity I max of the whole block point cloud, and the calculation formula is: I max = av_I+δ scal ×std; wherein av I , std are the intensity mean value and the mean square error size of all grids in the block point cloud, and δ scal is a set proportion threshold.

[0056] S2.3 Same as S2.2, the pixel gray values are calculated for the elevation image of the ground point cloud and the density image of the non-ground point cloud, respectively; and the gray value calculation is performed by using the normalization and the truncated threshold normalization method, respectively, and converted into a two-dimensional projection image, which is the bird's eye view ground image.

[0057] S3 The lane marking pixel points are extracted from the two-dimensional projection image, and are back-projected to the point cloud space to obtain the lane marking point cloud.

[0058] Specifically:

[0059] The two-dimensional projection image obtained in the previous step is subjected to median filtering to remove salt and pepper noise; a canny operator is used for edge detection to generate an intensity gradient image; then, based on the maximum entropy threshold segmentation, a density and elevation binary image is generated, which is then combined with the ground intensity gradient image to extract the road surface pixel points, and then based on the dynamic adaptive threshold segmentation and morphological erosion and expansion, the lane marking pixel points are finely extracted.

[0060] The dynamic adaptive threshold segmentation first divides the intensity gradient image into multiple smaller sub-images to alleviate the problem of uneven intensity distribution; then, adaptive threshold segmentation is performed on each sub-image to obtain a binary image of the road marking.

[0061] S4 adopts a point cloud contour matching method based on a template knowledge base to classify lane marking point clouds.

[0062] The template knowledge base is specifically a template library matching of incremental rotation ICP.

[0063] Specifically:

[0064] S4.1, based on connected components, optimizes analysis, merges and optimizes each marking line by calculating the nearest point between markings lines using kd-tree, and then marks, the marking adopts a seed filling algorithm: first traverse the pixel graph, start from a foreground pixel, push its foreground pixel into the stack; perform 3X3 neighborhood traversal search; until the neighborhood does not contain foreground pixels, pop the top pixel of the stack, and mark it with the same label, until the stack is empty.

[0065] S4.2, according to S2.1, extracts lane marking point clouds, and uses alpha shape concave algorithm to extract the contour information of each marking line point cloud, alpha takes the experience value 0.7, and the minimum circumscribed rectangle size of each marking line point cloud contour is counted.

[0066] S4.3, first, according to the minimum circumscribed rectangle size information, calculate the geometric feature information of the rectangular boundary box, the length of the long and short sides, the aspect ratio, the angle of the diagonal line and the contour area as the classification standard, and perform lane marking rough classification to extract solid and dashed lines and other irregular road markings;

[0067] The formula of the diagonal line of the quadrilateral is b^2+d^2-a^2-c^2=2m*n*cos(α), where a, b, c, d are the lengths of the four sides, m, n are the lengths of the diagonal lines, and α is the angle of the diagonal line representing the flatness.

[0068] S4.4, for irregular road marking point clouds, a point cloud matching method based on a template knowledge base is used for classification; that is, the contour of the irregular candidate road marking point cloud is sequentially matched with the road marking point cloud contour in the template library to solve the transformation matrix (4dof) by incremental rotation ICP registration, and then the best pose transformation matrix is used to convert the coordinates of the template point cloud.

[0069] S4.5, using KdTreeFLANN to establish a kdtree for road marking point clouds, using KNN algorithm to calculate the ratio of the nearest neighbor point cloud number to the total template point cloud number as the matching overlap rate, and the overlap rate greater than the set threshold is classified as the template class.

[0070] S5, obtain road image data, use traffic sign detection and classification neural network to extract traffic signs in road image data, get the confidence position of traffic signs contained in the current position, the confidence position includes bounding box information and classification information.

[0071] The traffic sign detection and classification neural network is specifically a deep learning convolutional neural network YOLOv5.

[0072] The bounding box information refers to the maximum circumscribed rectangular frame occupied by the traffic sign plate in the image identified by the deep learning convolutional neural network, and the category information refers to the category of the standing traffic sign plate such as a signal lamp, a speed limit plate, a no parking plate and the like identified by the deep learning convolutional neural network.

[0073] S6 fuses the non-ground road point cloud, the camera and radar calibration parameter file and the traffic sign plate to establish a point cloud region of interest;

[0074] Specifically, it includes:

[0075] The calibration parameter file of the camera and the radar can realize the conversion of the image pixel point coordinates and the laser point cloud coordinates, so as to realize the fusion of the point cloud data and the image data. The fusion mainly adds the RGB information, the confidence position and the category information of the traffic sign in the image to the point cloud data.

