CN113077473B - Three-dimensional laser point cloud pavement segmentation method, system, computer equipment and medium - Google Patents

Abstract

The present invention discloses a three-dimensional laser point cloud road segmentation method, comprising: constructing a sector grid map; performing connected domain clustering processing on the sector grid map to construct grid clusters; performing characteristic value analysis on grid clusters that meet preset conditions, extracting grid clusters that meet line features and surface features as ground grid clusters; performing smoothness inspection on the ground grid clusters in the radial direction of the sector grid map, and extracting ground grid clusters that meet the smoothness requirements. The present invention also discloses a three-dimensional laser point cloud road segmentation system, a computer device, and a computer-readable storage medium. The present invention realizes efficient and accurate point cloud road segmentation for complex terrain by analyzing the geometric features of the three-dimensional laser point cloud, and has a significant improvement in accuracy and robustness.

Description

Three-dimensional laser point cloud road segmentation method, system, computer equipment and medium

Technical Field

The present invention relates to the field of unmanned driving technology, and in particular to a three-dimensional laser point cloud road surface segmentation method, a three-dimensional laser point cloud road surface segmentation system, a computer device and a computer-readable storage medium.

Background technique

Unmanned vehicles or autonomous driving are one of the most valuable industries in the current artificial intelligence industry. Environmental perception, as the core research content of unmanned vehicles, is the basis for intelligent vehicles to achieve autonomous decision-making and path planning. The accuracy of environmental perception directly determines the level of intelligence of the car. However, current technology still cannot accurately and quickly perceive the most effective external information in complex environments, and there are still many key technologies that need to be broken through.

In the existing technology, laser radar is generally used to detect obstacles and drivable areas. Among them, road segmentation is the premise of laser radar detection. At present, the more typical laser point cloud road segmentation methods are:

First, the point cloud is built into an undirected graph, and the road surface is segmented by setting a loss function to solve the undirected graph model. However, this method is highly dependent on the accuracy of the loss function. If the loss function is not suitable, it will cause low accuracy.

Second, establish one or more fitted road surface models through the RANSAC method. However, this method is difficult to adapt to irregular road conditions such as uneven roads and multiple slopes.

Third, perform gradient screening in a specific direction and then filter the ground point cloud through a smoothing function. However, this method is difficult to comprehensively utilize the combined features presented by neighboring points, which will result in a waste of feature information.

Summary of the invention

The technical problem to be solved by the present invention is to provide a three-dimensional laser point cloud road surface segmentation method, system, computer equipment and medium, which can achieve efficient and accurate point cloud road surface segmentation for complex terrain by analyzing the geometric features of the three-dimensional laser point cloud.

In order to solve the above technical problems, the present invention provides a three-dimensional laser point cloud road segmentation method, comprising: constructing a sector grid map; performing connected domain clustering processing on the sector grid map to construct grid clusters; performing feature value analysis on grid clusters that meet preset conditions, and extracting grid clusters that meet line features and surface features as ground grid clusters; performing smoothness inspection on the ground grid clusters in the radial direction of the sector grid map, and extracting ground grid clusters that meet the smoothness requirements.

As an improvement of the above scheme, the three-dimensional laser point cloud road segmentation method also includes: in the radial direction of the fan-shaped grid map, constructing a smooth constraint according to the ground grid cluster, and taking the grid that meets the smooth constraint in the grid cluster that does not meet the preset conditions as the ground grid.

As an improvement of the above scheme, the step of constructing a fan-shaped grid map includes: projecting the three-dimensional laser point cloud of the lidar into the fan-shaped grid map, the fan-shaped grid map is composed of multiple independent grids; calculating the maximum height difference of all points in each grid respectively, if the maximum height difference is greater than a preset threshold, the grid is an obstacle grid, and the obstacle grid is deleted.

As an improvement of the above scheme, the step of performing connected domain clustering processing on the sector grid map to construct a grid cluster includes: S11, in the sector grid map, taking a grid as the search center to establish a grid cluster; S12, searching for grids that meet the gradient requirements in a preset field, and adding the grids that meet the gradient requirements to the grid cluster; S13, within the grid cluster, taking another grid that is not used as the search center as a new search center, entering step S12, until all grids in the grid cluster are searched; S14, outside the grid cluster, taking another grid that is not used as the search center as a new search center, establishing a new grid cluster, entering step S12, until all grids in the sector grid map are searched.

As an improvement of the above scheme, the step of determining whether a grid cluster meets preset conditions includes: converting each grid in the grid cluster into a point and constructing a minimum enclosing rectangle for the grid cluster; determining whether the number of points in the grid cluster is less than a preset number or whether the diagonal length of the minimum enclosing rectangle corresponding to the grid cluster is less than a preset length, and when the determination is yes, the grid cluster does not meet the preset conditions, and when the determination is no, the grid cluster meets the preset conditions.

