Large-scale point cloud denoising method based on longitudinal interlayer density and spatial connectivity constraint

Abstract

The invention relates to a large-scale point cloud denoising method based on longitudinal interlayer density and spatial connectivity constraint, and the method comprises the steps: carrying out the hierarchical division of large-scale point cloud data, obtaining a point cloud set of different height spaces, and solving a plane projection area; performing discrete noise removal on the point cloud set by using point density and connectivity clustering to obtain a denoised point cloud set; performing plane mapping on the denoised point cloud set to obtain a corresponding relation between each point cloud and a plane projection pixel; solving a density feature operator of each pixel of plane projection based on the corresponding relation, and calculating a point cloud interlayer projection matrix based on the density feature operators; summarizing all the point cloud interlayer projection matrixes and performing point cloud interlayer connectivity check to obtain an updated point cloud interlayer projection matrix; and performing point cloud restoration and splicing based on the updated point cloud interlayer projection matrix, and outputting large-scale point cloud data after noise reduction. According to the invention, the effect of the point cloud denoising technology is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud denoising processing, and in particular to a large-scale point cloud denoising method based on longitudinal inter-layer density and spatial connectivity constraints. Background Art

[0002] Point cloud data, as a set of three-dimensional discrete points describing spatial objects, is obtained by irregularly sampling continuous surfaces and has been widely used in computer vision, remote sensing technology, autonomous driving, robotics and other fields. Although point cloud data can capture fine spatial details, the irregularity and discontinuity of its structure make data processing complicated. In addition, due to the physical limitations of sensors, discontinuities in the boundaries between three-dimensional features, multiple reflections, occlusions and other factors, point cloud data often contains non-surface point noise. These noises seriously affect the downstream tasks of point cloud processing, such as point cloud and color registration, point cloud data segmentation, target detection, etc. Especially in large-scale point cloud data, such as substation scenes, unremoved noise will affect the accuracy of subsequent tasks such as classification or segmentation, resulting in missed or false detection of faults, which in turn affects the normal operation of the substation. Therefore, noise removal of large-scale point cloud data is an important issue that needs to be solved urgently.

[0003] For the denoising task of point cloud data, existing studies have summarized a variety of denoising methods for different surface types and noise models. The exploration of denoising methods initially projected points onto estimated local surfaces, and then bilateral filtering, non-local means, and sparse coding and other image processing methods were widely used in point cloud processing. However, due to the non-ordered structure of point clouds, these image processing-based methods do not have an effective denoising effect. With the emergence of PointNet, point cloud denoising methods based on deep learning have begun to develop. Neural projection, PointCleanNet, and total denoising are pioneers of deep learning point cloud denoising methods, and have achieved good denoising effects. However, these methods are mainly aimed at point cloud denoising at small scenes or component levels, and mainly for Gaussian noise on discrete surfaces.

[0004] In fact, for actual large-scale point cloud data, its complex spatial features and distribution characteristics are difficult to describe with deep learning models. For such point cloud data in large scenes, Charron et al. proposed a dynamic radius filtering algorithm, which dynamically adjusts the search radius to achieve point cloud denoising based on the density change and spatial position relationship of the point cloud. However, this method is sensitive to parameters and it is difficult to balance the filtering effects in different environments. Park et al. constructed a radius filtering algorithm for low-intensity point clouds, but since lidar is easily affected by the environment, this method does not consider the adjustment of parameters in different scenarios. Wu et al. improved the density-based clustering algorithm and improved the accuracy of the occlusion effect in snowy days, but its complex model has high requirements for computing performance. Duan et al. proposed an adaptive radius outlier removal filter based on principal component analysis. Although the paper uses principal component analysis to achieve the effect of dimensionality reduction, it cannot effectively remove high-frequency noise when the noise density is uneven. Shi et al. proposed a point cloud denoising algorithm based on the gravity feature function, which filters the noise in the point cloud through the gravity feature function and threshold, but this method has not been verified in actual application scenarios and has not been compared in detail with other denoising methods, making it difficult to evaluate its effect.

[0005] The above-mentioned deep learning methods mainly focus on reducing the diffuse Gaussian noise of component-level point clouds. For the denoising methods of point clouds obtained in complex detection environments of large scenes, most of them also target the discrete noise dispersed in space. For scenarios with different density noises, no effective method of "removing noise while retaining valid data" has been designed. In actual large-scale point cloud datasets, there are not only discrete noises, but also high-density noises that are not desired to be acquired or not desired for downstream tasks, such as moving pedestrians, flying birds, garbage, small columns and other objects in fixed scenes. These noises have characteristics such as high density and various shapes, and this kind of high-density noise will have a greater impact on subsequent tasks such as point cloud scene segmentation and digital twin. The current denoising methods can perform a certain degree of high-density denoising. In order to better remove high-density noise, limited by the setting of high thresholds or the limitations of feature extraction, a large amount of valid data is easily deleted. Therefore, in order to ensure that valid data is not deleted as much as possible, most people choose to manually delete high-density noise, which greatly reduces efficiency and consumes manpower. Therefore, the present invention proposes a large-scale point cloud denoising method based on longitudinal inter-layer density and spatial connectivity constraints. Summary of the Invention

[0006] The object of the present invention is to provide a large-scale point cloud denoising method based on vertical interlayer density and spatial connectivity constraints. First, the point cloud data is stratified based on the vertical distribution of the point cloud, discrete noise is preliminarily removed based on an improved DBSCAN method, and noise reduction in the plane projection area is completed through two-dimensional projection of the stratified data and interlayer density and connectivity constraints. Finally, the point cloud data is restored to achieve the purpose of point cloud denoising.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A large-scale point cloud denoising method based on vertical interlayer density and spatial connectivity constraints, comprising:

[0009] Performing hierarchical partitioning on the large-scale point cloud data to obtain point cloud sets in different height spaces and solving the plane projection area;

[0010] Removing discrete noise from the point cloud sets by using point density and connectivity clustering to obtain the denoised point cloud sets;

[0011] Performing plane mapping on the denoised point cloud sets to obtain the correspondence between each point cloud and plane projection pixels;

[0012] Solving the density feature operator of each pixel in the plane projection based on the correspondence and calculating the interlayer projection matrix of the point cloud based on the density feature operator;

[0013] Summarizing all the interlayer projection matrices of the point cloud and performing interlayer connectivity check of the point cloud to obtain the updated interlayer projection matrix of the point cloud;

[0014] Restoring and splicing the point cloud based on the updated interlayer projection matrix of the point cloud and outputting the large-scale denoised point cloud data.

[0015] Optionally, performing hierarchical partitioning on the large-scale point cloud data to obtain point cloud sets in different height spaces and solving the plane projection area includes:

[0016] Setting the stratification height according to the size, distribution characteristics, and high-frequency noise conditions of the large-scale point cloud data;

[0017] Dividing the large-scale point cloud data into several point cloud sets based on the stratification height, selecting the maximum and minimum values of the x coordinate and y coordinate in each point cloud set, and calculating the plane projection area of each point cloud set.

[0018] Optionally, performing plane mapping on the denoised point cloud sets to obtain the correspondence between each point cloud and plane projection pixels includes:

[0019] Set the plane projection resolution according to the high-frequency noise situation of the large-scale point cloud data, and perform grid division on the plane projection area based on the plane projection resolution to obtain plane projection pixels.

[0020] Project the denoised point cloud set onto the corresponding interlayer plane, complete the correspondence between pixels and point clouds, and obtain the correspondence between each point cloud and plane projection pixels.

[0021] Optionally, set the plane projection resolution as:

[0022]

[0023] Wherein, is the average length of the high-density noise of the point cloud data in the plane projection, and resolution is the plane projection resolution.

[0024] Optionally, solving the density feature operator of each pixel in the plane projection based on the correspondence, and calculating the interlayer projection matrix of the point cloud based on the density feature operator includes:

[0025] Calculate the local density normalization operator, the connectivity normalization operator in the x-axis direction, the connectivity normalization operator in the y-axis direction, and the connectivity normalization operator in the z-axis direction of each pixel in the plane projection respectively;

[0026] Perform point cloud feature normalization based on the local density normalization operator, the connectivity normalization operator in the x-axis direction, the connectivity normalization operator in the y-axis direction, and the connectivity normalization operator in the z-axis direction;

[0027] Assign the normalized result to the corresponding pixel, and sequentially complete the calculation of the interlayer projection plane in the interlayer order to obtain the interlayer projection matrix of the point cloud.

[0028] Optionally, the local density normalization operator is:

[0029]

[0030] The connectivity normalization operator in the x-axis direction is:

[0031]

[0032] The connectivity normalization operator in the y-axis direction is:

[0033]

[0034] The connectivity normalization operator in the z-axis direction is:

[0035]

[0036] Wherein, Θi is the interlayer local density operator, PointNum i (j,k) is the pixel Square i is the number of point clouds of the pixel Square (j,k), is the connectivity condition in the negative x-axis direction of the pixel Square i (j,k) is the connectivity condition in the positive x-axis direction of the pixel Square i (j,k) is the connectivity condition in the negative y-axis direction of the pixel Square i (j,k) is the connectivity condition in the positive y-axis direction of the pixel Square i (j,k) is the connectivity condition in the negative x-axis direction of the pixel Square i (j,k) is the connectivity condition in the positive x-axis direction of the pixel Square i (j,k).

[0037] Optionally, the normalized result is:

[0038]

[0039] wherein, respectively represent the connectivity normalization operator in the x-axis direction, the connectivity normalization operator in the y-axis direction, and the connectivity normalization operator in the z-axis direction; represents the local density normalization operator.

[0040] Optionally, summarizing all the interlayer projection matrices of the point clouds and performing an interlayer connectivity check of the point clouds to obtain the updated interlayer projection matrix of the point clouds includes:

[0041] Setting the strongly connected and weakly connected within the point cloud region, and obtaining the strongly connected result and the weakly connected result of each pixel;

[0042] Setting a threshold ConnectNum. If the strongly connected result of a pixel is greater than the threshold ConnectNum, then retain the result of the feature operator corresponding to the pixel; if not satisfied, retrieve the overall situation of the strong connection and the weak connection. If the overall situation is greater than the threshold ConnectNum, then retain the result of the feature operator corresponding to the pixel; if neither is satisfied, then consider the pixel as a non-connected noise region.

