A multi-stage joint point cloud adaptive filtering method and system

Abstract

The application discloses a kind of multi-stage joint point cloud adaptive filtering method and system, adopt four-stage progressive denoising strategy: first, through octree voxel downsampling, adopt nearest neighbor original point to replace the center of mass of the voxel to retain geometric characteristics;Then, based on the modified Z-score and generalized median deviation GMAD, the automatic threshold rough denoising is realized by improving RANSAC, and large-scale outlier noise is removed;Again, through the second-order gradient edge-preserving fine denoising, adaptive smoothing is realized according to the first-order and second-order gradient to distinguish flat area, edge area and transition area;Finally, through the density-distance double-weighted Gaussian weighted fusion correction, the surface and corner geometric details are restored.The polished volume detection verification shows that the volume error of the filtered point cloud and the real point cloud is only 0.06%, with strong robustness, high fidelity and automatic adaptive advantages.

Description

Technical Field

[0001] This invention relates to the field of 3D laser point cloud data preprocessing technology, specifically to an adaptive point cloud filtering method and system that balances efficient noise reduction, edge enhancement, and geometric fidelity. It is suitable for dense point cloud data preprocessing in high-precision scenarios such as industrial inspection, 3D reconstruction, and grinding volume assessment. Background Technology

[0002] With the rapid development of 3D laser scanning technology, point cloud data has been widely used in industrial inspection, 3D reconstruction, reverse engineering, autonomous driving, and other fields. In industrial polishing scenarios, the quality of point cloud data directly determines the accuracy of subsequent volume measurement and quality assessment. Because polishing areas typically have high density and complex geometric structures, the raw point cloud obtained by scanning inevitably contains various types of noise, such as multipath reflection noise, surface disturbance noise, and boundary blurring noise. If these noises are not effectively removed, they will lead to significant deviations in subsequent volume calculations and 3D reconstruction.

[0003] Currently, point cloud denoising methods can be mainly classified into the following categories: (1) Statistical filtering methods. For example, filtering methods based on the statistical characteristics of neighborhood distance calculate the average neighborhood distance of each point and remove outliers that exceed a certain standard deviation multiple of the overall average distance. Although this type of method can initially remove some outlier noise, it is difficult to cope with complex environments where multiple noise sources coexist, and it is prone to over-smoothing when processing edge regions, resulting in the loss of geometric features.

[0004] (2) Radius filtering and clustering methods. For example, point cloud denoising methods based on local density and improved fuzzy C-means use parameter-adaptive radius filtering to achieve denoising. This type of method is sensitive to parameter settings and is difficult to simultaneously remove large-scale outlier noise and local dense noise.

[0005] (3) Fitting-based reconstruction methods. For example, Moving Least Squares (MLS) smoothly reconstructs the surface by fitting local polynomials. Although this type of method has the ability to smoothly reconstruct, it is prone to excessive smoothing of edges or loss of geometric features when dealing with large-scale noise, and the volume error can reach more than 14%.

[0006] (4) Multi-stage filtering method. Existing technologies include denoising methods that combine statistical outlier elimination with adaptive bilateral filtering, as well as 3D point cloud denoising methods based on multi-stage information interaction. However, existing methods still have shortcomings in multi-stage collaborative denoising, adaptive threshold calculation, and edge fidelity, making it difficult to fully preserve the geometric structure information of point clouds such as edges, corners, and surfaces while efficiently denoising.

[0007] (5) Improve the RANSAC method. Existing studies have combined the RANSAC algorithm with dynamic threshold estimation for point cloud filtering, but existing methods are mostly designed for specific scenarios, and threshold calculation still requires manual intervention or depends on specific scenario assumptions, lacking a universal fully automatic adaptive threshold mechanism.

[0008] In summary, existing single-source filtering methods are insufficient to handle complex environments with multiple noise sources, while algorithms based on fitting and reconstruction are prone to excessive edge smoothing or loss of geometric features when dealing with large-scale noise. Therefore, there is an urgent need for an adaptive point cloud filtering method that can efficiently remove multiple types of noise, completely preserve geometric structure, and requires no manual parameter tuning throughout the process. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by providing a multi-stage joint point cloud adaptive filtering method and system. Through a hierarchical progressive filtering strategy, different types of noise points are removed step by step. While efficiently removing noise, the geometric structure information such as edges, corners, and curved surfaces of the original point cloud is preserved to the maximum extent, significantly improving the accuracy and reliability of point cloud data. It is suitable for high-precision point cloud processing scenarios such as industrial grinding.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a multi-stage joint point cloud adaptive filtering method, comprising the following steps: S1. Acquire raw point cloud data. Obtain raw, dense point cloud data of industrial scenes using structured light scanners and 3D laser scanners.

