An optimized point cloud segmentation method based on center homogenization clustering fusion

Abstract

The present invention discloses an optimized point cloud segmentation method with center homogenization and cluster fusion. First, the present invention constructs a relationship model between scanning distance, angular resolution and number of neighboring points, and gives a cutoff distance calculation formula. Secondly, the local density and distance function are normalized, and the cluster center is determined according to the size of the product of the two. At the same time, in order to avoid the influence of repeated or similar cluster centers on clustering, the present invention eliminates repeated or very close cluster centers, thereby achieving the homogenization of cluster centers. Finally, according to the cluster center, the cluster affiliation of each data is determined one by one by using the distance from the data point to the cluster center, and the fusion between neighboring clusters is performed based on the comparison between the distance between neighboring cluster data groups and the cutoff distance, thereby achieving fine segmentation of point cloud data. The present invention has obvious improvements in segmentation reliability and accuracy. It provides a preliminary data foundation for the construction of laser point cloud semantic models and semantic attribute annotation.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud segmentation, and in particular relates to an optimized center homogenization clustering fusion point cloud segmentation method. Background Art

[0002] Point cloud segmentation is the basis of target semantic labeling and attribute labeling. At present, the technology for point cloud segmentation mainly adopts clustering methods. Clustering methods are mainly divided into three categories according to different principles: clustering based on partitioning method, hierarchical clustering, and density-based clustering. Clustering based on partitioning method is to divide the data set D of n objects into k clusters, each partition represents a class, and k≤n. This method starts with the initial partition of the data set D and uses an iterative control strategy to optimize the objective function. The disadvantage of this method is that the quality of the sampled samples will directly affect the clustering quality. In view of this, Zheng improved the search characteristics in the combination algorithm of K-Means and K-modes, and proposed a global search strategy with evolutionary characteristics, thereby improving the clustering efficiency. However, the disadvantage of this method is that the clustering effect is greatly affected by the cluster center and is easy to fall into the local optimum. In view of this, Cheng Weiqing used the change of the error square sum SSE corresponding to different K values ​​to determine the optimal k value. Although this method can determine the K value more accurately, it still requires manual analysis and decision-making to determine the optimal k value. The hierarchical K-means clustering algorithm proposed by the author Hu Wei can adaptively obtain the optimal number of clusters k. The difference between K-medoids and K-means algorithms is that this method does not use the average value of the object as the reference point, but uses the center of the object as the reference point. The similarity is that the number of clusters needs to be determined in advance. Although the partition-based method has a high execution efficiency, it does not consider the data characteristics. In response to this, Zhang proposed a hierarchical clustering method. This algorithm clusters the data of different nodes through the initially established data layer set. This method is better than K-means and CLARANS in terms of stability, but the established data layer set is more suitable for spherical or convex data, while the clustering effect of other shapes is poor. The above methods are mainly clustered based on the distance between objects, without considering the influence of the target density and the similarity of surrounding objects on the clustering accuracy. Therefore, Fan proposed a density-based clustering method, which clusters according to the density distribution of the target data. DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a relatively classic density-based clustering algorithm. Unlike partitioning and hierarchical clustering methods, it defines clusters as the largest set of density-connected points, can divide areas with sufficiently high density into clusters, and can find clusters of any shape in noisy spatial databases, with certain noise resistance. The disadvantage of this method is that it is sensitive to the two parameters of domain size and the number of points in the domain. In the process of clustering using DBSCAN, various parameters need to be set, and it is very sensitive to the two parameters of Eps and Minpts.

[0003] The above clustering segmentation methods are mainly applicable to certain special cases, such as the BIRCH clustering method is more suitable for the segmentation of spherical data clusters, but not for the segmentation of irregular data clusters, k-medoids, DBSCAN clustering methods are too time-consuming, etc. At the same time, the above methods are more suitable for small-scale and low-dimensional data, but the clustering segmentation effect is poor for large-scale and high-dimensional data. In 2014, Rodriguez proposed a clustering segmentation algorithm (CFDP) based on fast search of density peaks. This method can quickly and accurately realize the clustering segmentation of two-dimensional data, but this method is only applicable to the segmentation of two-dimensional images, and there is the same local density in the data, which causes the repeatability of the determination of the cluster center, and ultimately leads to local erroneous segmentation. In view of this, the present invention improves this method, extends this method to point cloud segmentation, and improves the stability and accuracy of clustering by optimizing the judgment of cluster centers and the attribution of cluster attributes, thereby providing a stable point cloud segmentation result.