[0043] The beneficial effects of the present invention are:

[0044] After the point cloud data is stratified, the point cloud data is projected into a two-dimensional space. The projection matrix of a certain layer is calculated according to the density operator between layers and the connectivity check. Then, through the connectivity check between different layers, valid data is retained as much as possible. Finally, the point cloud data is restored and the point clouds between different layers are spliced to achieve noise reduction of the point cloud data. The present invention not only improves the effect of point cloud denoising technology, but also provides a more reliable basis for the further processing and analysis of point cloud data by reducing the misdeletion of key data. It has important practical significance and application value for the subsequent processing and application of point cloud data, especially in precise modeling and fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0046] Figure 1 It is a flowchart of a large-scale point cloud denoising method based on longitudinal interlayer density and spatial connectivity constraints according to an embodiment of the present invention;

[0047] Figure 2 It is a schematic diagram of the unified calculation results of local density and connectivity operators according to an embodiment of the present invention. Among them, (a) is the normalized plane result without using discrete noise removal, and (b) is the normalized plane result with interlayer discrete noise removal;

[0048] Figure 3 It is a schematic diagram of the result after interlayer connectivity check according to an embodiment of the present invention. Among them, (a) is the result without using discrete noise removal, and (b) is the result with using discrete noise removal;

[0049] Figure 4 It is a schematic diagram of the result of interlayer point cloud restoration according to an embodiment of the present invention. Among them, (a) is the original point cloud data and the label situation, and (b) is the result after noise reduction;

[0050] Figure 5 It is a schematic diagram of the point cloud denoising result of long wire accessories according to an embodiment of the present invention. Among them, (a) is the point cloud data of long wire accessories, (b) is the result of DBSCAN denoising of the point cloud data of long wire accessories, and (c) is the result of denoising the point cloud data of long wire accessories using the present method;

[0051] Figure 6Schematic diagram of the denoising result of the connection line attachment point cloud in the embodiment of the present invention. Among them, (a) is the connection line attachment point cloud data, (b) is the denoising result of the connection line attachment point cloud data using DBSCAN, and (c) is the denoising result of the connection line attachment point cloud data using the method of the present invention. Detailed implementation manners

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0053] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0054] Large-scale point cloud data refers to a collection of spatial data containing a large number of three-dimensional coordinate points. These data usually come from devices such as laser scanners (LiDAR), depth cameras (such as Kinect, RGB-D cameras), or stereo vision technologies, and are used to capture the three-dimensional information of objects, scenes, or terrains. With the development of 3D technology, the scale and complexity of point cloud data are increasing continuously. Therefore, the processing of large-scale point cloud data has become an important research topic in multiple fields such as computer vision, geographic information system (GIS), and autonomous driving.

[0055] Characteristics of large-scale point cloud data:

[0056] Huge data volume: Large-scale point cloud data usually contains millions or even billions of data points. Each point contains spatial coordinates (x, y, z), and sometimes contains additional information (such as color, intensity, etc.).

[0057] Sparsity and irregularity: Point cloud data is usually unevenly distributed in space, with holes and noises, which makes it challenging in storage, processing, and analysis.

[0058] Complex structure: Point clouds are usually composed of points in a three-dimensional space, rather than a grid structure. Although it can be converted into other forms through methods such as triangular meshes (such as Delaunay triangulation), the original point cloud data itself does not have a unified topological structure.

[0059] High-dimensional information: In addition to spatial position, point cloud data can also contain many additional information, such as reflection intensity, RGB color, timestamp, normal vector, etc.

[0060] This embodiment provides a large-scale point cloud denoising method based on vertical interlayer density and spatial connectivity constraints. As shown in Table 1, Figure 1 it includes:

[0061] Perform hierarchical partitioning on the large-scale point cloud data, obtain point clouds in different height spaces, and solve the planar projection area;

[0062] Use point density and connectivity clustering to remove discrete noise from the point clouds, and obtain the denoised point clouds;

[0063] Perform planar mapping on the denoised point clouds to obtain the correspondence between each point cloud and the planar projection pixels;

[0064] Solve the density feature operator of each pixel in the planar projection based on the correspondence, and calculate the interlayer projection matrix of the point cloud based on the density feature operator;

[0065] Summarize all the interlayer projection matrices of the point cloud and perform interlayer connectivity check of the point cloud to obtain the updated interlayer projection matrix of the point cloud;

[0066] Restore and splice the point cloud based on the updated interlayer projection matrix of the point cloud, and output the denoised large-scale point cloud data.

[0067] Table 1

[0068]

[0069] Specifically, after stratifying the point cloud data in this embodiment, the point cloud data is projected into a two-dimensional space. The calculation of the projection matrix for a certain layer is realized according to the interlayer density operator and connectivity check. Then, through the connectivity check between different layers, valid data is retained as much as possible. Finally, the point cloud data is restored and the point clouds between different layers are spliced to realize the denoising of the point cloud data. This embodiment not only improves the effect of the point cloud denoising technology, but also provides a more reliable basis for the further processing and analysis of the point cloud data by reducing the misdeletion of key data. It has important practical significance and application value for the subsequent processing and application of the point cloud data, especially in precise modeling and fault diagnosis.

[0070] Furthermore, performing hierarchical partitioning on the large-scale point cloud data, obtaining point clouds in different height spaces, and solving the planar projection area includes:

[0071] Set the stratification height according to the size, distribution characteristics, and high-frequency noise conditions of the large-scale point cloud data;

[0072] Based on the stratification height, divide the large-scale point cloud data into several point clouds, select the maximum and minimum values of the x and y coordinates in each point cloud, and calculate the planar projection area of each point cloud.

[0073] Specifically, it includes the following content:

[0074] First, set the large-scale point cloud dataset as Points = {p1,..., p N}, where N represents that there are N points in the point cloud dataset, and each point has coordinate values p = (x, y, z) in the actual space. First, set the hierarchical height dis z according to the size, distribution characteristics, and high-frequency noise situation of the point cloud data. Among them, the height of high-frequency noise needs to be considered emphatically, and the divided interval should be greater than the height of high-frequency noise as much as possible.

