A material point cloud missing area automatic filling method and system, medium

Abstract

The application provides a kind of material body point cloud missing area automatic completion method and system, medium, related to data processing technical field, method includes: using voxel grid downsampling is carried out to the data preprocessing of original point cloud, obtains denoising point cloud;Carrying out missing area boundary detection, obtains multiple missing area closed contour;Multiple geometric environment joint features are constructed;Carrying out missing area classification, obtains multiple same type missing area cluster;Feature space similarity quantification is carried out, and multiple parameter template samples are screened;Carrying out completion parameter optimization, obtains multiple parameter template library;Self-adaptive adjustment completion parameter, carries out completion operation to the remaining missing area closed contour, and outputs the completion point cloud model.The application solves the technical problem that the point cloud completion of prior art mainly depends on the geometric morphology of point cloud and the interpolation of adjacent points, ignores environmental features and other information, resulting in low completion quality, missing details, and unable to accurately restore the geometric shape of complex objects.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method, system, and medium for automatically filling in missing areas of a material point cloud. Background Technology

[0002] With the rapid development of laser scanning, 3D imaging technology, and sensors, point cloud data is widely used in various fields. Point cloud data can accurately describe the geometry of an object's surface. However, in practical applications, due to environmental factors, equipment limitations, or special data acquisition conditions, point cloud data often suffers from missing data. These missing data often affect subsequent point cloud processing and analysis processes, such as object modeling, shape restoration, and path planning. Therefore, how to effectively and automatically fill in missing areas and restore the integrity of point cloud data is an important research topic.

[0003] Existing completion methods mainly rely on the geometry of point clouds and interpolation of adjacent points for completion, ignoring environmental features and other information. These methods perform poorly when dealing with complex missing regions, especially for irregularly shaped missing regions or regions lacking sufficient support from surrounding point clouds. The completion results are often not accurate enough, resulting in low completion quality, missing details, and an inability to accurately restore the geometry of complex objects. Summary of the Invention

[0004] This application provides an automatic completion method, system, and medium for missing areas in a point cloud of a material, aiming to solve the technical problem that existing point cloud completion mainly relies on the geometric shape of the point cloud and the interpolation of adjacent points, ignoring environmental features and other information, resulting in low completion quality, missing details, and inability to accurately restore the geometric shape of complex objects.

[0005] The first aspect of this application discloses an automatic method for filling in missing regions of a point cloud. The method includes: preprocessing the original point cloud using voxel lattice downsampling to obtain a denoised point cloud; fusing RGB images to perform boundary detection on the denoised point cloud to obtain multiple closed contours of missing regions; extracting boundary geometric attributes, local environment attributes, and missing cause attributes of the multiple closed contours of missing regions to construct multiple joint geometric environment features; classifying the missing regions based on the multiple joint geometric environment features to obtain multiple clusters of similar missing regions; performing feature space similarity quantification on the multiple clusters of similar missing regions to filter multiple parameter template samples; optimizing the completion parameters of the multiple parameter template samples to obtain a multiple parameter template library; adaptively adjusting the completion parameters in the multiple parameter template library according to the joint geometric environment features of the missing regions; performing a completion operation on the remaining closed contours of the missing regions in the multiple clusters of similar missing regions; and outputting a completed point cloud model.

[0006] The second aspect of this application discloses an automatic completion system for missing regions in a material point cloud. This system is used in the aforementioned automatic completion method for missing regions in a material point cloud. The system includes: a data preprocessing module for preprocessing the original point cloud using voxel downsampling to obtain a denoised point cloud; a boundary detection module for performing missing region boundary detection on the denoised point cloud by fusing RGB images to obtain multiple closed contours of missing regions; a joint feature construction module for extracting the boundary geometric attributes, local environment attributes, and missing cause attributes of the multiple closed contours of missing regions to construct multiple joint geometric environment features; and missing region classification. The system comprises the following modules: a module for classifying missing regions based on the joint features of multiple geometric environments to obtain multiple clusters of missing regions of the same type; a similarity quantification module for performing feature space similarity quantification on the multiple clusters of missing regions of the same type to filter multiple parameter template samples; a completion parameter optimization module for optimizing the completion parameters of the multiple parameter template samples to obtain a multiple parameter template library; and a point cloud model output module for adaptively adjusting the completion parameters in the multiple parameter template library according to the joint features of the geometric environment of the missing regions, performing completion operations on the closed contours of the remaining missing regions in the multiple clusters of missing regions of the same type, and outputting a completed point cloud model.

[0007] The third aspect disclosed in this application provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the automatic completion method for missing regions of a material point cloud in the first aspect.

[0008] One or more technical solutions provided in this application have at least the following beneficial effects: By using voxel lattice downsampling to reduce the amount of point cloud data while eliminating noise points and retaining effective information, the efficiency of subsequent processing is improved, and the impact of noise on the point cloud completion results is reduced, ensuring clearer and cleaner point cloud data and laying the foundation for subsequent processing. Combining RGB images with point cloud data, and using color and texture information from the images to assist in the boundary detection of missing regions, enhances the accuracy of boundary detection. This provides more accurate boundaries for missing regions in subsequent completion operations, effectively improving the reliability of the completion process. By extracting the geometric, local environment, and causal attributes of missing regions and integrating them into comprehensive features, the morphology and surrounding environment of missing regions can be described more accurately. This provides comprehensive information for the classification and completion of missing regions, helping subsequent completion operations to be more accurate and efficient. By combining the geometric environment features of missing regions with local environment information, clustering algorithms are used to group missing regions with similar features into the same category. This classification method improves the accuracy of missing region classification. The targeted processing ensures that similar regions use the same or similar completion methods, improving the efficiency and accuracy of completion. By quantifying the feature space similarity of clusters of missing regions of the same type, the parameter template sample that best fits each cluster is selected. This ensures that the missing regions within each cluster can use the most suitable completion template, thereby improving the completion quality and reducing erroneous completions or unnatural transitions. By optimizing the parameter template samples, each template can be finely adjusted for specific types of missing regions. Through optimization, each template in the template library becomes more accurate and efficient, capable of handling more diverse missing regions, thus providing more accurate completion results. By adaptively adjusting the completion parameters in the template library, combined with the specific features of the missing regions, the accuracy of completion is further improved. The output completed point cloud model maintains realism while smoothly transitioning to the surrounding areas, ensuring a natural transition in geometry and density between the completed region and the surrounding point cloud, achieving a high-quality completion effect.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0010] Figure 1 This is a schematic flowchart of a method for automatically filling in missing areas of a material point cloud, provided in an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of an automatic filling system for missing areas of a material point cloud, provided in an embodiment of this application.

[0012] Figure labeling: Data preprocessing module 10, boundary detection module 20, joint feature construction module 30, missing region classification module 40, similarity quantification module 50, completion parameter optimization module 60, point cloud model output module 70. Detailed Implementation

[0013] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0014] Example 1, as Figure 1 As shown in the figure, this application embodiment provides a method for automatically filling in missing regions of a material point cloud, the method including: Voxel lattice downsampling was used to preprocess the original point cloud data to obtain a denoised point cloud.

