Photogrammetric point cloud clustering noise filtering method and system

Abstract

The present application belongs to the technical field of satellite remote sensing photogrammetry point cloud processing, and particularly relates to a photogrammetry point cloud clustering noise filtering method and system, which comprises the following steps: obtaining an original point cloud and setting unconstrained cloth simulation parameters and morphological detection parameters; inverting the point cloud along the elevation, converting the clustering noise below the ground surface into a local convex shape; constructing a cloth grid above the inverted point cloud, discretizing the cloth grid into particles and removing the spring constraint, controlling the particles to fall and performing collision detection, and generating a continuous filtering reference surface; detecting local maximum points on the reference surface as candidate abnormal peak points, mapping the candidate points that simultaneously satisfy the gradient threshold and the curvature threshold back to the original point cloud, removing the points falling within the range as the clustering noise, and outputting the point cloud after noise removal. The present application can effectively remove the clustering noise below the ground surface and in the form of a sheet, a strip or a cluster in the satellite-borne photogrammetry point cloud, and improve the quality of the point cloud.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing photogrammetry point cloud processing technology, specifically relating to a method and system for filtering clustered noise in photogrammetric point clouds. Background Technology

[0002] With the rapid development of high-resolution Earth observation technology, sub-meter-level optical stereo mapping satellites, represented by Gaofen-7 (GF-7), have gradually become important data sources for large-scale urban 3D reconstruction, topographic mapping, and feature height extraction due to their advantages such as wide coverage and strong surface elevation acquisition capabilities. Through stereo photogrammetry and dense matching techniques, 3D point cloud data of urban areas can be quickly acquired, providing fundamental support for urban spatial information extraction.

[0003] In recent years, researchers have proposed various denoising methods to address the issue of improving the quality of point cloud data. These methods include those based on statistical principles (such as SOR filtering), local geometric fitting, global surface approximation, clustering of connected components (such as HDBSCAN), and general anomaly detection (such as LOF and iForest). The core steps of these methods are as follows: First, the original 3D point cloud data of the area to be processed is measured using the Gaofen-7 satellite imagery. Then, the spatial coordinates of each point are extracted from the 3D point cloud data. The average distance between each point and its multiple neighboring points is calculated. This average distance is compared with a standard distance threshold set based on global distribution. If the distance exceeds the threshold, it is judged as noise and filtered out. Alternatively, a plane or surface (such as a conical model) is fitted within the local neighborhood of the point cloud. Noise points are identified by analyzing the distance from the point to the fitted surface or local geometric properties (slope, curvature, etc.).

[0004] However, while existing methods are effective in processing lidar point clouds or general random outliers, they are less effective in the case of large coverage areas, complex scene features, uneven point density distribution, and the fact that noise caused by occlusion, shadows, weak textures, and parallax errors in complex urban areas is often distributed in sheet-like, band-like, or cluster-like continuous patterns. Their local density and geometric relationships may not show obvious anomalies, making it difficult to effectively identify such clustered noise with pseudo-geometric connectivity, which can easily lead to missed detections or misjudgments. Summary of the Invention

[0005] To address the problems existing in the prior art, the purpose of this invention is to provide a method and system for filtering clustered noise in photogrammetric point clouds, which can effectively remove clustered noise in spaceborne photogrammetric point clouds that is below the Earth's surface and appears as sheet-like, band-like, or clustered continuous noise, thereby improving the quality of the point clouds.

[0006] The technical solution of this invention is: A photogrammetric point cloud clustering noise filtering method includes the following steps: The original 3D point cloud data of the area to be processed is obtained, and the unconstrained cloth simulation parameters and morphological detection parameters including gradient feature threshold and Laplacian curvature feature threshold are set accordingly. The original three-dimensional point cloud data is inverted along the elevation direction, so that the clustered noise below the real ground surface in the original three-dimensional point cloud data is converted into a local convex shape in the inverted space, and the inverted point cloud is obtained. Based on the unconstrained cloth simulation parameters, a regular cloth mesh is constructed above the inverted point cloud. The cloth mesh is discretized into several cloth particles. The spring constraints and stiffness constraints between adjacent cloth particles are removed. The cloth particles are controlled to fall along the elevation direction and perform local collision detection with the inverted point cloud to obtain a continuous filter reference surface. Local maxima detection is performed on the filter reference surface. All detected local maxima points are taken as candidate abnormal peak points. The local gradient features and Laplacian curvature features at each candidate abnormal peak point are determined. Candidate abnormal peak points that simultaneously satisfy the gradient feature threshold and the Laplacian curvature feature threshold are taken as cluster noise center points. The cluster noise center points are mapped back to the original 3D point cloud data. A local spatial search range is set in the original 3D point cloud data with each cluster noise center point as the center. Points falling within the local spatial search range are removed as cluster noise points, and the denoised point cloud is output.

