A robust adaptive point cloud internal hole detection method

Abstract

The present application relates to the technical field of point cloud data processing, in particular to a robust adaptive point cloud internal cavity detection method, comprising: defining target point cloud data under multi-view summary; according to the average statistics of the near neighbor of the target point cloud data, the outlier points of the target point cloud are distinguished, and after the outlier points are distinguished, the extended bounding box and the voxel grid space of the voxel mapping are called; the edge region where the extended bounding box is located is laid out with seed points, the candidate cavity region is established by judging the reverse filling of the seed points; the number of voxel grids and the shape factor of the candidate cavity region are used to perform screening processing, the effective candidate cavity is delimited, and the internal missing point cloud path is solved according to the profile intersection point of the effective candidate cavity mapping; according to the data identification of the profile intersection point on the point cloud path, the internal missing area is restored, and the local features of the internal missing area are integrated according to the corresponding point cloud quantity of the internal missing area. The processing efficiency and accuracy of point cloud data are improved.

Description

Technical Field

[0001] This invention relates to the field of point cloud data processing technology, specifically a robust adaptive method for detecting internal holes in point clouds. Background Technology

[0002] With the rapid development and widespread adoption of 3D scanning technology, point cloud data has become a core carrier for representing the 3D geometric information of objects, playing an irreplaceable role in many fields such as industrial manufacturing, construction engineering, cultural relic protection, and digital twins. However, in the actual point cloud acquisition process, due to the influence of various objective factors, the acquired point cloud data is often incomplete, and closed voids are commonly formed inside the 3D point cloud models constructed by camera scanning.

[0003] Specifically, the main reasons include: the limitation of the scanning angle, which makes it impossible to scan areas occluded by objects; self-occlusion of the object's own structure, which makes it impossible to obtain point cloud data for internal or hidden parts; the reflective and light-absorbing characteristics of the object's surface, which affect the signal acquisition of the scanning equipment and cause point cloud defects; in addition, the accuracy limitations of the scanning equipment, environmental interference and other factors may also lead to data integrity issues.

[0004] For example, Chinese Patent Publication No. CN120635087A discloses a catheter detection method based on scanning point clouds, including: a dynamic self-calibrating optical system to acquire the initial point cloud of the measurement tube; segmenting the initial point cloud to form a complete tube shape; distinguishing straight segments and bending segments by the change in inclination angle of adjacent cylindrical axes; aligning bending points using the bending point ICP algorithm and obtaining matrix transformation relationships; correcting bending points based on these relationships and solving for process data; restoring catheter feature data; and finally aligning the reconstructed tube and the model tube using the bending point ICP algorithm to determine the measurement deviation.

[0005] For example, Chinese Patent Publication No. CN120876799A discloses a three-dimensional laser point cloud decluttering method for defect detection of moving steel pipes, which relates to the field of machine vision inspection technology. The method includes: Step 1, fixing three fixed reference marker points at the beginning, end and middle of the axial direction of the moving steel pipe surface, collecting the original point cloud data of the steel pipe surface containing the marker points, separating the point cloud of the fixed reference marker point area, calculating the geometric center coordinates of the point cloud of each marker point area, and outputting the three-dimensional coordinate set of the marker points; Step 2, based on the original point cloud data, dividing it into an original 3D ring point cloud set along the axial direction of the steel pipe; and fitting the axis of the ideal motion trajectory of the steel pipe through spatial linear regression based on the three-dimensional coordinate set.

[0006] Existing technologies use normal vectors and point distance relationships to set seed points, and then perform point cloud calibration at bending points according to the point cloud trend of the seed points; or they use axial center alignment to complete the global point cloud processing. However, these processing methods are biased towards data processing under extrinsic parameter alignment, and easily overlook the internal holes that appear in the point cloud itself during acquisition, thus causing a decrease in point cloud processing accuracy. Summary of the Invention

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a robust adaptive point cloud internal hole detection method, including: S1, fixing marker points on the outer shell of the object, and defining target point cloud data under multi-view summary according to the three-dimensional coordinates and regional morphology of the missing images and viewpoints.

[0008] S2, based on the nearest neighbor average statistics of the target point cloud data, distinguish outliers in the target point cloud, and after distinguishing outliers, retrieve the extended bounding box and voxel grid space of the voxel mapping.

[0009] S3: By placing seed points in the edge region of the expanded bounding box, extending the boundary along the voxel grid from the seed points, and performing reverse filling judgment based on the voxel values ​​of the voxel grid, candidate hole regions are established.