[0015] Voxel lattice downsampling is used to divide point cloud data into a uniform grid of small cubes, or voxels, and retain a representative point within each voxel—the centroid of all points within that voxel. This helps reduce the number of points in the point cloud, thus accelerating subsequent processing and removing invalid points caused by measurement noise. Through voxel lattice downsampling, redundant parts of the point cloud are removed, and the remaining points more accurately represent the true geometry, effectively eliminating measurement errors and environmental noise, making subsequent processing more accurate.

[0016] By fusing RGB images, the boundary detection of missing regions in the denoised point cloud is performed to obtain multiple closed contours of missing regions.

[0017] Combining RGB images with point cloud data—for example, projecting an image onto a point cloud using camera calibration information, or inferring a point cloud from an RGB image using depth information—provides richer environmental information and helps improve the accuracy of missing region boundary detection. By performing boundary detection on the denoised point cloud, the boundaries of missing regions are found. These missing regions are caused by factors such as water reflection, blind spots, or object occlusion. Boundary detection analyzes the brightness, color, or geometric features in the point cloud and image data to identify the contours of these missing regions. The detected missing region boundaries may be incomplete or broken. The contour closure step uses curve fitting, interpolation, and other methods to connect these boundaries, forming a closed contour, which is the basis for subsequent completion operations.

[0018] Extract the boundary geometric attributes, local environment attributes, and missing cause attributes of the closed contours of the multiple missing regions, and construct multiple joint geometric environment features.

[0019] Extract the geometric features of the closed contour of the missing region, such as roundness, rectangularity, curvature, and the degree of bending of the boundary. These features describe the shape characteristics of the missing region and can reveal the specific form of the missing region. For example, some missing regions are regular, while other regions are random. Specifically, calculate the roundness and rectangularity of the boundary, calculate the curvature along the boundary point sequence, and evaluate the smoothness and curvature of the boundary.

[0020] Analyzing the environmental characteristics of the missing region, such as point cloud density distribution and normal direction, can help us better understand the causes of the missing region and the difficulty of its completion. Specifically, calculating the density ratio between the missing region and the surrounding area, i.e., comparing the point cloud density of the missing region with that of the surrounding area, helps us understand the density difference between the missing region and the surrounding area and infer the cause of the missing region. Calculating the proportion of the principal direction of the normal direction of the boundary of the missing region helps us analyze the spatial directional characteristics of the missing region.

[0021] By combining information from RGB images, such as brightness and color, we can further analyze the causes of the missing areas. For example, the missing areas may be caused by factors such as reflection, shadows, or viewing angle issues. Therefore, we need to infer the causes of the missing areas based on the characteristics of RGB images, such as highlights.

[0022] The geometric attributes, local environment attributes, and missing cause attributes mentioned above are normalized and concatenated to form a multi-dimensional feature vector, called the geometric environment joint feature. These features are used for subsequent missing region classification and completion operations.

[0023] Based on the joint features of the multiple geometric environments, missing regions are classified to obtain multiple clusters of missing regions of the same type.

[0024] This study analyzes the joint geometric and environmental features of missing regions, categorizing different types of missing regions into multiple classes. For example, some missing regions are caused by water reflection, unlike those caused by occlusion or blind spots. Traditional clustering algorithms such as K-means and spectral clustering can be used for classification. Specifically, a feature matrix is ​​constructed, where each row represents a missing region and each column represents a geometric environmental feature, such as roundness or density ratio. Based on these features, clustering algorithms are used to classify the missing regions, resulting in multiple clusters of the same type. Each cluster represents a group of missing regions with similar features. The missing regions within each cluster have similar geometric and environmental features, requiring similar processing methods. Depending on the number of clusters and the classification, subsequent completion work can be more targeted.

[0025] The feature space similarity quantification is performed on the multiple clusters of missing regions of the same type to filter multiple parameter template samples.

[0026] Measuring the similarity between regions within multiple clusters of similar missing regions aims to identify the most suitable completion template by quantifying the feature differences between regions within each cluster. Specifically, the joint feature vector of the geometric environment is extracted from the missing regions in each cluster. By calculating the similarity between these feature vectors, such as Euclidean distance or cosine similarity, the similarity between regions within the cluster is quantified, thereby determining which missing regions are similar and which can be completed using the same or similar templates.

[0027] Based on similarity quantification, the most suitable parameter template sample for each cluster is selected. The parameter template sample is the basic unit used to complete the missing regions. Each template contains a set of completion parameters. The goal of selecting templates is to find the completion scheme that best suits the current cluster features. Specifically, the feature vectors of each missing region within the cluster are weighted and averaged to determine the feature center of the cluster. Based on these feature centers and the feature differences of individuals within the cluster, the parameter template sample that is closest to the cluster center is selected.

[0028] The parameters of the multiple parameter template samples are completed and optimized to obtain a multiple parameter template library.

[0029] After selecting suitable template samples for each cluster, the completion parameters of these templates are further optimized to ensure the quality of the completion results. The optimization process aims to reduce the error between the completed results and the actual point cloud, such as the reconstruction error of point cloud boundaries, by adjusting the template parameters. Specifically, initial completion parameters are extracted for each template, and point cloud completion operations are performed based on these parameters to generate preliminary completion results. The difference between the original point cloud and the completed results is compared using the Hausdorff distance as a metric, which quantifies the maximum distance difference between two point sets. The completion parameters are optimized by minimizing the Hausdorff distance. In this process, a gradient descent algorithm is used to gradually adjust the completion parameters by calculating the gradient of the loss function until convergence to the optimal value. The loss function can incorporate metrics such as Hausdorff distance and curvature difference to ensure the accuracy and geometric continuity of the completion results. After optimizing the completion parameters, an optimized parameter template library is finally obtained. This parameter template library contains multiple optimized completion templates, each providing an accurate completion solution for a specific type of missing region.

[0030] Based on the joint features of the geometric environment of the missing region, the completion parameters in the multiple parameter template libraries are adaptively adjusted, and the completion operation is performed on the closed contours of the remaining missing regions in the multiple clusters of missing regions of the same type, and the completed point cloud model is output.

[0031] In the parameter template library, each completion template contains a set of parameters suitable for a specific missing region type. However, since the shape and environment of the actual missing region can vary, fine-tuning is required based on the actual geometric characteristics of each missing region. This adjustment ensures that the completion result does not solely depend on the initial parameters of the template, but can automatically adjust the parameters according to the specific situation of the missing region. The adjustment process depends on the shape characteristics of the missing region (such as roundness, curvature, etc.), the surrounding environment (such as density distribution, normal direction), and the causal characteristics (such as reflection, occlusion, etc.). Based on these characteristics, template parameters, such as the smoothness and fill density of the completed region, are adjusted.

[0032] After the above adaptive adjustments, the remaining missing regions are filled using completion parameters from the parameter template library. The boundaries of these regions have been identified and closed, but the missing data within them still needs to be filled. Based on the template parameters, the completion operation automatically generates point cloud data within the missing regions according to the surrounding environmental features. During the completion process, new point cloud data is generated within the missing regions based on the adjusted completion parameters. This includes inferring the point cloud positions within the missing regions based on neighborhood information. After the completion operation is complete, a complete point cloud model is output. This model includes the original point cloud data and the completed point cloud of the missing regions. The completed point cloud model is more complete and accurate, accurately reflecting the three-dimensional structure of the object.

[0033] Furthermore, the parameters of the multiple parameter template samples are completed and optimized to obtain multiple parameter template libraries, including: Extract the first initial completion parameter set from the first parameter template sample, perform point cloud completion operation, and output the first point cloud missing completion result; extract the original complete region point cloud from the original point cloud; perform gradient descent optimization to update the completion parameters based on the Hausdorff distance between the original complete region point cloud and the first point cloud missing completion result, until the convergence condition is met, and output the first optimized completion parameter set, wherein the convergence condition includes Hausdorff distance change rate < 0.01 and curvature difference < 0.05, and inject geometric conformal constraints and density continuity constraints during the optimization process; deconstruct the first optimized completion parameter set based on the missing region type to obtain a first parameter template library including multiple parameter key values ​​and multiple applicable scenario descriptions.

[0034] An analysis object, termed the first parameter template sample, is selected from the parameter template library, and its corresponding preset completion parameters, termed the first initial completion parameter set, are extracted as the starting parameters for the completion operation. The point cloud completion operation is performed using the first initial completion parameter set. The core objective of the completion operation is to fill in the missing regions, ensuring a natural connection between the completed point cloud and its surrounding environment. This is achieved through interpolation based on boundary features, or by using techniques such as triangulation and surface fitting. After the completion operation is complete, a first point cloud missing completion result containing the completed data is generated. This result reflects the missing regions after the completion operation.

[0035] Extract the original complete region point cloud related to the current completion region from the original point cloud. The point cloud data of the complete region is complete and is used to provide a reference for the completion region. For example, the normal direction, point cloud density, and local curvature of the complete region can help infer the completion strategy and parameters of the missing region.

[0036] Hausdorff distance is a method for measuring the difference between two point sets, used to compare the shape similarity between point clouds. In this step, Hausdorff distance is used to measure the difference between the original complete region point cloud and the missing completion result of the first point cloud. Ideally, the completed region should match the original region as closely as possible; therefore, the goal is to minimize the Hausdorff distance. By calculating the Hausdorff distance between the original complete region point cloud and the missing completion result of the first point cloud, the quality of the current completion result is evaluated. A smaller Hausdorff distance indicates that the completed result is closer to the original region, while a larger distance indicates that the completed result differs significantly from the original region.

[0037] To optimize the completion parameters, gradient descent is used. Gradient descent is an optimization algorithm that adjusts parameters by backpropagating errors. In this step, gradient descent is used to continuously adjust the completion parameters so that the completed result is closer to the original point cloud. The optimization process continues until the following convergence conditions are met: the rate of change of Hausdorff distance < 0.01, meaning that the change of Hausdorff distance is very small, indicating that the completion result is close to the optimal solution and further optimization will not bring significant improvement; and the curvature difference < 0.05. The curvature difference represents the difference between the curvature change of the completed region and the original region. If the curvature difference is small, it means that the curvature of the completed region matches that of the original region, thus ensuring the geometric continuity of the completed region.

[0038] During the optimization process, geometric conformance constraints and density continuity constraints are injected. Geometric conformance constraints are used to ensure the continuity of the geometric shape of the completed region and avoid unnatural deformation. By introducing geometric conformance constraints, it can be ensured that the geometric features of the completed region, such as surface smoothness and boundary shape, will not undergo abrupt changes and conform to the real physical form. Density continuity constraints are used to ensure that the point cloud density of the completed region is consistent with the surrounding region. This helps to avoid the phenomenon of uneven density in the completed region, that is, the density of the point cloud matches the physical density of the real object.

[0039] After optimization, the first optimized completion parameter set is output. The first optimized completion parameter set contains the parameter values ​​after gradient descent optimization. These parameters can maximize the accuracy and naturalness of the completion result while satisfying the convergence condition.

[0040] The first optimized completion parameter set is deconstructed, extracting multiple parameter key-value pairs. Each key-value pair represents a key parameter required for a completion operation. These parameters include: geometric parameters, such as boundary smoothness, curvature constraints, and normal direction; filling parameters, such as point cloud density and interpolation method; and local environment parameters, such as the density ratio of the missing region to the surrounding region and the proportion of normal direction. In addition, multiple applicable scenario descriptions are extracted. These descriptions are generated based on the geometric features, environmental features, and causes of the missing region. For example, the description template is suitable for missing regions with circular missing boundaries or obvious highlight spots.

[0041] Furthermore, by fusing RGB images to perform missing region boundary detection on the denoised point cloud, multiple closed contours of missing regions are obtained, including: Edge detection is performed on the RGB image to generate a binary edge image; boundary feature points are extracted by traversing the denoised point cloud to obtain a point cloud boundary feature point set, wherein the types of boundary feature points include normal abrupt change points and density anomaly points; a projection transformation model is constructed based on camera calibration parameters and point cloud acquisition device pose; using the projection transformation model as a spatial mapping reference, the point cloud boundary feature point set is projected onto the binary edge image plane, edge distance matching is performed, and spatially consistent point sets are selected; topological reconstruction is performed on the spatially consistent point sets to generate an initial boundary point set, and then contour smoothing is performed to form the closed contours of the multiple missing regions.

[0042] Edge detection algorithms, such as the Sobel operator and Canny edge detection, extract edge information from RGB images. These algorithms identify the boundaries of objects in an image, such as the boundaries of object surfaces, shadows, or reflective areas. The algorithms determine edges by calculating the gradient of pixel intensity changes in the image. Edge regions are areas in the image where grayscale values ​​change drastically; these regions correspond to the contours or depth changes of objects. After edge detection, the image is converted into a binary image, where edge parts are white (255) and non-edge parts are black (0), thus clearly defining the boundaries of objects.

[0043] Traversing the denoised point cloud, the point cloud data contains the three-dimensional coordinates of each point on the object's surface. Some points in the point cloud are located on the object's boundary, and the geometric features of these points are different from those of the points inside the point cloud. Therefore, in the process of extracting boundary feature points, we focus on the following types of feature points: abrupt change in normal direction, where the normal direction changes drastically at the boundary of the object's surface. For example, when a flat surface intersects with a raised or recessed area, the change in normal direction is abrupt; density anomaly points, where points on the boundary have different point cloud densities. The point cloud density at the edge is usually lower, so the density features of these points are abnormal compared to the neighboring areas. Density anomaly points are another characteristic feature of the boundary.

