A lane line automatic extraction method based on vehicle-mounted radar point cloud

Abstract

The application discloses a lane line automatic extraction method based on a vehicle-mounted radar point cloud, relates to the technical field of lane line automatic extraction, and is characterized in that: for each ordered scattered point set, the first point in the set is taken as a starting point, a least square method is used to fit a multi-segment line section by section, and in the fitting process, a preset node distance threshold value, a filtering intensity threshold value, a filtering distance threshold value and an interpolation node quantity threshold value are used to judge the position of a segmented node and filter noise points until the fitting of all points in the ordered scattered point set is completed. The application adopts a double Gaussian function to fit a point cloud intensity distribution curve, takes a double Gaussian function wave trough value as a lane line contour edge point cloud intensity value, adopts a directional triangulation network growth algorithm to calculate a target region triangulation network, reduces the calculation amount in the point cloud triangulation network construction process, and adopts a fitting node distance threshold value, a filtering intensity threshold value, a filtering distance threshold value and an interpolation node quantity threshold value to jointly participate in the lane line multi-segment line fitting, so that the fitting effect is better.

Description

Technical Field

[0001] This invention relates to the field of automatic lane line extraction technology, and in particular to an automatic lane line extraction method based on vehicle-mounted radar point clouds. Background Technology

[0002] In vehicle-mounted laser point cloud data, each measurement point contains not only three-dimensional coordinate information (x, y, h) but also echo intensity information. Echo intensity is influenced by factors such as the target's material, color, surface roughness, and moisture content. In the point cloud file, the intensity information is stored as a 16-bit unsigned integer, ranging from 0 to 65535. In highway engineering, because lane markings and ordinary road surfaces have different materials, the echo intensity of the laser point cloud differs; the echo intensity within lane markings is significantly higher than that within ordinary road surfaces.

[0003] Furthermore, since the radar is closer to the target when the vehicle-mounted laser scanning is performed, its point cloud density and accuracy are superior to those of the airborne point cloud. If the echo intensity of the vehicle-mounted point cloud is displayed in grayscale, the lane lines can be clearly identified.

[0004] The manual lane line extraction method based on vehicle radar point cloud involves drawing lane lines using the grayscale image of point cloud echo intensity as the base map. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an automatic lane line extraction method based on vehicle-mounted radar point clouds. The following technical solution is adopted:

[0006] An automatic lane line extraction method based on vehicle-mounted radar point clouds includes the following steps:

[0007] Step 1: Obtain the echo intensity value distribution of the road surface point cloud, use a double Gaussian function to fit the intensity distribution curve, and set the intensity value corresponding to the trough value of the double Gaussian function fitting curve as the target intensity value of the lane line outline edge.

[0008] Step 2: Define points in the point cloud whose echo intensity value is greater than the target intensity value as strong echo points and points whose echo intensity value is less than the target intensity value as weak echo points. Based on the planar coordinates and echo intensity values ​​of the strong and weak echo points, construct an oriented triangulation network that includes only the target intensity value using the triangulation growth method. During the construction process, skip all triangulation operations that do not involve the target intensity value within the triangles.

[0009] Step 3: Extract all edges connecting strong echo points and weak echo points in the directional triangulation network as target edges. Linearly interpolate the positions on each target edge where the intensity value is equal to the target intensity value as scattered points. Sort and number all scattered points to form a sorted scattered point set.

[0010] Step 4: For each sorted scatter point set, starting from the first point in the set, use the least squares method to fit a multi-segment line segment by segment. During the fitting process, determine the position of the segment nodes and filter noise based on the preset node distance threshold, filter intensity threshold, filter distance threshold and interpolation node number threshold, until the fitting of all points in the sorted scatter point set is completed, and output each segment of the multi-segment line as the lane line edge contour.

[0011] Optionally, in step 1, when statistically analyzing the point cloud echo intensity values, frequency statistics are performed using a preset fixed intensity range as the unit, and the fitting object of the double Gaussian function is the distribution histogram after frequency statistics.

[0012] Optionally, in step 1, the double Gaussian function model is:

[0013] ;

[0014] in It is the amplitude of the Gaussian function model at ordinary measuring points on the road surface. It is the peak position of the Gaussian function model of a common measuring point on the road surface. It is the standard deviation of the Gaussian function model of ordinary measuring points on the road surface. It is the amplitude of the Gaussian function model of the road lane line measuring point. It is the peak position of the Gaussian function model of the road lane line measuring point. It is the standard deviation of the Gaussian function model of the road lane line measuring points. It is the independent variable, representing the echo intensity value corresponding to a single point cloud measurement point. It is the dependent variable, representing the point cloud statistical frequency under the corresponding echo intensity value;

[0015] Based on the least squares method for parameters , , , , , Perform fitting, and take the fitted curve at... and The x-axis coordinate value corresponding to the minimum point between the two values ​​is taken as the target intensity value.

[0016] Optionally, in step 2, the construction of the oriented triangulation includes two stages: generating the initial edge and expanding the triangle edge by edge.

[0017] The method for generating the starting edge is as follows: traverse all strong echo points, calculate the point with the closest Euclidean distance to each strong echo point, and if the closest point is a weak echo point, then take the line connecting the strong echo point and the weak echo point as a starting edge.

[0018] The method for constructing triangles by expanding edge by edge is as follows: add the starting edge to an edge set, and check whether a triangulation has been formed on both sides of each edge in the edge set. If a triangulation has not been formed, find the third point that satisfies the empty circle property on each edge to form a triangulation. At the same time, add the newly added edge that forms the triangulation to the edge set. Repeat the above process until the triangulation is completed on both sides of all edges.

[0019] Optionally, in step 2, the method for determining whether to skip the mesh construction operation is as follows: if both ends of the current edge to be expanded are strong echo points and the third point found is also a strong echo point, then it is determined that the triangle does not pass through the target intensity value, and the corresponding edge is marked as having completed the triangulation construction without actually generating a triangle; if both ends of the current edge to be expanded are weak echo points and the third point found is also a weak echo point, it is also determined to skip the mesh construction.

[0020] Optionally, in step 3, the method for linear interpolation on each target edge is as follows: obtain the spatial coordinates of the two endpoints of the target edge and their respective echo intensity values, and calculate the planar coordinates of the interpolation point on the target edge whose intensity value is equal to the target intensity value based on the ratio of the difference between the intensity values ​​of the two endpoints and the target intensity value.

[0021] Optionally, in step 3, the specific method for sorting and numbering all scattered points is as follows: store all extracted target edges into an edge set. If two target edges belong to the same Deloni triangle, connect the two target edges sequentially according to the vertex order of the triangle and assign them continuously increasing numbers, so that the scattered points on the target edges belonging to the same connected region form a set of sequentially connected points.

[0022] Optionally, in step 4, the method for polyline fitting is as follows:

[0023] Points are sequentially taken from the sorted scatter set. The first point is taken as the starting point of the current segment and its coordinates are fixed. Subsequent points are added to the current fitted point set one by one. The least squares method is used to calculate the line parameters of the current fitted point set. When the segmentation stopping condition is met, the intersection of the current fitted line and the line connecting the first and last points of the selected points or the projection point of the current last point on the fitted line is taken as the next node of the polyline. Then, the corresponding node is taken as the new starting point of the segmentation and the fitting continues until all points in the point set have been processed.

[0024] Optionally, the segmentation stopping condition is jointly determined by preset node distance threshold, filter strength threshold, filter distance threshold, and interpolation node number threshold:

[0025] The node distance threshold is used to limit the maximum length of a single segment along the lane line direction. If the cumulative distance or the farthest distance from the points already selected in the current fitted point set to the start point of the segment exceeds the threshold, point selection is stopped and nodes are added.

[0026] The filter strength threshold is used to specify the number n outliers to be removed each time the line is fitted. Before each calculation of the line parameters, the residual distance from each point in the current set of fitted points to the fitted line is calculated. The points are sorted from largest to smallest residual distance. After removing the n points with the largest distance values, the least squares fitting is performed again with the remaining points.

[0027] The filtering distance threshold is used to determine whether the lane line profile has changed direction. It calculates the vertical distance from each point in the current fitted point set to the current fitted line. If the maximum vertical distance exceeds the filtering distance threshold, it stops taking points and adds nodes.

[0028] The interpolation node number threshold is used to filter out isolated noise points. If the total number of scattered points in the current sorted scattered point set is less than the interpolation node number threshold, the set is determined to be an invalid set corresponding to isolated strong reflection noise points and will not participate in the polyline fitting.

[0029] Optionally, in step 2, when constructing the directional triangulation, each strong echo point region independently constructs its own triangulation set, and each triangulation set corresponds to an independent connected region. In step 3, each connected region generates an independent sorted scatter point set. In step 4, each independent sorted scatter point set is fitted with an independent polyline, and multiple polylines together form a complete lane line edge profile.

[0030] In summary, the present invention has at least one of the following beneficial technical effects:

[0031] This invention provides an automatic lane line extraction method based on vehicle-mounted radar point clouds. It employs a double Gaussian function to fit the point cloud intensity distribution curve, taking the valley value of the double Gaussian function as the point cloud intensity value of the lane line contour edge. A directional triangular mesh growth algorithm is used to calculate the triangular mesh of the target region, significantly reducing the computational load in the point cloud triangular mesh construction process. Furthermore, by using fitting node distance thresholds, filtering intensity thresholds, filtering distance thresholds, and interpolation node number thresholds together in the lane line multi-segment fitting, better fitting results are achieved. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating an automatic lane line extraction method based on vehicle-mounted radar point clouds according to the present invention.

[0033] Figure 2 This is a schematic diagram of the echo intensity of vehicle point clouds in the existing technology of manual lane line extraction method based on vehicle radar point clouds;

[0034] Figure 3 This is a partially enlarged schematic diagram of the lane lines in a specific embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the distribution of measured laser echo intensity values ​​in a specific embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of a point cloud triangulation according to a specific embodiment of the present invention;

[0037] Figure 6 This is a partially enlarged schematic diagram of the initial edge construction result in a specific embodiment of the present invention;

[0038] Figure 7 This is a schematic diagram of the directional triangulation generation result in a specific embodiment of the present invention;

[0039] Figure 8 This is a schematic diagram of the target intensity scatter plot generation result in a specific embodiment of the present invention;

[0040] Figure 9 This is a schematic diagram of noise at a strong reflection point in a specific embodiment of the present invention;

[0041] Figure 10 This is a schematic diagram illustrating the lane line fitting effect of a specific embodiment of the present invention;

[0042] Figure 11 This is a partially enlarged schematic diagram of the lane line fitting effect in a specific embodiment of the present invention; Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings.

