Hybrid filtering denoising method and device based on adaptive parameters

Abstract

The invention provides a hybrid filtering denoising method and device based on adaptive parameters, and relates to the technical field of three-dimensional point cloud processing. The method comprises the following steps: acquiring three-dimensional point cloud data of a target object; performing voxel segmentation processing on the three-dimensional point cloud data to obtain a plurality of voxels; according to the segmented three-dimensional point cloud data, combining a plurality of voxels by using a density similarity model to obtain a voxel set with different density distributions; based on the merged voxels, a radius filtering parameter and a statistical filtering parameter of each voxel set are adaptively determined, the radius filtering parameter is used for filtering isolated noise in the three-dimensional point cloud data, and the statistical filtering parameter is used for filtering random noise in the three-dimensional point cloud data; performing random noise removal processing on the three-dimensional point cloud data according to the statistical filtering parameters; and performing isolated noise removal processing on the three-dimensional point cloud data after random noise removal according to the radius filtering parameter.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional point cloud processing, and more particularly, to a hybrid filtering denoising method and device based on adaptive parameters. BACKGROUND

[0002] A point cloud is a collection of three-dimensional samples of an object surface, each three-dimensional point containing position information, normal and tangent vectors, point cloud density, and other features. By analyzing the local or global features of the point cloud, the shape and semantics of the target can be obtained, and data analysis can be performed. For example, using the obtained three-dimensional point cloud for online detection of intelligent manufacturing products can improve product quality.

[0003] The surface of a complex structure workpiece has various materials and local occlusions. Due to the different optical absorption and reflection characteristics of multi-material surfaces, and the large curvature changes in the edge regions of complex workpieces, combined with the non-uniformity of light refraction and reflection, it will lead to uneven distribution of three-dimensional point cloud density obtained based on optical detection technology, resulting in various types of noise in the three-dimensional point cloud of the measured target surface.

[0004] However, the existing denoising methods for complex components have the problems of slow denoising speed and easy loss of detail information, which makes it difficult to be applied to online detection requirements of intelligent manufacturing. SUMMARY

[0005] Therefore, the present application provides a hybrid filtering denoising method and device based on adaptive parameters, aiming to solve the technical problem that the existing denoising method for complex components is difficult to be applied to online detection requirements of intelligent manufacturing.

[0006] One aspect of the present application provides a hybrid filtering denoising method based on adaptive parameters, comprising: obtaining three-dimensional point cloud data of a target object; performing voxel segmentation processing on the three-dimensional point cloud data to obtain a plurality of voxels, wherein the point clouds in each voxel are uniformly distributed; according to the segmented three-dimensional point cloud data, using a density similarity model to perform merging processing on the plurality of voxels to obtain a voxel set with different density distributions, wherein in response to the difference between the densities of the current voxel and the adjacent voxel being less than a preset density similarity threshold, the current voxel and the adjacent voxel are merged to obtain the corresponding voxel set; based on the merged voxels, respectively adaptively determining the radius filtering parameter and the statistical filtering parameter of each voxel set, wherein the radius filtering parameter is used to determine the isolated noise in the three-dimensional point cloud data, and the statistical filtering parameter is used to determine the random noise in the three-dimensional point cloud data; according to the statistical filtering parameter, performing random noise removal processing on the three-dimensional point cloud data; according to the radius filtering parameter, performing isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing.

[0007] According to an embodiment of the present application, the voxel segmentation processing is performed on the three-dimensional point cloud data to obtain a plurality of voxels, including: performing normalization processing on the three-dimensional point cloud data along the directions of the three axes x, y and z respectively according to the density and quantity of the three-dimensional point cloud data; performing uniform segmentation along the directions of the three axes x, y and z respectively based on the normalized three-dimensional point cloud data to obtain n*n*n voxels.

[0008] According to an embodiment of the present application, the voxel set with different density distributions is obtained by merging the plurality of voxels according to the segmented three-dimensional point cloud data and using a density similarity model, including: calculating the normalized density of each voxel and adjacent voxels respectively, wherein the normalized density represents the ratio of the number of point clouds in the voxel to the volume of the voxel; calculating the normalized density similarity between each voxel and adjacent voxels using the density similarity model according to the normalized density; and merging the current voxel and adjacent voxels in response to the normalized density similarity being less than a preset density similarity threshold.

[0009] According to an embodiment of the present application, the density similarity model is:

[0010]

[0011] wherein Δρ ij represents the normalized density similarity, ρ i represents the normalized density of the current voxel, ρ j represents the normalized density of the adjacent voxel.

[0012] According to an embodiment of the present application, the radius filtering parameter and the statistical filtering parameter of each voxel set are adaptively determined based on the merged voxels, including: determining the minimum number of neighborhood points according to the point cloud size parameter of each voxel set; determining the neighborhood average distance according to the minimum number of neighborhood points; and determining the search radius in the radius filtering parameter and the standard deviation in the statistical filtering parameter according to the neighborhood average distance.