[0010] S4 performs a screening process based on the number of voxel meshes and shape factor of the candidate cavity region, delineates the valid candidate cavities, and solves the missing point cloud path according to the contour intersection of the valid candidate cavity mapping.

[0011] S5 reconstructs the missing internal regions based on the data identifiers of the intersection points of the contours on the point cloud path, and integrates the local features of the missing internal regions based on the number of point clouds corresponding to the missing internal regions.

[0012] The beneficial effects of this invention are as follows: First, this invention summarizes the collected viewpoints based on the number of marker points within a single viewpoint, determines the overlap rate of marker points in adjacent viewpoints, continuously marks the collected viewpoints according to the overlap rate, and constructs a target point cloud by combining the image content. This ensures the global consistency of multi-viewpoint point cloud registration, avoids problems such as misaligned point cloud stitching and broken object outlines caused by the lack of correlation between adjacent viewpoints, and improves the data accuracy of the initial point cloud acquisition.

[0013] II. This invention calculates the average local nearest distance of a target point using a set of local nearest neighbor points, removes outliers based on global statistical features, adaptively determines voxel sizes, constructs axis-aligned bounding boxes, and extends them to form extended bounding boxes. Finally, it completes the 3D mesh division according to voxel sizes to form a voxel mesh space. Simultaneously, when multiple point cloud blocks overlap, point cloud blocks on the same physical plane are grouped according to the scan index. For bounding boxes with non-empty intersections within the groups, adjustments are made based on the average local nearest distance of points in the overlapping area, outputting the final axis-aligned bounding box. This avoids the problems of excessive computation due to overly dense voxels and false detections due to contour distortion caused by overly sparse voxels; while ensuring detection accuracy, it improves detection efficiency and adaptability to different point cloud data.

[0014] Third, this invention uses a voxel mesh space as a guide, selects the corner points of the extended bounding box as initial seed points, and performs clustering of voxels with the same label based on the voxel value difference between the seed point and its neighboring voxels. It then calibrates the boundaries of the filled voxels and extracts regions that meet the closure condition as candidate void regions. This avoids the problem of incorrect filling space labeling caused by seed points falling inside the object; further ensuring the authenticity of the candidate void region distribution and the target range.

[0015] IV. This invention identifies valid candidate holes by using the number of voxel grids and shape factor as screening indicators. It then performs multi-dimensional planar mapping on these valid candidate holes, determines the existence of contour intersections, and generates point cloud paths for different scenes: when there are contour intersections co-occurring on at least two planes, the paths are connected using these intersections as nodes; when there are no contour intersections, the paths are formed using the hole center point as a node. This clarifies the clustering phenomenon under the distribution of hole regions, transforming hole detection from a three-dimensional validity process to a two-dimensional discrete clustering phenomenon, providing a data foundation for the local features of subsequently missing internal regions. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] Figure 1 This is a flowchart illustrating a robust adaptive point cloud internal hole detection method.

[0018] Figure 2 This is a flowchart illustrating step S2 of a robust adaptive point cloud internal hole detection method.

[0019] Figure 3 This is a flowchart illustrating step S3 of a robust adaptive point cloud internal hole detection method.

[0020] Figure 4 This is a flowchart illustrating step S4 of a robust adaptive point cloud internal hole detection method. Detailed Implementation

[0021] The embodiments of the present invention are described in detail below. The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Where specific techniques or conditions are not specified in the embodiments, they shall be performed in accordance with the techniques or conditions described in the literature in the art or in accordance with the product manual.

[0022] See Figure 1 A robust adaptive point cloud internal hole detection method includes: S1, fixing marker points on the object shell, and defining target point cloud data under multi-view summary based on the images and viewpoints acquired by the markers and the missing three-dimensional coordinates and regional morphology of the images.

[0023] S2, based on the nearest neighbor average statistics of the target point cloud data, distinguish outliers in the target point cloud, and after distinguishing outliers, retrieve the extended bounding box and voxel grid space of the voxel mapping.

[0024] S3: By placing seed points in the edge region of the expanded bounding box, extending the boundary along the voxel grid from the seed points, and performing reverse filling judgment based on the voxel values ​​of the voxel grid, candidate hole regions are established.

[0025] S4 performs a screening process based on the number of voxel meshes and shape factor of the candidate cavity region, delineates the valid candidate cavities, and solves the missing point cloud path according to the contour intersection of the valid candidate cavity mapping.

[0026] S5 reconstructs the missing internal regions based on the data identifiers of the intersection points of the contours on the point cloud path, and integrates the local features of the missing internal regions based on the number of point clouds corresponding to the missing internal regions.