[0044] This invention discloses an automatic lane line extraction method based on vehicle-mounted radar point clouds.

[0045] Reference Figures 1-11 Example 1: An automatic lane line extraction method based on vehicle-mounted radar point clouds, comprising the following steps:

[0046] Step 1: Obtain the echo intensity value distribution of the road surface point cloud, use a double Gaussian function to fit the intensity distribution curve, and set the intensity value corresponding to the trough value of the double Gaussian function fitting curve as the target intensity value of the lane line outline edge.

[0047] Step 2: Define points in the point cloud whose echo intensity value is greater than the target intensity value as strong echo points and points whose echo intensity value is less than the target intensity value as weak echo points. Based on the planar coordinates and echo intensity values ​​of the strong and weak echo points, construct an oriented triangulation network that includes only the target intensity value using the triangulation growth method. During the construction process, skip all triangulation operations that do not involve the target intensity value within the triangles.

[0048] Step 3: Extract all edges connecting strong echo points and weak echo points in the directional triangulation network as target edges. Linearly interpolate the positions on each target edge where the intensity value is equal to the target intensity value as scattered points. Sort and number all scattered points to form a sorted scattered point set.

[0049] Step 4: For each sorted scatter point set, starting from the first point in the set, use the least squares method to fit a multi-segment line segment by segment. During the fitting process, determine the position of the segment nodes and filter noise based on the preset node distance threshold, filter intensity threshold, filter distance threshold and interpolation node number threshold, until the fitting of all points in the sorted scatter point set is completed, and output each segment of the multi-segment line as the lane line edge contour.

[0050] By adopting the above technical solution, the vehicle-mounted LiDAR can simultaneously record the three-dimensional coordinates and echo intensity values ​​when collecting road point clouds. The reflective properties of lane line materials are better than those of ordinary road surfaces. There is a stable numerical difference in the echo intensity of the two types of objects. The edge of the lane line is the dividing point between high and low intensity values. The entire solution relies on this physical characteristic to complete fully automatic extraction and execute a four-layer progressive processing flow.

[0051] The first layer of processing relies on the statistical distribution of the intensity of the point cloud across the entire road surface to distinguish between two types of reflective subjects: the road surface and the lane lines. It uses double Gaussian fitting to lock the standard value corresponding to the intensity transition boundary, which serves as the criterion for identifying the edge of the lane line.

[0052] The second-layer processing divides the reflection points into two types based on the baseline intensity. It abandons the high-computing-power scheme of constructing triangular meshes from the global point cloud and only generates triangular mesh structures in the boundary area between strong and weak measurement points. It filters out pure measurement point triangles of the same type without intensity transition, compresses the overall computational data volume, and only retains the triangular mesh containing the edge position of the lane line.

[0053] The three-layer processing filters the connecting edges between strong and weak measuring points inside the triangulation network. Linear interpolation is used to calculate the plane points with strength equal to the reference value on each boundary edge. Based on the topological connectivity of the triangulation network, the points are grouped in an orderly manner to form a continuously arranged sequence of edge points.

[0054] The four-layer processing performs straight line fitting for each group of continuous points, introduces four types of constraint thresholds to simultaneously complete noise removal and line segmentation determination, generates a polyline profile that fits the direction of the lane line segment by segment, and combines all polylines to obtain the complete lane line boundary.

[0055] In Example 2, in step 1, when statistically analyzing the point cloud echo intensity values, frequency statistics are performed using a preset fixed intensity interval as the unit, and the fitting object of the double Gaussian function is the distribution histogram after frequency statistics.

[0056] By adopting the above technical solution, the range of vehicle-mounted point cloud echo intensity values ​​covers 0-65535. The original discrete intensity data has no regular distribution pattern, and direct curve fitting is prone to local distortion and peak identification deviation.

[0057] In this embodiment, a fixed and uniform intensity segmentation interval is set before statistically analyzing the intensity distribution. All point cloud intensity values ​​are assigned to the corresponding intervals to complete the frequency statistics of the number of points, generating a regular intensity distribution histogram. The histogram can smooth out the numerical fluctuations caused by discrete measurement points, clearly present the concentrated intensity peaks corresponding to ordinary road surfaces and lane lines, and provide a stable and continuous data source for subsequent fitting of the double Gaussian function, avoiding interference from single-point abnormal intensity values ​​in solving for peak and trough locations.

[0058] Example 3, in step 1, the double Gaussian function model is:

[0059] ;

[0060] in It is the amplitude of the Gaussian function model at ordinary measuring points on the road surface. It is the peak position of the Gaussian function model of a common measuring point on the road surface. It is the standard deviation of the Gaussian function model of ordinary measuring points on the road surface. It is the amplitude of the Gaussian function model of the road lane line measuring point. It is the peak position of the Gaussian function model of the road lane line measuring point. It is the standard deviation of the Gaussian function model of the road lane line measuring points. It is the independent variable, representing the echo intensity value corresponding to a single point cloud measurement point. It is the dependent variable, representing the point cloud statistical frequency under the corresponding echo intensity value;

[0061] Based on the least squares method for parameters , , , , , Perform fitting, and take the fitted curve at... and The x-axis coordinate value corresponding to the minimum point between the two values ​​is taken as the target intensity value.