[0013] According to an embodiment of the present application, the random noise removal processing is performed on the three-dimensional point cloud data according to the statistical filtering parameter, including: calculating the neighborhood average distance corresponding to the q minimum neighborhood points in the neighborhood of each point cloud in response to there being q minimum neighborhood points in the neighborhood of any point cloud; performing random noise removal processing on the three-dimensional point cloud data using a probability density function model according to the neighborhood average distance; and determining the current point cloud as a random noise point cloud and removing the current point cloud in response to the neighborhood average distance of the current point cloud being greater than a distance threshold, wherein the distance threshold represents the sum of the mean value of the neighborhood average distances of all point clouds in the neighborhood and k times the standard deviation of the neighborhood average distance.

[0014] According to an embodiment of the present application, the probability density function model is:

[0015]

[0016] wherein, denotes a probability density function, denotes the neighborhood average distance corresponding to q minimum neighborhood points of each point cloud, μ denotes the mean of the neighborhood average distances of all point clouds in the neighborhood, and σ denotes the standard deviation of the neighborhood average distances; in response to , the current point cloud is determined to be a random noise point cloud.

[0017] According to an embodiment of the present application, the isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing according to the radius filtering parameter includes: determining the number of point clouds within the search radius range according to the search radius of the current point cloud; performing isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing by using a radius filtering model according to the number of point clouds within the search radius range; and in response to the number of point clouds within the search radius range being less than the minimum number of neighborhood points, determining the current point cloud to be an isolated noise point cloud and removing the current point cloud.

[0018] According to an embodiment of the present application, the radius filtering model is:

[0019]

[0020] wherein, P' denotes the point cloud set reserved after the radius filtering, denotes the number of point clouds in the point cloud set within the search radius r, and Th min denotes a density threshold.

[0021] Another aspect of the present application provides a hybrid filtering denoising device based on adaptive parameters, comprising: an acquisition module configured to acquire three-dimensional point cloud data of a target object; a segmentation module configured to perform voxel segmentation processing on the three-dimensional point cloud data to obtain a plurality of voxels, wherein the point clouds in a single voxel are uniformly distributed; a merging module configured to perform merging processing on the plurality of voxels by using a density similarity model according to the segmented three-dimensional point cloud data to obtain voxel sets with different density distributions, wherein in response to the difference between the densities of a current voxel and an adjacent voxel being less than a preset density similarity threshold, the current voxel and the adjacent voxel are merged to obtain a corresponding voxel set; a determination module configured to adaptively determine radius filtering parameters and statistical filtering parameters of each voxel set based on the merged voxels, wherein the radius filtering parameters are used to determine isolated noise in the three-dimensional point cloud data, and the statistical filtering parameters are used to determine random noise in the three-dimensional point cloud data; a random noise removal module configured to perform random noise removal processing on the three-dimensional point cloud data according to the statistical filtering parameters; and an isolated noise removal module configured to perform isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing according to the radius filtering parameters.

[0022] Compared with the prior art, the adaptive parameter-based hybrid filtering denoising method and device provided by the embodiment of the present application has at least the following beneficial effects:

[0023] (1) The adaptive parameter-based hybrid filtering denoising method and device provided by the embodiment of the present application reduces the number of times of adaptive determination of hybrid filtering parameters by performing voxel merging on segmented voxels based on density similarity, and sequentially performs statistical filtering and hybrid filtering on different density distribution point clouds by using the adaptive determined hybrid filtering parameters, so as to finally realize fast denoising of unevenly distributed three-dimensional point clouds and effectively retain the detail information of complex components, greatly meeting the needs of intelligent manufacturing online detection.

[0024] (2) The adaptive parameter-based hybrid filtering denoising method and device provided by the embodiment of the present application does not need to manually set hybrid filtering parameters of different density distribution regions, and is suitable for three-dimensional point cloud denoising of various complex components.

[0025] (3) Compared with the optimization iterative denoising method, the adaptive parameter-based hybrid filtering denoising method and device provided by the embodiment of the present application has fast calculation speed, further improves the denoising speed through voxel merging, and can be applied to intelligent manufacturing online detection needs. BRIEF DESCRIPTION OF DRAWINGS

[0026] The above and other objects, features and advantages of the present application will become more apparent from the following description of embodiments of the present application taken in conjunction with the accompanying drawings, in which:

[0027] Figure 1 A flowchart schematically showing an adaptive parameter-based hybrid filtering denoising method according to an embodiment of the present application is shown;

[0028] Figure 2 A principle diagram schematically showing merging processing of a plurality of voxels by using a density similarity model according to an embodiment of the present application is shown;

[0029] Figure 3 A curve diagram schematically showing the size of a point cloud changing with an index according to an embodiment of the present application is shown;

[0030] Figure 4 A denoising effect comparison diagram of an adaptive parameter-based hybrid filtering denoising method and other methods according to an embodiment of the present application is shown;

[0031] Figure 5 A structure block diagram schematically showing an adaptive parameter-based hybrid filtering denoising device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0032] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. It is to be understood, however, that these descriptions are merely exemplary and are intended to illustrate the scope of the present application, not to limit it in any way. In the following detailed description of embodiments, numerous specific details are set forth in order to provide a thorough understanding of embodiments of the present application. However, it will be apparent to one skilled in the art that one or more embodiments can be practiced without these specific details. In other instances, well-known structures and

[0033] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used herein, the term "includes" and tautological expressions thereof, such as "including," "includes," "include," "contains," "containing," and so forth, shall not be taken to exclude

[0034] All terms used herein including technical and scientific terms have the same meanings as commonly understood by one of ordinary skill in the art unless otherwise defined herein. It should be noted that the terms used herein are defined as having meanings that are consistent with the context of the specification in which the terms are used, and should not be interpreted in an idealized or overly formal way.