[0027] The current solution detects internal voids by identifying internal missing areas caused by surface discontinuities leading to filling leakage, scanning view limitations, and equipment errors. For these areas, bounding boxes are used to extend inwards, and three-dimensional voxels, multi-boundary selection, and 6-neighbor connectivity are simultaneously applied to suppress the oblique processing of internal void extension. This process determines the internal missing areas within the space, ultimately forming a dataset containing the potential causes of voids (such as scanning blind spots, material reflection, and structural self-occlusion) and the distribution of voids.

[0028] In one embodiment of the present invention, step S1 is used to fix a mark on the outer shell of the object, and determine the currently occurring internal missing area by using the image and viewpoint acquired by the mark, and define the target point cloud data for the current analysis according to the three-dimensional coordinates and region shape of the missing image.

[0029] Specifically, the markers are permanently fixed to the rigid, non-deformable area of ​​the object's outer shell using adhesive / clamps, serving as anchor points for global coordinates.

[0030] One implementation of step S1 includes: S11, summarizing the data of each view based on the number of marker points contained in each acquisition view, and determining the overlap rate of marker points between adjacent acquisition views; specifically, firstly, counting the total number of marker points detected in each view, and establishing a mapping relationship between the view and the set of marker points within the view; for two spatially adjacent views, finding the set of marker points they share through marker point matching, and then determining the value of the overlap rate.

[0031] S12, adjacent acquisition viewpoints are continuously marked according to the overlap rate of the marker points, and the target point cloud data is constructed from the real-time acquired image content.

[0032] Furthermore, a threshold for the overlap rate of marker points is set. When the overlap rate exceeds this threshold, the viewpoint is considered continuous, and the image content acquired from the corresponding viewpoint is used as the data basis for the output of the target point cloud.

[0033] Furthermore, the overlap rate threshold for marker points is set at 30% to ensure that continuous point cloud data have a corresponding continuous basis when stitching together.

[0034] In one embodiment of the present invention, a nearest neighbor search is performed on the target point cloud data, and data with outliers are filtered out by the local nearest neighbor average distance of the nearest neighbor points. Then, a three-dimensional voxel mesh is constructed to provide a bounding box basis for subsequent hole detection inside the point cloud.

[0035] Specifically, such as Figure 2 As shown, one implementation of step S2 includes: S21, for each target point in the target point cloud data, obtaining the local nearest neighbor set of the target point, and setting multiple local nearest average distances according to the average Euclidean distance between each point in the local nearest neighbor set.

[0036] S22, calculate the average and standard deviation of the average distances of all local nearest neighbors, remove outliers that meet the outlier determination threshold, and then determine the voxel size of the target point cloud data.

[0037] Specifically, the nearest neighbor average statistics are implemented as follows: For each target point in the target point cloud data, the k-nearest neighbor search algorithm is used, where k is 5-10, to search for the k nearest neighbors of the target point in the three-dimensional space, forming the local nearest neighbor set of the target point.

[0038] Secondly, calculate the Euclidean distance from the target point to each of its k nearest neighbors, sum all the Euclidean distances and take the average to obtain the local nearest neighbor average distance of the target point. This distance can reflect the point cloud density distribution around the target point.

[0039] Then, the local nearest neighbor average distances of all point clouds are traversed, and the average and standard deviation of all local nearest neighbor average distances are calculated. An outlier detection threshold is set, with a value ranging from 1.0 to 2.0, which can be adaptively adjusted according to the noise content of the original point cloud (a value close to 2.0 is used when the noise content is high, and a value close to 1.0 is used when the noise content is low). For each target point, if its local nearest neighbor average distance is greater than the average value + outlier detection threshold × standard deviation, the point is determined to be an outlier noise point. All point clouds are traversed, and all coordinate points determined to be outlier noise points are removed in batches. The remaining point cloud data is the denoised point cloud data. This data effectively eliminates invalid noise points generated during the scanning process due to environmental interference, equipment errors, etc., ensuring the accuracy of subsequent detection steps.

[0040] Furthermore, the voxel size is implemented by: obtaining the global average spacing corresponding to the local nearest neighbor average distance for the target point cloud data after outlier filtering; and determining the voxel size based on the product of the global average spacing and the scaling factor.

[0041] Specifically, the k-nearest neighbor algorithm is still used to set the voxel size for the point cloud data. The mean of the average distances of all local nearest neighbors is obtained as the global average spacing, or the median is calculated as the global average spacing; the scaling factor k ranges from 1.5 to 2.5. After setting the voxel size for each voxel grid, the target point cloud is arranged along the X, Y, and Z axes using voxel grids of equal size, forming point cloud data contained within the voxel grid.

