Point cloud completion method and system based on proxy point and neighborhood feature aggregation

Abstract

The application relates to a point cloud completion method and system based on an agent point and neighborhood feature aggregation, wherein the point cloud completion method comprises the following steps: determining a target missing area of to-be-completed point cloud data and generating an agent point; in the original point cloud data, neighborhood searching is performed on each agent point to obtain a first neighborhood point set; enhanced features are generated based on the geometric features of each neighborhood point, a completed point corresponding to each agent point and a completed point position are determined; and the completed point is inserted into the target missing area to complete the target missing area. Through the application, the problems that in the related art, point cloud completion is difficult to accurately restore geometric details and maintain boundary continuity in a complex scene, and is strongly dependent on data and a scene and has insufficient generalization capability are solved, and the beneficial effects that the sparse and incomplete point cloud in the complex scene is adaptively completed with high fidelity and high consistency without relying on large-scale labeled training data are achieved.

Description

Technical Field

[0001] This application relates to the field of 3D point cloud processing technology, and in particular to a point cloud completion method and system based on surrogate point and neighborhood feature aggregation. Background Technology

[0002] Point clouds, as an important data form representing the spatial geometry of objects, can directly reflect the structural information of three-dimensional space compared to traditional two-dimensional images. Therefore, they have irreplaceable value in applications such as intelligent transportation, industrial inspection, and spatial mapping. With the continuous advancement of sensors such as LiDAR, structured light, and depth cameras, the application status of point cloud technology in industrial and engineering fields is becoming increasingly prominent. However, due to factors such as complex acquisition environments, limited sensor resolution, differences in target surface reflectivity, and occlusion at observation angles, the actual acquired point cloud data often suffers from sparseness, voids, or local missing data. For example, in bridge inspection tasks, steel beams and cable structures may obstruct the scanning path, resulting in missing data in the back area; in open-air material stockpile measurement, the point cloud in the edge area is often undersampled due to the influence of lighting shadows and the slope angle of the stockpile. These defects not only affect the complete representation of target boundaries and details but also directly weaken the reliability of subsequent modeling, recognition, and precision measurement. Traditional geometric reconstruction methods, such as those based on triangulation or Poisson reconstruction, can fill in missing areas to some extent, but they often generate pseudo-points in the edge regions, resulting in the reconstructed surface not matching the real structure and insufficient preservation of local geometric details, making it difficult to meet the high-precision requirements of complex scenes.

[0003] With the rise of artificial intelligence, point cloud completion methods based on deep learning have gradually attracted widespread attention. These methods typically learn the patterns of missing points in a point cloud through neural networks, and can recover complex and empty areas to a certain extent, especially in standardized datasets such as furniture completion and human body modeling. However, in practical engineering, these methods still have obvious limitations: (1) they are heavily dependent on large-scale and diverse labeled samples, while high-quality training data is often difficult to obtain in industrial scenarios; (2) the model has insufficient generalization ability when applied across scenarios. At the same time, although traditional methods based on neighborhood interpolation are simple to implement and have low computational overhead, they often generate sparse and uneven points in large missing areas, especially at the boundary positions, where they are prone to blurring and distortion, and lack geometric consistency.

[0004] In summary, existing point cloud completion techniques suffer from several drawbacks, including strong dependence on training data, limited cross-scene generalization ability, and sparse and distorted reconstruction results in large missing areas, making it difficult to maintain geometric details and structural consistency. To overcome these shortcomings, a new solution is urgently needed that can achieve high-fidelity and high-consistency completion of sparse and missing point clouds through adaptive geometric reasoning and local context awareness, thereby improving the robustness and accuracy of reconstruction in complex engineering scenarios. Summary of the Invention

[0005] This application provides a point cloud completion method and system based on the aggregation of surrogate points and neighborhood features, which at least solves the problems in related technologies where point cloud completion is difficult to accurately restore geometric details and maintain boundary continuity in complex scenarios, and has strong dependence on data and scenarios and insufficient generalization ability.

[0006] In a first aspect, embodiments of this application provide a point cloud completion method based on surrogate points and neighborhood feature aggregation, comprising: determining a target missing region of point cloud data to be completed, and generating surrogate points within the target missing region, wherein the surrogate points are used to characterize reference points for completion; performing a neighborhood search on each surrogate point in the original point cloud data corresponding to the point cloud data to obtain a first neighborhood point set corresponding to each surrogate point; aggregating and generating enhanced features for each surrogate point based on the geometric features of each neighborhood point in the first neighborhood point set, determining a completion point and its position corresponding to each surrogate point according to the enhanced features; and inserting the completion point into the target missing region according to the position of the completion point to complete the target missing region.

[0007] Secondly, embodiments of this application provide a point cloud completion system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the point cloud completion method based on surrogate point and neighborhood feature aggregation as described in the first aspect.

[0008] Compared to related technologies, the point cloud completion method and system based on surrogate points and neighborhood feature aggregation provided in this application embodiment determines the target missing region of the point cloud data to be completed and generates surrogate points within the target missing region, wherein the surrogate points are used to represent the reference points for completion; in the original point cloud data corresponding to the point cloud data, a neighborhood search is performed on each surrogate point to obtain a first neighborhood point set corresponding to each surrogate point; based on the geometric features of each neighborhood point in the first neighborhood point set, enhanced features of each surrogate point are aggregated to generate enhanced features; and based on the enhanced features, the features of each surrogate point are determined to be related to the reference points for completion. The method identifies the corresponding completion points and their positions. Based on the completion point positions, the completion points are inserted into the target missing region to complete the target missing region. This method solves the problems in related technologies where point cloud completion is difficult to accurately recover geometric details and maintain boundary continuity in complex scenes, and is highly dependent on data and scene with insufficient generalization ability. It achieves the beneficial effect of high-fidelity and high-consistency adaptive completion of sparse and incomplete point clouds in complex scenes without relying on large-scale labeled training data, effectively improving the reliability and practicality of point clouds in subsequent modeling, recognition, and precision measurement tasks.

[0009] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a hardware structure block diagram of a terminal for a point cloud completion method based on agent point and neighborhood feature aggregation according to an embodiment of this application. Figure 2 This is a flowchart of a point cloud completion method based on the aggregation of surrogate points and neighborhood features according to an embodiment of this application; Figure 3 This is a flowchart of a point cloud completion method based on the aggregation of surrogate points and neighborhood features according to a preferred embodiment of this application; Figure 4 This is a schematic diagram of the defective point cloud before point cloud completion according to a preferred embodiment of this application; Figure 5 This is a schematic diagram illustrating the determination of the target missing region according to a preferred embodiment of this application. Figure 6 This is a schematic diagram of missing boundary point detection and target missing region construction according to a preferred embodiment of this application; Figure 7 This is a schematic diagram of the agent point generation process according to a preferred embodiment of this application; Figure 8 This is a schematic diagram of a neighborhood search method according to a preferred embodiment of this application; Figure 9 This is a schematic diagram of neighborhood feature weighted aggregation according to a preferred embodiment of this application; Figure 10 This is a schematic diagram of the complete point cloud after point cloud completion according to a preferred embodiment of this application; Figure 11 This is a structural block diagram of a point cloud completion device based on the aggregation of surrogate points and neighborhood features according to an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application. Furthermore, it is understood that although the efforts made in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, modifications to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0012] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0013] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application means two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The terms “first,” “second,” “third,” etc., used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0014] The method embodiments provided in this example can be executed on a terminal, computer, or similar computing device. Taking running on a terminal as an example, Figure 1 This is a hardware structure block diagram of the terminal for the point cloud completion method based on the aggregation of surrogate points and neighborhood features according to an embodiment of this application. Figure 1 As shown, terminal 10 may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. Optionally, the terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the terminal described above. For example, terminal 10 may also include components that are larger than those described above. Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0015] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the point cloud completion method based on surrogate point and neighborhood feature aggregation in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0016] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0017] This embodiment provides a point cloud completion method based on the aggregation of surrogate points and neighborhood features, which runs on the aforementioned terminal. Figure 2 This is a flowchart of a point cloud completion method based on surrogate point and neighborhood feature aggregation according to an embodiment of this application, as follows: Figure 2 As shown, the process includes the following steps: Step S201: Determine the target missing region of the point cloud data to be completed, and generate surrogate points within the target missing region, wherein the surrogate points are used to represent the reference points for completion.