[0004] At present, point cloud segmentation still has the following shortcomings:

[0005] 1. It can only segment planes or curved surfaces, but cannot segment complete objects;

[0006] 2. The number of segments needs to be set before segmentation, and automatic object segmentation cannot be achieved;

[0007] 3. Currently, the reliability and speed of point cloud segmentation cannot be guaranteed. Summary of the invention

[0008] In order to solve the problems existing in the prior art, the present invention provides an optimized center homogenization clustering fusion point cloud segmentation method to improve the reliability and speed of point cloud segmentation.

[0009] The technical solution adopted by the present invention is: an optimized center homogenization clustering fusion point cloud segmentation method, characterized in that it includes the following steps:

[0010] Step 1: 3D laser scanning to obtain point cloud data;

[0011] Step 2: Use Gaussian kernel function to determine the local density of point cloud data;

[0012] Step 3: Improve the distance parameter δ i Determine the method;

[0013] Step 4: Determine the point cloud cluster center based on local density and distance parameters;

[0014] Step 5: Determine the initial clustering of the point cloud based on the distance from the data point to the cluster center;

[0015] Step 6: Initial cluster fusion is achieved by calculating the shortest distance between adjacent clusters, and point cloud segmentation is achieved based on the cluster fusion results.

[0016] As a preferred embodiment, in step 2, the point cloud dataset is The present invention uses Gaussian kernel function to determine the local density ρ i

[0017]

[0018] Where, d ij For point P i and point P j The distance between Cut-off distance

[0019] As a preference, the cut-off distance The specific calculation process is: the laser scanning distance is D, the number of points adjacent to the laser scanning center point is k, and the laser scanning angle resolution is α, then the cutoff distance for:

[0020]

[0021] As a preference, in step 3, the distance parameter δ is improved i The determination method is: when point P i The corresponding local density ρ i is the maximum local density, then the distance parameter δ i It's point P i and Point The maximum value of all distances in. i The corresponding local density ρ i is not the maximum local density, δ i It's point P j The local density is greater than ρ i The minimum distance is shown below.

[0022]

[0023] In the formula,

[0024] As a preference, in step 4, the specific method for determining the point cloud cluster center is as follows: first consider the local density ρ i and the distance function δ i If the scales are different, they are normalized to

[01] to obtain the normalized local density and distance function Secondly, calculate the normalized local density and distance function The multiplication result is shown in the following formula

[0025]

[0026] Finally, based on The corresponding point is the initial cluster center point, and the distance between adjacent cluster center points is calculated. Compare this distance to the cutoff distance Compare, if Adjacent cluster center points are merged to finally obtain evenly distributed cluster center points.

[0027] Preferably, in step 5, the method for determining the initial clustering of the point cloud is:

[0028] Point cloud data has m cluster centers In addition to the cluster center, there are other data points The coordinates of its data points are shown below.

[0029]

[0030] For non-central data points Non-central data points and other data points The distance between

[0031]

[0032] The cluster affiliation of the point is determined according to the cluster center point corresponding to the minimum distance, as shown in the following formula.

[0033]

[0034] if If the minimum distance requirement is met, then the point and Belonging to the same class, and so on, to complete the initial clustering of all data points.

[0035] Preferably, in step 6, point cloud segmentation is achieved based on the initial clustering fusion result:

[0036] Based on the initial clustering of the data set, the clustering of all points is obtained, as shown in the following formula

[0037]

[0038] Use KNNS (k-NearestNeighbor searching) to calculate the shortest distance between neighboring clusters And compare this distance with the cutoff distance The comparison is shown in the following formula.

[0039]

[0040] if Cluster points and If the two clusters belong to the same class, the clusters of the two clusters are merged into one cluster. Otherwise, there is no need to merge them, so that the clustering of the entire data set can be obtained, and then the point cloud data set can be segmented.

[0041] Compared with the prior art, the present invention has the following main advantages:

[0042] 1. The present invention provides a method for determining the cut-off distance

[0043] The key to clustering is to analyze the cluster affiliation of each point, and the basis for judging the cluster affiliation of each point is the distance from the non-center point to the center point. The point affiliation is judged based on the distance index, which has always been a difficult point. The present invention establishes a relationship model among scanning distance, angular resolution and number of neighboring points, and provides a method for determining the cutoff distance.