As an improvement of the above scheme, the step of performing feature analysis on the grid clusters that meet the preset conditions and extracting the grid clusters that meet the line features and the surface features as the ground grid clusters includes: constructing a covariance matrix based on the points in the grid clusters that meet the preset conditions; calculating the eigenvalues of the covariance matrix; extracting the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue in the features, and making a judgment based on the difference between the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue, wherein when the difference between the minimum eigenvalue and the intermediate eigenvalue is not within the preset range, and the intermediate eigenvalue is When the difference between the minimum eigenvalue and the maximum eigenvalue is within the preset range, the grid cluster that meets the preset conditions is a planar grid cluster; when the difference between the minimum eigenvalue and the intermediate eigenvalue is within the preset range, and the difference between the minimum eigenvalue and the maximum eigenvalue is not within the preset range, the grid cluster that meets the preset conditions is a linear grid cluster; when the difference between the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue is within the preset range, the grid cluster that meets the preset conditions is a spherical grid cluster; the planar grid cluster and the linear grid cluster are regarded as ground grid clusters.

As an improvement of the above scheme, the step of checking the smoothness of the ground grid clusters in the radial direction of the sector grid map and extracting the ground grid clusters that meet the smoothness requirements includes: S21, taking the ground point where the laser radar is located as the starting point;

S22, in the radial direction of the sector grid map, sequentially calculate the gradient between adjacent grids in the ground grid cluster. If the gradient does not meet the gradient requirement, the previous grid in the adjacent grid corresponding to the gradient is the termination grid, and the termination grid and the previous grids are marked as ground grids; S23, the grids after the termination grid are marked as non-ground grids until the height of the current grid is lower than the height of the previous non-ground grid; S24, determine whether the height difference between the current grid and the previous termination grid is less than the preset height difference. If the judgment is yes, take the current grid as the new starting point and enter step S22. If the judgment is no, mark the current grid as a non-ground grid and continue to check the next grid and enter step S24; S25, until all grids in the ground grid cluster are checked, determine whether the number of non-ground grids in the ground grid cluster is greater than the number of ground grids. If the judgment is yes, delete the ground grid cluster. If the judgment is no, retain the ground grid cluster.

As an improvement of the above scheme, the step of constructing a smooth constraint according to the ground grid cluster in the radial direction of the sector grid map and taking the grid that meets the smooth constraint in the grid cluster that does not meet the preset conditions as the ground grid includes: constructing a smooth curve according to the radius length and height of the grid in the ground grid cluster in the radial direction of the sector grid map; smoothing all grids in the radial direction according to the smooth curve to generate a smooth function; substituting the radial length of the grid in the grid cluster that does not meet the preset conditions into the smooth function to calculate the theoretical height of the current grid; judging whether the difference between the theoretical height and the actual height of the current grid is within a preset difference range, if it is judged to be yes, the current grid is a ground grid, and if it is judged to be no, the current grid is a non-ground grid.

Correspondingly, the present invention also provides a three-dimensional laser point cloud road segmentation system, comprising: a map construction module, used to construct a sector grid map; a clustering processing module, used to perform connected domain clustering processing on the sector grid map to construct grid clusters; a feature analysis module, used to perform feature value analysis on grid clusters that meet preset conditions, and extract grid clusters that meet line features and surface features as ground grid clusters; a smoothness check module, used to perform smoothness check on the ground grid clusters in the radial direction of the sector grid map, and extract ground grid clusters that meet the smoothness requirements.

As an improvement of the above scheme, the three-dimensional laser point cloud road segmentation system also includes: a smooth constraint module, which is used to construct a smooth constraint according to the ground grid cluster in the radial direction of the fan-shaped grid map, and use the grid that meets the smooth constraint in the grid cluster that does not meet the preset conditions as the ground grid.

Correspondingly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the steps of the above-mentioned three-dimensional laser point cloud road surface segmentation method.

Correspondingly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the three-dimensional laser point cloud road surface segmentation method are implemented.

The implementation of the present invention has the following beneficial effects:

The present invention performs cluster analysis on the road point cloud to divide the road point cloud into multiple point cloud clusters, performs eigenvalue analysis on the geometric features of the road point cloud, and then uses the smoothness formula to smoothly fit the grid, ultimately achieving efficient and accurate point cloud road segmentation for complex terrain. Therefore, compared with the prior art, the present invention has greatly improved accuracy and robustness.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a first embodiment of a three-dimensional laser point cloud road segmentation method according to the present invention;

FIG2 is a schematic diagram of a sector grid map in the present invention;

FIG3 is a schematic diagram of the distribution of grids in a certain radial direction in the present invention;

FIG. 4 is a schematic diagram of another distribution of grids in a radial direction in the present invention.

