A point cloud strip noise LTCF elimination method and system for a subsidence water area

Abstract

The present application relates to the technical field of three-dimensional point cloud data processing, and provides a point cloud strip noise LTCF removal method and system for subsidence water area, which comprises the steps of point cloud data acquisition, preprocessing, filtering processing and optimized point cloud output, and the filtering processing comprises the following steps: local covariance feature extraction: for each target point in the preliminary optimized point cloud obtained through preprocessing, a k-neighborhood is constructed, and the local covariance contribution value of each target point in X, Y and Z three spatial directions is calculated.The present application has the following beneficial effects: solving the industry pain point of subsidence water area scene, innovatively introducing a three-direction covariance joint determination mechanism, designing a special recognition logic for the strip noise of "low elevation, only containing XY direction covariance" which is specific to subsidence water area, and accurately capturing the sheet-shaped strip noise which cannot be recognized by traditional algorithms, thus fundamentally solving the removal problem of core noise of point cloud in subsidence water area.

Description

Technical Field

[0001] This invention is a method and system for removing point cloud strip noise in subsidence water areas using Local Three-dimensional Covariance Filtering (LTCF), which belongs to the field of three-dimensional point cloud data processing technology. Background Technology

[0002] As a core carrier of spatial geographic information, 3D point cloud data can accurately represent the 3D spatial morphology of a target scene and is widely used in fields such as monitoring of mining subsidence areas, aquatic ecological surveys, topographic mapping, and 3D modeling. Among these, the spatial morphology identification, boundary extraction, and dynamic monitoring of mining subsidence water areas highly rely on high-precision 3D point cloud data. However, during the point cloud acquisition process for subsidence water scenes, the raw point cloud inevitably introduces a large number of noise points due to multiple factors, including the multipath effect of the water surface, false echoes caused by wind and waves, UAV flight path stitching errors, equipment positioning accuracy deviations, and environmental interference.

[0003] The noise in the aforementioned subsidence water area scene exhibits significant scene specificity, with low-elevation strip noise as its core characteristic. These noise points are distributed in thin, elongated strips along the scanning flight strip, exhibiting distribution characteristics only in the XY plane and almost no spatial distribution in the Z direction. Furthermore, the Z-axis elevation is significantly lower than that of normal ground and water surface points, and they are mostly false echo points below the ground, redundant water surface reflection points, and artifacts caused by misalignment of the flight strip splicing. The spatial characteristics of this special noise are subtly different from those of the effective point cloud, making it difficult for traditional filtering algorithms to accurately identify and completely remove it. This severely interferes with the accuracy and reliability of subsequent processing such as subsidence water area boundary extraction, terrain inversion, and 3D modeling.

[0004] In existing technologies, the mainstream filtering algorithms for point cloud denoising mainly include the following categories, but they all have significant shortcomings in the scenario of processing strip noise in subsidence water areas:

[0005] Straight-through filtering: By setting a fixed threshold range for a single coordinate dimension to remove noise, it can only handle outliers in regular structured scenes. It cannot identify strip noise distributed in the XY plane with only elevation anomalies. Furthermore, the fixed threshold is very likely to cause the false deletion of effective low-elevation points or incomplete noise removal. It has extremely poor adaptability to point clouds in unstructured subsidence water areas.

[0006] Statistical filtering (SOR filtering): Based on the distance distribution characteristics within the neighborhood of a point cloud, a global fixed threshold is used to identify discrete outliers. It has a certain effect on processing sparse noise points floating in the air, but it is almost impossible to identify strip noise that is attached to the effective point cloud and is distributed in a planar manner. The noise removal rate is extremely low, and the fixed threshold is difficult to adapt to the uneven density of point clouds in submerged water areas, which easily leads to filtering or under-filtering problems.

[0007] Radius filtering: It distinguishes noise from valid points by determining the number of points in the neighborhood of a specified radius. It is only effective for isolated discrete noise points and cannot identify noise points that are continuously distributed in strips. At the same time, it is very easy to mistakenly delete low-density valid boundary points at the edge of subsided water areas and in areas with abrupt changes in terrain, resulting in the loss of target features.

[0008] Surface fitting-based filtering algorithms, while possessing high filtering accuracy, have extremely high computational complexity, stringent hardware performance requirements, and extremely low efficiency when processing point clouds of large-scale subsidence water areas. They cannot meet the needs of engineering-scale batch processing and still have insufficient accuracy in recognizing sheet-like strip noise.

[0009] In summary, existing point cloud filtering algorithms cannot specifically address the challenge of removing low-elevation strip noise unique to subsidence water areas. They suffer from technical defects such as insufficient special noise recognition capabilities, low noise removal rates, high false deletion rates of effective features, poor scene adaptability, and an inability to balance accuracy and efficiency. These defects severely restrict the subsequent application and analysis accuracy of point cloud data from subsidence water areas. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for removing point cloud strip noise (LTCF) from subsidence water areas.

