Charger shell screw hole rapid positioning method based on three-dimensional point cloud recognition

Abstract

The application discloses a charger shell screw hole rapid positioning method based on three-dimensional point cloud recognition, comprising: collecting polarized structured light three-dimensional point cloud and preprocessing, establishing a workpiece coordinate system, and obtaining robust point cloud; calculating a homology bar chart in the point cloud neighborhood, screening candidate regions according to a threshold value, generating a candidate mask and a topological confidence; performing principal component analysis and polar coordinate projection on the candidate regions, and outputting local point cloud; projecting the local point cloud and correcting, obtaining a completed point cloud and an implicit field feature; using random sample consensus to find a hole axis, Gaussian mixture fitting to output a center and a hole diameter; fusing topological and geometric scores into joint confidence, and determining whether to enter execution; mapping the center and the axis to a robot coordinate system, compensating and updating parameters in assembly, and outputting results. Through three-dimensional point cloud recognition and multi-stage geometric topological analysis, the application realizes rapid, accurate positioning and assembly adaptive correction of the charger shell screw hole.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud processing and intelligent assembly technology, and in particular to a method for rapid positioning of screw holes in charger housings based on three-dimensional point cloud recognition. Background Technology

[0002] Currently, in the field of electronic product assembly, the rapid positioning of screw holes in charger housings mainly relies on image recognition or conventional 3D point cloud detection methods. Traditional 2D image methods are susceptible to interference from surface reflections, partial occlusions, and complex textures, resulting in insufficient recognition rates and positioning accuracy. With the development of 3D point cloud technology, researchers are increasingly using structured light or laser scanning to acquire 3D information of workpiece surfaces, and then using geometric fitting or template matching methods to detect and locate screw holes. However, existing methods often suffer from localized missing points or noise accumulation in the point cloud when dealing with highly reflective surfaces, leading to blurred screw hole boundaries and difficulty in obtaining reliable positioning results. Conventional point cloud fitting methods often rely on single constraints of geometric features, which can easily lead to misidentification or missed detection when defects or occlusions exist in the point cloud, reducing overall robustness.

[0003] On the other hand, existing methods often rely solely on local geometric information for judgment during the screening and feature extraction of candidate screw hole regions, lacking global constraints on topological structures. This makes it difficult to effectively distinguish screw holes from similar recesses or cavities in complex backgrounds. Even with template matching as a supplement, its reliance on the integrity and similarity of the template still fails to adapt to various changes under actual assembly conditions. More importantly, most existing technologies treat recognition and assembly as two independent processes, lacking an adaptive correction mechanism based on assembly feedback. This prevents real-time correction of positioning deviations caused by coordinate system offsets, clamping errors, or minor hole deformations during assembly, resulting in inconsistent overall accuracy.

[0004] Therefore, how to provide a method for rapid positioning of screw holes in charger housings based on 3D point cloud recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a rapid screw hole localization method for charger housings based on 3D point cloud recognition. This invention fully utilizes multi-angle polarized structured light acquisition, point cloud topology and geometric feature fusion analysis, and robot assembly feedback compensation technology. It details the complete process from point cloud acquisition and preprocessing, screw hole candidate region screening, local point cloud completion and feature generation, to joint confidence assessment and robot coordinate system mapping. This invention ensures the accuracy of screw hole recognition through weighted fusion of topological constraints and geometric consistency scores; improves point cloud integrity under conditions of high reflectivity and local defects by introducing reversible projection completion and implicit field feature generation; and effectively enhances the system's adaptability by combining end-effector force feedback during assembly to achieve compliance compensation and online parameter updates. This invention possesses advantages such as high positioning accuracy, strong robustness, and high automation, significantly improving the assembly efficiency and reliability of charger housings.

[0006] A method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to an embodiment of the present invention includes:

[0007] A three-dimensional point cloud containing spatial coordinates and intensity channels is obtained by using multi-angle polarized structured light. The three-dimensional point cloud is preprocessed and a workpiece coordinate system is established to obtain a robust point cloud.

[0008] A long-lasting homology bar is calculated within the sliding neighborhood of a robust point cloud. Candidate regions for screw holes are selected based on single-ring topology determination and topology threshold, and candidate masks and topology confidence are generated.

[0009] Principal component analysis is performed on the candidate region of screw holes to obtain the initial value of the local principal axis. Polar coordinate projection is performed on the candidate local point cloud in the orthogonal section, the polar sector density histogram is calculated, the candidate mask is updated according to the radius peak value and angular symmetry constraint, and the candidate local point cloud is output.

[0010] The candidate local point cloud is projected as an annular domain map, and invertible projection is used for completion. Bidirectional mapping and cycle consistency correction are performed between the projection and inverse projection to obtain the completed local point cloud and generate implicit field features.

[0011] Based on implicit field characteristics, random sampling consistency is used to solve the screw hole axis and obtain the geometric consistency score. Gaussian mixture fitting is performed on the polar sector density in the cross section orthogonal to the screw hole axis to output the screw hole center coordinates and hole diameter parameters.

[0012] The topological confidence score and geometric consistency score are weighted and fused into a joint confidence score according to a preset coefficient. When the joint confidence score is lower than the threshold, an adaptive re-acquisition or template re-matching process is triggered; otherwise, the localization result execution stage is entered.

[0013] Based on the workpiece coordinate system, the center coordinates of the screw hole, the axis of the screw hole and the diameter parameters are mapped to the robot coordinate system. During the assembly process, the end force is collected to implement compliance compensation, and the topology threshold and polar sector parameters are updated online with a preset learning rate, and the calibrated positioning results are output.

[0014] Optionally, the method of using multi-angle polarized structured light to obtain a three-dimensional point cloud containing spatial coordinates and intensity channels refers to scanning the charger casing with polarized structured light at multiple shooting angles to obtain point cloud data with spatial coordinate information and light intensity information, thereby forming a multi-view three-dimensional point cloud containing information from different perspectives.

[0015] Optionally, the preprocessing of the 3D point cloud refers to denoising, resampling, normal vector estimation, and coordinate normalization of the 3D point cloud.

[0016] Optionally, establishing a workpiece coordinate system and obtaining a robust point cloud refers to determining the spatial position and orientation of the workpiece by fitting the three reference planes of the shell to the three-dimensional point cloud, establishing a workpiece coordinate system, and uniformly mapping the point clouds from each viewpoint to the workpiece coordinate system to obtain a robust point cloud with consistent coordinates and suppressed noise.

[0017] Optionally, the step of calculating a long-maintained homology bar chart within the sliding neighborhood of the robust point cloud, filtering candidate regions for screw holes based on single-ring topology determination and topology threshold, and generating candidate masks and topology confidence scores includes:

[0018] In a robust point cloud, a sliding neighborhood with a fixed radius is constructed for each sampling point to obtain the corresponding neighborhood point set. The neighborhood point set is gradually expanded under different scale parameters to construct a complex sequence based on the distance relationship between points and record the topological changes of the neighborhood at each scale.

