Dynamic threshold determination method based on point cloud density and spacing and Euclidean clustering method

Abstract

The dynamic threshold determination method and Euclidean clustering method based on point cloud density and spacing belong to the field of autonomous driving environment perception technology. In order to solve the problem that the use of a unified distance threshold for Euclidean clustering for obstacle recognition in the entire range has a high under-segmentation rate of long-distance obstacle clustering due to the fixed threshold. The present invention first establishes a KD-tree based on the laser point cloud after ground segmentation; based on the established KD-tree, the average Euclidean distance between the point clouds in each set is calculated #imgabs0#, and the compensation factor e is determined based on the average Euclidean distance #imgabs1# of the point cloud set; the initial threshold d is determined based on two points in the cluster set. i ; Finally, the distance threshold d is determined based on the compensation factor and the initial threshold t =ed+d i , where d is the distance from the point cloud to the center of the laser radar. The present invention is mainly used for clustering point clouds.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving environment perception and relates to a laser point cloud threshold determination method and a clustering method. Background Art

[0002] Clustering algorithm is the most commonly used target detection algorithm for laser point cloud. It has the advantages of simple algorithm, fast speed and no need for training. Clustering is to divide point clouds with similar features into one category, and divide a dense point set into several clusters and classes based on distance or density. Euclidean clustering is often called Euclidean clustering. Euclidean clustering is a clustering algorithm based on Euclidean distance measurement. The principle of the Euclidean distance algorithm is to select a point in the input point cloud space, and find the points adjacent to it with this point as the center. The Euclidean distance between the adjacent points and the center point must be less than a given threshold. All adjacent points that meet the threshold condition are added to set A. If the number of laser point clouds in the set also meets the quantity threshold, the point cloud in the set is the clustering target. In this article, the target is the obstacle target. Then select the point outside the set A as the center point, and repeat the above actions until all points outside the set become the center point. All sets that become targets are regarded as obstacle point clouds. That is, they are divided into different obstacles. The threshold setting of traditional Euclidean clustering is very important, and it mainly depends on two thresholds: one is the distance threshold, and the other is the point cloud number threshold. These two thresholds are important factors in determining whether a point cloud cluster can be fitted. For normal point clouds, the number of point clouds projected on the target is large, and the point cloud clustering effect is good. However, due to the fixed emission angle of the laser radar, the distance between the point clouds will increase at a distance, so it cannot meet the distance threshold requirement, so the clustering segmentation rate for distant targets is low.

[0003] Euclidean clustering is more accurate for clustering close-range point clouds and densely packed point clouds. Within a sensing distance of 30 m (based on the laser radar coordinate system), the over-segmentation rate and under-segmentation rate are as follows: Figure 1 As shown. It can be seen from the figure that at a long distance, small objects (such as bicycles, people, trees, etc.) are prone to under-segmentation of clustering, that is, they cannot be clustered. At this time, due to the laser angle of the lidar, the number of point clouds on distant objects is small, and the spacing of the point clouds increases with the increase of the target distance. When the target distance is too large, the number of point clouds formed by the target reflection is less than the quantity threshold, or the point cloud spacing of the target is too large to meet the conditions of Euclidean clustering, then it cannot be clustered and under-segmentation occurs. In response to the under-segmentation phenomenon, the present invention proposes a Euclidean clustering optimization method based on dynamic thresholds of density and spacing. Summary of the invention

[0004] The present invention aims to solve the problem that a fixed threshold is used to identify obstacles in the entire range for Euclidean clustering, and a high under-segmentation rate of obstacles clustering at a long distance is caused.

[0005] The dynamic threshold determination method based on point cloud density and spacing includes the following steps:

[0006] Establish a KD-tree based on the laser point cloud after ground segmentation; calculate the average Euclidean distance between point clouds in each set based on the established KD-tree Average Euclidean distance based on point cloud collection Determine the compensation factor Where l is the number of laser radar lines, and ρ is the point cloud density in the point cloud set;

[0007] Determine the initial threshold based on two points in the cluster set Among them, p 1 and p 2 The two points in the cluster set with the highest point cloud density and meeting the point cloud quantity threshold;

[0008] Finally, the distance threshold d is determined according to the compensation factor e and the initial threshold t =ed+d i , where d is the distance from the point cloud to the center of the lidar.