[0011] S2. Octree voxel downsampling. Octree voxel downsampling is performed on the original point cloud to simplify the point cloud while preserving geometric features.

[0012] The octree voxel downsampling uses the nearest neighbor original point in Euclidean distance to replace the voxel center as the voxel output point. Specifically, a minimum bounding box is constructed for the point cloud P={p1,p2,,pn}: Set the voxel side length 'a', and calculate the voxel index of each point: For each voxel V, calculate the voxel centroid CV of the original point set PV={p1,p2,,pm}: Calculate the Euclidean distance dk from each origin point to the centroid, and select the origin point with the smallest distance as the voxel representative point: P *As voxel output points, feature-preserving downsampling is performed. This rule ensures that the downsampled point cloud comes entirely from the original scan points, without introducing virtual points, shifting positions, or damaging edges and corners.

[0013] S3. Improved RANSAC Adaptive Threshold Coarse Denoising. Improved RANSAC adaptive threshold coarse denoising is performed on the downsampled point cloud to automatically remove large-scale outlier noise points. Specifically, this includes: S31. Randomly select 3 points to construct a planar model Ax+By+Cz+D=0, and calculate the distance di from each point to the plane. S32. Calculate the distance to the median M = median(|di|) and the absolute median deviation MAD = median(|diM|), and then calculate the improved generalized median deviation: in, ; S33. Calculate the corrected Z-score: If |zi| > zs, it is identified as outlier noise and removed.

[0014] S4. Second-order gradient edge-preserving fine denoising. Second-order gradient edge-preserving fine denoising is performed on the point cloud after coarse denoising, distinguishing between flat and edge regions to achieve adaptive smoothing. Specifically, this includes: S41. Construct a local neighborhood for each point and calculate the magnitude of the first-order gradient. Second gradient and the maximum gradient difference in the neighborhood ; S42. Denoising is performed according to the following three-level adaptive filtering rules: Flat area: when When using neighborhood mean smoothing; Edge region: when At this time, the pixel value remains unchanged, forcibly preserving the edge; Transition region: when the gradient exceeds the limit in only a single direction and n left n right When the time comes, the mean of the side with more points is updated.

[0015] S5, Density-Distance Dual-Weight Correction. Density-distance dual-weight correction is performed on the finely denoised point cloud to restore the geometric details of surfaces and edges. Specifically, this includes: S51. Calculate the deviation of each point from the neighborhood mean: S52. Calculate weights based on Gaussian distribution: Where is the Gaussian kernel parameter; S53. Merge neighboring points by weight and output the final result: S6. Output the final denoised point cloud. Output a structure-preserving, high-precision final denoised point cloud.

[0016] The present invention also provides a multi-stage joint point cloud adaptive denoising system, comprising: The data input module is used to acquire raw point cloud data; The octree downsampling module is used to perform the octree voxel downsampling; An improved RANSAC coarse denoising module is used to perform the improved RANSAC adaptive threshold coarse denoising. The second-order gradient edge-preserving filtering module is used to perform the second-order gradient edge-preserving fine noise reduction. A density-distance dual-weight correction module is used to perform the density-distance dual-weight correction; Point cloud output module, used to output the final denoised point cloud; The octree downsampling module, the improved RANSAC coarse denoising module, the second-order gradient edge-preserving filtering module, and the density-distance dual-weight correction module are cascaded in sequence.

[0017] Compared with the prior art, the present invention has the following beneficial effects: A four-stage progressive denoising approach ensures thorough noise removal. It employs an octree voxel downsampling-improved RANSAC adaptive coarse denoising strategy with second-order gradient edge-preserving fine denoising and density-distance dual-weight correction. This progressive strategy removes large-scale outlier noise, locally dense noise, and boundary noise at each stage, with clear division of labor and efficient collaboration between stages.

[0018] It is fully adaptive and requires no manual parameter tuning. The improved RANSAC stage achieves fully automatic threshold calculation by correcting the Z-score and generalized median deviation (GMAD), overcoming the shortcomings of traditional methods that rely on manual threshold setting, and is suitable for batch processing in industrial production lines.