[0019] One-dimensional homology is calculated on the complex sequence to obtain information on the generation and disappearance of local loop structures, and a long-lasting homology bar chart is formed, where the bar length represents the duration of the loop structure.

[0020] Perform single-ring topology determination in a long-term homology bar graph. The ring with the longest duration must be unique in the neighborhood and its duration must be no less than a preset topology threshold. If the condition is met, the neighborhood is marked as a candidate neighborhood for screw holes; otherwise, it is discarded.

[0021] For the marked candidate screw hole neighborhood, the corresponding persistent loop point set is extracted to generate the candidate screw hole region, and a candidate mask corresponding to the region position is formed on the point cloud.

[0022] Based on the proportion of the duration of the persistent loop to the neighborhood search scale range, normalization is performed to obtain the topological confidence value between 0 and 1, which is then output together with the candidate mask.

[0023] All candidate regions are spatially merged and redundantly removed to form a set of candidate regions for screw holes, and the corresponding topological confidence is retained.

[0024] Optionally, the step of performing principal component analysis on the candidate screw hole region to obtain initial values ​​of the local principal axes, performing polar coordinate projection on the candidate local point cloud within the orthogonal section, calculating the polar sector density histogram, updating the candidate mask according to the radius peak value and angular symmetry constraints, and outputting the candidate local point cloud includes:

[0025] Select the target region from the candidate region of screw holes, extract the corresponding candidate local point cloud, and record the topological confidence of the target region;

[0026] Calculate the centroid and covariance matrix of the candidate local point cloud, obtain the largest eigenvector, and get the initial value of the local principal axis;

[0027] Construct a cross-sectional plane that passes through the centroid and is orthogonal to the initial value of the local principal axis. Perform a rigid body transformation on the candidate local point cloud and project it onto the cross-sectional plane to obtain the polar radius and polar angle of the projected points.

[0028] The number of angular bins is set to an even number, and the angular resolution and radius resolution are preset. A weighted normalized polar sector density histogram is constructed based on the polar radius and polar angle of the projection point. The weight of each projection point consists of three parts: the topological confidence weight, the intensity channel normalization weight of the projection point, and the radius prior weight around the median of the polar radius of the region. The sector count is normalized according to the sector area.

[0029] Circular smoothing is performed on the polar sector density histogram in the angular dimension. A pre-defined symmetrical discrete smoothing kernel is used to perform weighted averaging of adjacent sectors and symmetrical enhancement processing is performed. The densities of opposing sectors that are 180 degrees apart are averaged to obtain a symmetrically enhanced polar sector density histogram.

[0030] Calculate the radial cumulative density and determine the radial cumulative peak radius on the symmetric enhanced polar sector density histogram, and calculate the angular symmetry index at the radial cumulative peak radius;

[0031] Set the radius range and symmetry threshold. When the radial cumulative peak radius is within the radius range and the angular symmetry is not lower than the symmetry threshold, update the candidate mask as valid and output the candidate local point cloud, the initial value of the local principal axis and the radial cumulative peak radius; otherwise, mark it as invalid and discard it.

[0032] Optionally, the generation of implicit field features includes:

[0033] Obtain candidate local point clouds, initial values ​​of local principal axes, and radial cumulative peak radius; establish a local coordinate system with the centroid of the candidate local point cloud as the origin and the initial values ​​of local principal axes as the normal; and transform the candidate local point cloud into the local coordinate system.

[0034] In a cross-sectional plane orthogonal to the initial value of the local principal axis, the transformed candidate local point cloud is forward-projected and discretized into a ring domain raster according to the preset angular resolution and radius resolution.

[0035] Based on the topological confidence, the intensity channel normalized value of the candidate local point cloud, and the radius prior around the radial cumulative peak radius, weighting coefficients are assigned to the projection points, and the grids of each annular domain map are normalized and accumulated according to the grid area to generate the forward annular domain map.

[0036] Angular circular smoothing and occlusion interpolation are performed on the forward annular domain graph, and symmetrical enhancement is performed on the opposing sectors that are 180 degrees apart to obtain a symmetrically enhanced annular domain graph.

[0037] Inverse projection is performed on the symmetric-enhanced annular map to reconstruct the effective raster into 3D sample points. Depth offset estimation is applied in the local principal axis direction, and the reconstructed samples are fused with the candidate local point cloud to obtain the completed local point cloud.

[0038] The completed local point cloud is projected forward again and compared with the symmetrically enhanced annular map. When the consistency deviation exceeds the preset threshold, the width of the angular smoothing kernel, the depth offset and the weighting coefficient are adjusted and the forward projection, symmetrical enhancement and reverse projection are repeated until the consistency deviation does not exceed the preset threshold.

[0039] A bounded spatial region is constructed around the completed local point cloud, and implicit field features are generated by describing the bounded distance from the spatial location to the reconstructed boundary and the direction of local change.

[0040] Optionally, based on implicit field characteristics, the screw hole axis is solved using random sampling consistency to obtain a geometric consistency score. Gaussian mixture fitting is performed on the polar sector density within a section orthogonal to the screw hole axis to output the screw hole center coordinates and hole diameter parameters, including:

[0041] Load the completed local point cloud, implicit field features, initial values ​​of local principal axes, and radial cumulative peak radius, and extract the direction samples and effective point set for axis fitting;

[0042] The minimum sample is repeatedly randomly selected from the effective point set to form a candidate line. The distance from each point to the candidate line is calculated and the inner point and outer point are divided according to the residual threshold. The proportion of inner points and the average residual are recorded. At the end of the iteration, the candidate line with the highest proportion of inner points and the smallest average residual is selected as the screw hole axis. The geometric consistency score is obtained by combining the proportion of inner points and the average residual.

[0043] Within a cross section orthogonal to the screw hole axis and passing through the local centroid, the completed local point cloud is projected onto the cross section and a polar coordinate grid is established to statistically analyze the polar sector density distribution.

[0044] Gaussian mixture fitting is performed on the density distribution of the polar sector. The number of Gaussian components and the initial mean, variance and weight parameters of each component are set. The expectation-maximization iterative method is adopted to continuously update the weight, mean and variance of each component until the fitting result converges. The principal component that responds most significantly to the ring structure is selected as the target component, and the center position and radius parameters of the circle in the cross section are determined.

[0045] The consistency of the axis, center, and radius is checked, including the deviation constraint from the initial value of the local principal axis, the inclusion constraint with the candidate mask, and the aperture range constraint. If the conditions are not met, the calculation is returned to be re-estimated. If the conditions are met, the screw hole axis, geometric consistency score, screw hole center coordinates, and aperture parameters are output.

[0046] Optionally, the topological confidence score and geometric consistency score are weighted and fused into a joint confidence score according to a preset coefficient. When the joint confidence score is lower than a threshold, an adaptive re-acquisition or template re-matching process is triggered; otherwise, the localization result execution stage is entered, including:

[0047] The topological confidence and geometric consistency scores are received and normalized to the range of 0 to 1 according to a unified standard.