[0011] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0012] A method for removing line-of-sight noise (LTCF) from point clouds in subsidence water areas includes the steps of point cloud data acquisition, preprocessing, filtering, and optimizing the point cloud output. The filtering process includes the following steps:

[0013] Local covariance feature extraction: For each target point in the pre-processed preliminary optimized point cloud, a k-nearest neighbor region is constructed, and the local covariance contribution value of each target point in the three spatial directions of X, Y, and Z is calculated.

[0014] Unidirectional low threshold filtering: Set a low threshold for Z-axis elevation, and filter points with Z-axis coordinates lower than the low threshold as low elevation candidate points;

[0015] Three-directional covariance joint determination: The three-directional covariance joint determination is performed on the low elevation candidate points to identify and remove low elevation strip noise points that only contain XY direction covariance in the subsidence water scene, and obtain a first-filtered point cloud.

[0016] Multi-scale neighborhood construction: For each target point in the filtered point cloud, both the k-nearest neighbor and the spherical neighborhood are constructed simultaneously, and the average distance parameter in the k-nearest neighbor and the normal vector consistency parameter in the spherical neighborhood are calculated respectively.

[0017] Adaptive thresholding and curvature filtering: Based on the average distance parameter and normal vector consistency parameter, a personalized filtering threshold is generated for each target point. Combined with the curvature value of the target point, residual noise is removed to obtain an optimized point cloud.

[0018] Preferably, the preprocessing specifically includes two steps: coordinate normalization and preliminary outlier removal. The coordinate normalization process involves normalizing the X, Y, and Z coordinates of the point cloud data of the subsidence area to be processed to the interval [-1, 1].

[0019] The initial outlier removal process involves calculating the mean and standard deviation of the point cloud coordinates along the X, Y, and Z axes, respectively. Points exceeding the mean ± 3 times the standard deviation are identified as outliers and removed, resulting in a preliminary optimized point cloud.

[0020] Preferably, the specific steps for local covariance feature extraction are as follows: for each target point in the preliminary optimized point cloud, a k-nearest neighbor region with a k value range of 10-30 is constructed using the kd-tree search algorithm; a 3×3 local covariance matrix is ​​constructed for the k-nearest neighbor region of the target point, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved; based on the solved eigenvalues ​​and eigenvectors, the local covariance contribution values ​​of the target point in the X, Y, and Z directions are calculated respectively.

[0021] Preferably, the joint determination condition for the three-directional covariance is to simultaneously satisfy the following three conditions: Condition 1: The Z-axis coordinate of the target point is lower than the Z-axis elevation low threshold;

[0022] Condition 2: The sum of the covariance contributions of the target point in the XY directions is greater than the covariance contribution in the Z direction;

[0023] Condition 3: The Z-direction covariance contribution of the target point is less than 20% of the mean Z-direction covariance of the preliminary optimized point cloud.

[0024] Preferably, in the multi-scale neighborhood construction step, the value of k for the k-nearest neighbor is in the range of 10-30, the radius of the spherical neighborhood is 1.5-2.5 times the average point spacing of the point cloud to be processed, and the normal vector consistency parameter is the average value of the cosine of the angle between the normal vector of all points in the spherical neighborhood and the normal vector of the target point.

[0025] Preferably, the adaptive threshold and curvature filtering are implemented as follows: a personalized filtering threshold is generated based on a weighted adaptive threshold model, the expression of which is: In the formula, The personalized filtering threshold for the i-th target point. Let k be the normalized value of the average distance between the k nearest neighbors of the i-th target point. Let α be the consistency parameter of the normal vector of the spherical neighborhood of the i-th target point, and let α be the weight coefficient, with a value range of 0.3-0.7.

[0026] Perform PCA principal component analysis on the k-nearest neighbor region of the target point, and solve for the three descending eigenvalues ​​λ1≥λ2≥λ3 of the covariance matrix. Then, use the formula... The curvature value C of the target point is calculated. i ;

[0027] If the curvature value C i Greater than the personalized filter threshold T i The point is determined to be a residual noise point and is removed. If the curvature value C i Less than or equal to the personalized filter threshold T i The point is determined to be a valid target point and is retained.

[0028] According to another aspect of the present invention, a point cloud stripe noise LTCF removal system for subsidence water areas is provided, which is used in the aforementioned point cloud stripe noise LTCF removal method for subsidence water areas. The system includes a data acquisition and preprocessing module, a covariance feature extraction module, a unidirectional low-threshold filtering module, a three-directional covariance joint determination module, a multi-scale neighborhood construction module, an adaptive filtering module, and a data output module, which are sequentially connected in communication.

[0029] The data acquisition and preprocessing module is used to collect or import point cloud data containing three-dimensional coordinate information in the subsidence water scene, and to complete the preprocessing to obtain a preliminary optimized point cloud.

[0030] The covariance feature extraction module is used to construct a k-nearest neighbor region for each target point in the preliminary optimized point cloud and calculate the local covariance contribution value of each target point in the three spatial directions of X, Y, and Z.

