A ground segmentation method for laser point clouds

Abstract

A ground segmentation method for laser point clouds includes the following steps: S101: using a multi-line laser radar to obtain a three-dimensional point cloud of a target detection area, dividing the target detection area into concentric circles of different radii in a non-uniformly spaced manner with the laser radar as the center, associating the calculation of the concentric circle radius with the distribution of the point cloud data scanned by the laser radar, performing regional segmentation based on the correspondence between the points and the regions of the point cloud, and obtaining the concentric circle radius of the corresponding region; dividing each concentric circle into multiple sector-shaped regions according to a predetermined angle; S102: when a new frame of point cloud is input, obtaining the radial index of the region where each point is located based on the correspondence between each point of the point cloud and the region; and obtaining the lateral index of the region where each point is located based on the angle corresponding to the sector of each region, and obtaining the sector region index corresponding to each point based on the radial index and the lateral index. The present invention can improve the calculation accuracy of the point cloud height threshold, save computing power, and reduce missegmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud segmentation, in particular to a ground segmentation method of laser point clouds. Background Art

[0002] LiDAR sensors play a crucial role in unmanned intelligent systems, such as robotics. Processing the laser point clouds they collect enables high-precision mapping, localization, and obstacle avoidance. Most existing robots use 16-line LiDARs, which produce sparser point clouds compared to 64- and 128-line LiDARs. Most ground point cloud segmentation algorithms proposed in the literature either offer high accuracy but high computational overhead, only applicable to denser point clouds with more than 64 lines (e.g., Patchwork++: Fast and Robust Ground Segmentation Solving Partial Under-Segmentation Using 3D Point Cloud | IEEE Conference Publication | IEEE Xplore); or offer high speed but poor performance and are unsuitable for scenes with slopes (e.g., Fast segmentation of 3D pointclouds: A paradigm on LiDAR data for autonomous vehicle applications | IEEE Conference Publication | IEEE Xplore).

[0003] To address the above problems, the paper Fast Ground Segmentation for 3D LiDAR Point CloudBased on Jump-Convolution-Process (JCP) [Remote.Sens.13(16):3239(2021)] proposed a method, namely the point cloud segmentation method of JCP. Figure 1 As shown: With the laser radar as the center, concentric circles of different radii are divided at equal intervals ( Figure 1 a), and divide each concentric circle into multiple "sector" areas according to a certain angle ( Figure 1 b). Calculate the correspondence between each point in the point cloud scanned by the lidar for the fan-shaped area ( Figure 1c), each point has six values: t, c, x, y, z, and label. Among them, t is the radial index of the concentric circles from the inside to the outside (Note: t in the following content has this meaning), and c is the horizontal index of the cth sector area in the same circle according to the scanning direction. After obtaining the t and c of the point, we can know which "sector" area the point corresponds to. [t, c] is the index coordinate of the area; x, y, z are the original spatial data of the point cloud, and label is the label used to classify whether it is a ground point. Then calculate the minimum value of the point cloud z in each area as the height threshold thres of this area. z , and adjust the height threshold thres of area t according to the slope threshold parameter s of area [t-1, c] and area [t, c] in the radial direction z Then, we calculate whether each point in the point cloud is a ground point based on the height threshold, and then convert it into an elevation map. We use the dilation and convolution algorithms to recalculate the labels of the point clouds with lower confidence, and finally obtain the point cloud after ground segmentation.

[0004] However, when using the calculation method described in the aforementioned literature with a 16-line lidar, the point cloud is too sparse, so at greater distances, a point can differ significantly from its surrounding points. This can lead to mis-segmentation when recalculating labels for low-confidence points during dilation and convolution operations, which in turn reduces segmentation effectiveness. Furthermore, when generating concentric circles at equal intervals, there is a high probability of no point cloud in distant areas (e.g., over 60 meters). In this case, only the slope threshold is used when calculating the height threshold, leading to mis-segmentation of distant points.

[0005] Therefore, based on the above-mentioned JCP method, the present invention proposes a new point cloud ground segmentation method based on fan-shaped areas to improve it, and obtains a fast and more accurate ground segmentation method that can be applied to 16-line sparse point clouds, and this method is also applicable to lidar point clouds with more lines. Summary of the Invention

[0006] The purpose of the present invention is to overcome the above-mentioned deficiencies in the prior art and provide a ground segmentation method for laser point clouds.

