A three-dimensional point cloud clustering method, system and computer program product

Abstract

The present invention discloses a point cloud clustering method, system and computer program product in three-dimensional space, the method comprising: obtaining point cloud data of the surrounding environment space, and dividing the three-dimensional space into cubic grids; removing the cubic grids that do not contain objects according to preset noise filtering conditions; removing the cubic grids that contain the ground based on the height threshold and the ground plane fitting method; extracting features from the point cloud data of the cubic grid based on the graph convolution method to obtain feature vectors; clustering the feature vectors based on the distance threshold and the hierarchical clustering method to obtain clustering results. The technical solution provided by the present invention divides the three-dimensional space into cubic grids, filters out noise and the ground, extracts feature vectors and clusters; uses progressive ground fitting to improve the fitting accuracy of the ground plane; uses the graph convolution method to extract features from the point cloud data, can obtain the local geometric features of the point cloud data, and improve the accuracy of clustering.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud processing, and in particular to a point cloud clustering method, system and computer program product in three-dimensional space. Background Art

[0002] During the operation of a robot or self-driving vehicle, its built-in detection device will collect point cloud data of the surrounding space, and then the point cloud data needs to be detected and analyzed to identify obstacles. If each point in the point cloud data is directly regarded as an obstacle point, it will be seriously interfered by noise, ground, etc., and it will require a huge amount of calculation and take a lot of time. However, if the point cloud is clustered into circular, linear, and polygonal obstacles, the calculation efficiency and recognition accuracy will be improved. In addition, the robot or self-driving vehicle also needs to obtain the specific location and size of the obstacle based on the point cloud clustering results to determine the subsequent path planning, obstacle avoidance and other methods and strategies. Summary of the invention

[0003] The present invention provides a point cloud clustering method in three-dimensional space, which realizes clustering of point cloud data by dividing the three-dimensional space corresponding to the surrounding environment space into cubic grids, filtering out noise and the ground, extracting feature vectors and clustering the point cloud data in the cubic grids.

[0004] To achieve the above object, the embodiments of the present invention adopt the following technical solutions: A point cloud clustering method in three-dimensional space comprises the following steps: Obtain point cloud data of the surrounding environment space, and divide the three-dimensional space corresponding to the surrounding space into cubic grids; According to the preset noise filtering conditions, remove the cubic grid that does not contain objects; Based on the method of height threshold and ground plane fitting, the cube grid containing the ground is removed; Based on the graph convolution method, feature extraction is performed on the point cloud data of the environment space corresponding to the remaining cubic grid to obtain a feature vector; Based on the distance threshold and hierarchical clustering method, the feature vectors are clustered to obtain the clustering data set corresponding to the point cloud data.

[0005] According to one aspect of the present invention, the three-dimensional point cloud clustering method further includes: Downsample the point cloud data of the surrounding space.

[0006] According to one aspect of the present invention, dividing the three-dimensional space corresponding to the environment space into cubic grids comprises: The three-dimensional space can be segmented into different sizes according to the actual situation of the surrounding space and the distribution of the point cloud.

[0007] According to one aspect of the present invention, removing the cubic grid that does not contain the object according to the preset noise filtering condition includes: If the number of point clouds or the point cloud density of the environment space corresponding to the cube grid is lower than a preset threshold, it is considered that it does not contain an object.

[0008] According to one aspect of the present invention, the method based on height threshold and ground plane fitting to remove the cubic grid containing the ground includes: According to the height of the point cloud detection device, a height threshold is set; According to the height threshold, a set of point clouds are selected as the initial ground points; Select point cloud data within a certain height, angle and plane distance from the initial ground point, and use the least squares method to fit the ground plane of the initial segment; Repeatedly select point cloud data within a certain height, angle and plane distance from the fitted ground plane, and use the least squares method to progressively fit all ground planes; If the point cloud of the environment space corresponding to the cube grid is within a certain height threshold from the ground plane, it is considered to contain the ground.