[0031] The unidirectional low threshold filtering module is used to set a low threshold for Z-axis elevation and filter out low elevation candidate points.

[0032] The three-direction covariance joint determination module is used to perform three-direction covariance joint determination on low elevation candidate points, identify and remove low elevation strip noise points that only contain XY direction covariance in the subsidence water scene, and obtain a first-filtered point cloud.

[0033] The multi-scale neighborhood construction module is used to construct k-nearest neighbors and spherical neighbors for target points in a single-filtered point cloud, and calculate the average distance parameter and the normal vector consistency parameter.

[0034] The adaptive filtering module is used to generate a personalized filtering threshold for the target point, and combine it with the curvature value of the target point to remove residual noise and obtain an optimized point cloud.

[0035] The data output module is used to integrate and output optimized point cloud data.

[0036] Preferably, the preprocessing operations performed by the data acquisition and preprocessing module include coordinate normalization processing to normalize the three-axis coordinates of the point cloud of the subsidence water area to the interval [-1,1], and preliminary optimization processing to remove outliers based on the coordinate mean and standard deviation.

[0037] Preferably, the three-direction covariance joint determination module has three built-in synchronous determination conditions: the target point's Z-axis coordinate is lower than a low threshold, the sum of the XY direction covariance contribution values ​​is greater than the Z direction covariance contribution value, and the Z direction covariance contribution value is less than 20% of the global Z direction covariance mean.

[0038] Preferably, the k-nearest neighbor k value of the multi-scale neighborhood construction module is in the range of 10-30, and the radius of the spherical neighborhood is 1.5-2.5 times the average point spacing of the point cloud to be processed; the adaptive filtering module has a built-in weighted adaptive threshold model, which can generate a personalized filtering threshold for each target point, and combine it with the curvature value of the target point to complete noise removal.

[0039] The beneficial effects of this invention are:

[0040] This invention addresses the industry pain points in subsidence water scenarios by innovatively introducing a three-directional covariance joint judgment mechanism. It designs a dedicated recognition logic for the unique characteristics of subsidence water, namely "low elevation and strip noise containing only XY direction covariance". This logic can accurately capture thin strip noise that traditional algorithms cannot recognize at all, thus solving the problem of removing core noise from point clouds in subsidence water from the root. Experiments have verified that the recognition rate of this invention for target noise far exceeds that of existing algorithms.

[0041] The design employs a dual judgment strategy of "single-directional low-threshold initial screening and three-directional covariance precise judgment." This strategy not only identifies high-noise areas by using a low threshold to improve processing efficiency, but also accurately distinguishes low-elevation noise points from normal low-elevation ground and water points through joint judgment of three-dimensional covariance features. This approach ensures a high noise removal rate while significantly reducing the false deletion rate of valid points and avoiding filtering problems.

[0042] By integrating multi-scale neighborhood features with a weighted adaptive threshold model, this invention overcomes the limitations of traditional algorithms with fixed thresholds. It can generate personalized filtering thresholds for each point based on the density and morphological features of different regions of the point cloud. It has excellent removal effects on various residual noises such as discrete outliers, edge artifacts, and floating points, and is highly adaptable to different scenarios. Experiments show that the point cloud removal rate of this invention reaches 72.9% within the region of interest, which is 2.90 times that of SOR filtering and 1.56 times that of radius filtering.

[0043] The invention achieves an optimal balance between "high-precision noise reduction" and "feature integrity preservation." Experiments verify that the noise point removal rate reaches 84.8%, while the boundary point ratio reaches 51.8%. The boundary point enrichment capability is 2.15 times that of the original point cloud. While significantly reducing data redundancy, it completely preserves key features such as the boundary of the subsidence area and abrupt terrain changes. The external point cloud is reduced by 48% while maintaining its original shape, providing high-quality point cloud data support for subsequent subsidence area identification, 3D modeling, and other operations.

[0044] The algorithm has a reasonable design and low computational complexity. It can quickly process large-scale point clouds without relying on high-performance hardware, combining accuracy and efficiency. It can be directly adapted to point cloud data of submerged water areas collected by mainstream equipment such as UAV oblique photography and airborne LiDAR, and has extremely high engineering application value. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 is a schematic flowchart of the point cloud strip noise LTCF removal method for subsidence water areas according to the present invention.

[0047] Figure 2 shows a comparison of the number of point clouds after processing with different filtering methods;

[0048] Figure 3 is a comparison of the point cloud removal rate and the proportion of boundary points in the boundary points using different filtering methods.