[0035] In the case of using expressions similar to "at least one of A, B, and C, etc.", it is generally to be understood that the expression is to be interpreted to include any of the natural inclusive meanings of the term "at least one" (i.e., at least one and / or one or more), in addition to the meaning explicitly set forth herein (i.e., at least one of A, B, and C, etc.). Thus, the expression "at least one of A, B, and C, etc." would typically be understood to allow for the inclusion of only A, only B, only C, or a combination thereof such as A and B, A and C, B and C, or A, B, and C, etc., in, for example, a system having at least one of A, B, and C, etc.

[0036] In embodiments of the present application, the collection, updating, analysis, processing, use, transmission, provision, disclosure, storage, etc. of data (for example, including but not limited to user personal information) involved are in compliance with relevant laws and regulations, are used for legal purposes, and do not violate public order and good customs. In particular, necessary measures are taken to prevent illegal access to user personal information data, and to maintain user personal information security and network security.

[0037] A point cloud is a collection of three-dimensional samples of points on the surface of an object, each three-dimensional point containing position information, normal and tangent vectors, point cloud density, and other features. By analyzing the local or global features of the point cloud, the shape and semantics of the target can be obtained, and data analysis can be performed, for example, using the obtained three-dimensional point cloud for online detection of intelligent manufacturing products, to improve product quality.

[0038] The surface material of a complex structure workpiece is various and there is local shielding. Due to the different optical absorption and reflection characteristics of the multi-material surface, the large curvature variation of the complex workpiece edge and corner, and the non-uniform light refraction and reflection, the density distribution of the three-dimensional point cloud obtained based on the optical detection technology is non-uniform, and the three-dimensional point cloud of the measured target surface has various types of noise. According to the different mechanisms of generating noise, the three-dimensional point cloud noise of the complex structure workpiece includes isolated noise, noise cluster and small-scale random noise accompanying the main point cloud. Noise affects the accuracy of three-dimensional reconstruction based on three-dimensional point cloud and defect detection based on point cloud features, so it is necessary to remove noise from the three-dimensional point cloud of the complex structure workpiece.

[0039] In the prior art, the following methods are usually used for noise removal: (1) a filtering noise removal method based on optimization, which needs to first construct a target function item and a related constraint regular item about the original input point cloud, the target function is established based on distance, statistics, density and clustering, and filtering noise removal is completed through multiple iterations. For the three-dimensional point cloud of a complex component with a sharp curvature change, a local feature set of a point cloud neighborhood surface normal vector or a vertex normal vector is used for neighborhood grid noise removal, which is suitable for point cloud noise removal of components with complex geometric features and irregular sampling. (2) a sparsity noise removal method using first-order information of a normal vector, which can preserve the sharp curvature change characteristics of the point cloud surface, but a step effect will occur in the smooth surface area with uniform density distribution. (3) a L0 gradient minimization method, which can remove noise of a sharp feature surface. The above methods achieve noise removal through multiple optimizations, which is low in calculation efficiency and easy to lose information in the edge and other areas with sharp feature changes.

[0040] Statistical filtering can remove low-frequency noise, but does not consider noise removal in areas with sharp curvature changes, and is suitable for processing large amounts of three-dimensional point clouds with statistical characteristics. Radius filtering is suitable for removing isolated points or isolated cluster noise. Mixed filtering using statistical filtering and radius filtering can remove isolated noise while preserving the detailed features of the three-dimensional point cloud. However, the parameter setting of statistical filtering and radius filtering is related to the density distribution of the point cloud and the geometric structure of the complex component. When the three-dimensional point cloud is smooth and the density distribution is uniform, fixed parameter mixed filtering can achieve good noise removal effect. However, in the case of obvious detailed features and sparse point cloud in the edge and corner areas of the complex structure, fixed parameter mixed filtering is easy to lose the detailed information of the workpiece.

[0041] In summary, the existing noise removal methods for complex components have the problems of slow noise removal speed and easy loss of detailed information, which makes it difficult to be applied to online detection requirements of intelligent manufacturing.

[0042] Therefore, the embodiments of the present application provide a mixed filtering noise removal method and device based on adaptive parameters, which aims to solve the technical problem that the existing noise removal methods for complex components are difficult to be applied to online detection requirements of intelligent manufacturing.

[0043] In order to make the objects, technical solutions and advantages of the present application clearer, further specific embodiments will be described below in detail with reference to the accompanying drawings.

[0044] Figure 1 A flowchart of the adaptive parameter-based hybrid filtering denoising method according to an embodiment of the present application is schematically shown.