[0009] According to one aspect of the present invention, the step of extracting features from the point cloud data of the environment space corresponding to the remaining cubic grid based on the graph convolution method to obtain the feature vector includes: The point cloud data of the environment space corresponding to the cube grid is represented as a graph structure; The graph convolution method is used to extract features from the graph structure and obtain feature vectors.

[0010] According to one aspect of the present invention, clustering the feature vectors based on the distance threshold and the hierarchical clustering method to obtain a clustering data set of corresponding point cloud data includes: According to the adjacent relationship of the cube grid, the cube grid is segmented; Based on the hierarchical clustering method, the feature vectors of adjacent cube grids are clustered and layered according to the preset distance threshold to obtain the clustering data set of the corresponding point cloud data.

[0011] According to one aspect of the present invention, the three-dimensional point cloud clustering method further includes: Obstacle location and size data are obtained based on the clustered data set of point cloud data.

[0012] A three-dimensional point cloud clustering system, based on the three-dimensional point cloud clustering method as described above, comprises: The acquisition module is used to obtain the point cloud data of the surrounding environment space and divide the three-dimensional space corresponding to the environment space into cubic grids; A denoising module, used to remove the cubic grids that do not contain objects according to a preset noise filtering condition; A ground removal module is used to remove the cube grid containing the ground based on a height threshold and ground plane fitting method; A feature extraction module is used to extract features from the point cloud data of the environment space corresponding to the remaining cubic grid based on the graph convolution method to obtain a feature vector; The clustering module is used to cluster the feature vectors based on the distance threshold and hierarchical clustering method to obtain the clustering data set corresponding to the point cloud data.

[0013] A computer program product comprises a computer program, wherein when the computer program is executed, the steps of the three-dimensional point cloud clustering method described above are implemented.

[0014] Advantages of the present invention: The present invention provides a three-dimensional point cloud clustering method, which realizes the clustering of point cloud data by dividing the three-dimensional space corresponding to the surrounding environment space into cubic grids, filtering out noise and the ground, extracting feature vectors and clustering the point cloud data in the cubic grid. The method downsamples the point cloud data to reduce the amount of calculation; uses progressive ground fitting to improve the fitting accuracy of the ground plane; uses a graph convolution method to extract features from the point cloud data, which can obtain local geometric features of the point cloud data and improve the accuracy of clustering; compared with traditional point cloud data processing methods, it can reduce the amount of point cloud data for clustering, improve processing efficiency, and reduce the amount of calculation and processing time. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 This is a flow chart of a point cloud clustering method in three-dimensional space according to the first embodiment of the present invention; Figure 2 This is a flow chart of a point cloud clustering method in three-dimensional space described in Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Embodiment 1

[0019] like Figure 1 As shown, a point cloud clustering method in three-dimensional space includes the following steps: S1: Obtain point cloud data of the surrounding environment space, and divide the three-dimensional space corresponding to the surrounding space into cubic grids.

[0020] In actual applications, the detection devices of robots or autonomous vehicles, usually laser scanners, cameras, 3D scanners, etc., will collect point cloud data of the surrounding space environment. Point cloud data records information about the surrounding environment in the form of points. Each point contains 3D coordinates (X, Y, Z) and can carry other information about the attributes of the point, such as color (R, G, B), reflectivity, intensity, etc.

[0021] Preferably, the three-dimensional point cloud clustering method further includes: Downsample the point cloud data of the surrounding space.

[0022] In practical applications, the amount of point cloud data obtained may be large. Downsampling can reduce the number of points, reduce data complexity, and improve processing effects. Specifically, uniform sampling, voxel sampling, sampling based on geometric features, etc. can be adopted.

[0023] The step of dividing the three-dimensional space corresponding to the environment space into cubic grids comprises: The three-dimensional space can be segmented into different sizes according to the actual situation of the surrounding space and the distribution of the point cloud.