[0049] Figure 4 shows a comparison of the point cloud removal rates of different filtering methods within the region of interest;

[0050] Figure 5 is a comparison of the proportion of boundary points for different filtering methods;

[0051] Figure 6 shows a comparison of the feature preservation efficiency of different filtering methods;

[0052] Figure 7 shows the radar chart of the comprehensive performance scores of different filtering methods;

[0053] Figure 8 is a comparison of the point cloud effects before and after filtering in an embodiment of the present invention, wherein... Figure 8 (a) is the original point cloud image without filtering. Figure 8 (b) is the point cloud image after SOR statistical filtering. Figure 8 (c) is the point cloud effect after radius filtering. Figure 8(d) is the point cloud effect diagram after LTCF filtering processing according to the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1, please refer to Figure 1-8 The test object in this embodiment is 3D point cloud data from UAV oblique photography of a subsidence area in a mine in Panji District, Huainan City. This area is a typical coal mining subsidence area, and the point cloud data contains a large number of low-elevation strip noise caused by multipath effects on the water surface, artifacts caused by flight strip splicing errors, and discrete outliers, which is a typical scenario for point cloud processing in subsidence areas. In this embodiment, the total number of original point clouds is 26,731, and each point contains 3D coordinate (X, Y, Z) information.

[0056] The point cloud stripe noise LTCF removal method in this embodiment has the following specific steps:

[0057] Step 1: Point cloud data acquisition. Import the point cloud data to be processed from the above-mentioned subsidence water scene, read the X, Y, Z three-dimensional coordinate information of each point, and complete the data loading.

[0058] Step 2: Perform coordinate normalization on the acquired point cloud data, transforming the coordinates of all points to the interval [-1, 1]. The coordinate normalization formula is: In the formula, The coordinates are normalized and take values ​​in the range [-1, 1]. , These represent the maximum and minimum values ​​of the original point cloud's X-axis coordinates, respectively. The normalization method for the Y and Z-axis coordinates is similar. Then, the mean values ​​of the point cloud's coordinates along the X, Y, and Z axes are calculated. and standard deviation Taking the X direction as an example, the formula for the mean value of the coordinates is: Standard deviation formula: ; will exceed Points within the range are identified as outliers and removed. In this embodiment, after preprocessing, approximately 28 outliers were removed, leaving 26,703 valid point clouds.

[0059] Step 3: Local covariance feature extraction. For each preprocessed target point, construct a k-nearest neighbor region (k=20) and calculate the local covariance contribution value of the target point in the X, Y, and Z directions. Specifically, construct a covariance matrix for the k-nearest neighbor region of each target point, solve for the eigenvalues ​​and eigenvectors, and then calculate the covariance contribution value in each direction. The specific formula is as follows (taking the X direction as an example; the Y and Z directions are similar):

[0060] S31 calculates the mean of the k nearest neighbor center.

[0061] For target point The set of k nearest neighbors First, calculate the mean (center point) of the three-dimensional coordinates: In the formula, It is the j-th nearest neighbor. coordinate.

[0062] S32 constructs a 3×3 local covariance matrix;

[0063] The covariance matrix C is a symmetric matrix, and its elements represent the covariance between different dimensions. The formula is:

[0064] In the formula, It is a k×3 matrix (each row is the x-coordinate of a nearest neighbor point); It is a k×3 matrix (each row is...) ).

[0065] Solving for the eigenvalues ​​of the S33 covariance matrix;

[0066] |C-λI|=0; where λ is the eigenvalue (the three-dimensional covariance matrix has eigenvalues, denoted as λ1≥λ2≥λ3); I is a 3×3 identity matrix; |·| represents the matrix determinant.

[0067] S34 calculates the covariance contribution value in the X direction;

[0068] The eigenvalues ​​correspond to the variance (covariance contribution) of each principal direction. The eigenvectors need to be aligned with the X-axis before calculating the contribution value for each axis.

[0069] S341 first solves for the eigenvectors using the characteristic equation. , , (corresponding to λ1, λ2, λ3 respectively);

[0070] S342 eigenvectors Indicates the angle between the principal direction and the X-axis;

[0071] The final covariance contribution value of S343 in the X direction is: .

[0072] Step 4: Set a low threshold for Z-axis elevation, and select points with Z-axis coordinates below this threshold as candidate points for low elevation. In this embodiment, the minimum Z-axis value is 19.9225, therefore 20.0225 is selected.

[0073] Step 5: Joint determination of three-directional covariance. For the low-elevation candidate points selected in Step 4, a joint determination of the three-directional covariance is performed:

[0074] S51 calculates the mean global covariance: ;

[0075] S52 Judgment Criteria: ;in The threshold for determining covariance (in this embodiment) =0.2).

[0076] Points that simultaneously meet the above three conditions are identified as low elevation noise points containing only XY direction covariance and are removed.

[0077] In this embodiment, 7,128 special noise points were eliminated through this step, most of which were false measurement points below the ground.

