Panoramic three-dimensional measurement method and system for high-reflective aviation components based on reference and single binocular fusion

Abstract

This invention discloses a panoramic 3D measurement method and system for highly reflective aerospace components based on reference objects and monocular / binocular fusion, belonging to the field of 3D measurement and machine vision technology. The method includes the following steps: constructing a multi-view measurement system composed of a projector and a camera, and rigidly fixing the aerospace component to a stereo reference object with a globally unique ID; performing multi-view scanning by changing the relative pose of the measurement system and the measured assembly, obtaining robust phase information at each viewpoint using a multi-step phase-shifting method with pixel-by-pixel saturation frame culling, and reconstructing local point clouds using a pixel-level monocular / binocular point cloud fusion algorithm; simultaneously identifying and reconstructing the 3D coordinates of feature points on the reference object, and using the Kabsch algorithm to achieve high-precision alignment between the local point cloud and the global coordinate system; further repairing point cloud noise using an adaptive optimization algorithm based on normal vector noise evaluation, and finally fusing all aligned point clouds to obtain a complete high-fidelity 3D model of the aerospace component. This invention effectively solves the problems of missing data on highly reflective surfaces and difficulty in stitching together complex geometric features, achieving fully automatic high-precision 3D measurement without spraying or manual marking, significantly improving measurement efficiency and automation level.

Description

Technical Field

[0001] This invention relates to the field of precision three-dimensional measurement technology for aerospace components, and more specifically, to a panoramic three-dimensional measurement method and system for highly reflective aerospace components based on reference objects and monocular / binocular fusion. Background Technology

[0002] Aerospace components (such as aero-engine blades and composite panel panels) are core components that determine the aerodynamic performance, structural strength, and safety reliability of aircraft. These components generally possess complex free-form surfaces, exhibiting extreme geometric characteristics such as large twist angles, thin walls, and variable cross-sections, placing extremely stringent requirements on machining accuracy and surface quality. Micrometer-level shape deviations or subsurface micro-defects can disrupt aerodynamic design, leading to decreased efficiency, increased vibration, and even structural fatigue failure. Therefore, achieving high-precision, high-efficiency, and non-destructive full-field measurement of the three-dimensional morphology of aerospace components has become an indispensable key link in their precision manufacturing, online inspection, and service maintenance.

[0003] Among numerous 3D measurement technologies, structured light measurement is widely regarded as an ideal choice for industrial precision measurement due to its advantages such as non-contact operation, full-field coverage, high efficiency, and high precision. This method projects an coded light pattern onto the measured surface, analyzes its deformation after being modulated by the surface morphology, and achieves 3D reconstruction by combining it with the principles of triangulation. However, when directly applied to aerospace components, it still faces two inherent technical bottlenecks: first, overexposure and loss of phase information caused by highly reflective surfaces, resulting in gaps in point cloud data; second, the complex geometric features of the components and the limited overlapping areas between multiple viewpoints cause traditional point cloud stitching algorithms to fail.

[0004] While existing structured light measurement methods offer advantages such as non-contact operation, high efficiency, and high precision, they also have significant limitations. When dealing with high reflectivity, traditional methods involving the application of matte developer introduce additional steps, contaminate the workpiece, and introduce measurement errors due to coating thickness, violating the principles of non-destructive and in-situ measurement. When addressing splicing challenges, solutions relying on high-precision turntables require extremely high equipment accuracy and clamping consistency, resulting in poor system flexibility. Adding manual markings undermines the automation and efficiency of the measurement process. Furthermore, existing technologies often address reflectivity or splicing issues in isolation, lacking a systematic solution. For example, patent CN113091647B does not solve the splicing problem, and patent CN108151668A does not address the acquisition of complete point clouds under high reflectivity conditions, failing to simultaneously resolve the industry-wide problems of incomplete measurement and inaccurate splicing. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a panoramic 3D measurement method and system for highly reflective aerospace components based on reference objects and monocular / binocular fusion. The method constructs a measurement system consisting of a projector and two cameras, deconstructing it into a binocular structured photonics system and two monocular structured photonics systems. During measurement, the component and a globally unique 3D reference object are rigidly fixed to maintain their relative pose. Multi-view scanning is performed by changing the relative pose of the measurement system and the measured assembly. In each viewpoint, a robust phase is obtained using a multi-step phase-shifting method with pixel-by-pixel saturation frame culling. Then, a pixel-level correspondence-based monocular / binocular point cloud fusion algorithm is used to integrate the advantages of the three subsystems, reconstructing a complete and high-precision local point cloud of the component. Simultaneously, the 3D coordinates of feature points on the reference object are identified and reconstructed. Using their globally unique IDs, the Kabsch algorithm is used to calculate the rigid body transformation matrix from the current local point cloud to the global coordinate system for precise alignment. Finally, the local point clouds transformed to the global coordinate system from all views are merged to obtain a complete 3D model of the component. This invention overcomes high reflectivity by "redundant information acquisition and intelligent fusion" and achieves precise stitching without markers by "fixed reference objects and absolute coordinate mapping".