[0007] Preferably, the original 3D point cloud data is inverted along the elevation direction, including the following steps: Extract the elevation value of each point in the original 3D point cloud data; The elevation value of each point is reversed along the elevation axis, so that the clustered noise below the real ground surface in the original 3D point cloud data is converted into local convex structures in the inverted point cloud, thus forming the inverted point cloud.

[0008] Preferably, the unconstrained cloth simulation parameters include cloth mesh resolution, initial cloth height, gravity step size, and maximum number of iterations; obtaining a continuous filter reference surface includes the following steps: Based on the cloth mesh resolution, a regular cloth mesh is constructed above the inverted point cloud. The cloth mesh is discretized into several independent cloth particles. The spring constraints and stiffness constraints between adjacent cloth particles are removed, so that the internal stiffness of each cloth particle approaches zero and can independently respond to the local convex shape in the inverted point cloud. The initial position of the fabric particles is set according to the initial height of the fabric, and the fabric particles are controlled to fall along the elevation axis and perform local collision detection with the inverted point cloud according to the gravity step size and the maximum number of iterations. Record the stable positions of the fabric particles that have effective contact with the inverted point cloud. Based on preset elevation tolerance or effective support conditions, remove fabric particles that lack effective point cloud support. Perform spatial interpolation and smoothing on the retained fabric particles to generate a continuous filter reference surface.

[0009] Preferably, the step of using candidate anomaly peak points that simultaneously satisfy the gradient feature threshold and the Laplacian curvature feature threshold as cluster noise center points includes the following steps: Local maxima are searched on the filter reference surface using a two-dimensional sliding window, and all the searched local maxima are used as candidate abnormal peak points. Calculate the first-order change of the filter reference surface in the plane coordinate direction to obtain the local gradient features at each candidate anomaly peak point; The second-order change of the filter reference surface is calculated to obtain the Laplace curvature characteristics at each candidate anomaly peak point; Candidate abnormal peak points that simultaneously satisfy both gradient feature thresholds and Laplacian curvature feature thresholds are determined as the cluster noise center points.

[0010] Preferably, the step of removing points falling within the local spatial search range as clustered noise points includes the following steps: The two-dimensional coordinates of the cluster noise center point are mapped to the spatial coordinates of the original three-dimensional point cloud data, and a local spatial search range is established in the original three-dimensional point cloud data with the cluster noise center point as the reference. Points located within the local spatial search range and whose elevation is below a set threshold are identified as clustered noise points, and these clustered noise points are removed from the original 3D point cloud data.

[0011] Preferably, the original three-dimensional point cloud data is generated by dense matching of spaceborne optical stereo images.

[0012] Preferably, the clustering noise is anomaly clusters below the actual ground surface generated during the dense matching process of spaceborne optical stereo images, and the anomaly clusters are continuously distributed in a sheet-like, band-like or cluster-like manner in space.

[0013] Preferably, the local gradient features are determined according to the following formula: , In the formula, G The filtering reference surface is in coordinates ( x , y Local gradient eigenvalues ​​at () Z For the filter reference surface in coordinates ( x , y The elevation value at ().

[0014] Preferably, the Laplace curvature characteristic is determined according to the following formula: , In the formula, C The filtering reference surface is in coordinates ( x , y The Laplace curvature eigenvalue at () is shown. Z For the filter reference surface in coordinates ( x , y The elevation value at point ) is the Laplace operator for the elevation function.

[0015] A photogrammetric point cloud clustering noise filtering system includes: The data acquisition and parameter setting module is used to acquire the original 3D point cloud data of the area to be processed, and to set the unconstrained cloth simulation parameters and morphological detection parameters based on the acquired original 3D point cloud data; the morphological detection parameters include gradient feature threshold and Laplacian curvature feature threshold. The elevation inversion module is used to invert the original three-dimensional point cloud data along the elevation direction, so that the clustered noise in the original three-dimensional point cloud data that is lower than the real ground surface is converted into a local convex shape in the inverted space to obtain the inverted point cloud. An unconstrained cloth reference surface construction module is used to construct a regular cloth mesh above the inverted point cloud based on the unconstrained cloth simulation parameters, discretize the cloth mesh into several cloth particles, remove the spring constraints and stiffness constraints between adjacent cloth particles, and control the cloth particles to fall along the elevation direction and perform local collision detection with the inverted point cloud to obtain a continuous filter reference surface. The noise center detection module is used to perform local maximum detection on the filter reference surface, take all detected local maximum points as candidate abnormal peak points, determine the local gradient features and Laplacian curvature features at each candidate abnormal peak point, and take the candidate abnormal peak points that simultaneously satisfy the gradient feature threshold and the Laplacian curvature feature threshold as clustered noise center points. The abnormal point cluster removal and output module is used to map the cluster noise center point back to the original three-dimensional point cloud data, set a local spatial search range in the original three-dimensional point cloud data with each cluster noise center point as the center, remove the points falling into the local spatial search range as cluster noise points, and output the denoised point cloud.

[0016] Compared with the prior art, the photogrammetric point cloud clustering noise filtering method and system of the present invention has the following beneficial effects: This invention, through elevation inversion processing, transforms the "concave" noise in the original point cloud into a "convex" shape in the inverted space, solving the problem of clustered noise being difficult to identify due to its pseudo-connection to the real surface in space. Furthermore, it introduces unconstrained cloth simulation, removing spring and stiffness constraints between particles to allow each particle to independently respond to local convexities in the inverted point cloud, constructing a filtering reference surface that can sensitively fit anomalous shapes. Compared to traditional constrained cloth simulation, this avoids the problem of smoothing or ignoring small-scale anomalies caused by spring or stiffness constraints, significantly improving the ability to capture complex morphological noise. Based on this, it combines local maxima, Multidimensional morphological detection using gradient features and Laplacian curvature features can accurately locate the center of clustered noise, effectively distinguish noise protrusions from real terrain undulations, and significantly reduce the false positive rate. Finally, by mapping the noise center back to the original space and removing abnormal point clusters in the local neighborhood, stable and reliable filtering of clustered noise is achieved. This effectively solves the problem of missed detection or false positives in spaceborne photogrammetric point clouds caused by the pseudo-geometric connectivity of clustered noise, which is difficult to reliably identify using traditional methods. It significantly reduces the interference of continuous gross errors below the ground surface on ground filtering and feature height extraction, and improves the reliability of spaceborne point cloud preprocessing in complex urban areas and the accuracy of subsequent applications. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method in an embodiment of the present invention. Figure 2 This is a diagram summarizing the principle of the method in this embodiment of the invention; Figure 3 This is a framework diagram of the parameter preprocessing and initialization stage of the method flow in this embodiment of the invention; Figure 4 This is a framework diagram of the collision detection stage of the method flow in an embodiment of the present invention; Figure 5 This is a framework diagram of the noise removal stage of the method flow in an embodiment of the present invention; Figure 6 This is a comparison of the accuracy of the preprocessing results of the denoising method in this embodiment of the invention after inputting them into the CSF algorithm with the preprocessing results of other denoising methods. Figure 7 This is an example of the ground point extraction results after the preprocessing results of the denoising method and other denoising methods are input into the CSF algorithm in this embodiment of the invention. Detailed Implementation

[0018] 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 embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0020] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0021] Example 1 See Figures 1 to 5 As shown, in order to effectively remove cluster noise below the actual ground surface in densely matched point clouds of spaceborne optical stereo images, this embodiment provides a method for filtering cluster noise in photogrammetric point clouds, including the following steps: S1. Obtain the original 3D point cloud data of the area to be processed, and set the unconstrained cloth simulation parameters and morphological detection parameters based on the obtained original 3D point cloud data.

[0022] Obtain the raw 3D point cloud data of the area to be processed. The raw 3D point cloud data is generated from spaceborne optical stereo imagery (e.g., stereo imagery from the Gaofen-7 satellite) through dense matching. Each point in the raw 3D point cloud data contains at least planar coordinate information. X , Y ) and elevation information ( Z After acquiring the data, the unconstrained cloth simulation parameters and morphological detection parameters are set based on the acquired original 3D point cloud data.