[0019] The system prioritizes structural preservation, ensuring complete retention of geometric features. Octree downsampling uses nearest-neighbor real points to replace the volume center, avoiding the introduction of virtual points; second-order gradient edge-preserving filtering accurately distinguishes flat areas, edge areas, and transition areas using first-order / second-order gradients; dual-weight correction further restores surface and angular details.

[0020] Highly efficient and stable. Octree voxel downsampling can significantly reduce the number of point clouds, improve the efficiency of subsequent processing, and is suitable for pipeline operations.

[0021] The accuracy is significantly better than existing methods. Verification experiments on point cloud denoising in industrial grinding areas show that the error between the denoised point cloud volume and the actual point cloud volume using the method of this invention is only 0.06%, far superior to single statistical filtering and MLS moving least squares method. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the multi-stage joint point cloud adaptive filtering method of the present invention. Figure 2 This is a schematic diagram of the original point cloud. Figure 3 This is a schematic diagram of a real point cloud (baseline). Figure 4 This is a schematic diagram of the point cloud after a single statistical filter. Figure 5 This is a schematic diagram of the point cloud after MLS moving least squares filtering; Figure 6 This is a schematic diagram of the point cloud after filtering by the algorithm of this invention. Detailed Implementation

[0023] The present invention will now be further described with reference to the accompanying drawings.

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] Example: Comparative Experiment on Point Cloud Noise Reduction Effect in Industrial Grinding Areas This embodiment uses a single statistical filter, MLS moving least squares method, and the four-stage joint denoising algorithm of this invention to process the original point cloud of the industrial grinding area. By calculating the volume error between the denoised point cloud volume and the real point cloud (reference point cloud), the geometric structure fidelity of this algorithm is verified.

[0026] Point cloud data description Point cloud type Points (pieces) Volume (mm) illustrate Original point cloud 28933 206.025 Structured light scanning acquisition, including multipath reflection noise, surface disturbance noise, and boundary blurring noise. Real point cloud (benchmark) 25550 161.304 Noise-free reference point cloud after manual calibration and high-precision denoising Experimental environment Hardware: Intel i9-13900HX CPU, 32GB RAM Software: PCL 1.14.1, VS2022, C++ implementation Volume calculation method: A unified volume calculation method based on layered convex profiles is used. Comparison algorithm and parameter configuration Single statistical filtering (Comparison Group 1) Parameter configuration: number of neighboring points meanK=50, standard deviation multiple threshold stddevMulThresh=1.0.

[0027] Implementation steps: Input the original point cloud; construct a kd-tree and search for 50 nearest neighbors for each point; calculate the average neighborhood distance of each point and calculate the overall average distance and standard deviation; remove points whose average neighborhood distance exceeds the overall average distance + 1.0 standard deviation; output the processed point cloud and calculate the volume and error.

[0028] Experimental results: 26,681 points were obtained after treatment; processing time was 15.669 s; and the volume was 171.198 mm². 3 The volume error was 6.13%.

[0029] (2) MLS moving least squares method (comparison group 2) Parameter configuration: search radius 2.0mm, polynomial order 2, projection iterations 3.

[0030] Implementation steps: Input the original point cloud; construct a local neighborhood with a radius of 2.0 mm for each point; perform second-order polynomial fitting on the neighborhood points to construct a local surface; project the original points onto the fitted surface and update the 3D coordinates of the points; output the smoothed point cloud and calculate the volume and error.

[0031] Experimental results: 27,421 points were processed; processing time was 11.971 s; and the volume was 184.541 mm². 3 The volume error was 14.41%.

[0032] (3) The four-stage joint denoising algorithm of the present invention (experimental group) Parameter configuration stage parameter Value Octree downsampling voxel side length a 0.5mm Improved RANSAC Number of plane fitting iterations 100 times Improved RANSAC Z-score threshold z_s 3 Second-order gradient edge-preserving filter <![CDATA[Flat region threshold 1]]> 0.05 Second-order gradient edge-preserving filter <![CDATA[Edge region threshold 2]]> 0.02 Second-order gradient edge-preserving filter <![CDATA[Transition zone threshold 3]]> 0.08 Double weighting correction Gaussian kernel parameters 0.01mm Implementation steps: Phase 1: Construct the minimum bounding box of the point cloud, set the voxel side length to 0.5mm; calculate the centroid within each voxel, select the original point closest to the centroid as the representative point; output the downsampled point cloud.

[0033] Phase 2: Randomly select 3 points to fit the planar model, calculate the distance from all points to the plane; calculate GMAD and corrected Z-score based on the distance distribution, adaptively identify and remove outliers; output coarsely denoised point cloud.