[0044] The camera calibration process uses a known calibration board to calculate the camera's intrinsic and extrinsic parameters. Intrinsic parameters include focal length and optical center position, while extrinsic parameters include the camera's pose, i.e., rotation and translation relative to the world coordinate system. These parameters are used to project the 3D point cloud data onto the 2D image plane. The point cloud acquisition device pose refers to the device's position and orientation in space, used to transform the acquired point cloud data from the device coordinate system to the world coordinate system or other reference coordinate systems. Based on the camera calibration parameters and the point cloud acquisition device pose, a projection transformation model is constructed. This model converts the 3D coordinates in the point cloud into 2D coordinates in the image. Specifically, the projection transformation model is based on the principle of perspective projection, using known camera intrinsic and extrinsic parameters to map each 3D point cloud point onto the image plane.

[0045] Using a projection transformation model, the set of boundary feature points in the point cloud is projected onto the image plane. Through this projection process, the boundary feature points that were originally in three-dimensional space are transformed into their corresponding positions in the two-dimensional edge image. The goal of this step is to match the boundary information in the point cloud data with the edge information in the RGB image. Since the spatial distribution of the image and the point cloud data are different, the projection transformation makes the two comparable on the same plane.

[0046] Edge distance matching is performed by calculating the distance between the point cloud boundary feature points projected onto the binary edge image plane and the edges in the image. Through matching, it can be determined whether the boundary features in the point cloud are consistent with the edges in the image. The distance from each point cloud boundary feature point to its corresponding position in the binary edge image is calculated and filtered using a predetermined threshold. If a feature point in the point cloud matches an edge in the image, these point cloud boundary feature points are considered spatially consistent with the edge in the image. Based on the edge distance matching results, those point cloud boundary feature points that match the image edges are selected, and these points form a spatially consistent point set.

[0047] Topological reconstruction refers to reconstructing the boundary structure of a missing region by analyzing the connectivity between spatially consistent point sets. In this step, topological reconstruction restores the boundary shape of the spatially consistent point sets, making it more accurately reflect the actual geometric features of the missing region. The topological reconstruction process organizes the point sets into a graph or mesh structure and reconstructs the boundary based on the connectivity between points. This is achieved through Delaunay triangulation, Voronoi diagrams, or other point-set-based geometric methods. After topological reconstruction, an initial set of boundary points is formed, which is a coarse representation of the boundary of the missing region.

[0048] Contour smoothing aims to eliminate rough and irregular boundaries generated during topological reconstruction. By smoothing the initial boundary point set, the boundaries are made smoother, eliminating subtle fluctuations and making the closed contours more consistent with the actual shape of the object. Smoothing methods include using algorithms such as B-splines and Kalman filters to smooth the boundaries, maintaining the overall shape of the boundaries while reducing local irregularities. After smoothing, multiple closed contours of missing regions are formed. These contours represent the accurate boundaries of the missing regions in the point cloud and provide a foundation for subsequent completion operations.

[0049] Furthermore, after topological reconstruction of the spatially consistent point set to generate an initial boundary point set, contour smoothing is performed to form closed contours of the multiple missing regions, including: Based on a preset distance threshold, the spatially consistent point set is segmented using Euclidean distance to obtain multiple spatial clusters; α-shape boundary extraction is performed on the multiple spatial clusters to obtain multiple initial boundary point sets; after convex hull calculation of the multiple initial boundary point sets, the convex hull vertices are sorted to identify non-convex regions and perform interpolation connections to obtain multiple ordered boundary point sequences; after uniform cubic B-spline curve fitting using the multiple ordered boundary point sequences as control points, samples are taken at equal intervals along the fitted curve to construct the closed contours of the multiple missing regions.

[0050] Euclidean distance is a method for measuring the distance between two points. Based on the Euclidean distance between points in a spatially consistent point set, points in a point cloud can be divided into different clusters. In this step, a preset distance threshold is set to segment the spatially consistent point set. The distance threshold determines whether two points belong to the same cluster; if the distance between two points is less than the threshold, they are considered to belong to the same cluster. Through Euclidean distance segmentation, the spatially consistent point set is divided into multiple spatial clusters, each cluster representing a region in the point cloud and corresponding to a specific missing region boundary.

[0051] Alpha-shape is a technique for extracting boundaries by analyzing the local geometry of a point set. Unlike convex hull, alpha-shape not only analyzes the external boundaries of a point set but also captures the shape details within the point set. By selecting an appropriate alpha parameter, the fineness of the extracted boundaries can be adjusted. When the alpha value is small, the extracted boundaries are tighter and can capture the details of the point set; when the alpha value is large, the boundaries become smoother and are suitable for describing approximate edges. After applying alpha-shape boundary extraction to each spatial cluster, multiple initial boundary point sets are obtained. These point sets represent the boundaries of each cluster and provide the foundation for subsequent convex hull calculations.

[0052] A convex hull is the smallest convex polygon generated from a set of points, containing all points in that set. Convex hull calculation identifies the external boundary of the point set and provides a reference for further boundary fitting. In this step, convex hull calculation is performed on each initial boundary point set to obtain the external contour of each boundary. These convex hulls provide a preliminary framework for subsequent boundary optimization. After calculating the convex hull, the vertices are sorted. Sorting helps to clarify the order of the vertices, making the boundary construction more standardized. These vertices are connected in sequence to form a complete boundary.

[0053] Outside the convex hull boundary, non-convex regions may exhibit irregular corners or concave areas. To make the boundary more closely resemble the actual shape, these non-convex regions need to be processed. During the processing of non-convex regions, interpolation connection methods are used to smooth and repair them, aiming to eliminate existing cracks or discontinuities and make the boundary smoother. After interpolation connection and sorting, multiple ordered boundary point sequences are finally obtained. These point sequences represent the closed boundaries of the missing regions, possessing good coherence and smoothness, and are used for the completion operation.

[0054] Cubic B-spline curves are a mathematical tool for smooth curve fitting. They define the curve by the relationships between control points, ensuring good continuity at each point, including position, tangent, and curvature. By using multiple ordered boundary points as control points and connecting them using cubic B-spline curve fitting, a smooth closed contour line is generated. This curve not only accurately reflects the relationships between boundary points but also provides a smoother boundary shape, avoiding abrupt boundaries caused by the irregularities in point cloud data.

[0055] After fitting, evenly spaced sampling is performed along the B-spline curve. Evenly spaced sampling involves selecting a number of sampling points uniformly on the curve to ensure sufficient point cloud density at the boundaries of the missing region. This sampling method ensures a uniform distribution of boundary points in the missing region, thus providing accurate boundary definitions for subsequent point cloud completion operations. After B-spline curve fitting and evenly spaced sampling, a closed contour of the missing region is constructed. This contour, composed of multiple uniformly distributed sampling points, accurately represents the boundary of the missing region and provides a reference for subsequent point cloud completion.