[0078] Step 6: Multi-scale neighborhood construction for each target point P in the remaining point cloud i Construct k-nearest neighbors and spherical neighbors respectively:

[0079] S61k nearest neighbor region: Set k to 20 (range 10-30), use the kd-tree search algorithm to find the 20 nearest neighbors of each point, and calculate the distance between all points in this neighborhood and the target point P. i The average Euclidean distance D i For the target point in d-dimensional space and any point in its neighborhood (j=1,2,...,k, where k=20), Euclidean distance between two points The formula is: In the formula, d is the number of spatial dimensions; For target point P i The coordinates in the m-th dimension; For the neighborhood point P j The coordinates in the m-th dimension; Point P i With point P j The Euclidean distance between them. The average Euclidean distance D of the neighborhood. i The formula is: In the formula: k is the number of nearest neighbors, with a value of 20 (range 10~30); Di For target point P i The average Euclidean distance within the neighborhood.

[0080] S62 Spherical Neighborhood: Set the neighborhood radius R to twice the average point spacing of the point cloud to be processed (in this embodiment, the average point spacing is 0.0144m, so R=0.0288m), search all points within the spherical range, and calculate the normal vector of each point in the neighborhood and the target point P. i The cosine of the included angle of the normal vector is used as the average of all the cosine values ​​of the included angles, which is taken as the normal vector consistency parameter N. i The formula is: In the formula: m is the number of points in the spherical neighborhood; The normal vector of the target point; Let be the normal vector of the j-th point in the neighborhood.

[0081] Step 7: Adaptive threshold calculation, based on D i and N i Construct an adaptive filtering threshold model. First, for D... i Normalization is performed to obtain D _norm_i Then calculate the personalized threshold for each point. In this embodiment, the weighting coefficient α is set to 0.5 (with a value range of 0.3-0.7).

[0082] Step 8: Curvature filtering, calculate the curvature value C at each point. i (By performing PCA analysis on the k-nearest neighbor region, the ratio of the minimum eigenvalue to the sum of eigenvalues ​​is taken), the formula is: In the formula, C i Let λ be the curvature value at the i-th point; λ1, λ2, and λ3 are the three eigenvalues ​​of the covariance matrix (arranged in descending order).

[0083] The curvature value C i With personalized threshold T i Comparison:

[0084] S81 If C i >T i If a point is identified as a noise point (such as a discrete point on the edge or a floating noise point in the air), it is removed.

[0085] S82 if C i ≤T i If the point is determined to be a target point (such as a smooth area or ground point), it will be retained.

[0086] In this embodiment, approximately 7125 noise points were eliminated through this step.

[0087] Step 9: Optimize the point cloud output, integrate all retained target points, and restore the 3D point cloud data to the original coordinate range. The output format is TXT (including XYZ header), which will be used for subsequent 3D modeling, terrain analysis, and other applications. A total of 12,440 valid points are ultimately retained.

[0088] Example 2, please refer to Figure 1-8 Example of filtering effect verification:

[0089] To verify the advantages of this invention, the traditional statistical filtering algorithm (hereinafter referred to as SOR) and the radius filtering algorithm were selected as control groups. To objectively evaluate the filtering performance of the algorithm of this invention, the boundary points of the point cloud were marked manually (divided into inside and outside the region of interest, and...). Figure 8 (a) The boundary line is used to divide the data into inside and outside the boundary line. Points inside the boundary line are the remaining points to be removed, and points outside the boundary line are the remaining points to be thinned. The boundary line consists of 826 boundary points. In the comparative test with the algorithm of this invention, the test dataset was the same set of UAV point cloud data. The number of points in each cloud after filtering is shown in the figure. Figure 2 .

[0090] To verify the filtering effect of this invention, the following six indicators were used for comparative verification:

[0091] 1. Noise Removal Rate: To evaluate the filtering effect of each algorithm, the removal rate of the remaining points to be removed within the boundary line is calculated for each algorithm. The formula is as follows:

[0092]

[0093] In the formula, The original (unfiltered) number of points to be removed, and the total number of noise points that have been correctly identified and need to be removed (in this embodiment, there are 5061 remaining points to be removed). These are the points to be filtered after filtering. These are boundary points. See the results for the rejection rates of each algorithm. Figure 3 .

[0094] 2. Original Removal Rate: To demonstrate the proportion of points removed by the three filtering methods within the region of interest, the original point cloud is used as a benchmark to intuitively show the absolute removal capability of each method. The original point cloud represents the unprocessed, raw state. A removal rate of 0% is the ideal baseline, eliminating the mutual influence between different methods and directly reflecting the compression degree of each method relative to the original data. The formula is as follows:

[0095]

[0096] In the formula, This represents the number of remaining points to be removed in the original (unfiltered) state. See the removal rate results below. Figure 4 .

[0097] 3. Percentage of Boundary Points: Boundary points are the most critical feature information in the point cloud. Therefore, the percentage of boundary points for each filtering algorithm is calculated. See the results below. Figure 3 The formula is as follows:

[0098]

[0099] 4. Enrichment Ability: Furthermore, comparing the proportion of boundary points in the total number of filtered points can reflect the feature information enrichment of each algorithm. The calculation formula is as follows:

[0100]

[0101] In the formula, This represents the total number of remaining points under each filtering state. (From...) Figure 5 As can be seen, the boundary point ratio of LTCF is the highest at 6.64%, which is 2.15 times that of the original point cloud. The higher the ratio, the more concentrated the feature information is under the same amount of data. The radius filter is the second highest at 6.12%, and the SOR filter is only slightly higher than the original point cloud at 3.48%. This proves that LTCF has the strongest feature enrichment ability.