[0062] By adopting the above technical solution, the intensity distribution of road surface point cloud has two independent concentrated intervals, corresponding to low-reflection road surface measuring points and high-reflection lane line measuring points, respectively. The distribution of both types of measuring points satisfies the Gaussian normal distribution law. A single Gaussian function cannot simultaneously fit the two sets of peak distributions. Therefore, a double Gaussian superposition model is used to match the intensity distribution characteristics.

[0063] The model contains six unknown parameters. , , Corresponding to the distribution characteristics of road surface measuring points, parameters , , Based on the distribution characteristics of the lane line measuring points, all parameters are iteratively adjusted using the least squares method to reduce the numerical deviation between the fitted curve and the measured histogram, thus obtaining a fitted curve that closely matches the actual intensity distribution.

[0064] There is a unique minimum point in the interval between the two sets of Gaussian peak values, which corresponds to the boundary where the strength of the road surface and the lane line completely transitions. The horizontal axis strength value of the minimum point is read as the unified target strength value for identifying the edge of the lane line.

[0065] In Example 4, step 2, the construction of the directional triangulation includes two stages: generating the initial edge and expanding the triangle edge by edge.

[0066] The method for generating the starting edge is as follows: traverse all strong echo points, calculate the point with the closest Euclidean distance to each strong echo point, and if the closest point is a weak echo point, then take the line connecting the strong echo point and the weak echo point as a starting edge.

[0067] The method for constructing triangles by expanding edge by edge is as follows: add the starting edge to an edge set, and check whether a triangulation has been formed on both sides of each edge in the edge set. If a triangulation has not been formed, find the third point that satisfies the empty circle property on each edge to form a triangulation. At the same time, add the newly added edge that forms the triangulation to the edge set. Repeat the above process until the triangulation is completed on both sides of all edges.

[0068] By adopting the above technical solution, the construction of the Deloni triangulation network of the global point cloud will generate a large number of redundant triangles. Most of the grids do not involve the lane line intensity transition area, resulting in computing power loss. The directional triangulation network only retains the grid at the intersection of strong and weak measurement points. The whole process is divided into two stages: the generation of the initial edge and the edge-by-edge expansion of the network.

[0069] In the initial edge generation stage, all high-reflection points are traversed, the planar Euclidean distance between each high-reflection point and the other measurement points is calculated, and the nearest measurement point is selected. If the neighboring measurement point is a low-reflection point, the line connecting the two points is the initial edge of the strong-weak boundary. The initial edge serves as the basic framework for the growth of the triangular mesh, ensuring that all meshes are built outward from the edge area of ​​the lane line.

[0070] In the edge-by-edge expansion phase, all initial edges are stored in a unified edge set. Single edges are retrieved from the set in a loop, and each edge is checked to see if a triangle has been constructed on both sides. If not, a third measuring point that satisfies the empty circle property is searched to form a triangle. The empty circle property ensures that the generated mesh conforms to the Deloni triangulation topology specification. Newly generated side edges are synchronously added to the edge set and iterated continuously until the mesh construction is completed on both sides of all boundaries.

[0071] In Example 5, in step 2, the method for determining whether to skip the mesh construction operation is as follows: if both ends of the current edge to be expanded are strong echo points and the third point found is also a strong echo point, then it is determined that the triangle does not pass through the target intensity value, and the corresponding edge is marked as having completed the triangulation construction without actually generating a triangle; if both ends of the current edge to be expanded are weak echo points and the third point found is also a weak echo point, it is also determined to skip the mesh construction.

[0072] By adopting the above technical solution, three types of measurement point combinations will appear in the triangulation expansion process. Only the triangle formed by the combination of strong and weak measurement points has an intensity transition boundary inside. The triangle formed by pure high reflection points and pure low reflection points does not have a target intensity boundary inside and does not need to participate in subsequent interpolation calculations. Skipping the generation of this type of grid in advance can further reduce the data scale.

[0073] In Example 6, step 3, the method for linear interpolation on each target edge is as follows: obtain the spatial coordinates of the two endpoints of the target edge and their respective echo intensity values, and calculate the planar coordinates of the interpolation point on the target edge whose intensity value is equal to the target intensity value based on the ratio of the difference between the intensity values ​​of the two endpoints and the target intensity value.

[0074] By adopting the above technical solution, the two ends of the target edge connected by the strong and weak measurement points have two kinds of echo intensities, high and low respectively. The intensity value shows a linear and uniform change along the line segment, and this linear change law serves as the theoretical support for interpolation calculation.

[0075] Extract the planar coordinates and corresponding echo intensity values ​​of the two endpoints of the target edge, calculate the ratio of the difference between the intensity of the two endpoints and the target intensity value, use the ratio coefficient to split the planar coordinates of the line segment, and accurately locate the planar point on the line segment whose intensity value is exactly equal to the target intensity. This point is the discrete sampling point of the lane line edge. After interpolation of all target edges, a complete set of discrete edge points is obtained.