[0075] Statistically calculate the size of the projection plane of the point cloud data, that is, statistically calculate the x coordinates and y coordinates of all points, and select the maximum and minimum values {x max , x min , y max , y min}, then the plane projection area is:

[0076] Square = (x max - x min ) × (y max - y min ) (1);

[0077] Layer the point cloud data according to the hierarchical height dis z . Assume that the highest height of the point cloud data is z max , then the total number of layers of the point cloud data is Num z = z max / / dis z + 1, where / / is considered as the floor division operation, then this large-scale point cloud data can be divided into multiple sub-datasets.

[0078]

[0079] In formula (2), P i represents the point cloud data of the i-th layer after being segmented.

[0080] Furthermore, discrete noise removal is performed on the point cloud set by using point density and connectivity clustering to obtain a denoised point cloud set, which specifically includes the following content:

[0081] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering is widely used in the industrial field due to its advantages of noise resistance and non-preset clustering. The DBSCAN algorithm uses the number of points contained in the retrieval neighborhood eps and the point set (the distances between them are less than each other) as the criteria for judging the attributes of points. Through this rule, the minimum number of points in the domain is defined as MinPts. Each point is distinguished as belonging to a high-density area or a low-density area based on the density around it, and low-density areas are usually marked as outliers.

[0082] In this embodiment, DBSCAN is selected as the pre-basic algorithm, which can largely suppress the influence of discrete noise on the entire denoising process. In fact, this embodiment does not require data clustering, but uses the idea of retrieving the local density and connectivity of points in DBSCAN to denoise discrete points, that is, if the number of neighborhood points of a certain point with a size of eps is less than MinPts, it is considered a discrete noise point. Therefore, the DBSCAN method is modified and simplified. For each layer of point cloud data after segmentation, discrete noise point clouds are removed by retrieving the density and connectivity of the point cloud data.

[0083] After removing discrete noise points, the update of each layer of point cloud data is completed. At this time, each layer of point cloud data is P i '. Among them, the number of point clouds in P i ' should be less than or equal to P i , and it should be noted that this process will not change the position of the existing point cloud data or add additional data.

[0084] Furthermore, performing a plane mapping on the denoised point cloud set to obtain the correspondence between each point cloud and the plane projection pixels includes:

[0085] Setting the plane projection resolution according to the high-frequency noise situation of the large-scale point cloud data, performing grid division on the plane projection area based on the plane projection resolution to obtain plane projection pixels;

[0086] Projecting the denoised point cloud set onto the corresponding interlayer plane, completing the correspondence between pixels and point clouds, and obtaining the correspondence between each point cloud and the plane projection pixels.

[0087] Specifically, it includes the following content:

[0088] Set the resolution r of the projection plane according to the high-density noise situation of the large-scale point cloud data. Divide the projection area Square S into two directions and perform grid division according to the plane projection rate, and finally divide it into m rows and n columns of grids. Among them, the plane resolution is a key parameter, and its size setting directly affects the subsequent noise reduction effect. When setting, consider the plane projection area of high-frequency noise. Although setting a smaller value can retain as much effective data as possible, it will affect the noise reduction effect. Let the average value of the length of the high-density noise of the point cloud data in the plane projection be Usually set the plane resolution to satisfy:

[0089]

[0090] Project the interlayer point cloud data P i ' onto the corresponding interlayer plane, where the interlayer plane is Square i , i ∈ [1, Num z . Then for each pixel of the interlayer plane, such as Square i (j, k), at this time, it contains all the point cloud data that meet the following conditions:

[0091] The height meets the interlayer height, that is, z ∈ (z min ×i, z min ×i + dis z ) of the point cloud;

[0092] The horizontal and vertical coordinates meet the divided pixel intervals, that is, the horizontal and vertical coordinates meet the point cloud of {(x, y)|x ∈ (j, j + resolution), y ∈ (k, k + resolution)};

[0093] Through this operation, the corresponding relationship between each pixel of the plane projection matrix and the point cloud is completed at this time.

[0094] Furthermore, based on the corresponding relationship, solve the density feature operator of each pixel of the plane projection, and calculate the interlayer projection matrix of the point cloud based on the density feature operator, including:

[0095] Calculate the local density normalization operator, the connectivity normalization operator in the x-axis direction, the connectivity normalization operator in the y-axis direction, and the connectivity normalization operator in the z-axis direction of each pixel of the plane projection respectively;

[0096] Perform point cloud feature normalization based on the local density normalization operator, the connectivity normalization operator in the x-axis direction, the connectivity normalization operator in the y-axis direction, and the connectivity normalization operator in the z-axis direction;

[0097] Assign the normalized result to the corresponding pixel, and calculate the inter-layer projection plane in sequence according to the inter-layer order to obtain the inter-layer projection matrix of the point cloud.

[0098] Specifically, it includes the following content:

[0099] The above content completes the correspondence between the three-dimensional space and the plane. However, only doing this cannot perform subsequent noise reduction processing. Therefore, relevant algorithms are designed for data projection, and the spatial features of a certain projection plane pixel can also be extracted and mapped to the corresponding pixel. Next, for single-layer data, a feature extraction method based on the local density between layers of the point cloud and the connectivity constraint of the corresponding space of the plane pixel is designed. As shown in Table 2, the feature extraction undergoes normalization calculation to complete the numerical mapping of 0-1 and is projected onto the corresponding pixel. This result is described by (Density and ConnectivityOperator), where i represents the corresponding layer number, and (j,k) represents the corresponding pixel coordinates. Each pixel 's size describes its proximity to the valid data area. The closer its value is to 1, the more likely the point cloud data in this block is valid data, and the closer it is to 0, the more likely it is a noise area. Specifically, it includes:

[0100] (1) Local density normalization operator:

[0101] For each hierarchical interval, first calculate the local density Θ i of the point cloud between layers. Count the number of point clouds PointNum i ' of the point cloud data P i of each layer, and obtain the inter-layer local density operator according to the number of point clouds and the area of the projection plane:

[0102] Θ i = PointNum i / Square, i = 1,..., Num z (4);

[0103] This local density Θ i describes the point cloud density situation of a certain layer of point cloud interval. Use PointNum i (j,k) to describe the number of point clouds in the pixel block Square i (j,k) of this layer. If PointNum i (j,k) is less than this situation, it is considered that the local density of this layer is not satisfied. If it is greater than this local density value, it is considered that the pixel density meets the requirements. Next, a normalization operator is used to describe the density situation of the pixel. Here, the sigmoid function is used for normalization. The formula of the sigmoid function is as follows:

[0104]

[0105] Based on formula (4), we can get:

[0106]

[0107] Through the above calculation, the normalized point cloud density of each pixel in each level interval can be obtained The reason for using subtraction and ratio is to reduce the impact of gradient abrupt changes on the normalized result when the difference between the point cloud density and the local density between layers is not too large, and this method can also better describe the relationship between the current pixel point cloud density and the overall density of the point cloud layer. Where i represents the i-th level interval, j and k represent the pixel coordinates of the corresponding projection plane. The normalized result is used to describe the point cloud density of the pixel. If it is greater than 0.5, it means that it is greater than the average point cloud density. If it is less than 0.5, it means that it is less than the average point cloud density.

[0108] (2) Connectivity normalization operator:

[0109] The density of point clouds in different pixel intervals can be obtained through the above density operator, which is convenient for subsequent processing to better filter out point cloud clusters with lower density. However, in fact, this method cannot solve the problem of incorrectly removing low-density valid data, such as wires, corner lines, etc., so it is necessary to consider extracting their corresponding features and retaining such low-density point cloud valid data as much as possible.

[0110] Taking the point cloud data in the substation as an example, although the point cloud data density of wires is relatively small, they all have strong connectivity, that is, a certain plane in space has a long data continuity. i (j,k), obtain all point cloud information in the area, and obtain the three-axis connectivity normalization operators on the x, y, and z axes respectively. The x axis is used as an example below.

[0111] For a certain projection plane pixel block Square i (j,k), the actual x-axis data range is x∈(j,j+resolution), and the x-axis information of all point clouds is counted to obtain the minimum x-axis value in the point cloud of the pixel area. and the maximum value of the x-axis At the same time, set the minimum and maximum threshold interval ranges respectively. For example, you can set the minimum threshold interval to MinLevel low =0.05,MaxLevel low =0.35, the threshold interval of the maximum value is MinLevel high =0.65,MaxLevelhigh If =0.95, the upper and lower bounds of the connectivity of the pixel block in the x-axis direction can be expressed as follows:

[0112]

[0113]

[0114] The result of formula (6) is the connectivity of the pixel block in the negative x-axis direction, and the result of formula (7) is the connectivity of the pixel block in the positive x-axis direction.

[0115] For the connectivity normalization operator of the x-axis of the pixel block, it can be obtained by simple weighting:

[0116]

[0117] Similarly, according to the above x-axis example, the connectivity normalization operators of the y-axis and z-axis can be calculated and

[0118]

[0119]

[0120] where is the connectivity of the negative y-axis of pixel Square i (j,k), is the connectivity of the positive y-axis of pixel Square i (j,k), is the connectivity of the negative x-axis of pixel Square i (j,k), is the connectivity of the positive x-axis of pixel Square i (j,k).

[0121] (3) Feature extraction and normalization:

[0122] From the description of the above two operators, it can be seen that each pixel block of the point cloud projection at each level can obtain 4 operator calculation results:

[0123] One of them is i.e., the local density normalization operator, which is used to describe the local density level of the point cloud in the corresponding area;

[0124] The other three are and i.e., the connectivity normalization operators in three coordinate directions, which are used to describe the connectivity level of the point cloud in the corresponding area.

[0125] Finally, considering the situations of these four operators comprehensively: the regional density of the point cloud can well characterize that the region "is not" a valid region, that is, describe it as noise. However, due to the existence of high-density data in space, if only this method is used, it is easy to confuse high-density noise with valid data.

[0126] The connectivity checks in three axial directions can well characterize that the region "is" a valid region. In actual calculations, the axial data with the best connectivity can be selected to describe the "validity" of the region.

[0127] Therefore, summarize the results of the above four operators to form a better feature result and normalize it to describe the confidence level that a certain pixel block "is valid data". Finally, the result of each pixel region is represented using (Densityand Connectivity Operator):

[0128]

[0129] Assign the normalized result to the corresponding pixel, and complete the calculation of the interlayer projection plane in sequence according to the interlayer order.

[0130] Table 2

[0131]

[0132] Furthermore, summarize all the interlayer projection matrices of the point cloud layers and perform interlayer connectivity checks on the point cloud to obtain the updated interlayer projection matrix of the point cloud, including:

[0133] Set the strong connectivity and weak connectivity within the point cloud region, and obtain the strong connectivity result and weak connectivity result of each pixel;

[0134] Set the threshold ConnectNum. If the strong connectivity result of a pixel is greater than the threshold ConnectNum, then retain the result of the feature operator corresponding to the pixel; if not satisfied, retrieve the overall situation of strong connectivity and weak connectivity. If the overall situation is greater than the threshold ConnectNum, then retain the result of the feature operator corresponding to the pixel; if neither is satisfied, then consider this pixel as a non-connected noise region.

