A method and system for quantitatively evaluating point cloud denoising algorithms

Abstract

The application belongs to the technical field of data preprocessing, and particularly relates to a method and system for quantitatively evaluating a point cloud denoising algorithm. The method comprises the following steps: taking a group of original point cloud data as a reference point cloud; generating simulated noise points through noise configuration parameters; the noise configuration parameters include a noise generation method, corresponding specific parameters and a number of noise points; merging the reference point cloud and the simulated noise points to generate a point cloud to be processed, and then performing denoising processing through denoising configuration parameters to obtain a main point cloud remaining after denoising and outlier points removed after denoising; the denoising configuration parameters include a denoising algorithm to be tested and corresponding key parameters; and calculating a quantitative evaluation index according to at least one of the main point cloud and the outlier points and at least one of the reference point cloud and the simulated noise points to quantitatively evaluate the point cloud denoising algorithm. The method can quantitatively evaluate the degree of restoring a real image from a denoised image, and provides a data-driven scientific basis for parameter selection of the denoising algorithm.

Description

Technical Field

[0001] This invention belongs to the field of data preprocessing technology, specifically relating to a method and system for quantitatively evaluating point cloud denoising algorithms. Background Technology

[0002] With the widespread application of 3D laser scanning and photogrammetry technologies, point cloud data plays a crucial role in fields such as autonomous driving, high-precision mapping, digital twins, and power line inspection. The quality of point cloud data directly affects the accuracy of subsequent analysis and decision-making; therefore, denoising, as the primary step in data preprocessing, is of paramount importance.

[0003] Mainstream denoising algorithms in the industry, such as radius filtering and statistical filtering, rely heavily on the setting of a series of key parameters for performance. However, existing technologies generally suffer from the following serious shortcomings when evaluating the effectiveness of these algorithms and selecting parameters: The evaluation methods are subjective and non-quantitative: they generally rely on visual inspection by operators to judge the noise reduction effect, lacking objective and reproducible quantitative indicators. They cannot accurately answer the two core questions: how much noise was removed and how many effective points were retained.

[0004] Blind and inefficient parameter selection: Due to the lack of a quantitative feedback mechanism, parameter adjustment relies heavily on the operator's experience and repeated trial and error, which is not only inefficient, but also makes it difficult to find the optimal parameter combination that achieves the best balance between noise reduction and fidelity.

[0005] Lack of standardized comparison benchmarks: When comparing the performance of different algorithms or software, inconsistent test data and evaluation standards lead to a lack of fairness and credibility in the comparison results, which is not conducive to technology selection and algorithm optimization.

[0006] Currently, existing technologies lack methods and systems for objectively, quantitatively, and systematically evaluating point cloud denoising algorithms, which has become a technical bottleneck restricting the improvement of point cloud data processing accuracy and efficiency. For example, Chinese invention patent application CN114972066A discloses an automatic optimization method and apparatus for point cloud denoising algorithm parameters. The method includes: substituting each parameter in a pre-configured parameter group into an existing initial point cloud denoising algorithm to obtain an intermediate point cloud denoising algorithm; denoising the point cloud to be denoised to obtain a denoising result; evaluating the smoothness and local density of the denoising result to obtain an evaluation result; adjusting each parameter of the denoising algorithm using the evaluation result to obtain an adjusted denoising algorithm; and returning to execute the step of denoising the point cloud to be denoised to obtain a denoising result, until the number of adjustments reaches a preset number of adjustments; and forming a target parameter group from the parameters under the preset number of adjustments. As can be seen, evaluating the denoising effect and obtaining the denoising evaluation result as feedback for optimizing denoising, updating the parameters of the point cloud denoising algorithm, and finally obtaining the optimized parameters greatly saves the time of manual parameter tuning. Although this method can evaluate the point cloud denoising algorithm by quantifying the quality of the denoised image, it can only quantify the degree of relative clarity and smoothness of the image, but cannot quantify the degree to which the denoised image restores the real image. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for quantitatively evaluating point cloud denoising algorithms, which solves the problem that it is difficult to objectively, quantitatively, and systematically evaluate point cloud denoising algorithms in the prior art.

[0008] To achieve the above objectives, the present invention provides a method for quantitatively evaluating point cloud denoising algorithms, comprising: A set of raw point cloud data is used as the baseline point cloud; simulated noise points are generated through noise configuration parameters; the noise configuration parameters include the noise generation method, its corresponding specific parameters, and the number of noise points. The baseline point cloud is merged with simulated noise to generate a point cloud to be processed. Then, denoising is performed using denoising configuration parameters to obtain the remaining main point cloud after denoising and the outliers that are removed after denoising. The denoising configuration parameters include the denoising algorithm to be tested and its corresponding key parameters. Based on at least one of the main point cloud and outlier points, and at least one of the baseline point cloud and simulated noise points, a quantitative evaluation index is calculated to quantitatively evaluate the point cloud denoising algorithm.

[0009] Furthermore, the calculated quantitative evaluation index includes at least two different evaluation parameters.

[0010] Furthermore, the calculated quantitative evaluation metrics are used to provide feedback on the values ​​of the noise configuration parameters and the denoising configuration parameters.

[0011] Furthermore, it also includes: visually displaying the final quantitative evaluation indicators in tabular form, and displaying the intermediate process data used to calculate the quantitative evaluation indicators; the intermediate process data includes the number of reference point cloud points, the number of generated simulated noise points, the total number of noise points removed, and the number of reference point cloud points that were mistakenly deleted.

[0012] Furthermore, it also includes: overlaying or displaying the baseline point cloud, simulated noise, the remaining main point cloud after denoising, and the outliers removed after denoising through a visualization rendering window, and distinguishing them by different colors.

[0013] Furthermore, methods for generating simulated noise by configuring noise parameters include: The noise generation method in the noise configuration parameters is the radial displacement outlier generation method. When the specific parameters corresponding to the radial displacement outlier generation method are the noise source point and the noise generation radius, a virtual sphere space with the noise generation radius is defined with the coordinates of the noise source point as the center. A candidate noise point is randomly generated within the global bounding box of the reference point cloud and outside the virtual sphere space. Then, the finally generated candidate noise point that meets the conditions is output as a simulated noise point. The noise generation method in the noise configuration parameters is the statistical outlier generation method. Given the specific parameters for this method—the number of neighboring points to be analyzed and the standard deviation factor—a spatial index is constructed for the baseline point cloud. The average nearest neighbor distance of all points in the baseline point cloud is calculated, and the global mean and global standard deviation of these average distances are further calculated. Based on the global mean, global standard deviation, and standard deviation factor, an average distance threshold for a target outlier is defined. A candidate noise point is generated, and its average distance to the nearest neighbor of the analyzed neighboring points in the baseline point cloud is calculated. If this average distance is less than the average distance threshold, the point is discarded and regenerated until it meets the statistical outlier condition that the average distance is not less than the average distance threshold. The candidate noise point that finally meets the condition is then output as a simulated noise point. The noise generation method in the noise configuration parameters is the Gaussian surface noise generation method. Given the specific parameters for the Gaussian surface noise generation method are the noise source point, the number of neighboring points, and the standard deviation of the Gaussian distribution, the method searches for the nearest neighbor point of the set number of neighboring points, centered on the noise source point, to form a neighborhood point set. Principal component analysis is performed on this point set to calculate the local surface normal vector at the noise source point. A value is randomly sampled from a normal distribution with a mean of 0 and a standard deviation equal to the set standard deviation of the Gaussian distribution, and this value is used as the perturbation distance along the normal direction. Based on the perturbation distance, the local surface normal vector, and the noise source point, the coordinates of the generated noise point are calculated, and the noise point is output. The noise generation method in the noise configuration parameters is the simulated multipath reflection cluster generation method. Under the specific parameters of the simulated multipath reflection cluster generation method, namely the noise cluster center point, the number of points in the cluster, and the standard deviation of the cluster distribution, the noise cluster center point is taken as the center in the sparse region of the reference point cloud or the external space. Following a three-dimensional Gaussian distribution with a standard deviation equal to the standard deviation of the cluster distribution, the coordinates of the number of points in the cluster are randomly generated. All generated points are output, which together constitute an independent point cluster as noise.

[0014] Furthermore, the methods for denoising the generated point cloud to be processed by configuring denoising parameters include: The denoising algorithm to be tested in the denoising configuration parameters is the radius filtering algorithm. Under the condition that the key parameters of the radius filtering algorithm are the search radius and the minimum number of neighboring points, a spatial index is constructed for the point cloud to be processed to efficiently perform neighborhood search. Each point in the point cloud to be processed is traversed, and a radius search is performed on each point using the spatial index to find all neighboring points whose Euclidean distance to the point is within the search radius. The number of neighboring points found is counted. If the number of neighboring points is less than the minimum number of neighboring points, the point is classified as an outlier. Otherwise, the point is kept in the remaining main point cloud after denoising. Finally, the main point cloud and the outlier are returned. The denoising algorithm to be tested in the denoising configuration parameters is a statistical filtering algorithm. Given that the key parameters for the statistical filtering algorithm are the number of neighboring points analyzed and the standard deviation multiple threshold, a kd-tree is constructed for the point cloud to be processed. For each point in the point cloud, its nearest neighbor (based on the number of neighboring points analyzed) is found, and the average distance from that point to these nearest neighbors is calculated. The global average and global standard deviation of all average distances are then calculated. Based on the global average, global standard deviation, and standard deviation multiple, a global distance threshold is calculated. For each point in the point cloud to be processed, if the average distance from that point to its nearest neighbor is greater than the global distance threshold, that point is classified as an outlier; otherwise, it is retained in the main point cloud. Finally, the main point cloud and the outlier points are returned. The denoising algorithm to be tested in the denoising configuration parameters is a bilateral filtering algorithm. Given that the key parameters for the bilateral filtering algorithm are spatial standard deviation and feature standard deviation, a spatial index is constructed for the point cloud to be processed. Each point in the point cloud is traversed, and a neighborhood is determined for that point. For each point within the neighborhood, a spatial weight is calculated based on the Euclidean distance between that point and the corresponding point in the point cloud, and a Gaussian function of the spatial standard deviation. A local tangent plane is then calculated at the point in the point cloud. The perpendicular distance from the points in the neighborhood to this tangent plane is calculated, and feature weights are calculated based on the Gaussian function of the perpendicular distance and the feature standard deviation. Based on the feature weights and spatial weights, the final weights of the points in the neighborhood are calculated. The new position of a point in the point cloud is the weighted average of all its neighboring points based on the final weights. The resulting main point cloud, composed of all the new positions and smoothed with denoising, is returned.