[0045] As shown in the adaptive parameter-based hybrid filtering denoising method of the embodiment, for example, it can include operations S1-S6. Figure 1

[0046] In operation S1, three-dimensional point cloud data of a target object is acquired.

[0047] In operation S2, voxel segmentation processing is performed on the three-dimensional point cloud data to obtain a plurality of voxels, wherein the point cloud within each voxel is uniformly distributed.

[0048] In operation S3, according to the segmented three-dimensional point cloud data, the plurality of voxels are merged using a density similarity model to obtain a voxel set with different density distributions, wherein in response to the difference between the density of a current voxel and that of an adjacent voxel being less than a preset density similarity threshold, the current voxel and the adjacent voxel are merged to obtain a corresponding voxel set.

[0049] In operation S4, based on the merged voxels, a radius filtering parameter and a statistical filtering parameter of each voxel set are adaptively determined, wherein the radius filtering parameter is used to determine isolated noise in the three-dimensional point cloud data, and the statistical filtering parameter is used to determine random noise in the three-dimensional point cloud data.

[0050] In operation S5, according to the statistical filtering parameter, random noise removal processing is performed on the three-dimensional point cloud data.

[0051] In operation S6, according to the radius filtering parameter, isolated noise removal processing is performed on the three-dimensional point cloud data after the random noise removal processing.

[0052] According to an embodiment of the present application, operation S1 acquires three-dimensional point cloud data of a target object, which can be, for example, three-dimensional point cloud data of the surface of an arbitrary complex structure workpiece obtained by optical detection technology.

[0053] In the present embodiment, due to the material and structure of the surface of the complex structure workpiece itself, as well as the characteristics of uneven light reflection and refraction, the three-dimensional point cloud data is unevenly distributed and contains various types of noise, such as random noise and isolated noise.

[0054] Random noise (i.e., small-scale random noise) generally refers to noise points that are distributed near the main body of the target object, have a small number, and have a slight deviation in spatial position.​

[0055] Isolated noise (i.e. isolated points or small noise clusters), usually refers to noise points distributed in clusters or single isolated distribution in space away from the main point cloud distribution area.

[0056] According to an embodiment of the present application, operation S2 voxel segmentation processing is performed on the three-dimensional point cloud data to obtain a plurality of voxels, which can specifically include operations S21-S22:

[0057] In operation S21, according to the density and quantity of the three-dimensional point cloud data, normalization processing is performed along the directions of the three axes x, y, and z of the three-dimensional point cloud data, respectively;

[0058] In operation S22, based on the normalized three-dimensional point cloud data, uniform segmentation is performed along the directions of the three axes x, y, and z to obtain n*n*n voxels.

[0059] In this embodiment, for example, according to the density and quantity of the three-dimensional point cloud, voxel segmentation can be performed on the three-dimensional point cloud along the x, y, and z directions, specifically:

[0060] First, normalization processing can be performed on the three-dimensional point cloud along the three mutually perpendicular directions of the x, y, and z axes. Taking the x direction as an example, the normalization mapping can be defined as:

[0061] (1)

[0062] Where min x and max x represent the maximum and minimum coordinates along the positive and negative directions of the x axis, respectively, and x i and x i respectively represent the x-axis coordinates of the point cloud before and after mapping.

[0063] After mapping the point cloud to the interval [-1, 1] according to formula (1), the three coordinate axes of the mapping space are further uniformly divided into n parts, and the three-dimensional point cloud can be uniformly divided into n*n*n initial voxels to form an equal-size voxel grid.

[0064] In this embodiment, the number of voxel segmentation can be 10*10*10, and finally 1000 voxels can be segmented in the x, y, and z directions, wherein a single voxel is considered to be uniformly distributed, and a fixed parameter hybrid filtering denoising is used for the voxel.

[0065] According to an embodiment of the present application, operation S3 utilizes a density similarity model to perform merging processing on the plurality of voxels according to the segmented three-dimensional point cloud data to obtain a voxel set with different density distributions, which can specifically include operations S31-S33:

[0066] In operation S31, the normalized density of each voxel and its adjacent voxels is calculated respectively, where the normalized density represents the ratio of the number of point clouds in a voxel to the voxel volume.

[0067] In operation S32, based on the normalized density, the normalized density similarity between each voxel and its neighboring voxels is calculated using the density similarity model.

[0068] In operation S33, in response to the normalized density similarity being less than a preset density similarity threshold, the current voxel is merged with its neighboring voxels.

[0069] In this embodiment, the computational complexity of adaptively determining the hybrid filtering parameters for individual voxels with uneven density distribution is significant. Therefore, to reduce computational complexity, voxels can be merged using a density similarity model. Subsequently, the corresponding parameters are determined based on the merged voxels for hybrid filtering. Specifically, the process of merging multiple voxels using a density similarity model is described in [link to documentation]. Figure 2 As shown.

[0070] Figure 2 The diagram illustrates the principle of merging multiple voxels using a density similarity model according to an embodiment of the present invention.