[0007] The technical solution of the present invention is: a ground segmentation method of laser point cloud, comprising the following steps:

[0008] S101: Using a multi-line laser radar to obtain a three-dimensional point cloud of the target detection area, dividing the target detection area into concentric circles of different radii at unequal intervals with the laser radar as the center, correlating the calculation of the concentric circle radius with the distribution of the point cloud data scanned by the laser radar, performing region segmentation based on the correspondence between the points in the point cloud and the region, obtaining the concentric circle radius of the corresponding region, generating concentric circles based on the radius with the center as the reference, and gradually increasing the radius from the inside to the outside in a nonlinear manner; and dividing each concentric circle into multiple sector-shaped regions at a predetermined angle;

[0009] S102: When a new frame of point cloud is input, the radial index of the area where each point is located is obtained based on the correspondence between each point and the area of ​​the point cloud; and the lateral index of the area where each point is located is obtained based on the angle corresponding to the sector of each area, and the sector area index corresponding to each point is obtained based on the radial index and the lateral index.

[0010] Furthermore, the method further comprises the following steps:

[0011] S103: After obtaining the sector area division format, calculate the minimum value of the height z in the point cloud within each sector area as the height threshold thres of this area z , and adjust the height threshold thres of the [t,c] region according to the slope threshold parameter s of the [t-1,c] region and the [t,c] region in the radial direction z ; Then determine the height z of each point in the point cloud and the height threshold thres of the corresponding area z The size relationship, if z is greater than thres z , this point is a non-ground point; if z is less than thres z , this point is the ground point.

[0012] Furthermore, in step S101, the region segmentation method includes: forming a right triangle between the laser radar, the ground, and a point at a predetermined position far away from the ground, with the installation position of the laser radar as the vertex and the point on the ground perpendicular to the installation position of the laser radar as the origin; dividing the vertex of the right triangle into multiple equal parts, each with an angle of θ; each equal part has a corresponding base to the origin, and the base is the radius of the corresponding concentric circle of the region segmentation

[0013] Furthermore, the radius of the concentric circles for the region segmentation is obtained by the following formula:

[0014]

[0015] Where h is the height parameter of the lidar, θ is the angle parameter, and t is the radial index coefficient of the tth region from the inside to the outside in the radial direction, t = 1, 2, 3..., r t is the radius from the tth region to the center of the concentric circle.

[0016] Furthermore, in step S102, the radial index and the lateral index are obtained by the following formula:

[0017]

[0018]

[0019]

[0020] Where h and θ are the parameters set in advance; ang is the angle parameter for the horizontal division of the segmented area, indicating the angle size corresponding to each sector area; x and y are the vertical and horizontal distances of the points; r xy is the radius of the point; t is calculated by the reverse calculation of the concentric circle radius of the aforementioned calculation area segmentation, thereby obtaining the radial index of the area where each point belongs; c is the horizontal index of the area where each point is located; the fan area index [t,c] corresponding to the point cloud is calculated by the above formula.

[0021] Furthermore, the concentric circles of different radii are divided in a non-equally spaced manner, and the radius of the area closer to the center of the circle increases slowly to obtain a dense point cloud; the radius of the area farther from the center of the circle increases faster, so that the area gradually increases in the radial direction, and a sparse point cloud is obtained in the distant area.

[0022] Furthermore, the labels of the non-ground points are set to 1, and the labels of the ground points are set to 0; then they are converted into an elevation map, and the labels of the point clouds with lower confidence are recalculated using the dilation and convolution algorithms, and finally the point cloud after ground segmentation is obtained.

[0023] Furthermore, for the sparse point cloud, dilation and convolution operations are removed.

[0024] Furthermore, when the multi-line laser radar is a 16-line laser radar, the expansion and convolution operations are removed; when the multi-line laser radar is a 32-line laser radar or above, the expansion and convolution operations are used to update the segmentation results adjacent to the ground and non-ground point clouds.

[0025] The beneficial effects of the present invention are as follows: (1) when generating concentric circles, the radii of the concentric circles are not generated at equal intervals, but the radius of each circle is generated in a nonlinear increasing manner, which ensures that points in each area are divided, making the calculation of the height threshold more accurate and improving the ground segmentation effect; (2) this radius calculation method does not require tedious table lookup operations and has a small amount of calculation; (3) for sparse point clouds, the expansion and convolution operations are removed, saving computing power and reducing missegmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is the principle diagram of point cloud ground segmentation using the JCP method in the existing technology;

[0027] Figure 2 is a schematic diagram of a formula for calculating radial radius according to an embodiment of the present invention;

[0028] Figure 3 Schematic diagram of the evolution trend of the concentric circle radius according to an embodiment of the present invention;

[0029] Figure 4This is a schematic diagram of the principles of concentric circle generation and sector-shaped area division according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] A ground segmentation method for laser point cloud comprises the following steps:

[0032] S101: Use a multi-line laser radar to obtain a three-dimensional point cloud of the target detection area, divide the target detection area into concentric circles of different radii with the laser radar as the center in a non-uniformly spaced manner, associate the calculation of the concentric circle radius with the distribution of the point cloud data scanned by the laser radar, perform regional segmentation according to the correspondence between the points and the areas of the point cloud, obtain the concentric circle radius of the corresponding area, generate each concentric circle with the center of the circle as the reference according to the radius, and gradually increase the radius from the inside to the outside in a nonlinear manner; divide each concentric circle into multiple sector-shaped areas according to a predetermined angle.

