Multi-source feature lane edge line identification method and system based on scoring mechanism

Abstract

The application relates to the technical field of lane edge detection, and provides a multi-source feature lane edge line identification method and system based on a scoring mechanism, which comprises the following steps: acquiring road point cloud data and carrying out pretreatment to obtain road region point cloud data; fusing color information and laser intensity information to score, processing the road region point cloud to obtain lane line point cloud; aiming at the lane line point cloud, extracting boundary points based on a central main axis and normal feature jump scoring to obtain the edge line of each lane line; and smoothing and fitting the edge line of the lane line to obtain the fitted lane edge line. The application constructs a color and laser intensity double-scoring function, adaptively determines an intensity threshold value through an OTSU algorithm, and fuses the intensity threshold value and an RGB color difference in a weighted manner, so that the recognition accuracy and continuity of lane marking points are effectively improved, and the problems of extraction breakage and misrecognition in traditional methods are overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lane edge detection, in particular to a multi-source feature lane edge line identification method and system based on a scoring mechanism. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] With the rapid development of transportation infrastructure, the number and scale of highway reconstruction projects continue to grow, and the requirements for road structure precision and safety are constantly improving. As a key information carrier in road geometric features, lane edge lines not only carry the definition function of lane space boundaries, but also have important application value in the fields of road asset digital modeling, intelligent driving assistance systems, and transportation infrastructure operation. Especially in the maintenance and reconstruction of multi-lane highways, accurate extraction and geometric description of lane edge lines have become an important support for high-precision map construction and road condition perception. With the development of Mobile Laser Scanning (MLS) technology, point cloud data has become one of the core data sources for collecting highway geometric information due to its high density and three-dimensional spatial expression capability. However, how to efficiently and accurately extract multi-lane edge lines from complex highway point clouds remains a technical challenge in the industry.

[0004] The existing lane edge line extraction has the following problems: on the one hand, the traditional extraction method uses a laser reflection intensity threshold method, which is difficult to deal with the problem of uneven reflection caused by different road surface materials, lighting conditions, and reflection interference, often causing edge line extraction to break or become fuzzy, or even misjudgment. On the other hand, the noise, uneven density, and small-scale disturbances in point cloud data seriously affect the accurate positioning of edge points and the continuity of the curve, resulting in "burr" and "drift" phenomena in the line segment fitting process. At the same time, most current algorithms only focus on the overall profile or single-sided extraction of lane lines, making it difficult to meet the modeling needs of structural levels. SUMMARY

[0005] To solve the above problems, the present application proposes a multi-source feature lane edge line identification method and system based on a scoring mechanism, which constructs a color and laser intensity double scoring function, adaptively determines the intensity threshold value through the OTSU algorithm, and weights the fusion with the RGB color difference, effectively improving the recognition accuracy and continuity of lane marker points, overcoming the problems of extraction breakage and misidentification in traditional methods.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] The first aspect of the present application provides a multi-source feature lane edge line identification method based on a scoring mechanism, comprising the following steps:

[0008] Obtaining road point cloud data and preprocessing to obtain road region point cloud data;

[0009] Fusing color information and laser intensity information to calculate weighted scores, processing the road region point cloud, segmenting the lane line and the background, and obtaining lane line point cloud;

[0010] For the lane line point cloud, based on the center main axis and the normal feature jump score, taking the center main axis as the reference, the normal feature jump score is calculated based on the intensity and color change in the normal direction, and the boundary points are extracted to obtain the edge line of each lane line;

[0011] The edge line of the lane line is smoothed and fitted to obtain the fitted lane edge line.

[0012] The second aspect of the present application provides a multi-source feature lane edge line identification system based on a scoring mechanism, comprising:

[0013] The road point cloud extraction module is configured to obtain road point cloud data and preprocess to obtain road region point cloud data;

[0014] The lane line extraction module is configured to fuse color information and laser intensity information to calculate weighted scores, process the road region point cloud, segment the lane line and the background, and obtain lane line point cloud;

[0015] The multi-lane edge line extraction module is configured to, for the lane line point cloud, based on the center main axis and the normal feature jump score, taking the center main axis as the reference, the normal feature jump score is calculated based on the intensity and color change in the normal direction, and the boundary points are extracted to obtain the edge line of each lane line;

[0016] The multi-lane edge line curve fitting module is configured to smooth and fit the edge line of the lane line to obtain the fitted lane edge line.

[0017] The third aspect of the present application provides a multi-source feature lane edge line identification system based on a scoring mechanism, comprising a point cloud data acquisition device and a processor;

[0018] The point cloud data acquisition device is used to obtain road point cloud data;

[0019] The processor is configured to execute the steps in the multi-source feature lane edge line identification method based on the scoring mechanism described above;

[0020] Compared with the prior art, the present application has the following advantages:

[0021] The method of the present application fuses color information and laser reflection intensity features in point cloud data for lane line extraction, selects lane line points by weighting and fusing color scores and intensity scores, realizes joint analysis of point cloud data in color domain and intensity domain, and effectively overcomes misjudgment problems caused by uneven color distribution and intensity abnormalities.

[0022] Meanwhile, aiming at the problems of boundary blur, edge jitter and misjudgment caused by sparse point cloud in the process of lane marking line edge extraction, the present application proposes a double boundary extraction method based on center main axis and normal feature jump score. This method takes the fitted center main axis curve as the reference, analyzes the intensity and color mutation in the normal direction, accurately locates the left and right boundary points, and effectively improves the stability and geometric precision of boundary point extraction.