[0042] This method ensures that the voxel size can accurately capture the details of the point cloud surface, allowing the voxel mesh to enclose and expand the point cloud data, thereby improving the processing accuracy of subsequent inverse filling judgment.

[0043] S23, construct an axis-aligned bounding box that surrounds all point clouds, and extend the voxel size according to the corresponding dimension of the axis-aligned bounding box to form an extended bounding box.

[0044] Furthermore, based on the size of the voxel mapping and the point cloud data, the current bounding box setting is determined according to the distribution of the point cloud data.

[0045] For point cloud data that has completed voxel mapping, it is used as the data basis for constructing bounding boxes. All point cloud coordinates are traversed to obtain the minimum and maximum values ​​of the point cloud in the X, Y, and Z dimensions. An axis-aligned bounding box is constructed to enclose the entire point cloud to determine the spatial distribution range of the point cloud.

[0046] Specifically, when constructing the axis-aligned bounding box that surrounds all point clouds, it is also necessary to determine the overlapping position of multiple point clouds in the voxel mesh space to complete the setting process of the axis-aligned bounding box.

[0047] Furthermore, the implementation of the axis-aligned bounding box also includes: calculating the initial axis-aligned bounding box for each point cloud based on the location and spatial topology of the target point cloud data; the initial axis-aligned bounding box forms a preliminary bounding box containing the point cloud data based on the minimum and maximum values ​​in the three dimensions of X, Y, and Z; this bounding box can completely contain the point cloud data. If the current input target point cloud consists of multiple point clouds, it is necessary to determine the overlap of the multiple point clouds to avoid subsequent boundary extension to the location of the point cloud data.

[0048] Based on the index information during point cloud scanning, point cloud blocks belonging to the same physical plane are grouped. This index information represents the data index for point cloud data extraction from the object surface, indicating the scanning order and relative position. The point cloud blocks are subsets of point cloud data collected from different scanning perspectives, corresponding to the same physical plane, and are used to verify the internal voids in the current point cloud data acquisition.

[0049] For point cloud blocks within the same group, if the intersection of the corresponding initial axis-aligned bounding boxes is not empty, the local nearest neighbor set is used to extract the average local nearest neighbor distance of all points in the overlapping area, and the point cloud block with the smallest average local nearest neighbor distance is selected as the main body, and the initial axis-aligned bounding box in the current group is adjusted.

[0050] Specifically, the adjustment method here includes deduplicating other point cloud block data within the duplicate region, and taking the point cloud block with the smallest local nearest neighbor average distance at that location as the main body, and resetting the axis alignment bounding box for each point cloud block according to the deduplicated point cloud, until there are no more duplicate regions.

[0051] Use the adjusted initial axis-aligned bounding box as the output axis-aligned bounding box.

[0052] If the intersection of the corresponding initial axis-aligned bounding boxes is empty, then the currently generated initial axis-aligned bounding box is output as the axis-aligned bounding box constructed here.

[0053] Furthermore, the extended bounding box is constructed by spatially expanding the aforementioned axis-aligned bounding box. The smallest corner point of the bounding box is expanded by at least 3 voxels along each of the three negative directions of the X, Y, and Z axes, and the largest corner point is expanded by at least 3 voxels along each of the three positive directions of the X, Y, and Z axes, ultimately forming the extended bounding box. The extended bounding box does not perform any data operations; it only serves as a boundary expansion of the axis-aligned bounding box. Its edge region does not overlap with the point cloud, ensuring stable selection of external seed points and preventing detection failure due to seed points being covered by the point cloud.

[0054] S24. Using the obtained voxel size as the basic unit, perform three-dimensional meshing on the location of the extended bounding box, uniformly dividing it into several voxel meshes to form a voxel mesh space.

[0055] In the voxel mesh space, all voxels are initially marked as -1, which represents an unknown region that is not occupied or filled by the point cloud. Then, all point cloud data is traversed, and voxels containing point cloud data are marked as 0. Subsequently, a morphological dilation operation is performed on all voxels marked as 0, and the voxels that are adjacent to the voxels marked as 0 and marked as -1 are marked as 0.5. These voxels marked as 0.5 are marked as the region form of surface expansion.

[0056] Furthermore, all voxel markers within the voxel grid interval need to undergo adaptive dilation to form a continuously labeled voxel grid space, and the expanded bounding box under the markers is adjusted accordingly.