[0033] Specifically, the laser point cloud is first obtained from the robot control system data set. With the laser radar as the center, concentric circles of different radii are divided in a non-uniform manner.

[0034] The closer the target is to the LiDAR, the denser its point cloud becomes, and vice versa. Therefore, based on the characteristic that the point cloud changes from dense to sparse as the horizontal distance increases, the outer diameter of each concentric circle from the inside to the outside should be divided in a nonlinear way that gradually increases:

[0035] a. The radius closer to the center of the circle increases slowly, making the dense point cloud segmentation more refined. Figure 3 As shown in FIG, the area where the concentric circles are closer is smaller than the area where the concentric circles are farther away, and the ground segmentation threshold obtained in the corresponding area is more accurate, so the segmentation is finer.

[0036] b. The radius increases faster for areas farther from the center of the circle, so that the area gradually increases in the radial direction, resulting in sparse point clouds in the distant area and accurate calculation of the height threshold.

[0037] c. Based on points a and b above, we can conclude that the radius calculation method should be associated with the data distribution of the LiDAR scan. This allows the point-to-region correspondence to be directly calculated based on the spatial distribution of each point in the point cloud. If this is not associated with the point cloud data, the radius of each concentric circle must be calculated in advance, a data table must be generated, and then the data table must be queried based on the horizontal distance of the current point cloud to determine the region correspondence, resulting in a computational complexity increase of at least nlog2n times.

[0038] The radial radius formula for calculating region segmentation is proposed as follows:

[0039]

[0040] Where h is the height parameter of the lidar, θ is the angle parameter, Figure 2 A schematic diagram of the area division of this formula is given. The laser radar, the ground and a distant point form a right triangle. The angle with the laser radar as the vertex is divided into multiple equal parts by θ. Each equal part has a corresponding base r to the origin O. t (i.e. the radial radius of the region segmentation), Figure 2 The cases of t = 1, 2, 3, 4, etc. are given in the figure. Since the vertical point cloud angle distribution and installation height of the actual scene LiDAR are different, the point cloud distribution is also different. The h and θ parameters can be adjusted according to the actual situation to obtain a suitable region segmentation. t is the radial index coefficient of the tth region from the inside to the outside in the radial direction, r t is the radius from the tth region to the center of the concentric circle. For example, region 1 is the first concentric circle region close to the center of the circle, region 2 is the second concentric circle region adjacent to the first concentric circle region, and so on. The larger t is, the farther it is from the center of the circle.

[0041] For example, when the parameters are set to h = 3.5 and θ = 11°, as the t value increases, the radius r evolves as follows: Figure 3 As shown: the horizontal coordinate of each point is the radius, and the vertical coordinate is the difference between the radius of the [t, c]th area and the radius of the [t-1, c]th area, that is, the straight-line distance between two adjacent concentric circles. Figure 3 It can be seen that the distribution of all points just meets the requirement of gradually increasing distribution in a nonlinear manner.

[0042] S102: When a new frame of point cloud is input, the radial index of the area where each point is located is obtained based on the correspondence between each point and the area of ​​the point cloud; and the lateral index of the area where each point is located is obtained based on the angle corresponding to the sector of each area, and the sector area index corresponding to each point is obtained based on the radial index and the lateral index.

[0043] like Figure 4 As shown: Concentric circles generated according to the radius, Figure 4 The values ​​marked in are the angle and radius lengths.

[0044] After setting h and θ, we get the threshold for region division. When a new frame of point cloud is input, we need to calculate the correspondence between each point and the region. The formula is as follows:

[0045]

[0046]

[0047]

[0048] In the formula, h and θ are the parameters that need to be set in advance as mentioned above; ang is the angle parameter of the horizontal division of the segmented area, which indicates the angle size corresponding to each sector of the area. x and y are the vertical and horizontal distances of the point, r xy is the radius of the point. t is calculated by reversing the above-mentioned formula (1) for calculating the radial radius of the region segmentation, thereby obtaining the radial index of the region to which each point belongs; c is the horizontal index of the region corresponding to the point cloud. This formula can be used to quickly calculate the "sector" region index [t, c] corresponding to the point cloud in a non-query manner.

[0049] S103: Obtaining the point cloud after ground segmentation.