[0018] In this embodiment, by identifying the target missing region of the point cloud data to be completed and generating surrogate points within that region, the accuracy and boundary continuity of the restoration are improved. Since point cloud acquisition relies on sensors such as LiDAR, depth cameras, or structured light, these sensors are often affected by various uncertainties in real-world applications, such as limited viewing angles, occlusion by target objects, significant differences in surface reflectivity, and limited sensor accuracy. These issues result in sparse, void, or missing regions commonly found in the acquired point cloud data. Identifying missing regions in the point cloud data improves robustness and is more conducive to restoring the geometric details of the point cloud and maintaining boundary continuity.

[0019] Step S202: In the original point cloud data corresponding to the point cloud data, perform a neighborhood search for each proxy point to obtain the first neighborhood point set corresponding to each proxy point.

[0020] In this embodiment, by performing a neighborhood search on each surrogate point in the original point cloud data corresponding to the point cloud data, the first neighborhood point set corresponding to each surrogate point is obtained. This accurately locates the physical space that needs to be filled, providing clear geometric anchor points for the completion process. This effectively avoids the over-filling or under-filling (failure to completely fill holes) problems caused by the blurred boundaries of missing regions in traditional methods. By searching for neighborhood points in the outer region with the surrogate point as the center, the generation of completion points directly inherits and continues the geometric features of the real known region, significantly improving boundary continuity and detail fidelity. Furthermore, it reduces the dependence on labeled data, enhancing the generalization and practicality of the method.

[0021] Step S203: Based on the geometric features of each neighboring point in the first neighboring point set, aggregate to generate enhanced features for each proxy point, and determine the completion point and its location corresponding to each proxy point based on the enhanced features.

[0022] In this embodiment, enhanced features of each surrogate point are generated by aggregating the geometric features of each neighboring point in the first neighborhood point set. Based on the enhanced features, the completion point and its position corresponding to each surrogate point are determined, ensuring boundary continuity and detail fidelity. The calculation is based on the local geometric features of the input point cloud itself, without relying on any pre-trained generation model. Therefore, the computational resource consumption is low, and it can process large-scale point cloud data in real time or near real time, making it highly practical.

[0023] Step S204: Based on the position of the completion point, insert the completion point into the target missing region to complete the target missing region.

[0024] In this embodiment, by precisely inserting the generated completion points into the missing regions, the gaps in the geometric structure are directly filled, achieving spatial integrity of the point cloud data. This effectively eliminates the sense of discontinuity on the model surface, providing a dense and coherent data foundation for subsequent surface reconstruction and normal optimization, and is a key operation for restoring the geometric continuity of the 3D model.

[0025] Through steps S201 to S204, the target missing region of the point cloud data to be completed is determined, and surrogate points are generated within the target missing region. These surrogate points represent the reference points for completion. In the original point cloud data corresponding to the point cloud data, a neighborhood search is performed on each surrogate point to obtain a first neighborhood point set corresponding to each surrogate point. Based on the geometric features of each neighboring point in the first neighborhood point set, enhanced features of each surrogate point are aggregated. Based on the enhanced features, the completion point and its position corresponding to each surrogate point are determined. According to the completion point position, the completion point is inserted into the target missing region to complete the target missing region. This solves the problems in related technologies where point cloud completion is difficult to accurately recover geometric details and maintain boundary continuity in complex scenes, and is highly dependent on data and scene, with insufficient generalization ability. It achieves the beneficial effect of high-fidelity, high-consistency adaptive completion of sparse and incomplete point clouds in complex scenes without relying on large-scale labeled training data, effectively improving the reliability and practicality of point clouds in subsequent modeling, recognition, and precision measurement tasks.

[0026] It should be noted that in this embodiment, by introducing surrogate points in the missing regions and combining them with neighborhood density and geometric constraints, the location of holes in the point cloud can be accurately located, avoiding over- or under-filling caused by misjudgment in traditional methods. A weighted aggregation mechanism based on distance decay and normal consistency enables surrogate points to effectively integrate neighborhood information, maintaining the continuity and consistency of the geometric structure during the completion process and significantly improving boundary accuracy. The candidate point generation stage combines adaptive step size with surface fitting methods, ensuring not only a uniform distribution of completed points but also enhancing the naturalness of the point cloud distribution by introducing slight perturbations, making the results closer to real sampling.

[0027] After the completion process, this invention improves the smoothness and geometric detail preservation of the overall point cloud through fusion and post-processing operations, including redundant point removal, smoothing filtering, and normal consistency optimization. Experimental results show that in typical applications such as bridge inspection and material stacking measurement, the completed point cloud improves boundary sharpness and surface continuity by approximately 10%–15% compared to traditional geometric interpolation methods and pure deep learning methods, and improves the fidelity of local details by approximately 8%, while exhibiting better generalization robustness across different scenarios.

[0028] In some embodiments, before determining the target missing region of the point cloud data to be completed, the following steps are also included: Step 21: Obtain the original point cloud data. Based on the neighborhood distance between each point in the original point cloud data and its corresponding neighboring points, determine the noise points in the original point cloud data and remove the noise points from the original point cloud data. The noise points are used to represent points with large neighborhood distances in the original point cloud data.

[0029] In this embodiment, outlier noise points are accurately identified and removed through neighborhood distance statistics, effectively eliminating abnormal outliers introduced by scanning errors or environmental interference, and significantly improving the signal-to-noise ratio of point cloud data. This provides a cleaner and more reliable geometric basis for subsequent missing region detection and completion, avoiding interference from noise on hole boundary identification and feature calculation.

[0030] Step 22: Using the voxel grid method, downsample and equalize all points in the original point cloud data after removing noise points to obtain downsampled point cloud data with a preset point cloud density.

[0031] In this embodiment, if the original point cloud is used directly, the large density differences between points in different regions will increase the complexity of subsequent calculations and cause uneven distribution of completed points during the completion process. However, by using the voxel mesh method to downsample the point cloud after removing noise points, the amount of data can be significantly reduced while maintaining the overall geometric structure of the point cloud, resulting in a more balanced point cloud distribution and reducing the impact of density differences.

[0032] Step 23: Determine the neighborhood point set of each point in the downsampled point cloud data, and after calculating the covariance matrix of each point based on the neighborhood point set, perform eigenvalue decomposition on each covariance matrix, and use the eigenvector corresponding to the smallest eigenvalue obtained by decomposition as the normal vector of the corresponding point to obtain the point cloud data.

[0033] In this embodiment, covariance matrix decomposition can accurately extract the normal vectors of local surfaces from a set of discrete points, providing crucial geometric constraints for subsequent missing region detection, feature aggregation, and surface reconstruction. This statistical feature-based method is naturally robust to noise and density inhomogeneity, effectively avoiding normal estimation errors caused by local perturbations.

[0034] The above steps involve acquiring raw point cloud data. Based on the neighborhood distance between each point and its corresponding neighboring points, noise points are identified and removed from the raw point cloud data. These noise points represent points with excessively large neighborhood distances. Using a voxel grid method, all points in the raw point cloud data after noise removal are downsampled and equalized to obtain downsampled point cloud data with a preset density. The neighborhood set of each point in the downsampled point cloud data is determined, and after calculating the covariance matrix of each point based on the neighborhood set, eigenvalue decomposition is performed on each covariance matrix. The eigenvector corresponding to the smallest eigenvalue obtained from the decomposition is used as the normal vector of the corresponding point, thus obtaining the point cloud data. This significantly improves the geometric quality and feature consistency of the raw point cloud, laying the foundation for subsequent accurate completion. Specifically, outlier removal based on neighborhood statistics effectively filters measurement noise and abnormal reflection points, preventing noise from being misjudged as missing or interfering with feature extraction in subsequent steps. Using a voxel grid for downsampling reduces data size and forces a uniform spatial point density while maintaining global geometric structure, thus avoiding completion bias caused by density differences and improving computational efficiency. Normal vector estimation based on covariance matrix eigenvalue decomposition provides robust geometric properties describing local surface orientation and curvature for the point cloud, serving as a crucial basis for subsequent missing region identification, feature weighted aggregation, and surface reconstruction. The entire preprocessing workflow improves data quality while maintaining the algorithm's lightweight characteristics, making it suitable for the practical processing needs of large-scale engineering point clouds.