[0044] 2. The present invention provides a uniform clustering center

[0045] In the process of determining the cluster center, the present invention determines the initial cluster center by multiplying the local density based on the Gaussian kernel function and the improved distance parameter. After the initial cluster center is determined, the neighboring cluster centers are calculated and compared with the cutoff distance, so as to avoid the centralization of the cluster center, and finally obtain a uniform cluster center, such as Figure 1 shown.

[0046] 3. The present invention integrates the clustering of the initial data set

[0047] The present invention gives the clustering result of the initial data set based on the relationship between the distance from the non-center point to the cluster center point and the cutoff distance. After completing the clustering of the initial data set, the distance characteristics between adjacent clusters are calculated and compared with the cutoff distance. The initial clustering is fused based on the comparison result, and the accurate segmentation of the point cloud data is completed based on the cluster fusion result.

[0048] In summary, in the process of point cloud data segmentation based on clustering, the present invention overcomes the problem that the traditional clustering-based point cloud segmentation method cannot obtain the cutoff distance, and at the same time overcomes the problem of non-uniformity of the obtained cluster centers. After the initial data set is clustered and fused, it helps to improve the reliability of point cloud segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The cluster centers are determined by two different methods: the present invention and the traditional method;

[0050] Figure 2 is the relationship between the neighboring points of the center point o and the angular resolution and the scanning distance D during the implementation of the present invention;

[0051] Figure 3 are target point clouds of different resolutions scanned in an embodiment of the present invention;

[0052] Figure 4 It is to verify the segmentation effect of different methods on point clouds with different resolutions in the embodiments of the present invention;

[0053] Figure 5 It is to verify the segmentation of the point cloud of the apartment, bedroom, conference room, lobby, and attic in the embodiment of the present invention; DETAILED DESCRIPTION

[0054] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0055] In step 2, the local density of the point cloud data is determined using the Gaussian kernel function combined with the cutoff distance

[0056] The point cloud dataset is The disadvantage of the traditional method for determining the local density of point clouds is that there are overlapping local densities, that is, point P i Local density and point P m The local density may be the same. In other words, many different data points may have the same local density. To address this, the present invention proposes to use a Gaussian kernel function to determine the local density ρ of the point cloud. i .

[0057]

[0058] Where, d ij For point P i and point P j The distance between is the cut-off distance.

[0059] Cut-off distance It can be obtained based on the relationship between the laser scanning angle resolution α, the scanning distance D, and the number of adjacent points k. Figure 2 The relationship between the neighboring points of a certain center point o and the angular resolution and scanning distance D is given.

[0060] Assuming that the number of points on each scan line in the vicinity of the center point is m, the relationship model between the number of neighboring points k and the number of scan line points m is:

[0061] k=m 2 -1 (2)

[0062] according to Figure 2 The cut-off distance can be obtained as shown in The relationship between m, α and D is shown in formula (3):

[0063]

[0064] According to equations (2) and (3), the cutoff distance can be obtained as:

[0065]

[0066] Step 3: Optimize the distance parameter δ based on the local density of different points i Determine the method;

[0067] The distance parameter δ can be determined according to the size of the local density. i , when point P i Corresponding to the maximum local density, the distance parameter δ i It's point P i and Point All distances d in ij The maximum value of i It does not correspond to the maximum local density, δ i It's point P j The local density is greater than The minimum distance is shown in formula (5).

[0068]

[0069] In the formula, The different points thus obtained may have the same maximum local density, that is, point P i and P m have the same maximum local density, as shown in formula (6).

[0070]

[0071] In order to eliminate the influence of the same maximum local density, first, the distance d between the locals with the same maximum local density is calculated. im , compare the distance d im and cutoff distance As shown in formula (7).

[0072]

[0073] If Δd≤0, then point P i and P m Belong to the same class, the local density of the two only needs to determine and retain one, otherwise, point P i and P mBelonging to two different classes, the local densities of the two are still two.

[0074] Step 4: Determine the uniform point cloud cluster center based on local density and distance parameters Consider local density ρ i and the distance function δ i If the scales are different, they are normalized to between

[01] , as shown in formula (8):

[0075]

[0076] From formula (8), we can get

[0077]

[0078] According to formula (9) The size of can determine the initial cluster center, that is, Corresponding cluster centers. Assuming that the initial cluster centers are determined and Calculate the distance between the initial cluster centers if The two clusters belong to the same cluster and are therefore merged together, otherwise they are two different clusters. The cluster centers are determined by their size, such as Figure 1 shown.

[0079] Step 5: Determine the clustering of each point in the point cloud based on the distance from the data point to the cluster center; based on the m cluster centers determined in step 4 In addition to the cluster center, there are other data points The coordinates of its data points are shown in formula (10).

