A point cloud segmentation method based on feature clustering in natural scenes

Abstract

This invention discloses a point cloud segmentation method based on feature clustering in natural scenes, comprising: acquiring point cloud data with feature filtering, removing points containing ground features from the point cloud data to obtain a point cloud dataset free of ground feature interference and conforming to feature clustering; performing coarse segmentation on the point cloud dataset based on a feature clustering algorithm to obtain segmentation results; denoising the segmentation results based on a shape distribution algorithm to obtain denoised point cloud data; and performing fine segmentation on the denoised point cloud data based on a feature clustering algorithm to output the finely segmented point cloud data. The point cloud model segmentation method of this invention combines feature clustering and other related technologies in natural scenes, designs a complete segmentation technique, effectively reduces the computational complexity and computational resource overhead of the algorithm, and significantly improves accuracy. It also outperforms other methods in terms of real-time performance, scalability, and load balancing.

Description

Technical Field

[0001] This invention belongs to the field of point cloud segmentation, and specifically relates to a point cloud segmentation method based on feature clustering in natural scenes. Background Technology

[0002] Point clouds, as an important data format capable of fully representing the semantic information of the surrounding scene, have become the preferred research object for developing application technologies in various fields due to their crucial characteristic of preserving original geometric information intact in three-dimensional space. In recent years, an increasing number of researchers have focused on the exploration of point clouds and rapidly applied them to many fields such as computer vision, autonomous driving, and robotics. Therefore, how to modify and reuse point clouds as a research object has become an important issue.

[0003] Segmentation technology refers to the process of dividing points in three-dimensional space into smaller, coherent, and connected subsets. After segmentation, points with similar attributes are grouped together, possessing stronger semantic information. In recent years, with continuous exploration by researchers, point cloud segmentation algorithms have made some progress. For example, Mario Richtsfeld proposed a point cloud segmentation method based on radial reflection, and Aleksey Golovinskiy proposed several point cloud segmentation algorithms based on minimum cut. However, many existing point cloud segmentation algorithms are often targeted at niche application scenarios and suffer from problems such as excessive algorithm complexity and long running time, resulting in low algorithm efficiency. In addition, point cloud segmentation methods based on deep learning have also received increasing attention, but the large amount of data required for deep learning training prevents its application to specific application scenarios. At the same time, deep learning requires large-scale computation, thus placing high demands on computing resources and making it unsuitable for direct application in display scenarios. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the prior art, the present invention provides a point cloud segmentation method based on feature clustering in natural scenes. It combines feature clustering and other related technologies in natural scenes and designs a complete segmentation technology, which effectively reduces the computational complexity and computational resource overhead of the algorithm, while significantly improving the accuracy. It is also superior to other methods in terms of real-time performance, scalability, and load balancing.

[0005] Technical solution: In a first aspect, the present invention provides a point cloud segmentation method based on feature clustering in natural scenes, comprising:

[0006] Obtain point cloud data with feature filtering, and remove point clouds containing ground features to obtain a point cloud dataset without ground feature interference factors and conforming to feature clustering.

[0007] The point cloud dataset is coarsely segmented based on a feature clustering algorithm to obtain the segmentation results;

[0008] The segmentation results are denoised using a shape distribution algorithm to obtain denoised point cloud data.

[0009] The denoised point cloud data is finely segmented based on a feature clustering algorithm, and the finely segmented point cloud data is output.

[0010] In a further embodiment, the method for obtaining point cloud data based on feature filtering, and removing point clouds containing ground features to obtain a point cloud dataset free from ground feature interference and conforming to feature clustering is as follows:

[0011] The image is randomly downsampled from the original point cloud data to obtain point cloud data. During the sampling process, feature filtering is performed to remove isolated points in the original point cloud data, and the point cloud data is inverted to obtain feature-filtered point cloud data.

[0012] The feature-filtered point cloud data is imported into the cloth simulation algorithm for processing, and the point cloud data containing ground features is filtered out.

[0013] Point clouds containing non-ground features are preserved to obtain a point cloud dataset that is free from ground feature interference and conforms to feature clustering.

[0014] In a further embodiment, the target width and target height of the image downsampling are both set to 33 in the random downsampling of the image; the contrast truncation threshold of the feature filtering is set to 40.0.