[0056] Furthermore, the boundary geometric attributes, local environment attributes, and missing cause attributes of the closed contours of the multiple missing regions are extracted to construct multiple joint geometric environment features, including: The shape features of the first contour are quantified by calculating the circularity and rectangularity of the closed contour of the first missing region. After calculating the local curvature sequence along the contour point sequence of the closed contour of the first missing region, the average curvature and maximum curvature of the contour points are statistically analyzed to quantify the curvature features of the first boundary. Local environmental features are extracted from the closed contour of the first missing region to obtain the density ratio inside and outside the first contour and the proportion of the first principal direction of the normal. The first image region of the closed contour of the first missing region is extracted from the RGB image, the average brightness and brightness variance are calculated, and the proportion of the first highlight spot is detected. The shape features of the first contour, the curvature features of the first boundary, the density ratio inside and outside the first contour, the proportion of the first principal direction of the normal, and the proportion of the first highlight spot are normalized and stitched together to output the joint features of the geometric environment.

[0057] Circularity and rectangularity are two geometric shape characteristics used to describe the shape of a closed contour. Circularity measures how similar a contour is to a circle; a high circularity indicates a close resemblance to a circle, while a low circularity indicates an ellipse or other complex shape. Circularity is calculated by comparing the radius of the contour's smallest circumscribed circle to the actual area or perimeter of the contour. Rectangularity measures how similar a contour is to a rectangle; a high rectangularity indicates a close resemblance to a rectangle. Rectangularity is measured by comparing the aspect ratio of the contour's circumscribed rectangle to the actual width and height of the contour. A rectangularity value closer to 1 indicates a closer resemblance to a rectangle.

[0058] Curvature is a geometric quantity that describes the degree of bending of a curve or surface. In a sequence of boundary points, each point has a local curvature value, representing the degree of bending of the curve at that point. Local curvature is obtained by calculating the change in the normal direction at that point. By calculating the geometric features of the neighborhood of each contour point, such as the change in the normal direction of adjacent points, the local curvature is obtained. A larger local curvature indicates that the curve bends more sharply at that point, while a smaller local curvature indicates that the curve is flatter.

[0059] The average curvature of a contour point is the average of the curvatures of all points in a contour point sequence, reflecting the overall curvature of the entire boundary. The average curvature of the contour is obtained by calculating the average of the local curvatures of all contour points. A smaller average curvature indicates a smoother boundary, while a larger average curvature indicates greater boundary curvature.

[0060] Maximum curvature is the maximum value of local curvature in the sequence of contour points. It reflects the most curved part of the contour. By comparing the local curvature values ​​of each contour point, the point with the largest curvature is found, and the maximum curvature value is obtained. A larger maximum curvature value indicates that there are sharp curved parts in the contour.

[0061] By statistically analyzing the average curvature and maximum curvature, the curvature characteristics of the closed contour of the first missing region are quantified. These characteristics can reflect the geometric shape of the region and provide guidance for subsequent completion operations.

[0062] Local environmental features refer to the geometric features describing the environment surrounding the missing region. These features help analyze the relationship between the missing region and its surroundings, providing clues for completion. The inside-outside density ratio represents the difference in point cloud density between the boundary of the missing region and its surrounding areas. The missing region has a lower density, while the surrounding areas have a higher density. It is calculated by statistically analyzing the number of point clouds inside and outside the boundary, or by estimating the local density of the point cloud, for example, using the K-nearest neighbor method to calculate the local density, and then obtaining the ratio. The normal direction proportion reflects the consistency between the normal direction of the missing region boundary and the normal direction of the surrounding environment. The normal direction of the missing region differs from that of the surrounding environment; therefore, it is necessary to analyze the proportion of these normal directions to provide more suitable normal directions for the completion operation. It is calculated by calculating the principal direction of the normal of the missing region boundary and comparing it with the principal direction of the normal of the surrounding area. The proportion of the principal normal direction represents the similarity between the principal directions of the normal of the missing region and the normal of the surrounding environment.

[0063] RGB image region extraction involves extracting information about the corresponding region from the RGB image based on the previously established closed contour of the first missing region. This is done by matching the point cloud boundaries with the image to select the portion of the image corresponding to the missing point cloud region. This process relies on camera calibration parameters and a projection transformation model to map the spatial information of the point cloud onto the image plane, extracting the image region corresponding to the missing region.

[0064] Average brightness is the average brightness of all pixels in an extracted image region, reflecting the overall brightness of the region. Brightness variance measures the degree of brightness variation within an image region; a smaller brightness variation results in a lower variance, while a larger variation results in a higher variance. Calculation method: Calculate the brightness of pixels in the image region, which can be the brightness in a grayscale image or the average brightness in an RGB image. Then calculate the average and variance of these brightness values.

[0065] Highlights are areas of high brightness in an image caused by reflection or changes in lighting. Highlights in missing areas provide important information for completion, especially on reflective surfaces or in transparent areas. Highlight ratio refers to the proportion of highlight areas within an image region. It is determined by thresholding the bright areas in the image (areas with brightness greater than a certain value) and then calculating the proportion of these highlight areas within the total image region to obtain the first highlight ratio.

[0066] To combine features from different sources, these features are first normalized. Normalization aims to eliminate dimensional differences between features, making them comparable under the same dimension and preventing certain features from dominating the splicing process. Normalization employs standardization methods, such as subtracting the mean, dividing by the standard deviation, or minimax normalization. The normalized features are then spliced ​​into a multidimensional vector. The output geometric environment joint feature is a feature vector that combines information from multiple aspects, comprehensively describing the geometric shape and environmental characteristics of the missing region.

[0067] Furthermore, local environmental features are extracted from the closed contour of the first missing region to obtain the density ratio inside and outside the first contour and the proportion of the first principal normal direction, including: Based on the closed contour boundary of the first missing region, a first annular region is constructed by expanding the noise-reduced point cloud. The average density of the first region and the average density of the first annular region are calculated respectively within the closed contour of the first missing region and the first annular region. The density ratio inside and outside the first contour is compared and output. The standard deviation angle of the point cloud normal direction is calculated for the first annular region to output the proportion of the first normal principal direction.

[0068] Using the closed contour boundary of the first missing region as a reference, a ring-shaped region is created outside this boundary. Outward expansion refers to extending a certain distance outward from the original boundary to form a ring-shaped region. This ring-shaped region includes the boundary of the missing region and extends into the surrounding area, providing a neighborhood region for analyzing the geometric and density relationships between the missing region and its surrounding environment. The outward expansion distance is a preset constant or dynamically calculated based on the local density of the point cloud.

[0069] Density calculation analyzes the local density distribution of point clouds. First, it calculates the point cloud density within the closed contour of the first missing region, i.e., the point cloud distribution inside the boundary. Density calculation methods include estimating the density by the number of neighboring points of each point, such as the K-nearest neighbor method. This density value reflects the sparsity of the missing region; typically, the density of the missing region is low because point cloud data is missing in that area. Next, it calculates the average density of the first annular region, i.e., the distribution density of the point cloud within the outer boundary region. Since the annular region contains more point cloud data, its density is usually higher. Similarly, the density calculation method is achieved through neighborhood statistics of the point cloud.