[0048] Set preset weighting coefficients, all of which are in the range of 0 to 1 and sum to 1. Based on the weighting coefficients, perform weighted fusion of topological confidence and geometric consistency scores to obtain joint confidence.

[0049] Set a joint confidence threshold, and determine the joint confidence level. When the joint confidence level is lower than the threshold, enter the low confidence branch; when the joint confidence level is not lower than the threshold, enter the execution branch.

[0050] In the low-confidence branch, trigger adaptive re-collection or template re-matching:

[0051] First, the acquisition parameters, including exposure, polarization angle, projection frequency, and viewing angle, are adjusted to reacquire the point cloud and generate a new, robust point cloud.

[0052] Secondly, select candidate templates from the template library and perform rigid body registration and scale and orientation correction to obtain new candidate masks and candidate local point clouds;

[0053] The screw hole center coordinates, screw hole axis and hole diameter parameters, along with the joint confidence score, are submitted to the next execution stage, and the topological confidence score, geometric consistency score, weighting coefficient, joint confidence score and branch selection results are recorded.

[0054] The beneficial effects of this invention are:

[0055] This invention effectively alleviates the problem of missing points cloud caused by high reflectivity and local defects on the shell surface in existing technologies by introducing multi-angle polarized structured light and point cloud preprocessing methods. By denoising, resampling, and registering the multi-view point clouds, and establishing a coordinate system based on the shell reference surface, robust and uniform point cloud data is obtained, resulting in higher stability and accuracy in identification, thus laying the foundation for accurate selection of screw hole candidate areas.

[0056] This invention combines long-term homology topology analysis and geometric feature constraints of polar coordinate projection in the screening and positioning of candidate screw hole regions. This not only distinguishes screw holes from similar recessed structures in complex backgrounds but also effectively solves the boundary ambiguity problem caused by incomplete point clouds through reversible projection completion and implicit field feature generation. Furthermore, by solving for the screw hole axis using random sampling consistency and combining it with Gaussian mixture fitting, the accurate extraction of the screw hole center coordinates, axis, and hole diameter parameters is ensured, improving robustness under local defect conditions.

[0057] This invention introduces a weighted fusion mechanism of topological confidence and geometric consistency scoring in the verification and assembly execution stages of the identification results. By triggering adaptive re-acquisition or template matching processes through joint confidence determination, the reliability of the results is ensured. During assembly, compliance compensation is implemented using end-effector force feedback, and online updates of topological thresholds and polar sector parameters are achieved, enabling the system to adaptively correct itself based on actual assembly conditions. Therefore, this invention not only improves the accuracy and stability of rapid screw hole positioning but also enhances its adaptability and intelligence in real-world assembly environments. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of the rapid positioning method for screw holes in a charger housing based on three-dimensional point cloud recognition proposed in this invention;

[0060] Figure 2 This is a schematic diagram of the reversible projection completion and implicit field feature generation of the fast positioning method for screw holes in charger housing based on three-dimensional point cloud recognition proposed in this invention. Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0062] refer to Figure 1 and Figure 2A method for rapid positioning of screw holes in a charger housing based on 3D point cloud recognition includes:

[0063] A three-dimensional point cloud containing spatial coordinates and intensity channels is obtained by using multi-angle polarized structured light. The three-dimensional point cloud is preprocessed and a workpiece coordinate system is established to obtain a robust point cloud.

[0064] A long-lasting homology bar is calculated within the sliding neighborhood of a robust point cloud. Candidate regions for screw holes are selected based on single-ring topology determination and topology threshold, and candidate masks and topology confidence are generated.

[0065] Principal component analysis is performed on the candidate region of screw holes to obtain the initial value of the local principal axis. Polar coordinate projection is performed on the candidate local point cloud in the orthogonal section, the polar sector density histogram is calculated, the candidate mask is updated according to the radius peak value and angular symmetry constraint, and the candidate local point cloud is output.

[0066] The candidate local point cloud is projected as an annular domain map, and invertible projection is used for completion. Bidirectional mapping and cycle consistency correction are performed between the projection and inverse projection to obtain the completed local point cloud and generate implicit field features.

[0067] Based on implicit field characteristics, random sampling consistency is used to solve the screw hole axis and obtain the geometric consistency score. Gaussian mixture fitting is performed on the polar sector density in the cross section orthogonal to the screw hole axis to output the screw hole center coordinates and hole diameter parameters.

[0068] The topological confidence score and geometric consistency score are weighted and fused into a joint confidence score according to a preset coefficient. When the joint confidence score is lower than the threshold, an adaptive re-acquisition or template re-matching process is triggered; otherwise, the localization result execution stage is entered.

[0069] Based on the workpiece coordinate system, the center coordinates of the screw hole, the axis of the screw hole and the diameter parameters are mapped to the robot coordinate system. During the assembly process, the end force is collected to implement compliance compensation, and the topology threshold and polar sector parameters are updated online with a preset learning rate, and the calibrated positioning results are output.

[0070] In this embodiment, the method of using multi-angle polarized structured light to obtain a three-dimensional point cloud containing spatial coordinates and intensity channels refers to scanning the charger housing with polarized structured light at multiple shooting angles to obtain point cloud data with spatial coordinate information and light intensity information, thereby forming a multi-view three-dimensional point cloud containing information from different perspectives.

[0071] In this embodiment, the preprocessing of the 3D point cloud refers to denoising, resampling, normal vector estimation, and coordinate normalization of the 3D point cloud.

[0072] In this embodiment, establishing a workpiece coordinate system and obtaining a robust point cloud means determining the spatial position and orientation of the workpiece by fitting the three reference planes of the shell to the three-dimensional point cloud, establishing a workpiece coordinate system, and uniformly mapping the point clouds from each viewpoint to the workpiece coordinate system to obtain a robust point cloud with consistent coordinates and suppressed noise.

[0073] In this embodiment, the step of calculating a long-maintained homology bar chart within the sliding neighborhood of the robust point cloud, filtering candidate regions for screw holes based on single-ring topology determination and topology threshold, and generating candidate masks and topology confidence scores includes:

[0074] In a robust point cloud, a sliding neighborhood with a fixed radius is constructed for each sampling point to obtain the corresponding neighborhood point set. The neighborhood point set is gradually expanded under different scale parameters to construct a complex sequence based on the distance relationship between points and record the topological changes of the neighborhood at each scale.

[0075] One-dimensional homology is calculated on the complex sequence to obtain information on the generation and disappearance of local loop structures, and a long-lasting homology bar chart is formed, where the bar length represents the duration of the loop structure. Specifically, the calculation of one-dimensional homology on the complex sequence involves:

[0076] Based on candidate point clouds and their neighborhoods, a simplex complex sequence with different radii is constructed sequentially. At each step, points whose distance does not exceed the current radius are grouped into a simplex, and the complex structure is gradually expanded as the radius increases.

[0077] At each scale of the complex sequence, the generation of new loops between point sets and their filling or disappearance as the scale increases are tracked, and the radii corresponding to the first appearance and filling or disappearance of each loop structure are recorded respectively.