[0023] The advantages of the present application and the advantages of the additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0024] The drawings accompanying the specification of the present application form a part thereof and serve to provide further understanding of the present application, the illustrative embodiments of the present application and its description serve to explain the present application, and do not constitute limitations thereof.

[0025] Figure 1 is a flowchart of the lane edge line recognition method of embodiment 1 of the present application;

[0026] Figure 2 is a whole architecture diagram of the lane edge line recognition method of embodiment 1 of the present application;

[0027] Figure 3 is a method flowchart of the lane line extraction of embodiment 1 of the present application;

[0028] Figure 4 is a lane name schematic diagram of embodiment 1 of the present application;

[0029] Figure 5 is an example road marking clustering result schematic diagram of embodiment 1 of the present application;

[0030] Figure 6 is an effect schematic diagram of lane line point cloud extraction of embodiment 1 of the present application;

[0031] Figure 7 is an edge line extraction effect schematic diagram of the lane line of embodiment 1 of the present application;

[0032] Figure 8 is a lane edge line fitting effect diagram of embodiment 1 of the present application;

[0033] Figure 9 is a curve after vectorization processing of embodiment 1 of the present application;

[0034] Figure 10is a schematic diagram of the extraction point in the boundary point identification process of the lane edge line of the embodiment 1 of the present application;

[0035] Figure 11 is a boundary point identification result diagram of the lane edge line of the embodiment 1 of the present application; DETAILED DESCRIPTION

[0036] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0037] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0038] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and furthermore, it should be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof. It should be noted that the various embodiments and features in the present application can be combined with each other without conflict. The embodiments will be described in detail below in conjunction with the accompanying drawings.

[0039] The present application aims at the problems of difficult extraction and poor continuity of lane lines and edge lines in road point clouds, and proposes a multi-source feature lane edge line high-precision identification method based on CRI-Score scoring mechanism. In the aspect of lane marking line extraction, the CRI-Score fusion method is innovatively proposed, which comprehensively utilizes the RGB color and laser intensity features, constructs a scoring function and weighted fusion, effectively improves the segmentation accuracy of lane lines and road surface background, and solves the problems of discontinuity and high misjudgment rate of traditional intensity method. In the aspect of edge line extraction, the orthogonal normal is constructed in combination with the center principal axis of the lane line, the color and intensity jump scores are calculated by sampling and calculating in the normal direction, so as to stably extract the left and right boundary points. This method is suitable for complex situations such as lane line width change and point cloud sparseness, and improves the continuity and accuracy of edge extraction. The algorithm structure of the present application is clear, high in precision and strong in adaptability, meets the engineering application requirements of high-precision extraction and expression of lane edge lines in multi-lane complex scenes, and the following specific embodiments are described.

[0040] Embodiment 1

[0041] In one or more embodiments disclosed in the technical solutions, as shown in Figures 1 to 11 A multi-source feature lane edge line identification method based on a scoring mechanism includes the following steps:

[0042] Step 1, obtain road point cloud data and pre-process to obtain road region point cloud data;

[0043] Step 2, calculate weighted score by fusing color information and laser intensity information, process road region point cloud, segment lane line and background, and obtain lane line point cloud;

[0044] Step 3, for lane line point cloud, based on center main axis and normal feature jump score, take center main axis as reference, calculate normal feature jump score in normal direction based on intensity and color change, extract boundary points to obtain edge line of each lane line;

[0045] Step 4, smooth fit the edge line of lane line to obtain the fitted lane edge line;

[0046] The method of the embodiment utilizes RGB color information and laser reflection intensity features in point cloud data, proposes CRI-Score (Colour-Reflectivity-Intensity Score) fusion algorithm for lane line extraction, fuses multiple source features by weighted fusion of color score and intensity score, selects lane line points, realizes joint analysis of point cloud data in color domain and intensity domain, and effectively overcomes misjudgment problems caused by uneven color distribution and intensity anomalies.

[0047] At the same time, for the problems of boundary blur, edge jitter and misjudgment caused by sparse point cloud in the process of lane marking line edge extraction, the embodiment proposes a double boundary extraction method based on center main axis and normal feature jump score. This method takes the fitted center main axis curve as reference, analyzes intensity and color mutation in the normal direction, accurately locates left and right boundary points, and effectively improves the stability and geometric precision of boundary point extraction.

[0048] RGB color is a method of representing color by mixing red (Red), green (Green) and blue (Blue) three basic colors;

[0049] In step 1, point cloud data can be collected by MLS system. MLS (Mobile Laser Scanning) is a system that installs laser radar equipment on a mobile carrier (such as a car, a drone) to collect three-dimensional data.

[0050] The pre-processing of point cloud data can include filtering and denoising and interference point removal in sequence to obtain road region point cloud data;

[0051] Optionally, the filtering denoising method adopts a statistical filtering-based method: a statistical filter is used to analyze the neighborhood distance distribution of each point, and points with a distance mean deviation exceeding a set threshold are removed, and outliers are automatically determined according to the set threshold;

[0052] The filtering denoising method of the above embodiment can remove dust, raindrops and other random and isolated non-structural noise points, and significantly improve the overall stability and continuity of the point cloud data.