[0015] In a further embodiment, the point cloud dataset is coarsely segmented based on a feature clustering algorithm to obtain the segmentation results, including:

[0016] A point in the point cloud dataset that does not belong to the already clustered point set A1 is randomly selected as the reference point. The cluster set S1 around the reference point is set to be empty. In the coarse segmentation feature clustering algorithm, the lower limits of the already clustered point set A1, the feature aggregation radius r1, and the number of clustered point clouds are preset.

[0017] If a point is randomly selected from the point cloud data, and the selected point does not exist in the clustered set A1, then the l2 norm between the selected point and the reference point is calculated as the distance. If the distance is less than the feature aggregation radius r1, then the selected point is added to the cluster set S1 around the reference point and added to the clustered set A1. Then, all points are selected from the point cloud data in a loop and added to the cluster set S1 around the reference point until the distance between the un-clustered points and the reference point is greater than the feature clustering radius r1.

[0018] Calculate the number of points in the cluster set S1 surrounding the reference point. If it is greater than the lower limit of the number of points in the clustered point cloud, mark set S1' as the standard set. Otherwise, mark the surrounding cluster set S1 as the failure set and select the output standard set.

[0019] The standard set of output is selected iteratively until all points in the point cloud data are located in the clustered set A1. All standard clusters are output, and all points in the point cloud data are clustered based on the clusters to obtain the coarse segmentation result.

[0020] In a further embodiment, the feature aggregation radius r is set to 8; the lower limit of the number of clustered point clouds is set to 1000.

[0021] In a further embodiment, the segmentation results are denoised based on a shape distribution algorithm to obtain denoised point cloud data.

[0022] The coarse segmentation results are randomly downsampled, and the coordinates of the coarse segmentation results are transformed to obtain the coordinates of all points in the point cloud data;

[0023] Calculate the center coordinates of the point cloud data, then calculate the distribution distance of each point in the point cloud data from the center coordinates, and finally calculate the mean of the distribution distance as l. -

[0024] Noisy points are filtered based on the mean of the distribution distance, and points in the point cloud data whose distance from the center coordinates is greater than twice the mean distribution distance are considered. - The point cloud data points are deleted to obtain denoised point cloud data.

[0025] In a further embodiment, the method for finely segmenting the denoised point cloud data based on a feature clustering algorithm and outputting the finely segmented point cloud data is as follows:

[0026] A point in the denoised point cloud data that belongs to the unclustered point set A2 and does not belong to the unsuccessful set B is randomly selected as the reference point, and the cluster set S2 around the reference point is set to empty; the feature clustering algorithm in fine segmentation presets the feature aggregation radius r2, the lower limit of the number of clustered point clouds, the similarity distance l, the unclustered point set A2, and the unsuccessful set B;

[0027] From the denoised point cloud data, select any point. If the selected point exists in the unclustered set A2, calculate the l2 norm between the point and the reference point as the distance. If the distance is less than the feature aggregation radius r2, calculate the similarity distance l' between the point and the reference point. If l' is less than the reference similarity distance l, add the selected point to the cluster set S2 around the reference point and delete the point from the unclustered set A2. Repeat this process for all points in the denoised point cloud data until the distance between the unclustered points and the reference point is greater than the feature clustering radius r2.

[0028] Calculate the number of points in the cluster set S2 surrounding the reference point. If it is greater than the lower limit of the number of points in the clustered point cloud, mark set S2' as the standard set and output the standard set. If the number of points in the cluster set S2 surrounding the reference point is less than the lower limit of the number of points in the clustered point cloud, add all points in the cluster set S2 to the unsuccessful set B, and put all point cloud data points in the cluster set S2 back into the unclustered point set A2. Repeat the process of selecting all points to add to the cluster set S2 surrounding the reference point until the distance between the unclustered points and the reference point is greater than the feature cluster radius r2. Then recalculate the number of points in the cluster set S2 surrounding the reference point until the standard set is output.

[0029] The standard set of output is selected iteratively until all points in the point cloud data are no longer in the unclustered set A2. All standard clusters are then output. Based on the standard clusters, all points in the point cloud data are clustered to obtain finely segmented point cloud data.