[0057] Specifically, the operation process is an expansion operation, which is used to fill the microscopic gaps on the current point cloud surface to enhance the continuity of the point cloud surface, thereby improving the accuracy and relative effect of subsequent seed point placement; the specific steps are as follows: 1. Use a 3×3×3 structuring element, with the target voxel being the center of the structuring element, and only containing 6 neighboring voxels in the top, bottom, left, right, front, and back directions, to avoid misoperation caused by oblique expansion.

[0058] 2. Traverse all target voxels marked as 0 in the voxel grid. Using each target voxel as the center, search for its 6 neighboring voxels through a 3×3×3 structuring element. If a neighboring voxel is marked as -1, then correct the mark of the neighboring voxel from -1 to 0.5. If a neighboring voxel is 0, then keep the mark unchanged.

[0059] 3. By statistically analyzing whether the 0.5 voxel shows a continuous boundary, if there are still voxels with voxel markings of -1 in the six directions of the target voxel, repeat the above steps until the data under the 0.5 voxel marking shows a continuous boundary.

[0060] Furthermore, the implementation of the extended bounding box also includes setting a voxel marker for each voxel mesh corresponding to the extended bounding box.

[0061] Using data with the same voxel label as the statistical basis, a 6-neighborhood directional dilation process is performed to record the voxel mesh updated with the voxel label.

[0062] Using the updated voxel mesh as the basis for judgment, the edge region of the voxel mesh is delineated. Based on the update of the edge region in the voxel mesh space each time, the number of expanded voxels in the expanded bounding box is determined. Here, the edge region refers to the area still marked as -1, representing the boundary after the axis-aligned bounding box is expanded; the number of expanded voxels represents the number of voxel meshes located in the edge region after the voxel markers marked as 0.5 form a complete continuous line segment.

[0063] Specifically, when voxel markers with a value of 0.5 form a continuous boundary and there are no voxel markers with a value of -1 in the 6-neighborhood of the target voxel, the dilation process is stopped, and the number of voxels newly added during the dilation process is counted as the number of expanded voxels in the expanded bounding box. This number of expanded voxels will serve as an auxiliary parameter for subsequent 6-neighborhood backfilling to help determine the location of internal voids.

[0064] Based on a determined number of expanded voxels, the final expanded bounding box is generated. The final expanded bounding box includes voxel markers, edge regions, and other components, which serve as the data for subsequent padding processing.

[0065] In one embodiment of the present invention, based on the 6-neighbor connectivity method, connectivity analysis is performed on the region filled by the seed point in reverse, and only closed regions that satisfy 6-neighbor connectivity are retained as candidate holes.

[0066] Specifically, such as Figure 3 As shown, one implementation of step S3 includes: S31, using the current voxel mesh space as a guide, selecting the corner points of the extended bounding box as initial seed points, and performing reverse filling processing on the extended bounding box. Specifically, a 6-neighborhood processing method is used, starting from the 8 corner points of the extended bounding box, to extend the voxel mesh boundary and determine the voxel boundaries containing both closed and non-closed boundaries.

[0067] Specifically, starting from the selected initial seed point, all reachable voxel meshes marked as -1 are marked as 1, where 1 represents the external space region. After filling, the regions still marked as -1 are subjected to closure verification: if no voxel mesh can be connected to the external space region marked as 1 through 6 neighborhoods, and all boundaries of the cluster are completely surrounded by voxels marked as 0 or 0.5, then the region satisfies the closure condition and is considered a potential candidate hole; if the closure condition is not met, then the region does not belong to the internal hole, is marked as 1, and is excluded from the candidate holes.

[0068] S32, based on the difference in voxel values ​​between the initial seed point and neighboring voxels, performs clustering processing on the voxel grid corresponding to the same voxel value, and marks the voxel boundary after the initial seed point is filled.

[0069] Specifically, when performing clustering processing on voxel grids corresponding to the same voxel value, the implementation method includes: for the voxel value of the neighboring voxel, the voxel value is the voxel label set above; the initial seed point and the neighboring voxels are connected to determine the preliminary connected region corresponding to the initial seed point; the preliminary connected region represents all voxel grids that the initial seed point can reach and are labeled as -1. These regions will be used as the external space region under the preliminary connection, and their labels will be adjusted to 1 as the boundary to be distinguished first.

[0070] It should be noted that the current scheme uses 6-neighbor connectivity to determine all possible boundaries by removing diagonal connections.