[0071] like Figure 2 As shown, (a) represents the initial voxel and voxel segmentation, (b) represents voxel merging, and (c) represents the voxel merging result. For each voxel's six adjacent neighboring voxels (up, down, left, right, front, and back), the current voxel V is calculated. i and adjacent voxel V j normalized density ρ i and ρ j The normalized density is the number of point clouds within a voxel / the volume of a voxel.

[0072] In this embodiment, the similarity of the normalized densities of adjacent voxels can be calculated using a density similarity model, which is as follows:

[0073] (2)

[0074] Where, Δρ ij Represents normalized density similarity, ρ i ρ represents the normalized density of the current voxel. j This represents the normalized density of adjacent voxels.

[0075] If the normalized density similarity Δρ ij If the density similarity is less than a preset threshold, the current voxel is merged with its neighboring voxels. In this embodiment, the density threshold can be set to, for example, 0.2.

[0076] In addition, if the similarity of the adjacent voxels is all less than the density similarity threshold, the adjacent voxel with the minimum similarity is preferentially merged, and the process is repeated until there is no voxel pair less than the density similarity threshold. The number of merged voxels is reduced to n*(log(n)) 3 . 2 .

[0077] According to an embodiment of the present application, operation S4 adaptively determines the radius filtering parameter and the statistical filtering parameter of each voxel set respectively based on the merged voxels, and can specifically include operations S41-S43:

[0078] In operation S41, the minimum neighborhood point number is determined according to the point cloud size parameter of each voxel set.

[0079] In operation S42, the neighborhood average distance is determined according to the minimum neighborhood point number.

[0080] In operation S43, the search radius in the radius filtering parameter and the standard deviation in the statistical filtering parameter are determined according to the neighborhood average distance.

[0081] In this embodiment, when the three-dimensional point cloud density distribution is uneven, fixed radius filtering can cause different denoising effects in different point cloud density regions, and the three-dimensional point cloud with sharp edges or small structures in sparse distribution can be mistakenly removed as noise, resulting in loss of detailed information of complex components.

[0082] Therefore, in this embodiment, different neighborhood radii are used for radius filtering for voxels with different point cloud densities, and the minimum neighborhood point number q of radius filtering significantly affects the effect of radius filtering. The neighborhood radius parameter of radius filtering is adaptively determined according to the point cloud density of the merged voxel, which can better preserve the detailed information.

[0083] In order to avoid overfitting (neighborhood too large) of small point clouds or underfitting (neighborhood too small) of large point clouds, the neighborhood point number can be dynamically adjusted based on the point cloud density of the merged voxel. The curve of the point cloud voxel size N changing with the index k (k<1) is shown in Figure 3 .

[0084] Figure 3 A curve graph showing the point cloud size changing with the index according to an embodiment of the present application is schematically shown.

[0085] As shown in Figure 3 , the larger the point cloud size index, the higher the data compression rate; small-scale voxels are prone to over-smoothing, which prevents the target three-dimensional points from being mistakenly removed. As the voxel size increases, the index grows slowly, which can ensure a good neighborhood to stabilize the statistical characteristics and reduce noise interference. In this embodiment, the index is comprehensively determined, for example, it can be selected as 0.3 to ensure that the neighborhood point number increases slowly with N, avoiding overfitting (neighborhood too large) of small point clouds or underfitting (neighborhood too small) of large point clouds.

[0086] Therefore, the minimum number of neighborhood points q of the radius filtering can be determined by the exponential k of the point cloud size N of the current merged voxel:

[0087] (3)

[0088] wherein a represents an empirical minimum threshold value, ensuring that the neighborhood point number distribution is reasonable when the point cloud density is too small, and a = 5 in general.

[0089] After determining the minimum number of neighborhood points q of the radius filtering, the neighborhood average distance is calculated , and the search radius of the radius filtering and the standard deviation of the statistical filtering are determined according to the neighborhood average distance.

[0090] The calculation formula of the neighborhood average distance is as follows:

[0091] (4)

[0092] wherein, for the three-dimensional point cloud P i in the voxel, there are N-1 neighboring points, N is the number of three-dimensional points in the voxel, d ij represents the Euclidean distance between the point P i and the neighborhood point P j .

[0093] The calculation formula of the search radius r of the radius filtering is as follows:

[0094] (5)

[0095] Further, the calculation formula of the standard deviation σ of the statistical filtering is determined according to the neighborhood average distance .

[0096] (6)

[0097] wherein the constant term can ensure that in the dense area, even if the neighborhood average distance is close to zero, the noise significantly deviating from the neighborhood average distance can still be filtered out. The scaling coefficient k before the neighborhood average distance can be appropriately adjusted according to the point cloud density, so as to avoid that the correct three-dimensional points in the sparse area of the point cloud are removed due to the large distance fluctuation. In the embodiment, the scaling coefficient k can be set as k = 0.5, for example.

[0098] According to the embodiment of the present application, the operation S5 performs the random noise removal processing on the three-dimensional point cloud data according to the statistical filtering parameters, which can specifically include operations S51-S53:

[0099] In operation S51, the neighborhood average distance corresponding to the q minimum neighborhood points of each point cloud is calculated in response to that there are q minimum neighborhood points in the neighborhood of any point cloud;

[0100] In operation S52, random noise removal processing is performed on the 3D point cloud data based on the average neighborhood distance and using a probability density function model.