[0030] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0031] The method of this invention combines feature clustering and other related technologies in natural scenes to design a complete segmentation technique, which effectively reduces the computational complexity and computational resource overhead of the algorithm, while significantly improving accuracy. It is also superior to other methods in terms of real-time performance, scalability, and load balancing. Attached Figure Description

[0032] Figure 1 A detailed flowchart of the point cloud model segmentation method of the present invention. Detailed Implementation

[0033] To better understand the technical content of the present invention, the technical solution of the present invention will be further introduced and explained below with reference to specific embodiments, but is not limited thereto.

[0034] Combination Figure 1 This embodiment further illustrates a point cloud segmentation method based on feature clustering in a natural scene, including:

[0035] S1: Obtain the point cloud data with feature filtering, and remove the point cloud data containing ground features to obtain a point cloud dataset without ground feature interference factors and conforming to feature clustering.

[0036] S2: The point cloud dataset is coarsely segmented based on a feature clustering algorithm to obtain the segmentation results;

[0037] S3: The segmentation results are denoised based on the shape distribution algorithm to obtain denoised point cloud data;

[0038] S4: The denoised point cloud data is finely segmented based on the feature clustering algorithm, and the finely segmented point cloud data is output.

[0039] The method for obtaining point cloud data based on feature filtering, and removing points containing ground features to obtain a point cloud dataset free from ground feature interference and conforming to feature clustering is as follows:

[0040] S1-1: Random downsampling of the image from the original point cloud data to obtain point cloud data; wherein, during the sampling process, feature filtering is performed to remove isolated points in the original point cloud data, and the point cloud data is reversed to obtain feature-filtered point cloud data;

[0041] S1-2: Import the feature-filtered point cloud data into the cloth simulation algorithm for processing, and filter out the point cloud data containing ground features.

[0042] S1-3: Preserve non-ground features in the point cloud data to obtain a point cloud dataset that is free from ground feature interference and conforms to feature clustering.

[0043] The fabric simulation algorithm is as follows:

[0044] S1-2-1: Initial cloth grid, sets the grid resolution (gr). The initial "cloth" position is usually above the highest point.

[0045] S1-2-2: Project all points and grid particles in the point cloud onto the same horizontal plane, find the nearest neighbor (corresponding point, cp) of each particle, and record its elevation before projection (intersection heightvalue, ihv).

[0046] S1-2-3: For each movable grid "particle", calculate its displacement caused by gravity and compare it with the ihv of the current particle's corresponding cp point. If the particle's height is lower than or equal to ihv, set the particle's height to ihv and set it as an immovable point.

[0047] S1-2-4: For each grid "particle", calculate its displacement caused by the influence of internal driving factors.

[0048] S1-2-5: Repeat steps S1-2-3 and S1-2-4 until the maximum high commitment change of all particles is small enough or the number of iterations reaches the user's preset value, then stop the simulation process.

[0049] S1-2-6: Calculate the height difference between point cloud data and grid particles.

[0050] S1-2-7: Distinguish between ground points and non-ground points. If the distance between a point in the point cloud and the simulated particle is less than the preset threshold hcc, it is considered a ground point; otherwise, it is considered a non-ground point. Non-ground points are retained and output.

[0051] In this embodiment, the target width and target height of the image downsampling are both set to 33 in the random downsampling of the image; the contrast truncation threshold of the feature filtering is set to 40.0.

[0052] Point cloud datasets are coarsely segmented using feature clustering algorithms. Methods for obtaining segmentation results include:

[0053] S2-1: Randomly select a point in the point cloud dataset that does not belong to the already clustered point set A1 as the reference point, and set the cluster set S1 around the reference point to be empty. In the coarse segmentation feature clustering algorithm, the lower limit values ​​of the already clustered point set A1, the feature aggregation radius r1, and the number of clustered point clouds are preset.

[0054] S2-2: Select any point from the point cloud data. If the selected point does not exist in the clustered set A1, calculate the l2 norm between the selected point and the reference point as the distance. If the distance is less than the feature aggregation radius r1, add the selected point to the cluster set S1 around the reference point and add it to the clustered set A1. Then, iteratively select all points from the point cloud data and add them to the cluster set S1 around the reference point until the distance between the un-clustered points and the reference point is greater than the feature clustering radius r1.

[0055] S2-3: Calculate the number of points in the surrounding cluster set S1 of the reference point. If it is greater than the lower limit of the number of clustered point clouds, mark set S1' as the standard set. Otherwise, mark the surrounding cluster set S1 as the failure set and select the output standard set.