[0078] The generation and disappearance intervals of all ring structures are represented by bars. The starting point of the bar represents the scale at which the ring structure is first generated, and the ending point represents the scale at which the ring structure is filled or disappears. A long-term homology bar chart is drawn to reflect the stability of the topological structure of the candidate region.

[0079] In a sustained homology bar graph, a single-loop topology determination is performed, requiring that the longest-lasting loop is uniquely present in its neighborhood, and its duration is not less than a preset topology threshold. If the condition is met, the neighborhood is marked as a candidate neighborhood for screw holes; otherwise, it is discarded. Specifically, the single-loop topology determination in the sustained homology bar graph is as follows:

[0080] Analyze the long-lasting homology bar chart within each sliding neighborhood, count the duration of all ring structures, and select the ring with the longest duration as the main ring;

[0081] Determine whether the main loop is unique among all loop structures, i.e., it has the longest duration and no other loop structure has the same duration.

[0082] Check if the duration of the main loop is greater than or equal to the preset topology threshold. If the only main loop meets the duration threshold requirement, mark the neighborhood as a candidate neighborhood for screw holes; otherwise, remove the neighborhood.

[0083] For the marked candidate screw hole neighborhood, the corresponding persistent loop point set is extracted to generate the candidate screw hole region, and a candidate mask corresponding to the region position is formed on the point cloud.

[0084] Based on the proportion of the duration of the persistent loop to the neighborhood search scale range, normalization is performed to obtain the topological confidence value between 0 and 1, which is then output together with the candidate mask.

[0085] All candidate regions are spatially merged and redundantly removed to form a set of candidate regions for screw holes, and the corresponding topological confidence is retained.

[0086] This invention constructs a complex sequence based on the distance relationship between points within a robust point cloud sliding neighborhood, combined with persistent homology analysis, to accurately track the entire process of local loop structure generation and disappearance, and visually reflects the stability of topological features using a bar chart. Unlike traditional screw hole initial screening that relies solely on geometric or density features, this invention innovatively introduces single-loop topological determination, requiring the longest-lasting main loop to be uniquely present in the neighborhood and its duration not less than a set threshold. This effectively avoids misjudgments caused by local defects, reflective interference, or multiple adjacent holes, improving the accuracy and robustness of candidate region selection. By normalizing the duration of persistent loops, the topological confidence is quantified, providing a stable and reliable topological criterion for multi-source feature fusion and result determination. Finally, through spatial merging and redundancy removal, an accurate set of screw hole candidate regions is formed, laying a solid foundation for high-precision positioning in practical assembly scenarios such as point cloud deficiencies and complex backgrounds, and improving the recognition efficiency and anti-interference capability of automated assembly systems.

[0087] In this embodiment, the step of performing principal component analysis on the candidate screw hole region to obtain the initial value of the local principal axis, performing polar coordinate projection on the candidate local point cloud in the orthogonal section, calculating the polar sector density histogram, updating the candidate mask according to the radius peak value and angular symmetry constraint, and outputting the candidate local point cloud includes:

[0088] Select the target region from the candidate region of screw holes, extract the corresponding candidate local point cloud, and record the topological confidence of the target region;

[0089] Calculate the centroid and covariance matrix of the candidate local point cloud to obtain the largest eigenvector and the initial value of the local principal axis. Specifically, the calculation of the centroid and covariance matrix of the candidate local point cloud is as follows:

[0090] The centroid of the point cloud in the screw hole candidate region is calculated by averaging the spatial coordinates of all points in the candidate region and used to represent the center position of the point cloud.

[0091] Using the centroid as a reference, calculate the coordinate deviation of all points in the candidate region relative to the centroid, calculate the covariance of each coordinate component, and construct a three-dimensional covariance matrix.

[0092] Perform eigenvalue decomposition on the covariance matrix and select the eigenvector corresponding to the largest eigenvalue as the initial value of the local principal axis;

[0093] A cross-sectional plane passing through the centroid and orthogonal to the initial value of the local principal axis is constructed. The candidate local point cloud is then subjected to rigid body transformation and projected onto the cross-sectional plane to obtain the polar radius and polar angle of the projected point. The construction of the cross-sectional plane passing through the centroid and orthogonal to the initial value of the local principal axis means that in the candidate local point cloud, with the calculated centroid as the center point of the cross-sectional plane, a spatial plane perpendicular to the direction of the initial value of the region's principal axis is selected as the reference plane for intercepting and analyzing the candidate local point cloud. The reference plane is used to unfold the point cloud data in the orthogonal direction of the principal axis.

[0094] The number of angular bins is set to an even number, and the angular resolution and radius resolution are preset. A weighted normalized polar sector density histogram is constructed based on the polar radius and polar angle of the projection point. The weight of each projection point consists of three parts: the topological confidence weight, the intensity channel normalization weight of the projection point, and the radius prior weight around the median of the polar radius of the region. The sector count is normalized according to the sector area.

[0095] A circular smoothing process is performed on the polar sector density histogram along the angular dimension. A weighted average of adjacent sectors is calculated using a pre-defined symmetrical discrete smoothing kernel, followed by symmetrical enhancement processing. The densities of opposing sectors differing by 180 degrees are averaged to obtain a symmetrically enhanced polar sector density histogram. Specifically, the symmetrical enhancement processing involves:

[0096] For each sector in the polar sector density histogram, find the opposite sector that is 180 degrees away from the sector, and the density values ​​corresponding to polar coordinate angles α and α+180 degrees respectively.

[0097] Calculate the average density of the sector and the opposite sector, and replace the original density values ​​of the two sectors with the obtained average value to enhance the symmetry of the density distribution of the opposite sector.

[0098] The operation was repeated for all sectors, and finally a symmetrically enhanced polar sector density histogram was obtained, which provides a more robust density feature for radius peak extraction and circular structure analysis.

[0099] Calculate the radial cumulative density and determine the radial cumulative peak radius on the symmetrically enhanced polar sector density histogram, and calculate the angular symmetry index at the radial cumulative peak radius, specifically:

[0100] For the polar sector density histogram after symmetry enhancement, the density values ​​of all sectors corresponding to each radius are accumulated to obtain the radial cumulative density. By traversing all radius intervals in turn, the radial cumulative density curve is obtained.

[0101] In the radial cumulative density curve, find the radius with the largest cumulative density value, and use it as the radial cumulative peak radius, which represents the most likely radius position of the screw hole;

[0102] Based on the radial cumulative peak radius, the density distribution of all sectors under the statistical radius is analyzed, and the densities of the opposite sectors are compared. The average or variance of the density difference is calculated as an index of angular symmetry.

[0103] Set the radius range and symmetry threshold. When the radial cumulative peak radius is within the radius range and the angular symmetry is not lower than the symmetry threshold, update the candidate mask as valid and output the candidate local point cloud, the initial value of the local principal axis and the radial cumulative peak radius; otherwise, mark it as invalid and discard it.