[0023] In this embodiment, to improve the efficiency of subsequent processing, preferably, a spatial index structure for local spatial retrieval of the point cloud is established after the parameters are set. As a specific example, a KD-Tree (K-DimensionalTree) spatial index structure can be used to improve the computational efficiency of subsequent collision detection between cloth particles and the point cloud, as well as neighborhood point search.

[0024] Unconstrained cloth simulation parameters include, but are not limited to, cloth mesh resolution (cloth_resolution), initial cloth height (cloth_height), gravity step size (gravity_step), and maximum number of iterations (max_iterations). Among these:

[0025] Cloth mesh resolution: Used to control the spatial scale of the subsequently generated filter reference surface. Its value needs to be set in conjunction with the point cloud density and the scale of the target ground features. For example, for urban areas, it can be set to 1 meter to 5 meters.

[0026] Initial cloth height: Set to a certain distance above the highest point of the inverted point cloud to ensure that the cloth particles can fall from above.

[0027] Gravity step size: controls the distance the mass falls within each iteration step. The smaller the step size, the more detailed the simulation, but the greater the computational cost.

[0028] Maximum number of iterations: Sets the simulation termination condition to prevent infinite loops.

[0029] Morphological detection parameters include gradient feature threshold (G_threshold) and Laplacian curvature feature threshold (C_threshold). These thresholds can be adaptively determined based on the statistical distribution of the filtering reference surface. For example, the mean of the gradient values ​​at all points on the reference surface plus twice the standard deviation can be used as the gradient feature threshold, and the positive mean of the curvature values ​​can be used as the Laplacian curvature feature threshold, thus enhancing the algorithm's adaptability to different surface undulations.

[0030] S2. The original 3D point cloud data is inverted along the elevation direction, so that the clustered noise in the original 3D point cloud data that is lower than the real ground surface is transformed into a local convex shape in the inverted space, and the inverted point cloud is obtained.

[0031] The original 3D point cloud data is inverted along the elevation direction.

[0032] This step aims to transform subsurface cluster noise (manifesting as "pits" or "depressions") that is difficult to identify directly in the original point cloud into easily detectable local bulges or "peaks" in inverted space, thereby enhancing its detectability. Specifically, the elevation value is extracted for each point in the original 3D point cloud data. Z Then, the elevation values ​​of each point are reversed in sign along the elevation axis, and the inverted point cloud elevation is obtained according to the following formula. Z ':

[0033] Z' = -Z Through the above processing, the clustered noise below the actual ground surface in the original 3D point cloud data is transformed into a locally raised structure in the inverted point cloud, while the actual ground area appears as a relatively flat region, resulting in the inverted point cloud. This inverted point cloud will serve as the input data for subsequent unconstrained cloth simulations.

[0034] S3. Based on the unconstrained cloth simulation parameters, a regular cloth mesh is constructed above the inverted point cloud. The cloth mesh is discretized into several cloth particles, and the spring and stiffness constraints between adjacent cloth particles are removed. The cloth particles are controlled to fall along the elevation direction and perform local collision detection with the inverted point cloud to obtain a continuous filtering reference surface. Based on the unconstrained cloth simulation parameters, a regular cloth mesh is constructed above the inverted point cloud, and unconstrained cloth simulation is performed to obtain a continuous filtering reference surface. This step aims to generate a continuous surface that can fit the abnormal convex shape in the inverted point cloud.

[0035] Specifically, the method for generating a continuous filtering reference surface based on unconstrained cloth simulation parameters includes the following sub-steps: S3-1: Construct a regular cloth mesh above the inverted point cloud based on the cloth mesh resolution. This mesh is located above the inverted point cloud, and its initial height is determined by the cloth's initial height parameter.

[0036] S3-2: Discretize the cloth mesh into several independent cloth particles. This removes the spring and stiffness constraints between adjacent cloth particles in traditional cloth simulations, making the internal stiffness of each cloth particle approach zero. This "unconstrained" or "extremely flexible" cloth model allows each cloth particle to become an independent entity, capable of responding independently to local elevation changes in the point cloud below, without being affected by the stiffness of neighboring particles and smoothing out local convex details.