[0056] S2-4: Iteratively select the output standard set until all points in the point cloud data are located in the clustered set A1, output all standard clusters, and cluster all points in the point cloud data based on the clusters to obtain the coarse segmentation result.

[0057] In this embodiment, the feature aggregation radius r1 is set to 8; the lower limit of the number of clustered point clouds is set to 1000.

[0058] S3-1: The segmentation results are denoised using a shape distribution algorithm. The method for obtaining denoised point cloud data is as follows:

[0059] The coarse segmentation results are randomly downsampled, and the coordinates of the coarse segmentation results are transformed to obtain the coordinates of all points in the point cloud data;

[0060] S3-2: Calculate the center coordinates of the point cloud data, calculate the distribution distance of each point in the point cloud data from the center coordinates, and calculate the mean of the distribution distance as l. -

[0061] S3-3: Noise points are filtered based on the mean of the distribution distance, and for point cloud data, points whose distance from the center coordinates is greater than twice the mean distribution distance l - The point cloud data points are deleted to obtain denoised point cloud data.

[0062] S4-1: The method for refining the denoised point cloud data based on the feature clustering algorithm and outputting the finely segmented point cloud data is as follows:

[0063] S4-2: Randomly select a point in the denoised point cloud data that belongs to the unclustered point set A2 and does not belong to the unsuccessful set B as the reference point, and set the cluster set S2 around the reference point to be empty; where the feature clustering algorithm in fine segmentation presets the feature aggregation radius r2, the lower limit of the number of clustered point clouds, the similarity distance l, the unclustered point set A2, and the unsuccessful set B;

[0064] S4-3: Select any point from the denoised point cloud data. If the selected point exists in the unclustered set A2, calculate the l2 norm between the point and the reference point as the distance. If the distance is less than the feature aggregation radius r2, calculate the similarity distance l' between the point and the reference point. If l' is less than the reference similarity distance l, add the selected point to the cluster set S2 around the reference point and delete the point from the unclustered set A2. Iteratively select all points from the denoised point cloud data and add them to the cluster set S2 around the reference point until the distance between the unclustered points and the reference point is greater than the feature clustering radius r2.

[0065] S4-4: Calculate the number of points in the cluster set S2 surrounding the reference point. If it is greater than the lower limit of the number of points in the clustered point cloud, mark set S2' as the standard set and output the standard set. If the number of points in the cluster set S2 surrounding the reference point is less than the lower limit of the number of points in the clustered point cloud, add all points in the cluster set S2 to the unsuccessful set B, and put all point cloud data points in the cluster set S2 back into the unclustered point set A2. Then proceed to S4-3.

[0066] S4-5: Repeat S4-3 and S4-4 until all points in the point cloud data are no longer in the unclustered set A2. Output all standard clusters. Based on the standard clusters, cluster all points in the point cloud data to obtain finely segmented point cloud data.

[0067] In summary, the method of this invention combines feature clustering and other related technologies in natural scenes to design a complete segmentation technique, which effectively reduces the computational complexity and computational resource overhead of the algorithm, while significantly improving accuracy. It also outperforms other methods in terms of real-time performance, scalability, and load balancing.