[0076] In Example 7, step 3, the specific method for sorting and numbering all scattered points is as follows: all extracted target edges are stored in the edge set. If two target edges belong to the same Deloni triangle, the two target edges are connected sequentially according to the triangle vertex order and assigned continuously increasing numbers, so that the scattered points on the target edges belonging to the same connected region form a set of sequentially connected points.

[0077] By adopting the above technical solution, the discrete points obtained by direct interpolation have no sequential relationship. Unordered point fitting will cause problems such as lane line breakage and disordered direction. The points are grouped in an orderly manner by relying on the topological relationship of the Deloni triangulation network.

[0078] Store all strong and weak connection target edges into the edge set. Traverse the triangular network topology. If two target edges belong to the same triangle, it proves that the two edges are continuously connected in space. Assign continuously increasing numbers according to the arrangement order of triangle vertices. Connect all target edges in the same connected lane line area in sequence. The interpolated scattered points on the corresponding edges synchronously follow the numbering order. Finally, generate multiple sets of independent and ordered continuous scattered points. Each set corresponds to a continuous lane line edge.

[0079] Example 8, in step 4, the method for polyline fitting is as follows:

[0080] Points are sequentially taken from the sorted scatter set. The first point is taken as the starting point of the current segment and its coordinates are fixed. Subsequent points are added to the current fitted point set one by one. The least squares method is used to calculate the line parameters of the current fitted point set. When the segmentation stopping condition is met, the intersection of the current fitted line and the line connecting the first and last points of the selected points or the projection point of the current last point on the fitted line is taken as the next node of the polyline. Then, the corresponding node is taken as the new starting point of the segmentation and the fitting continues until all points in the point set have been processed.

[0081] By adopting the above technical solution, the road lane lines include various line shapes such as straight lines, transition curves, and circular curves. A single straight line cannot completely match the entire route. Therefore, a segmented least squares fitting of multiple line segments is used, and the complete lane line outline is restored by splicing together the straight lines segment by segment.

[0082] Select the first point in the ordered scatter set as the fixed starting point of the current segment, and add scatter points sequentially to construct a temporary fitting point set. Use the least squares method to solve the optimal straight line equation of the point set. Continue to add scatter points until the segmentation stop judgment condition is triggered. Calculate the intersection point of the current fitting line and the line connecting the points, or the vertical projection point of the last point on the fitting line as the segment node. Use this node as the starting point of the next segment and repeat the operation until all points in the current scatter set have been processed. A single set of ordered scatter points generates a continuous multi-segment line.

[0083] Example 9: The segmented stopping condition is jointly determined by preset node distance threshold, filter strength threshold, filter distance threshold, and interpolation node number threshold.

[0084] The node distance threshold is used to limit the maximum length of a single segment along the lane line direction. If the cumulative distance or the farthest distance from the points already selected in the current fitted point set to the start point of the segment exceeds the threshold, point selection is stopped and nodes are added.

[0085] The filter strength threshold is used to specify the number n outliers to be removed each time the line is fitted. Before each calculation of the line parameters, the residual distance from each point in the current set of fitted points to the fitted line is calculated. The points are sorted from largest to smallest residual distance. After removing the n points with the largest distance values, the least squares fitting is performed again with the remaining points.

[0086] The filtering distance threshold is used to determine whether the lane line profile has changed direction. It calculates the vertical distance from each point in the current fitted point set to the current fitted line. If the maximum vertical distance exceeds the filtering distance threshold, it stops taking points and adds nodes.

[0087] The interpolation node number threshold is used to filter out isolated noise points. If the total number of scattered points in the current sorted scattered point set is less than the interpolation node number threshold, the set is determined to be an invalid set corresponding to isolated strong reflection noise points and will not participate in the polyline fitting.

[0088] By adopting the above technical solution, a single judgment condition cannot simultaneously solve the four types of problems: lane length limitation, reflection noise, linear steering, and isolated invalid points. The four types of thresholds work together to complete the segmentation and noise reduction processing.

[0089] The node distance threshold controls the maximum extension length of a single fitted straight line, preventing long-distance straight line fitting from ignoring the slight curvature of the road surface. If the distance between the farthest point in the current point set and the starting point of the segment exceeds the threshold, the fitting of this segment is immediately terminated and a new node is added.

[0090] The filtering intensity threshold is used to remove random noise generated by laser scanning. Before each fitting, the vertical residual distance from all points to the fitted line is calculated, and the points are sorted from largest to smallest distance. After removing a preset number of points with the largest residual, the line is refitted to eliminate the linear offset caused by abnormal reflections at single points.

[0091] The filter distance threshold identifies lane line directional changes. It calculates the maximum vertical distance from all points in the current point set to the fitted straight line. If the distance exceeds the threshold, it means that the point distribution deviates from the straight line, the lane line enters a curve segment, the fitting of this segment is stopped, and segment nodes are added.

[0092] The threshold for the number of interpolated nodes filters out invalid areas with strong reflections, such as isolated metal signs and wall reflections. If the total number of ordered scattered points in a single group is less than the threshold, the points in that group are determined not to belong to the continuous lane line and are directly discarded and not included in the fitting calculation.