[0024] When segmenting three-dimensional space, this method is different from the usual voxelization method. This method can be divided into cuboids of different sizes according to the characteristics of the spatial environment and the distribution of the point cloud, without being limited to cubes of the same size. For example, for point cloud data distributed in a carpet-like manner on the ground, when setting the size of the segmented cuboid, the length and width values ​​can be set much larger than the height. This segmentation method can improve the processing efficiency of subsequent point cloud data and reduce the amount of calculation and processing time.

[0025] S2: According to the preset noise filtering conditions, remove the cubic grid that does not contain the object.

[0026] Specifically, if the number of point clouds or the point cloud density of the environment space corresponding to the cubic grid is lower than a preset threshold, it is considered that it does not contain an object.

[0027] In addition, it is also possible to determine whether a point is a noise point based on the reflectivity, intensity and other information of the point in the point cloud data.

[0028] S3: Based on the height threshold and ground plane fitting method, the cube grid containing the ground is removed.

[0029] Specifically, step S3 includes: S31: Setting a height threshold according to the height of the point cloud detection device.

[0030] The amount of point cloud data is large. If the ground plane is directly fitted, there will be many unnecessary calculations and errors. The coordinate position of the ground point cloud can be estimated according to the height of the detection device of the robot or autonomous driving vehicle, and a certain height threshold can be set. Within a certain range of this height threshold, it is the ground in front of the robot or autonomous driving vehicle.

[0031] S32: According to the height threshold, a group of point clouds are selected as initial ground points.

[0032] A set of point clouds in front of the robot or autonomous driving vehicle within a certain range of the height threshold are selected as the initial ground points.

[0033] S33: Select point cloud data within a certain height, angle and plane distance from the initial ground point, and use the least squares method to fit the ground plane of the initial segment.

[0034] Ground points are usually concentrated in a certain range of the z-axis, and the angle between the ground point cloud and the horizontal plane is small, while the non-ground point cloud may have a larger angle. Since the ground may have slope changes, if only one fitting is performed, there may be a large deviation, so a segmented and progressive fitting can be adopted to improve the fitting accuracy. Therefore, the range of height, angle and plane distance can be set, and the point cloud data within a certain height, angle and plane distance from the initial ground point is first selected, and the ground plane of the initial segment is fitted using the least squares method.

[0035] The specific method of fitting the ground plane by the least squares method is: construct the plane equation and error function of the ground, and by minimizing the error function, a linear equation system can be obtained, and the solution of the equation system is the parameters of the fitting plane.

[0036] S34: repeatedly selecting point cloud data within a certain height, angle and plane distance from the fitted ground plane, and progressively fitting all ground planes using the least squares method.

[0037] After fitting the ground plane of the initial segment in step S33, the point cloud data of the next segment of the ground can be selected for fitting according to the height of the terminal ground of the initial segment to obtain the ground plane of the next segment. The selection and fitting operations are repeated to progressively fit all the ground planes.

[0038] S35: If the point cloud of the environment space corresponding to the cubic grid is within a certain height threshold from the ground plane, it is considered to contain the ground.

[0039] In practical applications, the collected point cloud data will have certain errors. A height threshold is set. If the distance between the point cloud and the ground plane is within the height threshold range, this part of the point cloud is considered to be the point cloud of the ground.

[0040] S4: Based on the graph convolution method, feature extraction is performed on the point cloud data of the environment space corresponding to the remaining cubic grid to obtain a feature vector.

[0041] Specifically, step S4 includes: S41: Represent the point cloud data of the environment space corresponding to the cubic grid as a graph structure.

[0042] A graph structure is generated based on the spatial position relationship of the point cloud data, where each point is a node of the graph and the edges between nodes are determined based on the distance or similarity between the points.

[0043] S42: Use the graph convolution method to extract features from the graph structure and obtain a feature vector.