[0068] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0070] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A point cloud segmentation method based on feature clustering in natural scenes, characterized in that, include: Obtain point cloud data with feature filtering, and remove point clouds containing ground features to obtain a point cloud dataset without ground feature interference factors and conforming to feature clustering. The point cloud dataset is coarsely segmented based on a feature clustering algorithm to obtain the segmentation results; The segmentation results are denoised using a shape distribution algorithm to obtain denoised point cloud data. The denoised point cloud data is finely segmented based on a feature clustering algorithm, and the finely segmented point cloud data is output. Point cloud datasets are coarsely segmented using feature clustering algorithms. Methods for obtaining segmentation results include: Randomly select points in the point cloud dataset that do not belong to the already clustered set. One point is used as a reference point, and a cluster set is set around the reference point. Empty, where; the pre-set set of clustered points in the coarse segmentation feature clustering algorithm. Feature aggregation radius And the lower limit of the number of clustered point clouds; Select any point from the point cloud data. If the selected point does not exist in the already clustered set... In the middle, the distance between the selected point and the reference point is calculated. Norm as distance, if the distance is less than the feature aggregation radius Then, any chosen point will be added to the cluster set around the reference point. and add to the already clustered set. In the middle; and iteratively select all points from the point cloud data to add them to the cluster set around the reference point. The distances to the non-clustered points and the reference point are all greater than the feature cluster radius. ; Calculate the cluster set around the reference point If the number of midpoints is greater than the lower limit of the number of points in the clustered point cloud, then the set is marked. If it is the standard set, then the surrounding cluster set. Mark as the failure set and select the output standard set; The process continues, iteratively selecting the output standard set until all points in the point cloud data fall within the clustered set. In the process, all standard clusters are output, and all points in the point cloud data are clustered based on the clusters to obtain coarse segmentation results; The method for refining point cloud data based on feature clustering algorithm and outputting the refined point cloud data is as follows: Randomly select the set of unclustered points from the denoised point cloud data. And not belonging to the unsuccessful set Using one point as a reference point, set up a cluster set around the reference point. Empty; where the feature clustering algorithm in fine segmentation presets the feature aggregation radius. Lower limit of the number of clustered point clouds, similarity distance Unclustered point set Unsuccessful assembly ; Select any point from the denoised point cloud data. If the selected point exists in the unclustered set... In the middle, the distance between the point and the reference point is calculated. Norm as distance, if the distance is less than the feature aggregation radius Simultaneously, calculate the similarity distance between this point and the reference point. ,like Less than the baseline similarity distance Then, any chosen point will be added to the cluster set around the reference point. and in the unclustered set Delete the point; iteratively select all points from the denoised point cloud data and add them to the cluster set around the reference point. Until the distance between each unclustered point and the baseline point is greater than the feature cluster radius. ; Calculate the cluster set around the reference point If the number of midpoints exceeds the lower limit of the number of points in the clustered point cloud, then the set is labeled. Given a standard set, output the standard set; where the set of clusters surrounding the reference point is... If the number of midpoints is less than the lower limit of the number of clustered point clouds, then the surrounding clusters will be grouped together. All points were added to the unsuccessful set. In the middle, it simultaneously gathers the surrounding clusters. All point cloud data points were replaced with the set of unclustered points. In the process, all points are selected again in a loop and added to the cluster set around the reference point. Until the distance between each unclustered point and the baseline point is greater than the feature cluster radius. And recalculate the cluster set surrounding the benchmark point. The number of midpoints is counted until the standard set is output; The output standard set is repeated until all points in the point cloud data are no longer located in the unclustered set. In the process, all standard clusters are output, and all points in the point cloud data are clustered based on the standard clusters to obtain finely segmented point cloud data.

2. The point cloud segmentation method based on feature clustering in natural scenes according to claim 1, characterized in that, The method for obtaining point cloud data based on feature filtering, removing points containing ground features, and obtaining a point cloud dataset free from ground feature interference and conforming to feature clustering is as follows: The image is randomly downsampled from the original point cloud data to obtain point cloud data. During the sampling process, feature filtering is performed to remove isolated points in the original point cloud data, and the point cloud data is inverted to obtain feature-filtered point cloud data. The feature-filtered point cloud data is imported into the cloth simulation algorithm for processing, and the point cloud data containing ground features is filtered out. Point clouds containing non-ground features are preserved to obtain a point cloud dataset that is free from ground feature interference and conforms to feature clustering.

3. The point cloud segmentation method based on feature clustering in natural scenes according to claim 2, characterized in that, In the random downsampling of the image, the target width and target height of the image downsampling are both set to 33; the contrast truncation threshold of the feature filtering is set to 40.

0.

4. The point cloud segmentation method based on feature clustering in natural scenes according to claim 1, characterized in that, The feature aggregation radius r is set to 8; the lower limit of the number of clustered point clouds is set to 1000.

5. A point cloud segmentation method based on feature clustering in natural scenes according to claim 1, characterized in that, The segmentation results are denoised using a shape distribution algorithm. The method for obtaining denoised point cloud data is as follows: The coarse segmentation results are randomly downsampled, and the coordinates of the coarse segmentation results are transformed to obtain the coordinates of all points in the point cloud data; Calculate the center coordinates of the point cloud data, then calculate the distribution distance of each point in the point cloud data from the center coordinates, and finally calculate the mean of the distribution distances. ; Noisy points are filtered based on the mean of the distribution distance, and points in the point cloud data whose distance from the center coordinates is greater than twice the mean of the distribution distance are considered noise points. The point cloud data points are deleted to obtain denoised point cloud data.

Images