A method for denoising 3D point cloud data in complex scenes

Abstract

This invention discloses a method for denoising 3D point cloud data in complex scenes, belonging to the field of data processing technology. The method includes: screening candidate ground points; performing ground fitting algorithm to fit the ground and segment the ground point cloud; clustering the non-ground point cloud after removing the ground point cloud to obtain point cloud clusters belonging to each obstacle; constructing a 3D bounding box estimation model based on graphical envelope; estimating the 3D bounding box of the clustering results to obtain the minimum 3D bounding box of each obstacle point cloud; identifying and filtering dust, water, and high-inverse-dilation crosstalk noise in the scene to obtain high-precision point cloud data. This invention can effectively detect and filter noise in complex scenes and has good real-time performance and robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a three-dimensional point cloud data denoising method in a complex scene. BACKGROUND

[0002] With the rapid development of laser scanning technology, laser radar has been widely used in unmanned driving, geographic information system, robot perception and other fields. However, when laser radar collects three-dimensional point cloud data, it is affected by environmental noise, equipment error and other factors, and the collected point cloud data often contains a large amount of noise, especially in complex outdoor scenes, road dust and water and high reflectivity object surfaces will seriously affect the accuracy of the point cloud and the reliability of subsequent processing.

[0003] Existing point cloud denoising algorithms mainly rely on the geometric information or density information of point cloud, and the denoising effect is not good for scenes containing complex noise points. Although the existing DBSCAN algorithm can automatically determine the number of clusters and effectively identify outliers, the clustering effect is poor when processing point cloud data with uneven distances. Therefore, how to effectively filter out road dust and water noise points and high reflection blooming crosstalk noise points in a complex scene has become a problem to be solved in the field of point cloud denoising. SUMMARY

[0004] To solve the above technical problems, the present application provides a three-dimensional point cloud data denoising method in a complex scene, which can effectively detect and filter out noise points in a complex scene to obtain high-precision point cloud data.

[0005] A three-dimensional point cloud data denoising method in a complex scene, comprising: Step S1, screening ground candidate points, using a ground fitting algorithm to perform ground fitting and segmenting ground point cloud; Step S2, clustering the non-ground point cloud after removing the ground point cloud to obtain point cloud clusters belonging to each obstacle; Step S3, constructing a three-dimensional bounding box estimation model based on a graph envelope to estimate the three-dimensional bounding box of the clustering result and obtain the minimum three-dimensional bounding box of each obstacle point cloud; Step S4, identifying and filtering out dust and water noise points and high reflection blooming crosstalk noise points in the scene to obtain high-precision point cloud data.

[0006] Optionally, in step S1, screening ground candidate points comprises: Only point clouds within a range of 30 degrees left and right in front of the vehicle are selected as candidate points; The point cloud is sorted according to the height value, the outliers in the height value that do not meet the requirements are removed, and the average height value of the lowest part of the points is calculated as the initial ground height.

[0007] Optionally, in step S1, the step of using a ground fitting algorithm to perform ground fitting and segment the ground point cloud includes: Randomly select several candidate points from the ground candidate points to construct overdetermined equations, solve the plane equations, and solve the ground plane equations in the least squares sense; Calculate the distance from each point in the original point cloud to the plane, and classify points whose distance is less than a certain threshold as inliers; The number of inliers fitted in each iteration is counted. The iteration stops when the number of inliers exceeds a certain threshold or the number of iterations reaches a preset value. The plane equation with the most inliers is used to approximate the ground.

[0008] Optionally, in step S2, the non-ground point cloud after removing the ground point cloud is clustered using the DBSCAN clustering algorithm based on adaptive neighborhood radius and distance; during clustering, density clustering is performed on the point cloud in four dimensions: x, y, z and intensity, and intensity judgment is used as one of the selection conditions for initializing core points.

[0009] Optionally, step S3 includes: The convex hull of each obstacle point cloud cluster is calculated using the Graham scan method; The rotating caliper algorithm is used to solve for the minimum bounding rectangle of the convex hull of each obstacle point cloud cluster, thereby obtaining the length, width and height of the obstacle and realizing the minimum 3D bounding box estimation of each obstacle point cloud.