[0080]

[0081] For non-central data points Calculate non-central data points and other data points The distance between

[0082]

[0083] The cluster affiliation of the point is determined according to the cluster center point corresponding to the minimum distance, as shown in formula (12).

[0084]

[0085] According to formula (12), point Hedianji The distance between for

[0086]

[0087] For formula (13), if If the minimum distance requirement of formula (12) is satisfied, then point and Belonging to the same class, and so on, the initial clustering of all point cloud data points is completed.

[0088] Step 6: Initial cluster fusion is achieved by calculating the shortest distance between adjacent clusters, and accurate point cloud segmentation is achieved based on the cluster fusion results.

[0089] Based on the initial clustering of the data set, the clustering of all points is obtained, as shown in formula (14):

[0090]

[0091] Use KNNS (k-NearestNeighbor searching) to calculate the shortest distance between neighboring clusters And compare this distance with the cutoff distance For comparison, as shown in formula (15).

[0092]

[0093] if Cluster points and If the two clusters belong to the same class, the two clusters are merged into one cluster. Otherwise, there is no need to merge them, so that the segmentation of the entire point cloud dataset can be obtained.

[0094] In order to verify the effectiveness of the present invention, a 3D laser scanner is used to collect point clouds with resolutions of 0.01m, 0.02m and 0.03m, respectively. Figure 3 shown.

[0095] Using the present invention Figure 3 Point clouds with different resolutions are segmented and the segmentation results are compared with the clustering results of CFDP, K-means and DBSCAN, such as Figure 4 shown.

[0096] Depend on Figure 4 It can be seen that under different point cloud resolution conditions, the present invention can segment almost every target. From the segmentation results of different resolutions in the figure, it can be seen that the point cloud clustering segmentation effect of the present invention under these three distance resolutions is the same. Figure 4It can be seen from the white rectangle that when the distance resolution is 0.02, the CFDP method divides the shoe rack point cloud into two areas, while when the distance resolution is 0.01 and 0.03, the shoe rack point cloud is accurately segmented as a whole. Figure 4 The white ellipse area shows the effect of distance resolution on k-means clustering. For the DBSCAN method, although the clustering segmentation effects of point clouds with different resolutions are basically the same, none of them can achieve accurate segmentation of each target. Figure 4 (c) shows that the target segmentation results of the CFDP method are basically consistent with those of the present invention, and the segmentation effect is better than that of the other two methods. Figure 4 (b) It can be seen that the segmentation effect of the CFDP method is similar to that of the K-means method, but many objects are not segmented by the DBSCAN method. Figure 4 The rounded rectangle on the left shows that the performance of the present invention in the field of large target point cloud segmentation is better than the other three methods. Therefore, the above clustering-based point cloud segmentation example analysis shows that the present invention has a relatively stable application value and can accurately segment point clouds of different resolutions.

[0097] The segmentation effects of different clusters are comprehensively evaluated by calculating the accuracy, as shown in formula (16):

[0098]

[0099] In the formula, k is the number of data set classes, m is i is the correct number of divisions, n i is the number of samples in the i-th class, m i / n i is the accuracy in the i-th class.

[0100] According to formula (16), the segmentation accuracy of different methods for different resolution point clouds is calculated, as shown in Table 1.

[0101] Table 1 Segmentation accuracy of different methods for point clouds with different resolutions

[0102]

[0103] It can be clearly seen from Table 1 that the segmentation accuracy of the present invention and the DBSCAN method for point clouds with different resolutions is the same, and the segmentation accuracy of the two methods is not affected by the resolution of the point cloud. However, it is obvious that the segmentation accuracy of the present invention is very high, while the segmentation accuracy of the DBSCAN method is much lower than that of the present invention, and it is impossible to accurately segment point clouds with different resolutions. The CFDP method is greatly affected by the resolution of the point cloud. When the resolution is 0.03m, the segmentation accuracy of this method is basically the same as that of the present invention, and when the resolution is 0.01m, the segmentation accuracy becomes very low; while the K-means method has little difference in the segmentation accuracy of point clouds with different resolutions, and the segmentation accuracy is the highest when the resolution is 0.02m. It is determined that the point cloud segmentation of the present invention is more stable and accurate, and the segmentation effect is significantly better than the other three methods.

