High-universality farthest point down-sampling algorithm based on multi-level adaptive grid

Through multi-level adaptive grid division and farthest point sampling algorithm, the problem that point cloud downsampling in existing technology cannot fully retain global features is solved, and high-efficiency and accurate point cloud data processing is achieved.

CN120599290APending Publication Date: 2025-09-05SHANDONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510735088.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-06-04
Publication Date
2025-09-05

AI Technical Summary

Technical Problem

Existing downsampling technologies cannot achieve both high efficiency and complete preservation of global features, resulting in a decrease in the accuracy of subsequent point cloud data processing.

Method used

The farthest point downsampling algorithm of multi-level adaptive grid is adopted. The point cloud is divided into multi-level grids step by step. The point capacity is set according to the number of idle CPU cores. Multi-threaded computing is used for farthest point sampling to ensure the integrity of the global and local features of the point cloud.

Benefits of technology

It achieves the high-efficiency retention of global and local features of point clouds, improves the accuracy and computational efficiency of subsequent point cloud data processing, and is suitable for large-scale, high-precision point cloud processing tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120599290A_ABST
    Figure CN120599290A_ABST
Patent Text Reader

Abstract

The invention discloses a high-universality farthest point down-sampling algorithm based on multi-level adaptive grids, belongs to the technical field of data processing, is used for point cloud data processing, and comprises the following steps: determining point capacity in each grid, acquiring idle CPU (Central Processing Unit) core number of current processing equipment, dividing original point cloud into multi-level grids step by step, and carrying out farthest point sampling on the grid, if the point number after farthest point sampling is greater than the point number capacity, repeatedly executing until the point number after farthest point sampling is less than the point number capacity, and carrying out farthest point sampling on the point cloud after farthest point sampling again. According to the method, the high efficiency of a down-sampling algorithm is ensured, and the method has obvious advantages in large-scale and high-precision point cloud processing tasks; global features and local features of the point cloud are reserved as completely as possible, and it is guaranteed that feature information as complete as possible can be obtained through a subsequent point cloud data processing algorithm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention discloses a farthest point downsampling algorithm based on a multi-level adaptive grid with high universality, belonging to the technical field of data processing. Background Art

[0002] With the iteration of 3D point cloud hardware acquisition technology, the widespread application of 3D point cloud technology in fields such as robotics, autonomous driving, augmented reality and virtual reality, and the popularization of high-precision laser scanners, the volume of 3D point cloud data is growing, requiring higher-precision and real-time point cloud data processing methods. Downsampling, as a necessary step in point cloud data processing, can reduce computational pressure, remove redundant information, and maintain complete feature information. Among them, the global features of the point cloud are a very important part of the feature information. However, existing downsampling technologies find it difficult to strike a balance between high efficiency and the complete preservation of global features. When the downsampled point cloud data cannot retain complete global feature information, the accuracy of subsequent point cloud data processing methods such as point cloud registration and point cloud classification will drop significantly. Summary of the Invention

[0003] The purpose of the present invention is to provide a highly universal farthest point downsampling algorithm based on a multi-level adaptive grid to solve the problem in the prior art that the downsampled point cloud data cannot retain complete global feature information.

[0004] A highly universal farthest point downsampling algorithm based on a multi-level adaptive grid, including:

[0005] S1. For each grid point capacity N, obtain the number of idle CPU cores of the current processing device;

[0006] S2, dividing the original point cloud into multi-level grids step by step;

[0007] S3, for i grids Q i Perform sampling at the farthest point;

[0008] S4, if the number of points after the farthest point sampling is N t Satisfy N t > N, then repeat S1 to S3 until N is satisfied t <N;

[0009] S5. Perform the farthest point sampling again on the point cloud after the farthest point sampling.

[0010] The point capacity N in each grid is set to the product of the number of idle CPU cores and the inverse of the CPU frequency.

[0011] The original point cloud is divided into multi-level grids step by step. When the number of points retained in the previous level grid is greater than N, it is divided again until the number of points in the i-th grid is N.i Satisfy N i <N.