[0053] Optionally, the interference point removal can adopt a region growing method, including the following steps:

[0054] Step 11, initial seed point selection: selecting a plane point with an angle between the surface normal and the Z-axis less than an angle set value and a height lower than a height set value as a seed;

[0055] In a three-dimensional space, each point cloud point can estimate a local fitting plane through its neighborhood points (i.e. surrounding points); the local fitting plane is assumed to be a local model of the surface of the object to which the point belongs; the "surface normal" refers to a unit vector perpendicular to the local fitting plane, which represents the spatial inclination degree and direction of the surface where the point is located; which can reflect the spatial inclination degree and direction of the surface where the point is located; the Z-axis is the vertical direction, i.e. the "upward" or "perpendicular to the ground" direction; in the point cloud, the direction of the Z-axis is usually used to represent the height or altitude;

[0056] Two thresholds are set: one is the angle threshold between the normal and the vertical direction (Z-axis), and the other is the height threshold of the point. When the angle between the surface normal and the Z-axis is less than the set angle threshold; the Z coordinate (height) of the point is lower than the set height threshold, which is set as a seed point.

[0057] Step 12, taking the seed point as the starting point, taking the normal similarity and the distance between points as the growth conditions, adding the adjacent points meeting the growth conditions to the growing region for expansion, and iterating until no further growth is possible, finally extracting a complete road point cloud region with similar geometric characteristics and continuity;

[0058] The normal similarity is that the angle between the surface normal of the plane where the seed point is located and the surface normal of the current point is less than the set threshold;

[0059] The distance between points is the spatial distance from the seed point, which is less than the set distance;

[0060] In this embodiment, based on the region growing algorithm, the flat geometric characteristics and the normal direction consistency of the road surface are combined to exclude vegetation and buildings and other interference, and efficient segmentation of the road region point cloud is realized.

[0061] To address the difficulties, inaccuracies, and discontinuities in lane line extraction caused by uneven distribution of point cloud intensity features, step 2 of this embodiment employs a filtering extraction method that fuses RGB color and intensity. Lane lines typically possess high-brightness color features (such as white or yellow) and strong reflectivity. Therefore, scoring functions are constructed for color and intensity respectively to quantify the closeness of each point to the ideal lane line features. The color score is calculated based on the difference between the point's RGB color and the target color, while the intensity score is based on the OTSU algorithm to adaptively determine the threshold for reflectivity, highlighting high-intensity areas. Finally, a comprehensive score is obtained by weighted fusion of the two scores, and a threshold is set to filter out highly reliable candidate lane line points, achieving an organic unity and complementary extraction of color and intensity information.

[0062] Among them, the OTSU algorithm (translated as: Otsu thresholding algorithm) is an adaptive global thresholding method based on image gray-level histogram analysis.

[0063] Step 2 involves lane line extraction. The road area point cloud data extracted in Step 1 is processed to identify lane lines. A scoring method (CRI-Score) is used, fusing color information and laser intensity information to segment the lane lines from the background, resulting in the lane line point cloud. This method includes the following steps:

[0064] Step 21: Calculate the color score: For the point cloud data of the road area, extract the RGB color of each point cloud point, calculate the color difference between the RGB color of each point cloud data and the target color of the lane, and obtain the color score;

[0065] The target lane color is the standard lane line color, which is generally white or yellow.

[0066] Each point in a point cloud The color is represented as Define the target lane line color, specifically including: white lane lines. Yellow lane lines ;

[0067] Step 211: Calculate the color difference between the RGB color of each point and the target color of the lane using Euclidean distance;

[0068] Iterate through and extract the RGB color information of each point in the regional road point cloud data, and use Euclidean distance to separate the color differences. Defined as:

[0069] ;

[0070] in, include and , The smaller the color difference value is, the closer the color of the point is to the color of the lane line.

[0071] Step 212, calculating the color score according to the ratio of the color difference to the maximum value of the color difference in the RGB space;

[0072] The color difference is changed into the color score, and the score range is [0, 1], the higher the score is, the closer the color is to the lane line, and the color score formula is:

[0073] ;

[0074] Wherein: is the maximum distance in the RGB space;

[0075] Step 22, calculating the intensity score: for the road area point cloud data, the reflection intensity of the point cloud is mapped to the gray value, and the OTSU algorithm is used to adaptively determine the intensity segmentation threshold; the normalized score function is constructed according to the relationship between the intensity value of each point and the intensity segmentation threshold , the intensity score of each point cloud close to the lane line feature degree is obtained, and the process is as follows:

[0076] Step 221, for the road area point cloud data, the reflection intensity of the point cloud is mapped to the gray value, and the threshold method is used for partitioning;

[0077] Specifically, the reflection intensity of the point cloud is mapped to the gray value, the threshold method is used for partitioning, and the point cloud is divided into a high-intensity region and a low-intensity region to distinguish potential lane lines and non-lane regions;

[0078] For example, assuming that the gray value range of a gray image is between 0 and 255, the image can be divided into two regions by setting a threshold: one contains pixels with a gray value less than or equal to the threshold, and the other contains pixels with a gray value greater than the threshold; each region is a class;

[0079] Step 222, calculating the inter-class variance between each partition, and extracting the maximum inter-class variance;

[0080] The inter-class variance is calculated , and the formula is as follows:

[0081] ;

[0082] Wherein, T is the candidate threshold; L is the total number of gray levels; is the probability distribution of the gray level ;