[0076] The projection conversion between the world coordinate system and the camera coordinate system is as follows:

[0077]

[0078] The projection conversion from the camera coordinate system to the image coordinate system is as follows:

[0079]

[0080] u=I width -(α / 2π)*I width v=(β / 2π)*I height

[0081] Wherein R and T are the calibration rotation matrix and translation vector of the laser radar respectively. Alpha and beta respectively represent the angle between the connecting line of a certain point in the camera coordinate system and the origin and the Z coordinate axis, the angle between the projection of the connecting line on the XOY plane and X, Iwidth and Iheight represent the width and height of the image.

[0082] S7 processes the point cloud region of interest to extract candidate traffic sign point clouds, establishes a local projection system, projects the candidate traffic sign point clouds to a two-dimensional projection plane, and generates a gray image

[0083] According to the bounding box and the category information, the center pixel coordinates (u, v) of the sign are calculated, and a buffer is increased on the basis of the center pixel to establish a point cloud region of interest. Then, clustering analysis is performed on the point cloud region of interest, and candidate traffic sign point clouds are finely extracted by using intensity filtering; according to the candidate traffic sign point clouds obtained by clustering, a normal vector obtained by principal component analysis is used as a Z-axis, a vertical direction is used as an X-axis, and a direction perpendicular to the ZOX plane is used as a third axis, a local projection system c-xyz is established with the center of the candidate point cloud block c as an origin, and the candidate sign point clouds in the region of interest are projected onto a two-dimensional projection plane to generate a gray image.

[0084] Further, the two-dimensional projection plane is specifically: a local projection coordinate system is established with the normal vector direction of the traffic sign obtained by principal component analysis as a projection direction, a vertical direction and a direction perpendicular to the two directions, and a two-dimensional image of the traffic sign is obtained with the center of the traffic sign as an origin

[0085] The formula for projecting the point cloud of the world coordinate system into the local coordinate system c-xyz is as follows:

[0086]

[0087] The formula for projecting the point cloud from the local coordinate system onto the XOY plane is as follows:

[0088] u=[x max -(x-x min )]*L pixel

[0089] v=[y max -(y-y min )]*L pixel .

[0090] S8 extracts the three-dimensional center position of the traffic sign based on the image matching method of the template knowledge base according to the generated gray image;

[0091] Specifically, first, image enhancement is realized by using dilation operation, a gradient image is obtained by using a sobel edge detection algorithm, and a constructed planar template is used as a convolution kernel to perform convolution operation with the gradient image. When the template just covers the sign, the intensity value is maximum, and the corresponding center is the accurate center position of the traffic sign.

[0092] S9 calibrates data according to GPS high-precision positioning data and point cloud three-dimensional coordinates, then performs polynomial parameter fitting on the calibrated lane marking point cloud to vectorize it; then, according to the center position and category information of the extracted traffic sign, data correlation is performed according to relevant traffic rules, and finally, a high-precision vector map for automatic driving is exported.

[0093] The point cloud three-dimensional coordinates are three-dimensional coordinates measured by the laser radar based on a laser radar coordinate system.

[0094] The method classifies and extracts lane markings according to intensity information and geometric structure features of laser point clouds, segments the ground, non-ground point clouds through ground filtering, converts the three-dimensional point cloud space into a two-dimensional projection image through block gridding, extracts road marking pixels through image filtering, edge extraction operators, dynamic adaptive threshold segmentation, morphological analysis and other image processing operations, extracts marking point clouds through back projection based on the mapping relationship of point cloud images, classifies and identifies markings through a rotating incremental point cloud contour matching method based on a knowledge template library, detects and classifies traffic signs through a neural network method, establishes a region of interest in a point cloud space through image point cloud fusion, filters candidate point clouds through clustering, establishes a local projection system through principal component analysis, extracts and classifies traffic signs through point cloud matching based on a knowledge template library, and finally calibrates lane markings, performs parameter fitting, vectorization, and data correlation and fusion of traffic signs based on high-precision positioning data, to complete construction of a high-precision vector map.

[0095] The above embodiments are preferred embodiments of the present application, but the embodiments of the present application are not limited by the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application are equivalent replacement methods and are included in the protection scope of the present application.

