Point cloud lightweight method for noise robustness and detail fidelity

Abstract

The invention discloses a noise robustness and detail fidelity-oriented point cloud lightening method, which relates to the technical field of three-dimensional point cloud processing, and comprises the following steps: S1, carrying out multi-scale noise suppression on original point cloud data by adopting a self-adaptive sliding window polynomial fitting filtering algorithm to obtain smooth point cloud data; s2, constructing a spatial index structure for the smooth point cloud data by adopting a multi-level axial bounding box adaptive recursive subdivision method, and realizing sparse region memory optimization and high-density region feature retention; s3, extracting a skeleton structure from the point cloud data processed in the step S2 based on an improved central axis transformation algorithm; s4, performing normal estimation and correction on the point cloud data processed in the step S3 so as to retain and enhance sharp features, by means of the adaptive sliding window polynomial fitting filtering method, key details and feature distribution of a local geometric structure of the point cloud are reserved while noise interference is effectively suppressed, and the method has high robustness. And high-fidelity point cloud data is provided for subsequent normal estimation and three-dimensional reconstruction.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud processing technology, and in particular to a lightweight point cloud method that aims for noise robustness and detail fidelity. Background Technology

[0002] Point cloud data in complex terrain scenes commonly suffers from noise interference, uneven density distribution, and blurred details, which restricts the accuracy of semantic segmentation and reconstruction efficiency. Traditional lightweight point cloud methods struggle to balance noise robustness, detail integrity, and computational efficiency: existing denoising methods easily destroy local differential structures, making it difficult to distinguish noise from feature points; skeleton extraction algorithms are sensitive to noise, often producing pseudo-branches and breaks, affecting topological consistency; while uniform spatial partitioning strategies ignore density heterogeneity, leading to wasted memory in sparse areas and loss of features in high-density areas, limiting their effectiveness in high-precision applications such as 3D geographic information modeling.

[0003] However, existing methods still suffer from significant drawbacks when processing ground features, such as noise sensitivity, feature imbalance, and computational inefficiency. (1) Point clouds generally contain non-uniform noise. Existing denoising methods cannot avoid destroying the local differential structure of point clouds, making it impossible to accurately distinguish between "noise points" and "detail feature points" during lightweight processing. (2) Although the analysis method based on geometric features can partially compensate for the shortcomings of the distance metric method in noise resistance through local attribute analysis, its computational efficiency and ability to preserve geometric details still need to be systematically improved through the structured optimization strategy of spatial partitioning method. (3) In large-scale point cloud processing, traditional spatial partitioning strategies ignore the density heterogeneity of point clouds. Uniform bounding box partitioning leads to a waste of memory resources in sparse regions, while the geometric feature retention rate in high-density regions is seriously insufficient. To address these issues, a lightweight point cloud quantization method oriented towards noise robustness and detail fidelity is proposed. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the prior art by proposing a lightweight point cloud method that is both noise robust and detail-preserving.

[0005] A lightweight point cloud simplification method for noise robustness and detail preservation includes the following steps: S1. An adaptive sliding window polynomial fitting filtering algorithm is used to suppress noise at multiple scales in the original point cloud data to obtain smooth point cloud data. S2. A multi-level axial bounding box adaptive recursive subdivision method is used to construct a spatial index structure for smooth point cloud data to achieve memory optimization in sparse regions and feature preservation in high-density regions. S3. Extract the skeleton structure from the point cloud data processed in step S2 based on the improved median transformation algorithm; S4. Perform normal estimation and correction on the point cloud data processed in step S3 to preserve and enhance sharp features.

[0006] Preferably, in step S1, the multi-scale noise suppression using an adaptive sliding window polynomial fitting filtering algorithm includes: Define the sliding window size With polynomial order Satisfying the constraints ,in Using empirical formulas Sure; For each point in the point cloud , construct a containing A data window containing points, window width polynomial order , ; For the center point of the window Perform PCA analysis, construct a local tangent plane, and project neighborhood points onto this local coordinate system; Construct the data matrix and perform least squares fitting, then calculate the window center point. The filtered value; A multi-resolution processing mechanism is introduced to dynamically adjust the filtering results of different scale windows through coarse-scale weights and fine-scale weights; A two-level window strategy is adopted in the feature edge region to dynamically optimize the filtering parameters.