[0083] The maximum inter-class variance is calculated, and the goal of OTSU threshold segmentation is to find the threshold that makes the inter-class variance maximum threshold value The maximum inter-class variance is calculated as follows:

[0084] ;

[0085] Step 223, based on the reflection intensity of each point and the maximum inter-class variance, the intensity score of each point of the point cloud is calculated;

[0086] According to the threshold value output by OTSU, a score function based on the relative intensity difference is designed , and linearly normalized to the interval [0, 1]:

[0087] ;

[0088] Wherein, represents the intensity of the current point, represents the maximum intensity of the road region point cloud;

[0089] If the intensity of the point cloud data point is higher than the threshold value , it enters the scoring area, and the higher the intensity, the closer to 1; if it is lower than the threshold value , it is directly set to 0 and does not participate in scoring;

[0090] Step 23, the color score and intensity score of each point of the point cloud are weighted and fused to obtain a fusion score, and the calculation formula is as follows:

[0091] ;

[0092] Wherein: and are the set weights, and ;

[0093] Step 24, screening candidate points: based on the threshold method, the point cloud points with a fusion score greater than a set threshold value are screened out as marker line point clouds;

[0094] Set a threshold value , only keep the points with a fusion score higher than this threshold value, traverse all road point clouds, , will be classified as marker line point clouds, thereby realizing accurate extraction of the markers in the lane.

[0095] Step 25, marker line point cloud segmentation: based on the Euclidean distance of the point cloud, the marker line point cloud after screening is clustered to separate the lane line from the internal marker of the road, and the lane line point cloud is obtained;

[0096] This step is to group the point cloud data using the Euclidean clustering method to extract the lane marking and other lane marking area features. This method clusters based on the Euclidean distance of the point cloud, generates multiple clustering areas, and filters out effective clusters by setting appropriate clustering radius and minimum cluster point number to separate lane lines and internal road markings, with the lane middle area being the area between two lane lines.

[0097] As shown in Figure 5 is a road marking clustering diagram, the markings on the road include lane lines and markings in the area between lane lines, wherein, Figure 5 The two blue straight lines are lane lines, and the green line and green arrow are internal road markings. Through clustering in step 25, internal road markings can be filtered out, and lane line point cloud data, i.e., lane line point cloud, is retained. Figure 5 In (a), the road markings include lane line markings and internal markings, Figure 5 In (b), the lane line markings, Figure 5 and in (c), the internal road markings.

[0098] In the implementation of step 2 of the present embodiment, an RGB color and intensity fusion filtering extraction method is used to accurately extract lane line point cloud, a color and laser intensity double scoring function is constructed, an intensity threshold is adaptively determined by OTSU algorithm, and is weighted and fused with RGB color difference, effectively improving the recognition accuracy and continuity of lane marking points, and realizing high-precision segmentation of lane line point cloud and road surface background. Compared with traditional lane line extraction, after determining the optimal segmentation threshold by maximizing the inter-class variance, the present embodiment weightedly fuses RGB color information, effectively solving the problems of discontinuity and inaccuracy caused by uneven intensity distribution.

[0099] Specific extraction examples are shown in Figure 6 , which shows the effect changes in the process of extracting lane lines. Figure 6 In (a), the point cloud data after executing step 1 filtering and denoising, i.e., road surface point cloud data, is shown. Figure 6 In (b), the road area point cloud after executing step 1 interference point removal based on the region growing method is shown. Figure 6 In (c), the lane line point cloud obtained by executing step 2 processing is shown.

[0100] In step 3, a multi-lane edge line extraction step is performed based on the obtained lane lines to identify the edges of the lane lines. As shown in Figure 6 (c), the lane line point cloud, and Figure 5 the blue line in (c) is the lane line point cloud. Figure 5 There are two lane lines, each with a certain width, so each lane line has two edge lines, and step 3 is to identify the edge lines of the lane lines.

[0101] For lane line point clouds, a method for extracting the edge lines of each lane line based on the central principal axis and normal feature jump scoring, using the central principal axis as a reference, and performing normal feature jump scoring based on intensity and color changes in the normal direction, includes the following steps:

[0102] Step 31, PCA Center Principal Axis Extraction: Principal component analysis is used to determine the direction of the lane lines and extract the center principal axis of the lane lines;

[0103] Step 311: Principal Component Analysis (PCA) is used to estimate the direction of the lane line point cloud set, and the first principal component direction vector is obtained. As the overall direction of the lane lines, in the direction vector The above uses a fixed step size Perform projection slicing to obtain point cloud slices;

[0104] Each lane line marker point cloud appears as a band-like structure, approximately 0.15 meters wide. The morphological centerline needs to be extracted from this set of points as a reference for subsequent normal analysis and boundary extraction.

[0105] Optionally, principal component analysis (PCA) can be used to analyze the lane line point cloud set. Perform direction estimation and take the direction vector of the first principal component. As the overall direction of the lane lines. Then in the direction vector. The above uses a fixed step size Projection slicing is performed, with each slice representing a transverse section of the lane line, and the transverse section and direction vector... Vertically, the points contained in the k-th slice are denoted as... .

[0106] Step 312: Calculate the average lateral coordinates of each projected slice as the center point to obtain an ordered set of center points. ;

[0107] The formula for calculating the horizontal average coordinate is:

[0108] ;

[0109] Step 313: Perform cubic B-spline curve fitting on the set of center points to obtain the central principal axis o of the lane line, denoted as... , which is the normalized arc length, and the parameter s takes the value [0,1].