Claims

1. A method for automatically generating a high-precision vector map, characterized by, The application relates to a method for extracting lane marking and traffic sign point clouds from original road laser point cloud data. The original road laser point cloud data is processed to separate ground point clouds and non-ground point clouds, and then the ground point clouds are fitted to obtain a ground surface; The ground point clouds are subjected to block gridding processing, an index mapping relationship between pixel points and point clouds is established and saved, pixel gray values are calculated according to the point cloud intensity in each block grid, the elevation image of the ground point clouds and the density image of the non-ground point clouds, and the pixel gray values are converted into a two-dimensional projection image; Lane marking pixel points are extracted from the two-dimensional projection image, and are inversely projected into the point cloud space to separate lane marking point clouds, specifically as follows: The two-dimensional projection image is subjected to median filtering to remove salt and pepper noise; an edge detection ground intensity gradient image is generated by using a canny operator; then, density and elevation binary images are generated based on maximum entropy threshold segmentation, and the road surface pixel points are extracted in combination with the ground intensity gradient image; and then, lane marking pixel points are finely extracted based on dynamic adaptive threshold segmentation and morphological erosion and expansion; Optimization analysis is carried out based on connected components, and then marking is carried out; According to the index mapping relationship between the pixel points and the point clouds, the lane marking point clouds are inversely projected into the point cloud space to be extracted. The lane marking point clouds are classified: Road image data is acquired, and traffic sign detection and classification neural networks are used to extract traffic signs in the road image data; The non-ground point clouds, camera and radar calibration parameter files and traffic signs are fused to establish a point cloud region of interest; The point cloud region of interest is processed to extract candidate traffic sign point clouds, and a local projection system is established, and the candidate traffic sign point clouds are projected onto a two-dimensional projection plane to generate a gray image, specifically as follows: According to the bounding box and category information, the center pixel coordinates (u, v) of the sign are calculated, a buffer zone is added on the basis of the center pixel to establish a point cloud region of interest, then the point cloud space of interest is subjected to clustering analysis, and candidate traffic sign point clouds are finely extracted by using intensity filtering; according to the candidate traffic sign point clouds obtained through clustering, a normal vector obtained through principal component analysis is used as the Z-axis, the vertical direction is used as the X-axis, and a direction perpendicular to the ZOX plane is used as the third axis, a local projection system c-xyz is established with the candidate point cloud block center c as the origin, and the candidate sign point clouds in the region of interest are projected onto a two-dimensional plane to generate a gray image; According to the generated gray image, a three-dimensional center position of the traffic sign is extracted based on an image matching method of a template knowledge base; According to GPS high-precision positioning data and point cloud three-dimensional coordinates, data calibration is carried out, then the calibrated lane marking point clouds are subjected to polynomial parameter fitting to be vectorized; then, data correlation is carried out according to the center position and category information of the extracted traffic sign, and a high-precision vector map used for automatic driving is exported.

2. The automatic generation method according to claim 1, characterized in that, The original road laser point cloud data is separated by using an elevation difference double-threshold ground filtering method to obtain ground point clouds and non-ground road point clouds, and then a random sample consensus algorithm is used to finely fit the ground point clouds to obtain a ground surface, specifically as follows: all points in a plane with a fitting error less than a set threshold are counted as local points, then the area of an outer convex polygon of the local points is calculated, and a fitting plane with the largest local point area is selected as the ground.

3. The automatic generation method according to claim 1, characterized in that, The ground point cloud is divided into blocks and gridded, the mapping relationship between the pixel points and the point cloud indexes is saved, the pixel gray value is calculated according to the point cloud intensity in each block, the elevation image of the ground point cloud and the density image of the non-ground point cloud, and the two-dimensional projection image is converted, specifically as follows: First, the ground point cloud is divided into blocks and gridded, and the mapping relationship between the pixel points and the point cloud indexes is saved using a two-dimensional matrix; Then, the gray value is calculated based on the image gray value calculation formula, and the gray value is enhanced: The pixel gray value of the elevation image of the ground point cloud and the density image of the non-ground point cloud is calculated.

4. The automatic generation method according to claim 1, characterized in that, It also includes using the alpha shape rolling algorithm to extract the contour information of each lane marking point cloud, and determining the minimum circumscribed rectangle size of each lane marking point cloud.

5. The automatic generation method according to claim 1, characterized in that, The lane marking point cloud is classified, specifically as follows: First, the geometric features of the minimum circumscribed rectangle of the outer contour are used as the coarse classification standard to extract solid and dashed lines and other irregular road markings; then, a point cloud contour matching method based on a template knowledge base is used to classify the irregular road marking point cloud; specifically, the candidate point cloud is sequentially subjected to a rotating incremental ICP template library point cloud contour matching, coordinates are unified, and then a KNN algorithm is used to calculate the nearest neighbor point cloud number, and the ratio of the number to the total number of template point clouds is used as a classification standard for classification.

6. The automatic generation method according to claim 1, characterized in that, A deep learning convolutional neural network is used for traffic sign detection and classification to obtain a traffic sign including a current confidence position, and the confidence position includes bounding box information and classification information.

7. The automatic generation method according to claim 1, characterized in that, According to the generated gray image, a template knowledge base-based image matching method is used to extract the three-dimensional center position of the traffic sign, specifically as follows: First, an expansion operation is used to realize image enhancement, a sobel edge detection algorithm is used to obtain a gradient image, and a constructed planar template is used as a convolution kernel to perform convolution operation with the gradient image, when the template just covers the sign, the intensity value is maximum, and the corresponding center is the accurate center position of the traffic sign.

8. The automatic generation method according to claim 1, characterized in that, The traffic sign detection and classification neural network is a deep learning convolutional neural network YOLOv5.

Images