[0007] Preferably, the introduction of a multi-resolution processing mechanism, which dynamically adjusts the filtering results of different scale windows through coarse-scale weights and fine-scale weights, includes: Coarse-scale weights:

[0008] in, Indicates coarse-scale weights. This represents the variance of the data within the coarse-scale window. It is a very small constant; Fine-scale weights:

[0009] in, Indicates fine-scale weights. This represents the gradient magnitude within a fine-scale window.

[0010] Preferably, in step S2, the multi-level axial bounding box adaptive recursive partitioning method includes: Calculate the minimum and maximum values ​​of the point cloud projections on each coordinate axis, determine the bounding box vertices, scan point by point and iteratively update the extreme values ​​to obtain the boundary of the minimum bounding box; The minimum bounding box is recursively divided into multiple sub-bounding boxes. For each sub-bounding box, a weight factor is introduced, and points are selected and retained by combining distance and weight. Calculate the sub-bounding box weight entropy value based on the weight distribution within the sub-bounding box. The segmentation granularity is dynamically adjusted; if the entropy value is higher than the dynamic threshold... If the sub-bounding boxes are not recursively split, continue recursively; otherwise, stop splitting.

[0011] Preferably, the dynamic threshold is:

[0012] in, The average entropy value of the current level bounding box. The standard deviation of the entropy value. and This is an empirical coefficient.

[0013] Preferably, in step S3, extracting the skeleton structure from the point cloud data based on the improved median transformation algorithm includes: locating the distance field maxima point by combining dynamic radius neighborhood search, and processing each point... Calculate its distance to the nearest set of surface points. Perform central axis screening if If a value is a local maximum, it is marked as a central axis point. Curvature-weighted sampling is used to connect the central axis points to form a simplified skeleton, and sampling weights are assigned along the skeleton according to curvature.

[0014] Preferably, the central axis transformation algorithm uses the moving average method to initially smooth the central axis point cloud, performs secondary refinement based on Gaussian filtering, uses the Douglas-Peucker algorithm for topology simplification, and dynamically retains key feature points and removes redundant nodes by iteratively calculating the maximum deviation between the sampling points and the baseline.

[0015] Preferably, in step S4, the normal estimation and correction of the smoothed point cloud data to preserve and enhance sharp features includes: Calculate the geometric feature statistics within the local neighborhood and filter out sharp feature points; A feature propagation path is constructed based on dihedral angle geometric constraints, and the normal vector direction of the path nodes is corrected through backpropagation.

[0016] Preferably, the feature propagation path construction based on dihedral angle geometric constraints includes: iteratively extending along the direction of the edge vector of the candidate point's neighborhood, simultaneously detecting the consistency of the normal vector direction of consecutive extension points, and determining that the propagation has reached the stable region of the plane and terminating the propagation when the difference in normal vectors between adjacent extension points is lower than a preset tolerance. The formula for checking the consistency of normal vector direction is:

[0017] in, , This represents the normal vector of adjacent extension points in the propagation path. This represents the tolerance threshold for the difference between normal vectors.

[0018] Compared with existing technologies, the advantages of this invention are: 1. This invention uses an adaptive sliding window polynomial fitting filtering method to effectively suppress noise interference while preserving key details and feature distributions of the local geometric structure of the point cloud, providing high-fidelity point cloud data for subsequent normal estimation and 3D reconstruction.

[0019] 2. The present invention improves the central axis transformation algorithm, which effectively ensures the integrity of the geometric features of the ground features and the connectivity of the ground feature topology, and significantly reduces the risk of redundant branches and breaks in the ground feature skeleton.

[0020] 3. This invention develops an adaptive recursive spatial partitioning algorithm based on multi-level axial bounding boxes to achieve adaptive partitioning of multi-scale spatial constraint domains and efficient structured recombination of point cloud distributions. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the principle of the present invention.

[0022] Figure 2 This is a flowchart illustrating the segmentation process of the adaptive sliding window polynomial fitting filtering algorithm of the present invention.

[0023] Figure 3 This is a flowchart of the adaptive recursive partitioning process of the multi-level axial bounding box of the present invention.

[0024] Figure 4 The flowchart shows the skeleton extraction process of the improved central axis transformation algorithm of this invention. Detailed Implementation

[0025] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0026] Reference Figure 1-2 As shown, a lightweight point cloud quantization method for noise robustness and detail preservation includes the following steps: S1. An adaptive sliding window polynomial fitting filtering algorithm is used to suppress noise at multiple scales in the original point cloud data to obtain smooth point cloud data. S2. A multi-level axial bounding box adaptive recursive subdivision method is used to construct a spatial index structure for smooth point cloud data to achieve memory optimization in sparse regions and feature preservation in high-density regions. S3. Extract the skeleton structure from the point cloud data processed in step S2 based on the improved median transformation algorithm; S4. Perform normal estimation and correction on the point cloud data processed in step S3 to preserve and enhance sharp features.