[0010] Optionally, the criteria for identifying dust and water noise points in the scene include: The Euclidean distance from the geometric center of the cluster to the ground is less than 0.25m; Point clouds with a reflection intensity below 20 account for more than 90%, while point clouds with a reflection intensity above 100 account for less than 1%. The noise point clusters caused by dust and water pollution exhibit significantly greater variance in the x or y direction than in the z direction in the lidar coordinate system. The roughness of the dust and water noise clusters is greater than the threshold.

[0011] Optionally, the formula for calculating the Euclidean distance from the cluster geometric center to the ground is: ; In the formula, d is the Euclidean distance from the geometric center of the cluster to the ground; x, y, and z are the coordinates of the geometric center of the cluster in the lidar coordinate system; and A, B, C, and D are the four parameters of the plane equation.

[0012] Optionally, the algorithm for calculating the roughness of dust and water noise point clusters is as follows: Calculate the Euclidean distance from each laser point in the 3D cluster bounding box to the geometric center of the cluster and store it in a one-dimensional array; Sliding window smoothing filter for a one-dimensional array; The difference between the array after sliding window smoothing and the original array is obtained to get the difference array; Calculate the standard deviation of the difference array, use this standard deviation to characterize the roughness of the clustered point cloud, and calculate the roughness of each cluster.

[0013] Optionally, the algorithm for calculating high inverse dilation crosstalk noise points is as follows: Each frame of 3D point cloud is mapped into two M×N pixel intensity maps and distance maps respectively; A depth-first search region labeling algorithm is used to label the high reflectivity point clouds of the four adjacent channels (up, down, left, and right) in two-dimensional space as one class, and the label is incremented for each connected region starting from 0. After obtaining each highly reflective region with a label, two one-dimensional arrays are used to record the upper and lower boundaries of the region, respectively. After marking the high-reflectivity regions on the 2D image, crosstalk noise points are identified and eliminated in the upper and lower neighborhood spaces of the high-reflectivity regions.

[0014] Optionally, when determining crosstalk noise, the geometric and physical characteristics of the crosstalk noise include: The crosstalk noise intensity in the trailing region is close to zero; The distance between the trailing area and the high-reflection area is close; The number of valid points in each row of the trailing region decreases.

[0015] Compared with the prior art, this application has at least the following beneficial effects: This invention can effectively filter out noise points from road dust and water spray, as well as high-inverse-expansion crosstalk noise points in complex scenarios, effectively improving the noise filtering effect and thus obtaining high-precision point cloud data. Attached Figure Description

[0016] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 A schematic diagram of the DBSCAN algorithm; Figure 3 This is a schematic diagram illustrating the process of obtaining the convex hull using the Graham scan method. Figure 4 This is a schematic diagram of the marked two-dimensional connected region; Figure 5 A schematic diagram marking the upper and lower boundaries of the high-reflection region. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] like Figure 1 As shown, a method for denoising 3D point cloud data in complex scenes includes the following steps: Step S1: Filter ground candidate points, use ground fitting algorithm to perform ground fitting and segment the ground point cloud.

[0019] The process of selecting candidate ground points includes: selecting only point clouds within a 30° radius to the left and right of the vehicle's direct front as candidate points; sorting the point clouds by height values, removing outliers that do not meet the height requirements, and calculating the average height of the lowest-ranking points as the initial ground height. Outliers that do not meet the requirements can be identified based on expert experience; they are generally very few in number and are located at a certain distance from other point clusters, indicating that they are points or point clusters in clearly different locations.

[0020] The step of using a ground fitting algorithm to fit the ground and segment the ground point cloud includes: Step S101: Randomly select several candidate points from the ground candidate points to construct overdetermined equations, solve the plane equations, and solve the ground plane equations in the least squares sense.

[0021] Step S102: Calculate the distance from each point in the original point cloud to the plane, and classify points whose distance is less than a certain threshold as interior points.