[0015] Furthermore, the methods for obtaining raw point cloud data include: receiving the file corresponding to the raw point cloud data uploaded by the user through a file selection dialog box provided to the user, and obtaining the raw point cloud data; The noise configuration parameters can be obtained through a drop-down menu provided to the user for selecting a noise generation method, at least one input box for inputting specific parameter values ​​corresponding to the noise generation method, and a button or input box for specifying the number of noise points generated. The methods for obtaining denoising configuration parameters include: providing users with a drop-down menu for selecting the denoising algorithm to be tested, and at least one input box for setting the values ​​of the key parameters corresponding to the denoising algorithm.

[0016] Furthermore, the calculated quantitative evaluation metrics include both denoising rate and fidelity.

[0017] The above-described technical solution of this invention provides a novel method for quantitatively evaluating point cloud denoising algorithms, the beneficial effects of which include: By introducing actively generated simulated noise into the original point cloud data and combining it with the denoised results, core indicators for evaluating point cloud denoising algorithms are calculated, achieving a comprehensive, objective, and reproducible quantitative evaluation of algorithm performance. Since both noisy and noise-free original point clouds are simulated as configurable quantitative data, the degree to which the denoised image restores the true image can be quantitatively evaluated. This provides a data-driven scientific basis for parameter selection in denoising algorithms; by comparing the quantitative evaluation results under different parameters, the optimal parameter combination that achieves the best balance between denoising effect and data fidelity can be quickly found. Through a configurable parameterized process, this method can adapt to various types and scenarios of point cloud data, transforming the traditional inefficient trial-and-error process into an efficient and quantifiable testing and analysis process. This makes the method a fair and consistent algorithm testing framework, enabling impartial performance comparisons of different algorithms or parameter combinations under the same testing conditions, providing reliable data support for technology selection and algorithm optimization.

[0018] This invention also provides a system for quantizing and evaluating point cloud denoising algorithms, comprising a processor storing executable program instructions, which are executed to implement a method for quantizing and evaluating point cloud denoising algorithms, specifically including: A set of raw point cloud data is used as the baseline point cloud; simulated noise points are generated through noise configuration parameters; the noise configuration parameters include the noise generation method, its corresponding specific parameters, and the number of noise points. The baseline point cloud is merged with simulated noise to generate a point cloud to be processed. Then, denoising is performed using denoising configuration parameters to obtain the remaining main point cloud after denoising and the outliers that are removed after denoising. The denoising configuration parameters include the denoising algorithm to be tested and its corresponding key parameters. Based on at least one of the main point cloud and outlier points, and at least one of the baseline point cloud and simulated noise points, a quantitative evaluation index is calculated to quantitatively evaluate the point cloud denoising algorithm.

[0019] Furthermore, the calculated quantitative evaluation index includes at least two different evaluation parameters.

[0020] Furthermore, the calculated quantitative evaluation metrics are used to provide feedback on the values ​​of the noise configuration parameters and the denoising configuration parameters.

[0021] Furthermore, the method for quantifying and evaluating point cloud denoising algorithms also includes: visually displaying the final quantification evaluation index in tabular form, and displaying intermediate process data used to calculate the quantification evaluation index; the intermediate process data includes the number of reference point cloud points, the number of generated simulated noise points, the total number of noise points removed, and the number of reference point cloud points that were mistakenly deleted.

[0022] Furthermore, methods for quantitatively evaluating point cloud denoising algorithms also include: overlaying or displaying the baseline point cloud, simulated noise, the remaining main point cloud after denoising, and the outliers removed after denoising through a visualization rendering window, and distinguishing them by different colors.

[0023] Furthermore, methods for generating simulated noise by configuring noise parameters include: The noise generation method in the noise configuration parameters is the radial displacement outlier generation method. When the specific parameters corresponding to the radial displacement outlier generation method are the noise source point and the noise generation radius, a virtual sphere space with the noise generation radius is defined with the coordinates of the noise source point as the center. A candidate noise point is randomly generated within the global bounding box of the reference point cloud and outside the virtual sphere space. Then, the finally generated candidate noise point that meets the conditions is output as a simulated noise point. The noise generation method in the noise configuration parameters is the statistical outlier generation method. Given the specific parameters for this method—the number of neighboring points to be analyzed and the standard deviation factor—a spatial index is constructed for the baseline point cloud. The average nearest neighbor distance of all points in the baseline point cloud is calculated, and the global mean and global standard deviation of these average distances are further calculated. Based on the global mean, global standard deviation, and standard deviation factor, an average distance threshold for a target outlier is defined. A candidate noise point is generated, and its average distance to the nearest neighbor of the analyzed neighboring points in the baseline point cloud is calculated. If this average distance is less than the average distance threshold, the point is discarded and regenerated until it meets the statistical outlier condition that the average distance is not less than the average distance threshold. The candidate noise point that finally meets the condition is then output as a simulated noise point. The noise generation method in the noise configuration parameters is the Gaussian surface noise generation method. Given the specific parameters for the Gaussian surface noise generation method are the noise source point, the number of neighboring points, and the standard deviation of the Gaussian distribution, the method searches for the nearest neighbor point of the set number of neighboring points, centered on the noise source point, to form a neighborhood point set. Principal component analysis is performed on this point set to calculate the local surface normal vector at the noise source point. A value is randomly sampled from a normal distribution with a mean of 0 and a standard deviation equal to the set standard deviation of the Gaussian distribution, and this value is used as the perturbation distance along the normal direction. Based on the perturbation distance, the local surface normal vector, and the noise source point, the coordinates of the generated noise point are calculated, and the noise point is output. The noise generation method in the noise configuration parameters is the simulated multipath reflection cluster generation method. Under the specific parameters of the simulated multipath reflection cluster generation method, namely the noise cluster center point, the number of points in the cluster, and the standard deviation of the cluster distribution, the noise cluster center point is taken as the center in the sparse region of the reference point cloud or the external space. Following a three-dimensional Gaussian distribution with a standard deviation equal to the standard deviation of the cluster distribution, the coordinates of the number of points in the cluster are randomly generated. All generated points are output, which together constitute an independent point cluster as noise.

[0024] Furthermore, the methods for denoising the generated point cloud to be processed by configuring denoising parameters include: The denoising algorithm to be tested in the denoising configuration parameters is the radius filtering algorithm. Under the condition that the key parameters of the radius filtering algorithm are the search radius and the minimum number of neighboring points, a spatial index is constructed for the point cloud to be processed to efficiently perform neighborhood search. Each point in the point cloud to be processed is traversed, and a radius search is performed on each point using the spatial index to find all neighboring points whose Euclidean distance to the point is within the search radius. The number of neighboring points found is counted. If the number of neighboring points is less than the minimum number of neighboring points, the point is classified as an outlier. Otherwise, the point is kept in the remaining main point cloud after denoising. Finally, the main point cloud and the outlier are returned. The denoising algorithm to be tested in the denoising configuration parameters is a statistical filtering algorithm. Given that the key parameters for the statistical filtering algorithm are the number of neighboring points analyzed and the standard deviation multiple threshold, a kd-tree is constructed for the point cloud to be processed. For each point in the point cloud, its nearest neighbor (based on the number of neighboring points analyzed) is found, and the average distance from that point to these nearest neighbors is calculated. The global average and global standard deviation of all average distances are then calculated. Based on the global average, global standard deviation, and standard deviation multiple, a global distance threshold is calculated. For each point in the point cloud to be processed, if the average distance from that point to its nearest neighbor is greater than the global distance threshold, that point is classified as an outlier; otherwise, it is retained in the main point cloud. Finally, the main point cloud and the outlier points are returned. The denoising algorithm to be tested in the denoising configuration parameters is a bilateral filtering algorithm. Given that the key parameters for the bilateral filtering algorithm are spatial standard deviation and feature standard deviation, a spatial index is constructed for the point cloud to be processed. Each point in the point cloud is traversed, and a neighborhood is determined for that point. For each point within the neighborhood, a spatial weight is calculated based on the Euclidean distance between that point and the corresponding point in the point cloud, and a Gaussian function of the spatial standard deviation. A local tangent plane is then calculated at the point in the point cloud. The perpendicular distance from the points in the neighborhood to this tangent plane is calculated, and feature weights are calculated based on the Gaussian function of the perpendicular distance and the feature standard deviation. Based on the feature weights and spatial weights, the final weights of the points in the neighborhood are calculated. The new position of a point in the point cloud is the weighted average of all its neighboring points based on the final weights. The resulting main point cloud, composed of all the new positions and smoothed with denoising, is returned.

[0025] Furthermore, the methods for obtaining raw point cloud data include: receiving the file corresponding to the raw point cloud data uploaded by the user through a file selection dialog box provided to the user, and obtaining the raw point cloud data; The noise configuration parameters can be obtained through a drop-down menu provided to the user for selecting a noise generation method, at least one input box for inputting specific parameter values ​​corresponding to the noise generation method, and a button or input box for specifying the number of noise points generated. The methods for obtaining denoising configuration parameters include: providing users with a drop-down menu for selecting the denoising algorithm to be tested, and at least one input box for setting the values ​​of the key parameters corresponding to the denoising algorithm.

[0026] Furthermore, the calculated quantitative evaluation metrics include both denoising rate and fidelity.

[0027] The technical solution of the system for quantizing point cloud denoising algorithm described above in this invention can achieve the same beneficial effects as the method for quantizing point cloud denoising algorithm described above. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the method for quantifying and evaluating point cloud denoising algorithms in the implementation of the present invention. Figure 2 This is an example graphical user interface diagram corresponding to the implementation method of the point cloud denoising algorithm for quantitative evaluation in the present invention. Detailed Implementation

[0029] 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.