[0110] Step 32: Determine the normal direction of the center axis o of the lane line, and perform normal sampling to obtain candidate boundary points;

[0111] Step 321: Construct the normal direction of the central axis o of the lane line;

[0112] The normal direction of the lane line is the orthogonal direction of the central main axis, and the normal direction of the lane line is the central main axis At each point s on the central main axis of the lane line, the tangent direction is calculated and the corresponding unit normal vector is constructed, that is, the tangent vector is rotated counterclockwise by 90 degrees to obtain the unit normal vector, and the formula is as follows:

[0113] ;

[0114] ;

[0115] Wherein, is the unit vector of the tangent direction; is the unit normal vector of the normal direction, is the tangent direction angle.

[0116] Step 322, at each center point, a plurality of points are sampled in the normal direction to obtain left and right boundary candidate points;

[0117] Given the maximum expansion distance , the sampling interval , M sampling points are generated in the left and right directions:

[0118] The sampling formula of the left boundary sampling point is:

[0119] ;

[0120] The sampling formula of the right boundary sampling point is:

[0121] ;

[0122] Wherein, ， ;

[0123] The left sampling point and the right sampling point constitute the left and right boundary candidate points, and each candidate point corresponds to a position in space, and the RGB color information and intensity information of the point can be obtained by mapping, as shown in Figure 10 and Figure 11 , the center point on the central main axis o is represented as a green point, and the points other than the green point in the normal direction of the green point include blue and red points as boundary candidate points;

[0124] Step 33, for the obtained boundary candidate points, the feature jump score is calculated based on the intensity and color change, and the point with the highest score is selected as the boundary point; that is, the color and intensity are compared point by point to see if they have changed, and the point where the change occurs is the demarcation point of the road surface and the edge line, as shown in Figure 10 , the red point is the boundary point;

[0125] At the edge of the lane line, there is a significant difference between the road surface and the lane line, the road surface is gray, and the lane line is white or yellow; the RGB color information and intensity information of the point cloud points at the edge are prone to mutation, and the edge points are prone to mis-extraction, breakage or noise, etc., and only distance is easy to misjudge;

[0126] In this embodiment, the boundary score function is innovatively constructed Auxiliary screening is performed, and the left and right sampling points are calculated respectively, so that the boundary score function is:

[0127] ;

[0128] Among them, represents a color vector, and the color vector of the j-th sampling point in the boundary candidate point; represents the reflection intensity of the j-th sampling point in the boundary candidate point; and are the weights of the reflection intensity and the color change score, respectively, is the Euclidean distance of the two RGB color vectors; +1, j and j-1 are adjacent points sampled in the normal direction of the central main axis o; Specifically, as shown in and

[0129] , M points are extracted in the normal direction of the left and right directions of the central main axis, and the numbering is sequentially increased from the central axis, such as j-1, j, j+1. If the outermost point is j, j+1 does not exist, and its data is 0. In this case, the score function value is the maximum, that is, a mutation occurs. In order to illustrate, Figure 10 in the figure, j+1 is drawn outside the lane edge line; Figure 11 Figure 10 Let the sampling point be the signed distance of the sampling point relative to the normal

[0130] , wherein the distance of the left point is negative, and the distance of the right point is positive. In order to avoid the interference of color mutation or abnormal reflection intensity in the area near the normal in the judgment, in the first K sampling points farthest from the normal on each side, the score function is The maximum one is selected as the boundary point, and the specific process is as follows: Left boundary point extraction, formula:

[0131] ;

[0132] ;

[0133] ;

[0134] Right boundary point extraction, formula: ​

[0135] ;

[0136] ;

[0137] wherein, represents the jth boundary candidate point; is the offset distance of the sampling point relative to the lane center principal axis, on the left side of the normal line, on the right side of the normal line; is the boundary score function; represents the index set of the first K points farthest from the normal line on the left side; represents the index set of the first K points farthest from the normal line on the left and right sides; represents the first K points selected by distance from the points that meet the conditions; 、 represents the sampling point number finally selected as the left and right boundary points.

[0138] The range of the window range K is 3 by default, which can be adjusted to adapt to different point densities; the final selection 、 as the left and right boundary points.

[0139] Step 34, cluster and sort all the boundary points on both sides of the lane lines to obtain the inside and outside edge point clouds of the two lane lines, i.e. the lane edge lines.

[0140] As shown in Figure 7 , the point cloud of the two lane lines obtained by step 2 is in the form of points; as Figure 11 the left boundary point and the right boundary point in the middle, i.e. the lane edge line of a lane line; with the middle yellow line as the dividing line, it is divided into the right lane line and the left lane line; the right edge line of the right lane line and the left edge line of the left lane are the outside edge lines. The left edge line of the right lane and the right edge line of the left lane are called the inside edge line.