5 is a flow chart of a second embodiment of a three-dimensional laser point cloud road segmentation method according to the present invention;

6 is a schematic structural diagram of a first embodiment of a three-dimensional laser point cloud road segmentation system according to the present invention;

FIG. 7 is a schematic structural diagram of a second embodiment of a three-dimensional laser point cloud road segmentation system according to the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

According to the imaging characteristics of 3D laser radar, a large part of the returned point cloud is reflected by the road surface, and this type of point cloud is called ground point cloud. The purpose of road surface segmentation is to separate the ground point cloud from the entire point cloud. Its significance lies in that on the one hand, the drivable area can be extracted based on the separated ground, and on the other hand, since the ground points have no meaning for obstacle detection, a large number of useless points can be reduced when performing obstacle detection. Therefore, road surface segmentation is the prerequisite for using laser radar for obstacle detection, drivable area detection and other related functions.

Referring to FIG. 1 , FIG. 1 shows a flow chart of a first embodiment of a three-dimensional laser point cloud road segmentation method of the present invention, which includes:

S101, constructing a sector grid map.

Specifically, the step of constructing a sector grid map includes:

(1) Projecting the three-dimensional laser point cloud of the laser radar into a fan-shaped grid map, wherein the fan-shaped grid map is composed of a plurality of independent grids.

(2) Calculate the maximum height difference of all points in each grid respectively. If the maximum height difference is greater than a preset threshold, the grid is an obstacle grid and the obstacle grid is deleted.

The maximum height difference refers to the height difference between the highest point and the lowest point in each grid. For example, if a grid has point A (height 2mm), point B (height 9mm) and point C (height 2.5mm), the maximum height of the grid is 9mm, the minimum height is 2mm, and the maximum height difference is 7mm.

During the segmentation process, the three-dimensional laser point cloud of the laser radar must first be projected into the grid as shown in Figure 2, where any number of laser radar points can exist in the same grid; the maximum height difference Hmm of all points in the grid is calculated. If the maximum height difference Hmm is greater than the preset threshold, it means that this grid is an obstacle grid, and this grid is deleted.

S102, performing connected domain clustering processing on the sector grid map to construct grid clusters.

Since the geometric features of the ground in reality are only related to the adjacent ground features, and the ground is relatively flat, the gradient change between adjacent grids is relatively small, so the present invention uses clustering to cluster grids with small adjacent gradient differences into one category. Specifically, the steps of performing connected domain clustering processing on the sector grid map to construct grid clusters include:

(1) In a sector grid map, a grid cluster is established with one grid as the search center.

Create a new grid cluster with a random grid as the search center.

(2) Searching for grids that meet the gradient requirements in a preset area, and adding the grids that meet the gradient requirements to the grid cluster.

As shown in Figure 2, if the preset horizontal range Nh=1 (i.e., based on the search center, it extends horizontally by 1 grid), and the preset vertical range Np=2 (i.e., based on the search center, it extends vertically by 2 grids), then the preset areas are "Grid 1", "Grid 2", "Grid 3", "Grid 4", "Grid 5" and "Grid 6".

The gradient refers to the ratio of the height difference Hg of two grids to the Euclidean distance D of the coordinates of the two grids, that is, the gradient G = Hg/D. If the gradient G is less than the preset gradient Gt, the grid is considered to meet the gradient requirements. Among them, the grid height refers to the average height of all points in each grid.

(3) Within the grid cluster, another grid that is not used as the search center is used as a new search center, and the process goes to step (2) until all grids in the grid cluster have completed the search.

After the gradient judgment of the current search center within the preset range is completed, another new grid is used as the search center in the grid cluster (this grid has never been used as the search center point), and the search operation in step (2) is repeated until all grids that meet the gradient requirements are added to the grid cluster. Finally, when there are no more grids in the grid cluster that have not been used as the search center, the expansion of the grid cluster is completed.

(4) Outside the grid cluster, a new grid cluster is established with another grid that is not used as the search center as a new search center, and the process goes to step (2) until all grids in the fan-shaped grid map have been searched.

A new grid cluster is established with the grid outside the grid cluster as the center, and the search operations of step (2) and step (3) are repeated until there is no starting grid for establishing a new grid cluster. At this time, several grid clusters are formed and clustering is completed.

Therefore, through step S102, effective clustering of grids can be achieved to construct one or more grid clusters, so as to achieve separate processing of grids with strong pertinence.

S103, performing feature value analysis on grid clusters meeting preset conditions, and extracting grid clusters meeting line features and surface features as ground grid clusters.

Specifically, the step of determining whether the grid cluster meets the preset conditions includes:

(1) Convert each grid in a grid cluster into a point and construct a minimum enclosing rectangle for the grid cluster.

Each grid in the grid cluster is approximated as a point (x, y, z), where x is the x-axis coordinate of the grid in the Cartesian coordinate system, y is the y-axis coordinate of the grid in the Cartesian coordinate system, and z is the grid height (the grid height refers to the average height of all points in each grid).