[0030] Quantitative Evaluation Methods and Implementation Implementation of Point Cloud Denoising Algorithms This embodiment presents a technical solution for quantitatively evaluating point cloud denoising algorithms. By configuring parameters for noise, simulated noise with known features is added to the benchmark point cloud data, and then the algorithm under test is called for processing. Finally, based on the simulated noise and the benchmark point cloud, the known true values ​​are compared to accurately evaluate the algorithm performance, thereby providing scientific data support for algorithm optimization and parameter selection.

[0031] The method includes: A set of original point cloud data is used as the baseline point cloud; simulated noise points are generated by noise configuration parameters; the noise configuration parameters include the noise generation method, the specific parameters corresponding to the noise generation method, and the number of noise points. The baseline point cloud and simulated noise points are merged to generate a point cloud to be processed. Then, the generated point cloud to be processed is denoised using denoising configuration parameters to obtain the remaining main point cloud after denoising and the outliers removed after denoising. The denoising configuration parameters include the denoising algorithm to be tested and the key parameters corresponding to the denoising algorithm. Based on at least one of the remaining main point cloud after denoising and the outliers removed after denoising, as well as at least one of the baseline point cloud and simulated noise, a quantitative evaluation index is calculated to quantitatively evaluate the point cloud denoising algorithm.

[0032] Therefore, by adding simulated noise with known features to the benchmark point cloud data through a process that responds to external parameter configuration, and then calling the algorithm under test for processing, this method effectively overcomes the drawbacks of subjective evaluation. By introducing actively generated simulated noise into the original point cloud data and combining it with the denoised results, the core indicators for evaluating the point cloud denoising algorithm are calculated, achieving a comprehensive, objective, and reproducible quantitative evaluation of algorithm performance. Since it simulates two configurable quantitative data points—noise-filled and noise-free original point clouds—it can quantitatively evaluate the degree to which the denoised image restores the real image, providing a data-driven scientific basis for the selection of denoising algorithm parameters. By comparing the quantitative evaluation results under different parameters, the optimal parameter combination that achieves the best balance between denoising effect and data fidelity can be quickly found. Through a configurable parameterized process, this method can adapt to point cloud data of various types and scenarios, and transform the traditional inefficient trial-and-error process into an efficient and quantifiable test and analysis process. This makes the method a fair and consistent algorithm testing framework, enabling fair performance comparison of different algorithms or different parameter combinations under the same test conditions, providing reliable data support for technology selection and algorithm optimization.

[0033] In a preferred embodiment, in order to accurately distinguish between the reference point cloud and the simulated noise in the point cloud to be processed, the generated simulated noise is assigned a corresponding label; when selecting the reference point cloud, the original point cloud data with higher quality and less noise should be selected as much as possible.

[0034] In this embodiment, to achieve comprehensive quantitative evaluation from more dimensions, the calculated quantitative evaluation index includes at least two different evaluation parameters. Specifically, the calculated quantitative evaluation index includes both denoising rate and fidelity, that is, by introducing the dual-dimensional core indicators of denoising rate and fidelity, a more comprehensive and objective quantitative evaluation of algorithm performance is further achieved. Specifically, the calculation methods for denoising rate and fidelity are as follows: Based on the labels assigned to the simulated noise in the preceding steps, the total number of simulated noise in the original input is denoted as N_total_noise, and the total number of reference point clouds in the original input is denoted as N_total_base.

[0035] After denoising, statistics are performed on the outlier file E that was removed after denoising and the main point cloud file D that remains after denoising: In the outlier file E, the number of points labeled as simulated noise is denoted as N_removed_noise, which represents the noise points that were correctly removed. In the main point cloud file D, the label shows the number of points in the baseline point cloud, denoted as N_kept_base, which represents the valid points that were correctly preserved.

[0036] Based on the above statistical data, the calculation formula is as follows: (1) Denoising rate (R_denoise): Characterizes the algorithm's ability to remove noise, and is calculated using the following formula: R_denoise = ( N_removed_noise / N_total_noise ) × 100% (2) Fidelity (R_fidelity): Characterizes the algorithm's ability to retain effective point cloud data. The calculation formula is: R_fidelity = ( N_kept_base / N_total_base ) × 100% Reference Figure 1 In this embodiment, the calculated quantitative evaluation index is used to provide feedback on the values ​​of the noise configuration parameters and the denoising configuration parameters. This allows for simultaneous iterative optimization of both the noise simulation and denoising algorithms, avoiding the influence of subjectively set noise configuration parameters on the optimization of the denoising configuration parameters, while also improving the evaluation results.

[0037] In a preferred embodiment, the overall flow of the above-described method for quantizing and evaluating point cloud denoising algorithms is as follows: Figure 1 As shown, the main steps include: 1) Data preparation steps: Receive a high-quality, low-noise raw point cloud file as the reference point cloud data A.

[0038] 2) Controllable Noise Generation Step: Receive a set of configuration parameters for defining noise generation, including the selection of the noise generation method, the setting of the number of noise points, and specific parameters corresponding to the selected method. Based on the received set of configuration parameters, generate a simulated noise file B and assign it a unique classification label.

[0039] 3) Data integration step: Merge the baseline point cloud A with the simulated noise point B to generate the point cloud file C to be processed.

[0040] 4) Denoising Processing Steps: Receive a set of configuration parameters for defining the denoising process, including the selection of the denoising algorithm to be tested and the key parameters corresponding to the selected algorithm. Based on the received configuration parameters, denoise the point cloud file C to obtain the denoised main point cloud file D and the file E containing the removed outliers.

[0041] 5) Multi-dimensional quantitative evaluation step: Based on the processing results, a dual-dimensional core quantitative evaluation index is automatically calculated and output, such as denoising rate and fidelity. This evaluation index can be used for feedback to adjust the configuration parameters input in steps 2) and 4), thereby performing iterative optimization.

[0042] In addition, the method for quantifying and evaluating point cloud denoising algorithms also includes: visually displaying the final quantification evaluation index in tabular form, and displaying the intermediate process data used to calculate the quantification evaluation index; the intermediate process data includes the number of reference point cloud points, the number of generated simulated noise points, the total number of noise points removed, and the number of reference point cloud points that were mistakenly deleted.

[0043] To further enhance the visualization effect and facilitate a more intuitive viewing of the quantitative evaluation results, the method also includes: overlaying or displaying the baseline point cloud, simulated noise, the remaining main point cloud after denoising, and the outliers removed after denoising through a visualization rendering window, and distinguishing them by different colors.

[0044] The abstract denoising process and results are presented intuitively through a visualization window. This evaluation method, which combines qualitative and quantitative approaches, along with convenient parameter input and instant feedback mechanisms, provides users with the most intuitive and comprehensive understanding of algorithm performance, greatly improving the experience and efficiency of algorithm debugging and parameter optimization.

[0045] To facilitate users in configuring the relevant parameters required for the quantitative evaluation point cloud denoising algorithm, in this embodiment, the method of obtaining raw point cloud data includes: receiving the file corresponding to the raw point cloud data uploaded by the user through a file selection dialog box provided to the user, and obtaining the raw point cloud data. The noise configuration parameters can be obtained through a drop-down menu provided to the user for selecting a noise generation method, at least one input box for inputting specific parameter values ​​corresponding to the noise generation method, and a button or input box for specifying the number of noise points generated. The methods for obtaining denoising configuration parameters include: providing users with a drop-down menu for selecting the denoising algorithm to be tested, and at least one input box for setting the values ​​of the key parameters corresponding to the denoising algorithm.

[0046] In one specific embodiment, this method is implemented through software that includes a graphical user interface, as follows: a) Data loading module: Provides a file selection dialog box for receiving the benchmark point cloud file A uploaded by the user.

[0047] b) Noise generation configuration module: Provides a set of interactive controls for receiving noise generation parameters, including a drop-down menu for selecting the noise type, one or more numeric input boxes for inputting method-specific parameters, and a set of radio buttons and input boxes for specifying the noise quantity.

[0048] c) Denoising Processing Configuration Module: Provides a set of interactive controls for receiving denoising processing parameters, including a drop-down menu for selecting the denoising algorithm to be tested; and one or more numerical input boxes for setting key parameters of the algorithm.

[0049] d) Processing and execution module: contains one or more execution buttons, used to execute core algorithms such as noise generation, data integration and denoising processing according to the current configuration.

[0050] e) Quantitative evaluation and result display module, including an evaluation calculation unit that automatically calculates the denoising rate and fidelity in the background, and a result display table that clearly displays the final quantitative evaluation indicators in tabular form. The result display table can show the intermediate process data used to calculate these indicators, such as "original point cloud points", "number of generated noise points", "total number of noise points removed" and "number of original points accidentally deleted".

[0051] f) 3D visualization module: Provides a 3D rendering window that can overlay or display the original point cloud, injected noise, denoised main point cloud D, and removed outliers E, and can be distinguished by color.

[0052] The above examples of graphical user interfaces are as follows: Figure 2 As shown.