[0141] In the above implementation of the present embodiment, a bilateral boundary extraction method based on the central principal axis and the normal feature jump score is used. This method takes the fitting central principal curve as the reference, analyzes the intensity and color mutation in the normal direction, accurately locates the left and right boundary points, and effectively improves the stability and geometric precision of boundary point extraction. For the problem of lane marking line boundary recognition, the central principal axis of the lane line point cloud is first extracted through directional slicing and average position calculation, and the derivative thereof is used to construct the orthogonal normal direction. Boundary candidate points are sampled at fixed intervals in the normal direction, and a boundary score function is constructed combining the intensity and color gradient, and the boundary points with the most obvious mutation are extracted in each normal direction. Finally, the continuous and smooth lane line bilateral contour is obtained by fitting the left and right boundary point sets, which is suitable for various actual situations such as lane line bending, slight width change and boundary loss.

[0142] In the implementation of step 3, the normal direction is constructed based on the central principal axis, and the left and right edge points of the lane line are identified and fitted with high precision by sampling and feature analysis in the normal direction. Unlike traditional contour extraction methods, the structural features of the lane line are fully utilized, and a combination strategy of "central principal axis + normal extension + feature jump score" is adopted, which can realize stable and continuous boundary extraction in the case of limited width, unclear boundary or uneven point cloud distribution.

[0143] In step 4, multi-lane edge line curve fitting is performed: the edge line of the lane is fitted to obtain the fitted lane edge line; specifically, the lane edge line extracted in step 3 is used, and an improved cubic B-spline curve fitting algorithm and a least squares curve approximation method are used to smooth fit the lane edge to generate a continuous curve, i.e., the fitted lane edge line, including the following steps:

[0144] Step 41, data point parameterization: the data points of the lane edge line are parameterized using the centripetal parameterization method to obtain the parameter value of each data point, so that the point cloud is reasonably distributed in the parameter space;

[0145] The centripetal parameterization method is used to parameterize the point cloud data points of the lane edge line, and the process is as follows:

[0146] Step 411, according to the coordinate values of the point cloud, the chord length between adjacent data points is calculated:

[0147] ;

[0148] ;

[0149] Step 412, the cumulative chord length is taken as the parameter value of the centripetal parameterization, and the calculation formula is:

[0150] ;

[0151] wherein, is the mth data point; is the total number of points; is the chord length between the mth data point and the previous data point; is the cumulative chord length between all data points; is the parameter value corresponding to the data point ; is the adjustment parameter, which is set to a value. Step 42, construct a node vector, determine the starting node and the ending node, and the point cloud data points between the starting node and the ending node are intermediate points. Based on the starting node, the position of each intermediate point is calculated through the parameterized value of the data point.

[0152] Step 43, normalize the obtained parameter value, so that all parameter values are distributed in the interval ;

[0153] Step 44, select control points, establish a relationship matrix with the control points, determine the initial position of the control points by solving the corresponding linear equation system, and generate a preliminary B-spline curve. The linear equation system is as follows:

[0154] ;

[0155] ;

[0156] In the formula, represents the mth lane edge line point cloud data point, such as the red point in Figure 10 ; represents the nth control point, which is a data point on the lane edge line selected according to the distance; represents the B-spline basis function value.

[0157] Step 46, for the obtained B-spline curve, the position of the control points is further optimized by using the least square method, so that the finally generated curve approximates the real edge line, and the fitted lane edge line is obtained, realizing accurate expression of the edge line shape characteristics.

[0158] The B-spline-least square fitting model is:

[0159] ;

[0160] The least square optimization objective function is:

[0161] ;

[0162] In the formula, represents the fitting value at parameter t under the influence of the ith sample point and its surrounding control points; ​The weighting coefficient corresponding to the kth control point is determined by least squares; is the fitting residual; is the kth control point; (t) is an n-order B-spline basis function; ;

[0163] This step improves the point cloud data of the edge line of each lane line extracted by step 3 using a cubic B-spline curve and a least squares fusion algorithm, realizes high-precision smooth fitting and vectorization output of multi-lane edge lines, and effectively solves the curve segment drift and burr problem by optimizing the control point weight and smoothing constraint condition in combination with the centripetal parameterization method. Figure 8 Fig. (a) shows the fitting effect of the lane edge line, wherein the blue curve represents the fitted lane edge line, and the Figure 7 points are connected as a curve; Figure 8 Fig. (b) shows the fitting of the lane edge line and the road marking line after fitting, wherein the red transverse line segment between the two lane edge lines represents the lane line point cloud data obtained in step 2.

[0164] To verify the curve fitting accuracy, 20 fitting points in the test section are randomly selected, and their coordinates are compared with the corresponding original lane line point coordinates. As shown in Table 1, the results show that the average error is less than 1 cm, the fitted curve is highly consistent with the original point cloud, and there is no burr phenomenon, which fully verifies the high precision and good stability of the method in lane line extraction.

[0165] Table 1 Comparison of fitted coordinates and original coordinates of test section;

[0166]

[0167] The fitted lane edge line output by step 4 is discrete data composed of points, i.e. point cloud data format, and further includes a vectorization processing step that converts the discrete data composed of points into a vector form with clear topology and geometric structure, outputting a curve;

[0168] Vectorization is to derive the curve into DXF format, not point cloud format. After vectorization, it can be used across platforms and can be opened with existing drawing software;

[0169] Optionally, the vectorization processing adopts KD-Tree (K-Dimensional Tree) spatial indexing to topologically sort the fitted lane edge line, i.e. point cloud curve, establishes accurate point cloud adjacency relationship based on adjacent point search, and obtains the lane line after vectorization processing, to ensure the topological connectivity of vector geometry.