[0070] The inner-outer density ratio is obtained by comparing the density within the missing region with the density outside the annular region. Typically, the density of the missing region is lower, while the density of the annular region is higher. The inner-outer density ratio of the first contour is output, which reflects the sparsity of the missing region relative to its surrounding environment.

[0071] In point cloud processing, the normal direction refers to the orientation of each point in three-dimensional space, used to describe the shape of an object's surface. Each point cloud has a corresponding normal vector, which is estimated through the geometric relationship between adjacent points. For the first annular region, the normal vector of each point is calculated, and based on these normal vectors, the standard deviation angle of the normal direction of the entire annular region is calculated. The standard deviation angle represents the degree of dispersion of the normal vector distribution, that is, the degree of variation in the normal direction. If the standard deviation angle is small, it indicates that the surface of the region is relatively smooth and the normal direction is consistent; while if the standard deviation angle is large, it indicates that the surface of the region has greater variation and the normal direction changes more drastically.

[0072] The proportion of the principal normal direction reflects the consistency of normal directions. If the proportion is large, it indicates that the surface of the region is relatively flat, the normal directions are consistent, and the completion of missing areas is relatively simple. If the proportion is small, it indicates that the region has many curved surfaces or irregular shapes, requiring more surface fitting or smoothing during completion. Calculating the proportion of the principal normal direction involves analyzing the principal direction of the normal vectors and outputting the proportion of this direction to the total normal vector directions. This value guides subsequent completion methods and helps optimize surface continuity during the completion process.

[0073] Furthermore, based on the joint features of the multiple geometric environments, missing regions are classified to obtain multiple clusters of missing regions of the same type, including: A joint feature matrix of the environment is constructed based on the joint features of the multiple geometric environments; a similarity matrix of missing regions is calculated based on the joint feature matrix of the environment; spectral clustering analysis is performed on the similarity matrix of missing regions to generate multiple initial region clusters; the multiple initial region clusters are optimized based on the contour coefficient, and the multiple missing region clusters of the same type are output.

[0074] The joint geometric and environmental features describe the geometric and environmental characteristics of the missing regions, providing a set of numerical descriptions for each missing region. The joint environmental feature matrix integrates the features of all missing regions into a single matrix. Assuming multiple missing regions, each with a set of geometric and environmental features, each row of this matrix represents a feature vector for a missing region, and each column represents a single feature. The joint environmental feature matrix provides a complete feature representation of all missing regions and forms the basis for subsequent similarity calculations and cluster analysis.

[0075] The similarity matrix is ​​obtained by comparing the similarity between the feature vectors of the missing regions in the joint feature matrix of the environment. This matrix is ​​used to quantify the degree of similarity between the missing regions. The similarity matrix is ​​generated by calculating the distance between the feature vectors of each pair of missing regions in the matrix. Each element in the matrix represents the similarity between two missing regions. A high similarity means that the two missing regions have similar geometric shapes and environmental features.

[0076] Spectral clustering is a graph-based clustering method that divides missing regions into multiple clusters by performing eigenvalue decomposition (spectral decomposition) on the similarity matrix. Spectral clustering effectively handles complex nonlinear relationships and performs well in high-dimensional data. Specifically, the similarity matrix is ​​viewed as an adjacency matrix of a graph, where nodes represent missing regions and edge weights between nodes represent their similarity. By performing eigenvalue decomposition on the Laplacian matrix of the graph, a low-dimensional spatial representation is obtained. Then, standard clustering methods, such as K-means, are applied to cluster these low-dimensional features, thereby assigning missing regions to different clusters. Through spectral clustering analysis, missing regions are divided into multiple clusters, each representing a type of missing region, and regions within a cluster share similar features.

[0077] The silhouette coefficient is a metric for evaluating clustering quality. It measures how closely each data point is clustered with other points in its cluster and how closely it is separated from the midpoint of its nearest neighbor cluster. The silhouette coefficient ranges from -1 to 1. A larger value indicates better clustering, while a smaller value indicates less than ideal clustering.

[0078] Based on the initial spectral clustering results, silhouette coefficients are used to optimize multiple initial region clusters. By adjusting clustering parameters, such as the number of clusters and the similarity threshold, the silhouette coefficient is maximized. The optimization goal is to make the regions within each cluster as similar as possible and as different as possible from the regions in other clusters. For example, if a cluster has a low silhouette coefficient, it means that the cluster contains missing regions with inconsistent features. In this case, the clustering strategy is adjusted to make the regions within the cluster more consistent. After silhouette coefficient optimization, the final clusters of missing regions of the same type are output. The missing regions in these clusters have high similarity in features and are suitable for using the same completion method.

[0079] Furthermore, feature space similarity quantification is performed on the multiple clusters of missing regions of the same type to filter multiple parameter template samples, including: Extract the M joint geometric environment feature vectors of the M missing regions within the first cluster of missing regions of the same type, and calculate the arithmetic mean to locate the feature center of the first cluster; based on the M weighted Euclidean distances between the M joint geometric environment feature vectors and the feature center of the first cluster, sort the first parameter template samples in ascending order.

[0080] Each missing region within the first cluster of similar missing regions possesses a set of joint geometric-environment feature vectors describing its geometric and environmental characteristics. These feature vectors include data across multiple dimensions, such as contour shape, curvature, density ratio, normal direction, and image brightness features. For the M missing regions within a cluster, after extracting their joint geometric-environment feature vectors, the average feature of these feature vectors is calculated using an arithmetic mean. The purpose of this process is to identify the central feature of the cluster, serving as its representative. The arithmetic mean is calculated by averaging the feature values ​​across each dimension to obtain a new feature vector representing the feature center of the cluster. This step simplifies and unifies the features of all missing regions within the cluster, providing a vector representing the overall characteristics of the cluster.

[0081] Weighted Euclidean distance is a metric used to measure the difference between two feature vectors. Unlike ordinary Euclidean distance, weighted Euclidean distance considers the different importance of different features in the distance calculation. For each feature vector of a missing region, the weighted Euclidean distance with the feature center of the cluster is calculated using the following formula: , among which, among which, D j w is the weighted Euclidean distance between the feature vector of the j-th missing region and the cluster center. i It is the weight of the i-th feature, v ji It is the i-th feature of the j-th missing region, C i The i-th feature of the cluster center, and the weight w of the weighted Euclidean distance. i The settings are based on the importance of each feature.