[0037] S3-3: Control the cloth particles to fall along the elevation direction. In each iteration step, update the position of the particles according to the gravity step size. At the same time, based on the pre-established spatial index structure, perform local collision detection between the cloth particles and the inverted point cloud.

[0038] S3-4: When a cloth particle comes into contact with a support point in the inverted point cloud within its collision search radius, it is considered a valid collision, and the stable position of the cloth particle (i.e., the position where it stops falling) is recorded.

[0039] S3-5: After completing all iterations, remove cloth particles (i.e., suspended invalid particles) that lack effective point cloud support according to the preset elevation tolerance or effective support conditions.

[0040] S3-6: Spatial interpolation and smoothing are performed on the retained effective fabric particles to generate a continuous filter reference surface. Interpolation methods can include inverse distance weighted (IDW), Kriging interpolation, or spline interpolation. Smoothing can employ moving average or Gaussian filtering to obtain a smooth and continuous reference surface. Due to the characteristics of unconstrained fabric, this reference surface can well fit the local bulges formed by clustered noise in the inverted point cloud.

[0041] S4. Perform local maximum detection on the filtering reference plane, and take all detected local maximum points as candidate abnormal peak points. Determine the local gradient features and Laplacian curvature features at each candidate abnormal peak point. Take the candidate abnormal peak points that simultaneously satisfy the gradient feature threshold and the Laplacian curvature feature threshold as the cluster noise center points. In this embodiment, the specific method for obtaining the cluster noise center point includes the following sub-steps: S4-1: Search for local maxima on the filter reference surface using a two-dimensional sliding window. For the center point of the window, if its elevation value is greater than the elevation values ​​of all other points within the window, it is marked as a candidate anomalous peak point.

[0042] Local maxima satisfy the following conditions: in, For coordinates ( x , y The local search neighborhood centered on ) For the search radius, For the set of candidate noise peak points, For the filter reference surface in coordinates ( x , y The elevation value at point ) This is a preset elevation threshold. Within this local neighborhood... The coordinates of any pixel.

[0043] S4-2: Calculate the first-order change of the filtered reference surface in the planar coordinate direction to obtain the local gradient features at each candidate anomaly peak point. This feature characterizes the intensity of the slope change of the reference surface at that point. The larger the gradient value, the more drastic the elevation change near that location, and the more likely it corresponds to the boundary of an anomaly protrusion. The local gradient features can be determined according to the following formula:

[0044] , In the formula, G The filtering reference surface is in coordinates ( x , y Local gradient eigenvalues ​​at () Z For the filter reference surface in coordinates ( x , y The elevation value at ().

[0045] S4-3: Calculate the second-order change of the filter reference surface to obtain the Laplacian curvature characteristics at each candidate anomaly peak point. This characteristic is used to describe the curvature of the local surface. Clustered noise, after inversion, appears as sharp peaks, with a larger curvature value in the central region. The Laplacian curvature characteristics can be determined according to the following formula:

[0046] , In the formula, C The filtering reference surface is in coordinates ( x , y The Laplace curvature eigenvalue at () is shown.Z For the filter reference surface in coordinates ( x , y The elevation value at point ) is the Laplace operator for the elevation function.

[0047] S4-4: Candidate anomaly peaks that simultaneously satisfy both the gradient feature threshold and the Laplacian curvature feature threshold are identified as cluster noise centers. That is, when the local gradient feature value at a candidate anomaly peak is greater than the threshold value, the cluster noise center is determined. G threshold (Gradient feature threshold) and Laplacian curvature eigenvalue > C threshold Only when the Laplacian curvature feature threshold is reached is it ultimately identified as the center point of the clustered noise. This multi-dimensional feature joint determination method can effectively reduce the misclassification of the slow undulations of real terrain as noise and improve the accuracy of identification.

[0048] S5. Map the cluster noise center points back to the original 3D point cloud data. Set a local spatial search range in the original 3D point cloud data with each cluster noise center point as the center. Remove the points that fall within the local spatial search range as cluster noise points and output the denoised point cloud.

[0049] S5-1: The two-dimensional coordinates of the cluster noise center points determined in step S4 ( x , y Map back to the spatial coordinate system of the original 3D point cloud data.