[0135] Specifically, it includes the following content:

[0136] Through interlayer segmentation, plane projection, and density and connectivity constraints of the pixel corresponding point cloud data, the noise reduction of the point cloud plane region is completed. Subsequently, for downstream tasks, data reduction needs to be performed according to the noise-reduced plane result. Specifically, given a threshold gate, by comparing the plane pixel operator and the magnitude of the threshold value to determine whether to retain the point cloud data in the corresponding projection area.

[0137] However, it should be noted that since the point cloud data is mapped to a two-dimensional plane for related processing and threshold analysis, it is very easy to have the problem of loss of connectivity between different pixel blocks. In the calculation process of the pixel feature description result in the above plane projection area, this embodiment considers the spatial connectivity of the point cloud inside a certain pixel. During the current restoration process, the connectivity of the corresponding three-dimensional space between different pixel blocks also needs to be considered.

[0138] Taking high connectivity as an example for analysis, for instance, a certain point cloud data is divided into upper and lower parts after layer height division. Since the division is based on equal intervals, it is possible that the segmentation result of the point cloud object is not average, and there may be a situation where there are more point clouds in the lower part and fewer point clouds in the upper part.

[0139] Then, after the operator calculation, for the same pixel block in the two plane projections, it is possible that: the corresponding result Square i (j,k)≥gate in the lower layer, and the result Square i+1 (j,k)<gate at the same pixel position in the upper layer. Then, according to the requirements, the point cloud in the lower layer will be restored, while the point cloud in the upper layer will be removed, which obviously does not meet the requirements of point cloud restoration.

[0140] The reason for such a problem is that only the situation of a certain pixel area is considered for restoration, without considering the spatial connectivity of the overall point cloud data. Therefore, when performing restoration, a neighborhood connectivity retrieval scheme is designed: traverse the upper and lower layers and the 3*3 neighborhood situation, that is, examine the 26-neighborhood of a certain point in space to determine whether to perform regional point cloud restoration.

[0141] Regarding the form of the object's existence, that is, most creations in three-dimensional space are perpendicular to the horizontal plane. Especially for substation equipment, therefore, the region itself, 2 vertically connected neighborhoods, and 8 same-layer neighborhood connections are defined as strong connectivity (Strong Connectivity), and the remaining 16 connections are defined as weak connectivity (Weak Connectivity). Different connectivities are given different weights. The weight of strong connectivity is 1, and the weight of weak connectivity is 0.5. According to the connectivity situation, a threshold ConnectNum is set as the condition for checking whether the regional connectivity situation meets the requirements.

[0142] If the strongly connected result of a certain pixel is greater than the threshold, the result of the feature operator is retained; if not, the overall situation of strong and weak connectivity is retrieved. If it is greater than the threshold, the result of the feature operator is retained; if neither is satisfied, the block is considered a non-connected "noise area". During the process, the boundary conditions of the data in the first and last layers need to be considered, and the connectivity threshold should be appropriately reduced for these two layers. The specific connectivity check algorithms for different layers are shown in Table 3.

[0143] Table 3

[0144]

[0145]

[0146] Furthermore, based on the updated inter-layer projection matrix of the point cloud, point cloud restoration and stitching are performed to output the denoised large-scale point cloud data, which specifically includes the following content:

[0147] After completing the connectivity check, point cloud restoration and stitching are carried out. This part requires traversing the entire point cloud data, corresponding the point cloud to the area to be restored, and restoring the entire space of the point cloud.

[0148] During the restoration process, the point cloud can be retrieved through the pixel area of each layer; or each point can be mapped to the corresponding area. The difference is that the first method is to find the corresponding plane area and retrieve the point cloud, that is, retrieve unordered data for an ordered sequence; the second method is to retrieve for each point and for the situation of the pixel area, that is, perform an ordered retrieval for unordered data. Generally speaking, the first method will consume more time. By pre-encoding the point cloud data with an octree, according to the locality principle, the point cloud data in the same area is more likely to be concentrated in the same space of the octree, and the retrieval is faster, so the retrieval time is not disadvantaged. At the same time, due to the pre-conducted connectivity check, the logic of point cloud restoration can be integrated into the connectivity check to reduce the number of traversals. The specific point cloud restoration and stitching algorithms are shown in Table 4, and a threshold threshold is set. Generally, threshold = gate can be taken.

[0149] Table 4

[0150]

[0151] The following uses the measured large-scale point cloud data of a substation in the southeast of China to specifically illustrate the use and effect verification of the method proposed in this embodiment:

[0152] Since the complete point cloud data volume of the substation is too large, part of the data is selected for algorithm verification. A total of 37,611,708 points are intercepted from the partial point cloud, and the minimum distance between points is 5 mm.

[0153] To facilitate the verification of the effectiveness of this method, label sticking was additionally performed on the point cloud data. It should be noted that during the actual operation of the noise reduction algorithm model, no supervised training or noise label division is required. The purpose of label sticking is to facilitate the subsequent evaluation of the point cloud denoising results.

[0154] High-density noise includes the point cloud data generated by inspectors during the acquisition process, as well as some point cloud data that do not want to be registered or used for downstream tasks, such as columns, incomplete acquisition point cloud data, moving objects, etc.