[0104] This invention innovatively integrates rigid body transformation based on principal axis analysis, polar coordinate projection, and multi-weighted polar sector density histogram modeling to address the 3D spatial distribution characteristics of local point clouds in screw holes. This achieves highly robust recognition of circular features under complex working conditions. Through annular smoothing and symmetry enhancement processing, the ability of the polar sector density distribution to suppress reflections, defects, and noise is improved. Furthermore, by combining radial cumulative density and angular symmetry as joint criteria, stable extraction of circular structural features of screw holes is ensured, effectively avoiding misjudgments and missed detections that are common with traditional single density or geometric threshold methods. The introduction of multiple normalized weights enables the system to adaptively balance topology, intensity, and prior information, improving the targeting and generalization of feature extraction. Finally, a validity judgment mechanism based on range and symmetry is employed to further filter invalid regions, ensuring the reliability and accuracy of the output candidate results and improving the accuracy of screw hole positioning and the system's environmental adaptability in automated assembly scenarios.

[0105] In this embodiment, the generation of implicit field features includes:

[0106] Obtain candidate local point clouds, initial values ​​of local principal axes, and radial cumulative peak radius; establish a local coordinate system with the centroid of the candidate local point cloud as the origin and the initial values ​​of local principal axes as the normal; and transform the candidate local point cloud into the local coordinate system.

[0107] Within a cross-sectional plane orthogonal to the initial value of the local principal axis, the transformed candidate local point cloud is forward-projected and discretized into a ring-shaped raster according to a preset angular resolution and radius resolution. Specifically, the forward projection of the transformed candidate local point cloud involves:

[0108] Using the calculated initial value of the local principal axis as the normal vector, construct a cross-sectional plane orthogonal to the normal vector, and map the candidate local point cloud into the cross-sectional plane through rigid body transformation;

[0109] Using the centroid of the point cloud in the cross-sectional plane as the origin of polar coordinates, calculate the radial distance from each point to the centroid and the angle with the reference direction, and convert the three-dimensional coordinates of all points into a two-dimensional point set represented by polar coordinates;

[0110] According to the preset angular resolution and radius resolution, the polar coordinate system is divided into several annular and sector grids. The number of points or weighted density in each grid is counted to form the corresponding annular domain map. The preset angular resolution is 10 degrees and the radius resolution is 0.2 millimeters.

[0111] Based on the topological confidence, the intensity channel normalized value of the candidate local point cloud, and the radius prior around the radial cumulative peak radius, weighting coefficients are assigned to the projection points, and the grids of each annular domain map are normalized and accumulated according to the grid area to generate the forward annular domain map.

[0112] Angular toroidal smoothing and occlusion interpolation are performed on the forward annular domain graph, and symmetrical enhancement is applied to opposing sectors that are 180 degrees apart, resulting in a symmetrically enhanced annular domain graph, specifically:

[0113] In the angular dimension of the forward loop domain graph, a weighted average is performed on each sector and its adjacent sectors using a preset smoothing kernel to achieve loop smoothing and reduce the impact of extreme outliers.

[0114] For sectors with missing data or extremely low density in the annular domain graph, the occluded or missing areas are filled by neighborhood interpolation to restore the continuous density distribution.

[0115] For all sectors, find the opposite sector that is 180 degrees away from the circumference, calculate the average density of the current sector and the opposite sector, and replace the original density values ​​of both with the average density to achieve symmetrical enhancement processing, and finally obtain the symmetrically enhanced annular domain map.

[0116] Inverse projection is performed on the symmetric-enhanced annular map to reconstruct the effective grid into 3D sample points. Depth offset estimation is applied in the local principal axis direction. The reconstructed samples are fused with the candidate local point cloud to obtain the completed local point cloud. The depth offset estimation refers to the spatial depth value assigned to each inverse projection reconstructed point according to the local principal axis direction.

[0117] The completed local point cloud is projected forward again and compared with the symmetrically enhanced annular map. When the consistency deviation exceeds the preset threshold, the width of the angular smoothing kernel, the depth offset and the weighting coefficient are adjusted and the forward projection, symmetrical enhancement and reverse projection are repeated until the consistency deviation does not exceed the preset threshold.

[0118] A bounded spatial region is constructed around the completed local point cloud, and implicit field features are generated by describing the bounded distance from the spatial location to the reconstructed boundary and the direction of local change.

[0119] This invention addresses the issues of missing or partially occluded point clouds in screw holes under complex working conditions. It innovatively proposes a forward-inverse reversible projection completion method based on a local polar coordinate annular domain graph. Through rigid body transformation and orthogonal section mapping along the principal axes, it achieves efficient 2D flattening and fine rasterization of candidate local point clouds. The method integrates multiple weighting mechanisms, including topological confidence, intensity normalization, and radius prior, to enhance the expressive power of the forward annular domain graph. Angular annular smoothing, occlusion interpolation, and symmetry enhancement effectively suppress noise and compensate for data gaps caused by point cloud occlusion or reflection, ensuring the density consistency of circular features. Inverse projection combined with depth offset estimation achieves spatial completion of point clouds in missing regions. Dynamic parameter adaptive optimization using cyclic consistency constraints significantly improves the realism and robustness of the completed point cloud. Finally, implicit field features are generated around the completed point cloud, improving the accuracy of circular structure recognition.

[0120] In this embodiment, based on implicit field characteristics, the screw hole axis is solved using random sampling consistency to obtain a geometric consistency score. Gaussian mixture fitting is performed on the polar sector density within a cross-section orthogonal to the screw hole axis to output the screw hole center coordinates and hole diameter parameters, including:

[0121] Load the completed local point cloud, implicit field features, initial values ​​of local principal axes, and radial cumulative peak radius, and extract the direction samples and effective point set for axis fitting;

[0122] The minimum sample is repeatedly randomly selected from the effective point set to form a candidate line. The distance from each point to the candidate line is calculated and the inner point and outer point are divided according to the residual threshold. The proportion of inner points and the average residual are recorded. At the end of the iteration, the candidate line with the highest proportion of inner points and the smallest average residual is selected as the screw hole axis. The geometric consistency score is obtained by combining the proportion of inner points and the average residual.

[0123] Within a cross section orthogonal to the screw hole axis and passing through the local centroid, the completed local point cloud is projected onto the cross section and a polar coordinate grid is established to statistically analyze the polar sector density distribution.

[0124] Gaussian mixture fitting is performed on the density distribution of the polar sector. The number of Gaussian components and the initial mean, variance, and weight parameters of each component are set. An expectation-maximization iterative method is used to continuously update the weight, mean, and variance of each component until the fitting result converges. The principal component that has the most significant response to the ring structure is selected as the target component, and the center position and radius parameters of the circle in the cross section are determined.

[0125] The consistency of the axis, center, and radius is checked, including deviation constraints from the initial local principal axis value, inclusion constraints with the candidate mask, and aperture range constraints. If these constraints are not met, the estimation is returned to the previous step. If they are met, the screw hole axis, geometric consistency score, screw hole center coordinates, and aperture parameters are output, where:

[0126] Deviation constraint refers to comparing the direction of the screw hole axis with the initial value of the local spindle and calculating the angle between the two. When the angle is less than the preset threshold, the axis direction is considered to meet the requirements of the initial value of the spindle. Otherwise, it is judged that the deviation exceeds the limit and needs to be re-estimated.