[0035] In some embodiments, noise points in the original point cloud data are determined based on the neighborhood distance between each point and its corresponding neighboring points, including the following steps: Step 31: Based on the distance between two adjacent points, detect neighboring points in all points of the original point cloud data whose distance to the target point is less than a preset radius, and obtain multiple neighboring points corresponding to each target point.

[0036] In this embodiment, a local geometric context is constructed for each point through a fixed-radius neighborhood search. This can stably extract local surface features (such as normals and curvature), providing an accurate local support domain for subsequent noise detection, hole identification, and feature aggregation, which is the foundation for all subsequent local geometric processing.

[0037] Step 32: Calculate the average distance between each target point and its corresponding multiple neighboring points to obtain the neighborhood distance between each target point and its corresponding neighboring points.

[0038] In this embodiment, by calculating the local average distance, the abstract "discrete set of points" is transformed into a quantified "local density index". It can accurately depict the local geometric distribution of each point, providing key data for subsequent noise identification and adaptive parameter adjustment.

[0039] Step 33: Determine whether the neighborhood distance corresponding to each target point is greater than the neighborhood distance threshold. If the neighborhood distance is greater than the neighborhood distance threshold, the target point is determined as a noise point. The neighborhood distance threshold is determined based on the mean and standard deviation of the neighborhood distance.

[0040] In this embodiment, a threshold is dynamically set based on statistical principles, which can adapt to density differences in different regions and accurately identify and remove outlier noise points far from the main distribution. This effectively avoids the problem of fixed thresholds failing to identify sparse or dense regions, significantly improving the quality of point cloud data and the robustness of subsequent processing.

[0041] In some embodiments, determining the target missing region of the point cloud data to be completed includes the following steps: Step 41: Perform a neighborhood search on each point in the point cloud data to obtain the second neighborhood point set corresponding to each point, and determine the local density corresponding to each point in the point cloud data based on the second neighborhood point set and the neighborhood volume corresponding to the second neighborhood point set.

[0042] In this embodiment, reference Figure 5 Detection of sparse points Figure 5 The dashed circle represents the neighborhood range defined by the target point, and the point cloud points in the dashed circle constitute the second neighborhood point set corresponding to the target point.

[0043] In this embodiment, discrete point clouds are transformed into a continuous density field through local volume density estimation. This accurately quantifies the local sampling density of each point, providing crucial density prior information for subsequent adaptive hole detection, multi-scale feature calculation, and non-uniform data processing.

[0044] In this embodiment, reference Figure 5 It can be seen that within the same or approximately the same neighborhood, the local density is higher when there are more neighborhood points and lower when there are fewer neighborhood points.

[0045] Step 42: Obtain the normal vector of each point in the point cloud data. Based on the normal vector of each point and each neighboring point in the second neighboring point set, determine the geometric continuity and curvature corresponding to each point. The geometric continuity is used to characterize the neighborhood direction difference determined by the angle between the normal vectors of each point and the corresponding neighboring points.

[0046] In this embodiment, the degree of geometric abrupt changes in the local surface is quantified from the discrete point set through normal vector consistency analysis. It can accurately identify discontinuous regions such as edges and hole boundaries, providing crucial topological information for subsequent accurate hole segmentation and preservation of geometric features.

[0047] Step 43: After detecting sparse points with local density less than a preset density threshold from the point cloud data, detect missing boundary points in the point cloud data where the neighborhood direction difference is greater than the difference threshold or where there is a sudden change in curvature.

[0048] Step 44: Construct the target missing region based on all sparse points and missing boundary points.

[0049] In this embodiment, by using the dual criteria of density filtering and geometric feature locking, false missing regions that are geometrically flat but simply due to sparse sampling are effectively eliminated. Only true boundary points with geometric discontinuities (abrupt normals, high curvature) are retained, providing accurate boundary constraints for subsequent structure-aware completion and avoiding over-completeness or structural distortion. (Reference) Figure 6 In this embodiment, sparse points are obtained through local density analysis, and missing boundary points are detected through normal difference and curvature abrupt change. Based on the sparse points and missing boundary points, the target missing region corresponding to the dashed outline is constructed. After obtaining the sparse points, the local geometric features of the point cloud are further analyzed. Specifically, the neighborhood direction difference is determined based on the difference in normal vector direction between the target point and its neighboring points. At the same time, the curvature abrupt change position is identified by combining the local curvature change situation. When a certain region has both obvious normal difference and curvature abrupt change, the relevant points near the region can be identified as missing boundary points. Subsequently, the sparse points obtained from the aforementioned local density analysis and the missing boundary points identified in this step are used together as constraints to construct the target missing region. It should be understood that the target missing region in this embodiment is not simply determined by low-density points, but is jointly determined by "sparse points + missing boundary points". Sparse points are used to reflect the low-density distribution near insufficient point cloud sampling or holes, and missing boundary points are used to reflect the geometric boundaries caused by discontinuity in the normal direction or abrupt curvature changes, thereby improving the accuracy of target missing region identification and avoiding misjudging ordinary sparse regions as missing regions that need to be filled.

[0050] In this embodiment, the discrete set of outliers is transformed into a structured missing region; it accurately defines the geometric range and internal filling space of the hole, providing a clear processing target for subsequent structure-aware completion and effectively avoiding misoperation of non-missing regions.

[0051] By performing a neighborhood search on each point in the point cloud data as described above, a second neighborhood point set corresponding to each point is obtained. Based on the neighborhood volume corresponding to the second neighborhood point set, the local density corresponding to each point in the point cloud data is determined. The normal vector of each point in the point cloud data is obtained. Based on the normal vector of each point and each neighboring point in the second neighborhood point set, the geometric continuity and curvature corresponding to each point are determined. Geometric continuity is used to characterize the neighborhood direction difference determined by the angle between the normal vectors of each point and the corresponding neighboring points. After detecting sparse points with local density less than a preset density threshold from the point cloud data, missing boundary points with neighborhood direction difference greater than the difference threshold or curvature abrupt changes are detected in the point cloud data. The target missing region is constructed based on all sparse points and missing boundary points. This avoids misjudging uniform sparse points as holes. By utilizing normal difference and curvature abrupt changes, complex geometric boundaries such as holes and edges can be accurately captured, preserving sharp features. This provides a complete input for subsequent completion algorithms, including both spatial range and geometric constraints.

[0052] It should be noted that in this embodiment, the missing region is not determined directly out of thin air. Instead, the local density is first calculated by the number of neighboring points and the neighborhood volume, and then sparse points are identified based on the density threshold to provide a density basis for the subsequent construction of the target missing region.

[0053] In some embodiments, generating proxy points within the target missing region includes the following steps: Step 51: Divide the target missing region into multiple voxel units, and use the center point corresponding to each voxel unit as the initial proxy point. The voxel in-voxel points of each voxel unit include at least one of sparse points and missing boundary points.

[0054] In this embodiment, the dense and disordered original point cloud is transformed into regularly distributed surrogate points, significantly reducing the subsequent computational load. By replacing the original scattered points with voxel center points, the amount of data is reduced while smoothing out minor noise within the voxels and preserving the macroscopic geometric features of missing regions. The voxel structure inherently possesses spatial partitioning characteristics, providing an efficient indexing foundation for subsequent operations such as neighborhood search and normal vector estimation, significantly improving algorithm efficiency.

[0055] Step 52: Based on each initial proxy point, generate proxy points with a preset adaptive step size to generate multiple candidate proxy points. The adaptive step size is used to represent the spacing between proxy points and is dynamically adjusted according to the local density corresponding to the missing boundary points.

[0056] In this embodiment, the step size is dynamically adjusted using the local density of missing boundary points to ensure that the spacing between generated surrogate points is consistent with the density of the surrounding real point cloud, avoiding oversampling in sparse areas or undersampling in dense areas. Through adaptive step size control, the generated candidate surrogate points can better conform to the geometric distribution of the original point cloud, preventing unnatural concavity and convexity deformations in the completed area. This avoids computational redundancy caused by a globally uniform small step size and improves generation efficiency by dynamically adjusting the sampling density.

[0057] Step 53: Based on the distance between two adjacent candidate proxy points, use the Poisson disk sampling method to filter all candidate proxy points, so as to select all candidate proxy points whose distance between two adjacent candidate proxy points is greater than the preset distance threshold, and obtain the proxy points corresponding to the target missing region.