[0044] In practical applications, graph convolution methods can be used to perform convolution operations on graph structures to extract local geometric features in point clouds. The specific approach is: for any two points in the point cloud, define an edge feature, which is usually calculated by a nonlinear function. This function can contain learnable parameters to capture the complex relationship between points; aggregate the edge features of all neighboring points of any point to obtain a new feature representation. This aggregation operation can be summation, maximum pooling, etc.; use the aggregated feature representation as the feature encoding of the point; construct a feature extraction network, such as a one-dimensional convolutional network FENet, a fully connected network, an attention mechanism layer, etc., to extract higher-level features from the feature encoding of each point in the cubic grid, and finally obtain the feature vector of the cubic grid.

[0045] In practical applications, it is necessary to use labeled point cloud datasets to train edge feature aggregation and feature extraction networks to learn the mapping relationship from point cloud data to feature vectors.

[0046] S5: Based on the distance threshold and hierarchical clustering method, the feature vectors are clustered to obtain the clustering data set corresponding to the point cloud data.

[0047] Specifically, step S5 includes: S51: Segment the cubic grid according to the adjacent relationship of the cubic grid.

[0048] After steps S2 and S3, the cube grids that do not contain objects and contain the ground are removed. If the remaining cube grids are not adjacent, the point cloud data therein do not belong to the same cluster. Therefore, only the adjacent cube grids need to be clustered, which can reduce redundant calculations.

[0049] S52: Based on a hierarchical clustering method, according to a preset distance threshold, the feature vectors of adjacent cubic grids are clustered and layered to obtain a clustering data set of corresponding point cloud data.

[0050] Adjacent cubic grids may or may not belong to the same cluster, and a split hierarchical clustering method can be used to achieve accurate clustering stratification. According to the characteristics of the feature vector obtained in step S4, an appropriate distance threshold is set, and adjacent cubic grids are recursively split until the set distance threshold is reached.

[0051] Through the clustering results of the cube grid, the clustered dataset of the point cloud data can be obtained.

[0052] The beneficial effects of this embodiment are as follows: this method realizes clustering of point cloud data by dividing the three-dimensional space corresponding to the environmental space into cubic grids, filtering out noise and the ground, extracting feature vectors and clustering the point cloud data in the cubic grid. This method can reduce the amount of calculation by downsampling the point cloud data; using progressive ground fitting, it can improve the fitting accuracy of the ground plane; using the graph convolution method to extract features from point cloud data, it can obtain the local geometric features of the point cloud data and improve the accuracy of clustering. Compared with traditional point cloud data processing methods, this method can reduce the amount of point cloud data for clustering, improve processing efficiency, and reduce the amount of calculation and processing time.

[0053] Embodiment 2

[0054] like Figure 2 As shown, a point cloud clustering method in three-dimensional space includes the following steps: S1: Obtain point cloud data of the surrounding environment space, and divide the three-dimensional space corresponding to the surrounding space into cubic grids.

[0055] In actual applications, the detection devices of robots or autonomous vehicles, usually laser scanners, cameras, 3D scanners, etc., will collect point cloud data of the surrounding space environment. Point cloud data records information about the surrounding environment in the form of points. Each point contains 3D coordinates (X, Y, Z) and can carry other information about the attributes of the point, such as color (R, G, B), reflectivity, intensity, etc.

[0056] Preferably, the three-dimensional point cloud clustering method further includes: Downsample the point cloud data of the surrounding space.

[0057] In practical applications, the amount of point cloud data obtained may be large. Downsampling can reduce the number of points, reduce data complexity, and improve processing effects. Specifically, uniform sampling, voxel sampling, sampling based on geometric features, etc. can be adopted.

[0058] The step of dividing the three-dimensional space corresponding to the environment space into cubic grids comprises: The three-dimensional space can be segmented into different sizes according to the actual situation of the surrounding space and the distribution of the point cloud.

[0059] When segmenting three-dimensional space, this method is different from the usual voxelization method. This method can be divided into cuboids of different sizes according to the characteristics of the spatial environment and the distribution of the point cloud, without being limited to cubes of the same size. For example, for point cloud data distributed in a carpet-like manner on the ground, when setting the size of the segmented cuboid, the length and width values ​​can be set much larger than the height. This segmentation method can improve the processing efficiency of subsequent point cloud data and reduce the amount of calculation and processing time.