(2) determining whether the number of points within the grid cluster is less than a preset number or whether the diagonal length of the minimum enclosing rectangle corresponding to the grid cluster is less than a preset length; if so, the grid cluster does not meet the preset condition; and if not, the grid cluster meets the preset condition.

It should be noted that when the number of points in the grid cluster is less than the preset number or the diagonal length of the minimum enclosing rectangle corresponding to the grid cluster is less than the preset length, it means that the scale of the grid cluster is small and is not suitable for feature analysis. Therefore, the present invention only performs feature analysis on grid clusters with larger scales, and does not process or analyzes grid clusters with smaller scales in other ways.

Accordingly, the step of performing feature analysis on grid clusters meeting preset conditions and extracting grid clusters meeting line features and surface features as ground grid clusters includes:

(1) Constructing a covariance matrix based on the points in the grid cluster that meet the preset conditions. Specifically, the present invention can use the x, y, and z values of each point in the grid cluster to establish a covariance matrix.

(2) Calculating the eigenvalues of the covariance matrix to obtain the minimum eigenvalue, the middle eigenvalue and the maximum eigenvalue among the eigenvalues.

(3) extracting the minimum eigenvalue, the middle eigenvalue and the maximum eigenvalue from the features, and making a judgment based on the difference between the minimum eigenvalue, the middle eigenvalue and the maximum eigenvalue, wherein:

When the difference between the minimum eigenvalue and the intermediate eigenvalue is not within the preset range, and the difference between the intermediate eigenvalue and the maximum eigenvalue is within the preset range, the grid cluster that meets the preset conditions is a planar grid cluster; that is, when the minimum eigenvalue is very small compared to the other two eigenvalues (the intermediate eigenvalue and the maximum eigenvalue), and the other two eigenvalues (the intermediate eigenvalue and the maximum eigenvalue) are not much different, then the point cloud in the grid cluster is roughly planar, and the normal vector can be in any direction.

When the difference between the minimum eigenvalue and the intermediate eigenvalue is within a preset range, and the difference between the minimum eigenvalue and the maximum eigenvalue is not within the preset range, the grid cluster that meets the preset conditions is a linear grid cluster; that is, when the minimum eigenvalue is not much different from the intermediate eigenvalue, and the minimum eigenvalue is significantly different from the maximum eigenvalue, the point cloud in the grid cluster is roughly linear, and the normal vector can be in any direction.

When the difference between the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue is within a preset range, the grid cluster that meets the preset conditions is a spherical grid cluster; that is, when the three values (minimum eigenvalue, intermediate eigenvalue and maximum eigenvalue) are roughly equal, the point cloud in the grid cluster is roughly spherical.

(4) The planar grid cluster and the linear grid cluster are used as ground grid clusters.

Therefore, step S103 utilizes two-dimensional feature information, rather than just combining information in a single direction or two directions, and has very high accuracy in feature extraction. It can effectively classify grid clusters and perform feature value analysis on larger grid clusters to extract surface grid clusters and linear grid clusters as ground grid clusters with high precision and accuracy.

S104, performing a smoothness check on the ground grid clusters in the radial direction of the sector grid map, and extracting ground grid clusters that meet the smoothness requirement.

It should be noted that after the screening in step S103, the grid clusters whose characteristic values meet the conditions are retained as ground grid clusters, but it is necessary to further check the ground grid clusters, because if a higher platform appears in front of the vehicle (for example, a container-type truck suddenly appears in front), the point cloud projected by the laser radar on such objects is also relatively flat, so it is necessary to eliminate the misjudged grids in similar situations.

The present invention proposes to eliminate misjudged grids by calculating the gradient in the radial direction. Specifically, the step of performing a smoothness check on the ground grid clusters in the radial direction of the sector grid map and extracting the ground grid clusters that meet the smoothness requirements includes:

(1) The ground point where the laser radar is located is taken as the starting point.

(2) The gradients between adjacent grids in the ground grid cluster are calculated in sequence in the radial direction of the sector grid map. If the gradient does not meet the gradient requirement, the previous grid in the adjacent grids corresponding to the gradient is the termination grid, and the termination grid and the previous grids are marked as ground grids.

As shown in FIG3 , the ground point where the laser radar is located is taken as the starting point, and the gradient α of two adjacent grids is calculated in sequence. If the gradient α is greater than the preset gradient α _max , then the previous grid of the two adjacent grids is the termination grid, and the termination grid and the previous grids are marked as ground grids, wherein the gradient is calculated by dividing the height difference of the two adjacent grids by the length difference in the radial direction. For example, grid C in FIG3 is the termination grid, and grids A, B, and C are all ground grids.

(3) Marking the grids after the termination grid as non-ground grids until the height of the current grid is lower than the height of the previous non-ground grid.