[0050] S5-2: Establish a local spatial search radius (e.g., a sphere with radius R or a cylinder with a specific height and width) in the original 3D point cloud data, centered on each cluster noise center point. This search radius can be set according to the scale of the outlier clusters, for example, 1.5 to 2 times the cloth mesh resolution.

[0051] S5-3: Points falling within the local spatial search range are identified as clustered noise points and removed. Preferably, an elevation determination condition can be further added, for example, only those points located below the center point of the clustered noise (i.e., whose elevation is below a set threshold) are removed to avoid mistakenly deleting real features. These clustered noise points are then removed from the original point cloud as a whole.

[0052] S5-4: Output denoised point cloud data. This point cloud data retains the coordinates and other attribute information of the original point cloud and can be used as high-quality input data for subsequent processing (such as ground point extraction, digital terrain model generation, normalized digital surface model generation, or ground feature height inversion).

[0053] Example 2 See Figure 6As shown, this embodiment provides a photogrammetric point cloud clustering noise filtering system to implement the photogrammetric point cloud clustering noise filtering method and system in Embodiment 1, including: Data acquisition and parameter setting module: used to acquire the original 3D point cloud data of the area to be processed, and set the unconstrained cloth simulation parameters and morphological detection parameters based on the acquired original 3D point cloud data; the morphological detection parameters include gradient feature threshold and Laplacian curvature feature threshold.

[0054] Elevation Inversion Module: This module inverts the original 3D point cloud data along the elevation direction, transforming clustered noise below the actual ground surface in the original 3D point cloud data into local convex shapes in the inverted space, thus obtaining the inverted point cloud.

[0055] Unconstrained Cloth Reference Surface Construction Module: This module is used to construct a regular cloth mesh on top of the inverted point cloud based on the unconstrained cloth simulation parameters. The cloth mesh is discretized into several cloth particles, and the spring and stiffness constraints between adjacent cloth particles are removed. The cloth particles are controlled to fall along the elevation direction and perform local collision detection with the inverted point cloud to obtain a continuous filtered reference surface.

[0056] Noise center detection module: Used to detect local maxima on the filter reference plane, take all detected local maxima points as candidate abnormal peak points, determine the local gradient features and Laplacian curvature features at each candidate abnormal peak point, and take the candidate abnormal peak points that simultaneously satisfy the gradient feature threshold and the Laplacian curvature feature threshold as the clustered noise center points.

[0057] The outlier point cluster removal and output module is used to map the center points of clustered noise back to the original 3D point cloud data. It sets a local spatial search range in the original 3D point cloud data with each clustered noise center point as the center, removes the points that fall within the local spatial search range as clustered noise points, and outputs the denoised point cloud.

[0058] Application Examples and Effect Verification To further verify the effectiveness of the technical solution of this invention, the applicant conducted an experiment in a typical urban built-up area of ​​Qingdao. The original three-dimensional point cloud data of this area was generated by dense matching of stereo images from the Gaofen-7 satellite. It includes various land features such as roads, buildings, trees, and water bodies, and contains a large amount of clustered noise below the true ground surface caused by factors such as weak texture and shadow occlusion. The clustered noise is distributed continuously in the space in the form of patches, bands, or clusters.

[0059] In this application embodiment, firstly, the original 3D point cloud data is preprocessed according to the method of the present invention. In S1, the cloth mesh resolution is set to 2 meters, the gravity step size is 0.01, the maximum number of iterations is 500, and the gradient feature threshold and Laplacian curvature feature threshold are adaptively determined according to the statistical distribution of the filtering reference surface. In S2, the point cloud elevation is inverted. In S3, an unconstrained cloth model is constructed and a filtering reference surface is generated. In S4, multiple morphological features are combined to determine the cluster noise center points. Finally, in S5, with each cluster noise center point as the center, a local spatial search range radius of 3 meters is set, and cluster noise points falling within this range are removed.

[0060] To verify its effectiveness, the method of this invention is compared with commonly used point cloud denoising methods such as SOR, HDBSCAN, LOF, and iForest.