[0053] In this embodiment, the method of generating simulated noise points by configuring noise parameters includes: The noise generation method in the noise configuration parameters is the radial displacement outlier generation method. When the specific parameters corresponding to the radial displacement outlier generation method are the noise source point and the noise generation radius, a virtual sphere space with the noise generation radius is defined with the coordinates of the noise source point as the center. A candidate noise point is randomly generated within the global bounding box of the reference point cloud and outside the virtual sphere space. Then, the finally generated candidate noise point that meets the conditions is output as a simulated noise point. The noise generation method in the noise configuration parameters is the statistical outlier generation method. Given the specific parameters for this method—the number of neighboring points to be analyzed and the standard deviation factor—a spatial index is constructed for the baseline point cloud. The average nearest neighbor distance of all points in the baseline point cloud is calculated, and the global mean and global standard deviation of these average distances are further calculated. Based on the global mean, global standard deviation, and standard deviation factor, an average distance threshold for a target outlier is defined. A candidate noise point is generated, and its average distance to the nearest neighbor of the analyzed neighboring points in the baseline point cloud is calculated. If this average distance is less than the average distance threshold, the point is discarded and regenerated until it meets the statistical outlier condition that the average distance is not less than the average distance threshold. The candidate noise point that finally meets the condition is then output as a simulated noise point. The noise generation method in the noise configuration parameters is the Gaussian surface noise generation method. Given the specific parameters for the Gaussian surface noise generation method are the noise source point, the number of neighboring points, and the standard deviation of the Gaussian distribution, the method searches for the nearest neighbor point of the set number of neighboring points, centered on the noise source point, to form a neighborhood point set. Principal component analysis is performed on this point set to calculate the local surface normal vector at the noise source point. A value is randomly sampled from a normal distribution with a mean of 0 and a standard deviation equal to the set standard deviation of the Gaussian distribution, and this value is used as the perturbation distance along the normal direction. Based on the perturbation distance, the local surface normal vector, and the noise source point, the coordinates of the generated noise point are calculated, and the noise point is output. The noise generation method in the noise configuration parameters is the simulated multipath reflection cluster generation method. Under the specific parameters of the simulated multipath reflection cluster generation method, namely the noise cluster center point, the number of points in the cluster, and the standard deviation of the cluster distribution, the noise cluster center point is taken as the center in the sparse region of the reference point cloud or the external space. Following a three-dimensional Gaussian distribution with a standard deviation equal to the standard deviation of the cluster distribution, the coordinates of the number of points in the cluster are randomly generated. All generated points are output, which together constitute an independent point cluster as noise.

[0054] In one specific embodiment, the detailed implementation of the various user-selectable noise generation methods described above is as follows: Method for generating radial displacement outlier points: Principle and Simulation Scenario: This method aims to generate isolated outliers that are spatially distinct from the main body of the local point cloud. The method is not based on specific physical noise causes, but rather directly follows the working principle of denoising algorithms based on spatial neighborhood distance, such as radius filtering. Its core idea is to create a point that, within its specified neighborhood radius, contains no neighboring points from the original point cloud.

[0055] Test Objective: This method is used to verify the basic effectiveness of algorithms such as radius filtering. By correlating the noise generation radius with the algorithm's filtering radius parameter, it is possible to accurately and quantitatively evaluate the algorithm's theoretical detection limit and denoising effectiveness under specific parameters.

[0056] The implementation details are as follows: Input: reference point cloud A, noise source point P, noise generation radius R.

[0057] Implementation steps: a. Define a virtual spherical space S with radius R centered on the coordinates of the noise source point P.

[0058] b. Within the global bounding box of point cloud A, but outside the virtual sphere space S, randomly generate a candidate noise point P'.

[0059] c. Output the final generated P' that meets the conditions as a simulated noise point.

[0060] Methods for generating statistical outliers: Principle and Simulation Scenario: This method is designed based on the working principle of statistical filtering algorithms. The core of statistical filtering is to identify points whose neighborhood distance distribution exhibits anomalies. Therefore, this method aims to create a point whose average distance to its neighbors significantly deviates from the normal distribution characteristics of the entire point cloud in a statistically significant manner.

[0061] Test Objective: This test is specifically designed to validate the fundamental effectiveness of statistical filtering algorithms. By generating noise points that conform to the definition of statistical outliers, the recognition accuracy of the algorithm can be accurately evaluated under specific parameters of the number of neighboring points k and the standard deviation multiple f, thus transforming the performance evaluation of such algorithms from qualitative judgment to quantitative calculation.

[0062] The implementation details are as follows: Input: Baseline point cloud A, number of neighboring points to be analyzed k, standard deviation factor f.

[0063] Implementation steps: a. Construct spatial indexes such as kd-trees for the baseline point cloud A.

[0064] b. Calculate the average k-nearest neighbor distance for all points in point cloud A, and further calculate the global mean μ_d and global standard deviation σ_d of these average distances.

[0065] c. Define an average distance threshold for target outliers, T_target = μ_d + f × σ_d.

[0066] d. Generate a candidate noise point \(P_{candidate}\) and calculate the average distance \(d_{candidate}\) from it to the \(k\) nearest neighbor points in point cloud \(A\).

[0067] e. Iterative verification: If \(d_{candidate}<T_{target}\), discard this point and regenerate it until it meets the statistical outlier condition of \(d_{candidate}\geq T_{target}\).

[0068] f. Output the finally verified \(P_{candidate}\) as a simulated noise point.

[0069] Gaussian surface noise generation method: Principle and simulation scenario: This method simulates the tiny measurement errors attached near the real surface of an object. The ranging error of a high-precision laser scanner is often considered to approximately follow a Gaussian distribution with a mean of 0 in the direction perpendicular to the object surface in the physical model. This kind of noise is common in high-precision scanning tasks of smooth curved surfaces, and macroscopically shows the phenomenon of point cloud thickening or burrs.

[0070] Test purpose: It is mainly used to evaluate the data fidelity and fine feature preservation ability of the algorithm. An ideal algorithm should be able to smooth this kind of noise, that is, reduce the variance of the point cloud in the normal direction, rather than deleting it as an outlier. This test can quantify the balance ability of the algorithm between denoising and maintaining the fine geometric details of the object.

[0071] The implementation details are as follows: Input: Reference point cloud \(A\), noise source point \(P\), number of neighborhood points \(k\), standard deviation \(\sigma_g\) of the Gaussian distribution.

[0072] Implementation steps: a. Taking point \(P\) as the center, search for its \(k\) nearest neighbor points to form a neighborhood point set \(N_p\). Through principal component analysis of this point set, calculate the local surface normal vector \(n\) at point \(P\).

[0073] b. Randomly sample a value \(d\) from a normal distribution with a mean of 0 and a standard deviation of \(\sigma_g\) as the perturbation distance along the normal direction.

[0074] c. Calculate the coordinates of the generated noise point \(P'\): \(P' = P + d\times n\).

[0075] d. Output the noise point \(P'\).

[0076] Simulated multipath reflection cluster generation method: Principle and Simulation Scenario: This method simulates the multipath reflection effect commonly seen in complex environments such as indoor corners, vehicle undercarriages, and under vegetation canopies for lidar. This effect refers to the laser beam undergoing one or more reflections between multiple surfaces before hitting its final target, causing the sensor to receive a delayed signal, thus forming misplaced "ghost images" in space. These ghost images typically appear as small, low-density clusters of independent points.

[0077] Test Objective: This test primarily evaluates the algorithm's ability to distinguish locally clustered noise. Simple outlier removal algorithms may be ineffective against such noise clusters, and might even misidentify them as real small objects. This test measures the algorithm's intelligence, specifically its ability to effectively identify and remove these pseudo-feature point clusters generated by physical effects without mistakenly deleting real small objects.

[0078] The implementation details are as follows: Input: Baseline point cloud A, center point C of noise cluster, number of points within the cluster N_c, standard deviation of cluster distribution σ_c.

[0079] Implementation steps: a. Define the center point C of a noise cluster in the sparse region or external space of point cloud A.

[0080] b. With C as the center, follow a three-dimensional Gaussian distribution (whose covariance matrix can define the shape of the cluster, which is isotropic in the simplest case and has a standard deviation of σ_c), and randomly generate the coordinates of N_c points.

[0081] c. Output all generated points, which together form an independent point cluster as noise.

[0082] Furthermore, in this embodiment, the method of denoising the generated point cloud to be processed by configuring denoising parameters includes: The denoising algorithm to be tested in the denoising configuration parameters is the radius filtering algorithm. Under the condition that the key parameters of the radius filtering algorithm are the search radius and the minimum number of neighboring points, a spatial index is constructed for the point cloud to be processed to efficiently perform neighborhood search. Each point in the point cloud to be processed is traversed, and a radius search is performed on each point using the spatial index to find all neighboring points whose Euclidean distance to the point is within the search radius. The number of neighboring points found is counted. If the number of neighboring points is less than the minimum number of neighboring points, the point is classified as an outlier. Otherwise, the point is kept in the remaining main point cloud after denoising. Finally, the main point cloud and the outlier are returned. The denoising algorithm to be tested in the denoising configuration parameters is a statistical filtering algorithm. Given that the key parameters for the statistical filtering algorithm are the number of neighboring points analyzed and the standard deviation multiple threshold, a kd-tree is constructed for the point cloud to be processed. For each point in the point cloud, its nearest neighbor (based on the number of neighboring points analyzed) is found, and the average distance from that point to these nearest neighbors is calculated. The global average and global standard deviation of all average distances are then calculated. Based on the global average, global standard deviation, and standard deviation multiple, a global distance threshold is calculated. For each point in the point cloud to be processed, if the average distance from that point to its nearest neighbor is greater than the global distance threshold, that point is classified as an outlier; otherwise, it is retained in the main point cloud. Finally, the main point cloud and the outlier points are returned. The denoising algorithm to be tested in the denoising configuration parameters is a bilateral filtering algorithm. Given that the key parameters for the bilateral filtering algorithm are spatial standard deviation and feature standard deviation, a spatial index is constructed for the point cloud to be processed. Each point in the point cloud is traversed, and a neighborhood is determined for that point. For each point within the neighborhood, a spatial weight is calculated based on the Euclidean distance between that point and the corresponding point in the point cloud, and a Gaussian function of the spatial standard deviation. A local tangent plane is then calculated at the point in the point cloud. The perpendicular distance from the points in the neighborhood to this tangent plane is calculated, and feature weights are calculated based on the Gaussian function of the perpendicular distance and the feature standard deviation. Based on the feature weights and spatial weights, the final weights of the points in the neighborhood are calculated. The new position of a point in the point cloud is the weighted average of all its neighboring points based on the final weights. The resulting main point cloud, composed of all the new positions and smoothed with denoising, is returned.

[0083] In one specific embodiment, the detailed implementation of the various user-selectable denoising algorithms to be tested described above is as follows: (1) Radius filtering algorithm: This algorithm is an outlier removal technique based on local point density. Its core idea is that for a given point, if the number of its neighbors within a certain spatial radius is less than a threshold, the point is considered sparse and isolated, and should be judged as noise and removed.