[0022] Step S103: Count the number of interior points fitted each time. Stop iterating when the number of interior points exceeds a certain threshold or the number of iterations reaches a preset value. Use the plane equation with the most interior points to approximate the ground.

[0023] Step S2: Cluster the non-ground point clouds after removing the ground point clouds to obtain point cloud clusters belonging to each obstacle.

[0024] The DBSCAN clustering algorithm based on adaptive neighborhood radius and distance is used to cluster the non-ground point cloud after removing the ground point cloud. During clustering, density clustering is performed on the point cloud in four dimensions: x, y, z and intensity. At the same time, intensity judgment is used as one of the selection conditions for initial core points.

[0025] Step S3: Construct a 3D bounding box estimation model based on graph envelope, estimate the 3D bounding box of the clustering results, and obtain the minimum 3D bounding box of each obstacle point cloud.

[0026] Step S301: Calculate the convex hull of each obstacle point cloud cluster using the Graham scan method.

[0027] Step S302: Use the rotating caliper algorithm to solve for the minimum bounding rectangle of the convex hull of each obstacle point cloud cluster, and obtain the length, width and height of the obstacle to achieve the minimum three-dimensional bounding box estimation of each obstacle point cloud.

[0028] Step S4: Identify and filter out dust and water noise and high-inversion-expansion crosstalk noise in the scene to obtain high-precision point cloud data.

[0029] Among them, dust and water noise and high-inflection dilation crosstalk noise are both noise at the cluster level. The identification conditions for dust and water noise in the scene include: The Euclidean distance from the geometric center of the cluster to the ground is less than 0.25m; Point clouds with a reflection intensity below 20 account for more than 90%, while point clouds with a reflection intensity above 100 account for less than 1%. The noise point clusters caused by dust and water pollution exhibit significantly greater variance in the x or y direction than in the z direction in the lidar coordinate system. The roughness of the dust and water noise clusters is greater than the threshold.

[0030] It should be noted that if one or more of the above conditions are met, the point cluster can be considered as dust and water noise points in the scene.

[0031] Furthermore, the formula for calculating the Euclidean distance from the cluster geometric center to the ground is: ; In the formula, d is the Euclidean distance from the geometric center of the cluster to the ground; x, y, and z are the coordinates of the geometric center of the cluster in the lidar coordinate system; and A, B, C, and D are the four parameters of the plane equation.

[0032] The algorithm for calculating the roughness of dust and water noise point clusters is as follows: Calculate the Euclidean distance from each laser point in the 3D cluster bounding box to the geometric center of the cluster and store it in a one-dimensional array; Sliding window smoothing filter for a one-dimensional array; The difference between the array after sliding window smoothing and the original array is obtained to get the difference array; Calculate the standard deviation of the difference array, use this standard deviation to characterize the roughness of the clustered point cloud, and calculate the roughness of each cluster.

[0033] When judging crosstalk noise, the geometric and physical characteristics of crosstalk noise include: The crosstalk noise intensity in the trailing region is close to zero; The distance between the trailing area and the high-reflection area is close; The number of valid points in each row of the trailing region decreases.

[0034] The algorithm for calculating high inverse dilation crosstalk noise points is as follows: Each frame of 3D point cloud is mapped into two M×N pixel intensity maps and distance maps respectively; A depth-first search region labeling algorithm is used to label the high reflectivity point clouds of the four adjacent channels (up, down, left, and right) in two-dimensional space as one class, and the label is incremented for each connected region starting from 0. After obtaining each highly reflective region with a label, two one-dimensional arrays are used to record the upper and lower boundaries of the region, respectively. After marking the high-reflectivity regions on the 2D image, crosstalk noise points are identified and eliminated in the upper and lower neighborhood spaces of the high-reflectivity regions. Example

[0035] Screening ground candidate points is the core step in ground point cloud segmentation. Traditional ground candidate point screening relies on the radar installation height, which is used to obtain the initial ground height, and then ground candidate points are extracted within a certain range above and below this height.