[0093] In Example 10, when constructing the directional triangulation in step 2, each strong echo point region independently constructs its own triangulation set, and each triangulation set corresponds to an independent connected region. In step 3, each connected region generates an independent sorted scatter point set. In step 4, each independent sorted scatter point set is fitted with an independent polyline, and multiple polylines together form a complete lane line edge profile.

[0094] By adopting the above technical solution, there are multiple independent parallel boundaries of lane lines in the road. The edges of different lane lines belong to high-reflection areas that are not connected to each other. A unified network across the entire area will confuse the boundary points of different lanes. A regional independent processing scheme is adopted to distinguish and extract multiple lane lines.

[0095] In the directional triangulation construction stage, each spatially independent and unconnected high-reflectivity point region generates its own exclusive triangulation set, with no topological relationship between the sets; in the interpolation stage, each independent triangulation set generates a set of exclusive ordered scattered points; in the fitting stage, each set of scattered points generates an independent polyline, and after all independent polylines are spliced ​​together in space, the inner and outer edge contours of all lane lines on the road surface are completely restored.

[0096] The following specific embodiments illustrate the implementation principle of the present invention:

[0097] The automatic lane line extraction method based on vehicle-mounted laser point clouds is based on the following idea: A key-value pair (x, y, i) is established using the point cloud planar coordinates (x, y) as the key and the laser echo intensity (i) as the value (which can also be understood as a three-dimensional elevation point with x, y as planar coordinates and i as elevation). A Delaunay triangulation is constructed based on the key-value pair (x, y, i). The scattered coordinates (x, y) of the lane line contour edge are obtained by interpolation based on the triangulation and the point cloud intensity values ​​of the lane line contour edge. The lane line edge contour is then fitted based on the scattered coordinates to obtain the lane line edge contour.

[0098] The specific process is as follows:

[0099] Calculate the point cloud intensity value of the lane outline edge:

[0100] In the point cloud echo intensity map within the range of the road surface and lane lines, the point cloud intensity shows a clear contrast, exhibiting a characteristic of being either strong or weak. The measurement points required for lane line contour fitting are the measurement points at the junction of strong and weak points.

[0101] Figure 3 Enlarged view of a lane line section: To obtain the echo intensity values ​​at the points where strong and weak light oscillations intersect, it is necessary to statistically analyze the laser echo intensity values ​​of the point cloud in the target area. The laser echo intensity values ​​range from 0 to 65535. The following analysis uses 100 units as intervals to statistically analyze the laser echo intensity of a sample dataset. The results are as follows:

[0102] Figure 4 The measured distribution of laser echo intensity shows two peaks in the point cloud intensity value, corresponding to the intensity and quantity distribution of echo points on the ordinary road surface and the lane line. The trough between these two peaks represents the intensity and quantity distribution of echo points along the lane line contour. The horizontal axis represents the laser point cloud echo intensity value of the desired contour line points (hereinafter referred to as the target intensity). To accurately calculate the target intensity value, a double Gaussian function curve is used to fit the laser echo intensity distribution curve. The function model is as follows:

[0103] ;

[0104] The function parameters were fitted using the least squares method, and the fitting results are as follows:

[0105] Example data fitting results:

[0106] =1019.21, =92.11, =9.37

[0107] =82.26, =168.18, =9.34

[0108] The horizontal axis corresponding to the trough value is That is, the target intensity value is 13300. Figure 4 The result is a fitting plot of the measured echo intensity distribution.

[0109] Generate a directional triangulated mesh containing point cloud edge intensity values:

[0110] A key-value pair (x, y, i) is established using the point cloud planar coordinates (x, y) as the key and the laser echo intensity (i) as the value. A Delaunay triangulation is constructed based on the key-value pair (x, y, i) to interpolate the target intensity value. Figure 5 For example, points with echo intensity greater than the target intensity are defined as strong echo points, represented by solid circles; points with echo intensity less than the target intensity are defined as weak echo points, represented by hollow circles; the lane line outline is located at the junction of strong and weak echo points. Since the number of vehicle-mounted point cloud measurement points is extremely large, generating a triangulation network across the entire area would result in a huge computational load (the solid and dashed lines in the diagram). However, the triangulation network used for interpolating the target intensity points of the lane line is only the portion where strong and weak echo points intersect (the solid line portion in the diagram). Therefore, it is only necessary to generate a triangulation network that includes the target intensity points. The triangulation network containing only the target intensity value is referred to as a directional triangulation network.

[0111] Figure 5 Point cloud triangulation diagram

[0112] The specific steps for generating a directional triangulated network are as follows:

[0113] Generate the starting edge:

[0114] The construction of the directional triangulation network is based on the triangulation growth method. The calculation process of the triangulation growth method requires initial edges; that is, based on the initial edges, a third point satisfying the empty circle property is found step by step until all points complete the network construction. Since the regions formed by strong echo points may not be continuous, it is necessary to find the starting edge for each strong echo point region. The specific process is as follows: traverse all strong echo points and find the closest point to them; the line connecting these two points will be one side of the Deloni triangulation network. If the closest point is a weak echo point, then the two points form the initial edge. The result of the initial edge construction is shown in [the diagram]. Figure 6 The values ​​in the figure represent echo intensity values.