[0012] The sampling rate of the farthest point sampling of S3 is set to twice the target sampling rate.

[0013] The sampling rate of the farthest point sampling of S5 is set to be equal to the target sampling rate.

[0014] The farthest point sampling includes, given a point cloud set P:

[0015] P={p j ∈R 3 |i=1,...,M},

[0016] Where p j is the jth point, M is the total number of points, R 3 Represents three-dimensional space;

[0017] The farthest point sampling generates the sampling set S, Includes initialization and iterative sampling.

[0018] The initialization includes randomly selecting a seed point s0, s0∈P, and setting S={s0}.

[0019] The iterative sampling includes calculating the minimum Euclidean distance d(p, S) from each candidate point p to S, selecting the point s with the maximum distance and the minimum distance k Add S, update the set, and change S∪{s k} as the final S.

[0020] d(p, S) is:

[0021]

[0022] p∈P or S;

[0023] Where S is an element of S.

[0024] s k for:

[0025]

[0026] Compared with the existing technology, the present invention has the following beneficial effects: the present invention ensures the high efficiency of the downsampling algorithm and has obvious advantages in point cloud processing tasks with large-scale and high-precision requirements; it retains the global and local features of the point cloud as completely as possible, ensuring that the subsequent point cloud data processing algorithm can obtain the most complete feature information. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the overall flow chart of the present invention;

[0028] Figure 2 This is a schematic diagram of the farthest point sampling in the prior art;

[0029] Figure 3 This is a schematic diagram of multi-threaded computing;

[0030] Figure 4 This is a speed comparison chart for 12.5% ​​sampling rate;

[0031] Figure 5 This is a speed comparison chart for 6.25% sampling rate;

[0032] Figure 6 The speed comparison chart is for 3.125% sampling rate. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0034] Highly universal farthest point downsampling algorithm based on multi-level adaptive grid, such as Figure 1 ,include:

[0035] S1. For each grid point capacity N, obtain the number of idle CPU cores of the current processing device;

[0036] S2, dividing the original point cloud into multi-level grids step by step;

[0037] S3, for i grids Q i Perform sampling at the farthest point;

[0038] S4, if the number of points after the farthest point sampling is N t Satisfy N t > N, then repeat S1 to S3 until N is satisfied t <N;

[0039] S5. Perform the farthest point sampling again on the point cloud after the farthest point sampling.

[0040] The point capacity N in each grid is set to the product of the number of idle CPU cores and the inverse of the CPU frequency.

[0041] The original point cloud is divided into multi-level grids step by step. When the number of points retained in the previous level grid is greater than N, it is divided again until the number of points in the i-th grid is N. i Satisfy N i <N.

[0042] The sampling rate of the farthest point sampling of S3 is set to twice the target sampling rate.

[0043] The sampling rate of the farthest point sampling of S5 is set to be equal to the target sampling rate.

[0044] The farthest point sampling includes, given a point cloud set P:

[0045] P={p j ∈R 3 |i=1,...,M},

[0046] Where p j is the jth point, M is the total number of points, R 3 Represents three-dimensional space;

[0047] The farthest point sampling generates the sampling set S, Includes initialization and iterative sampling.

[0048] The initialization includes randomly selecting a seed point s0, s0∈P, and setting S={s0}.

[0049] The iterative sampling includes calculating the minimum Euclidean distance d(p, S) from each candidate point p to S, selecting the point s with the maximum distance and the minimum distance k Add S, update the set, and change S∪{s k} as the final S.

[0050] d(p, S) is:

[0051]

[0052] p∈P or S;

[0053] Where s is an element of S.