[0036] To improve the accuracy of ground fitting, the following screening process is used for ground candidate points in this embodiment: a) Horizontal angle range limitation: Only point clouds within a 30° range to the left and right of the vehicle's front are selected as candidate points to avoid misjudging point clouds on other horizontal planes as the ground.

[0037] b) Sorting based on Z-value: Sort the point cloud according to Z-value (height), remove outliers whose Z-values ​​do not meet the requirements, and calculate the average Z-value of the lowest part of the points as the initial ground height.

[0038] In addition, when there are too many candidate points on the ground, adaptive uniform downsampling of the point cloud is performed to control the number of candidate points and reduce the algorithm complexity.

[0039] After screening candidate points, an improved ground fitting algorithm based on the Random Sample Consensus (RANSAC) method is used for road segmentation. This algorithm selects candidate ground points through a certain strategy and calculates the four parameters of the plane equation by solving the overdetermined equations using the candidate points. The RANSAC approach ensures the robustness and accuracy of the solution by solving the equations multiple times and calculating scores, selecting the result with the highest score.

[0040] Ground fitting requires solving the plane equations based on three points: ; In the formula, A, B, and C together constitute the normal vector of the plane, D is the core parameter that determines the spatial orientation of the plane; and is a constant term in the equation, which only determines the spatial position (translation distance) of the plane and does not change the orientation of the plane.

[0041] Due to inherent radar noise and the presence of random outliers, selecting noisy points can lead to poor ground fitting results. This invention constructs an overdetermined equation based on 20 candidate ground points and solves the ground plane equation in the least squares sense.

[0042] ; The equation of a plane is of the form: The overdetermined equations of A are solved by the eigenvectors corresponding to the smallest eigenvalues ​​after SVD decomposition of A: ; The solution in the least squares sense is V The last column vector.

[0043] The method for ground fitting under the RANSAC strategy is as follows: randomly sample 20 points from the ground candidate points, solve a plane equation in the manner described above, and calculate the distance from each point in the original point cloud to the plane. Points with a distance less than a certain threshold are classified as inliers. The number of inliers needs to be counted for each fitting. When the number of inliers exceeds a certain threshold or the number of iterations reaches a preset value, the iteration stops. The plane equation with the most inliers is used to approximate the ground.

[0044] After the RANSAC algorithm removes ground point clouds, the remaining non-ground point clouds need to be clustered to obtain point cloud clusters of different objects, providing initial information for detecting water and dust noise and high-inverse dilation crosstalk noise. The DBSCAN algorithm is illustrated below. Figure 2 As shown.

[0045] This invention proposes the following improvements to the DBSCAN algorithm, taking into account the characteristics of lidar point clouds: Since the density of LiDAR point cloud data becomes increasingly sparse with increasing distance, a fixed distance threshold cannot simultaneously adapt to point cloud clustering for both near and far points. Therefore, an adaptive clustering threshold model is implemented by comprehensively considering the Euclidean distance between the LiDAR scan point and obstacles, as well as the LiDAR's scanning resolution.

[0046] In addition, intensity information of the point cloud is introduced to analyze the point cloud in... Density clustering is performed under four dimensions: intensity, etc. At the same time, intensity judgment is added to the selection condition of initial core points, which can accelerate the algorithm and improve the clustering accuracy of water and dust and high reflectivity areas.

[0047] Furthermore, to improve the real-time performance of the algorithm, a tree-like data structure, such as a KD-Tree, can be used to store instance points in K-dimensional space for fast retrieval, thereby reducing the time complexity of the DBSCAN algorithm from... Reduced to .