[0050] After obtaining the sector area division format, calculate the minimum value of the height z in the point cloud within each sector area as the height threshold thres of this area z , and adjust the height threshold thres of the [t,c] region according to the slope threshold parameter s of the [t-1,c] region and the [t,c] region in the radial direction z ; Then determine the height z of each point in the point cloud and the height threshold thres of the corresponding area z The size relationship, if z is greater than thres z , this point is a non-ground point (label label is set to 1); if z is less than thres z , this point is a ground point (the label is set to 0). It is then converted into an elevation map, and the labels of the point cloud with lower confidence are recalculated using the dilation and convolution algorithms, finally obtaining the ground segmented point cloud.

[0051] Since the radial distribution of the 16-line lidar point cloud is relatively sparse, it is necessary to remove the dilation and convolution operations in the original paper. If a 32-line or higher lidar is used, the resulting point cloud is denser, and the dilation and convolution operations in the original paper can be used to update the segmentation results of the adjacent areas of the ground and non-ground point clouds.

Claims

1. A ground segmentation method for laser point cloud, characterized in that: The following steps are involved: S101: Use multi-line laser radar to obtain the three-dimensional point cloud of the target detection area, with the laser radar as the center of the circle, Concentric circles of different radii are divided in a non-uniformly spaced manner. The calculation of the concentric circle radius is associated with the distribution of the point cloud data scanned by the lidar. The region is segmented according to the correspondence between the points and the regions of the point cloud to obtain the concentric circle radius of the corresponding region. The concentric circles are generated based on the radius with the center as the reference, and the radius increases nonlinearly from the inside to the outside. Each concentric circle is divided into multiple sector-shaped regions according to a predetermined angle. The radius of the concentric circle for the region segmentation is obtained by the following formula: ； Where h is the height parameter of the lidar, θ is the angle parameter, and t is the radial index coefficient of the tth region from the inside to the outside in the radial direction, t=1, 2, 3..., r t is the radius from the tth region to the center of the concentric circle; S102: When a new frame of point cloud is input, the radial index of the region where each point is located is obtained based on the correspondence between each point and the region of the point cloud; and the transverse index of the region where each point is located is obtained based on the angle corresponding to the sector of each region. The sector region index corresponding to each point is obtained based on the radial index and the transverse index; wherein the radial index and the transverse index are obtained by the following formula: ； Where, ang is the angle parameter for the horizontal division of the segmented area, indicating the angle size corresponding to each sector area; x and y are the vertical and horizontal distances of the points; r xy is the radius of the point; t is calculated by the reverse calculation of the concentric circle radius of the aforementioned calculation area segmentation, thereby obtaining the radial index of the area where each point belongs; c is the horizontal index of the area where each point is located; the sector area index [t, c] corresponding to the point cloud is calculated by the above formula; S103: After obtaining the sector area division format, calculate the minimum value of the height z in the point cloud within each sector area as the height threshold of this area , and adjust the height threshold of the [t, c] region according to the slope threshold parameter s of the [t-1, c] region and the [t, c] region in the radial direction ; Then determine the height z of each point in the point cloud and the height threshold of the corresponding area If z is greater than , this point is a non-ground point; if z is less than , this point is the ground point.

2. The ground segmentation method of laser point cloud according to claim 1, characterized in that: In step S101, the method for area segmentation includes: forming a right triangle between the laser radar, the ground, and a point at a predetermined position far away from the ground, with the installation position of the laser radar as the vertex and the point on the ground perpendicular to the installation position of the laser radar as the origin; dividing the vertex angle of the right triangle into multiple equal parts, each with an angle of θ; each equal part has a corresponding base to the origin, and the base is the radius of the corresponding concentric circle of the area segmentation.

3. The ground segmentation method of laser point cloud according to claim 1, characterized in that: The concentric circles of different radii are divided in a non-uniformly spaced manner, and the radius of the area closer to the center of the circle is increased slowly to obtain a dense point cloud; the radius of the area farther from the center of the circle is increased faster, so that the area gradually increases in the radial direction, and a sparse point cloud is obtained in the distant area.

4. The ground segmentation method of laser point cloud according to claim 3, characterized in that: The labels of the non-ground points are set to 1, and the labels of the ground points are set to 0; then they are converted into an elevation map, and the labels of the point clouds with lower confidence are recalculated using the dilation and convolution algorithms, and finally the point cloud after ground segmentation is obtained.

5. The ground segmentation method of laser point cloud according to claim 4, characterized in that: For the sparse point cloud, the dilation and convolution operations are removed.

6. The ground segmentation method of laser point cloud according to claim 5, characterized in that: When the multi-line laser radar is a 16-line laser radar, the expansion and convolution operations are removed; when the multi-line laser radar is a 32-line laser radar or above, the expansion and convolution operations are used to update the segmentation results adjacent to the ground and non-ground point clouds.

Images