[0102] 5. Feature Preservation Efficiency: To demonstrate the amount of feature information carried per unit of data and reflect the algorithm's ability to retain key features while simplifying data, the following formula is introduced:

[0103]

[0104] In the formula Feature retention efficiency is a core indicator for measuring the degree of "few but excellent" features. Retention efficiency can be found in [link to relevant documentation]. Figure 6 .

[0105] 6. Overall performance

[0106] Ultimately, radar charts can simultaneously display performance across multiple dimensions, avoiding the one-sidedness of single-metric evaluation and comprehensively reflecting the algorithm's capabilities. Calculation formulas: Simplification score = Method's rejection rate / Maximum rejection rate (LTCF) × 100; Boundary retention score = Method's boundary point percentage / Maximum boundary point percentage (LTCF) × 100; External morphology... Figure 8 The feature efficiency score is calculated as: (Method's feature preservation efficiency / Maximum feature efficiency (LTCF)) × 100; the overall score is the weighted average of the scores from each dimension. See the overall performance calculation results below. Figure 7 .

[0107] The experimental results of the comparative test are as follows:

[0108] Filtering Algorithm Noise removal rate Boundary point percentage Region of Interest Removal Rate Boundary point enrichment ratio Feature Preservation Efficiency Original point cloud 0% 100% 0% 3.09% - SOR filter 29.20% 18.70% 25.10% 3.48% 0.04 radius filtering 54.30% 26.30% 46.70% 6.12% 0.12 This invention provides LTCF filtering. 84.80% 51.80% 72.90% 6.64% 0.14

[0109] The experimental results clearly show that:

[0110] (1) The core noise removal capability of this invention far exceeds that of existing mainstream algorithms. The removal rate of strip noise unique to subsidence waters reaches 84.8%, which is 2.90 times that of SOR filtering and 1.56 times that of radius filtering. It completely solves the problem of insufficient recognition capability of traditional algorithms for low elevation strip noise.

[0111] (2) While achieving a high rejection rate, this invention also achieves excellent feature retention. The boundary point ratio reaches 51.8%, which is much higher than SOR filtering and radius filtering. The boundary point enrichment ratio reaches 6.64%, which is 2.15 times that of the original point cloud. This can maximize the retention of key features such as the boundary of the subsidence water area and avoid the loss of effective features.

[0112] (3) The feature retention efficiency of the present invention is 0.14, which is the highest among the three types of algorithms. It achieves the optimal balance between data simplification and feature retention, and ensures the accuracy of subsequent subsidence water area identification and analysis while greatly reducing data redundancy.

[0113] (4) For the background point cloud outside the region of interest, the present invention achieves a point cloud simplification rate of 48.0% while preserving the basic outline and geometric shape of the point cloud and without structural distortion, which greatly reduces the computational load of subsequent processing and improves the overall processing efficiency.

[0114] In particular, it is worth noting that the algorithm of this invention achieves a good recognition rate for special noise points with low elevations containing only XY direction covariance (such as spurious measurement points below ground level and sheet-like noise points generated by equipment reflections), while traditional statistical filtering and radius filtering are almost ineffective in identifying this type of noise. These noise points have the following characteristics: their Z-axis coordinates are significantly lower than normal ground points; they are distributed in sheet-like patterns in the XY plane but lack distribution in the Z direction (the covariance contribution is mainly concentrated in the XY direction). This invention, through a single-direction low threshold + three-direction covariance joint judgment mechanism, can accurately identify and remove these special noise points that are difficult for traditional algorithms to handle, significantly improving the overall quality of point cloud data.

[0115] The experimental results above demonstrate that the algorithm of this invention significantly outperforms traditional algorithms in both noise removal rate and target point retention rate, achieving an optimal balance between high-precision filtering and target integrity preservation. For the points to be thinned, from... Figure 8It can be clearly observed that after filtering by the method of this invention, the basic outline and geometric shape of points outside the region of interest remain intact, without structural distortion or loss of key features. Simultaneously, the number of external point clouds is significantly reduced from the original 20,844 to 10,844, a reduction rate of 48.0%. This result demonstrates that the method of this invention significantly reduces data redundancy while ensuring the integrity of external background information, effectively reducing the computational load of subsequent point cloud registration, segmentation, and 3D modeling processes, thereby improving the overall algorithm efficiency and processing speed.

[0116] Example 3, please refer to Figure 1-8 LTCF removal system for point cloud strip noise in subsidence waters:

[0117] This embodiment provides a point cloud strip noise LTCF removal system for subsidence water areas, used to implement the filtering method described in Embodiment 1. The system includes a data acquisition and preprocessing module, a covariance feature extraction module, a unidirectional low threshold screening module, a three-directional covariance joint determination module, a multi-scale neighborhood construction module, an adaptive filtering module, and a data output module that are connected in sequence.