[0048] The specific process includes: First, input the point cloud data to be processed and start KD-Tree. To ensure the effectiveness of segmentation, the algorithm first selects the dimension Ki with the largest variance in the current dataset. This dimension has the most dispersed data distribution, and using it as the segmentation axis makes the subtree data more balanced. Then, the median Kv is selected on the Ki dimension as the segmentation threshold, dividing the point cloud data into two groups: points with Ki dimension values ​​greater than Kv are assigned to the left subtree feature data, and the rest are assigned to the right subtree feature data. After completing a single segmentation, check whether the subtree contains only a single point or reaches the preset minimum size to confirm whether the segmentation is complete. If not, repeat the steps of "selecting the dimension with the largest variance and taking the median for segmentation" for the left and right subtrees respectively, recursively constructing subtrees; until all subtrees meet the segmentation conditions, the corresponding process is finally completed.

[0049] The above segmentation and clustering process completes the clustering point set of obstacle targets, resulting in a point cloud cluster belonging to each obstacle.

[0050] To more accurately represent the feature information of obstacle targets, it is necessary to construct a 3D bounding box estimation model based on graphical envelopes. This embodiment employs the Graham scan method, which has low time complexity and low computational cost. Figure 3 As shown. The core idea of ​​this algorithm is to use a stack data structure to push points onto the stack counterclockwise, sort them by polar angle, and then pop them from the stack. Finally, connecting the points in the stack sequentially yields the convex hull.

[0051] After calculating the convex hull, it is necessary to further solve for the minimum bounding rectangle of the obstacle point cloud, i.e., the minimum area bounding rectangle (MER) of the convex polygon. This embodiment uses an efficient rotating caliper algorithm to solve for the minimum bounding rectangle of the obstacle point cloud. The core idea is that the minimum bounding rectangle of a convex hull has at least one edge collinear with the convex polygon. The specific solution process is as follows: calculate the minimum bounding matrix of the obstacle point cloud. Once the minimum area bounding rectangle of the obstacle point cloud is determined, the length, width, and height of the obstacle can be obtained, ultimately achieving the 3D bounding box estimation of the obstacle point cloud. The height is the difference between the maximum and minimum height of the point cloud cluster.

[0052] After clustering using 4D DBSCAN and estimating the 3D bounding boxes of the clustering results, two types of clustering results were obtained: candidate classes for dust and water pollution noise points and candidate classes for high reflectivity dilatation crosstalk noise points. For the dust and water pollution noise point candidate classes, the geometric and physical information of the point set within the clustering box needs to be calculated to determine whether it is indeed a dust and water pollution noise point. For the high reflectivity dilatation crosstalk noise point candidate classes, the upper and lower neighborhoods of the high reflectivity point class need to be searched to determine whether it is indeed a high reflectivity dilatation crosstalk noise point.

[0053] For dust and water noise points, the Euclidean distance between the geometric center of the cluster and the ground is relatively small, generally less than 0.25m. The ground plane equation has already been fitted, and the formula for calculating the Euclidean distance from the cluster geometric center to the ground is as follows: ; In the formula, The coordinates of the cluster geometric center point in the lidar coordinate system. These are the four parameters of the plane equation.

[0054] Meanwhile, low-reflectivity point clouds account for a large proportion of the dust and water noise point clusters, with point clouds having a reflectivity of less than 20 generally accounting for more than 90%. Point clouds with a reflectivity of more than 100 account for a small proportion, generally less than 1%.

[0055] Furthermore, since most dust and water noise points are caused by vehicle movement, the clustering results show that the variance in the x or y direction is significantly greater than the variance in the z direction in the lidar coordinate system. The covariance of the point cloud in a cluster is: ; The variance of point clouds in dust and water noise point clusters needs to meet the following requirements: ; In the formula, This is the preset threshold.

[0056] Furthermore, due to the gaps between dust and water noise points, the roughness of the point cloud of dust and water noise points is significantly greater than that of ordinary objects. Roughness is defined as the smoothness of a one-dimensional vector data consisting of the Euclidean distances from each point in a point cloud to its cluster geometric center. The lower the smoothness, the greater the roughness. The calculation algorithm is as follows: Calculate the Euclidean distance from each laser point in the 3D cluster bounding box to the geometric center of the cluster and store it in a 1D array.