[0170] The lane edge line detection and fitting method of the embodiment has shown excellent performance in practical application to highway environment. Not only has it significantly improved the accuracy and efficiency of multi-lane edge line extraction, but also ensured smooth transition of new and old road sections in width, slope and overall geometry, thereby effectively reducing the error rate in the construction process.

[0171] To verify the effectiveness of the lane edge line identification method in the embodiment, a simulation experiment was carried out. For the lane edge line point cloud extracted in step 3, the existing RANSAC (RANdom SAmple Consensus) line fitting algorithm was used for fitting and comparative analysis. Figure 9(a) shows the overall effect of RANSAC line fitting, and Figure 9(b) is a local enlarged view. As can be seen from the figures, RANSAC line fitting has obvious drift and "burr" phenomenon in the curve section, and the fitting accuracy is poor in the curve section. Comparing the lane edge line fitting effect of the method of the embodiment in Figure 9 with Figure 8 It can be seen that the curvature change of the edge line obtained by using the method of the embodiment is more gentle, and there is no obvious drift, which has good smoothness and fitting accuracy.

[0172] Embodiment 2

[0173] Based on embodiment 1, the multi-source feature lane edge line identification system based on scoring mechanism is provided in the embodiment, which includes:

[0174] The road point cloud extraction module is configured to obtain road point cloud data and pre-process to obtain road region point cloud data;

[0175] The lane line extraction module is configured to fuse color information and laser intensity information to calculate weighted score, process the road region point cloud, and segment the lane line and the background to obtain lane line point cloud;

[0176] The multi-lane edge line extraction module is configured to, based on the center principal axis and the normal feature jump score, take the center principal axis as the reference, and perform normal feature jump score based on intensity and color change in the normal direction to extract boundary points to obtain the edge line of each lane line;

[0177] The multi-lane edge line curve fitting module is configured to perform smooth fitting on the edge line of the lane line to obtain the fitted lane edge line.

[0178] Further, it further includes a vectorization module, which uses KD-Tree spatial index to topologically sort the fitted lane edge line, establishes accurate point cloud adjacency relationship based on adjacent point search, and obtains the lane line after vectorization processing.

[0179] It should be noted that each module in the embodiment corresponds to each step in Embodiment 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0180] Embodiment 3

[0181] Based on Embodiment 1, this embodiment provides a multi-source feature lane edge line identification system based on a scoring mechanism, including a point cloud data acquisition device and a processor.

[0182] The point cloud data acquisition device is used to acquire road point cloud data.

[0183] The processor is configured to perform the steps in the multi-source feature lane edge line identification method based on the scoring mechanism described in Embodiment 1.

[0184] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0185] Although the specific embodiments of the present application are described above in combination with the accompanying drawings, it is not intended to limit the protection scope of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A multi-source feature lane edge line identification method based on a scoring mechanism, characterized in that, The method comprises the following steps: Obtain road point cloud data and preprocess to obtain road region point cloud data; Fuse color information and laser intensity information to calculate a weighted score, process the road region point cloud, and separate the lane line from the background to obtain lane line point cloud; For the lane line point cloud, based on the central principal axis and the normal feature jump score, take the central principal axis as the reference, and perform normal feature jump scoring in the normal direction based on intensity and color change, and extract the boundary points to obtain the edge line of each lane line; Smoothly fit the edge line of the lane line to obtain the fitted lane edge line; The method for fusing color information and laser intensity information to calculate a weighted score, processing the road region point cloud, and separating the lane line from the background to obtain lane line point cloud comprises the following steps: For the road region point cloud data, extract the RGB color value of each point cloud point, calculate the color distance between the RGB color value of each point and the lane target color, and obtain a color score; For the road region point cloud data, map the reflection intensity of the point cloud to a gray value, and adaptively determine the intensity segmentation threshold value by using the OTSU algorithm; construct a normalized score function according to the relationship between the intensity value of each point and the intensity segmentation threshold value to obtain the intensity score of each point cloud in terms of the degree of proximity to the lane line feature; and perform weighted fusion on the color score and the intensity score of each point of the point cloud to obtain a fusion score; Based on the threshold method, the point cloud points with a fusion score greater than a set threshold value are screened out as marker line point cloud; For the screened marker line point cloud, perform clustering based on the Euclidean distance of the point cloud to separate the lane line from the marker point cloud between the lanes to obtain lane line point cloud; The color score is calculated according to the formula: ； wherein: is the maximum distance possible in the RGB space, is the color distance, is the color of each point in the point cloud is the color of each point in the point cloud is the white lane line color, is the yellow lane line color; The intensity score is calculated according to the formula: ； wherein, represents the intensity of the current point, represents the maximum intensity of the road region point cloud, is the threshold corresponding to the maximum inter-class variance, is the inter-class variance, T is the candidate threshold, and L is the total number of gray levels, is the gray level probability distribution; The fusion score is calculated according to the formula: ； wherein and are set weights, and ; The method for extracting the lane edge line comprises the following steps: determine the direction of the lane line by using principal component analysis to extract the central principal axis of the lane line; determine the normal direction of the central principal axis of the lane line, perform normal sampling to obtain boundary candidate points; for the obtained boundary candidate points, perform feature jump scoring based on intensity and color change, select the point with the highest score as the boundary point, and obtain the extracted boundary point; and cluster and sort the boundary points on both sides of all lane lines to obtain the inside and outside edge point cloud of the lane line, i.e., the lane edge line.