[0084] Key adjustable parameters: Search radius R, which defines the size of the spherical neighborhood space used to count the number of neighbors. Minimum number of neighbors K, which defines the minimum number of neighbors required to determine that a point is "non-noise" (i.e., a point in a dense region).

[0085] Implementation steps: a. Construct spatial indexes such as kd-trees for the point cloud C to be processed, in order to perform neighborhood search efficiently.

[0086] b. Traverse every point Pi in the point cloud C.

[0087] c. Neighborhood search: For each Pi, perform a radius search using a spatial index to find all neighboring points whose Euclidean distance from Pi is within R.

[0088] d. Condition judgment: Count the number of found neighboring points (excluding Pi itself), denoted as Count.

[0089] e. Classification: If Count < K, classify the point Pi into the outlier (noise) file E; otherwise, retain Pi in the denoised main point cloud file D.

[0090] f. Output: Return the main point cloud file D and the outlier file E.

[0091] The performance of the radius filtering of this algorithm is completely determined by the two parameters R and K. The method of quantitatively evaluating the point cloud denoising algorithm in this embodiment provides a direct means of quantitative evaluation. By generating known noise through methods such as the method of injecting radial displacement outliers, this evaluation method can accurately quantify the denoising rate and fidelity rate under different (R, K) parameter combinations. This enables users to find the best balance between effectively removing outliers and avoiding misdeleting sparse feature points in a data-driven manner.

[0092] (2) Statistical filtering algorithm: This algorithm is an outlier removal technique based on the statistical characteristics of the local distance distribution of the point cloud. It assumes that the neighborhood distance distribution of most points should conform to a statistical law such as the Gaussian distribution. For a given point, if its average distance to neighboring points significantly deviates from the global statistical distribution of the entire point cloud, then this point is determined to be a statistical outlier and is removed.

[0093] Key adjustable parameters: The number of neighboring points N for analysis, which defines the neighborhood size for calculating the local average distance of each point. The standard deviation multiple threshold StdDev_Mul, which defines the strictness of determining outliers. The larger this value, the more distant a point must deviate from the global mean to be determined as noise.

[0094] Implementation steps: a. Construct a k-d tree for the point cloud C to be processed.

[0095] b. Local statistics: Traverse each point Pi in the point cloud C, find its N nearest neighboring points, and calculate the average distance d_i from Pi to these N points.

[0096] c. Global statistics: Calculate the global average μ_d and global standard deviation σ_d of all d_i values.

[0097] d. Threshold calculation: Calculate a global distance threshold T = μ_d + StdDev_Mul × σ_d based on the standard deviation multiple.

[0098] e. Classification: Iterate through each point Pi again. If its local average distance d_i > T, then classify Pi into the outlier file E; otherwise, keep it in the main point cloud file D.

[0099] f. Output: Returns the main point cloud file D and the outlier file E.

[0100] The selection of parameters for statistical filtering in this algorithm is highly dependent on experience. The quantitative evaluation method for point cloud denoising algorithms in this implementation transforms this process into a scientific quantitative analysis. By injecting noise generated by either the "statistical outlier generation method" or the "global random shot noise generation method," this evaluation method can clearly demonstrate the "denoising rate" and "fidelity" under different combinations of (N, StdDev_Mul) parameters, thereby helping users find the statistical filtering parameters most suitable for the current data distribution characteristics.

[0101] (3) Bilateral filtering algorithm: Bilateral filtering is a nonlinear, edge-preserving, smooth noise reduction technique. Unlike traditional filters that only consider spatial distance, it considers two factors simultaneously when calculating the neighborhood weighted average of a point: spatial proximity (the closer the neighboring points are to the center point, the higher their weight) and feature similarity (the smaller the feature differences between neighboring points and the center point, the higher their weight). In point cloud geometry processing, this "feature" is typically the distance from the neighboring points to the local tangent plane of the center point.

[0102] Through this dual weighting mechanism, bilateral filtering can effectively preserve geometric details such as edges and corners while smoothing flat areas (where neighboring points have high feature similarity). Points on both sides of the edge have low weights due to large feature differences and will not affect each other.

[0103] Key adjustable parameters: Spatial standard deviation σ_s controls the weight of spatial proximity. A larger value means that points farther away will also be included in the smoothing calculation, resulting in a stronger smoothing effect. Distance / feature standard deviation σ_r controls the weight of feature similarity. A larger value means that points with large feature differences (i.e., points farther away from the local tangent plane) will also be included in the calculation, leading to a decrease in edge preservation ability and an enhanced smoothing effect.

[0104] Implementation steps: a. Construct spatial indexes such as kd-trees for the point cloud C to be processed.

[0105] b. Traverse every point Pi in the point cloud C.

[0106] c. Neighborhood definition: Determine a neighborhood of Pi.

[0107] d. Weight Calculation: For each point Pj in the neighborhood, perform the following steps i-v: i. Calculate the spatial weights w_s, which are typically Gaussian functions based on the Euclidean distance between points Pi and Pj and the spatial standard deviation σ_s.

[0108] ii. Calculate the local tangent plane at Pi.

[0109] iii. Calculate the perpendicular distance d_plane from point Pj to the tangent plane.

[0110] iv. Calculate the feature weights w_r, typically a Gaussian function based on the distance d_plane and the feature standard deviation σ_r.

[0111] v. Calculate the final weight of Pj, w_final = w_s × w_r.

[0112] e. Position Update: The new position P'i of point Pi is the weighted average of all its neighbors based on the final weight w_final.

[0113] f. Output: Returns a smoothed and denoised main point cloud file D consisting of all new locations P'i.

[0114] The performance of this bilateral filtering algorithm is highly sensitive to the balance between the parameters σ_s and σ_r. The quantization evaluation method for the point cloud denoising algorithm in this embodiment provides a crucial quantization evaluation mechanism for this purpose. By injecting noise generated using a Gaussian surface noise generation method, this evaluation method can accurately quantify the noise smoothing and geometric feature preservation of the algorithm under different combinations of (σ_s, σ_r) parameters. This allows users to find the optimal parameter pair that achieves the best balance between noise smoothing and detail preservation through a data-driven approach.

[0115] System Implementation Method for Quantitative Evaluation of Point Cloud Denoising Algorithms This embodiment provides a technical solution for a system for quantizing and evaluating point cloud denoising algorithms, including a processor. The processor stores executable program instructions, which are executed to implement the following method for quantizing and evaluating point cloud denoising algorithms: A set of original point cloud data is used as the baseline point cloud; simulated noise points are generated by noise configuration parameters; the noise configuration parameters include the noise generation method, the specific parameters corresponding to the noise generation method, and the number of noise points. The baseline point cloud and simulated noise points are merged to generate a point cloud to be processed. Then, the generated point cloud to be processed is denoised using denoising configuration parameters to obtain the remaining main point cloud after denoising and the outliers removed after denoising. The denoising configuration parameters include the denoising algorithm to be tested and the key parameters corresponding to the denoising algorithm. Based on at least one of the remaining main point cloud after denoising and the outliers removed after denoising, as well as at least one of the baseline point cloud and simulated noise, a quantitative evaluation index is calculated to quantitatively evaluate the point cloud denoising algorithm.

[0116] Therefore, through a process that responds to external parameter configuration, simulated noise with known characteristics is added to the baseline point cloud data, and then the algorithm under test is invoked for processing. This system effectively overcomes the drawbacks of subjective evaluation. By introducing actively generated simulated noise into the original point cloud data and combining it with the denoised results, it calculates the core indicators for evaluating the point cloud denoising algorithm, achieving a comprehensive, objective, and reproducible quantitative evaluation of algorithm performance. Since it simulates both noisy and noise-free original point clouds—two configurable quantitative data points—it can quantitatively evaluate the degree to which the denoised image restores the true image, providing a data-driven scientific basis for parameter selection in the denoising algorithm. By comparing the quantitative evaluation results under different parameters, it can quickly find the parameter combination that achieves the optimal balance between denoising effect and data fidelity. Through a configurable parameterized process, the system can adapt to various types and scenarios of point cloud data, transforming the traditional inefficient trial-and-error process into an efficient and quantifiable testing and analysis process. This makes the system a fair and consistent algorithm testing framework, enabling impartial performance comparisons of different algorithms or parameter combinations under the same testing conditions, providing reliable data support for technology selection and algorithm optimization.

[0117] In a preferred embodiment, in order to accurately distinguish between the reference point cloud and the simulated noise in the point cloud to be processed, the generated simulated noise is assigned a corresponding label; when selecting the reference point cloud, the original point cloud data with higher quality and less noise should be selected as much as possible.

[0118] In this embodiment, to achieve comprehensive quantitative evaluation from more dimensions, the calculated quantitative evaluation index includes at least two different evaluation parameters. Specifically, the calculated quantitative evaluation index includes both denoising rate and fidelity, that is, by introducing the dual-dimensional core indicators of denoising rate and fidelity, a more comprehensive and objective quantitative evaluation of algorithm performance is further achieved. Specifically, the calculation methods for denoising rate and fidelity are as follows: Based on the labels assigned to the simulated noise in the preceding steps, the total number of simulated noise in the original input is denoted as N_total_noise, and the total number of reference point clouds in the original input is denoted as N_total_base.

[0119] After denoising, statistics are performed on the outlier file E that was removed after denoising and the main point cloud file D that remains after denoising: In the outlier file E, the number of points labeled as simulated noise is denoted as N_removed_noise, which represents the noise points that were correctly removed. In the main point cloud file D, the label shows the number of points in the baseline point cloud, denoted as N_kept_base, which represents the valid points that were correctly preserved.

[0120] Based on the above statistical data, the calculation formula is as follows: (1) Denoising rate (R_denoise): Characterizes the algorithm's ability to remove noise, and is calculated using the following formula: R_denoise = ( N_removed_noise / N_total_noise ) × 100% (2) Fidelity (R_fidelity): Characterizes the algorithm's ability to retain effective point cloud data. The calculation formula is: R_fidelity = ( N_kept_base / N_total_base ) × 100% In this embodiment, the calculated quantitative evaluation index is used to provide feedback on the values ​​of the noise configuration parameters and the denoising configuration parameters. This allows for simultaneous iterative optimization of both the noise simulation and denoising algorithms, thereby avoiding the influence of subjectively set noise configuration parameters on the optimization of the denoising configuration parameters, and also improving the evaluation results.