[0009] Furthermore, the average Euclidean distance of the point cloud set Among them, d n (p i ) is any point p in the point cloud set i The Euclidean distance to its nearest neighbor, N represents the total number of point clouds in the point cloud set.

[0010] Furthermore, the process of determining the point cloud density ρ in the point cloud set includes the following steps:

[0011] After obtaining the average Euclidean distance, the point cloud space is divided into cubic grids, that is, the space is voxelized. Assuming the side length of the grid is ⊿a×⊿b×⊿c, the point cloud space can be divided into s cubic grids;

[0012] s=a×b×c

[0013] Where a, b, c are the number of grid planes in each direction of the cube;

[0014] After dividing the point cloud space into cubic grids, if there is a point cloud in the grid, the grid is recorded as 1, and the number of point clouds n is recorded i , if there is no point cloud in the grid, it is recorded as 0; the point cloud density in the point cloud set is:

[0015]

[0016] Where s(1) represents the number of grids marked as 1, n i Indicates the number of point clouds in the grid.

[0017] Furthermore, the number of grid planes in each direction of the cube is as follows:

[0018]

[0019] Among them, int means round down; x max 、x min They are the maximum and minimum values ​​of the x-axis coordinates of the point cloud in the point cloud set, and the y max ,y min are the maximum and minimum values ​​of the y-axis coordinates of the point cloud in the point cloud set, and z max 、z min They are respectively the maximum and minimum values ​​of the z-axis coordinates of the point cloud in the point cloud set.

[0020] The dynamic threshold Euclidean clustering method based on point cloud density and spacing first determines the distance threshold using the dynamic threshold determination method based on point cloud density and spacing. After obtaining the distance threshold, clustering is performed using the Euclidean clustering method based on the set clustering parameters of the Euclidean clustering.

[0021] Furthermore, the process of clustering using the Euclidean clustering method includes the following steps:

[0022] S201. Create a set Q 1 , using KD-tree in o 1 Search for the nearest point outside the point and find n points p that meet the distance threshold 1 ,p 2 …,p n And point p joins the set Q 1 In, Q 1 =(o 1 ,p 1 ,p 2 …,p n ), where the set Q 1 The points in satisfy:

[0023]

[0024] where p i For the set Q 1 In addition to o 1 Any point outside i is the nearest search distance,

[0025] S202, in set Q 1Select another point as the center point and create another set Q 2 , find points that meet the distance threshold in the same way as step S201, traverse all input points, and repeat the above steps until all points become a set Q i The point or center point in the set Q is recorded. i The number of point clouds in t i ,

[0026] When the quantity threshold condition is met, that is, t min <t i <t max When the set Q i The point cloud inside is the obstacle point cloud, where t min and t max is the clustering quantity threshold; the point cloud in the set that does not meet the quantity threshold condition is an invalid point cloud.

[0027] Beneficial effects:

[0028] The present invention optimizes the distance threshold of Euclidean clustering by adopting a dynamic threshold method based on the point cloud density and average spacing of the point cloud cluster, and adopts different thresholds for point cloud clusters with different point cloud densities and average spacings, thereby achieving the purpose of improving the clustering accuracy of distant targets and reducing the under-segmentation rate. The method of the present invention can improve the recognition rate of distant targets and increase the overall clustering effect, and can effectively solve the problem of large spacing of distant target point clouds and high clustering under-segmentation rate in the existing Euclidean clustering method. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the relationship diagram between obstacle distance and segmentation rate;

[0030] Figure 2 It is a schematic diagram of the process of the present invention;

[0031] Figure 3(a) and Figure 3(b) are the effect diagram of Euclidean clustering and its local enlarged diagram respectively;

[0032] FIG. 4( a ) and FIG. 4( b ) are respectively the effect diagrams of the traditional Euclidean clustering and the Euclidean clustering after optimization of the present invention;