(4) Determine whether the height difference between the current grid and the previous end grid is less than the preset height difference. If it is judged to be yes, take the current grid as the new starting point and enter step (2). If it is judged to be no, mark the current grid as a non-ground grid and continue to check the next grid and enter step (4).

The grids after the end grid are marked as non-ground grids in turn until the height of the current grid is lower than the previous non-ground grid. At this time, the height of the current grid is compared with the height of the previous end grid. If the height difference between the current grid and the previous end grid is less than the preset height difference, the current grid is used as the new starting point, otherwise it is marked as a non-ground grid and the next grid is checked.

As shown in Figure 4, the previous ending grid is B. Even if grid E is lower than the previous non-ground grid D, grid E cannot become the new starting grid because the height h ₁ of grid E is much higher than the height of the previous ending grid B (h ₁ >h _min ); the next grid F is judged in sequence. At this time, grid F meets the requirements of becoming a new starting grid (that is, the height difference h ₂ between grid F and the previous ending grid B is less than the preset height difference h _min ), so grid F can be used as the new starting grid.

(5) After all grids in the ground grid cluster are checked, determine whether the number of non-ground grids in the ground grid cluster is greater than the number of ground grids. If so, delete the ground grid cluster; if not, retain the ground grid cluster.

Therefore, by calculating the gradient in the radial direction in step S104, misjudged grids can be effectively eliminated, further ensuring the accuracy of segmentation.

From the above, it can be seen that the present invention performs road segmentation by analyzing the geometric features (points, lines, and surfaces) of the three-dimensional laser point cloud, and can make a better judgment on the shape of the point cloud cluster, thereby judging whether it belongs to a ground point based on the shape of the point cloud cluster; therefore, compared with the prior art, the present invention has greatly improved accuracy and robustness.

Referring to FIG. 5 , FIG. 5 shows a flow chart of a second embodiment of a three-dimensional laser point cloud road segmentation method according to the present invention, which includes:

S201, constructing a sector grid map.

S202: Perform connected domain clustering processing on the sector grid map to construct grid clusters.

S203, performing feature value analysis on the grid clusters meeting the preset conditions, and extracting grid clusters meeting the line features and the surface features as ground grid clusters.

S204, performing a smoothness check on the ground grid clusters in the radial direction of the sector grid map, and extracting ground grid clusters that meet the smoothness requirement.

S205 , in the radial direction of the sector grid map, constructing a smooth constraint according to the ground grid cluster, and taking the grids that meet the smooth constraint in the grid cluster that does not meet the preset condition as the ground grid.

It should be noted that there are still many scattered sparse grids in the grid clusters that do not meet the preset conditions (i.e., the grid clusters that have not been subjected to eigenvalue analysis). There may be ground grids in these grids. Therefore, a smoothing function is set in the radial direction to establish a smoothing constraint. As long as the constraint conditions are met, they can be added to the ground grid class. In the present invention, the following method can be used to process sparse grids:

(1) constructing a smooth curve in the radial direction of the fan-shaped grid map according to the radius length and height of the grid in the ground grid cluster;

(2) performing smoothing processing on all grids in the radial direction according to the smooth curve to generate a smoothing function;

(3) Substituting the radial length of the grid in the grid cluster that does not meet the preset conditions into the smoothing function to calculate the theoretical height of the current grid;

(4) Determine whether the difference between the theoretical height and the actual height of the current grid is within a preset difference range. If so, the current grid is a ground grid; if not, the current grid is a non-ground grid.

In each radial direction, a cubic B-spline curve is established with the radius length of the grid in the ground grid cluster as the horizontal coordinate and the height as the vertical coordinate. The cubic B-spline curve is used as a smoothing cancellation and the grid in the entire radial direction is smoothed using the smoothing curve to obtain a smoothing function for each segment. Then, the radial length of the grid in the small-scale grid cluster (i.e., the grid cluster without eigenvalue analysis) is substituted into the smoothing function as the horizontal coordinate, and the obtained theoretical height H _s is compared with the actual height _Hi . If |H _s -H _i | is less than the preset difference H _diff , then this grid belongs to the ground grid, where the preset difference H _diff can be a difference threshold.

Therefore, different from the first embodiment shown in FIG. 1 , in this embodiment, the ground grid can be accurately extracted by adding further classification processing to the sparse grid.

Referring to FIG. 6 , FIG. 6 shows a first embodiment of a three-dimensional laser point cloud road segmentation system 100 of the present invention, which includes:

The map construction module 1 is used to construct a sector grid map. Specifically, the map construction module 1 projects the three-dimensional laser point cloud of the laser radar into the sector grid map, and then calculates the maximum height difference of all points in each grid respectively. If the maximum height difference is greater than a preset threshold, the grid is an obstacle grid and the obstacle grid is deleted.