[0121] In a preferred embodiment, the overall process of the above-described method for quantizing and evaluating point cloud denoising algorithms mainly includes the following steps: 1) Data preparation steps: Receive a high-quality, low-noise raw point cloud file as the reference point cloud data A.

[0122] 2) Controllable Noise Generation Step: Receive a set of configuration parameters for defining noise generation, including the selection of the noise generation method, the setting of the number of noise points, and specific parameters corresponding to the selected method. Based on the received set of configuration parameters, generate a simulated noise file B and assign it a unique classification label.

[0123] 3) Data integration step: Merge the baseline point cloud A with the simulated noise point B to generate the point cloud file C to be processed.

[0124] 4) Denoising Processing Steps: Receive a set of configuration parameters for defining the denoising process, including the selection of the denoising algorithm to be tested and the key parameters corresponding to the selected algorithm. Based on the received configuration parameters, denoise the point cloud file C to obtain the denoised main point cloud file D and the file E containing the removed outliers.

[0125] 5) Multi-dimensional quantitative evaluation step: Based on the processing results, a dual-dimensional core quantitative evaluation index is automatically calculated and output, such as denoising rate and fidelity. This evaluation index can be used for feedback to adjust the configuration parameters input in steps 2) and 4), thereby performing iterative optimization.

[0126] In addition, the above-mentioned method for quantitative evaluation of point cloud denoising algorithms also includes: visually displaying the final quantitative evaluation index in tabular form, and displaying the intermediate process data used to calculate the quantitative evaluation index; the intermediate process data includes the number of reference point cloud points, the number of generated simulated noise points, the total number of noise points removed, and the number of reference point cloud points that were mistakenly deleted.

[0127] To further enhance the visualization effect and facilitate a more intuitive viewing of the quantitative evaluation results, the aforementioned quantitative evaluation point cloud denoising algorithm also includes: overlaying or displaying the baseline point cloud, simulated noise, the remaining main point cloud after denoising, and the outliers removed after denoising through a visualization rendering window, and distinguishing them by different colors.

[0128] The abstract denoising process and results are presented intuitively through a visualization window. This evaluation method, which combines qualitative and quantitative approaches, along with convenient parameter input and instant feedback mechanisms, provides users with the most intuitive and comprehensive understanding of algorithm performance, greatly improving the experience and efficiency of algorithm debugging and parameter optimization.

[0129] To facilitate users in configuring the relevant parameters required for the quantitative evaluation point cloud denoising algorithm, in this embodiment, the method of obtaining raw point cloud data includes: receiving the file corresponding to the raw point cloud data uploaded by the user through a file selection dialog box provided to the user, and obtaining the raw point cloud data. The noise configuration parameters can be obtained through a drop-down menu provided to the user for selecting a noise generation method, at least one input box for inputting specific parameter values ​​corresponding to the noise generation method, and a button or input box for specifying the number of noise points generated. The methods for obtaining denoising configuration parameters include: providing users with a drop-down menu for selecting the denoising algorithm to be tested, and at least one input box for setting the values ​​of the key parameters corresponding to the denoising algorithm.

[0130] In one specific embodiment, the system includes software with a graphical user interface, as detailed below: a) Data loading module: Provides a file selection dialog box for receiving the benchmark point cloud file A uploaded by the user.

[0131] b) Noise generation configuration module: Provides a set of interactive controls for receiving noise generation parameters, including a drop-down menu for selecting the noise type, one or more numeric input boxes for inputting method-specific parameters, and a set of radio buttons and input boxes for specifying the noise quantity.

[0132] c) Denoising Processing Configuration Module: Provides a set of interactive controls for receiving denoising processing parameters, including a drop-down menu for selecting the denoising algorithm to be tested; and one or more numerical input boxes for setting key parameters of the algorithm.

[0133] d) Processing and execution module: contains one or more execution buttons, used to execute core algorithms such as noise generation, data integration and denoising processing according to the current configuration.

[0134] e) Quantitative evaluation and result display module, including an evaluation calculation unit that automatically calculates the denoising rate and fidelity in the background, and a result display table that clearly displays the final quantitative evaluation indicators in tabular form. The result display table can show the intermediate process data used to calculate these indicators, such as "original point cloud points", "number of generated noise points", "total number of noise points removed" and "number of original points accidentally deleted".

[0135] f) 3D visualization module: Provides a 3D rendering window that can overlay or display the original point cloud, injected noise, denoised main point cloud D, and removed outliers E, and can be distinguished by color.

[0136] In this embodiment, the method of generating simulated noise points by configuring noise parameters includes: The noise generation method in the noise configuration parameters is the radial displacement outlier generation method. When the specific parameters corresponding to the radial displacement outlier generation method are the noise source point and the noise generation radius, a virtual sphere space with the noise generation radius is defined with the coordinates of the noise source point as the center. A candidate noise point is randomly generated within the global bounding box of the reference point cloud and outside the virtual sphere space. Then, the finally generated candidate noise point that meets the conditions is output as a simulated noise point. The noise generation method in the noise configuration parameters is the statistical outlier generation method. Given the specific parameters for this method—the number of neighboring points to be analyzed and the standard deviation factor—a spatial index is constructed for the baseline point cloud. The average nearest neighbor distance of all points in the baseline point cloud is calculated, and the global mean and global standard deviation of these average distances are further calculated. Based on the global mean, global standard deviation, and standard deviation factor, an average distance threshold for a target outlier is defined. A candidate noise point is generated, and its average distance to the nearest neighbor of the analyzed neighboring points in the baseline point cloud is calculated. If this average distance is less than the average distance threshold, the point is discarded and regenerated until it meets the statistical outlier condition that the average distance is not less than the average distance threshold. The candidate noise point that finally meets the condition is then output as a simulated noise point. The noise generation method in the noise configuration parameters is the Gaussian surface noise generation method. Given the specific parameters for the Gaussian surface noise generation method are the noise source point, the number of neighboring points, and the standard deviation of the Gaussian distribution, the method searches for the nearest neighbor point of the set number of neighboring points, centered on the noise source point, to form a neighborhood point set. Principal component analysis is performed on this point set to calculate the local surface normal vector at the noise source point. A value is randomly sampled from a normal distribution with a mean of 0 and a standard deviation equal to the set standard deviation of the Gaussian distribution, and this value is used as the perturbation distance along the normal direction. Based on the perturbation distance, the local surface normal vector, and the noise source point, the coordinates of the generated noise point are calculated, and the noise point is output. The noise generation method in the noise configuration parameters is the simulated multipath reflection cluster generation method. Under the specific parameters of the simulated multipath reflection cluster generation method, namely the noise cluster center point, the number of points in the cluster, and the standard deviation of the cluster distribution, the noise cluster center point is taken as the center in the sparse region of the reference point cloud or the external space. Following a three-dimensional Gaussian distribution with a standard deviation equal to the standard deviation of the cluster distribution, the coordinates of the number of points in the cluster are randomly generated. All generated points are output, which together constitute an independent point cluster as noise.

[0137] In one specific embodiment, the detailed implementation of the various user-selectable noise generation methods described above is as follows: Method for generating radial displacement outlier points: Principle and Simulation Scenario: This method aims to generate isolated outliers that are spatially distinct from the main body of the local point cloud. The method is not based on specific physical noise causes, but rather directly follows the working principle of denoising algorithms based on spatial neighborhood distance, such as radius filtering. Its core idea is to create a point that, within its specified neighborhood radius, contains no neighboring points from the original point cloud.

[0138] Test Objective: This method is used to verify the basic effectiveness of algorithms such as radius filtering. By correlating the noise generation radius with the algorithm's filtering radius parameter, it is possible to accurately and quantitatively evaluate the algorithm's theoretical detection limit and denoising effectiveness under specific parameters.

[0139] The implementation details are as follows: Input: reference point cloud A, noise source point P, noise generation radius R.

[0140] Implementation steps: a. Define a virtual spherical space S with radius R centered on the coordinates of the noise source point P.

[0141] b. Within the global bounding box of point cloud A, but outside the virtual sphere space S, randomly generate a candidate noise point P'.

[0142] c. Output the final generated P' that meets the conditions as a simulated noise point.

[0143] Methods for generating statistical outliers: Principle and Simulation Scenario: This method is designed based on the working principle of statistical filtering algorithms. The core of statistical filtering is to identify points whose neighborhood distance distribution exhibits anomalies. Therefore, this method aims to create a point whose average distance to its neighbors significantly deviates from the normal distribution characteristics of the entire point cloud in a statistically significant manner.

[0144] Test Objective: This test is specifically designed to validate the fundamental effectiveness of statistical filtering algorithms. By generating noise points that conform to the definition of statistical outliers, the recognition accuracy of the algorithm can be accurately evaluated under specific parameters of the number of neighboring points k and the standard deviation multiple f, thus transforming the performance evaluation of such algorithms from qualitative judgment to quantitative calculation.

[0145] The implementation details are as follows: Input: Baseline point cloud A, number of neighboring points to be analyzed k, standard deviation factor f.

[0146] Implementation steps: a. Construct spatial indexes such as kd-trees for the baseline point cloud A.

[0147] b. Calculate the average k-nearest neighbor distance for all points in point cloud A, and further calculate the global mean μ_d and global standard deviation σ_d of these average distances.

[0148] c. Define an average distance threshold for target outliers, T_target = μ_d + f × σ_d.

[0149] d. Generate a candidate noise point \(P_{candidate}\), and calculate the average distance \(d_{candidate}\) from it to the \(k\) nearest neighbor points in point cloud \(A\).

[0150] e. Iterative verification: If \(d_{candidate}<T_{target}\), discard this point and regenerate it until it meets the statistical outlier condition of \(d_{candidate}\geq T_{target}\).