[0061] See Figure 6 As shown, on the constructed self-made dataset (which covers various urban landform types such as roads, buildings, trees, and water bodies, and contains clustered noise below the ground surface caused by factors such as weak texture, shadow occlusion, and disparity matching errors in some areas), the ground point extraction accuracy after inputting the preprocessing results of different denoising algorithms into CSF ​​varies significantly. Experimental results show that the point cloud data processed by the method of this invention has the lowest root mean square error and the highest coefficient of determination in CSF ground point extraction. Compared with other comparative methods, it can more effectively reduce the interference of below-surface clustered noise on the ground filtering results, enabling the CSF algorithm to more accurately identify real ground points.

[0062] See Figure 7 As shown in the examples of ground point extraction results in typical areas, methods such as SOR, HDBSCAN, LOF, and iForest exhibit varying degrees of problems in misidentifying clustered noise as ground points or misidentifying real ground points as noise and removing them, resulting in a large amount of residual noise or ground holes in the generated ground point cloud. The method of this invention, however, can effectively remove abnormal clustered noise while better preserving real terrain details (such as road boundaries, gentle slopes, and micro-topographic undulations), making the subsequent CSF-extracted ground point results more continuous, smooth, and reliable, significantly improving the preprocessing quality of spaceborne photogrammetric point clouds and the accuracy of subsequent ground feature height inversion. Figure 7 As shown, experimental results demonstrate that, compared to other comparative methods, the method of this invention can more effectively remove clustered noise below the true ground surface, reduce "pit" errors generated during ground filtering, and better preserve real terrain details while removing anomalous point clusters. The processed point cloud was used to generate a digital terrain model (DTM) and a normalized digital surface model (nDSM), and the accuracy of the height inversion results for ground features (such as buildings and trees) was superior to that of the comparative methods, proving the effectiveness and superiority of this invention as a preprocessing step for photogrammetric point clouds.

[0063] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for filtering clustered noise in photogrammetric point clouds, characterized in that, Includes the following steps: The original 3D point cloud data of the area to be processed is obtained, and the unconstrained cloth simulation parameters and morphological detection parameters including gradient feature threshold and Laplacian curvature feature threshold are set accordingly. The original three-dimensional point cloud data is inverted along the elevation direction, so that the clustered noise below the real ground surface in the original three-dimensional point cloud data is converted into a local convex shape in the inverted space, and the inverted point cloud is obtained. Based on the unconstrained cloth simulation parameters, a regular cloth mesh is constructed above the inverted point cloud. The cloth mesh is discretized into several cloth particles. The spring constraints and stiffness constraints between adjacent cloth particles are removed. The cloth particles are controlled to fall along the elevation direction and perform local collision detection with the inverted point cloud to obtain a continuous filter reference surface. Local maxima detection is performed on the filter reference surface. All detected local maxima points are taken as candidate abnormal peak points. The local gradient features and Laplacian curvature features at each candidate abnormal peak point are determined. Candidate abnormal peak points that simultaneously satisfy the gradient feature threshold and the Laplacian curvature feature threshold are taken as cluster noise center points. The cluster noise center points are mapped back to the original 3D point cloud data. A local spatial search range is set in the original 3D point cloud data with each cluster noise center point as the center. Points falling within the local spatial search range are removed as cluster noise points, and the denoised point cloud is output.

2. The photogrammetric point cloud clustering noise filtering method according to claim 1, characterized in that, The original 3D point cloud data is inverted along the elevation direction, including the following steps: Extract the elevation value of each point in the original 3D point cloud data; The elevation value of each point is reversed along the elevation axis, so that the clustered noise below the real ground surface in the original 3D point cloud data is converted into local convex structures in the inverted point cloud, thus forming the inverted point cloud.

3. The photogrammetric point cloud clustering noise filtering method according to claim 1, characterized in that, The unconstrained cloth simulation parameters include cloth mesh resolution, initial cloth height, gravity step size, and maximum number of iterations. The process of obtaining a continuous filter reference surface includes the following steps: Based on the cloth mesh resolution, a regular cloth mesh is constructed above the inverted point cloud. The cloth mesh is discretized into several independent cloth particles. The spring constraints and stiffness constraints between adjacent cloth particles are removed, so that the internal stiffness of each cloth particle approaches zero and can independently respond to the local convex shape in the inverted point cloud. The initial position of the fabric particles is set according to the initial height of the fabric, and the fabric particles are controlled to fall along the elevation axis and perform local collision detection with the inverted point cloud according to the gravity step size and the maximum number of iterations. Record the stable positions of the fabric particles that have effective contact with the inverted point cloud. Based on preset elevation tolerance or effective support conditions, remove fabric particles that lack effective point cloud support. Perform spatial interpolation and smoothing on the retained fabric particles to generate a continuous filter reference surface.