[0054] s k for:

[0055]

[0056] Point cloud data is a massive collection of three-dimensional spatial points formed by discrete sampling of the target surface. Each data point contains at least geometric coordinates in a three-dimensional Cartesian coordinate system and can be expanded to store multimodal attribute information (such as RGB color, normal vector, reflection intensity, timestamp, etc.). Its data acquisition method mainly relies on laser measurement (LiDAR) and multi-view photogrammetry technology, and has the characteristics of non-contact, high precision, and full-scene digital modeling. With the breakthrough progress in LiDAR hardware performance, the scale of single-frame point cloud data has exceeded one million sampling points, demonstrating irreplaceable technical value in cutting-edge fields such as point cloud classification, hierarchical management of large-scale point clouds, and browsing and display.

[0057] However, the exponential growth of point cloud data size presents dual challenges to core processing workflows (such as registration, segmentation, and object recognition): 1) the nonlinear growth of computational complexity, making it difficult to meet real-time requirements; and 2) a significant increase in storage and transmission costs. Against this backdrop, point cloud downsampling technology, a key enabling technology in the preprocessing phase, aims to achieve scientific reduction of data volume through intelligent dimensionality reduction strategies while maintaining the integrity of key geometric features, thereby improving the efficiency and robustness of subsequent algorithm processing.

[0058] In point cloud downsampling, the goal is to reduce the amount of point cloud data while preserving the geometric structure and key features as much as possible. Common downsampling methods include: First, regularized downsampling: random sampling, uniform sampling, and voxel sampling. These methods are simple to implement and highly efficient, but are prone to losing important edge or detail information. Second, downsampling based on geometric features: such as curvature sampling and farthest point sampling (FPS). These methods solve the problem of partial detail loss by preserving feature points in the point cloud, but the computational complexity is relatively high. Geometric feature-based downsampling algorithms (such as farthest point downsampling and curvature downsampling) are widely used in deep learning-based point cloud processing algorithms due to their excellent feature preservation effects. However, the FPS algorithm has become a standard preprocessing module for point cloud deep learning models due to its outstanding performance in uniform coverage of the feature space and topology preservation. However, its inherent greedy iteration mechanism limits the scalability of the algorithm, especially in engineering deployment under the CPU architecture, which faces severe challenges: existing optimization solutions are mostly limited to heterogeneous computing frameworks accelerated by GPU parallelization. Although they can improve local efficiency, they are limited by the tightly coupled characteristics of the preprocessing link and cannot meet the lightweight requirements of embedded systems and edge computing scenarios.

[0059] Therefore, innovative research is urgently needed: designing a novel iterative framework based on spatial segmentation and hierarchical sampling to overcome the performance bottleneck of traditional point-by-point distance calculations and balance computational efficiency with feature preservation; and building a memory access optimization model based on a CPU multi-level storage architecture to achieve deep synergy between algorithms and hardware features. Such breakthroughs will drive the evolution of point cloud processing technology towards low power consumption, high real-time performance, and strong generalization, providing core algorithmic support for edge intelligent devices in related fields.

[0060] Traditional Farthest Point Sampling (FPS) Principle Farthest point sampling is a classic geometric downsampling method that aims to select a portion of points with the most uniform spatial distribution from the point cloud. Figure 2First, randomly select a point in the point cloud as the initial point and add it to the sampling set. Calculate the minimum distance from the remaining points to the point in the sampling set, select the point with the farthest distance, and add it to the sampling set. Repeat the above process until the predetermined number of sampling points is reached. FPS performs well in preserving the edge features and geometric characteristics of point clouds. However, calculating the distance between points involves multiple traversals, and the amount of calculation increases exponentially with the number of point clouds. In addition, a large amount of point cloud feature information needs to be stored during processing, and a large number of constructor calls will be generated when calling related data elements, resulting in a large amount of memory usage and low operating efficiency.

[0061] Current FPS optimization efforts focus on hardware acceleration and algorithm tuning:

[0062] Hardware acceleration leverages the parallel computing capabilities of GPUs, using frameworks like CUDA to optimize FPS calculations. While this approach improves efficiency, it requires specific hardware support and is not universally applicable. QUICKFPS utilizes a bucket data structure to optimize runtime by reducing the number of point-to-point distance calculations and relying on CUDA for parallel operations, but its applicability is limited. Hardware acceleration relies on high-performance hardware and cannot be easily directly invoked by users.