[0151] f. Output the finally verified \(P_{candidate}\) as a simulated noise point.

[0152] Gaussian surface noise generation method: Principle and simulation scenario: This method simulates the tiny measurement errors attached near the real surface of an object. The ranging error of a high-precision laser scanner is often considered to approximately follow a Gaussian distribution with a mean of 0 in the direction perpendicular to the object surface in the physical model. This kind of noise is common in high-precision scanning tasks of smooth surfaces and macroscopically shows the phenomenon of point cloud thickening or burrs.

[0153] Test purpose: It is mainly used to evaluate the data fidelity and fine feature preservation ability of the algorithm. An ideal algorithm should be able to smooth this kind of noise, that is, reduce the variance of the point cloud in the normal direction, rather than deleting it as an outlier. This test can quantify the balance ability of the algorithm between denoising and maintaining the fine geometric details of the object.

[0154] Implementation details are as follows: Input: Reference point cloud \(A\), noise source point \(P\), number of neighborhood points \(k\), standard deviation \(\sigma_g\) of the Gaussian distribution.

[0155] Implementation steps: a. With point \(P\) as the center, search for its \(k\) nearest neighbor points to form a neighborhood point set \(N_p\). Through principal component analysis of this point set, calculate the local surface normal vector \(n\) at point \(P\).

[0156] b. Randomly sample a value \(d\) from a normal distribution with a mean of 0 and a standard deviation of \(\sigma_g\) as the perturbation distance along the normal direction.

[0157] c. Calculate the coordinates of the generated noise point \(P'\): \(P' = P + d\times n\).

[0158] d. Output the noise point \(P'\).

[0159] Simulated multipath reflection cluster generation method: Principle and Simulation Scenario: This method simulates the multipath reflection effect commonly seen in complex environments such as indoor corners, vehicle undercarriages, and under vegetation canopies for lidar. This effect refers to the laser beam undergoing one or more reflections between multiple surfaces before hitting its final target, causing the sensor to receive a delayed signal, thus forming misplaced "ghost images" in space. These ghost images typically appear as small, low-density clusters of independent points.

[0160] Test Objective: This test primarily evaluates the algorithm's ability to distinguish locally clustered noise. Simple outlier removal algorithms may be ineffective against such noise clusters, and might even misidentify them as real small objects. This test measures the algorithm's intelligence, specifically its ability to effectively identify and remove these pseudo-feature point clusters generated by physical effects without mistakenly deleting real small objects.

[0161] The implementation details are as follows: Input: Baseline point cloud A, center point C of noise cluster, number of points within the cluster N_c, standard deviation of cluster distribution σ_c.

[0162] Implementation steps: a. Define the center point C of a noise cluster in the sparse region or external space of point cloud A.

[0163] b. With C as the center, follow a three-dimensional Gaussian distribution (whose covariance matrix can define the shape of the cluster, which is isotropic in the simplest case and has a standard deviation of σ_c), and randomly generate the coordinates of N_c points.

[0164] c. Output all generated points, which together form an independent point cluster as noise.

[0165] Furthermore, in this embodiment, the method of denoising the generated point cloud to be processed by configuring denoising parameters includes: The denoising algorithm to be tested in the denoising configuration parameters is the radius filtering algorithm. Under the condition that the key parameters of the radius filtering algorithm are the search radius and the minimum number of neighboring points, a spatial index is constructed for the point cloud to be processed to efficiently perform neighborhood search. Each point in the point cloud to be processed is traversed, and a radius search is performed on each point using the spatial index to find all neighboring points whose Euclidean distance to the point is within the search radius. The number of neighboring points found is counted. If the number of neighboring points is less than the minimum number of neighboring points, the point is classified as an outlier. Otherwise, the point is kept in the remaining main point cloud after denoising. Finally, the main point cloud and the outlier are returned. The denoising algorithm to be tested in the denoising configuration parameters is a statistical filtering algorithm. Given that the key parameters for the statistical filtering algorithm are the number of neighboring points analyzed and the standard deviation multiple threshold, a kd-tree is constructed for the point cloud to be processed. For each point in the point cloud, its nearest neighbor (based on the number of neighboring points analyzed) is found, and the average distance from that point to these nearest neighbors is calculated. The global average and global standard deviation of all average distances are then calculated. Based on the global average, global standard deviation, and standard deviation multiple, a global distance threshold is calculated. For each point in the point cloud to be processed, if the average distance from that point to its nearest neighbor is greater than the global distance threshold, that point is classified as an outlier; otherwise, it is retained in the main point cloud. Finally, the main point cloud and the outlier points are returned. The denoising algorithm to be tested in the denoising configuration parameters is a bilateral filtering algorithm. Given that the key parameters for the bilateral filtering algorithm are spatial standard deviation and feature standard deviation, a spatial index is constructed for the point cloud to be processed. Each point in the point cloud is traversed, and a neighborhood is determined for that point. For each point within the neighborhood, a spatial weight is calculated based on the Euclidean distance between that point and the corresponding point in the point cloud, and a Gaussian function of the spatial standard deviation. A local tangent plane is then calculated at the point in the point cloud. The perpendicular distance from the points in the neighborhood to this tangent plane is calculated, and feature weights are calculated based on the Gaussian function of the perpendicular distance and the feature standard deviation. Based on the feature weights and spatial weights, the final weights of the points in the neighborhood are calculated. The new position of a point in the point cloud is the weighted average of all its neighboring points based on the final weights. The resulting main point cloud, composed of all the new positions and smoothed with denoising, is returned.

[0166] In one specific embodiment, the detailed implementation of the various user-selectable denoising algorithms to be tested described above is as follows: (1) Radius filtering algorithm: This algorithm is an outlier removal technique based on local point density. Its core idea is that for a given point, if the number of its neighbors within a certain spatial radius is less than a threshold, the point is considered sparse and isolated, and should be judged as noise and removed.

[0167] Key adjustable parameters: Search radius R, which defines the size of the spherical neighborhood space used to count the number of neighbors. Minimum number of neighbors K, which defines the minimum number of neighbors required to determine that a point is "non-noise" (i.e., a point in a dense region).

[0168] Implementation steps: a. Construct spatial indexes such as kd-trees for the point cloud C to be processed, in order to perform neighborhood search efficiently.

[0169] b. Traverse every point Pi in the point cloud C.

[0170] c. Neighborhood search: For each Pi, perform a radius search using a spatial index to find all neighboring points whose Euclidean distance from Pi is within R.

[0171] d. Condition judgment: Count the number of found neighboring points (excluding Pi itself), denoted as Count.

[0172] e. Classification: If Count < K, classify the point Pi into the outlier (noise) file E; otherwise, retain Pi in the denoised main point cloud file D.

[0173] f. Output: Return the main point cloud file D and the outlier file E.

[0174] The performance of the radius filtering of this algorithm is completely determined by the two parameters R and K. The system for quantitatively evaluating the point cloud denoising algorithm in this embodiment provides a direct means of quantitative evaluation. By injecting known noises generated by methods such as the radial displacement outlier generation method, this evaluation system can accurately quantify the denoising rate and fidelity rate under different (R, K) parameter combinations. This enables users to find the best balance between effectively removing outliers and avoiding misdeleting sparse feature points in a data-driven manner.

[0175] (2) Statistical filtering algorithm: This algorithm is an outlier removal technique based on the statistical characteristics of the local distance distribution of the point cloud. It assumes that the neighborhood distance distribution of most points should conform to a statistical law such as the Gaussian distribution. For a given point, if its average distance to neighboring points significantly deviates from the global statistical distribution of the entire point cloud, then this point is determined to be a statistical outlier and is removed.

[0176] Key adjustable parameters: The number of neighboring points N for analysis, which defines the neighborhood size used to calculate the local average distance of each point. The standard deviation multiple threshold StdDev_Mul, which defines the strictness of determining outliers. The larger this value, the more it means that only points that deviate further from the global mean will be determined as noise.

[0177] Implementation steps: a. Construct a k-d tree for the point cloud C to be processed.

[0178] b. Local statistics: Traverse each point Pi in the point cloud C, find its N nearest neighboring points, and calculate the average distance d_i from Pi to these N points.

[0179] c. Global statistics: Calculate the global average μ_d and global standard deviation σ_d of all d_i values.

[0180] d. Threshold calculation: Calculate a global distance threshold T = μ_d + StdDev_Mul × σ_d based on the standard deviation multiple.

[0181] e. Classification: Iterate through each point Pi again. If its local average distance d_i > T, then classify Pi into the outlier file E; otherwise, keep it in the main point cloud file D.

[0182] f. Output: Returns the main point cloud file D and the outlier file E.

[0183] The selection of parameters for statistical filtering in this algorithm is highly dependent on experience. The system for quantifying and evaluating point cloud denoising algorithms in this implementation transforms this process into a scientific quantitative analysis. By injecting noise generated by either the "statistical outlier generation method" or the "global random shot noise generation method," this evaluation system can clearly demonstrate the "denoising rate" and "fidelity" under different combinations of (N, StdDev_Mul) parameters, thereby helping users find the statistical filtering parameters most suitable for the current data distribution characteristics.

[0184] (3) Bilateral filtering algorithm: Bilateral filtering is a nonlinear, edge-preserving, smooth noise reduction technique. Unlike traditional filters that only consider spatial distance, it considers two factors simultaneously when calculating the neighborhood weighted average of a point: spatial proximity (the closer the neighboring points are to the center point, the higher their weight) and feature similarity (the smaller the feature differences between neighboring points and the center point, the higher their weight). In point cloud geometry processing, this "feature" is typically the distance from the neighboring points to the local tangent plane of the center point.

[0185] Through this dual weighting mechanism, bilateral filtering can effectively preserve geometric details such as edges and corners while smoothing flat areas (where neighboring points have high feature similarity). Points on both sides of the edge have low weights due to large feature differences and will not affect each other.

[0186] Key adjustable parameters: Spatial standard deviation σ_s controls the weight of spatial proximity. A larger value means that points farther away will also be included in the smoothing calculation, resulting in a stronger smoothing effect. Distance / feature standard deviation σ_r controls the weight of feature similarity. A larger value means that points with large feature differences (i.e., points farther away from the local tangent plane) will also be included in the calculation, leading to a decrease in edge preservation ability and an enhanced smoothing effect.