[0115] Figure 6 A magnified view of the initial edge construction result (target strength value 13100).

[0116] Find the third point edge by edge:

[0117] Add the initial edge to an edge set. For each edge in the set, check if a triangulation has been formed on both sides. If not, find a third point on each edge that satisfies the empty circle property to form a triangulation, and add the newly added edge to the edge set. Repeat this process until triangulations are completed on both sides of all edges. Note that if the set of edges to be constructed is a strong-strong edge (connecting strong reflection points), and the third point is a strong point (strong reflection point), then this triangulation will not contain the target strength point, and the corresponding triangulation will not participate in the construction. The direction of the corresponding edge is considered as having completed triangulation construction. The same applies to triangles formed by weak-weak edges and weak points.

[0118] After constructing all the triangulations within the edge set using the steps described above, multiple edge sets will be formed, as shown in the image. Figure 7 As shown.

[0119] Figure 7 Result of directional triangulation generation (target strength value 13100)

[0120] Scattered points obtained from target intensity values ​​through triangulation interpolation:

[0121] In the edge set of a triangular network, if the connecting edge is a strong-weak edge ( Figure 8 (As shown by the red edge), then there must be a target intensity interpolation point on that edge. Store the above strong-weak edges in sorted order, following the sorting principle: if two edges belong to the same Deloni triangle, then the two edges are connected sequentially. The sorting numbers are as follows: Figure 8 The red numbers are shown. The target strength value scatter points are obtained by interpolating the above strong-weak edges in order, and the scatter point sorting number is consistent with the edge number, forming a scatter point set sorted by target strength value.

[0122] After completing the interpolation of the target strength values ​​of strong and weak edges within the edge set by following the steps above, multiple point sets will be obtained.

[0123] Figure 8 Target intensity scatter plot generation results (target intensity value 13100)

[0124] Fit the lane line contour edge line based on the scatter plot of the target intensity value:

[0125] A polyline is used to fit the lane lines. Each set of points corresponds to a fitted polyline. The first point of the polyline is the first point of the corresponding set of points, and its coordinates are (x0, y0). The equation of the line segment passing through this point is:

[0126] y - y0 = k(x - x0);

[0127] The equation of the straight line contains only one position parameter k. Subsequent points are extracted one by one from the point set, and the parameter k is calculated using the least squares method to obtain the equation of the corresponding segment of the polyline. The projection point from the last point to the fitted line is taken as the second node. This process is repeated until all points in the point set are extracted, thus completing the calculation of each node of the polyline. The number of points extracted requires several thresholds for determination; the relevant thresholds are as follows:

[0128] Fitted node distance threshold: This threshold is the length threshold of the polyline node. When the distance from the measured point in the point set to the previous node exceeds this threshold, point selection is stopped.

[0129] Filtering intensity threshold: This threshold is the number of points n to be filtered out, used to filter out laser noise; calculate the distance from each point participating in the line fitting to the fitted line in turn, sort them from largest to smallest, delete the n points with the largest distance values, and then recalculate the parameter k;

[0130] Filtering distance threshold: This threshold is the maximum distance from the point participating in the line fitting to the fitted line. When the maximum value exceeds this threshold, it indicates that the lane line profile begins to turn, so point selection is stopped and multi-segment line nodes are added.

[0131] Interpolation Node Count Threshold: This threshold is the number of points involved in the line fitting process. It is used to filter out isolated, highly reflective noise points, such as... Figure 9 The image shows a typical strong reflection point noise. In this case, the scale of the directional triangulation network is obviously too small, and it can be filtered out by increasing the number of point sets.

[0132] Figure 9 Schematic diagram of noise at a strong reflection point (target intensity value 13100)

[0133] Complete the polyline fitting for each set of points using the steps described above. See section 10 for the fitting results. Figure 11 .

[0134] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic lane line extraction based on vehicle-mounted radar point clouds, characterized in that, Includes the following steps: Step 1: Obtain the echo intensity value distribution of the road surface point cloud, use a double Gaussian function to fit the intensity distribution curve, and set the intensity value corresponding to the trough value of the double Gaussian function fitting curve as the target intensity value of the lane line outline edge. Step 2: Define points in the point cloud whose echo intensity value is greater than the target intensity value as strong echo points and points whose echo intensity value is less than the target intensity value as weak echo points. Based on the planar coordinates and echo intensity values ​​of the strong and weak echo points, construct an oriented triangulation network that includes only the target intensity value using the triangulation growth method. During the construction process, skip all triangulation operations that do not involve the target intensity value within the triangles. Step 3: Extract all edges connecting strong echo points and weak echo points in the directional triangulation network as target edges. Linearly interpolate the positions on each target edge where the intensity value is equal to the target intensity value as scattered points. Sort and number all scattered points to form a sorted scattered point set. Step 4: For each sorted scatter point set, starting from the first point in the set, use the least squares method to fit a multi-segment line segment by segment. During the fitting process, determine the position of the segment nodes and filter noise based on the preset node distance threshold, filter intensity threshold, filter distance threshold and interpolation node number threshold, until the fitting of all points in the sorted scatter point set is completed, and output each segment of the multi-segment line as the lane line edge contour.

2. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 1, characterized in that, In step 1, when statistically analyzing the point cloud echo intensity values, frequency statistics are performed using a preset fixed intensity range as the unit, and the fitting object of the double Gaussian function is the distribution histogram after frequency statistics.

3. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 2, characterized in that, In step 1, the double Gaussian function model is as follows: ； in It is the amplitude of the Gaussian function model at ordinary measuring points on the road surface. It is the peak position of the Gaussian function model of a common measuring point on the road surface. It is the standard deviation of the Gaussian function model of ordinary measuring points on the road surface. It is the amplitude of the Gaussian function model of the road lane line measuring point. It is the peak position of the Gaussian function model of the road lane line measuring point. It is the standard deviation of the Gaussian function model of the road lane line measuring points. It is the independent variable, representing the echo intensity value corresponding to a single point cloud measurement point. It is the dependent variable, representing the point cloud statistical frequency under the corresponding echo intensity value; Based on the least squares method for parameters , , , , , Perform fitting, and take the fitted curve at... and The x-axis coordinate value corresponding to the minimum point between the two values ​​is taken as the target intensity value.

4. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 1, characterized in that, In step 2, the construction of the oriented triangulation network includes two stages: generating the initial edge and expanding the triangle edge by edge. The method for generating the starting edge is as follows: traverse all strong echo points, calculate the point with the closest Euclidean distance to each strong echo point, and if the closest point is a weak echo point, then take the line connecting the strong echo point and the weak echo point as a starting edge. The method for constructing triangles by expanding edge by edge is as follows: add the starting edge to an edge set, and check whether a triangulation has been formed on both sides of each edge in the edge set. If a triangulation has not been formed, find the third point that satisfies the empty circle property on each edge to form a triangulation. At the same time, add the newly added edge that forms the triangulation to the edge set. Repeat the above process until the triangulation is completed on both sides of all edges.

5. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 4, characterized in that, In step 2, the method for determining whether to skip the mesh construction operation is as follows: if both ends of the current edge to be expanded are strong echo points and the third point found is also a strong echo point, then it is determined that the triangle does not pass through the target intensity value, and the corresponding edge is marked as having completed the triangulation construction without actually generating a triangle; if both ends of the current edge to be expanded are weak echo points and the third point found is also a weak echo point, it is also determined to skip the mesh construction.

6. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 1, characterized in that, In step 3, the method for linear interpolation on each target edge is as follows: obtain the spatial coordinates of the two endpoints of the target edge and their respective echo intensity values, and calculate the planar coordinates of the interpolation point on the target edge whose intensity value is equal to the target intensity value based on the ratio of the difference between the intensity values ​​of the two endpoints and the target intensity value.

7. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 6, characterized in that, In step 3, the specific method for sorting and numbering all scattered points is as follows: store all extracted target edges into the edge set. If two target edges belong to the same Deloni triangle, connect the two target edges in sequence according to the vertex order of the triangle and assign them continuously increasing numbers, so that the scattered points on the target edges belonging to the same connected region form a set of points connected in sequence.

8. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 1, characterized in that, In step 4, the method for polyline fitting is as follows: Points are sequentially taken from the sorted scatter set. The first point is taken as the starting point of the current segment and its coordinates are fixed. Subsequent points are added to the current fitted point set one by one. The least squares method is used to calculate the line parameters of the current fitted point set. When the segmentation stopping condition is met, the intersection of the current fitted line and the line connecting the first and last points of the selected points or the projection point of the current last point on the fitted line is taken as the next node of the polyline. Then, the corresponding node is taken as the new starting point of the segmentation and the fitting continues until all points in the point set have been processed.

9. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 8, characterized in that, The segmentation stopping condition is jointly determined by preset node distance threshold, filter strength threshold, filter distance threshold, and interpolation node number threshold: The node distance threshold is used to limit the maximum length of a single segment along the lane line direction. If the cumulative distance or the farthest distance from the points already selected in the current fitted point set to the start point of the segment exceeds the threshold, point selection is stopped and nodes are added. The filter strength threshold is used to specify the number n outliers to be removed each time the line is fitted. Before each calculation of the line parameters, the residual distance from each point in the current set of fitted points to the fitted line is calculated. The points are sorted from largest to smallest residual distance. After removing the n points with the largest distance values, the least squares fitting is performed again with the remaining points. The filtering distance threshold is used to determine whether the lane line profile has changed direction. It calculates the vertical distance from each point in the current fitted point set to the current fitted line. If the maximum vertical distance exceeds the filtering distance threshold, it stops taking points and adds nodes. The interpolation node number threshold is used to filter out isolated noise points. If the total number of scattered points in the current sorted scattered point set is less than the interpolation node number threshold, the set is determined to be an invalid set corresponding to isolated strong reflection noise points and will not participate in the polyline fitting.

10. The method for automatic lane line extraction based on vehicle-mounted radar point clouds according to claim 1, characterized in that, In step 2, when constructing the directional triangulation, each strong echo point region independently constructs its own triangulation set, and each triangulation set corresponds to an independent connected region. In step 3, each connected region generates an independent sorted scatter point set. In step 4, each independent sorted scatter point set is fitted with an independent polyline, and multiple polylines together form a complete lane line edge profile.

Images