[0006] To achieve the above objectives, this invention provides a panoramic 3D measurement method for highly reflective aerospace components based on reference objects and monocular / binocular fusion, comprising the following steps: S100: Build a binocular structured light measurement system, divide the system into one binocular structured light system and two monocular structured light systems, and calibrate them separately; S200: The absolute phase table is preprocessed by a multi-step phase shift method that removes saturated frames pixel by pixel; S300: Uses a pre-processed phase table for monocular and binocular reconstruction, and establishes point-to-point mapping relationships through a 3D point correspondence retrieval algorithm in the pixel dimension; S400: Calculates affine transformation through point-to-point mapping relationship, searches for points that failed to match during stereo matching of binocular structured light system, performs monocular reconstruction and applies the affine transformation to perform mono- and binocular fusion to remove residual specular highlights, and denoises the edge noise of the fused point cloud. S500: Create a reference object for the size of the object being measured, calibrate the reference object, and simultaneously photograph the object being measured and the reference object during measurement. Change the relative spatial pose between the measurement system and the object being measured, measure again, and align the obtained point cloud to the calibrated point cloud to achieve complete point cloud reconstruction.

[0007] Further, the binocular structured light measurement system in step S100 includes: a digital projector, an industrial camera, a synchronous triggering device, and a computing unit. The industrial camera is symmetrically arranged on both sides of the digital projector, and the fields of view of the three are fully overlapped by adjustment. The binocular structured light measurement system is logically divided into three subsystems, including: a monocular subsystem A consisting of camera 1 and projector, a monocular subsystem B consisting of camera 2 and projector, and a binocular subsystem C consisting of camera 1, projector, and camera 2.

[0008] Furthermore, the multi-step phase shifting method in step S200 employs at least a six-step phase shifting sequence; the pixel-by-pixel removal of saturated frames specifically involves: for each pixel, identifying and removing image frames whose grayscale values ​​have reached saturation, and using only the remaining unsaturated frame data for phase calculation to obtain the absolute phase of the pixel.

[0009] Furthermore, the specific process of the 3D point correspondence algorithm described in step S300 includes: S310: The monocular and binocular subsystems are calibrated separately. The absolute phase table is obtained by decoding using a multi-step phase shift method with pixel-by-pixel saturation frame removal and stereo correction is performed to provide a geometrically consistent phase data basis for subsequent matching. S320: Perform stereo matching in the binocular subsystem to obtain the matching point set; use phase recovery and distortion correction to simultaneously derive the matching point sets corresponding to the two monocular subsystems and establish a one-to-one correspondence among the three. S330: For pixel areas not covered by binocular matching, the unmatched points in the monocular subsystem are retrieved through a reverse matching strategy to fill in the blind spots and highlight areas of the binocular field of view. S340: Based on the above matching point set, five sets of point clouds are reconstructed using the disparity-depth formula (binocular) and the projection equation (monocular), respectively, including the binocular reference point cloud, two corresponding monocular point clouds, and two completed point clouds; S350: To eliminate the scale error caused by multi-system calibration, the progressive threshold iteration method is used to calculate the affine transformation matrix from monocular point cloud to binocular reference point cloud, and the transformation accuracy is optimized by gradually eliminating noise points. S360: The affine transformation obtained by completing the point cloud is fused with the binocular reference point cloud to form a complete and seamless local point cloud, effectively filling the point cloud gaps caused by high reflectivity and blind spots.