4. The photogrammetric point cloud clustering noise filtering method according to claim 1, characterized in that, The step of using candidate anomaly peak points that simultaneously satisfy the gradient feature threshold and the Laplacian curvature feature threshold as cluster noise center points includes the following steps: Local maxima are searched on the filter reference surface using a two-dimensional sliding window, and all the searched local maxima are used as candidate abnormal peak points. Calculate the first-order change of the filter reference surface in the plane coordinate direction to obtain the local gradient features at each candidate anomaly peak point; The second-order change of the filter reference surface is calculated to obtain the Laplace curvature characteristics at each candidate anomaly peak point; Candidate abnormal peak points that simultaneously satisfy both gradient feature thresholds and Laplacian curvature feature thresholds are determined as the cluster noise center points.

5. The photogrammetric point cloud clustering noise filtering method according to claim 1, characterized in that, The step of removing points falling within the local spatial search range as clustered noise points includes the following steps: The two-dimensional coordinates of the cluster noise center point are mapped to the spatial coordinates of the original three-dimensional point cloud data, and a local spatial search range is established in the original three-dimensional point cloud data with the cluster noise center point as the reference. Points located within the local spatial search range and whose elevation is below a set threshold are identified as clustered noise points, and these clustered noise points are removed from the original 3D point cloud data.

6. The photogrammetric point cloud clustering noise filtering method according to claim 1, characterized in that, The original three-dimensional point cloud data was generated by dense matching of spaceborne optical stereo images.

7. The photogrammetric point cloud clustering noise filtering method according to claim 6, characterized in that, The cluster noise refers to anomalous point clusters below the actual ground surface generated during the dense matching process of spaceborne optical stereo images. These anomalous point clusters are distributed continuously in a sheet-like, band-like, or clump-like manner in space.

8. The photogrammetric point cloud clustering noise filtering method according to claim 4, characterized in that, The local gradient features are determined according to the following formula: ， In the formula, G The filtering reference surface is in coordinates ( x , y Local gradient eigenvalues ​​at () Z For the filter reference surface in coordinates ( x , y The elevation value at ().

9. A photogrammetric point cloud clustering noise filtering method according to claim 4, characterized in that, The Laplace curvature characteristic is determined according to the following formula: ， In the formula, C The filtering reference surface is in coordinates ( x , y The Laplace curvature eigenvalue at () is shown. Z For the filter reference surface in coordinates ( x , y The elevation value at point ) is the Laplace operator for the elevation function.

10. A photogrammetric point cloud clustering noise filtering system, characterized in that, include: The data acquisition and parameter setting module is used to acquire the original 3D point cloud data of the area to be processed, and to set the unconstrained cloth simulation parameters and morphological detection parameters based on the acquired original 3D point cloud data; the morphological detection parameters include gradient feature threshold and Laplacian curvature feature threshold. The elevation inversion module is used to invert the original three-dimensional point cloud data along the elevation direction, so that the clustered noise in the original three-dimensional point cloud data that is lower than the real ground surface is converted into a local convex shape in the inverted space to obtain the inverted point cloud. An unconstrained cloth reference surface construction module is used to construct a regular cloth mesh above the inverted point cloud based on the unconstrained cloth simulation parameters, discretize the cloth mesh into several cloth particles, remove the spring constraints and stiffness constraints between adjacent cloth particles, and control the cloth particles to fall along the elevation direction and perform local collision detection with the inverted point cloud to obtain a continuous filter reference surface. The noise center detection module is used to perform local maximum detection on the filter reference surface, take all detected local maximum points as candidate abnormal peak points, determine the local gradient features and Laplacian curvature features at each candidate abnormal peak point, and take the candidate abnormal peak points that simultaneously satisfy the gradient feature threshold and the Laplacian curvature feature threshold as clustered noise center points. The abnormal point cluster removal and output module is used to map the cluster noise center point back to the original three-dimensional point cloud data, set a local spatial search range in the original three-dimensional point cloud data with each cluster noise center point as the center, remove the points falling into the local spatial search range as cluster noise points, and output the denoised point cloud.

Images