[0082] After calculating the weighted Euclidean distance for each missing region, all missing regions are sorted in ascending order of their weighted Euclidean distance. This means that missing regions with higher similarity to the cluster center (i.e., regions with smaller weighted Euclidean distances) will be ranked higher. Based on the ranking results, the missing region samples most similar to the cluster feature centers are selected. Through this step, the most suitable first-parameter template samples are chosen. These first-parameter template samples are selected based on the principle of minimum weighted Euclidean distance, thus they can better fit the missing regions within the cluster and are used to complete other missing regions in the current cluster.

[0083] Example 2, based on the same inventive concept as the automatic completion method for missing areas of a material point cloud in the previous examples, such as... Figure 2 As shown in the figure, this application embodiment provides an automatic completion system for missing areas in a material point cloud, the system comprising: The data preprocessing module 10 is used to preprocess the original point cloud using voxel lattice downsampling to obtain a denoised point cloud; the boundary detection module 20 is used to fuse RGB images to perform missing region boundary detection on the denoised point cloud to obtain multiple missing region closed contours; the joint feature construction module 30 is used to extract the boundary geometric attributes, local environment attributes, and missing cause attributes of the multiple missing region closed contours to construct multiple geometric environment joint features; the missing region classification module 40 is used to classify missing regions based on the multiple geometric environment joint features to obtain multiple clusters of missing regions of the same type; the similarity metric module 50 is used to perform feature space similarity metric on the multiple clusters of missing regions of the same type to filter multiple parameter template samples; the completion parameter optimization module 60 is used to optimize the completion parameters of the multiple parameter template samples to obtain multiple parameter template libraries; the point cloud model output module 70 is used to adaptively adjust the completion parameters in the multiple parameter template libraries according to the geometric environment joint features of the missing regions, perform completion operations on the remaining missing region closed contours in the multiple clusters of missing regions of the same type, and output a completed point cloud model.

[0084] Furthermore, the parameter completion optimization module 60 is used to perform the following operation steps: Extract the first initial completion parameter set from the first parameter template sample, perform point cloud completion operation, and output the first point cloud missing completion result; extract the original complete region point cloud from the original point cloud; perform gradient descent optimization to update the completion parameters based on the Hausdorff distance between the original complete region point cloud and the first point cloud missing completion result, until the convergence condition is met, and output the first optimized completion parameter set, wherein the convergence condition includes Hausdorff distance change rate < 0.01 and curvature difference < 0.05, and inject geometric conformal constraints and density continuity constraints during the optimization process; deconstruct the first optimized completion parameter set based on the missing region type to obtain a first parameter template library including multiple parameter key values ​​and multiple applicable scenario descriptions.

[0085] Furthermore, the boundary detection module 20 is used to perform the following operation steps: Edge detection is performed on the RGB image to generate a binary edge image; boundary feature points are extracted by traversing the denoised point cloud to obtain a point cloud boundary feature point set, wherein the types of boundary feature points include normal abrupt change points and density anomaly points; a projection transformation model is constructed based on camera calibration parameters and point cloud acquisition device pose; using the projection transformation model as a spatial mapping reference, the point cloud boundary feature point set is projected onto the binary edge image plane, edge distance matching is performed, and spatially consistent point sets are selected; topological reconstruction is performed on the spatially consistent point sets to generate an initial boundary point set, and then contour smoothing is performed to form the closed contours of the multiple missing regions.

[0086] Furthermore, the boundary detection module 20 is used to perform the following operation steps: Based on a preset distance threshold, the spatially consistent point set is segmented using Euclidean distance to obtain multiple spatial clusters; α-shape boundary extraction is performed on the multiple spatial clusters to obtain multiple initial boundary point sets; after convex hull calculation of the multiple initial boundary point sets, the convex hull vertices are sorted to identify non-convex regions and perform interpolation connections to obtain multiple ordered boundary point sequences; after uniform cubic B-spline curve fitting using the multiple ordered boundary point sequences as control points, samples are taken at equal intervals along the fitted curve to construct the closed contours of the multiple missing regions.

[0087] Furthermore, the joint feature construction module 30 is used to perform the following operation steps: The shape features of the first contour are quantified by calculating the circularity and rectangularity of the closed contour of the first missing region. After calculating the local curvature sequence along the contour point sequence of the closed contour of the first missing region, the average curvature and maximum curvature of the contour points are statistically analyzed to quantify the curvature features of the first boundary. Local environmental features are extracted from the closed contour of the first missing region to obtain the density ratio inside and outside the first contour and the proportion of the first principal direction of the normal. The first image region of the closed contour of the first missing region is extracted from the RGB image, the average brightness and brightness variance are calculated, and the proportion of the first highlight spot is detected. The shape features of the first contour, the curvature features of the first boundary, the density ratio inside and outside the first contour, the proportion of the first principal direction of the normal, and the proportion of the first highlight spot are normalized and stitched together to output the joint features of the geometric environment.

[0088] Furthermore, the joint feature construction module 30 is used to perform the following operation steps: Based on the closed contour boundary of the first missing region, a first annular region is constructed by expanding the noise-reduced point cloud. The average density of the first region and the average density of the first annular region are calculated respectively within the closed contour of the first missing region and the first annular region. The density ratio inside and outside the first contour is compared and output. The standard deviation angle of the point cloud normal direction is calculated for the first annular region to output the proportion of the first normal principal direction.

[0089] Furthermore, the missing region classification module 40 is used to perform the following operation steps: A joint feature matrix of the environment is constructed based on the joint features of the multiple geometric environments; a similarity matrix of missing regions is calculated based on the joint feature matrix of the environment; spectral clustering analysis is performed on the similarity matrix of missing regions to generate multiple initial region clusters; the multiple initial region clusters are optimized based on the contour coefficient, and the multiple missing region clusters of the same type are output.

[0090] Furthermore, the similarity metric module 50 is used to perform the following operational steps: Extract the M joint geometric environment feature vectors of the M missing regions within the first cluster of missing regions of the same type, and calculate the arithmetic mean to locate the feature center of the first cluster; based on the M weighted Euclidean distances between the M joint geometric environment feature vectors and the feature center of the first cluster, sort the first parameter template samples in ascending order.

[0091] Through the foregoing detailed description of an automatic filling method for missing areas of a material point cloud, those skilled in the art can clearly understand the automatic filling system for missing areas of a material point cloud in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section.

[0092] Example 3 provides a storage medium on which a computer program is stored, which, when executed by a processor, implements any step of Example 1.