[0101] In operation S53, in response to the average neighborhood distance of the current point cloud being greater than a distance threshold, the current point cloud is determined to be a random noise point cloud and the current point cloud is removed. The distance threshold represents the sum of the mean of the average neighborhood distances of all point clouds in the neighborhood and the standard deviation of k times the average neighborhood distance.

[0102] In this embodiment, statistical filtering assumes that the 3D point cloud in the neighborhood of a certain point conforms to a statistical distribution law, while point clouds that do not conform to the statistical distribution law are considered noise points. It is applicable to small-scale random noise caused by sensor noise, environmental interference, or data acquisition errors. Specifically:

[0103] For any point cloud Pi, its q nearest neighbors can be determined through the above operation S4.

[0104] In the absence of significant external disturbances, it is assumed that the 3D point cloud follows a statistical distribution within a local area. Then, the average neighborhood distance d is calculated for the q nearest neighbors of each 3D point cloud. ij It is assumed that the expected value of its statistical model follows a normal distribution, and the probability density function model is used to remove random noise from the three-dimensional point cloud.

[0105] In this embodiment, the probability density function model is:

[0106] (7)

[0107] in, Represents the probability density function. Let μ represent the mean neighborhood distance of the q smallest neighboring points of each point cloud, μ represent the mean of the mean neighborhood distances of all point clouds within the neighborhood, and σ represent the standard deviation of the mean neighborhood distances.

[0108] Assuming a normal point cloud follows a normal distribution, the criterion for determining whether a current 3D point is a noise point is: That is, when hour( (Characteristic distance threshold), which considers the current point cloud as a random noise point cloud.

[0109] According to an embodiment of the present invention, operation S6 performs isolated noise removal processing on the three-dimensional point cloud data after random noise removal processing based on the radius filtering parameters, specifically including operations S61~S63:

[0110] In operation S61, the number of points in the search radius is determined based on the current search radius of the point cloud.

[0111] In operation S62, according to the number of point clouds in the search radius range, isolated noise denoising processing is performed on the three-dimensional point cloud data after random noise removal processing by using a radius filtering model.

[0112] In operation S63, in response to the number of point clouds in the search radius range being less than the minimum number of neighborhood points, it is determined that the current point cloud is an isolated noise point cloud, and the current point cloud is removed.

[0113] In this embodiment, the radius filtering is suitable for removing isolated points or small noise clusters in the point cloud.

[0114] Firstly, the search radius of the current three-dimensional point cloud can be set, and then the number of three-dimensional point clouds in the search radius range is counted.

[0115] Then, according to the number of point clouds in the search radius range, isolated noise denoising processing is performed on the three-dimensional point cloud data after random noise removal processing by using a radius filtering model.

[0116] In this embodiment, the radius filtering model is:

[0117]

[0118] wherein P' represents the point cloud set reserved after radius filtering, represents the number of point clouds in the point cloud set in the search radius r, and Th min represents the density threshold.

[0119] If the number of three-dimensional point clouds is lower than the minimum number of neighborhood points q, the current point is regarded as an isolated point and is removed, otherwise the point is reserved.

[0120] The mixed filtering denoising method based on adaptive parameters provided in the embodiment of the application reduces the number of times of adaptive determination of mixed filtering parameters by performing voxel merging on the segmented voxels based on density similarity, and sequentially performs statistical filtering and mixed filtering on point clouds with different density distributions by using the mixed filtering parameters determined adaptively, so as to finally realize fast denoising of three-dimensional point clouds with uneven distribution and effectively retain the detail information of complex components, greatly meeting the needs of online detection of intelligent manufacturing.

[0121] Figure 4 A denoising effect comparison diagram of the mixed filtering denoising method based on adaptive parameters according to the embodiment of the application and other methods is schematically shown.

[0122] As Figure 4As shown, to verify the performance of the adaptive parameter-based hybrid filtering denoising method of this invention, a classic 3D rabbit model was used as the original data. The point cloud of the rabbit's back region was simplified by a uniform downsampling of 0.1 times to simulate the sparse point cloud at the edges when scanning the point cloud of complex components in 3D measurement. The rabbit's head and legs were simplified by a uniform downsampling of 0.4 times to simulate the incomplete 3D data acquisition caused by occlusion during 3D measurement.

[0123] In the table, (a) shows the obtained 3D point cloud sample with uneven density distribution, i.e., the original point cloud; (b), (c), and (d) show the comparison of the denoising results using fixed-parameter hybrid filtering, statistical filtering, and the method of this invention (adaptive parameter hybrid filtering), respectively. The comparison results show that the fixed-parameter method, due to the use of dense point cloud distribution to determine fixed parameters for hybrid filtering, fails to remove some outlier noise, leading to excessive denoising at the back edges and leg occlusion areas of the sparse point cloud distribution, resulting in the loss of detailed information. Statistical filtering removes random noise near the target subject, but does not filter out isolated noise clusters. In contrast, the adaptive method of this invention determines the hybrid filtering parameters differently for dense and sparse point cloud regions, thus preserving detailed information. The specific 3D point cloud denoising results of different filtering methods are shown in Table 1.