[0033] FIG5(a), FIG5(b), and FIG5(c) are comparison diagrams of over-segmentation rate, under-segmentation rate, and number of clusters between traditional Euclidean clustering and the optimized method of the present invention. DETAILED DESCRIPTION

[0034] The present invention is a technology that uses sensors (such as laser radar and visual cameras) in autonomous driving to survey the surrounding environment and identify obstacles. The present invention uses a dynamic threshold Euclidean clustering optimization algorithm based on point cloud density to optimize the fixed threshold in the algorithm program, calculate the point cloud density and point cloud spacing of point cloud clusters in different areas, and calculate the value with a certain proportional coefficient to obtain different distance thresholds, thereby optimizing the threshold requirements in a dynamic threshold manner. The present invention is described in more detail below in conjunction with specific implementation methods.

[0035] Specific implementation method 1: Combination Figure 2 To explain this embodiment,

[0036] This embodiment is a method for determining a dynamic threshold based on point cloud density and spacing, comprising the following steps:

[0037] (1) KD-tree is established based on the laser point cloud after ground segmentation. The key to Euclidean distance is to find the nearest point cloud. This method relies on the nearest neighbor search algorithm of KD-tree for acceleration. KD-tree is a tree-like data structure that can sort a large number of disordered spatial point clouds into a certain order. It can more effectively perform nearest neighbor search on three-dimensional point clouds in space, thereby reducing clustering time.

[0038] (2) The distance threshold should be changed dynamically for obstacles at different distances. The threshold setting depends on the point cloud density per unit area and the interval distance. Since the dynamic setting requires the initial threshold to be determined first, the initial threshold needs to ensure that the detection rate is high and stable within a short distance range. When the distance is short, the point cloud density is high, the point cloud interval is short, and it is easy to segment. The initial threshold can be set as:

[0039]

[0040] Among them, p 1 and p 2 are two points in the cluster set with the highest point cloud density and meeting the point cloud quantity threshold.

[0041] (3) Calculate the impact of the average point cloud spacing on the threshold, that is, the impact of the average distance between point clouds in the point cloud cluster on the threshold. The average point cloud spacing is the average value of the Euclidean distance between each point in the point cloud cluster and the nearest point. Through experiments, it can be found that the average point cloud spacing of the cluster set is higher at 30m. Due to the limitation of the laser radar emission angle, the average point cloud spacing increases with the increase of the obstacle distance, so the distance threshold can be set as:

[0042] d t =ed+d i

[0043] Where e is the compensation factor, di is the initial clustering threshold, d is the distance from the point cloud to the center of the lidar; as the distance increases, the distance threshold needs to be increased accordingly; and it is ensured that small obstacles can be accurately clustered successfully at a longer distance.

[0044] (4) In addition to the average point cloud spacing, the point cloud density in the point cloud cluster also affects the selection of the threshold. The point cloud density is the number of point clouds per unit volume. The point cloud density has a greater impact on the clustering threshold. The farther the obstacle is, the sparser the point cloud density is. After point cloud preprocessing, the average Euclidean distance between point clouds in each set is calculated for the established KD-tree to represent the spacing of point clouds in each set. The calculation formula is as follows:

[0045]

[0046] in is the average Euclidean distance of the point cloud set; d n (p i ) is any point p in the point cloud set i The Euclidean distance to its nearest neighbor; N represents the total number of point clouds in the point cloud set.

[0047] (5) After obtaining the average Euclidean distance, the point cloud space is divided into cubic grids. That is, the space is voxelized, and the side length of the grid is set to ⊿a×⊿b×⊿c, then the point cloud space can be divided into s cubic grids. It can be obtained:

[0048] s=a×b×c

[0049] Where a, b, c are the number of grid planes in each direction of the cube, and the calculation formula is:

[0050]

[0051] Among them, int means round down; x max 、x min They are the maximum and minimum values ​​of the x-axis coordinates of the point cloud in the point cloud set, and the y max ,y min are the maximum and minimum values ​​of the y-axis coordinates of the point cloud in the point cloud set, and z max 、z min They are the maximum and minimum values ​​of the z-axis coordinates of the point cloud in the point cloud set;

[0052] After dividing the point cloud space into cubic grids, if there is a point cloud in the grid, the grid is recorded as 1, and the number of point clouds n is recorded i , if there is no point cloud in the grid, it is recorded as 0; the point cloud density in the point cloud set can be expressed as:

[0053]

[0054] Where ρ is the point cloud density in the point cloud set, s(1) represents the number of grids marked as 1, and n i Indicates the number of point clouds in the grid.