[0027] In step S1, the multi-scale noise suppression using the adaptive sliding window polynomial fitting filtering algorithm includes: Define the sliding window size With polynomial order Satisfying the constraints ,in Using empirical formulas Sure; For each point in the point cloud , construct a containing A data window containing points, window width polynomial order , ; For the center point of the window Perform PCA analysis, construct a local tangent plane, and project neighboring points onto this local coordinate system to eliminate normal dimension noise; Construct the data matrix and perform least squares fitting, then calculate the window center point. The filtered value; A multi-resolution processing mechanism is introduced, which dynamically adjusts the filtering results of different scale windows through coarse-scale weights and fine-scale weights, including: Coarse-scale weights:

[0028] in, Indicates coarse-scale weights. This represents the variance of the data within the coarse-scale window. It is a very small constant; Fine-scale weights:

[0029] in, Indicates fine-scale weights. This represents the gradient magnitude within a fine-scale window; A two-level window strategy is adopted in the feature edge region to dynamically optimize the filtering parameters.

[0030] In step S2, the multi-level axial bounding box adaptive recursive subdivision method includes: Calculate the minimum and maximum values ​​of the point cloud projections on each coordinate axis, determine the bounding box vertices, scan point by point and iteratively update the extreme values ​​to obtain the boundary of the minimum bounding box; Recursively divide the minimum bounding box into multiple sub-bounding boxes. For each sub-bounding box, a weight factor is introduced, and points are selected and retained by combining distance and weight. Calculate the sub-bounding box weight entropy value based on the weight distribution within the sub-bounding box. The segmentation granularity is dynamically adjusted; if the entropy value is higher than the dynamic threshold... If the sub-bounding boxes are not recursively split, continue recursively; otherwise, stop splitting.

[0031] The dynamic threshold:

[0032] in, The average entropy value of the current level bounding box. The standard deviation of the entropy value. and This is an empirical coefficient.

[0033] In step S3, extracting the skeleton structure from the point cloud data based on the improved median transformation algorithm includes: locating the distance field maxima point by combining dynamic radius neighborhood search, and processing each point... Calculate its distance to the nearest set of surface points. Perform central axis screening if If a value is a local maximum, it is marked as a central axis point. Curvature-weighted sampling is used to connect the central axis points to form a simplified skeleton, and sampling weights are assigned along the skeleton according to curvature.

[0034] The central axis transformation algorithm uses the moving average method to initially smooth the central axis point cloud, and then performs secondary refinement based on Gaussian filtering. It uses the Douglas-Peucker algorithm for topology simplification, and dynamically retains key feature points and removes redundant nodes by iteratively calculating the maximum deviation between the sampling points and the baseline.

[0035] In step S4, normal estimation and correction are performed on the smoothed point cloud data to preserve and enhance sharp features, including: Calculate the geometric feature statistics within the local neighborhood and filter out sharp feature points; The feature propagation path is constructed based on the dihedral angle geometric constraint. The normal vector direction of the path node is corrected by backpropagation, so as to achieve a continuous and smooth transition of the normal field at the feature edge.

[0036] The feature propagation path constructed based on dihedral angle geometric constraints includes: iteratively extending along the neighborhood edge vector direction of the candidate point (step size equals neighborhood radius), simultaneously detecting the consistency of the normal vector direction of continuous extension points, and determining that the propagation has reached the stable region of the plane and terminating the propagation when the difference in the normal vectors of adjacent extension points is lower than the preset tolerance. The formula for checking the consistency of normal vector direction is:

[0037] in, , This represents the normal vector of adjacent extension points in the propagation path. This represents the tolerance threshold for the difference between normal vectors.

[0038] Example S1, Adaptive sliding window polynomial fitting filter instance segmentation An adaptive sliding window polynomial fitting filtering algorithm is used for multi-scale noise suppression. First, the sliding window size is defined. With polynomial order The constraints are satisfied:

[0039] in Using empirical formulas Sure.

[0040] Then set the window width to... The order of the polynomial is And satisfy ( For each point in the point cloud. , construct a containing A data window containing points, where M is the half-width of the window and the center point of the window is... .matrix The line, number Column elements are That is, each point within the window Power of 1.