[0058] In this embodiment, the spacing between proxy points is required to be greater than a preset threshold, effectively preventing candidate points from being too dense or overlapping in local areas. Based on the boundary density constraints, the spatial distribution of the point set is further optimized, ensuring that the proxy points present a uniform, irregular, but gapless coverage state throughout the missing area. The filtered proxy point set exhibits excellent spatial independence, guaranteeing the smoothness of the completed surface.

[0059] The above steps involve segmenting the target missing region into multiple voxel units, using the center point of each voxel unit as an initial surrogate point; generating multiple candidate surrogate points based on each initial surrogate point with a preset adaptive step size; and filtering all candidate surrogate points using the Poisson disk sampling method based on the distance between two adjacent candidate surrogate points to select all candidate surrogate points whose distance between two adjacent candidate surrogate points is greater than a preset distance threshold, thus obtaining the surrogate points corresponding to the target missing region. This significantly reduces the data size and ensures that the distribution of generated candidate points strictly matches the density characteristics of the original point cloud.

[0060] In some embodiments, a neighborhood search is performed on each surrogate point in the original point cloud data corresponding to the point cloud data to obtain a first neighborhood point set corresponding to each surrogate point, including the following steps: Step 61: Among all the original points in the original point cloud data, perform a radius neighborhood search centered on the surrogate point to obtain multiple candidate neighborhood points.

[0061] In this embodiment, reverse neighborhood search is used to lock the real original points around each virtual proxy point located in the missing region, ensuring that the subsequent completion points can be projected and fused strictly according to the local features of the real surface. This is the key mechanism to ensure that the completion process does not deviate from the actual geometric constraints.

[0062] Step 62: Determine the number of candidate neighbor points, and if the number of candidate neighbor points is less than the set number, use K-nearest neighbor search to search for the corresponding original points from the original points other than the candidate neighbor points.

[0063] In this embodiment, the problem of neighborhood search failure in sparse regions is completely solved. It ensures that each proxy point obtains a sufficient number of reference points, preventing completion interruptions or fitting distortions caused by insufficient neighborhood points.

[0064] Step 63: Take the original point and multiple candidate neighbor points found by the K-nearest neighbor search as the first neighbor point set corresponding to the surrogate point.

[0065] In this embodiment, a complete and reliable first neighborhood point set is constructed through a hybrid strategy of radius-first and K-nearest neighbor complementation. This ensures that each surrogate point obtains a sufficient number of reference points, preventing complementation interruptions or fitting distortions caused by insufficient neighborhood points.

[0066] Through the above steps, a radius neighborhood search centered on the surrogate point is performed among all the original points in the original point cloud data to obtain multiple candidate neighbor points; the number of candidate neighbor points is determined, and if the number of candidate neighbor points is less than a set number, the K-nearest neighbor search is used to search for the corresponding original point from the original points other than the candidate neighbor points; the original point and multiple candidate neighbor points found by the K-nearest neighbor search are used as the first neighbor point set corresponding to the surrogate point; a stable and reliable first neighbor point set is constructed for each surrogate point, which effectively solves the problem of insufficient reference points at the edge of the missing region due to the sparse point cloud.

[0067] In some embodiments, based on the geometric features of each neighboring point in the first neighboring point set, enhanced features of each surrogate point are aggregated to generate enhanced features. Based on the enhanced features, the completion point and its location corresponding to each surrogate point are determined, including the following steps: Step 71: Obtain the geometric features of each neighboring point in the first neighboring point set, wherein the geometric features include the coordinate parameters, normal vector and local curvature of the corresponding neighboring point.

[0068] In this embodiment, by extracting the multidimensional geometric attributes of neighboring points, the completion position of the proxy point can not only fit the spatial distribution of the original point, but also strictly follow the normal and curvature characteristics of the local surface, thereby significantly improving the geometric realism and smoothness of the completion result.

[0069] Step 72: Based on the coordinate parameters and normal vector, calculate the feature weight corresponding to each neighboring point using a weighting function based on distance decay and direction consistency. Then, aggregate and weight the geometric features of all neighboring points according to the feature weights to generate enhanced features.

[0070] In this embodiment, intelligent aggregation of neighborhood geometric information is achieved through a dual weighting mechanism of distance attenuation and orientation consistency. This mechanism gives higher weight to reliable points that are close to each other and have the same orientation, thereby suppressing the interference of noise and outliers. The generated enhanced features retain the sharpness of local details while ensuring a smooth transition of the completed region.

[0071] Step 73: Using a mapping function that employs linear mapping and surface fitting, determine the spatial offset corresponding to the enhanced feature, and based on the spatial offset, the random perturbation term corresponding to the spatial offset, and the coordinate parameters corresponding to the surrogate point, generate the spatial coordinate parameters corresponding to the completion point, wherein the location of the completion point includes the spatial coordinate parameters.

[0072] In this embodiment, linear mapping and surface fitting ensure that the completed points strictly follow local geometric rules and fit the original surface trend; while the introduced random perturbation term simulates the inherent disorder and natural scattering characteristics of real point clouds, making the completion result more visually realistic while maintaining structural accuracy.

[0073] The steps described above involve obtaining the geometric features of each neighboring point in the first neighborhood point set; calculating the feature weights corresponding to each neighboring point based on the coordinate parameters and normal vectors using a weighting function based on distance decay and direction consistency; aggregating and weighting the geometric features of all neighboring points based on the feature weights to generate enhanced features; determining the spatial offset corresponding to the enhanced features using a mapping function employing linear mapping and surface fitting; and generating the spatial coordinate parameters corresponding to the completed points based on the spatial offset, the random perturbation term corresponding to the spatial offset, and the coordinate parameters corresponding to the surrogate points. This achieves a high degree of unity between geometric accuracy, structural continuity, and natural distribution in the completed point set.

[0074] In some embodiments, the feature weights corresponding to each neighboring point are calculated according to the following formula: Among them, w ij It is the feature weight, p i and p k All are neighborhood points, a j It is an agency point, n i It is point p i The normal vector, n k It is point p k The normal vector, n j Is it agent point A? j The average normal vector of the neighborhood, N(a) j ) is the neighborhood point set, and σ is the scale parameter.

[0075] In this embodiment, the feature weight formula has two functions: first, it makes closer points have a greater impact on the results through exponential decay; second, it suppresses points with excessively large geometric differences through normal dot product, thereby improving the geometric consistency of aggregation.

[0076] This feature weighting formula achieves a quantitative evaluation of the neighborhood geometric contribution through normalized weighting based on distance decay and normal consistency. The distance term ensures that the nearest point dominates feature aggregation, fitting the local surface; the normal term filters geometric information on the same side, suppressing interference from opposite surfaces or noise; and the normalization process stabilizes the feature scale, enabling the enhanced features obtained by the surrogate point to have both local accuracy and noise robustness, providing high-quality geometric feature parameters for subsequent completion.

[0077] In some embodiments, inserting the completion point into the target missing region based on the completion point location includes the following steps: Step 81: Insert the corresponding completion point at the position corresponding to the completion point in the target missing area.

[0078] In this embodiment, point cloud misalignment or boundary conflicts that may be caused by traditional interpolation are avoided, and the perfect geometric and topological continuity between the newly added points and the original structure is ensured, so that the missing area is completely repaired in both visual and data structure terms.

[0079] Step 82: In the target missing region where the insertion and completion points are inserted, surface repair is performed using the local geometric reconstruction method to obtain the completed point cloud. The local geometric reconstruction method includes one of the following: moving least squares method and local triangulation method.

[0080] In this embodiment, discrete completion points are transformed into explicit surfaces by moving least squares (MLS) or local triangulation. By utilizing the smooth fitting characteristics of moving least squares or the local linear interpolation of triangulation, a strict geometric continuity is established between the completed region and the original point cloud, effectively eliminating splicing gaps and abrupt transitions, and ensuring that the reconstructed surface transitions naturally in terms of both visual appearance and geometric structure.

[0081] Step 83: Perform smoothing filtering and global normal adjustment on the completed point cloud to obtain the completed point cloud data.

[0082] In this embodiment, smoothing filtering is used to eliminate minor noise and irregular protrusions that may exist in the completed area due to discrete generation, ensuring a smooth surface; then, global normal adjustment is used to unify the normal consistency between the newly completed area and the original point cloud, eliminating the seam effect that may occur in lighting rendering or subsequent modeling.