The cluster processing module 2 is used to perform connected domain clustering processing on the sector grid map to construct a grid cluster. Specifically, the cluster processing module 2 establishes a grid cluster in the sector grid map with one grid as the search center; then, searches for grids that meet the gradient requirements in a preset area, and adds the grids that meet the gradient requirements to the grid cluster; then, within the grid cluster, another grid that is not used as the search center is used as the new search center to search for information again until all grids in the grid cluster are searched; finally, outside the grid cluster, another grid that is not used as the search center is used as the new search center to establish a new grid cluster, and search again until all grids in the sector grid map are searched.

The feature analysis module 3 is used to perform feature value analysis on the grid clusters that meet the preset conditions, and extract the grid clusters that meet the line features and surface features as the ground grid clusters. It should be noted that the method for determining whether the grid cluster meets the preset conditions includes: (1) converting each grid in the grid cluster into a point, and constructing a minimum enclosing rectangle for the grid cluster. (2) determining whether the number of points in the grid cluster is less than a preset number or whether the diagonal length of the minimum enclosing rectangle corresponding to the grid cluster is less than a preset length. If the judgment is yes, the grid cluster does not meet the preset conditions, and if the judgment is no, the grid cluster meets the preset conditions. Specifically, the feature analysis module 3 constructs a covariance matrix according to the points in the grid cluster that meets the preset conditions; then, the eigenvalues of the covariance matrix are calculated to obtain the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue in the eigenvalues; then, the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue in the features are extracted, and a judgment is made according to the difference between the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue, wherein: when the difference between the minimum eigenvalue and the intermediate eigenvalue is not within the preset range, and the difference between the intermediate eigenvalue and the maximum eigenvalue is within the preset range, the grid cluster that meets the preset conditions is a planar grid cluster; that is, when the minimum eigenvalue is very small compared to the other two eigenvalues (the intermediate eigenvalue and the maximum eigenvalue), and the other two eigenvalues (the intermediate eigenvalue and the maximum eigenvalue) are not much different, then the point cloud in the grid cluster is roughly planar, and the normal vector can be in any direction. When the difference between the minimum eigenvalue and the intermediate eigenvalue is within the preset range, and the difference between the minimum eigenvalue and the maximum eigenvalue is not within the preset range, the grid cluster that meets the preset conditions is a linear grid cluster; that is, when the minimum eigenvalue and the intermediate eigenvalue are not much different, and the minimum eigenvalue and the maximum eigenvalue are greatly different, the point cloud in the grid cluster is roughly linear, and the normal vector can be in any direction. When the difference between the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue is within the preset range, the grid cluster that meets the preset conditions is a spherical grid cluster; that is, when the three values (minimum eigenvalue, intermediate eigenvalue and maximum eigenvalue) are roughly equal, the point cloud in the grid cluster is roughly spherical; finally, the planar grid cluster and the linear grid cluster are used as ground grid clusters.

The smoothness checking module 4 is used to check the smoothness of the ground grid clusters in the radial direction of the sector grid map, and extract the ground grid clusters that meet the smoothness requirements. Specifically, the smoothing check module 4 takes the ground point where the laser radar is located as the starting point; then, the gradient between adjacent grids in the ground grid cluster is calculated in turn in the radial direction of the fan-shaped grid map. If the gradient does not meet the gradient requirement, the previous grid in the adjacent grid corresponding to the gradient is the termination grid, and the termination grid and the previous grids are marked as ground grids; then, the grids after the termination grid are marked as non-ground grids until the height of the current grid is lower than the height of the previous non-ground grid; then, it is determined whether the height difference between the current grid and the previous termination grid is less than the preset height difference. If it is determined to be yes, the current grid is used as a new starting point and the detection is performed again. If it is determined to be no, the current grid is marked as a non-ground grid and the next grid is checked; until all the grids in the ground grid cluster are checked, it is determined whether the number of non-ground grids in the ground grid cluster is greater than the number of ground grids. If it is determined to be yes, the ground grid cluster is deleted. If it is determined to be no, the ground grid cluster is retained.

From the above, it can be seen that the present invention performs road segmentation by analyzing the geometric features (points, lines, and surfaces) of the three-dimensional laser point cloud, and can make a better judgment on the shape of the point cloud cluster, thereby judging whether it belongs to a ground point based on the shape of the point cloud cluster; therefore, compared with the prior art, the present invention has greatly improved accuracy and robustness.

Referring to FIG. 7 , FIG. 7 shows a second embodiment of a three-dimensional laser point cloud road segmentation system 100 of the present invention. Different from the first embodiment shown in FIG. 6 , the three-dimensional laser point cloud road segmentation system 100 in this embodiment further includes: a smooth constraint module 5 for constructing a smooth constraint according to the ground grid cluster in the radial direction of the sector grid map, and using the grid that meets the smooth constraint in the grid cluster that does not meet the preset conditions as the ground grid.