[0071] Furthermore, the 6-neighbor closed boundary represents full blockage in multiple directions to enable rapid detection of internal cavities, emphasizing the size and shape of the closed region.

[0072] Regions outside the initially connected regions are selected as the main clustering subjects. Voxel values ​​are used as clustering values ​​to delineate multiple clusters after clustering. Each cluster represents a set of voxel grids labeled -1, 0, or 0.5, and the region boundaries are formed through 6-neighborhood connectivity clustering. Based on the formed region boundaries, they are used as the main subjects for subsequent closure condition verification to determine the connectivity between the internal void regions labeled -1 and those labeled 1, as well as the closure status of the grids labeled 0 and 0.5.

[0073] For each cluster-related voxel grid, a closure check is performed, and data that meets the closure check are considered as the output voxel boundaries.

[0074] In the context of closure verification, for a cluster labeled -1, if no voxel mesh can be connected to the external spatial region labeled 1 through a 6-neighborhood, and is surrounded by meshes labeled 0 or 0.5, this region is considered a closed region and is included as a candidate hole region in the output. If it can extend to meshes labeled 0 or 0.5, or is connected to a voxel mesh labeled 1, it is marked as not closed and is removed from the candidate holes.

[0075] S33 extracts data that meets the closure condition from the voxel boundary as candidate void regions for output.

[0076] In the current processing scenario, candidate void regions are used to explain the parts that do not contain point cloud data in multiple data fillings. These data will be used as internally identified missing regions.

[0077] In one embodiment of the present invention, data screening is performed on candidate void regions to clarify the size and shape of the candidate void regions, transform the missing situation into the intersection relationship of two-dimensional projection, and illustrate the missing accumulation phenomenon of adjacent regions in non-oblique cases through the point cloud path of contour mapping, thereby characterizing the feature form of internal missingness in local three-dimensional positions.

[0078] Specifically, three two-dimensional planes, XY, YZ, and XZ, are selected as mapping planes. The current voxel mesh is mapped to the corresponding planes. The intersection points of the contour lines are extracted according to the two-dimensional projection contours at each plane. When the contour intersection points appear in at least two two-dimensional planes, the corresponding contour intersection points are used as path nodes of the point cloud path. Finally, the point cloud path corresponding to each candidate hole region is generated.

[0079] like Figure 4 As shown, one implementation of step S4 includes: S41, using the number of voxel grids and shape factor corresponding to the candidate void region as screening indicators to delineate valid candidate voids.

[0080] Among them, when the voxel size is known, the volume of each candidate void region can be directly calculated based on the number of voxel meshes; secondly, the shape factor includes parameters such as the slenderness, flatness, sphericity and volume ratio relative to the bounding box corresponding to the candidate void region, and eliminates non-target candidate void regions such as slits and slender channels.

[0081] Furthermore, for slenderness and flatness, principal component analysis is required for the candidate void regions to calculate three eigenvalues, in descending order of magnitude: , and All three values ​​are positive real numbers.

[0082] The slender length is represented as A higher value indicates a finer object; flatness is expressed as... The closer this value is to 1, the more flattened it is; as for sphericity, it is expressed as... ;in, Represents pi (π). Indicates the volume of the candidate void region; The surface area of ​​the candidate cavity region can be estimated by the number of voxel meshes at the 3D contour. This value is used to eliminate cavities with too small a sphericity to filter candidate cavity regions of specific shapes and sizes.

[0083] Furthermore, when clustering candidate hole regions, the shape factor adjusts its parameter composition according to the shapes filtered out in the current scene, and performs a filtering operation according to each parameter value corresponding to the shape factor and the number of voxel grids. The currently obtained candidate hole regions are compared with historical data, and candidate hole regions that all meet the quartiles of historical data (values ​​at the 75th percentile after sorting) are regarded as the current valid candidate holes.

[0084] Furthermore, since relatively large values ​​need to be removed for slenderness and flatness, the data will be filtered based on their descending quartiles, selecting data smaller than these quartiles to eliminate long, flat regions of slits and narrow channels. As for sphericity and the volume ratio relative to the bounding box, relatively large values ​​need to be obtained, so ascending quartiles need to be selected, and the data larger than these quartiles will be chosen to complete the data filtering.

[0085] S42, perform two-dimensional plane mapping on the valid candidate voids to determine whether there are contour intersections of the valid candidate voids on multiple two-dimensional planes; where each point on the contour corresponds to a voxel mesh, thereby determining the internal void aggregation situation in the case of three-dimensional to two-dimensional conversion.