[0083] The above steps involve inserting corresponding fill points at positions corresponding to the fill point locations within the target missing region; then, surface repair is performed on the target missing region where fill points are inserted using local geometric reconstruction methods to obtain the filled point cloud. These local geometric reconstruction methods include one of the following: moving least squares method or local triangulation method; the filled point cloud is then smoothed and the global normal of the point cloud is adjusted to obtain the filled point cloud data; this achieves a complete closed loop from "point filling" to "surface reconstruction" and then to "global optimization," ensuring that the filled region meets industrial-grade application standards in terms of geometric accuracy, visual continuity, and data structure.

[0084] The embodiments of this application will be described and illustrated below through preferred embodiments.

[0085] Figure 3 This is a flowchart of a point cloud completion method based on the aggregation of surrogate points and neighborhood features according to a preferred embodiment of this application. Figure 3 As shown, the point cloud completion method based on the aggregation of surrogate points and neighborhood features includes the following steps: Step S301: Obtain the raw point cloud data. A point cloud, as a three-dimensional data set composed of a large number of discrete points, is usually denoted as: , where p i =(x i y i , z i Let x represent the i-th point in the point cloud P, where i ∈ {1, 2, ..., N}. i y i , z i They represent p respectively i The X, Y, and Z coordinates.

[0086] Point cloud acquisition relies on sensors such as LiDAR, depth cameras, or structured light. However, these sensors are often affected by various uncertainties in real-world applications, such as limited viewing angles, occlusion by the target object, significant differences in surface reflectivity, and limited sensor accuracy. These issues result in sparse, hollow, or missing regions commonly found in the acquired point cloud data. Figure 4 As shown. Meanwhile, due to environmental interference or equipment noise, point clouds may also contain isolated outliers or measurement errors. If these defects are not corrected, they will directly affect the stability of subsequent missing region detection and will be amplified during the completion process, leading to blurred boundaries, distorted geometry, and even large-area distortion. Therefore, before the data enters the missing region detection and completion process, the point cloud is first systematically preprocessed. Preprocessing includes three core steps: noise removal, downsampling equalization, and normal estimation.

[0087] Step S302, noise removal, namely: acquiring the original point cloud data, determining the noise points in the original point cloud data based on the neighborhood distance between each point in the original point cloud data and its corresponding neighboring points, and removing the noise points from the original point cloud data. The noise points are used to characterize the points in the original point cloud data with large neighborhood distances.

[0088] This embodiment employs a neighborhood statistics-based outlier identification method to remove noisy points. The basic idea is to determine whether a point is consistent with the overall distribution by analyzing its spatial relationship with its surrounding neighbors. Specifically, for any point p in the point cloud... i Find the neighborhood point set N(p) within a given radius r. i ): N(p i )={ p i ∈P| || p j - p i ||≤r}, and calculate the value of point p. i The average distance d between neighboring points i : ; By statistically analyzing the average distance d between the neighboring points of all points in the point cloud. i The global mean μ is obtained. d With standard deviation σ d When the neighborhood distance of a point is significantly larger, i.e., satisfying d i >μ d +ασ d If the threshold value is α, then the point is identified as an outlier and removed. This method effectively identifies isolated points, reflection interference points, and scanning artifacts, thus making the point cloud distribution closer to the true geometric shape.

[0089] Step S303, downsampling equalization, that is: using the voxel grid method, all points in the original point cloud data after removing noise points are downsampled and equalized to obtain downsampled point cloud data with a preset point cloud density. After denoising, this embodiment uses the voxel grid method to downsample the point cloud after removing noise points to reduce the data size and equalize the spatial distribution of points. If the original point cloud is used directly, the large density difference of points in different regions will increase the complexity of subsequent calculations and also cause uneven distribution of completed points during the completion process. This invention uses the voxel grid method to divide the three-dimensional space into sections with a side length of A cubic mesh, where points within each voxel are represented by their geometric centers: Where V is the set of points within a voxel, p vThis represents the point of the voxel; in this way, the amount of data can be greatly reduced while maintaining the overall geometric structure of the point cloud, making the point cloud distribution more balanced and reducing the impact of density differences.

[0090] Step S304, normal vector estimation, namely: determining the neighborhood point set of each point in the downsampled point cloud data, and after calculating the covariance matrix of each point based on the neighborhood point set, performing eigenvalue decomposition on each covariance matrix, and using the eigenvector corresponding to the smallest eigenvalue obtained by decomposition as the normal vector of the corresponding point, thus obtaining the point cloud data. This embodiment uses eigenvalue decomposition based on the neighborhood covariance matrix to estimate the normal vector of each point. The normal vector, as an important feature describing the local surface morphology, plays an important role in the subsequent detection of hole regions and the completion process. The specific method is: for each point p i The covariance matrix C is calculated using its neighborhood point set. i : ,in, Let T be the mean of the neighborhood points, and T be the transpose.

[0091] By analyzing matrix C i Perform eigenvalue decomposition to obtain the eigenvector corresponding to the smallest eigenvalue, which is used as the point p. i Normal vector: .

[0092] Step S305: Determine the target missing region of the point cloud data to be completed.

[0093] In this embodiment, due to factors such as occlusion, scanning blind spots, or surface reflection, sparse and hollow regions often appear in the point cloud. To identify these regions, this embodiment first calculates the local density of the points. Defined as: , where N(p i V is the neighborhood point set. r For the neighborhood volume, when Points below a set threshold are marked as sparse points. Subsequent corrections are made based on geometric continuity and curvature characteristics. If the difference in neighborhood direction is too large or the curvature changes abruptly, the point is identified as a missing boundary point. Through joint determination using multiple indicators, a set of missing regions Ω is finally obtained, providing target regions for subsequent proxy point generation.

[0094] Step S306: Generate proxy points in the missing regions.

[0095] This embodiment generates a set of proxy points A={a} within the missing region Ω. j |j=1,2,…,M} is used for subsequent neighborhood feature aggregation and candidate point generation; the generation of surrogate points follows the principles of coverage and uniformity; to achieve coverage, the missing region is first spatially partitioned, dividing Ω into voxel units V.k Let the side length of the voxel be h. Calculate the geometric center of the candidate point within each voxel: , where V k This represents the set of in-voxel points corresponding to the proxy point.

[0096] The above steps ensure global coverage of the proxy points. To achieve uniformity, this invention introduces adaptive step size control. This is based on the local density of points near the missing boundary. Dynamically adjust the generation spacing h(a) of proxy points j ): Where k is an adjustment coefficient. When local points are sparse, the step size is reduced to generate more surrogate points; when local points are dense, the step size is increased to reduce redundant surrogate points.

[0097] While ensuring uniform distribution, this embodiment adaptively adjusts the generation spacing of surrogate points based on the point density at the boundary of the missing region, and uses a Poisson disk sampling constraint to limit the minimum distance between surrogate points; for any two surrogate points a j a k If ||a j -a k ||<d min Then, discard the points generated after the replacement, until all proxy points satisfy the minimum spacing d. min This process avoids excessive aggregation of agency points; such as Figure 7 This is a schematic diagram of the proxy point generation process according to a preferred embodiment of this application. It shows the evolution of the proxy points from initial uniform coverage to adaptive adjustment, and then to Poisson disk constraint screening. It intuitively reflects the principle of this embodiment in generating a uniformly covered and reasonably distributed set of proxy points in the missing region.

[0098] Step S307: In the original point cloud data corresponding to the point cloud data, perform a neighborhood search for each proxy point to obtain the first neighborhood point set corresponding to each proxy point.

[0099] This embodiment focuses on the agent point set A={a j | j=1, 2, ..., M}, in the original point cloud P={p i Perform a neighborhood search in | i = 1, 2, ..., N} to obtain the local point set N(a) of each agent point. j ) .

[0100] Figure 8 This is a schematic diagram of a neighborhood search method according to a preferred embodiment of this application. The neighborhood search employs two methods: sphere radius search and K-nearest neighbor search. A sphere query is performed with the surrogate point as the center. If the number of returned points is insufficient, the search switches to searching the K points closest to the surrogate point.