[0104] In order to illustrate that the present invention is still useful for segmenting more complex point cloud scenes, the FARO 3D X330 three-dimensional laser scanner was used to collect point cloud data of an apartment, bedroom, conference room, lobby, and loft. The scanner has a scanning distance of 0.6 to 330 meters and a scanning accuracy of 0.1 mm within 10 meters. Each scene is scanned from multiple stations and finally spliced ​​together. The present invention is used to segment different point cloud data, and the segmentation results are shown in the figure. Figure 5 shown.

[0105] Depend on Figure 5 From the segmentation, we can see that the bedroom point cloud segmentation results mainly include the bed, chair, table and lamp, and they are displayed in different point cloud colors. From the overall segmentation results, we can see that almost all objects are segmented. Figure 5 The present invention can separate objects close to the wall from the wall, such as Figure 5 On the other hand, the present invention not only segments the objects in the room, but also segments the lights hanging from the ceiling, such as Figure 5 The present invention separates the sofa, table and chairs and displays them in different colors. For example, Figure 5 The dotted black box shows that the pure yellow color represents the sofa. Figure 5 The dotted hexagon shows that the method successfully segmented the flowers. The segmentation results of the lobby show that the present invention can segment about 50 objects. Almost all objects in the room, such as the ceiling, walls, and floor, are segmented. Only some objects that are close to each other cannot be segmented, such as chairs and tables. At the same time, the chandelier on the top of the lobby is also segmented from the top. Figure 5 as shown in the circle.

[0106] Based on the above analysis, it can be clearly found that the present invention has a better effect on the segmentation of point cloud objects in more complex and different scenes.

[0107] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0108] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. An optimized center homogenization clustering fusion point cloud segmentation method, characterized in that: The following steps are involved: Step 1: 3D laser scanning to obtain point cloud data; Step 2: Use Gaussian kernel function to determine the local density of point cloud data; Step 3: Improve the distance parameter δ i Determine the method; Step 4: Determine the point cloud cluster center based on local density and distance parameters; Step 5: Determine the initial clustering of the point cloud based on the distance from the data point to the cluster center; Step 6: Initial cluster fusion is achieved by calculating the shortest distance between adjacent clusters, and point cloud segmentation is achieved based on the cluster fusion results; Cut-off distance The specific calculation process is: the laser scanning distance is D, the number of points adjacent to the laser scanning center is k, and the laser scanning angle resolution is α, then the cutoff distance for: The specific method for determining the point cloud cluster center is: first consider the local density ρ i and the distance function δ i If the scales are different, they are normalized to [0 1] to obtain the normalized local density. and distance function Secondly, calculate the normalized local density and distance function The multiplication result is shown in the following formula Finally, based on The corresponding point is the initial cluster center point, and the distance between adjacent cluster center points is calculated. Compare this distance to the cutoff distance Compare, if Then the adjacent cluster center points are merged to finally obtain evenly distributed cluster center points; Point cloud segmentation is achieved based on the initial clustering fusion results: Based on the initial clustering of the data set, the clustering of all points is obtained, as shown in the following formula Use KNNS (k-Nearest Neighbor searching) to calculate the shortest distance between neighboring clusters And compare this distance with the cutoff distance For comparison, as shown below, if Cluster points and If the two clusters belong to the same class, the clusters of the two clusters are merged into one cluster; otherwise, there is no need to merge them, so that the clustering of the entire data set can be obtained, and then the point cloud data set can be segmented.

2. The optimized center homogenization clustering fusion point cloud segmentation method according to claim 1, characterized in that: In step 2, the point cloud dataset is The present invention uses Gaussian kernel function to determine the local density ρ i Where, d ij Point P i and point P j The distance between is the cut-off distance.

3. The optimized center homogenization clustering fusion point cloud segmentation method according to claim 1, characterized in that: In step 3, the distance parameter δ is improved i The determination method is: when point P i The corresponding local density ρ i is the maximum local density, then the distance parameter δ i It's point P i and Point The maximum value of all distances in; when point P i The corresponding local density ρ i is not the maximum local density, δ i It's point P j The local density is greater than ρ i The minimum distance is as follows: In the formula, 4. The optimized center homogenization clustering fusion point cloud segmentation method according to claim 1, characterized in that: In step 5, the method for determining the initial clustering of the point cloud is: Point cloud data has m cluster centers In addition to the cluster center, there are other data points The coordinates of its data points are shown below: For non-central data points Non-central data points and other data points The distance between The cluster affiliation of the point is determined according to the cluster center point corresponding to the minimum distance, as shown in the following formula: if If the minimum distance requirement is met, then the point and Belonging to the same class, and so on, to complete the initial clustering of all data points.

Images