[0127] The inclusion constraint refers to comparing the spatial position of the circular region determined by the center and radius of the fitted circle with the candidate mask. When the circular region is completely contained within the candidate mask or the overlap ratio between the two exceeds a preset threshold, it is determined that the inclusion constraint is satisfied; otherwise, it needs to be re-estimated.

[0128] Aperture range constraint refers to the interval judgment of the fitted aperture parameters. When the aperture value is within the preset allowable range of the physical size of the screw hole, it is determined that the aperture range constraint is met; otherwise, it needs to be re-estimated.

[0129] This invention addresses the problem of point cloud recognition and localization of screw holes in charger housings being susceptible to missing data, noise, and structural disturbances. It innovatively introduces a multi-stage joint solution mechanism involving directional sample extraction of the completed point cloud, random sampling consistency fitting, and Gaussian mixture modeling. Through repeated random sampling and residual internal / external point discrimination, robust fitting of the screw hole axis is achieved, significantly improving robustness against outliers and local occlusion. Gaussian mixture fitting and principal component extraction of the extreme sector density distribution greatly enhance the accurate reconstruction of the center and aperture parameters under complex contours. A unique triple consistency check of deviation constraints, inclusion constraints, and aperture range constraints effectively prevents misjudgments caused by fitting errors and false candidates, significantly improving the reliability of the output geometric parameters. This not only overcomes the dependence on complete point clouds and regular structures in traditional methods but also adapts to point cloud quality fluctuations under various working conditions, improving the accuracy and success rate of screw hole localization in automated assembly scenarios and ensuring the system's effectiveness in real-world complex environments.

[0130] In this embodiment, the topological confidence score and geometric consistency score are weighted and fused into a joint confidence score according to a preset coefficient. When the joint confidence score is lower than a threshold, an adaptive re-acquisition or template re-matching process is triggered; otherwise, the localization result execution stage is entered, including:

[0131] The topological confidence and geometric consistency scores are received and normalized to the range of 0 to 1 according to a unified standard.

[0132] Set preset weighting coefficients, all of which are in the range of 0 to 1 and sum to 1. Based on the weighting coefficients, perform weighted fusion of topological confidence and geometric consistency scores to obtain joint confidence.

[0133] Set a joint confidence threshold, and determine the joint confidence level. When the joint confidence level is lower than the threshold, enter the low confidence branch; when the joint confidence level is not lower than the threshold, enter the execution branch.

[0134] In the low-confidence branch, trigger adaptive re-collection or template re-matching:

[0135] First, the acquisition parameters, including exposure, polarization angle, projection frequency, and viewing angle, are adjusted to reacquire the point cloud and generate a new, robust point cloud.

[0136] Secondly, select candidate templates from the template library and perform rigid body registration and scale and orientation correction to obtain new candidate masks and candidate local point clouds;

[0137] The screw hole center coordinates, screw hole axis and hole diameter parameters, along with the joint confidence score, are submitted to the next execution stage, and the topological confidence score, geometric consistency score, weighting coefficient, joint confidence score and branch selection results are recorded.

[0138] This invention innovatively constructs a joint confidence criterion for screw hole positioning tasks by introducing a normalized weighted fusion of topological confidence and geometric consistency scores, enabling dynamic evaluation of point cloud data reliability. This method not only fully utilizes the complementary advantages of multi-source criteria but also achieves adaptive closed-loop correction of point cloud acquisition and geometric template matching by setting low-confidence and execution branches. For recognition uncertainties caused by reflections, occlusions, or defects in complex scenarios, it can dynamically adjust the acquisition and matching process to ensure the system consistently outputs highly reliable positioning results. This improves environmental adaptability and recognition robustness in automated assembly, reduces rework rates, and ensures stable system operation and assembly efficiency.

[0139] In this embodiment, the process of mapping the screw hole center coordinates, screw hole axis, and hole diameter parameters from the workpiece coordinate system to the robot coordinate system, collecting end-effector forces during assembly to implement compliance compensation, and updating the topology threshold and polar sector parameters online with a preset learning rate, and outputting the calibrated positioning results includes:

[0140] Receive the obtained screw hole center coordinates, screw hole axis and hole diameter parameters, and receive the output joint confidence level;

[0141] Based on the calibration relationship between the workpiece coordinate system and the robot coordinate system, the center coordinates, axis and aperture parameters are mapped to the robot coordinate system to generate the corresponding assembly pose and process parameters;

[0142] The assembly posture and process parameters are sent to the robot control system to drive the actuator to complete the alignment and tightening actions.

[0143] During the assembly process, end force signals are collected to make online micro-adjustments to the assembly posture and implement compliance compensation to eliminate deviations during the assembly process.

[0144] Based on the data collected during the assembly process, the topology threshold and polar sector parameters are adjusted online according to the preset update step size. The execution results and joint confidence are recorded, and the calibrated positioning results are output. Example

[0145] To verify the feasibility of this invention in practice, the research team applied it to an automated assembly workshop, selecting a charger casing of a certain model as the experimental object. The outer surface of this casing is made of high-gloss plastic, which is prone to specular reflection and localized scattering, leading to insufficient accuracy in screw hole recognition based on traditional two-dimensional images or single three-dimensional point cloud acquisition. Especially in mass production, common assembly deviations manifest as screw hole position misidentification, incorrect hole diameter judgment, and misjudgments caused by missing local point clouds. These problems cause alignment failures in automated screw-tightening robotic arms, resulting in production line stoppages or rework. To verify the effectiveness of the method of this invention, the team conducted a three-month assembly experiment at an electronics manufacturing plant from April to June 2025.

[0146] The experimental setup includes a multi-angle polarization structured light scanning system, an assembly workstation equipped with a six-degree-of-freedom robotic arm, and a data acquisition and analysis server. During the experiment, three-dimensional point cloud data of the shell was first acquired using multi-angle polarization structured light. This data contains spatial coordinates and intensity channel information. After denoising, resampling, and normalization preprocessing, a robust point cloud was obtained by establishing a workpiece coordinate system. Subsequently, a long-lasting homology bar graph was calculated within the sliding neighborhood of the point cloud. Single-ring topology judgment and topology thresholding were used to filter candidate regions for screw holes, thereby generating candidate masks and topology confidence scores.

[0147] Principal component analysis is further performed within the candidate region to obtain initial values ​​of the local principal axes. Then, polar coordinate projection is performed on the orthogonal cross section to calculate the polar sector density histogram. Unlike traditional methods, this invention introduces dynamic symmetry constraints and a radius peak comparison mechanism to effectively eliminate pseudo-hole structures caused by reflection. For local point cloud regions with missing data, this invention employs a reversible projection completion method, performing bidirectional mapping and cyclic consistency correction between projection and inverse projection to obtain the completed local point cloud and generate implicit field features. Finally, the random sampling consistency method is used to solve for the screw hole axis, and a Gaussian mixture model is combined to fit the cross-sectional density distribution of the screw hole, outputting the screw hole center coordinates and hole diameter parameters.