[0055] (6) By calculating the point cloud spacing and point cloud density, the compensation factor e can be obtained, where e is obtained by formula (1), and finally the distance threshold is obtained by formula (2). The distance threshold of clustering can be dynamically changed with the change of point cloud density and spacing to achieve the purpose of dynamic threshold and improve the clustering accuracy of long-distance targets.

[0056]

[0057]

[0058] Where l is the number of laser radar lines.

[0059] Specific implementation method 2: Combination Figure 2 To explain this embodiment,

[0060] This implementation is a dynamic threshold Euclidean clustering method based on point cloud density and spacing, including the following steps:

[0061] (1) KD-tree is established based on the laser point cloud after ground segmentation. The key to Euclidean distance is to find the nearest point cloud. This method relies on the nearest neighbor search algorithm of KD-tree for acceleration. KD-tree is a tree-like data structure that can sort a large number of disordered spatial point clouds into a certain order. It can more effectively perform nearest neighbor search on three-dimensional point clouds in space, thereby reducing clustering time.

[0062] (2) The distance threshold should be changed dynamically for obstacles at different distances. The threshold setting depends on the point cloud density per unit area and the interval distance. Since the dynamic setting requires the initial threshold to be determined first, the initial threshold needs to ensure that the detection rate is high and stable within a short distance range. When the distance is short, the point cloud density is high, the point cloud interval is short, and it is easy to segment. The initial threshold can be set as:

[0063]

[0064] Among them, p 1 and p 2 are two points in the cluster set with the highest point cloud density and meeting the point cloud quantity threshold.

[0065] (3) Calculate the impact of the average point cloud spacing on the threshold, that is, the impact of the average distance between point clouds in the point cloud cluster on the threshold. The average point cloud spacing is the average value of the Euclidean distance between each point in the point cloud cluster and the nearest point. Through experiments, it can be found that the average point cloud spacing of the cluster set is higher at 30m. Due to the limitation of the laser radar emission angle, the average point cloud spacing increases with the increase of the obstacle distance, so the distance threshold can be set as:

[0066] d t =ed+d i

[0067] Where e is the compensation factor, d i is the initial clustering threshold, d is the distance from the point cloud to the center of the lidar. As the distance increases, the distance threshold needs to be increased accordingly; and it is ensured that small obstacles can be accurately clustered successfully at a longer distance.

[0068] (4) In addition to the average point cloud spacing, the point cloud density in the point cloud cluster also affects the selection of the threshold. The point cloud density is the number of point clouds per unit volume. The point cloud density has a greater impact on the clustering threshold. The farther the obstacle is, the sparser the point cloud density is. After point cloud preprocessing, the average Euclidean distance between point clouds in each set is calculated for the established KD-tree to represent the spacing of point clouds in each set. The calculation formula is as follows:

[0069]

[0070] in is the average Euclidean distance of the point cloud set; d n (p i ) is any point p in the point cloud set i The Euclidean distance to its nearest neighbor; N represents the total number of point clouds in the point cloud set.

[0071] (5) After obtaining the average Euclidean distance, the point cloud space is divided into cubic grids. That is, the space is voxelized, and the side length of the grid is set to ⊿a×⊿b×⊿c, then the point cloud space can be divided into s cubic grids. It can be obtained:

[0072] s=a×b×c

[0073] Where a, b, c are the number of grid planes in each direction of the cube, and the calculation formula is:

[0074]

[0075] Among them, int means round down; x max 、x min They are the maximum and minimum values ​​of the x-axis coordinates of the point cloud in the point cloud set, and the y max ,y minare the maximum and minimum values ​​of the y-axis coordinates of the point cloud in the point cloud set, and z max 、z min They are the maximum and minimum values ​​of the z-axis coordinates of the point cloud in the point cloud set;

[0076] After dividing the point cloud space into cubic grids, if there is a point cloud in the grid, the grid is recorded as 1, and the number of point clouds n is recorded i , if there is no point cloud in the grid, it is recorded as 0; the point cloud density in the point cloud set can be expressed as:

[0077]

[0078] Where ρ is the point cloud density in the point cloud set, s(1) represents the number of grids marked as 1, and n i Indicates the number of point clouds in the grid.