[0086] S43, if contour intersections exist, select contour intersections that appear in at least two planes, and connect them sequentially according to the spatial coordinate order of the contour intersections to form a point cloud path; when contour intersections exist, it indicates that a specific two-dimensional plane accumulation has occurred, and its distribution will be concentrated at a specific location. When the point cloud path can close the path, the point cloud path will further emphasize the concentration effect of internal missing parts, and data marking will be performed on each valid candidate hole according to the current point cloud path connection form.

[0087] S44. If there is no contour intersection, the position mapped by the center point of each valid candidate hole is taken as the path node to form a point cloud path. If there is no contour intersection, it means that the missing holes inside are distributed discretely and there is no accumulation phenomenon in a specific two-dimensional plane. In this case, they need to be connected in order from the inside to the outside according to the Euclidean distance between their center points to explain the relative direction of the missing holes.

[0088] It should be noted that when setting the connection path, start from any marked path node and extend in the up, down, left and right directions of the two-dimensional plane. Record the connection branch every time you pass through a path node until all path nodes are connected. And do not connect to the path nodes that have already been recorded, except for the starting point.

[0089] In one embodiment of the present invention, the local features of the effective candidate holes in each identification are determined based on the changes of the point cloud path in multiple data identifications, and the corresponding region is taken as the internal missing region according to the identifier of the point cloud path. The local features are integrated with the contents such as the number of point clouds, voxel value, and path identifier.

[0090] One implementation of step S5 includes: S51, when the data identifier of the point cloud path changes, the candidate hole region corresponding to the point cloud path is regarded as an internal missing region.

[0091] This data identifier represents the existence of the contour intersection of valid candidate holes and the path closure identifier. If this data identifier changes after multiple processing, it indicates that the density of the corner / thin wall / deep cavity region is uneven, or that the workpiece has reflective or oily content, causing the point cloud alignment to shift and the voxel filling to deviate, thus reducing the recognition accuracy of the missing parts inside.

[0092] If the data identifiers of the point cloud data do not change after multiple processing steps, the corresponding candidate hole regions will be directly output and the relevant local features will be directly integrated.

[0093] S52 performs secondary verification on each internal missing region, and uses the data corresponding to the voxel grid and candidate hole regions after secondary verification as the output local features.

[0094] The secondary verification includes: regenerating the internal missing region and recording the data such as the grid, candidate holes, and labeled nearest neighbor set under voxel mapping of the point cloud data in two consecutive generations. These data are used as two sets of voxel grid mapping data sets, and the parts corresponding to the internal missing region in these two sets of data are used as the local features of the current output. In other words, multiple sets of intermediate data such as voxel grid, candidate hole region, and labeled nearest neighbor set corresponding to the internal missing region after multiple iterations are selected as the local features of the output to complete the relative processing and summarization of the corresponding data, so as to explain the relevant parameters that have reduced recognition accuracy, thereby improving the processing efficiency and backtracking effect for internal holes in the point cloud.

[0095] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention, which are still covered within the protection scope of the present invention.

Claims

1. A robust adaptive method for detecting internal holes in point clouds, characterized in that, include: S1, mark points are fixed on the outer shell of the object. Based on the images and viewpoints collected by the marks, the target point cloud data is defined under the sum of multiple viewpoints according to the missing three-dimensional coordinates and regional morphology of the images. The methods for implementing target point cloud data include: S11, count the total number of marker points detected in each viewpoint, establish the mapping relationship between the viewpoint and the set of marker points within the viewpoint, and for two spatially adjacent viewpoints, find the set of marker points they share, and obtain the marker point overlap rate of adjacent viewpoints. S12, set the threshold for the overlap rate of marker points. When the overlap rate is greater than the threshold, the view is considered continuous. Adjacent view points are continuously marked according to the overlap rate of marker points to obtain the output target point cloud data. S2, based on the nearest neighbor average statistics of the target point cloud data, distinguish outliers in the target point cloud, and after distinguishing outliers, retrieve the extended bounding box and voxel grid space of the voxel mapping. S3, by placing seed points in the edge region of the expanded bounding box, extending the boundary along the voxel mesh starting from the seed points, and performing reverse filling judgment based on the voxel values ​​of the voxel mesh, candidate hole regions are established. The methods for implementing candidate void regions include: S31, guided by the current voxel mesh space, select the corner points of the expanded bounding box as the initial seed points; S32, based on the difference in voxel values ​​between the initial seed point and the neighboring voxels, cluster the voxel grids corresponding to the same voxel value and mark the voxel boundaries after the initial seed point is filled. S33, extract data that meets the closure condition from the voxel boundary as candidate void regions for output; S4. Perform a screening process based on the number of voxel meshes and shape factor of the candidate cavity region to delineate the effective candidate cavities, and solve the missing point cloud path inside according to the contour intersection of the effective candidate cavity mapping. Point cloud paths can be implemented in the following ways: S41, using the number of voxel grids and shape factor corresponding to the candidate cavity region as screening indicators, delineate the effective candidate cavities; S42, Perform two-dimensional plane mapping on the valid candidate holes to determine whether there are contour intersections of the valid candidate holes on multiple two-dimensional planes; S43, If contour intersections exist, select contour intersections that appear at least in two planes and connect them to form a point cloud path; S44, if there are no contour intersections, then take the position mapped by the center point of each valid candidate hole as a path node to form a point cloud path; S5 reconstructs the missing internal regions based on the data identifiers of the intersection points of the contours on the point cloud path, and integrates the local features of the missing internal regions based on the number of point clouds corresponding to the missing internal regions.