[0118] The specific functions of each module are as follows:

[0119] The data acquisition and preprocessing module is used to collect or import point cloud data containing three-dimensional coordinate information in the subsidence water scene, and to complete the preprocessing to obtain a preliminary optimized point cloud.

[0120] The covariance feature extraction module is used to construct a k-nearest neighbor region for each target point in the preliminary optimized point cloud and calculate the local covariance contribution value of each target point in the three spatial directions of X, Y, and Z.

[0121] The unidirectional low threshold filtering module is used to set a low threshold for Z-axis elevation and filter out low elevation candidate points.

[0122] The three-direction covariance joint determination module is used to perform three-direction covariance joint determination on low elevation candidate points, identify and remove low elevation strip noise points that only contain XY direction covariance in the subsidence water scene, and obtain a first-filtered point cloud.

[0123] The multi-scale neighborhood construction module is used to construct k-nearest neighbors and spherical neighbors for target points in a single-filtered point cloud, and calculate the average distance parameter and the normal vector consistency parameter.

[0124] The adaptive filtering module is used to generate a personalized filtering threshold for the target point, and combine it with the curvature value of the target point to remove residual noise and obtain an optimized point cloud.

[0125] The data output module is used to integrate and output optimized point cloud data. It can restore the optimized point cloud to the original coordinate range and supports exporting point cloud files in mainstream formats such as TXT and LAS, adapting to various subsequent point cloud processing software and application scenarios.

[0126] Although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A point cloud stripe noise removal method for subsidence water areas, wherein LTCF is an abbreviation for Local Three-dimensional Covariance Filtering, the method comprising the steps of point cloud data acquisition, preprocessing, filtering, and optimizing point cloud output, characterized in that, The filtering process includes the following steps: Local covariance feature extraction: For each target point in the pre-processed optimized point cloud, a k-nearest neighbor region is constructed. For each target point's k-nearest neighbor region, a 3×3 local covariance matrix is ​​constructed. The eigenvalues ​​λ1≥λ2≥λ3 and the eigenvectors of the covariance matrix are then solved. , , Based on the obtained eigenvalues ​​and eigenvectors, the local covariance contribution values ​​of each target point in the X, Y, and Z spatial directions are calculated respectively. The local covariance contribution value in the X direction is... ; Unidirectional low threshold filtering: Set a low threshold for Z-axis elevation, and filter points with Z-axis coordinates lower than the low threshold as low elevation candidate points; Three-directional covariance joint determination: The three-directional covariance joint determination is performed on the low elevation candidate points to identify and remove low elevation strip noise points that only contain XY direction covariance in the subsidence water scene, and obtain a first-filtered point cloud. Multi-scale neighborhood construction: For each target point in the filtered point cloud, both the k-nearest neighbor and the spherical neighborhood are constructed simultaneously. The average Euclidean distance from all points in the k-nearest neighbor to the target point is calculated as the average distance parameter, and the average cosine of the angle between the normal vector of all points in the spherical neighborhood and the normal vector of the target point is calculated as the normal vector consistency parameter. Adaptive thresholding and curvature filtering: Based on the average distance parameter and normal vector consistency parameter, a personalized filtering threshold is generated for each target point. Combined with the curvature value of the target point, residual noise is removed to obtain an optimized point cloud.

2. The method for removing point cloud stripe noise (LTCF) in subsidence water areas according to claim 1, characterized in that, The preprocessing specifically includes two steps: coordinate normalization and preliminary outlier removal. The coordinate normalization process involves normalizing the X, Y, and Z coordinates of the point cloud data of the subsidence area to be processed to the interval [-1, 1]. The initial outlier removal process involves calculating the mean and standard deviation of the point cloud coordinates along the X, Y, and Z axes, respectively. Points exceeding the mean ± 3 times the standard deviation are identified as outliers and removed, resulting in a preliminary optimized point cloud.

3. The method for removing point cloud stripe noise (LTCF) in subsidence water areas according to claim 1, characterized in that, The specific steps for local covariance feature extraction are as follows: For each target point in the preliminary optimized point cloud, a k-nearest neighbor region with a k value range of 10-30 is constructed using the kd-tree search algorithm; a 3×3 local covariance matrix is ​​constructed for the k-nearest neighbor region of the target point, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved; based on the solved eigenvalues ​​and eigenvectors, the local covariance contribution values ​​of the target point in the X, Y, and Z directions are calculated respectively.

4. The method for removing point cloud stripe noise (LTCF) in subsidence water areas according to claim 1, characterized in that, The joint determination of the three-directional covariance is based on the following three conditions being met simultaneously: Condition 1: The Z-axis coordinate of the target point is lower than the Z-axis elevation low threshold. Condition 2: The sum of the covariance contributions of the target point in the XY directions is greater than the covariance contribution in the Z direction; Condition 3: The Z-direction covariance contribution of the target point is less than 20% of the mean Z-direction covariance of the preliminary optimized point cloud.