[0079] (6) By calculating the point cloud spacing and point cloud density, the compensation factor e can be obtained, where e is obtained by formula (1), and finally the distance threshold is obtained by formula (2). The distance threshold of clustering can be dynamically changed with the change of point cloud density and spacing to achieve the purpose of dynamic threshold and improve the clustering accuracy of long-distance targets.

[0080]

[0081]

[0082] Where l is the number of laser radar lines.

[0083] (7) After obtaining the distance threshold, set the clustering parameters. The specific parameter settings are shown in Table 1.

[0084] Table 1. European clustering parameter settings

[0085] Settings Parameter Value Maximum number of point clouds 60000 Minimum number of point clouds 20 Distance Threshold <![CDATA[d t ]]> Minimum cluster height -0.5m Maximum cluster height 2.2m The maximum left length of the detection box 5m The maximum left length of the detection box 5m

[0086] (8) After setting various parameters, create a set Q 1 , using KD-tree in o 1 Search for the nearest point outside the point and find n points p that meet the distance threshold 1 ,p 2 …,p n And point p joins the set Q 1 In, Q 1 =(o 1 ,p 1 ,p 2 …,p n ), where the set Q 1 The points in satisfy:

[0087]

[0088] where p i For the set Q 1 In addition to o 1 Any point outside i is the closest search distance.

[0089] (9) In the set Q 1 Select another point as the center point and create another set Q 2 , follow the method in step (8) to find points that meet the distance threshold and add them, traverse all input points, and repeat the above steps until all points become a set Q i The point or center point in the set Q is recorded. i The number of point clouds in t i .

[0090] (10) When the quantity threshold condition is met, that is, t min <t i <t max When the set Q i The point cloud inside is the obstacle point cloud, where t min and t max is the clustering threshold; the point cloud in the set that does not meet the threshold condition is an invalid point cloud. i If it is too large, it means that the object is not clustered normally or the object is a large object such as a building. i If it is too small, it means that the object is small or far away.

[0091] (11) Load the Euclidean clustering framework to cluster obstacles. The results are shown in Figure 3(a) and Figure 3(b), where Figure 3(a) and Figure 3(b) are the Euclidean clustering effect diagram and its local enlarged diagram, respectively. Euclidean clustering fits obstacles that meet the quantity threshold and distance threshold, and calculates the vertex coordinates of the three-dimensional detection box, as shown by the green line in Figure 3(b), and calculates the center position of the point cloud set, as shown by the blue dot in Figure 3(b). Finally, the distance from the center of gravity to the origin of the laser radar coordinate system is calculated, and the specific distance value of 15.8m is marked, as shown in Figure 3(b).

[0092] Based on the information shown in FIG3(a), it can be seen that the clustering frame decreases as the target distance increases. The dynamic threshold clustering optimization algorithm based on density and distance of the present invention is used, and compared with the traditional Euclidean clustering, the clustering effect comparison in the same scene is shown in FIG4(a) and FIG4(b). FIG4(a) shows that the number of clusters of the traditional Euclidean clustering is only 31, and the clustering effect of objects far from the center of the laser radar at the edge is poor. However, the number of clusters using the dynamic threshold clustering method of the present invention reaches 45, as shown in FIG4(b), the number of cluster fittings is greatly improved, and the recognition rate of objects far from the edge is improved.