[0041] For the center point of the window Perform PCA analysis:

[0042] in, Represents the covariance matrix. Indicates the number of points within the window. It is the center point of the window index, For the first in the window Feature vectors of points, For the set of points within the window The mean, It is the vector after removing the mean. This represents the transpose of a matrix.

[0043] Take the first two principal component vectors Construct a tangent plane and project neighboring points onto the local coordinates:

[0044] Solve the system of linear equations Least squares solution:

[0045]

[0046] in, For data matrix, A vector of data points within the window. For polynomial coefficients, Represent the transpose of a matrix. Find the... It is necessary to multiply both sides of the equals sign by . The inverse of the matrix is ​​obtained, but this inverse matrix may not exist. Therefore, a pseudo-inverse is used as a substitute, calculated by... pseudo-reversal This achieves optimal separation between noise distribution and geometric features.

[0047] Using the obtained polynomial coefficients Calculate the center point of the window Filtered value:

[0048] in, Indicates the filtered value. These are the polynomial fitting coefficients. This represents the position coordinates of the data points within the window relative to the center point.

[0049] A multi-resolution processing mechanism is introduced, and the filtering results of different scale windows are dynamically adjusted by designing weighting coefficients, so that noise suppression at the coarse scale and feature preservation at the fine scale can be effectively complementary: Coarse-scale weights:

[0050] in, Indicates coarse-scale weights. This represents the variance of the data within the coarse-scale window. It is a very small constant used to prevent the denominator from being zero.

[0051] Fine-scale weights:

[0052] in, Indicates fine-scale weights. This represents the gradient magnitude within a fine-scale window.

[0053] A two-level window strategy is employed in the feature edge region to dynamically optimize the filtering parameters, thereby balancing the requirements of noise suppression and feature preservation.

[0054] in, , , , This represents the filtered output value.

[0055] Finally, move the window one point to the right and repeat the above steps.

[0056] S2, Multi-level AABB Adaptive Recursive Decomposition To construct an efficient spatial index structure, a multi-level axial bounding box adaptive recursive partitioning method is adopted. For example... Figure 3 As shown, the starting point of the geometric center of the bounding box and each edge are determined based on Euclidean distance weighting: (1) Calculate the minimum and maximum values ​​of the point cloud projections on each coordinate axis:

[0057] (2) The vertices of the bounding box are:

[0058] Scan point by point and iteratively update the extreme values:

[0059]

[0060]

[0061] Among them, superscript express No. Spatial extremum coordinates at the next iteration It is the first The geometric importance weighting coefficient of each point Indicates the first The process processed during the nth iteration The original points Coordinate values. After the traversal is complete, by and Determine the diagonal vertices to obtain the boundary of the minimum bounding box.

[0062] If the point cloud is too large, recursively divide the minimum bounding box into sub-bounding boxes. For each sub-bounding box... Introducing weighting factors Based on a combination of distance and weight, select the retention points:

[0063] in, This represents the points retained within the sub-bounding box after weighted filtering. It is a point Weighting factors The center point of the child bounding box.

[0064] Subsequently, based on the weight distribution within the sub-bounding boxes, the segmentation granularity is dynamically adjusted to optimize computational resource allocation. The sub-bounding box weight entropy value is then calculated. :

[0065] in, Any point within the bounding box The weight, That is, all these points (including points) The sum of the weights of the bounding boxes themselves, and the entropy of the sub-bounding box weights. The higher the value, the more complex the weight distribution.

[0066] Dynamic threshold The core parameter controlling the adaptive adjustment of bounding box segmentation granularity is determined based on a local curvature entropy model, and is adaptively generated by quantizing the disorder of the weight distribution and the heterogeneity of point cloud density.

[0067] in, The average entropy value of the current level bounding box. The standard deviation of the entropy value. and This is an empirical coefficient. If... Continue recursively segmenting the sub-bounding boxes; otherwise, stop segmenting. If the weight difference within a sub-bounding box exceeds a threshold, check its internal weight distribution; specifically, calculate the coefficient of variation of the weights of all points within that bounding box.

[0068] in, This represents the standard deviation of the weights of the points within the sub-bounding box. This represents the mean weight of the points within the sub-bounding box. Greater than a preset mutation threshold If the weight difference within the bounding box is significant, non-uniform segmentation is triggered, and the high-weight region is segmented first.