[0124] Table 1. Noise denoising results of 3D point clouds using different filtering methods

[0125]

[0126] As shown in Table 1, the denoising rate and detail edge preservation rate of the method of this invention are significantly better than other methods. Compared with fixed parameter filtering, although the adaptive parameter hybrid filtering increases the computation time due to the adaptive determination of parameters, it reduces the number of times the adaptive parameters are determined by voxel merging based on density similarity. The time used for voxel merging and parameter determination is much less than that used by the denoising method. The time advantage is more obvious when the point cloud scale is larger.

[0127] Figure 5 A schematic diagram of a hybrid filtering and denoising device based on adaptive parameters according to an embodiment of the present invention is shown.

[0128] like Figure 5 As shown, the hybrid filtering denoising device 500 based on adaptive parameters according to an embodiment of the present invention includes: an acquisition module 510, a segmentation module 520, a merging module 530, a determination module 540, a random noise removal module 550, and an isolated noise removal module 560.

[0129] The acquisition module 510 is used to acquire the three-dimensional point cloud data of the target object.

[0130] The segmentation module 520 is configured to perform voxel segmentation processing on the three-dimensional point cloud data to obtain a plurality of voxels, wherein the point cloud in a single voxel is uniformly distributed.

[0131] The merging module 530 is configured to perform merging processing on the plurality of voxels according to the segmented three-dimensional point cloud data by using a density similarity model to obtain a voxel set with different density distributions, wherein in response to a difference between the density of a current voxel and the density of a neighboring voxel being less than a preset density similarity threshold, the current voxel and the neighboring voxel are merged to obtain a corresponding voxel set.

[0132] The determination module 540 is configured to adaptively determine a filtering radius parameter and a filtering statistical parameter of each voxel set based on the merged voxels, wherein the radius filtering parameter is used to determine isolated noise in the three-dimensional point cloud data, and the statistical filtering parameter is used to determine random noise in the three-dimensional point cloud data.

[0133] The random noise removal module 550 is configured to perform random noise removal processing on the three-dimensional point cloud data according to the statistical filtering parameter.

[0134] The isolated noise removal module 560 is configured to perform isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing according to the radius filtering parameter.

[0135] Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present application, or at least part of any one or more of the modules, sub-modules, units, sub-units can be implemented in one module. Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present application can be split into multiple modules for implementation. Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present application can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or any other reasonable way of hardware or firmware by integrating or packaging the circuit, or in any one of software, hardware and firmware or in an appropriate combination of any one or more of them. Alternatively, one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present application can be at least partially implemented as computer program modules that can perform corresponding functions when executed.

[0136] For example, any multiple of the acquisition module 510, the segmentation module 520, the merging module 530, the determination module 540, the random noise removal module 550 and the isolated noise removal module 560 can be combined in one module / unit / sub-unit, or any one of the modules / units / sub-units can be split into multiple modules / units / sub-units. Alternatively, at least part of the function of one or more of the modules / units / sub-units can be combined with at least part of the function of other modules / units / sub-units, and implemented in one module / unit / sub-unit. According to an embodiment of the present application, at least one of the acquisition module 510, the segmentation module 520, the merging module 530, the determination module 540, the random noise removal module 550 and the isolated noise removal module 560 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on board, a system on package, an application specific integrated circuit (ASIC), or any other reasonable manner of integrating or packaging a circuit, etc. hardware or firmware, or in any one of the three implementation manners of software, hardware and firmware or in any appropriate combination of any of the three. Alternatively, at least one of the acquisition module 510, the segmentation module 520, the merging module 530, the determination module 540, the random noise removal module 550 and the isolated noise removal module 560 can be at least partially implemented as a computer program module which can perform the corresponding function when executed.

[0137] It should be noted that the part of the adaptive parameter based hybrid filter denoising device in the embodiments of the present application corresponds to the part of the adaptive parameter based hybrid filter denoising method in the embodiments of the present application, and the description of the part of the adaptive parameter based hybrid filter denoising device is specifically referred to the part of the adaptive parameter based hybrid filter denoising method, which will not be repeated here.

[0138] The embodiments of the present application are described above. However, these embodiments are only for illustrative purposes, and are not intended to limit the scope of the present application. Although each embodiment is described above, this does not mean that the measures in each embodiment cannot be used advantageously in combination. Without departing from the scope of the present application, those skilled in the art can make various alternatives and modifications, which should fall within the scope of the present application.