[0034] Phase 3: Calculate the first and second gradients of the neighborhood point by point, and distinguish between flat areas, edge areas and transition areas; flat areas are smoothed using the neighborhood mean, edge areas keep the original points unchanged, and transition areas are updated with weighted averages based on the number of neighborhood points; output edge-preserving and noise-reduced point clouds.

[0035] Stage 4: Calculate the deviation of each point from the neighborhood mean, construct a density-distance dual weight based on Gaussian distribution; fuse neighborhood points according to weight, correct the surface and minor features; output the final denoised point cloud.

[0036] Experimental results: 25,564 points were processed; processing time was 16.789 s; and the volume was 161.557 mm². 3 The volume error is 0.06%.

[0037] 4. Comparison and Analysis of Experimental Results method Volume (mm) Volume error (%) Core issues / advantages Original point cloud 206.025 27.89 The noise is severe, and the volume calculation has a huge deviation. Single statistical filtering 171.198 6.13 It can remove some outliers, but excessive edge smoothing results in an inflated volume. MLS Moving Least Squares Method 184.541 14.41 Poor surface smoothing, easy edge expansion, and large volumetric error. Algorithm of this invention 161.557 0.06 With an error of only 0.06%, it is almost indistinguishable from real point clouds, balancing noise reduction and structure preservation. 5. Experimental Conclusions (1) Fidelity advantage: The volume error of the present invention is only 0.06%, which is much lower than that of the comparison algorithm, proving that it can preserve the geometric structure of the polishing area to the maximum extent during the denoising process.

[0038] (2) Robustness advantage: The algorithm parameters do not need to be manually adjusted repeatedly, and it can adaptively handle different types of noise, making it suitable for batch processing in industrial production lines.

[0039] (3) Efficiency balance: The processing time is slightly higher than that of the comparison algorithm, but the error is reduced by more than 100 times, which has significant practical value in high-precision industrial inspection scenarios.

[0040] 6. Parameter Selection Guide Voxel side length a: ranges from 0.3 to 0.8 mm, with smaller values ​​for high precision requirements and larger values ​​for high processing efficiency requirements; The score threshold z_s ranges from 2.5 to 3.5; the smaller the value, the stronger the noise reduction. Flat area threshold 1: The value ranges from 0.03 to 0.08. The smaller the value, the stricter the judgment of flat areas. Edge zone threshold 2: The value ranges from 0.01 to 0.05, and the smaller the value, the stronger the edge protection; Transition zone threshold 3: value range 0.05~0.12; Gaussian kernel parameter: The value ranges from 0.005 to 0.02 mm. The smaller the value, the lower the correction intensity.

[0041] 7. Industrial Application Examples Industrial grinding volume assessment: The point cloud of the grinding area is filtered to obtain high-precision point cloud data for volume calculation. 3D reconstruction: providing high-quality point cloud data for the digitization of cultural relics and the modeling of buildings; Reverse engineering: providing structure-preserving point cloud input for CAD model reconstruction; Industrial inspection: Provides a reliable data foundation for component dimensional measurement and defect detection.

[0042] The above description is only used to illustrate the technical solution of the present invention and is not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention, as long as they do not depart from the spirit and scope of the technical solution of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A multi-stage joint point cloud adaptive filtering method, characterized in that, Includes the following steps: S1. Acquire raw point cloud data; S2. Perform octree voxel downsampling on the raw point cloud to simplify the point cloud while preserving geometric features; S3. Perform improved RANSAC adaptive threshold coarse denoising on the downsampled point cloud to automatically remove large-scale outlier noise points; S4. Perform second-order gradient edge-preserving fine denoising on the coarsely denoised point cloud to distinguish between flat and edge regions and achieve adaptive smoothing; S5. Perform density-distance dual-weight correction on the finely denoised point cloud to restore surface and angular geometric details; S6. Output the final denoised point cloud.

2. The multi-stage joint point cloud adaptive filtering method according to claim 1, characterized in that... The octree voxel downsampling uses the nearest neighbor original point in Euclidean distance to replace the voxel centroid as the voxel output point. Specifically, it includes: constructing a minimum bounding box for the point cloud and setting the voxel side length a, and calculating the voxel index of each point; for the original point set P_V={p_1,p_2,...,p_m} within each voxel V, calculating the voxel centroid C_V=(1 / m(k=1)^m x_k, 1 / m(k=1)^m y_k, 1 / m(k=1)^m z_k); selecting the original point with the smallest Euclidean distance from the centroid C_V within the voxel as the voxel representative point p output, i.e., p=argmin{p_kP_V}‖p_k-C_V‖.