[0093] The traditional method has poor segmentation effect in the edge areas of the point cloud space and in areas at a long distance. Through the cluster optimization of the present invention, the adaptive threshold is used at different distances and different point cloud densities to reduce the cluster under-segmentation rate and increase the number of effective clusters. In addition, in order to test the degree of improvement of the clustering accuracy of the optimization algorithm, 5 groups of experiments are conducted to compare the effects of the two methods on the under-segmentation rate, over-segmentation rate and number of clusters. The results are shown in Figure 5 (a)-Figure 5 (c). The experimental results show that the optimization effect of the optimization algorithm improves with the increase of the target distance. When the target distance is greater than 30m, compared with the traditional method, the under-segmentation rate of each experimental group decreases, with an average decrease of 9.23%, and the number of clusters increases, with an average increase of 3.58, while the over-segmentation rate changes less, with an average decrease of 1.2%.

[0094] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A dynamic threshold determination method based on point cloud density and spacing, characterized in that: The following steps are involved: Establish a KD-tree based on the laser point cloud after ground segmentation; calculate the average Euclidean distance between point clouds in each set based on the established KD-tree Average Euclidean distance based on point cloud collection Determine the compensation factor Where l is the number of laser radar lines, and ρ is the point cloud density in the point cloud set; Determine the initial threshold based on two points in the cluster set Among them, p1 and p2 are two points in the cluster set with the highest point cloud density and meeting the point cloud quantity threshold; Finally, the distance threshold d is determined according to the compensation factor e and the initial threshold t =ed+d i , where d is the distance from the point cloud to the center of the lidar.

2. The method for determining a dynamic threshold based on point cloud density and spacing according to claim 1, characterized in that: The average Euclidean distance of the point cloud set Among them, d n (p i ) is any point p in the point cloud set i The Euclidean distance to its nearest neighbor, N represents the total number of point clouds in the point cloud set.

3. The method for determining a dynamic threshold based on point cloud density and spacing according to claim 2, characterized in that: The process of determining the point cloud density ρ in the point cloud set comprises the following steps: After obtaining the average Euclidean distance, the point cloud space is divided into cubic grids, that is, the space is voxelized. Assuming the side length of the grid is Δa×Δb×Δc, the point cloud space can be divided into s cubic grids; s=a×b×c Where a, b, c are the number of grid planes in each direction of the cube; After dividing the point cloud space into cubic grids, if there is a point cloud in the grid, the grid is recorded as 1, and the number of point clouds n is recorded i , if there is no point cloud in the grid, it is recorded as 0; the point cloud density in the point cloud set is: Where s(1) represents the number of grids marked as 1, n i Indicates the number of point clouds in the grid.

4. The method for determining a dynamic threshold based on point cloud density and spacing according to claim 3, characterized in that: The number of grid planes in each direction of the cube is as follows: Among them, int means round down; x max 、x min They are the maximum and minimum values ​​of the x-axis coordinates of the point cloud in the point cloud set, and the y max ,y min are the maximum and minimum values ​​of the y-axis coordinates of the point cloud in the point cloud set, and z max 、z min They are respectively the maximum and minimum values ​​of the z-axis coordinates of the point cloud in the point cloud set.

5. Dynamic threshold Euclidean clustering method based on point cloud density and spacing, characterized in that: First, the distance threshold is determined by using the dynamic threshold determination method based on point cloud density and spacing as described in one of claims 1 to 4. After the distance threshold is obtained, clustering is performed using the Euclidean clustering method based on the set clustering parameters of the Euclidean clustering.

6. The dynamic threshold Euclidean clustering method based on point cloud density and spacing according to claim 5, characterized in that: The process of clustering using the Euclidean clustering method comprises the following steps: S201, create a set Q1, use KD-tree to search for the nearest point outside o1, and find n points p1, p2..., p1 that meet the distance threshold. n And point p is added to the set Q1, then Q1=(o1,p1,p2…,p n ), where the points in set Q1 satisfy: where p i is any point in the set Q1 except o1, d i is the nearest search distance, S202, select another point outside the set Q1 as the center point, create another set Q2, and find points that meet the distance threshold in the same way as step S201 to join, traverse all the input points, and repeat the above steps until all points become a set Q i The point or center point in the set Q is recorded. i The number of point clouds in t i , When the quantity threshold condition is met, that is, t min <t i <t max When the set Q i The point cloud inside is the obstacle point cloud, where t min and t max is the clustering quantity threshold; the point cloud in the set that does not meet the quantity threshold condition is an invalid point cloud.

Images