[0060] S2: According to the preset noise filtering conditions, remove the cubic grid that does not contain the object.

[0061] Specifically, if the number of point clouds or the point cloud density of the environment space corresponding to the cubic grid is lower than a preset threshold, it is considered that it does not contain an object.

[0062] In addition, it is also possible to determine whether a point is a noise point based on the reflectivity, intensity and other information of the point in the point cloud data.

[0063] S3: Based on the height threshold and ground plane fitting method, the cube grid containing the ground is removed.

[0064] Specifically, step S3 includes: S31: Setting a height threshold according to the height of the point cloud detection device.

[0065] The amount of point cloud data is large. If the ground plane is directly fitted, there will be many unnecessary calculations and errors. The coordinate position of the ground point cloud can be estimated according to the height of the detection device of the robot or autonomous driving vehicle, and a certain height threshold can be set. Within a certain range of this height threshold, it is the ground in front of the robot or autonomous driving vehicle.

[0066] S32: According to the height threshold, a group of point clouds are selected as initial ground points.

[0067] A set of point clouds in front of the robot or autonomous driving vehicle within a certain range of the height threshold are selected as the initial ground points.

[0068] S33: Select point cloud data within a certain height, angle and plane distance from the initial ground point, and use the least squares method to fit the ground plane of the initial segment.

[0069] Ground points are usually concentrated in a certain range of the z-axis, and the angle between the ground point cloud and the horizontal plane is small, while the non-ground point cloud may have a larger angle. Since the ground may have slope changes, if only one fitting is performed, there may be a large deviation, so a segmented and progressive fitting can be adopted to improve the fitting accuracy. Therefore, the range of height, angle and plane distance can be set, and the point cloud data within a certain height, angle and plane distance from the initial ground point is first selected, and the ground plane of the initial segment is fitted using the least squares method.

[0070] The specific method of fitting the ground plane by the least squares method is: construct the plane equation and error function of the ground, and by minimizing the error function, a linear equation system can be obtained, and the solution of the equation system is the parameters of the fitting plane.

[0071] S34: repeatedly selecting point cloud data within a certain height, angle and plane distance from the fitted ground plane, and progressively fitting all ground planes using the least squares method.

[0072] After fitting the ground plane of the initial segment in step S33, the point cloud data of the next segment of the ground can be selected for fitting according to the height of the terminal ground of the initial segment to obtain the ground plane of the next segment. The selection and fitting operations are repeated to progressively fit all the ground planes.

[0073] S35: If the point cloud of the environment space corresponding to the cubic grid is within a certain height threshold from the ground plane, it is considered to contain the ground.

[0074] In practical applications, the collected point cloud data will have certain errors. A height threshold is set. If the distance between the point cloud and the ground plane is within the height threshold range, this part of the point cloud is considered to be the point cloud of the ground.

[0075] S4: Based on the graph convolution method, feature extraction is performed on the point cloud data of the environment space corresponding to the remaining cubic grid to obtain a feature vector.

[0076] Specifically, step S4 includes: S41: Represent the point cloud data of the environment space corresponding to the cubic grid as a graph structure.

[0077] A graph structure is generated based on the spatial position relationship of the point cloud data, where each point is a node of the graph and the edges between nodes are determined based on the distance or similarity between the points.

[0078] S42: Use the graph convolution method to extract features from the graph structure and obtain a feature vector.

[0079] In practical applications, graph convolution methods can be used to perform convolution operations on graph structures to extract local geometric features in point clouds. The specific approach is: for any two points in the point cloud, define an edge feature, which is usually calculated by a nonlinear function. This function can contain learnable parameters to capture the complex relationship between points; aggregate the edge features of all neighboring points of any point to obtain a new feature representation. This aggregation operation can be summation, maximum pooling, etc.; use the aggregated feature representation as the feature encoding of the point; construct a feature extraction network, such as a one-dimensional convolutional network FENet, a fully connected network, an attention mechanism layer, etc., to extract higher-level features from the feature encoding of each point in the cubic grid, and finally obtain the feature vector of the cubic grid.

[0080] In practical applications, it is necessary to use labeled point cloud datasets to train edge feature aggregation and feature extraction networks to learn the mapping relationship from point cloud data to feature vectors.

[0081] S5: Based on the distance threshold and hierarchical clustering method, the feature vectors are clustered to obtain the clustering data set corresponding to the point cloud data.

[0082] Specifically, step S5 includes: S51: Segment the cubic grid according to the adjacent relationship of the cubic grid.

[0083] After steps S2 and S3, the cube grids that do not contain objects and contain the ground are removed. If the remaining cube grids are not adjacent, the point cloud data therein do not belong to the same cluster. Therefore, only the adjacent cube grids need to be clustered, which can reduce redundant calculations.

[0084] S52: Based on a hierarchical clustering method, according to a preset distance threshold, the feature vectors of adjacent cubic grids are clustered and layered to obtain a clustering data set of corresponding point cloud data.

[0085] Adjacent cubic grids may or may not belong to the same cluster, and a split hierarchical clustering method can be used to achieve accurate clustering stratification. According to the characteristics of the feature vector obtained in step S4, an appropriate distance threshold is set, and adjacent cubic grids are recursively split until the set distance threshold is reached.

[0086] In order to improve the accuracy of hierarchical clustering, you can also set multiple distance thresholds and perform cluster segmentation in various ways. Then compare the accuracy of cluster segmentation in different ways and select a more suitable distance threshold.

[0087] Through the clustering results of the cube grid, the clustered dataset of the point cloud data can be obtained.

[0088] S6: Obtain obstacle location and size data based on the clustered data set of the point cloud data.

[0089] After point cloud clustering is completed, point cloud data belonging to the same class belong to the same obstacle. The complete obstacle can be fitted based on the point cloud data, such as smoothing, bounding box, minimum convex hull method, etc., to obtain the position coordinates, size data, etc. of the obstacle.

[0090] The beneficial effects of this embodiment are as follows: this method can also perform cluster segmentation in various ways to improve clustering accuracy; based on the point cloud clustering results, the obstacle position and size data are obtained, which can facilitate subsequent path planning, forward operation, obstacle crossing, etc. of the robot or autonomous driving vehicle.

[0091] Embodiment 3

[0092] A three-dimensional point cloud clustering system, based on the three-dimensional point cloud clustering method as described in Embodiment 1 or 2, comprises: The acquisition module is used to obtain the point cloud data of the surrounding environment space and divide the three-dimensional space corresponding to the environment space into cubic grids; A denoising module, used to remove the cubic grids that do not contain objects according to a preset noise filtering condition; A ground removal module is used to remove the cube grid containing the ground based on a height threshold and ground plane fitting method; A feature extraction module is used to extract features from the point cloud data of the environment space corresponding to the remaining cubic grid based on the graph convolution method to obtain a feature vector; The clustering module is used to cluster the feature vectors of the cube grid based on the distance threshold and hierarchical clustering method to obtain the clustering data set of the corresponding point cloud data.

[0093] Preferably, the system further comprises: The sampling module is used to downsample the point cloud data of the surrounding environment space.

[0094] The obstacle acquisition module is used to obtain obstacle position and size data based on the clustered data set of point cloud data.

[0095] Embodiment 4

[0096] A computer program product comprises a computer program, wherein when the computer program is executed, the steps of the point cloud clustering method in three-dimensional space as described in the first or second embodiment are implemented.

[0097] Embodiment 5

[0098] A readable storage medium stores a computer program as described in Example 4, and when the computer program is executed, the steps of the three-dimensional point cloud clustering method as described in Example 1 or 2 are implemented.