3. The multi-stage joint point cloud adaptive filtering method according to claim 1, characterized in that, The improved RANSAC adaptive threshold coarse denoising achieves fully automatic threshold calculation by correcting the Z-score and generalized median deviation (GMAD). Specifically, it includes: randomly selecting 3 points to construct a planar model, calculating the distance d_i from each point to the plane; calculating the median distance M = median(|d_i|) and the median absolute deviation (MAD) = median(|d_i-M|), and then calculating GMAD = MAD(1-(V-n / 2) / V2), where V0 = e^(-u_i) 2 / 2), V2=u_i 2 e^(-u_i 2 / 2), u_i=|d_i-| / S_n^(0); calculate the corrected Z-score: z_i=(d_i-) / GMAD, if |z_i|>z_s, it is determined to be outlier noise and removed.

4. The multi-stage joint point cloud adaptive filtering method according to claim 1, characterized in that, The second-order gradient edge-preserving fine-tuning denoising method uses the first-order gradient to determine flatness and the second-order gradient to determine edge abrupt changes. Specifically, it includes: constructing a local neighborhood for each point, calculating the first-order gradient magnitude, the second-order gradient, and the maximum gradient difference _i,j^max in the neighborhood; and performing denoising according to a three-level adaptive filtering rule: when _i,j^max < 1, it is determined to be a flat region, and neighborhood mean smoothing is used; when max(| 2 When |) < 2, it is determined to be an edge region, and the pixel value remains unchanged; when only a single direction gradient exceeds the limit and n_left n_right, it is determined to be a transition region, and the average value of the side with more points is updated.

5. The multi-stage joint point cloud adaptive filtering method according to claim 1, characterized in that, The density-distance dual-weight correction employs Gaussian weighted fusion, specifically including: calculating the deviation of each point from the neighborhood mean p_k=|f(x_k,y_k)-f(x_k,y_k)|; and calculating the weight q_k=exp(-p_k) based on a Gaussian distribution. 2 / (2 2 )) / (2 2 The final result is output by f(i,j)=q_kf(x_k,y_k) / q_k after merging neighborhood points according to weight.

6. The multi-stage joint point cloud adaptive filtering method according to any one of claims 1-5, characterized in that, In step S2, the value of the voxel side length a ranges from 0.3 mm to 0.8 mm; in step S3, the value of the Z-score threshold z_s ranges from 2.5 to 3.5; in step S4, the value of threshold 1 ranges from 0.03 to 0.08, and the value of threshold 2 ranges from 0.01 to 0.05; in step S5, the value of the Gaussian kernel parameter ranges from 0.005 mm to 0.02 mm.

7. The multi-stage joint point cloud adaptive filtering method according to claim 6, characterized in that, The voxel side length a is 0.5 mm, the Z-score threshold z_s is 3, the threshold 1 is 0.05, the threshold 2 is 0.02, and the Gaussian kernel parameter is 0.01 mm.

8. The multi-stage joint point cloud adaptive filtering method according to claim 4, characterized in that, The local neighborhood is a 33-neighborhood, and the smoothing of the flat region adopts 8-neighborhood mean smoothing, with the formula f(i,j)=1 / 8_(m,n=-1)1 f(i+m,j+n); the transition region is also set with a threshold 3. When it is determined to be a transition region, if the number of second-order gradient over-limit directions is equal to 1 and n_leftn_right, then the mean of the side with more points is updated. The value range of the threshold 3 is 0.05 to 0.

12.

9. A multi-stage joint point cloud adaptive denoising system, characterized in that, include: The data input module is used to acquire raw point cloud data; An octree downsampling module is used to perform octree voxel downsampling as described in claim 2; An improved RANSAC coarse denoising module is provided for performing the improved RANSAC adaptive threshold coarse denoising as described in claim 3. The second-order gradient edge-preserving filtering module is used to perform the second-order gradient edge-preserving fine noise reduction as described in claim 4; A density-distance dual-weight correction module is used to perform the density-distance dual-weight correction as described in claim 5; Point cloud output module, used to output the final denoised point cloud; The octree downsampling module, the improved RANSAC coarse denoising module, the second-order gradient edge-preserving filtering module, and the density-distance dual-weight correction module are cascaded in sequence.

10. The application of the method according to any one of claims 1 to 5 in industrial grinding volume assessment, three-dimensional reconstruction or reverse engineering.

Images