2. The multi-source feature lane edge line recognition method based on the scoring mechanism according to claim 1, wherein: The preprocessing of the point cloud data comprises sequentially performing filtering denoising and interference point removal: the filtering denoising method adopts a statistical filtering method, and the interference point removal adopts a region growing method.

3. The multi-source feature lane edge line identification method based on a scoring mechanism according to claim 1, characterized in that: The method for determining the intensity score of each point cloud in terms of the degree of proximity to the lane line feature comprises the following steps: For the road region point cloud data, map the reflection intensity of the point cloud to a gray value, and perform partitioning based on the threshold method; Calculate the inter-class variance between each partition, and extract the maximum inter-class variance; Based on the reflection intensity of each point and the maximum inter-class variance, calculate the intensity score of each point of the point cloud.

4. The multi-source feature lane edge line identification method based on a scoring mechanism according to claim 1, characterized in that: Smoothly fit the edge line of the lane line to generate the fitted lane edge line, comprising the following steps: The point cloud data points of the lane edge line are parameterized by using a centripetal parameterization method to obtain parameter values of each point cloud data point; The obtained parameter values are normalized; A node vector is constructed, a starting node and an ending node are determined, and the point cloud data points between the starting node and the ending node are intermediate points. The position of each intermediate point is calculated based on the parameterization value of the data point and the starting node as a reference; Control points are selected, a relationship matrix between the control points is established, the initial positions of the control points are determined by solving the linear equation of the B-spline curve, and a preliminary B-spline curve is generated; The positions of the control points are further optimized by using the least square method based on the obtained B-spline curve, so that the finally generated curve approximates the real edge line, and the fitted lane edge line is obtained.

5. The multi-source feature lane edge line identification method based on a scoring mechanism according to claim 1, characterized in that: It also includes a vectorization processing step, which uses KD-Tree spatial indexing to topologically sort the fitted lane edge line, establishes an accurate point cloud adjacency relationship based on neighboring point search, and obtains the lane line after vectorization processing.

6. A multi-source feature lane edge line recognition system based on a scoring mechanism, characterized in that, It includes: A road point cloud extraction module configured to obtain road point cloud data and preprocess it to obtain road region point cloud data; A lane line extraction module configured to fuse color information and laser intensity information to calculate a weighted score, process the road region point cloud, and segment the lane line from the background to obtain lane line point cloud; A multi-lane edge line extraction module configured to extract the edge line of each lane line based on the center principal axis and the normal feature jump score in the normal direction based on the intensity and color change of the center principal axis; A multi-lane edge line curve fitting module configured to smooth and fit the edge line of the lane line to obtain the fitted lane edge line; The method of fusing color information and laser intensity information to calculate a weighted score, processing the road region point cloud, and segmenting the lane line from the background to obtain lane line point cloud includes the following steps: for the road region point cloud data, extracting the RGB color value of each point cloud point, calculating the color difference between the RGB color value of each point and the lane target color, and obtaining a color score; For the road region point cloud data, the reflectivity intensity of the point cloud is mapped to a gray value, and the OTSU algorithm is used to adaptively determine the intensity segmentation threshold; a normalized score function is constructed based on the relationship between the intensity value of each point and the intensity segmentation threshold to obtain the intensity score of each point cloud close to the lane line feature; the color score and the intensity score of each point of the point cloud are weighted and fused to obtain a fusion score; Based on the threshold method, the points of the point cloud with a fusion score greater than a set threshold are selected as marker line point clouds; For the screened marker line point cloud, clustering is performed based on the Euclidean distance of the point cloud to separate the lane line and the marker point cloud between the lanes, and the lane line point cloud is obtained. The color score is calculated by the formula: ； wherein: is the maximum distance possible in the RGB space, is the color distance, is the color of each point in the point cloud is the color of each point in the point cloud is the white lane line color, is the yellow lane line color; The intensity score is calculated by the formula: ； wherein, represents the intensity of the current point, represents the maximum intensity of the road region point cloud, is the threshold corresponding to the maximum inter-class variance, is the inter-class variance, T is the candidate threshold, and L is the total number of gray levels, is the probability distribution of the gray level . The fusion score is calculated by the formula: ； wherein, and are set weights, and ; The method for extracting lane edge lines comprises the following steps: determining the direction of lane lines by using a principal component analysis method, and extracting the central principal axis of the lane lines; determining the normal of the central principal axis of the lane lines, and performing normal sampling to obtain boundary candidate points; performing feature jump scoring based on intensity and color change on the obtained boundary candidate points, selecting the point with the highest score as a boundary point, and obtaining extracted boundary points; clustering and sorting the boundary points on both sides of all lane lines to obtain inside and outside edge point clouds of the lane lines, i.e. lane edge lines.

7. The multi-source feature lane edge line identification system based on a scoring mechanism of claim 6, wherein: The method further comprises a vectorization module, which performs topological sorting on the fitted lane edge lines by using a KD-Tree spatial index, establishes an accurate point cloud adjacency relationship based on a neighbor point search, and obtains vectorized lane lines.

8. A multi-source feature lane edge line identification system based on a scoring mechanism, characterized in that: The method comprises a point cloud data acquisition device and a processor. The point cloud data acquisition device is used to acquire road point cloud data. The processor is configured to perform the steps in the multi-source feature lane edge line identification method based on the scoring mechanism in any one of claims 1-5.

Images