[0187] Implementation steps: a. Construct spatial indexes such as kd-trees for the point cloud C to be processed.

[0188] b. Traverse every point Pi in the point cloud C.

[0189] c. Neighborhood definition: Determine a neighborhood of Pi.

[0190] d. Weight Calculation: For each point Pj in the neighborhood, perform the following steps i-v: i. Calculate the spatial weights w_s, which are typically Gaussian functions based on the Euclidean distance between points Pi and Pj and the spatial standard deviation σ_s.

[0191] ii. Calculate the local tangent plane at Pi.

[0192] iii. Calculate the perpendicular distance d_plane from point Pj to the tangent plane.

[0193] iv. Calculate the feature weights w_r, typically a Gaussian function based on the distance d_plane and the feature standard deviation σ_r.

[0194] v. Calculate the final weight of Pj, w_final = w_s × w_r.

[0195] e. Position Update: The new position P'i of point Pi is the weighted average of all its neighbors based on the final weight w_final.

[0196] f. Output: Returns a smoothed and denoised main point cloud file D consisting of all new locations P'i.

[0197] The performance of this bilateral filtering algorithm is highly sensitive to the balance between the parameters σ_s and σ_r. The quantization evaluation system for the point cloud denoising algorithm in this embodiment provides a crucial quantization evaluation mechanism for this purpose. By injecting noise generated using a Gaussian surface noise generation method, this evaluation system can accurately quantify the noise smoothing and geometric feature preservation of the algorithm under different combinations of (σ_s, σ_r) parameters. This allows users to find the optimal parameter pair that achieves the best balance between noise smoothing and detail preservation through a data-driven approach.

[0198] It should be understood that the above-described specific embodiments of the present invention are merely illustrative or explanatory of the principles of the present invention, and do not constitute a limitation thereof.

Claims

1. A method of quantitatively evaluating a point cloud denoising algorithm, characterized in that, include: Use a set of raw point cloud data as the baseline point cloud; Simulated noise is generated by configuring noise parameters; The noise configuration parameters include the noise generation method, its corresponding specific parameters, and the number of noise points. The baseline point cloud is merged with the simulated noise to generate the point cloud to be processed. Then, the noise is processed by the noise reduction configuration parameters to obtain the remaining main point cloud after noise reduction and the outliers that are removed after noise reduction. The denoising configuration parameters include the denoising algorithm to be tested and its corresponding key parameters; Based on at least one of the main point cloud and outlier points, and at least one of the baseline point cloud and simulated noise points, a quantitative evaluation index is calculated to quantitatively evaluate the point cloud denoising algorithm.

2. The method of quantitatively evaluating a point cloud denoising algorithm according to claim 1, wherein, The calculated quantitative evaluation index contains at least two different evaluation parameters.

3. The method of quantitatively evaluating a point cloud denoising algorithm according to claim 1 or 2, characterized in that, The calculated quantitative evaluation metrics are used to provide feedback on the values ​​of the noise configuration parameters and the denoising configuration parameters.

4. The method of quantitatively evaluating a point cloud denoising algorithm according to claim 1 or 2, characterized in that, Also includes: The final quantitative evaluation indicators are presented in tabular form, along with the intermediate process data used to calculate these indicators. The intermediate process data includes the number of reference point cloud points, the number of generated simulated noise points, the total number of noise points removed, and the number of reference point cloud points that were mistakenly deleted.

5. The method for quantizing and evaluating point cloud denoising algorithms according to claim 1 or 2, characterized in that, Also includes: The visualization rendering window allows you to overlay or display the baseline point cloud, simulated noise, the remaining main point cloud after denoising, and the outliers removed after denoising, and distinguish them by different colors.

6. The method for quantizing and evaluating point cloud denoising algorithms according to claim 1 or 2, characterized in that, Methods for generating simulated noise by configuring noise parameters include: The noise generation method in the noise configuration parameters is the radial displacement outlier generation method. When the specific parameters corresponding to the radial displacement outlier generation method are the noise source point and the noise generation radius, a virtual sphere space with the noise generation radius is defined with the coordinates of the noise source point as the center. A candidate noise point is randomly generated within the global bounding box of the reference point cloud and outside the virtual sphere space. Then, the finally generated candidate noise point that meets the conditions is output as a simulated noise point. The noise generation method in the noise configuration parameters is the statistical outlier generation method. Given the specific parameters for this method—the number of neighboring points to be analyzed and the standard deviation factor—a spatial index is constructed for the baseline point cloud. The average nearest neighbor distance of all points in the baseline point cloud is calculated, and the global mean and global standard deviation of these average distances are further calculated. Based on the global mean, global standard deviation, and standard deviation factor, an average distance threshold for a target outlier is defined. A candidate noise point is generated, and its average distance to the nearest neighbor of the analyzed neighboring points in the baseline point cloud is calculated. If this average distance is less than the average distance threshold, the point is discarded and regenerated until it meets the statistical outlier condition that the average distance is not less than the average distance threshold. The candidate noise point that finally meets the condition is then output as a simulated noise point. The noise generation method in the noise configuration parameters is the Gaussian surface noise generation method. Given the specific parameters for the Gaussian surface noise generation method are the noise source point, the number of neighboring points, and the standard deviation of the Gaussian distribution, the method searches for the nearest neighbor point of the set number of neighboring points, centered on the noise source point, to form a neighborhood point set. Principal component analysis is performed on this point set to calculate the local surface normal vector at the noise source point. A value is randomly sampled from a normal distribution with a mean of 0 and a standard deviation equal to the set standard deviation of the Gaussian distribution, and this value is used as the perturbation distance along the normal direction. Based on the perturbation distance, the local surface normal vector, and the noise source point, the coordinates of the generated noise point are calculated, and the noise point is output. The noise generation method in the noise configuration parameters is the simulated multipath reflection cluster generation method. Under the specific parameters of the simulated multipath reflection cluster generation method, namely the noise cluster center point, the number of points in the cluster, and the standard deviation of the cluster distribution, the noise cluster center point is taken as the center in the sparse region of the reference point cloud or the external space. Following a three-dimensional Gaussian distribution with a standard deviation equal to the standard deviation of the cluster distribution, the coordinates of the number of points in the cluster are randomly generated. All generated points are output, which together constitute an independent point cluster as noise.

7. The method for quantizing and evaluating point cloud denoising algorithms according to claim 1 or 2, characterized in that, The methods for denoising the generated point cloud to be processed by configuring denoising parameters include: The denoising algorithm to be tested in the denoising configuration parameters is the radius filtering algorithm. Under the condition that the key parameters of the radius filtering algorithm are the search radius and the minimum number of neighboring points, a spatial index is constructed for the point cloud to be processed to efficiently perform neighborhood search. Each point in the point cloud to be processed is traversed, and a radius search is performed on each point using the spatial index to find all neighboring points whose Euclidean distance to the point is within the search radius. The number of neighboring points found is counted. If the number of neighboring points is less than the minimum number of neighboring points, the point is classified as an outlier. Otherwise, the point is kept in the remaining main point cloud after denoising. Finally, the main point cloud and the outlier are returned. The denoising algorithm to be tested in the denoising configuration parameters is a statistical filtering algorithm. Given that the key parameters for the statistical filtering algorithm are the number of neighboring points analyzed and the standard deviation multiple threshold, a kd-tree is constructed for the point cloud to be processed. For each point in the point cloud, its nearest neighbor (based on the number of neighboring points analyzed) is found, and the average distance from that point to these nearest neighbors is calculated. The global average and global standard deviation of all average distances are then calculated. Based on the global average, global standard deviation, and standard deviation multiple, a global distance threshold is calculated. For each point in the point cloud to be processed, if the average distance from that point to its nearest neighbor is greater than the global distance threshold, that point is classified as an outlier; otherwise, it is retained in the main point cloud. Finally, the main point cloud and the outlier points are returned. The denoising algorithm to be tested in the denoising configuration parameters is a bilateral filtering algorithm. Given that the key parameters for the bilateral filtering algorithm are spatial standard deviation and feature standard deviation, a spatial index is constructed for the point cloud to be processed. Each point in the point cloud is traversed, and a neighborhood is determined for that point. For each point within the neighborhood, a spatial weight is calculated based on the Euclidean distance between that point and the corresponding point in the point cloud, and a Gaussian function of the spatial standard deviation. A local tangent plane is then calculated at the point in the point cloud. The perpendicular distance from the points in the neighborhood to this tangent plane is calculated, and feature weights are calculated based on the Gaussian function of the perpendicular distance and the feature standard deviation. Based on the feature weights and spatial weights, the final weights of the points in the neighborhood are calculated. The new position of a point in the point cloud is the weighted average of all its neighboring points based on the final weights. The resulting main point cloud, composed of all the new positions and smoothed with denoising, is returned.

8. The method for quantizing and evaluating point cloud denoising algorithms according to claim 1 or 2, characterized in that, The methods for obtaining raw point cloud data include: receiving the file corresponding to the raw point cloud data uploaded by the user through a file selection dialog box provided to the user, and obtaining the raw point cloud data; The noise configuration parameters can be obtained through a drop-down menu provided to the user for selecting a noise generation method, at least one input box for inputting specific parameter values ​​corresponding to the noise generation method, and a button or input box for specifying the number of noise points generated. The methods for obtaining denoising configuration parameters include: providing users with a drop-down menu for selecting the denoising algorithm to be tested, and at least one input box for setting the values ​​of key parameters corresponding to the denoising algorithm.

9. The method for quantizing and evaluating point cloud denoising algorithms according to claim 2, characterized in that, The calculated quantitative evaluation metrics include both denoising rate and fidelity.

10. A system for quantifying and evaluating point cloud denoising algorithms, comprising a processor, wherein the processor stores executable program instructions, characterized in that, The executable program instructions are executed to implement the method for quantizing and evaluating point cloud denoising algorithms as described in any one of claims 1-9.

Images