Specifically, the smoothing constraint module 5 constructs a smooth curve in the radial direction of the sector grid map according to the radius length and height of the grid in the ground grid cluster; then smoothes all the grids in the radial direction according to the smooth curve to generate a smooth function; then, substitutes the radial length of the grid in the grid cluster that does not meet the preset conditions into the smooth function to calculate the theoretical height of the current grid; finally, determines whether the difference between the theoretical height and the actual height of the current grid is within a preset difference range, and when the judgment is yes, the current grid is a ground grid, and when the judgment is no, the current grid is a non-ground grid.

Therefore, the present invention adds further classification processing to the sparse grid through the smoothing constraint module 5, so that the ground grid can be accurately extracted.

In this embodiment, by adding further classification processing to the sparse grid, the ground grid can be accurately extracted. Accordingly, the present invention also provides a computer device, including a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above-mentioned three-dimensional laser point cloud road surface segmentation method when executing the computer program. At the same time, the present invention also provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the above-mentioned three-dimensional laser point cloud road surface segmentation method when executed by the processor.

Therefore, the present invention performs cluster analysis on the road point cloud to divide the road point cloud into multiple point cloud clusters, performs eigenvalue analysis on the geometric features of the road point cloud, uses the smoothness formula to smoothly fit the grid, and further classifies the sparse grid, ultimately achieving efficient and accurate point cloud road segmentation for complex terrain.

The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims (12)