[0101] (1) Ball radius search, that is: to perform ball query with the proxy point as the center, given the radius parameter r, the proxy point a j The first neighborhood point set N r (a j ) is defined as: N r (a j ) ={ p i ∈P| || p i – a j The sphere radius search (||≤r}) ensures that all neighboring points are within a fixed spatial range and is suitable for regions with uniform point cloud density.

[0102] (2) K-nearest neighbor search, that is: search for the K points closest to the agent point.

[0103] When the point cloud density is uneven, a fixed number of neighbors K is used, and its first neighborhood point set N is used. K (a j )for: , where the point set is arranged according to a j The Euclidean distances, arranged in ascending order, satisfy: It is important to understand that while K-nearest neighbor search can ensure a stable number of neighboring points, the radius may be too large in sparse regions, thus introducing points far from the boundary.

[0104] In this embodiment, to balance stability and adaptability, this application combines two methods: sphere radius search is used in regions with high point density, and K-nearest neighbor search is used in sparse regions. The specific criterion is: if the number of points in the sphere's neighborhood |N r (a j )|<K min Then use Nearest neighbor search completion.

[0105] Step S308: Based on the geometric features of each neighboring point in the first neighboring point set, aggregate to generate enhanced features for each surrogate point, and determine the completion point and its position corresponding to each surrogate point according to the enhanced features.

[0106] In this embodiment, the coordinates and normal vector features of the points in the neighborhood point set are extracted; the geometric properties of the neighborhood are calculated, including the average point spacing d(a j ) and local covariance matrix C(a j ): ; ; Wherein, the average point spacing d(a) j ) is used to characterize the sparsity of the neighborhood, and the local covariance matrix C(a) j This is used for subsequent feature aggregation and surface approximation.

[0107] In this embodiment, feature weights are calculated and weighted aggregation is performed based on the distance and normal consistency between each neighboring point and the surrogate point to obtain enhanced features of the surrogate point that can characterize the local geometric structure. Figure 9 This is a schematic diagram of neighborhood feature weighted aggregation according to a preferred embodiment of this application, such as... Figure 9 As shown, for each agent point a j Its neighborhood point set is denoted as N(a j Let p be a neighboring point. i =(x i y i , z i Its characteristics can be expressed as f(p) i )=[ x i y i , z i n i k i ], where n i k is the normal vector calculated in step 302. i As a local curvature, this defined feature includes both spatial coordinates and geometric properties.

[0108] To ensure a reasonable contribution of different neighboring points to the proxy point, this embodiment designs a weighting function based on distance decay and direction consistency. Its feature weights are calculated using the following formula: Among them, w ij It is the feature weight, p i and p k All are neighborhood points, a j It is an agency point, n i It is point p i The normal vector, n k It is point p k The normal vector, n j Is it agent point A? j The average normal vector of the neighborhood, N(a) j ) is the neighborhood point set, and σ is the scale parameter. This feature weight calculation formula has two functions: first, it makes closer points have a greater impact on the result through exponential decay; second, it suppresses points with excessively large geometric differences through normal dot product, thereby improving the geometric consistency of aggregation.

[0109] Therefore, the enhanced features of the proxy point are represented as follows: .

[0110] The spatial offset is calculated based on the enhanced features; the spatial coordinates of the surrogate point and its enhanced features are then mapped to the candidate point space. Specifically, the candidate point coordinates can be represented as: , where △(a)j ) represents the spatial offset inferred from features.

[0111] The offset calculation is determined by both the enhanced features and geometric constraints. This embodiment employs a combination of linear mapping and surface fitting. Firstly, the mapping from features to offsets can be achieved using a small mapping function: △(a j )=W·F(a j )+b, where W is the mapping matrix and b is the bias parameter. In surface fitting, the principal direction of the neighborhood of the surrogate point is used as the local reference system. The neighborhood points are fitted with a plane or quadratic surface under the reference system. The fitting equation can be written as: z=αx+βy+γ, and the local insertion position of the candidate point is determined accordingly.

[0112] During the generation process, this embodiment also introduces diversity constraints. If all candidate points are strictly located on the fitted surface, the distribution of points will be too regular and lack natural perturbations that match the real point cloud. Therefore, this embodiment allows the addition of random perturbation terms to the offset when generating candidate points. This makes the spatial distribution of candidate points closer to the real point cloud: The final set of candidate points is denoted as: .

[0113] In this embodiment, the candidate completion point generation and insertion process includes three stages: generating offset vectors from surrogate points and generating initial candidate points; projecting candidate points onto a local fitting plane or surface to ensure geometric continuity and consistency; and inserting candidate points into the missing region and combining them with the original point cloud to form a complete point cloud, thereby achieving local completion of the defective region.

[0114] Step S309: Based on the position of the completion point, insert the completion point into the target missing region to complete the target missing region. Set the candidate completion points... Insert into the missing region and perform local geometric reconstruction to ensure the continuity of the point cloud in terms of boundary and overall geometry.

[0115] First, the candidate points are merged with the boundary of the missing region set Ω. To avoid abrupt distribution of candidate points near the boundary, the spatial relationship between the candidate points and the original points needs to be constrained. Let the candidate points... to the boundary point set The minimum distance is: ,like If the point deviates too far from the boundary, it will be removed or resampled to ensure the spatial rationality of the completion process.

[0116] After inserting candidate points, a local geometric reconstruction method (moving least squares or local triangulation) is used to reconstruct the surface in the region where the completed points are inserted, so as to ensure the geometric continuity between the completed region and the original point cloud.

[0117] First, a local triangulation is constructed based on the candidate points and their neighboring point sets. A commonly used method is Delaunay triangulation, whose principle is that, in a two-dimensional projection plane, the triangulation of a given point set should satisfy the condition that the circumcircle of any triangle does not contain any other points, thus obtaining a topologically reasonable mesh structure. Local surface approximation in three-dimensional space can be achieved by projecting the candidate points and their neighboring boundary points onto a tangent plane and performing Delaunay triangulation.

[0118] After obtaining the local topology, this invention further employs a surface fitting method to improve geometric accuracy. Specifically, Moving Least Squares (MLS) can be used to approximate the local point set as a continuous smooth surface. Let the local fitting function be... Its goal is to minimize the weighted residuals: ,in, The weighting function is typically a Gaussian weighting function. h is a smoothing parameter, which ensures that the local surface is spatially continuous and smooth after the candidate point is inserted.

[0119] In this embodiment, after local reconstruction is completed, boundary consistency constraints can also be applied. That is, at the boundary between the completed region and the original point cloud, continuity of the normal direction is required. If the candidate point set... If the difference between the normal distribution and the boundary normal is too large, the candidate points are iteratively adjusted until the condition is met. ,in, This is the normal consistency threshold.

[0120] Then, the candidate point set Compared with the original point cloud P={p i The points are then merged to form a complete point cloud P*, such as... Figure 10 This is a schematic diagram of the complete point cloud after point cloud completion according to a preferred embodiment of this application. First, the candidate point set is merged with the original point cloud to obtain: During the fusion process, this application uses a distance criterion to eliminate redundant points and sets candidate points. The closest distance to the original point cloud is: ,like If a candidate point is found to be duplicated from the original point, it is deleted to avoid data redundancy and over-density.

[0121] Step S308, post-processing step: Perform smoothing filtering and normal consistency optimization on the fused complete point cloud.

[0122] To further improve the smoothness of the point cloud, this embodiment introduces a smoothing filter operation on the fused point cloud. The Moving Least Squares (MLS) method is used to make the point cloud surface more continuous through locally weighted fitting. For point p, its smoothing position is given by the following formula: ,in, For local fitting of planes or surfaces, Using a Gaussian weighting function, smoothing filtering can effectively reduce noise interference while preserving boundary details.

[0123] In this embodiment, after smoothing filtering, normal consistency optimization is further performed. Specifically, the orientation of the point cloud surface is unified through global normal adjustment. For any point p in the point cloud... k If the angle between its normal and the average normal of its adjacent points is greater than the threshold θ max Then, its normal is flipped to make the overall normal distribution consistent: .

[0124] Finally, to facilitate the application of point clouds in downstream tasks, this embodiment can resample the point cloud to maintain uniform density and generate the final result dataset: P final =Resample(P*), where Resample(·) represents a resampling operator based on voxels or Poisson disks.