2. The robust adaptive point cloud internal hole detection method according to claim 1, characterized in that, The implementation methods for expanding the bounding box and voxel mesh space in step S2 include: S21, for each target point in the target point cloud data, obtain the local nearest neighbor set of the target point, and set multiple local nearest average distances according to the average Euclidean distance between each point in the local nearest neighbor set; S22, calculate the average and standard deviation of the average distances of all local nearest neighbors, remove outliers that meet the outlier determination threshold, and then determine the voxel size of the target point cloud data. S23, construct an axis-aligned bounding box that encloses all point clouds, and extend the voxel size according to the corresponding dimension of the axis-aligned bounding box to form an extended bounding box; S24. Using the obtained voxel size as the basic unit, perform three-dimensional meshing on the location of the extended bounding box, uniformly dividing it into several voxel meshes to form a voxel mesh space.

3. The robust adaptive point cloud internal hole detection method according to claim 2, characterized in that, The implementation methods for voxel size include: For the target point cloud data after outlier filtering, obtain the global average spacing corresponding to the local nearest neighbor average distance; determine the voxel size based on the product of the global average spacing and the scaling factor.

4. A robust adaptive point cloud internal hole detection method according to claim 2, characterized in that, Other methods for implementing axis-aligned bounding boxes include: Based on the location and spatial topology of the target point cloud data, calculate the initial axis-aligned bounding box for each point cloud block; Based on the index information during point cloud scanning, point cloud blocks belonging to the same physical plane are grouped. For point cloud blocks within the same group, if the intersection of the corresponding initial axis-aligned bounding boxes is not empty, the local nearest neighbor set is used to extract the average local nearest neighbor distance of all points in the overlapping area, and the point cloud block with the smallest average local nearest neighbor distance is selected as the main body, and the initial axis-aligned bounding box in the current group is adjusted. Use the adjusted initial axis-aligned bounding box as the output axis-aligned bounding box.

5. A robust adaptive point cloud internal hole detection method according to claim 2, characterized in that, Other ways to implement an extended bounding box include: For each voxel mesh corresponding to the expanded bounding box, set a voxel tag for each voxel mesh; Using data with the same voxel label as the statistical basis, a 6-neighborhood directional dilation process is performed, and the voxel mesh updated with the voxel label is recorded. Using the updated voxel mesh as the basis for judgment, the edge region of the voxel mesh is delineated, and the number of extended voxels of the extended bounding box is determined according to the update of the edge region in the voxel mesh space each time. Based on a determined number of extended voxels, the final output extended bounding box is generated.

6. A robust adaptive point cloud internal hole detection method according to claim 1, characterized in that, When performing clustering on voxel meshes corresponding to the same voxel value, the implementation methods include: For the voxel values ​​of neighboring voxels, the initial seed point and neighboring voxels are connected to determine the preliminary connected region corresponding to the initial seed point. Other regions outside the initial connected regions are selected as the main clustering subjects, and voxel values ​​are used as clustering values ​​to delineate multiple clusters after clustering; For each cluster-related voxel grid, a closure check is performed, and data that meets the closure check are considered as the output voxel boundaries.

7. A robust adaptive point cloud internal hole detection method according to claim 1, characterized in that, The implementation methods for local features in step S5 include: S51, when the data identifier of the point cloud path changes, the candidate hole area corresponding to the point cloud path is regarded as an internal missing area; S52 performs secondary verification on each internal missing region, and uses the data corresponding to the voxel grid and candidate hole regions after secondary verification as the output local features.

Images