[0063] Logical optimization is an optimization method that is widely used by researchers. There are various optimization modes, but the effects of the optimization schemes vary. In general, it is difficult to balance the advantages of FPS's global feature retention effect and the need for efficient computing.

[0064] The main idea behind the Farthest Point Downsampling (DFPS) algorithm is to determine the maximum number of points to retain in each grid Q based on the number of idle CPU cores on the current device, divide the original point cloud into a multi-level adaptive grid, and perform farthest point downsampling layer by layer. DFPS optimizes the way point cloud information is called within the multi-level adaptive grid division to reduce computational costs. First, it arrays the three-dimensional point cloud position information, reducing the number of constructor calls. Second, it establishes a correspondence between grid blocks and point clouds, enabling rapid search and call access to point clouds via the grid. This significantly reduces the need for excessive constructor calls, thereby improving computational efficiency, as shown in the example image. Compared to FPS_After, which only adds multi-threaded processing and point cloud information call optimization, the speed improvement is significant. Please refer to the attached speed comparison chart for details.

[0065] In addition, one of the design goals of DFPS is to implement a universal point cloud downsampling algorithm that does not perform heterogeneous operations, such as Figure 3, n represents the number of point cloud data in a grid. DFPS does not use GPU multi-threading operations such as CUDA, but uses OpenMP acceleration that supports multi-threaded computing on the CPU for the entire memory process. The present invention introduces a multi-level adaptive grid division method. By setting the maximum number of points retained in a single grid according to the number of idle CPU cores of the current device, the original point cloud is divided into multi-level grids, and multi-threading is used for acceleration, and downsampling is performed step by step. The present invention conducted three sets of experiments, and the speed comparison of the 12.5% ​​sampling rate was Figure 4 As shown, the 6.25% sampling rate speed comparison is Figure 5 As shown, the 3.125% sampling rate speed comparison is Figure 6 shown.

[0066] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A highly universal farthest point downsampling algorithm based on a multi-level adaptive grid, characterized by: include: S1. For each grid point capacity N, obtain the number of idle CPU cores of the current processing device; S2, dividing the original point cloud into multi-level grids step by step; S3, for i grids Q i Perform sampling at the farthest point; S4, if the number of points after the farthest point sampling is N t Satisfy N t >N, then repeat S1 to S3 until N is satisfied t <N; S5. Perform the farthest point sampling again on the point cloud after the farthest point sampling.

2. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 1 is characterized in that: The point capacity N in each grid is set to the product of the number of idle CPU cores and the inverse of the CPU frequency.

3. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 2 is characterized in that: The original point cloud is divided into multi-level grids step by step. When the number of points retained in the previous level grid is greater than N, it is divided again until the number of points in the i-th grid is N. i Satisfy N i <N.

4. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 3 is characterized in that: The sampling rate of the farthest point sampling of S3 is set to twice the target sampling rate.

5. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 4 is characterized in that: The sampling rate of the farthest point sampling of S5 is set to be equal to the target sampling rate.

6. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 5 is characterized in that: The farthest point sampling includes, given a point cloud set P: P={p j ∈R 3 |i=1,...,M}, Where p j is the jth point, M is the total number of points, R 3 Represents three-dimensional space; The farthest point sampling generates the sampling set S, Includes initialization and iterative sampling.

7. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 6 is characterized in that: The initialization includes randomly selecting a seed point s0, s0∈P, and setting S={s0}.

8. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 7 is characterized in that: The iterative sampling includes calculating the minimum Euclidean distance d(p,S) from each candidate point p to S, selecting the point s with the maximum distance and the minimum distance k Add S, update the set, and change S∪{s k } as the final S.

9. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 8, characterized in that: d(p,S) is: p∈P or S; Where s is an element of S.

10. The highly universal farthest point downsampling algorithm based on multi-level adaptive grid according to claim 9, characterized in that: s k for: