Self-adaptive neighborhood point cloud fusion method based on voxel pre-acceleration

Abstract

The invention discloses an adaptive neighborhood point cloud fusion method based on voxel pre-acceleration, and belongs to the technical field of three-dimensional data processing, and the method comprises the following steps: S1, collecting three-dimensional point clouds, and forming a total point cloud; s2, performing space division on the total point cloud to generate a Hash dictionary; s3, traversing the Hash dictionary to generate a neighborhood point cluster; and S4, calculating representative points for each neighborhood point cluster, and taking a set of all the representative points as a final fusion point cloud. The algorithm logic is clear, engineering implementation, deployment and parallelization acceleration are easy, and the method has high robustness and engineering value in practical application.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional data processing technology, specifically relating to an adaptive neighborhood point cloud fusion method based on voxel pre-acceleration. Background Technology

[0002] In previous research on 3D environment perception using multiple detector units (DMUs), fusing the point cloud data acquired by each DMU is a crucial subsequent processing step. However, existing fusion methods typically suffer from several problems: First, regarding fusion quality, simple point cloud data overlay, while capable of gathering all information, results in abnormally dense data points and significant redundancy in overlapping detector areas. Furthermore, minor measurement errors between different DMUs can lead to data inconsistencies, directly impacting the smoothness of subsequent surface reconstruction and the accuracy of feature extraction. Second, regarding fusion efficiency, nearest neighbor search algorithms used to effectively remove data redundancy typically involve enormous computational costs. Traditional global brute-force search algorithms are too complex to meet real-time processing requirements. While algorithms based on advanced spatial index structures can improve efficiency, they still suffer from high index construction overhead and complex implementation, hindering their application and deployment in practical engineering. Therefore, there is still considerable room for further research in this field to develop innovative algorithms that balance processing efficiency and fusion quality. Summary of the Invention

[0003] To address the above problems, this invention proposes an adaptive neighborhood point cloud fusion method based on voxel pre-acceleration.

[0004] The technical solution of this invention is: an adaptive neighborhood point cloud fusion method based on voxel pre-acceleration, comprising the following steps:

[0005] S1. Collect 3D point cloud data to form the total point cloud;

[0006] S2. Spatial partitioning of the total point cloud to generate a hash dictionary;

[0007] S3. Traverse the hash dictionary to generate neighborhood point clusters;

[0008] S4. Calculate representative points for each neighboring point cluster, and use the set of all representative points as the final fused point cloud.

[0009] Furthermore, in S1, the receiver Each detection unit is registered in the same global coordinate system. A three-dimensional point cloud is formed to create a total point cloud.

[0010] Furthermore, total point cloud The expression is:

[0011] ;

[0012] in, For each 3D point cloud, This represents the number of 3D point cloud components.

[0013] Furthermore, S2 includes the following sub-steps:

[0014] S21. Set the fusion distance threshold and pre-divided voxel size;

[0015] S22. Spatial division of the total point cloud using a 3D mesh with pre-divided voxel sizes;

[0016] S23. Calculate the voxel index for any point after spatial partitioning;

[0017] S24. Construct a hash dictionary with voxel indices as keys and the set of points falling into voxels as values.

[0018] Furthermore, in S21, the expression between the fusion distance threshold and the pre-divided voxel size is:

[0019] ;

[0020] in, To pre-divide voxel sizes, The fusion distance threshold, is a coefficient.

[0021] Furthermore, in S23, voxel indexing The expression is:

[0022] ;

[0023] in, Let x be the x-coordinate of any point. Let be the y-coordinate of any point. Let be the vertical coordinate of any point. To pre-divide voxel sizes, This is for floor function.

[0024] Furthermore, S3 includes the following sub-steps:

[0025] S31. Initialize the global collection;

[0026] S32. Traverse all non-empty voxel units of the hash dictionary;

[0027] S33. For any center point in a non-empty voxel cell in the hash dictionary that is not recorded in the global set, create a new point cluster and add the center point to the new point cluster;

[0028] S34. Traverse the remaining unrecorded points within the voxel unit and calculate their Euclidean distance from the center point.

[0029] S35. Add points whose Euclidean distance is less than the set threshold to a new point cluster to generate a neighborhood point cluster.

[0030] Furthermore, in S4, the representative point The expression is:

[0031] ;

[0032] in, For the neighborhood point cluster, the first One point, The number of points in the neighborhood cluster.

[0033] The beneficial effects of this invention are:

[0034] (1) The present invention adopts a two-stage strategy of “coarse division first and then refinement”. Through the “voxel pre-acceleration” in the first stage, the global and computationally complex nearest neighbor search problem is cleverly decomposed into a large number of parallel local search problems that are limited to a single voxel and have low computational complexity. This fundamentally avoids the huge computational overhead of brute global search in traditional algorithms and also avoids the large amount of preprocessing time required to build a complex spatial index structure. It is especially suitable for real-time processing scenarios of large-scale, high-density point clouds.

[0035] (2) Through the second-stage “adaptive neighborhood” fusion, this invention can intelligently identify and aggregate truly neighboring point clusters in three-dimensional space according to the set physical distance threshold. This method can effectively eliminate data redundancy generated by multiple detection units in overlapping areas, and generate a point cloud model with more uniform density and significantly reduced data volume. At the same time, since its fusion method generates representative points at the geometric center of local point clusters, it can retain the key geometric features of the original landform with high fidelity and avoid the feature blurring problem that may be caused by conventional voxel downsampling.

[0036] (3) The core parameter of this invention—the pre-division voxel size and the fusion distance threshold—has a direct mathematical coupling relationship and a clear physical meaning, making the algorithm adjustment simple and intuitive. The entire algorithm process is based on spatial hashing and iterative loops, and does not rely on complex recursive data structures (such as kd-trees and octrees). The algorithm logic is clear, easy to implement in engineering, deploy and accelerate in parallel, and has high robustness and engineering value in practical applications. Attached Figure Description

[0037] Figure 1 The flowchart shows the adaptive neighborhood point cloud fusion method based on voxel pre-acceleration.

[0038] Figure 2(a) is a global point cloud view of detection unit 1;

[0039] Figure 2(b) is a top view of the overall point cloud of detection unit 1;

[0040] Figure 3(a) is a global point cloud view of detection unit 2;

[0041] Figure 3(b) is a top view of the overall point cloud of detection unit 2;

[0042] Figure 4 This is a global terrain point cloud fusion result image;

[0043] Figure 5 This is a three-dimensional global real seabed topographic map. Detailed Implementation

[0044] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0045] like Figure 1 As shown, this invention provides an adaptive neighborhood point cloud fusion method based on voxel pre-acceleration, comprising the following steps:

[0046] S1. Collect 3D point cloud data to form the total point cloud;

[0047] S2. Spatial partitioning of the total point cloud to generate a hash dictionary;

[0048] S3. Traverse the hash dictionary to generate neighborhood point clusters;

[0049] S4. Calculate representative points for each neighboring point cluster, and use the set of all representative points as the final fused point cloud.

[0050] This invention proposes an adaptive neighborhood point cloud fusion method based on voxel pre-acceleration. Through a two-stage strategy of "coarse segmentation followed by refinement," it cleverly transforms a complex global search problem into multiple independent local search problems. The first stage, voxel pre-acceleration, significantly reduces computational complexity without constructing complex data structures. The second stage, adaptive neighborhood fusion, effectively eliminates redundant data while preserving key geometric features. This method can significantly improve processing speed and efficiency while maintaining high-quality fusion results, providing important technical support for data processing in multi-sensor collaborative detection.

[0051] In this embodiment of the invention, in S1, the receiver... Each detection unit is registered in the same global coordinate system. A three-dimensional point cloud is formed to create a total point cloud.

[0052] In this embodiment of the invention, the total point cloud The expression is:

[0053] ;

[0054] in, For each 3D point cloud, This represents the number of 3D point cloud components.

[0055] In this embodiment of the invention, S2 includes the following sub-steps:

[0056] S21. Set the fusion distance threshold and pre-divided voxel size;

[0057] S22. Spatial division of the total point cloud using a 3D mesh with pre-divided voxel sizes;

[0058] S23. Calculate the voxel index for any point after spatial partitioning;

[0059] S24. Construct a hash dictionary with voxel indices as keys and the set of points falling into voxels as values.

[0060] In this embodiment of the invention, in S21, the expression between the fusion distance threshold and the pre-divided voxel size is:

[0061] ;

[0062] in, To pre-divide voxel sizes, The fusion distance threshold, is a coefficient.

[0063] In this embodiment of the invention, in S23, the voxel index... The expression is:

[0064] ;

[0065] in, Let x be the x-coordinate of any point. Let be the y-coordinate of any point. Let be the vertical coordinate of any point. To pre-divide voxel sizes, This is for floor function.

[0066] In this embodiment of the invention, S3 includes the following sub-steps:

[0067] S31. Initialize the global collection;

[0068] S32. Traverse all non-empty voxel units of the hash dictionary;

[0069] S33. For any center point in a non-empty voxel cell in the hash dictionary that is not recorded in the global set, create a new point cluster and add the center point to the new point cluster;

[0070] S34. Traverse the remaining unrecorded points within the voxel unit and calculate their Euclidean distance from the center point.

[0071] S35. Add points whose Euclidean distance is less than the set threshold to a new point cluster to generate a neighborhood point cluster.

[0072] In this embodiment of the invention, S4 represents a point. The expression is:

[0073] ;

[0074] in, For the neighborhood point cluster, the first One point, The number of points in the neighborhood cluster.

[0075] The following description is based on specific embodiments.

[0076] First, receive and aggregate from Each detection unit, registered in the same global coordinate system 3D point cloud Each 3D point cloud Each is a point containing multiple three-dimensional coordinates. The collection forms a unified, unordered point cloud containing the coordinates of all data points. The global point cloud images of the two detection units are shown in Figure 2(a) and Figure 2(b), and their top views are shown in Figure 3(b).

[0077] Based on the user-defined fusion distance threshold Set a pre-division voxel size and using a size of The three-dimensional mesh of the total point cloud Divide the space.

[0078] For any point in the total point cloud Calculate its corresponding voxel index. .

[0079] Ultimately, a hash dictionary can be constructed with voxel indices as keys and the set of points falling into that voxel as values. .

[0080] Traverse the hash dictionary All non-empty voxel units are processed, and neighborhood fusion operations are performed independently within each unit: first, a global set is initialized to record the indices of processed points, and then for the hash dictionary... Any element within a non-empty voxel unit that is not recorded at the center point of the global set Create a new cluster of points And Add it to it, and at the same time, this point The index is used to record the data, and all other unrecorded points within the voxel unit are traversed. Calculate its distance from the center point Euclidean distance ,like Then the point Add point clusters In the middle, its index is recorded.

[0081] Final calculation of representative points Its coordinates are determined by the neighborhood point cluster. All Points The arithmetic mean of the coordinates is determined.

[0082] This completes the point cloud fusion. The global terrain point cloud fusion result is shown below. Figure 4 .

[0083] To more clearly and intuitively illustrate the implementation process of this invention, a specific technical example of collaborative mapping using dual detection units is given. In this embodiment, two underwater acoustic detection units, sonar 1 and sonar 2, are deployed. The global terrain function model used for the detection environment is a complex terrain model. This model, at a base depth of -50 meters, superimposes a linear slope, sinusoidal waves, two Gaussian hills of different sizes, and a Gaussian depression to simulate the complex undulations of the real seabed. An example of a three-dimensional global real seabed topographic map can be found... Figure 5The rotation center of detector unit 1 is located at global coordinates (0,0,-10), its initial installation position is (60,0,5), and its initial attitude vector of the Z-axis in its intrinsic coordinate system is (0.5,0,0.866), meaning the detector head is tilted 60° upwards and eastwards. The rotation center of detector unit 2 is located at (-10,5,-5), its initial installation position is (-60,35,5), and its initial attitude vector of the Z-axis in its intrinsic coordinate system is (-0.5,0,0.866), meaning the detector head is tilted 60° upwards and westwards. The scanning parameters for each detector unit are set as follows: total horizontal rotation range of 270°, rotation step size of 1°; instantaneous sector opening angle of 90°, and ray resolution within the sector of 1°. After completing their respective scanning detections, both units generate two original point clouds containing approximately 24,000 data points. These two point cloud data are used as input and combined to form a total point cloud containing approximately 49,000 points. Using the fusion method described in this invention, a fusion distance threshold T = 0.4 meters is set. Following the steps and process of this invention, the total point cloud is first subjected to voxel pre-acceleration processing, and then adaptive neighborhood fusion is performed on it, ultimately obtaining a simplified fused point cloud of approximately 36,000 points. This method, through a two-stage strategy of "coarse segmentation followed by refinement," efficiently eliminates redundancy in overlapping areas of the original data. Simultaneously, by taking the average value within local neighborhoods, it ensures high fidelity of the original terrain features in the fused point cloud. This method more efficiently and intelligently constructs a unified three-dimensional seabed topography model, strongly supporting research on multi-sensor collaborative mapping and environmental perception technologies, and more realistically and accurately realizing the fusion and simplification of multi-source heterogeneous point cloud data.

[0084] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A voxel pre-acceleration based adaptive neighborhood point cloud fusion method, characterized in that, The method comprises the following steps: S1, collecting three-dimensional point clouds to form total point clouds; S2, spatially dividing the total point clouds to generate a hash dictionary; S3, traversing the hash dictionary to generate neighborhood point clusters; S4, calculating representative points for each neighborhood point cluster, and taking the set of all representative points as a final fused point cloud.

2. The voxel-preacceleration-based adaptive neighborhood point cloud fusion method according to claim 1, characterized in that, In the S1, receiving a plurality of three-dimensional point clouds of the individual probe units, which are registered to the same global coordinate system to form a total point cloud.

3. The voxel-preacceleration-based adaptive neighborhood point cloud fusion method according to claim 2, characterized in that, The total point cloud The expression is: ； wherein, is the number of three-dimensional point clouds, is the number of three-dimensional point clouds.

4. The voxel-pre-acceleration based adaptive neighborhood point cloud fusion method according to claim 1, characterized in that, The S2 comprises the following sub-steps: S21, setting a fusion distance threshold and a pre-partition voxel size; S22, spatially dividing the total point clouds by a three-dimensional grid with the pre-partition voxel size; S23, calculating a voxel index for any point after spatial division; S24, constructing a hash dictionary with the voxel index as the key and the point set falling into the voxel as the value.

5. The voxel-preacceleration-based adaptive neighborhood point cloud fusion method according to claim 4, characterized in that, In the S21, the expression between the fusion distance threshold and the pre-partition voxel size is: ； wherein, is a pre-division voxel size, is a fusion distance threshold, is a coefficient.

6. The voxel-preacceleration-based adaptive neighborhood point cloud fusion method according to claim 4, characterized in that, In the S23, the expression of the voxel index is: ； wherein, is the horizontal coordinate of an arbitrary point, is the vertical coordinate of an arbitrary point, is the vertical coordinate of an arbitrary point, is the pre-division voxel size, is the floor operation.

7. The voxel-pre-acceleration based adaptive neighborhood point cloud fusion method according to claim 1, characterized in that, The S3 comprises the following sub-steps: S31, initializing a global set; S32, traversing all non-empty voxel units of the hash dictionary; S33, for any center point in the non-empty voxel units of the hash dictionary that has not been recorded in the global set, creating a new point cluster and adding the center point to the new point cluster; S34, traversing the remaining unrecorded points in the voxel unit to calculate the Euclidean distance between the points and the center point; S35, adding the points with the Euclidean distance less than the set threshold to the new point cluster to generate a neighborhood point cluster.

8. The voxel-pre-acceleration based adaptive neighborhood point cloud fusion method according to claim 1, characterized in that, In the S4, the representative point The expression is: ； wherein, is the number of points in the cluster of neighborhood points, is the number of points in the cluster of neighborhood points, is the number of points in the cluster of neighborhood points.

Images