[0125] Through the above fusion and post-processing, the completed point cloud and the original point cloud are seamlessly combined. The data avoids excessive density and redundancy, and is optimized in terms of smoothness and normal consistency, thereby generating a final point cloud with complete structure and high geometric accuracy, providing reliable input for subsequent modeling, detection and measurement.

[0126] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although the flowchart shows a logical order, in some cases the steps shown or described may be executed in a different order than that shown here.

[0127] This embodiment also provides a point cloud completion device based on the aggregation of surrogate points and neighborhood features. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0128] Figure 11This is a structural block diagram of a point cloud completion device based on the aggregation of surrogate points and neighborhood features according to an embodiment of this application, as shown below. Figure 11 As shown, the device includes: a generation module 91, a search module 92, an analysis module 93, and a completion module 94, wherein... The generation module 91 is used to determine the target missing region of the point cloud data to be completed, and to generate surrogate points within the target missing region, wherein the surrogate points are used to represent the reference points for completion; The search module 92, coupled to the generation module 91, is used to perform a neighborhood search on each proxy point in the original point cloud data corresponding to the point cloud data, and obtain the first neighborhood point set corresponding to each proxy point. The analysis module 93, coupled to the search module 92, is used to aggregate and generate enhanced features for each proxy point based on the geometric features of each neighboring point in the first neighborhood point set, and determine the completion point and its position corresponding to each proxy point based on the enhanced features. The completion module 94, coupled to the analysis module 93, is used to insert the completion point into the target missing region according to the position of the completion point, so as to complete the target missing region.

[0129] The point cloud completion device based on surrogate point and neighborhood feature aggregation provided in this application embodiment determines the target missing region of the point cloud data to be completed and generates surrogate points within the target missing region, where the surrogate points are used to represent the reference points for completion; in the original point cloud data corresponding to the point cloud data, a neighborhood search is performed on each surrogate point to obtain a first neighborhood point set corresponding to each surrogate point; based on the geometric features of each neighborhood point in the first neighborhood point set, enhanced features of each surrogate point are aggregated to generate enhanced features; and based on the enhanced features, the completion point and completion method corresponding to each surrogate point are determined. The system identifies the full point location and inserts the fill point into the missing region of the target to complete the missing region. This addresses the problems of point cloud completion in complex scenes, such as difficulty in accurately restoring geometric details and maintaining boundary continuity, strong dependence on data and scene, and insufficient generalization ability. It achieves the beneficial effect of high-fidelity and high-consistency adaptive completion of sparse and incomplete point clouds in complex scenes without relying on large-scale labeled training data, effectively improving the reliability and practicality of point clouds in subsequent modeling, recognition, and precision measurement tasks.

[0130] In some embodiments, the generation module 91 further includes: The first preprocessing unit is used to acquire the original point cloud data, determine the noise points in the original point cloud data based on the neighborhood distance between each point in the original point cloud data and its corresponding neighboring points, and remove the noise points from the original point cloud data. The noise points are used to characterize the points in the original point cloud data with large neighborhood distances.

[0131] The second preprocessing unit, coupled to the first preprocessing unit, is used to perform downsampling equalization processing on all points in the original point cloud data after removing noise points using the voxel grid method, so as to obtain downsampled point cloud data with a preset point cloud density.

[0132] The third preprocessing unit, coupled to the second preprocessing unit, is used to determine the neighborhood point set of each point in the downsampled point cloud data. After calculating the covariance matrix of each point based on the neighborhood point set, it performs eigenvalue decomposition on each covariance matrix and uses the eigenvector corresponding to the smallest eigenvalue obtained by decomposition as the normal vector of the corresponding point to obtain the point cloud data.

[0133] In some embodiments, the first preprocessing unit further includes: The detection component is used to detect neighboring points in the original point cloud data whose distance to the target point is less than a preset radius, based on the distance between two adjacent points, and to obtain multiple neighboring points corresponding to each target point.

[0134] The calculation component, coupled to the detection component, is used to calculate the average distance between each target point and its corresponding multiple neighboring points, thus obtaining the neighborhood distance between each target point and its corresponding neighboring points.

[0135] The decision component, coupled to the calculation component, is used to determine whether the neighborhood distance corresponding to each target point is greater than the neighborhood distance threshold. If the neighborhood distance is greater than the neighborhood distance threshold, the target point is determined to be a noise point. The neighborhood distance threshold is determined based on the mean and standard deviation of the neighborhood distance.

[0136] In some embodiments, the generation module 91 further includes: The density determination unit is used to perform a neighborhood search for each point in the point cloud data to obtain the second neighborhood point set corresponding to each point, and to determine the local density corresponding to each point in the point cloud data based on the second neighborhood point set and the neighborhood volume corresponding to the second neighborhood point set.

[0137] The curvature determination unit, coupled to the density determination unit, is used to obtain the normal vector of each point in the point cloud data. Based on the normal vector of each point and each neighboring point in the second neighboring point set, the geometric continuity and curvature of each point are determined. The geometric continuity is used to characterize the neighborhood direction difference determined by the angle between the normal vectors of each point and the corresponding neighboring points.

[0138] The missing detection unit, coupled with the curvature determination unit, is used to detect missing boundary points in the point cloud data after detecting sparse points with local density less than a preset density threshold, where the neighborhood directional difference is greater than a difference threshold or there is a sudden change in curvature.

[0139] The building unit, coupled to the missing detection unit, is used to construct the target missing region based on all sparse points and missing boundary points.

[0140] In some embodiments, the generation module 91 further includes: The first segmentation unit is used to segment the target missing region into multiple voxel units, and the center point corresponding to each voxel unit is used as the initial surrogate point. The voxel in-voxel points of each voxel unit include at least one of sparse points and missing boundary points.

[0141] The first generation unit, coupled to the first segmentation unit, is used to generate proxy points based on each initial proxy point with a preset adaptive step size, generating multiple candidate proxy points. The adaptive step size is used to represent the spacing between proxy points and is dynamically adjusted according to the local density corresponding to the missing boundary points.

[0142] The first filtering unit, coupled to the first generation unit, is used to filter all candidate proxy points based on the distance between two adjacent candidate proxy points using the Poisson disk sampling method, so as to filter out all candidate proxy points whose distance between two adjacent candidate proxy points is greater than a preset distance threshold, and obtain the proxy points corresponding to the target missing region.

[0143] In some embodiments, the search module 92 further includes: The radius neighborhood search unit is used to perform a radius neighborhood search centered on the surrogate point among all the original points in the original point cloud data to obtain multiple candidate neighborhood points.

[0144] The K-nearest neighbor search unit and the radius neighbor search unit are used to determine the number of candidate neighbor points. If the number of candidate neighbor points is less than a set number, the K-nearest neighbor search is used to search for the corresponding original point from the original points other than the candidate neighbor points.

[0145] The combination unit, coupled to the K-nearest neighbor search unit, is used to take the original point and multiple candidate neighbor points found by the K-nearest neighbor search as the first neighbor point set corresponding to the surrogate point.

[0146] In some embodiments, the analysis module 93 further includes: The first acquisition unit is used to acquire the geometric features of each neighboring point in the first neighborhood point set, wherein the geometric features include the coordinate parameters, normal vector and local curvature of the corresponding neighboring point.

[0147] The aggregation weighting unit, coupled to the first acquisition unit, is used to calculate the feature weights corresponding to each neighboring point based on the coordinate parameters and normal vectors using a weighting function based on distance decay and direction consistency. Based on the feature weights, the geometric features of all neighboring points are aggregated and weighted to generate enhanced features.

[0148] The spatial determination unit, coupled with the aggregation weighting unit, is used to determine the spatial offset corresponding to the enhanced feature using a mapping function that employs linear mapping and surface fitting. Based on the spatial offset, the random perturbation term corresponding to the spatial offset, and the coordinate parameters corresponding to the surrogate point, it generates the spatial coordinate parameters corresponding to the completion point, wherein the location of the completion point includes the spatial coordinate parameters.

[0149] In some embodiments, the feature weights corresponding to each neighboring point are calculated according to the following formula: Among them, w ij It is the feature weight, p i and p k All are neighborhood points, a j It is an agency point, n i It is point p i The normal vector, n k It is point p k The normal vector, n j Is it agent point A? j The average normal vector of the neighborhood, N(a) j ) is the neighborhood point set, and σ is the scale parameter.