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for automatically filling in missing regions of a point cloud of a material, characterized in that, The method includes: Voxel lattice downsampling was used to preprocess the original point cloud data to obtain a denoised point cloud; By fusing RGB images to perform boundary detection on the denoised point cloud, multiple closed contours of missing regions are obtained: Extract the boundary geometric attributes, local environment attributes, and missing cause attributes of the closed contours of the multiple missing regions, and construct multiple joint geometric environment features; Based on the joint features of the multiple geometric environments, missing regions are classified to obtain multiple clusters of missing regions of the same type. The feature space similarity quantification is performed on the multiple clusters of missing regions of the same type to filter multiple parameter template samples; The parameters of the multiple parameter template samples are completed and optimized to obtain a multiple parameter template library; Based on the joint features of the geometric environment of the missing region, the completion parameters in the multiple parameter template libraries are adaptively adjusted, and the completion operation is performed on the closed contours of the remaining missing regions in the multiple clusters of missing regions of the same type, and the completed point cloud model is output. Specifically, parameter completion and optimization are performed on the multiple parameter template samples to obtain multiple parameter template libraries, including: Extract the first initial completion parameter set from the first parameter template sample, perform point cloud completion operation, and output the first point cloud missing completion result; Extract the original complete region point cloud of the first parameter template sample from the original point cloud; Based on the Hausdorff distance between the original complete region point cloud and the missing completion result of the first point cloud, gradient descent optimization is performed to update the completion parameters until the convergence condition is met, and the first optimized completion parameter set is output. The convergence condition includes the Hausdorff distance change rate < 0.01 and the curvature difference < 0.

05. Geometric conformal constraints and density continuity constraints are injected during the optimization process. Based on the missing region type, the first optimized completion parameter group is deconstructed to obtain a first parameter template library that includes multiple parameter key values ​​and multiple applicable scenario descriptions.

2. The method for automatically filling in missing areas of a material point cloud as described in claim 1, characterized in that, By fusing RGB images to perform boundary detection on the denoised point cloud for missing regions, multiple closed contours of missing regions are obtained, including: Perform edge detection on the RGB image to generate a binary edge image; The boundary feature points are extracted by traversing the denoised point cloud to obtain a set of point cloud boundary feature points, wherein the types of the boundary feature points include normal abrupt change points and density anomaly points. A projection transformation model is constructed based on camera calibration parameters and the pose of the point cloud acquisition device. Using the projection transformation model as a spatial mapping reference, the point cloud boundary feature point set is projected onto the binary edge image plane, edge distance matching is performed, and spatially consistent point sets are selected. After topological reconstruction of the spatially consistent point set to generate an initial boundary point set, contour smoothing is performed to form closed contours of the multiple missing regions.

3. The method for automatically filling in missing areas of a material point cloud as described in claim 2, characterized in that, After performing topological reconstruction on the spatially consistent point set to generate an initial boundary point set, contour smoothing is performed to form closed contours of the multiple missing regions, including: Based on a preset distance threshold, the set of spatially consistent points is segmented by Euclidean distance to obtain multiple spatial clusters; α-shape boundary extraction is performed on the multiple spatial clusters to obtain multiple initial boundary point sets; After performing convex hull calculation on the multiple initial boundary point sets, the convex hull vertices are sorted to identify non-convex regions and perform interpolation connections to obtain multiple ordered boundary point sequences. After fitting a uniform cubic B-spline curve using the multiple ordered boundary point sequences as control points, sampling is performed at equal intervals along the fitted curve to construct the closed contours of the multiple missing regions.

4. The method for automatically filling in missing areas of a material point cloud as described in claim 1, characterized in that, Extract the boundary geometric attributes, local environment attributes, and missing cause attributes of the closed contours of the multiple missing regions, and construct multiple joint geometric environment features, including: The shape features of the first contour are quantified by calculating the roundness and rectangularity of the closed contour of the first missing region. After calculating the local curvature sequence along the contour point sequence of the closed contour of the first missing region, the average curvature and maximum curvature of the contour points are statistically analyzed to quantify the curvature characteristics of the first boundary. Local environmental features are extracted from the closed contour of the first missing region to obtain the density ratio inside and outside the first contour and the proportion of the first principal normal direction. Extract the first image region from the closed contour of the first missing region from the RGB image, calculate the average brightness and brightness variance, and detect the proportion of the first highlight spot. The first contour shape feature, the first boundary curvature feature, the first contour inner and outer density ratio, the proportion of the first normal principal direction, and the proportion of the first highlight spot are normalized and spliced ​​together to output the joint geometric environment feature.

5. The method for automatically filling in missing areas of a material point cloud as described in claim 4, characterized in that, Local environmental features are extracted from the closed contour of the first missing region to obtain the density ratio inside and outside the first contour and the proportion of the first principal normal direction, including: Based on the closed contour boundary of the first missing region, a first annular region is constructed by expanding the noise reduction point cloud outward; After calculating the average density of the first region and the average density of the first ring within the closed contour of the first missing region and the first ring within the first ring region respectively, the density ratio inside and outside the first contour is compared and output. The standard deviation angle of the point cloud normal direction is calculated for the first annular region to output the proportion of the first principal normal direction.

6. The method for automatically filling in missing areas of a material point cloud as described in claim 1, characterized in that, Based on the joint features of the multiple geometric environments, missing regions are classified to obtain multiple clusters of missing regions of the same type, including: Construct an environment joint feature matrix based on the joint features of the multiple geometric environments; Calculate the missing region similarity matrix based on the aforementioned joint environmental feature matrix; Spectral clustering analysis is performed on the similarity matrix of the missing regions to generate multiple initial region clusters; The multiple initial region clusters are optimized based on the contour coefficient, and the multiple missing region clusters of the same type are output.

7. The method for automatically filling in missing areas of a material point cloud as described in claim 1, characterized in that, The feature space similarity quantification is performed on the multiple clusters of missing regions of the same type to filter multiple parameter template samples, including: Extract the joint feature vectors of the geometric environment of M missing regions within the first cluster of missing regions of the same type, and calculate the arithmetic mean to locate the feature center of the first cluster; Based on the M joint feature vectors of the geometric environment and the M weighted Euclidean distances of the first cluster feature centers, the first parameter template samples are selected in ascending order.

8. An automatic completion system for missing areas in a point cloud of a material, characterized in that, For implementing the automatic completion method for missing regions of a point cloud of a material as described in any one of claims 1-7, the system comprises: The data preprocessing module is used to preprocess the raw point cloud data using voxel lattice downsampling to obtain a denoised point cloud; The boundary detection module is used to perform boundary detection of missing regions on the denoised point cloud by fusing RGB images, and obtain multiple closed contours of missing regions: The joint feature construction module is used to extract the boundary geometric attributes, local environment attributes, and missing cause attributes of the closed contours of the multiple missing regions, and construct multiple joint geometric environment features; The missing region classification module is used to classify missing regions based on the joint features of the multiple geometric environments, and obtain multiple clusters of missing regions of the same type. The similarity measurement module is used to perform feature space similarity measurement on the multiple clusters of missing regions of the same type in order to filter multiple parameter template samples; The parameter completion optimization module is used to complete and optimize the parameters of the multiple parameter template samples to obtain multiple parameter template libraries; The point cloud model output module is used to adaptively adjust the completion parameters in the multiple parameter template libraries according to the joint features of the geometric environment of the missing region, perform completion operation on the closed contour of the remaining missing region in the multiple missing region clusters of the same type, and output the completed point cloud model.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the automatic filling method for missing areas of a material point cloud according to any one of claims 1 to 7.

Images