[0099] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the art within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A point cloud clustering method in three-dimensional space, characterized in that: The following steps are involved: Obtain point cloud data of the surrounding environment space, and divide the three-dimensional space corresponding to the surrounding space into cubic grids; According to the preset noise filtering conditions, remove the cubic grid that does not contain objects; Based on the method of height threshold and ground plane fitting, the cube grid containing the ground is removed; Based on the graph convolution method, feature extraction is performed on the point cloud data of the environment space corresponding to the remaining cubic grid to obtain a feature vector; Based on the distance threshold and hierarchical clustering method, the feature vectors are clustered to obtain the clustering data set corresponding to the point cloud data.

2. The three-dimensional point cloud clustering method according to claim 1, characterized in that: The point cloud clustering method in the three-dimensional space also includes: Downsample the point cloud data of the surrounding space.

3. The point cloud clustering method in three-dimensional space according to claim 1, characterized in that: The step of dividing the three-dimensional space corresponding to the environment space into cubic grids comprises: The three-dimensional space can be segmented into different sizes according to the actual situation of the surrounding space and the distribution of the point cloud.

4. The point cloud clustering method in three-dimensional space according to claim 1, characterized in that: The step of removing the cubic grid that does not contain the object according to the preset noise filtering condition includes: If the number of point clouds or the point cloud density of the environment space corresponding to the cube grid is lower than a preset threshold, it is considered that it does not contain an object.

5. The three-dimensional point cloud clustering method according to claim 1, characterized in that: The method based on height threshold and ground plane fitting to remove the cubic grid containing the ground includes: According to the height of the point cloud detection device, a height threshold is set; According to the height threshold, a set of point clouds are selected as the initial ground points; Select point cloud data within a certain height, angle and plane distance from the initial ground point, and use the least squares method to fit the ground plane of the initial segment; Repeatedly select point cloud data within a certain height, angle and plane distance from the fitted ground plane, and use the least squares method to progressively fit all ground planes; If the point cloud of the environment space corresponding to the cube grid is within a certain height threshold from the ground plane, it is considered to contain the ground.

6. The point cloud clustering method in three-dimensional space according to claim 1, characterized in that: The graph convolution method is used to extract features from the point cloud data of the environment space corresponding to the remaining cubic grid to obtain a feature vector, including: The point cloud data of the environment space corresponding to the cube grid is represented as a graph structure; The graph convolution method is used to extract features from the graph structure and obtain feature vectors.

7. The three-dimensional point cloud clustering method according to claim 1, characterized in that: The method of clustering the feature vectors based on the distance threshold and the hierarchical clustering method to obtain the clustering data set of the corresponding point cloud data includes: According to the adjacent relationship of the cube grid, the cube grid is segmented; Based on the hierarchical clustering method, the feature vectors of adjacent cube grids are clustered and layered according to the preset distance threshold to obtain the clustering data set of the corresponding point cloud data.

8. The three-dimensional point cloud clustering method according to claim 1, characterized in that: The point cloud clustering method in the three-dimensional space also includes: Obstacle location and size data are obtained based on the clustered data set of point cloud data.

9. A three-dimensional point cloud clustering system, characterized in that: The point cloud clustering method in three-dimensional space according to any one of claims 1 to 8 comprises: The acquisition module is used to obtain the point cloud data of the surrounding environment space and divide the three-dimensional space corresponding to the environment space into cubic grids; A denoising module, used to remove the cubic grids that do not contain objects according to a preset noise filtering condition; A ground removal module is used to remove the cube grid containing the ground based on a height threshold and ground plane fitting method; A feature extraction module is used to extract features from the point cloud data of the environment space corresponding to the remaining cubic grid based on the graph convolution method to obtain a feature vector; The clustering module is used to cluster the feature vectors based on the distance threshold and hierarchical clustering method to obtain the clustering data set corresponding to the point cloud data.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed, implements the steps of the point cloud clustering method in three-dimensional space as claimed in any one of claims 1 to 8.

Images