[0057] Sliding window smoothing filter for a one-dimensional array: ; In the formula, Indicates the first The filtering results of the data. Indicates the first One data point, Represents the radius of the sliding window, set .

[0058] The difference between the array after smoothing with a sliding window and the original array is calculated to obtain the difference array.

[0059] Calculate the standard deviation of this difference array, and use this standard deviation to characterize the roughness of the clustered point cloud. Then calculate the... Roughness of each cluster: ; In the formula, For the first The roughness of each cluster, For the above difference array, the first... One element, The roughness is the mean of the above difference array, and is greater than a preset threshold. The point cloud clusters were marked as dust and water noise point clusters.

[0060] Crosstalk noise, particularly in high-reflectivity areas, typically occurs between a lidar sensor and a highly reflective object. Since crosstalk noise generally resides in the vertical space above and below the highly reflective object, point cloud-level image processing is the simplest approach for quickly and accurately identifying crosstalk noise regions, considering both data volume and signal processing architecture. This invention proposes a two-dimensional detection method based on point cloud projection to quickly and accurately determine crosstalk noise regions, as detailed below: Arrange the point cloud into an M×N matrix according to the point cloud ID. Each element in the matrix is ​​the intensity and the distance. That is, each frame of 3D point cloud can be mapped to two M×N pixel 2D images - an intensity map and a distance map.

[0061] Considering that projecting 3D high-reflectivity bounding boxes onto a 2D plane may result in some high-reflectivity points being missed or overlapping, a depth-first search region labeling algorithm is used. This algorithm labels adjacent high-reflectivity point clouds in the four channels (up, down, left, and right) of the 2D space as one class, and increments the label for each connected region starting from 0. Figure 4 As shown.

[0062] After obtaining each labeled high-reflectivity region, two one-dimensional arrays are used. and Record the upper and lower boundaries of the regions respectively. The first element of each array records the x-coordinate of the left and right boundaries, respectively, and subsequent elements are the y-coordinates of the upper and lower boundaries of each column of the target region. These also represent the lower boundary of the upper string noise point region and the upper boundary of the lower string noise point region, as shown below. Figure 5 As shown.

[0063] After marking the high-reflectivity regions on the 2D image, it is necessary to determine crosstalk noise points in the upper and lower neighborhood spaces (trailing regions) of the high-reflectivity regions. The geometric and physical characteristics of crosstalk noise points include the following three points: a) The crosstalk noise intensity in the trailing region is close to zero, that is: ; in, The intensity threshold, For the first A collection of point clouds in the upper and lower trailing regions of the high-contrast area.

[0064] b) The distance to the trailing area point is close to that of the high-reflection area, that is: ; in, It is the first The average distance of point clouds in high-reflectivity regions This is the distance difference threshold.

[0065] c) Count the number of valid points in each row of the trailing region. If the number of points decreases, it is a trailing crosstalk noise point.

[0066] The above-mentioned high-inverse-expansion crosstalk noise detection algorithm can be used to detect and filter crosstalk noise in the trailing region.

[0067] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0068] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for denoising 3D point cloud data in complex scenes, characterized in that, include: Step S1: Filter ground candidate points, use ground fitting algorithm to perform ground fitting and segment the ground point cloud; Step S2: Cluster the non-ground point cloud after removing the ground point cloud to obtain the point cloud cluster belonging to each obstacle; Step S3: Construct a 3D bounding box estimation model based on graph envelope, estimate the 3D bounding box of the clustering results, and obtain the minimum 3D bounding box of each obstacle point cloud. Step S4: Identify and filter out dust and water noise and high-inversion-expansion crosstalk noise in the scene to obtain high-precision point cloud data.

2. The method for denoising 3D point cloud data in complex scenes as described in claim 1, characterized in that, In step S1, screening ground candidate points includes: Only point clouds within a 30° radius to the left and right of the vehicle's front are selected as candidate points; The point cloud is sorted according to its height value, outliers that do not meet the height requirements are removed, and the average height value of the lowest part of the points is calculated as the initial ground height.