[0148] During the assembly phase, this invention obtains a joint confidence score through a weighted fusion of topological confidence and geometric consistency scores. When the joint confidence score falls below a set threshold, the system automatically triggers an adaptive re-acquisition or template re-matching process to ensure the reliability of the results. When the joint confidence score reaches or exceeds the threshold, the screw hole parameters are mapped to the robot coordinate system, and the robotic arm uses this information for positioning. During the assembly process, the robotic arm's end effector collects force feedback data in real time and fine-tunes the pose using a compliance compensation strategy. Simultaneously, the system dynamically updates the topological threshold and polar sector parameters based on the feedback.

[0149] Table 1. Comparative experimental results of the method of the present invention and the traditional method under different working conditions.

[0150]

[0151] Based on the data in Table 1, it is clear that the method of the present invention exhibits significantly superior performance compared to traditional methods under various working conditions. Under ordinary surface conditions, the screw hole recognition accuracy of the method of the present invention reaches 99.1%, a significant improvement compared to 93.2% of the traditional method. The average positioning error is reduced from 0.16mm to 0.05mm, the assembly success rate increases from 94.1% to 99.5%, the average time per hole is shortened to 2.0s, and the rework rate is significantly reduced to 0.6%. This demonstrates that the method of the present invention can guarantee high precision and high efficiency under normal conditions.

[0152] In environments with highly reflective surfaces, the traditional method's recognition accuracy drops to 85.7%, the average positioning error rises to 0.21 mm, and both assembly success rate and stability decrease significantly, with a rework rate increasing to 5.4%. In contrast, the method of this invention maintains a high recognition accuracy of 96.9%, an average positioning error controlled at 0.07 mm, an assembly success rate of 98.8%, and a rework rate of only 1.1%. This fully demonstrates the robustness advantage of this invention under strong reflective interference, effectively overcoming the problems of misidentification and insufficient positioning accuracy caused by surface reflection.

[0153] Under conditions of localized surface defects, traditional methods achieve an accuracy rate of only 88.6%, an assembly success rate of 93.5%, and a rework rate of 4.7%, demonstrating disadvantages such as sensitivity to defects and insufficient stability. The method of this invention, however, maintains a high accuracy rate of 98.2% under the same conditions, with an average positioning error of only 0.06 mm, an assembly success rate of 99.2%, and a rework rate reduced to 0.8%, demonstrating its advantage of efficiently completing rapid screw hole positioning and assembly even under defective conditions.

[0154] The method of this invention is significantly superior to traditional methods in three typical working conditions: normal, strong reflection, and local defects. It not only ensures higher recognition accuracy and assembly success rate, but also greatly reduces the average positioning error and rework rate, while improving work efficiency, demonstrating its practicality and reliability in complex assembly environments.

[0155] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for rapid positioning of screw holes in a charger housing based on 3D point cloud recognition, characterized in that, include: A three-dimensional point cloud containing spatial coordinates and intensity channels is obtained by using multi-angle polarized structured light. The three-dimensional point cloud is preprocessed and a workpiece coordinate system is established to obtain a robust point cloud. A long-lasting homology bar is calculated within the sliding neighborhood of a robust point cloud. Candidate regions for screw holes are selected based on single-ring topology determination and topology threshold, and candidate masks and topology confidence are generated. Principal component analysis is performed on the candidate region of screw holes to obtain the initial value of the local principal axis. Polar coordinate projection is performed on the candidate local point cloud in the orthogonal section, the polar sector density histogram is calculated, the candidate mask is updated according to the radius peak value and angular symmetry constraint, and the candidate local point cloud is output. The candidate local point cloud is projected as an annular domain map, and invertible projection is used for completion. Bidirectional mapping and cycle consistency correction are performed between the projection and inverse projection to obtain the completed local point cloud and generate implicit field features. Based on implicit field characteristics, random sampling consistency is used to solve the screw hole axis and obtain the geometric consistency score. Gaussian mixture fitting is performed on the polar sector density in the cross section orthogonal to the screw hole axis to output the screw hole center coordinates and hole diameter parameters. The topological confidence score and geometric consistency score are weighted and fused into a joint confidence score according to a preset coefficient. When the joint confidence score is lower than the threshold, an adaptive re-acquisition or template re-matching process is triggered; otherwise, the localization result execution stage is entered. Based on the workpiece coordinate system, the center coordinates of the screw hole, the axis of the screw hole and the diameter parameters are mapped to the robot coordinate system. During the assembly process, the end force is collected to implement compliance compensation, and the topology threshold and polar sector parameters are updated online with a preset learning rate, and the calibrated positioning results are output.

2. The method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to claim 1, characterized in that, The method of acquiring a three-dimensional point cloud containing spatial coordinates and intensity channels using multi-angle polarized structured light refers to scanning the charger casing with polarized structured light at multiple shooting angles to obtain point cloud data with spatial coordinate information and light intensity information, thus forming a multi-view three-dimensional point cloud containing information from different perspectives.

3. The method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to claim 1, characterized in that, The preprocessing of the 3D point cloud refers to the denoising, resampling, normal vector estimation, and coordinate normalization of the 3D point cloud.

4. The method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to claim 1, characterized in that, The establishment of the workpiece coordinate system and the acquisition of robust point cloud refers to determining the spatial position and orientation of the workpiece by fitting the three reference planes of the shell to the three-dimensional point cloud, establishing the workpiece coordinate system, and uniformly mapping the point clouds from each viewpoint to the workpiece coordinate system to obtain a robust point cloud with consistent coordinates and suppressed noise.

5. The method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to claim 1, characterized in that, The process of calculating a long-maintained homology bar chart within the sliding neighborhood of a robust point cloud, filtering candidate regions for screw holes based on single-ring topology determination and topology threshold, and generating candidate masks and topology confidence scores includes: In a robust point cloud, a sliding neighborhood with a fixed radius is constructed for each sampling point to obtain the corresponding neighborhood point set. The neighborhood point set is gradually expanded under different scale parameters to construct a complex sequence based on the distance relationship between points and record the topological changes of the neighborhood at each scale. One-dimensional homology is calculated on the complex sequence to obtain information on the generation and disappearance of local loop structures, and a long-lasting homology bar chart is formed, where the bar length represents the duration of the loop structure. Perform single-ring topology determination in a long-term homology bar graph. The ring with the longest duration must be unique in the neighborhood and its duration must be no less than a preset topology threshold. If the condition is met, the neighborhood is marked as a candidate neighborhood for screw holes; otherwise, it is discarded. For the marked candidate screw hole neighborhood, the corresponding persistent loop point set is extracted to generate candidate screw hole regions, and a candidate mask corresponding to the region position is formed on the point cloud. Based on the proportion of the duration of the persistent loop to the neighborhood search scale range, normalization is performed to obtain the topological confidence value between 0 and 1, which is then output together with the candidate mask. All candidate regions are spatially merged and redundantly removed to form a set of candidate regions for screw holes, and the corresponding topological confidence is retained.