1. A 3D laser point cloud road segmentation method, comprising: Construct sector grid maps; According to the gradient requirement, the sector grid map is subjected to connected domain clustering processing to construct a grid cluster; Performing eigenvalue analysis on grid clusters that meet preset conditions, extracting grid clusters that meet line features and surface features as ground grid clusters; the grid clusters that meet preset conditions refer to grid clusters whose number of points is less than a preset number or whose diagonal length of the minimum enclosing rectangle is less than a preset length; the specific steps include: constructing a covariance matrix based on the points in the grid clusters that meet the preset conditions; calculating the eigenvalues of the covariance matrix; extracting the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue in the features, and taking the grid clusters whose difference between the minimum eigenvalue and the intermediate eigenvalue is not within a preset range and the difference between the intermediate eigenvalue and the maximum eigenvalue is within a preset range, and the grid clusters whose difference between the minimum eigenvalue and the intermediate eigenvalue is within a preset range and the difference between the minimum eigenvalue and the maximum eigenvalue is not within a preset range as ground grid clusters; The ground grid clusters are checked for smoothness in the radial direction of the sector grid map, and ground grid clusters meeting the smoothness requirement are extracted. 2. The three-dimensional laser point cloud road segmentation method as described in claim 1 is characterized in that it also includes: in the radial direction of the fan-shaped grid map, constructing a smooth constraint based on the ground grid cluster, and using the grid that meets the smooth constraint in the grid cluster that does not meet the preset conditions as the ground grid. 3. The three-dimensional laser point cloud road segmentation method according to claim 1 or 2, characterized in that the step of constructing a sector grid map comprises: Projecting the three-dimensional laser point cloud of the laser radar into a fan-shaped grid map, wherein the fan-shaped grid map is composed of a plurality of mutually independent grids; The maximum height difference of all points in each grid is calculated respectively. If the maximum height difference is greater than a preset threshold, the grid is an obstacle grid and the obstacle grid is deleted. 4. The three-dimensional laser point cloud road segmentation method according to claim 1 or 2, characterized in that the step of performing connected domain clustering processing on the sector grid map to construct grid clusters comprises: S11, in the sector grid map, taking one grid as the search center, establishing a grid cluster; S12, searching for grids that meet the gradient requirements in a preset area, and adding the grids that meet the gradient requirements to the grid cluster; S13, in the grid cluster, taking another grid that is not used as the search center as a new search center, and entering step S12 until all grids in the grid cluster are searched; S14, outside the grid cluster, using another grid that is not used as the search center as a new search center, establishing a new grid cluster, and entering step S12 until all grids in the fan-shaped grid map have completed the search. 5. The three-dimensional laser point cloud road segmentation method according to claim 1 or 2, characterized in that the step of determining whether the grid cluster meets the preset conditions comprises: Convert each grid in the grid cluster into a point and construct a minimum enclosing rectangle for the grid cluster; Determine whether the number of points in the grid cluster is less than a preset number or whether the diagonal length of the minimum enclosing rectangle corresponding to the grid cluster is less than a preset length, When the judgment is yes, the grid cluster does not meet the preset condition. When the judgment is no, the grid cluster meets the preset condition. 6. The three-dimensional laser point cloud road segmentation method according to claim 5, characterized in that the step of performing feature analysis on grid clusters that meet preset conditions and extracting grid clusters that meet line features and surface features as ground grid clusters comprises: Constructing a covariance matrix based on the points in the grid cluster that meet the preset conditions; Calculating eigenvalues of the covariance matrix; Extract the minimum eigenvalue, the middle eigenvalue and the maximum eigenvalue from the features, and make a judgment based on the difference between the minimum eigenvalue, the middle eigenvalue and the maximum eigenvalue, wherein: When the difference between the minimum eigenvalue and the intermediate eigenvalue is not within the preset range, and the difference between the intermediate eigenvalue and the maximum eigenvalue is within the preset range, the grid cluster meeting the preset conditions is a planar grid cluster; When the difference between the minimum eigenvalue and the intermediate eigenvalue is within a preset range, and the difference between the minimum eigenvalue and the maximum eigenvalue is not within a preset range, the grid cluster meeting the preset condition is a linear grid cluster; When the difference between the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue is within a preset range, the grid cluster meeting the preset condition is a spherical grid cluster; The planar grid cluster and the linear grid cluster are used as ground grid clusters. 7. The three-dimensional laser point cloud road segmentation method according to claim 1 or 2, characterized in that the step of performing a smoothness check on the ground grid clusters in the radial direction of the sector grid map and extracting the ground grid clusters that meet the smoothness requirements comprises: S21, taking the ground point where the laser radar is located as the starting point; S22, sequentially calculating the gradient between adjacent grids in the ground grid cluster in the radial direction of the sector grid map, if the gradient does not meet the gradient requirement, the previous grid in the adjacent grids corresponding to the gradient is the termination grid, and the termination grid and the previous grids are marked as ground grids; S23, marking the grids after the end grid as non-ground grids until the height of the current grid is lower than the height of the previous non-ground grid; S24, determining whether the height difference between the current grid and the last ending grid is less than a preset height difference, If the answer is yes, the current grid is used as the new starting point and the process goes to step S22. If the result is negative, the current grid is marked as a non-ground grid and the next grid is checked, and the process goes to step S24; S25, until all grids in the ground grid cluster are checked, determine whether the number of non-ground grids in the ground grid cluster is greater than the number of ground grids, if it is determined to be yes, delete the ground grid cluster, if not, retain the ground grid cluster. 8. The three-dimensional laser point cloud road segmentation method according to claim 2, characterized in that the step of constructing a smooth constraint according to the ground grid cluster in the radial direction of the sector grid map and taking the grid that meets the smooth constraint in the grid cluster that does not meet the preset conditions as the ground grid comprises: In the radial direction of the sector grid map, a smooth curve is constructed according to the radius length and height of the grids in the ground grid cluster; Smoothing all grids in the radial direction according to the smooth curve to generate a smooth function; Substituting the radial length of the grid in the grid cluster that does not meet the preset condition into the smoothing function to calculate the theoretical height of the current grid; It is determined whether the difference between the theoretical height and the actual height of the current grid is within a preset difference range. If it is determined to be yes, the current grid is a ground grid; if it is determined to be no, the current grid is a non-ground grid. 9. A three-dimensional laser point cloud road segmentation system, characterized by comprising: Map construction module, used to construct sector grid map; A clustering processing module, used for performing connected domain clustering processing on the sector grid map to construct grid clusters according to gradient requirements; A feature analysis module is used to perform eigenvalue analysis on grid clusters that meet preset conditions, and extract grid clusters that meet line features and surface features as ground grid clusters; the grid clusters that meet the preset conditions refer to grid clusters whose number of points is less than a preset number or whose diagonal length of the minimum enclosing rectangle is less than a preset length; the feature analysis module constructs a covariance matrix based on the points in the grid clusters that meet the preset conditions; calculates the eigenvalues of the covariance matrix; extracts the minimum eigenvalue, the intermediate eigenvalue and the maximum eigenvalue in the features, and takes the grid clusters whose difference between the minimum eigenvalue and the intermediate eigenvalue is not within a preset range and the difference between the intermediate eigenvalue and the maximum eigenvalue is within a preset range, and the grid clusters whose difference between the minimum eigenvalue and the intermediate eigenvalue is within a preset range and the difference between the minimum eigenvalue and the maximum eigenvalue is not within a preset range as ground grid clusters; The smoothness checking module is used to check the smoothness of the ground grid clusters in the radial direction of the sector grid map, and extract the ground grid clusters that meet the smoothness requirements. 10. The three-dimensional laser point cloud road segmentation system as described in claim 9 is characterized by further comprising: a smooth constraint module, which is used to construct a smooth constraint according to the ground grid cluster in the radial direction of the fan-shaped grid map, and use the grid that meets the smooth constraint in the grid cluster that does not meet the preset conditions as the ground grid. 11. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein the processor implements the steps of the method according to any one of claims 1 to 8 when executing the computer program. 12. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.