3. The method for denoising 3D point cloud data in complex scenes as described in claim 2, characterized in that, In step S1, the step of using a ground fitting algorithm to fit the ground and segment the ground point cloud includes: Randomly select several candidate points from the ground candidate points to construct overdetermined equations, solve the plane equations, and solve the ground plane equations in the least squares sense; Calculate the distance from each point in the original point cloud to the plane, and classify points whose distance is less than a certain threshold as inliers; The number of inliers fitted in each iteration is counted. The iteration stops when the number of inliers exceeds a certain threshold or the number of iterations reaches a preset value. The plane equation with the most inliers is used to approximate the ground.

4. The method for denoising 3D point cloud data in complex scenes as described in claim 3, characterized in that, In step S2, the DBSCAN clustering algorithm based on adaptive neighborhood radius and distance is used to cluster the non-ground point cloud after removing the ground point cloud. During clustering, density clustering is performed on the point cloud in four dimensions: x, y, z and intensity. At the same time, intensity judgment is used as one of the selection conditions for initializing core points.

5. The method for denoising 3D point cloud data in complex scenes as described in claim 4, characterized in that, Step S3 includes: The convex hull of each obstacle point cloud cluster is calculated using the Graham scan method; The rotating caliper algorithm is used to solve for the minimum bounding rectangle of the convex hull of each obstacle point cloud cluster, thereby obtaining the length, width and height of the obstacle and realizing the minimum 3D bounding box estimation of each obstacle point cloud.

6. The method for denoising 3D point cloud data in complex scenes as described in claim 5, characterized in that, The criteria for identifying dust and water noise in a scene include: The Euclidean distance from the geometric center of the cluster to the ground is less than 0.25m; Point clouds with a reflection intensity below 20 account for more than 90%, while point clouds with a reflection intensity above 100 account for less than 1%. The noise point clusters caused by dust and water pollution exhibit significantly greater variance in the x or y direction than in the z direction in the lidar coordinate system. The roughness of the dust and water noise clusters is greater than the threshold.

7. The method for denoising 3D point cloud data in complex scenes as described in claim 6, characterized in that, The formula for calculating the Euclidean distance from the geometric center of a cluster to the ground is: ； In the formula, d is the Euclidean distance from the geometric center of the cluster to the ground; x, y, and z are the coordinates of the geometric center of the cluster in the lidar coordinate system; and A, B, C, and D are the four parameters of the plane equation.

8. The method for denoising 3D point cloud data in complex scenes as described in claim 7, characterized in that, The algorithm for calculating the roughness of dust and water noise point clusters is as follows: Calculate the Euclidean distance from each laser point in the 3D cluster bounding box to the geometric center of the cluster and store it in a one-dimensional array; Sliding window smoothing filter for a one-dimensional array; The difference between the array after sliding window smoothing and the original array is obtained to get the difference array; Calculate the standard deviation of the difference array, use this standard deviation to characterize the roughness of the clustered point cloud, and calculate the roughness of each cluster.

9. The method for denoising 3D point cloud data in complex scenes as described in claim 8, characterized in that, The algorithm for calculating high inverse dilation crosstalk noise points is as follows: Each frame of 3D point cloud is mapped into two M×N pixel intensity maps and distance maps respectively; A depth-first search region labeling algorithm is used to label the high reflectivity point clouds of the four adjacent channels (up, down, left, and right) in two-dimensional space as one class, and the label is incremented for each connected region starting from 0. After obtaining each highly reflective region with a label, two one-dimensional arrays are used to record the upper and lower boundaries of the region, respectively. After marking the high-reflectivity regions on the 2D image, crosstalk noise points are identified and eliminated in the upper and lower neighborhood spaces of the high-reflectivity regions.

10. The method for denoising 3D point cloud data in complex scenes as described in claim 9, characterized in that, When judging crosstalk noise, the geometric and physical characteristics of crosstalk noise include: The crosstalk noise intensity in the trailing region is close to zero; The distance between the trailing area and the high-reflection area is close; The number of valid points in each row of the trailing region decreases.

Images