[0069] S3. Improved central axis transformation algorithm for skeleton extraction. In skeleton extraction, an improved median transformation algorithm is used, such as... Figure 4 As shown, the extraction steps are as follows: First, the distance field is calculated for each point. Calculate the distance from this point to the nearest set of surface points. Distance:

[0070] Perform central axis screening if If a value is a local maximum within its neighborhood, it is marked as a central point. Curvature-weighted sampling is then performed, and the central points are connected to form a simplified skeleton. Sampling weights are then assigned along the skeleton according to curvature. It is the surface of a point cloud object. Any candidate point on the [the map].

[0071]

[0072] In the formula, For point curvature at that point Represents the sampling probability and point sampling weight value Proportional. Search for local maxima in the obtained distance field; these points are the central axis points of the point cloud.

[0073] Based on the above selection results of the central axis candidate points, the topology structure is generated as follows: First, a neighborhood traversal strategy is adopted to compare the distance field value of each sampling point with its neighboring nodes. The set of points that satisfy the local maximum condition is defined as the central axis candidate point. Then, when the distance between candidate points meets the threshold condition and the cosine value of the normal angle is greater than the set threshold, the vertex connection relationship is established to generate a geometrically robust central axis curve network and a topology-preserving central axis surface structure.

[0074] In terms of skeleton optimization, the moving average method is used to initially smooth the central axis point cloud, and the node coordinates are updated based on the arithmetic mean of the sample points in the neighborhood window to suppress high-frequency noise. Then, Gaussian filtering is used for secondary refinement, that is, the weight contribution of the nearest neighbor nodes is enhanced based on the spatial distance weighting strategy to eliminate the interference of discrete outliers. Finally, the Douglas-Peucker algorithm is used for topology simplification, and finally a simplified geometric representation of the central axis is obtained.

[0075]

[0076] In the formula, σ represents the standard deviation of the adaptive Gaussian filter kernel function. The weight of the current edge. This represents the weight of the edge containing the root node of the skeleton. The main radius is [missing information].

[0077] S4. Normal Estimation and Correction Although sliding window filtering can denoise and smooth point cloud data, it may affect the preservation of sharp features. Therefore, it is necessary to further screen sharp feature points and their nearby points and perform normal correction.

[0078] First, calculate the geometric feature statistics within the local neighborhood: the angle between normal vectors, the standard deviation of elevation, and the curvature distribution. Quantify the local curvature and terrain undulation intensity to construct a multi-dimensional feature space to describe local geometric characteristics. Calculate the geometric feature statistics within the local neighborhood:

[0079] In the formula, Indicates the intensity of terrain undulation. This indicates the number of valid points in the current local neighborhood that participate in the calculation of the elevation standard deviation. Represents the local neighborhood of the first Point elevation value, This represents the average elevation of the neighborhood.

[0080] Combining point cloud distribution characteristics with empirical data, an adaptive threshold adjustment mechanism is established—a loose threshold is used in areas with significant terrain undulations to enhance feature inclusiveness, while a tighter threshold is used in flat areas to reduce the false detection rate. Based on this, spatial proximity is used for extraction. The neighborhood point set consists of the selected sharp feature points.

[0081] Entering the normal correction stage, the feature propagation path is first constructed based on the dihedral angle geometric constraint: iteratively extending along the direction of the edge vector of the candidate point's neighborhood (step size equals neighborhood radius); simultaneously detecting the consistency of the normal vector direction of continuous extension points, when the difference in normal vectors between adjacent extension points is lower than the preset tolerance, it is determined that the plane has reached a stable region and the propagation is terminated.

[0082] The formula for checking the consistency of normal vector direction is as follows:

[0083] In the formula, , This represents the normal vector of adjacent extension points in the propagation path. This represents the tolerance threshold for the difference between normal vectors.

[0084] By correcting the normal vector direction of the path nodes through backpropagation, a continuous and smooth transition of the normal field at the feature edge is achieved.

[0085] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A lightweight point cloud method for achieving both noise robustness and detail fidelity, characterized in that: Includes the following steps: S1. An adaptive sliding window polynomial fitting filtering algorithm is used to suppress noise at multiple scales in the original point cloud data to obtain smooth point cloud data. S2. A multi-level axial bounding box adaptive recursive subdivision method is used to construct a spatial index structure for smooth point cloud data to achieve memory optimization in sparse regions and feature preservation in high-density regions. S3. Extract the skeleton structure from the point cloud data processed in step S2 based on the improved median transformation algorithm; S4. Perform normal estimation and correction on the point cloud data processed in step S3 to preserve and enhance sharp features.