5. The method for removing point cloud stripe noise (LTCF) in subsidence water areas according to claim 1, characterized in that, In the multi-scale neighborhood construction step, the value of k for the k-nearest neighbor is in the range of 10-30, and the radius of the spherical neighborhood is 1.5-2.5 times the average point spacing of the point cloud to be processed; the normal vector consistency parameter is the average value of the cosine of the angle between the normal vector of all points in the spherical neighborhood and the normal vector of the target point.

6. The method for removing point cloud stripe noise (LTCF) in subsidence water areas according to claim 1, characterized in that, The specific implementation of the adaptive threshold and curvature filtering is as follows: a personalized filtering threshold is generated based on a weighted adaptive threshold model, the expression of which is: In the formula, The personalized filtering threshold for the i-th target point. This is the normalized value of the average Euclidean distance from the i-th target point to all points in its k-nearest neighbor region. Let α be the consistency parameter of the normal vector of the spherical neighborhood of the i-th target point, and let α be the weight coefficient, with a value range of 0.3-0.

7. Perform PCA principal component analysis on the k-nearest neighbor region of the target point, and solve for the three descending eigenvalues ​​λ1≥λ2≥λ3 of the covariance matrix. Then, use the formula... The curvature value C of the target point is calculated. i ; If the curvature value C i Greater than the personalized filter threshold T i The point is determined to be a residual noise point and is removed. If the curvature value C i Less than or equal to the personalized filter threshold T i The point is determined to be a valid target point and is retained.

7. A point cloud stripe noise LTCF removal system for subsidence water areas, wherein LTCF is an abbreviation for Local Three-dimensional Covariance Filtering, characterized in that... The point cloud strip noise LTCF removal method for subsidence water areas as described in claim 1 includes, in sequence, a data acquisition and preprocessing module, a covariance feature extraction module, a unidirectional low threshold filtering module, a three-directional covariance joint determination module, a multi-scale neighborhood construction module, an adaptive filtering module, and a data output module: The data acquisition and preprocessing module is used to collect or import point cloud data containing three-dimensional coordinate information in the subsidence water scene, and to complete the preprocessing to obtain a preliminary optimized point cloud. The covariance feature extraction module is used to construct a k-nearest neighbor region for each target point in the preliminary optimized point cloud, construct a 3×3 local covariance matrix for the k-nearest neighbor region of each target point, solve the eigenvalues ​​and eigenvectors of the covariance matrix, and calculate the local covariance contribution values ​​of each target point in the three spatial directions of X, Y and Z based on the solved eigenvalues ​​and eigenvectors. The unidirectional low threshold filtering module is used to set a low threshold for Z-axis elevation and filter out low elevation candidate points. The three-direction covariance joint determination module is used to perform three-direction covariance joint determination on low elevation candidate points, identify and remove low elevation strip noise points that only contain XY direction covariance in the subsidence water scene, and obtain a first-filtered point cloud. The multi-scale neighborhood construction module is used to construct k-nearest neighbors and spherical neighbors for the target point in the first-filtered point cloud, calculate the average Euclidean distance from all points in the k-nearest neighbors to the target point as the average distance parameter, and calculate the average cosine of the angle between the normal vector of all points in the spherical neighborhood and the normal vector of the target point as the normal vector consistency parameter. The adaptive filtering module is used to generate a personalized filtering threshold for the target point, and combine it with the curvature value of the target point to remove residual noise and obtain an optimized point cloud. The data output module is used to integrate and output optimized point cloud data.

8. A point cloud stripe noise LTCF removal system for subsidence water areas according to claim 7, characterized in that, The preprocessing operations performed by the data acquisition and preprocessing module include coordinate normalization processing, which normalizes the three-axis coordinates of the point cloud of the subsidence water area to the interval [-1,1], and preliminary optimization processing, which removes outliers based on the coordinate mean and standard deviation.

9. A point cloud stripe noise LTCF removal system for subsidence water areas according to claim 7, characterized in that, The three-direction covariance joint determination module has three built-in synchronous determination conditions: the target point's Z-axis coordinate is lower than a low threshold, the sum of the XY direction covariance contribution values ​​is greater than the Z direction covariance contribution value, and the Z direction covariance contribution value is less than 20% of the mean Z direction covariance of the preliminary optimized point cloud.

10. A point cloud stripe noise LTCF removal system for subsidence water areas according to claim 7, characterized in that, The k-nearest neighbor k value of the multi-scale neighborhood construction module ranges from 10 to 30, and the radius of the spherical neighborhood is 1.5 to 2.5 times the average point spacing of the point cloud to be processed. The adaptive filtering module has a built-in weighted adaptive threshold model to generate a personalized filtering threshold for each target point, and combines the curvature value of the target point to complete noise removal.

Images