[0155] Noise reduction processing:

[0156] (1) Point cloud layering and projection matrix calculation:

[0157] For this point cloud data, parameter settings are first performed. Set the layer height dis z = 3m, the plane resolution is resolution = 0.4m, and plane projection is completed; then set the relevant parameters for discrete noise removal, refer to the parameters of DBSCAN clustering, set the search neighborhood distance to eps = 0.06m, and set the minimum number of search points to MinPts = 5; next, perform local density and connectivity analysis, analyze and extract the plane data, and set the minimum critical range to MinLevel low = 0.05, MaxxLevel low = 0.35, and set the maximum critical range to MinLevel high = 0.65, MaxLevel high = 0.95, that is, complete the calculation of the normalized results of each interlayer plane pixel points, and assign the normalized results to the corresponding pixels. Taking the point cloud mapping plane of the first layer as an example, the result after assignment Figure 2 is shown.

[0158] Among them, Figure 2 in (a) is the normalized plane result without using discrete noise removal, Figure 2 in (b) is the normalized plane result with interlayer discrete noise removal added. It can be seen that there are many discrete noise points in the normalized image without using discrete noise removal in (a), distributed at various positions in space, especially the noise points near the effective data. More complex removal methods need to be designed in subsequent processing. Therefore, discrete noise removal can first filter out the discrete noise distributed in space, facilitating subsequent processing.

[0159] (2) Point cloud interlayer connectivity check:

[0160] Check the connectivity between layers. Set the operator threshold gate of each pixel block to 0.7, set the number of connected pixels to ConnectNum = 3, and calculate according to the algorithm flow shown in Table 3. Take the first layer data as an example to check its connectivity with the second layer.

[0161] from Figure 3 It can be seen that the plane layout of the point cloud mapping of the plane has been further simplified, where Figure 3 (a) is the result without discrete noise removal. Figure 3 (b) shows the result of discrete noise removal. It can be seen that more discrete point cloud blocks are removed after the connectivity check in (b), further optimizing the effects of high-density noise and discrete noise removal.

[0162] (3) Point cloud restoration and comparison:

[0163] Next, the point cloud is reconstructed and stitched. Generally, the threshold is set to threshold = gate, that is, the threshold is shared with the connectivity check. In the actual process, the reconstruction and stitching of the point cloud are performed together with the connectivity check. This does not require repeated traversal of the same plane mapping matrix, reducing time overhead, and the corresponding matrix can be deleted after reconstruction to reduce memory overhead. Next, the result of processing the first layer of point cloud data is also used as an example. Figure 4 shown.

[0164] Figure 4 (a) shows the original point cloud data and labels. Figure 4 (b) is the result after denoising. Comparing the point cloud data of the two, it can be seen that most of the discrete noise and high-density noise have been effectively removed, and some low-density data, such as connecting lines, have been well preserved.

[0165] Comparison of different noise reduction methods:

[0166] like Figure 5 and Figure 6 As shown, this is the point cloud data of a part of the wire. Figure 5 Point cloud data of long wire attachments. Figure 5 (a) is the point cloud data of the long wire attachment. Figure 5 (b) shows the denoising result of the long wire attachment point cloud data using DBSCAN. Figure 5 (c) is the denoising result of the point cloud data of the long wire attachment using this method; Figure 6 For connecting line attachment point cloud data, Figure 6 (a) is the point cloud data of the connecting line attachment. Figure 6 (b) shows the denoising result of the connection line attachment point cloud data using DBSCAN. Figure 6In Figure (c), the denoising result of the wire connection attachment point cloud data using this method is shown. It can be seen that compared with the original image, since DBSCAN clustering is a density-based clustering method, it misinterprets the wire data as noise data and deletes it, and only a small part of the valid wire data is retained. The denoising result of this method almost retains all the wire data. There is a small number of discontinuous point cloud phenomena because this method uses relatively small retrieval parameters to remove discrete noise, and there is still a situation where overly sparse point cloud data is misdeleted. However, this method better retains valid data on the basis of removing high-density noise.

[0167] After that, different denoising algorithms are applied to the point cloud data, and statistical filtering, neighborhood filtering, DBSCAN clustering, and LDCCPointCloudDenoising are used for comparison respectively. Among them, a DBSCAN model with the same parameters as this model is added for comparison, and at the same time, a method that does not use discrete noise removal in this method is added for comparison. The main models for comparison are statistical filtering, neighborhood filtering, DBSCAN, DBSCAN-N (DBSCAN uses the parameters of LDCC Point Cloud Denoising), LDCC-PCD-U (LDCC Point Cloud Denoising without using DBSCAN), and LDCC-PCD (LDCCPoint Cloud Denoising). The main parameters for comparison are: the proportion of misdeleted points; the percentage of discrete noise removal; the percentage of high-density noise removal; the overall noise removal percentage. The comparison results are shown in Table 5.

[0168] Table 5

[0169]

[0170] According to the above results, it can be seen that the denoising effect of this method has been effectively verified. For large-scale point cloud data, when denoising, it is necessary to retain valid data as much as possible. Although the proportion of discrete noise removal by statistical filtering and neighborhood filtering is as high as more than 85%, and the proportion of misdeleted points is less than 2%, for large-scale data sets, a large amount of valid data is misdeleted. Especially for substation point cloud data, a large amount of data with low density such as wires is deleted, and these components are high-failure components and provide important references for future fault diagnosis. Therefore, the performance of the above traditional models in retaining data is not good.

[0171] Observing and comparing DBSCAN and this method, it can be seen that DBSCAN does have good performance in removing high-density noise, but as Figure 5 and Figure 6As shown, the misdeletion rate of 0.53% still deletes a relatively large amount of valid data. While this method has shown good performance in high-density noise removal and the proportion of misdeleted points even without using the discrete noise removal model, it has poor performance in discrete noise removal. This is mainly because this model takes pixel blocks as units and does not analyze and remove the features of each point. Therefore, an improved DBSCAN denoising method is integrated into our model to optimize the problem of poor discrete noise removal effect of this method, and finally a good denoising effect is obtained.

[0172] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A large-scale point cloud denoising method based on vertical inter-layer density and spatial connectivity constraints, characterized in that Including: Performing hierarchical partitioning on the large-scale point cloud data to obtain point cloud sets in different height spaces and solving the planar projection area; Removing discrete noise from the point cloud sets by using point density and connectivity clustering to obtain the denoised point cloud sets; Performing planar mapping on the denoised point cloud sets to obtain the correspondence between each point cloud and planar projection pixels; Solving the density feature operator of each pixel in the planar projection based on the correspondence, and calculating the inter-layer projection matrix of the point cloud based on the density feature operator; Summarizing all the inter-layer projection matrices of the point cloud and performing inter-layer connectivity check of the point cloud to obtain the updated inter-layer projection matrix of the point cloud; Performing point cloud restoration and stitching based on the updated inter-layer projection matrix of the point cloud, and outputting the denoised large-scale point cloud data.

2. The large-scale point cloud denoising method based on vertical interlayer density and spatial connectivity constraints according to claim 1, wherein Performing hierarchical partitioning on the large-scale point cloud data to obtain point cloud sets in different height spaces and solving the planar projection area includes: Setting the layer height according to the size, distribution characteristics and high-frequency noise situation of the large-scale point cloud data; Dividing the large-scale point cloud data into several point cloud sets based on the layer height, and selecting the maximum and minimum values of the x coordinate and y coordinate in each point cloud set to calculate the planar projection area of each point cloud set.

3. The large-scale point cloud denoising method based on longitudinal interlayer density and spatial connectivity constraints according to claim 1, wherein Performing planar mapping on the denoised point cloud sets to obtain the correspondence between each point cloud and planar projection pixels includes: Setting the planar projection resolution according to the high-frequency noise situation of the large-scale point cloud data, performing grid partitioning on the planar projection area based on the planar projection resolution to obtain planar projection pixels; Projecting the denoised point cloud sets onto the corresponding inter-layer planes, completing the correspondence between pixels and point clouds, and obtaining the correspondence between each point cloud and planar projection pixels.

4. The large-scale point cloud denoising method based on longitudinal interlayer density and spatial connectivity constraints according to claim 3, characterized in that Setting the planar projection resolution as: Among them, is the average length of the high-density noise of the point cloud data in the planar projection, and resolution is the planar projection resolution.

5. The large-scale point cloud denoising method based on longitudinal interlayer density and spatial connectivity constraints according to claim 1, characterized in that, Solving the density feature operator of each pixel in the planar projection based on the correspondence, and calculating the inter-layer projection matrix of the point cloud based on the density feature operator includes: Calculating the local density normalization operator, connectivity normalization operator in the x-axis direction, connectivity normalization operator in the y-axis direction, and connectivity normalization operator in the z-axis direction of each pixel in the planar projection respectively; Performing point cloud feature normalization based on the local density normalization operator, connectivity normalization operator in the x-axis direction, connectivity normalization operator in the y-axis direction, and connectivity normalization operator in the z-axis direction; Assigning the normalized result to the corresponding pixel, and sequentially completing the calculation of the inter-layer projection plane in the inter-layer order to obtain the inter-layer projection matrix of the point cloud.

6. The large-scale point cloud denoising method based on longitudinal inter-layer density and spatial connectivity constraints according to claim 5, characterized in that, The local density normalization operator is: The connectivity normalization operator in the x-axis direction is: The connectivity normalization operator in the y-axis direction is: The connectivity normalization operator in the z-axis direction is: Among them, Θ i is the interlayer local density operator, and PointNum i (j,k) is the number of point clouds of the pixel Square i (j,k). is the connectivity in the negative x-axis direction of the pixel Square i (j,k). is the connectivity in the positive x-axis direction of the pixel Square i (j,k). is the connectivity in the negative y-axis direction of the pixel Square i (j,k). is the connectivity in the positive y-axis direction of the pixel Square i (j,k). is the connectivity in the negative x-axis direction of the pixel Square i (j,k). is the connectivity in the positive x-axis direction of the pixel Square i (j,k).

7. The large-scale point cloud denoising method based on vertical interlayer density and spatial connectivity constraints according to claim 6, wherein The normalized result is: Among them, respectively represent the connectivity normalization operator in the x-axis direction, the connectivity normalization operator in the y-axis direction, and the connectivity normalization operator in the z-axis direction; represents the local density normalization operator.

8. The large-scale point cloud denoising method based on vertical interlayer density and spatial connectivity constraints according to claim 1, wherein Summarizing all the inter-layer projection matrices of the point cloud and performing inter-layer connectivity check of the point cloud to obtain the updated inter-layer projection matrix of the point cloud includes: Setting strong connectivity and weak connectivity within the point cloud area, and obtaining the strong connectivity result and weak connectivity result of each pixel; Set a connectivity threshold. If the strongly connected result of a pixel is greater than the connectivity threshold, retain the result of the feature operator corresponding to the pixel; if not satisfied, retrieve the overall situation of strong connectivity and weak connectivity. If the overall situation is greater than the connectivity threshold, retain the result of the feature operator corresponding to the pixel; if neither is satisfied, consider the pixel as a non-connected noise region.

Images