2. The lightweight point cloud method for noise robustness and detail preservation according to claim 1, characterized in that: In step S1, the multi-scale noise suppression using the adaptive sliding window polynomial fitting filtering algorithm includes: Define the sliding window size With polynomial order Satisfying the constraints ,in Using empirical formulas Sure; For each point in the point cloud , construct a containing A data window containing points, window width polynomial order , ; For the center point of the window Perform PCA analysis, construct a local tangent plane, and project neighborhood points onto this local coordinate system; Construct the data matrix and perform least squares fitting, then calculate the window center point. The filtered value; A multi-resolution processing mechanism is introduced to dynamically adjust the filtering results of different scale windows through coarse-scale weights and fine-scale weights; A two-level window strategy is adopted in the feature edge region to dynamically optimize the filtering parameters.

3. The lightweight point cloud method for noise robustness and detail preservation according to claim 2, characterized in that: The introduction of a multi-resolution processing mechanism, which dynamically adjusts the filtering results of different scale windows through coarse-scale weights and fine-scale weights, includes: Coarse-scale weights: in, Indicates coarse-scale weights. This represents the variance of the data within the coarse-scale window. It is a very small constant; Fine-scale weights: in, Indicates fine-scale weights. This represents the gradient magnitude within a fine-scale window.

4. The lightweight point cloud method for noise robustness and detail preservation according to claim 2, characterized in that: In step S2, the multi-level axial bounding box adaptive recursive subdivision method includes: Calculate the minimum and maximum values ​​of the point cloud projections on each coordinate axis, determine the bounding box vertices, scan point by point and iteratively update the extreme values ​​to obtain the boundary of the minimum bounding box; Recursively divide the minimum bounding box into multiple sub-bounding boxes, and for each sub-bounding box... We introduce a weighting factor and use a combination of distance and weighting to select and retain points. Calculate the sub-bounding box weight entropy value based on the weight distribution within the sub-bounding box. The segmentation granularity is dynamically adjusted; if the entropy value is higher than the dynamic threshold... If the sub-bounding boxes are not recursively split, continue recursively; otherwise, stop splitting.

5. A lightweight point cloud method for noise robustness and detail preservation according to claim 4, characterized in that: The dynamic threshold: in, The average entropy value of the current level bounding box. The standard deviation of the entropy value. and This is an empirical coefficient.

6. The lightweight point cloud method for noise robustness and detail preservation according to claim 1, characterized in that: In step S3, extracting the skeleton structure from the point cloud data based on the improved median transformation algorithm includes: locating the distance field maxima point by combining dynamic radius neighborhood search, and processing each point... Calculate its distance to the nearest set of surface points. Perform central axis screening if If a value is a local maximum, it is marked as a central axis point. Curvature-weighted sampling is used to connect the central axis points to form a simplified skeleton, and sampling weights are assigned along the skeleton according to curvature.

7. A lightweight point cloud method for noise robustness and detail preservation according to claim 6, characterized in that: The central axis transformation algorithm uses the moving average method to initially smooth the central axis point cloud, and then performs secondary refinement based on Gaussian filtering. It uses the Douglas-Peucker algorithm for topology simplification, and dynamically retains key feature points and removes redundant nodes by iteratively calculating the maximum deviation between the sampling points and the baseline.

8. A lightweight point cloud method for noise robustness and detail preservation according to claim 1, characterized in that: In step S4, normal estimation and correction are performed on the point cloud data to preserve and enhance sharp features, including: Calculate the geometric feature statistics within the local neighborhood and filter out sharp feature points; A feature propagation path is constructed based on dihedral angle geometric constraints, and the normal vector direction of the path nodes is corrected through backpropagation.

9. A lightweight point cloud method for noise robustness and detail preservation according to claim 8, characterized in that: The feature propagation path constructed based on dihedral angle geometric constraints includes: iteratively extending along the direction of the edge vector of the candidate point's neighborhood, simultaneously detecting the consistency of the normal vector direction of consecutive extension points, and determining that the propagation has reached the stable region of the plane and terminating the propagation when the difference in normal vectors between adjacent extension points is lower than the preset tolerance. The formula for checking the consistency of normal vector direction is: in, , This represents the normal vector of adjacent extension points in the propagation path. This represents the tolerance threshold for the difference between normal vectors.

Images