Claims

1. A hybrid filter denoising method based on adaptive parameters, characterized in that, The method comprises: acquiring three-dimensional point cloud data of a target object; performing voxel segmentation processing on the three-dimensional point cloud data to obtain a plurality of voxels, wherein the point cloud in each voxel is uniformly distributed; based on the segmented three-dimensional point cloud data, performing merging processing on the plurality of voxels using a density similarity model to obtain a voxel set with different density distributions, wherein in response to the difference between the density of a current voxel and that of an adjacent voxel being less than a preset density similarity threshold, the current voxel and the adjacent voxel are merged to obtain a corresponding voxel set; based on the merged voxels, adaptively determining a radius filtering parameter and a statistical filtering parameter of each voxel set, wherein the radius filtering parameter is used to filter out isolated noise in the three-dimensional point cloud data, and the statistical filtering parameter is used to filter out random noise in the three-dimensional point cloud data; performing random noise removal processing on the three-dimensional point cloud data according to the statistical filtering parameter; performing isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing according to the radius filtering parameter.

2. The method of claim 1, wherein, The voxel segmentation processing on the three-dimensional point cloud data to obtain a plurality of voxels comprises: performing normalization processing on the three-dimensional point cloud data along the directions of the three axes x, y, and z, respectively, according to the density and quantity of the three-dimensional point cloud data; based on the normalized three-dimensional point cloud data, performing uniform segmentation along the directions of the three axes x, y, and z to obtain n*n*n voxels.

3. The method of claim 1, wherein, The merging processing on the plurality of voxels using a density similarity model based on the segmented three-dimensional point cloud data to obtain a voxel set with different density distributions comprises: calculating the normalized density of each voxel and adjacent voxels, respectively, wherein the normalized density represents the ratio of the number of point clouds in the voxel to the volume of the voxel; calculating the normalized density similarity between each voxel and adjacent voxels using a density similarity model according to the normalized density; in response to the normalized density similarity being less than a preset density similarity threshold, merging the current voxel and the adjacent voxel.

4. The method of claim 3, wherein, The density similarity model is: where Δρ ij represents the normalized density similarity, p i represents the normalized density of the current voxel, p j represents the normalized density of the neighboring voxel.

5. The method of claim 1, wherein, The adaptively determining of the radius filtering parameter and the statistical filtering parameter of each voxel set based on the merged voxels comprises: determining the minimum number of neighborhood points according to the point cloud size parameter of each voxel set; determining the average distance of the neighborhood according to the minimum number of neighborhood points; determining the search radius in the radius filtering parameter and the standard deviation in the statistical filtering parameter according to the average distance of the neighborhood.

6. The method of claim 5, wherein, The random noise removal processing on the three-dimensional point cloud data according to the statistical filtering parameter comprises: in response to there being q minimum neighborhood points in the neighborhood of any point cloud, calculating the average distance of the neighborhood corresponding to the q minimum neighborhood points of each point cloud; performing random noise removal processing on the three-dimensional point cloud data using a probability density function model according to the average distance of the neighborhood; in response to the average distance of the neighborhood of the current point cloud being greater than a distance threshold, determining that the current point cloud is a random noise point cloud and removing the current point cloud, wherein the distance threshold represents the sum of the average of the average distance of the neighborhood of all point clouds in the neighborhood and k times the standard deviation of the average distance of the neighborhood.

7. The method of claim 6, wherein, The probability density function model is: wherein, denotes a probability density function, denotes the neighborhood average distance corresponding to the q smallest neighborhood points of each point cloud, μ denotes the mean of the neighborhood average distances of all point clouds within the neighborhood, and σ denotes the standard deviation of the neighborhood average distances. in response to , determining that the current point cloud is a random noise point cloud.

8. The method of claim 5, wherein, The isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing according to the radius filtering parameter comprises: determining the number of point clouds in the search radius range according to the search radius of the current point cloud; performing isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing by using a radius filtering model according to the number of point clouds in the search radius range; in response to the number of point clouds in the search radius range being less than the minimum number of neighbor points, determining that the current point cloud is an isolated noise point cloud, and removing the current point cloud.

9. The method of claim 8, wherein, The radius filtering model is: wherein P' represents the set of point clouds reserved after radius filtering, represents the number of point clouds within the search radius r Th min represents the density threshold.

10. An adaptive parameter based hybrid filter denoising device, characterized by, The device comprises: an acquisition module configured to acquire three-dimensional point cloud data of a target object; a segmentation module configured to perform voxel segmentation processing on the three-dimensional point cloud data to obtain a plurality of voxels, wherein the point clouds in a single voxel are uniformly distributed; a merging module configured to perform merging processing on the plurality of voxels by using a density similarity model according to the segmented three-dimensional point cloud data to obtain a voxel set with different density distributions, wherein in response to the difference between the densities of a current voxel and an adjacent voxel being less than a preset density similarity threshold, the current voxel and the adjacent voxel are merged to obtain a corresponding voxel set; a determination module configured to adaptively determine a radius filtering parameter and a statistical filtering parameter of each voxel set based on the merged voxels, wherein the radius filtering parameter is used to filter out isolated noise in the three-dimensional point cloud data, and the statistical filtering parameter is used to filter out random noise in the three-dimensional point cloud data; a random noise removal module configured to perform random noise removal processing on the three-dimensional point cloud data according to the statistical filtering parameter; an isolated noise removal module configured to perform isolated noise removal processing on the three-dimensional point cloud data after the random noise removal processing according to the radius filtering parameter.

Images