[0150] In some embodiments, the completion module 94 further includes: The insertion unit is used to insert the corresponding completion point at the position corresponding to the completion point position within the target missing region.

[0151] The geometric reconstruction unit, coupled with the insertion unit, is used to repair the surface of the target missing region of the inserted completion point using the local geometric reconstruction method to obtain the completed point cloud. The local geometric reconstruction method includes one of the following: moving least squares method and local triangulation method.

[0152] The adjustment and correction unit, coupled with the geometric reconstruction unit, is used to perform smoothing filtering and global normal adjustment and correction on the completed point cloud to obtain the completed point cloud data.

[0153] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0154] This embodiment also provides a point cloud completion system, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0155] Optionally, the point cloud completion system may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0156] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0157] S1, determine the target missing region of the point cloud data to be completed, and generate surrogate points within the target missing region, where the surrogate points are used to represent the reference points for completion.

[0158] S2, in the original point cloud data corresponding to the point cloud data, perform a neighborhood search for each proxy point to obtain the first neighborhood point set corresponding to each proxy point.

[0159] S3. Based on the geometric features of each neighboring point in the first neighboring point set, aggregate to generate enhanced features for each surrogate point. Based on the enhanced features, determine the completion point and its location corresponding to each surrogate point.

[0160] S4. Based on the position of the completion point, insert the completion point into the target missing region to complete the target missing region.

[0161] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0162] Furthermore, in conjunction with the point cloud completion method based on surrogate point and neighborhood feature aggregation in the above embodiments, this application embodiment can provide a storage medium for implementation. This storage medium stores a computer program; when executed by a processor, the computer program implements any of the point cloud completion methods based on surrogate point and neighborhood feature aggregation in the above embodiments.

[0163] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0164] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A point cloud completion method based on surrogate point and neighborhood feature aggregation, characterized in that, include: The target missing region of the point cloud data to be completed is determined, and surrogate points are generated within the target missing region, wherein the surrogate points are used to represent the reference points for completion; In the original point cloud data corresponding to the point cloud data, a neighborhood search is performed on each of the proxy points to obtain the first neighborhood point set corresponding to each of the proxy points; Based on the geometric features of each neighboring point in the first neighborhood point set, the enhanced features of each surrogate point are aggregated and generated. Based on the enhanced features, the completion point and the position of the completion point corresponding to each surrogate point are determined. Based on the location of the completion point, the completion point is inserted into the target missing region to complete the target missing region.

2. The method according to claim 1, characterized in that, Before determining the target missing region of the point cloud data to be completed, the method further includes: The original point cloud data is obtained. Based on the neighborhood distance between each point in the original point cloud data and its corresponding neighboring points, noise points in the original point cloud data are determined and removed from the original point cloud data. The noise points are used to characterize points in the original point cloud data with large neighborhood distances. Using the voxel grid method, all points in the original point cloud data after removing the noise points are downsampled and equalized to obtain downsampled point cloud data with a preset point cloud density. The neighborhood point set of each point in the downsampled point cloud data is determined, and after calculating the covariance matrix of each point based on the neighborhood point set, the eigenvalue decomposition is performed on each covariance matrix, and the eigenvector corresponding to the smallest eigenvalue obtained by decomposition is used as the normal vector of the corresponding point to obtain the point cloud data.

3. The method according to claim 2, characterized in that, Based on the neighborhood distance between each point in the original point cloud data and its corresponding neighboring points, noise points in the original point cloud data are determined, including: Based on the distance between two adjacent points, among all points in the original point cloud data, neighboring points whose distance to the target point is less than a preset radius are detected, thus obtaining multiple neighboring points corresponding to each target point; Calculate the average distance between each target point and its corresponding plurality of neighboring points to obtain the neighborhood distance between each target point and its corresponding neighboring points; It is determined whether the neighborhood distance corresponding to each target point is greater than a neighborhood distance threshold, and if it is determined that the neighborhood distance is greater than the neighborhood distance threshold, the target point is determined as the noise point, wherein the neighborhood distance threshold is determined based on the mean and standard deviation of the neighborhood distance.

4. The method according to claim 1, characterized in that, Identify the target missing regions in the point cloud data to be completed, including: For each point in the point cloud data, a neighborhood search is performed to obtain the second neighborhood point set corresponding to each point. Based on the second neighborhood point set and the neighborhood volume corresponding to the second neighborhood point set, the local density corresponding to each point in the point cloud data is determined. Obtain the normal vector of each point in the point cloud data. Based on the normal vector of each point and each neighboring point in the second neighboring point set, determine the geometric continuity and curvature of each point. The geometric continuity is used to characterize the neighborhood direction difference determined by the angle between the normal vectors of each point and the corresponding neighboring points. After detecting sparse points with local density less than a preset density threshold from the point cloud data, the missing boundary points with neighborhood direction difference greater than a difference threshold or curvature abrupt change are detected in the point cloud data. The target missing region is constructed based on all the sparse points and the missing boundary points.

5. The method according to claim 4, characterized in that, Generating proxy points within the target missing region includes: The target missing region is divided into multiple voxel units, and the center point corresponding to each voxel unit is used as the initial proxy point. The voxel in-voxel point of each voxel unit includes at least one of the sparse point and the missing boundary point. Based on each initial proxy point, proxy points are generated with a preset adaptive step size to generate multiple candidate proxy points. The adaptive step size is used to represent the spacing between proxy points and is dynamically adjusted according to the local density corresponding to the missing boundary point. Based on the distance between two adjacent candidate proxy points, the Poisson disk sampling method is used to filter all candidate proxy points to select those whose distance between two adjacent candidate proxy points is greater than a preset distance threshold, thereby obtaining the proxy points corresponding to the target missing region.

6. The method according to claim 1, characterized in that, In the original point cloud data corresponding to the point cloud data, a neighborhood search is performed on each surrogate point to obtain a first neighborhood point set corresponding to each surrogate point, including: In all the original points of the original point cloud data, a radius neighborhood search centered on the proxy point is performed to obtain multiple candidate neighborhood points; The number of candidate neighbor points is determined, and if the number of candidate neighbor points is less than a set number, K-nearest neighbor search is used to search for the corresponding original points from the original points other than the candidate neighbor points. The original point and multiple candidate neighbor points found by the K-nearest neighbor search are used as the first neighbor point set corresponding to the surrogate point.

7. The method according to claim 6, characterized in that, Based on the geometric features of each neighboring point in the first neighboring point set, enhanced features of each surrogate point are aggregated and generated. Based on these enhanced features, the completion point and its location corresponding to each surrogate point are determined, including: Obtain the geometric features of each neighboring point in the first neighboring point set, wherein the geometric features include the coordinate parameters, normal vector, and local curvature corresponding to the neighboring point; Based on the coordinate parameters and the normal vector, the feature weights corresponding to each neighborhood point are calculated using a weighting function based on distance decay and direction consistency. Based on the feature weights, the geometric features of all the neighborhood points are aggregated and weighted to generate the enhanced features. Using a mapping function that employs linear mapping and surface fitting, the spatial offset corresponding to the enhanced feature is determined. Based on the spatial offset, the random perturbation term corresponding to the spatial offset, and the coordinate parameters corresponding to the surrogate point, spatial coordinate parameters corresponding to the completion point are generated, wherein the location of the completion point includes the spatial coordinate parameters.

8. The method according to claim 6, characterized in that, The feature weights corresponding to each of the neighboring points are calculated using the following formula: Among them, w ij It is the feature weight, p i and p k All are neighborhood points, a j It is an agency point, n i It is point p i The normal vector, n k It is point p k The normal vector, n j Is it agent point A? j The average normal vector of the neighborhood, N(a) j ) is the neighborhood point set, and σ is the scale parameter.

9. The method according to claim 6, characterized in that, Based on the location of the completion point, inserting the completion point into the target missing region includes: Insert the corresponding completion point at the position corresponding to the completion point position within the target missing region; In the target missing region where the completion points are inserted, surface repair is performed using a local geometric reconstruction method to obtain a completed point cloud. The local geometric reconstruction method includes one of the following: moving least squares method and local triangulation method. The completed point cloud is then subjected to smoothing filtering and global normal adjustment correction to obtain the completed point cloud data.

10. A point cloud completion system, comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the point cloud completion method based on the aggregation of surrogate points and neighborhood features as described in any one of claims 1 to 9.

Images