6. The method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to claim 1, characterized in that, The process of performing principal component analysis on the candidate screw hole region to obtain initial values ​​of the local principal axes, performing polar coordinate projection on the candidate local point cloud within the orthogonal section, calculating the polar sector density histogram, updating the candidate mask according to the radius peak value and angular symmetry constraints, and outputting the candidate local point cloud includes: Select the target region from the candidate region of screw holes, extract the corresponding candidate local point cloud, and record the topological confidence of the target region; Calculate the centroid and covariance matrix of the candidate local point cloud, obtain the largest eigenvector, and get the initial value of the local principal axis; Construct a cross-sectional plane that passes through the centroid and is orthogonal to the initial value of the local principal axis. Perform a rigid body transformation on the candidate local point cloud and project it onto the cross-sectional plane to obtain the polar radius and polar angle of the projected points. The number of angular bins is set to an even number, and the angular resolution and radius resolution are preset. A weighted normalized polar sector density histogram is constructed based on the polar radius and polar angle of the projection point. The weight of each projection point consists of three parts: the topological confidence weight, the intensity channel normalization weight of the projection point, and the radius prior weight around the median of the polar radius of the region. The sector count is normalized according to the sector area. Circular smoothing is performed on the polar sector density histogram in the angular dimension. A pre-defined symmetrical discrete smoothing kernel is used to perform weighted averaging of adjacent sectors and symmetrical enhancement processing is performed. The densities of opposing sectors that are 180 degrees apart are averaged to obtain a symmetrically enhanced polar sector density histogram. Calculate the radial cumulative density and determine the radial cumulative peak radius on the symmetric enhanced polar sector density histogram, and calculate the angular symmetry index at the radial cumulative peak radius; Set the radius range and symmetry threshold. When the radial cumulative peak radius is within the radius range and the angular symmetry is not lower than the symmetry threshold, update the candidate mask as valid and output the candidate local point cloud, the initial value of the local principal axis and the radial cumulative peak radius; otherwise, mark it as invalid and discard it.

7. The method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to claim 1, characterized in that, The generated implicit field features include: Obtain candidate local point clouds, initial values ​​of local principal axes, and radial cumulative peak radius; establish a local coordinate system with the centroid of the candidate local point cloud as the origin and the initial values ​​of local principal axes as the normal; and transform the candidate local point cloud into the local coordinate system. In a cross-sectional plane orthogonal to the initial value of the local principal axis, the transformed candidate local point cloud is forward-projected and discretized into a ring domain raster according to the preset angular resolution and radius resolution. Based on the topological confidence, the intensity channel normalized value of the candidate local point cloud, and the radius prior around the radial cumulative peak radius, weighting coefficients are assigned to the projection points, and the grids of each annular domain map are normalized and accumulated according to the grid area to generate the forward annular domain map. Angular circular smoothing and occlusion interpolation are performed on the forward annular domain graph, and symmetrical enhancement is performed on the opposing sectors that are 180 degrees apart to obtain a symmetrically enhanced annular domain graph. Inverse projection is performed on the symmetric-enhanced annular map to reconstruct the effective raster into 3D sample points. Depth offset estimation is applied in the local principal axis direction, and the reconstructed samples are fused with the candidate local point cloud to obtain the completed local point cloud. The completed local point cloud is projected forward again and compared with the symmetrically enhanced annular map. When the consistency deviation exceeds the preset threshold, the width of the angular smoothing kernel, the depth offset and the weighting coefficient are adjusted and the forward projection, symmetrical enhancement and reverse projection are repeated until the consistency deviation does not exceed the preset threshold. A bounded spatial region is constructed around the completed local point cloud, and implicit field features are generated by describing the bounded distance from the spatial location to the reconstructed boundary and the direction of local change.

8. The method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to claim 1, characterized in that, Based on implicit field characteristics, the screw hole axis is solved using random sampling consistency to obtain a geometric consistency score. Gaussian mixture fitting is performed on the polar sector density within a section orthogonal to the screw hole axis to output the screw hole center coordinates and hole diameter parameters, including: Load the completed local point cloud, implicit field features, initial values ​​of local principal axes, and radial cumulative peak radius, and extract the direction samples and effective point set for axis fitting; The minimum sample is repeatedly randomly selected from the effective point set to form a candidate line. The distance from each point to the candidate line is calculated and the inner point and outer point are divided according to the residual threshold. The proportion of inner points and the average residual are recorded. At the end of the iteration, the candidate line with the highest proportion of inner points and the smallest average residual is selected as the screw hole axis. The geometric consistency score is obtained by combining the proportion of inner points and the average residual. Within a cross section orthogonal to the screw hole axis and passing through the local centroid, the completed local point cloud is projected onto the cross section and a polar coordinate grid is established to statistically analyze the polar sector density distribution. Gaussian mixture fitting is performed on the density distribution of the polar sector. The number of Gaussian components and the initial mean, variance and weight parameters of each component are set. The expectation-maximization iterative method is adopted to continuously update the weight, mean and variance of each component until the fitting result converges. The principal component that responds most significantly to the ring structure is selected as the target component, and the center position and radius parameters of the circle in the cross section are determined. The consistency of the axis, center, and radius is checked, including the deviation constraint from the initial value of the local principal axis, the inclusion constraint with the candidate mask, and the aperture range constraint. If the conditions are not met, the calculation is returned to be re-estimated. If the conditions are met, the screw hole axis, geometric consistency score, screw hole center coordinates, and aperture parameters are output.

9. The method for rapid positioning of screw holes in a charger housing based on three-dimensional point cloud recognition according to claim 1, characterized in that, The topological confidence score and geometric consistency score are weighted and fused into a joint confidence score according to a preset coefficient. When the joint confidence score is lower than a threshold, an adaptive re-acquisition or template re-matching process is triggered; otherwise, the localization result execution stage is entered, including: The topological confidence and geometric consistency scores are received and normalized to the range of 0 to 1 according to a unified standard. Set preset weighting coefficients, all of which are in the range of 0 to 1 and sum to 1. Based on the weighting coefficients, perform weighted fusion of topological confidence and geometric consistency scores to obtain joint confidence. Set a joint confidence threshold, and determine the joint confidence level. When the joint confidence level is lower than the threshold, enter the low confidence branch; when the joint confidence level is not lower than the threshold, enter the execution branch. In the low-confidence branch, trigger adaptive re-collection or template re-matching: First, the acquisition parameters, including exposure, polarization angle, projection frequency, and viewing angle, are adjusted to reacquire the point cloud and generate a new, robust point cloud. Secondly, select candidate templates from the template library and perform rigid body registration and scale and orientation correction to obtain new candidate masks and candidate local point clouds; The screw hole center coordinates, screw hole axis and hole diameter parameters, along with the joint confidence score, are submitted to the next execution stage, and the topological confidence score, geometric consistency score, weighting coefficient, joint confidence score and branch selection results are recorded.

Images