[0010] Furthermore, the affine transformation described in step S400 is calculated based on the point-to-point mapping relationship established in step S300, and is used to unify the scale of the monocular reconstructed point cloud and the binocular reference point cloud; the pixel regions not covered by the binocular stereo matching are retrieved through the inverse matching strategy, and the region is reconstructed by monoculars to obtain the completed point cloud. The completed point cloud is then fused with the binocular reference point cloud after the affine transformation to remove the remaining highlights and fill the blind spots in the field of view.

[0011] Furthermore, the step of denoising the edge noise of the fused point cloud in step S400 includes: S410: Based on the preliminary fused point cloud, the surface normal vector of each point is estimated through local neighborhood principal component analysis, and the standard deviation of the angle between the normal vectors in the neighborhood is calculated. Points exceeding a preset threshold are marked as high-noise points, thereby locating areas with poor reconstruction quality. S420: For marked high-noise points, retrieve the original reconstructed points of their corresponding coordinates in the two monocular subsystems as candidates, and re-evaluate the consistency of the normal vectors of the local region after replacement; if any candidate point can significantly reduce the noise index, then replace the original noise point with that monocular point to achieve data optimization based on viewpoint advantage. S430: If the noise still exceeds the threshold after replacement, the measurement data at that location is deemed invalid, and the noise point is removed to form a local void. An interpolation algorithm based on radial basis functions is used to smoothly interpolate based on the geometric features of the surrounding reliable area, thereby achieving continuous surface reconstruction of the void area.

[0012] Furthermore, the reference object mentioned in step S500 is a three-dimensional structure with an AprilTag code bearing a globally unique ID pasted on its surface. The size of the AprilTag code is adapted to the size of the three-dimensional structure of the reference object, and the overall size of the reference object is determined according to the size of the aircraft component being measured. During the measurement process, the reference object and the aircraft component being measured are rigidly fixed to ensure that their relative pose remains unchanged.

[0013] Furthermore, the specific process of the reference object calibration and alignment step in step S500 is as follows: S510: Uses a binocular system to capture stereo reference objects with AprilTag codes from multiple perspectives, extracts and reconstructs the 3D coordinates of feature points from each perspective, and simultaneously establishes a stable association between feature points and their unique IDs; S520: Select a point cloud from a certain viewpoint as the global coordinate system reference, use the Kabsch algorithm to solve the rigid body transformation using common ID feature points, gradually align all viewpoint point clouds, and form an initial global 3D model of the reference object feature points; S530: Using the initial model as the initial value, a weighted distance residual optimization function is constructed. Nonlinear optimization is performed by combining the observation confidence of feature points in different images to improve the accuracy and consistency of the global coordinates of feature points and obtain a high-precision reference model. S540: In actual measurement, each scan synchronously acquires the component point cloud and reference object feature points. The current feature point set is matched with the calibrated global model by the feature point ID, the optimal rigid body transformation matrix is ​​calculated, the component point cloud is transformed to a unified global coordinate system, and finally all viewpoint point clouds are fused to obtain a complete 3D model.

[0014] Furthermore, the formula for the weighted distance residual optimization function mentioned in step S530 is: , In the formula, , The total number of images, The total number of feature points on the reference object. Indicates the first If you see the first one on the picture If there are 1 points, the value is 1; otherwise, it is 0. The confidence level for this point is determined by the sharpness of the top corner of the tag. , Indicates the first The rotation matrix and translation vector corresponding to the image. In the The first image observed The original 3D coordinates of each point In the reference coordinate system, the first... The coordinates of the corner points to be optimized; the point cloud data optimized using this optimization function is used as the final data of the global 3D model of the reference object feature points, denoted as . .

[0015] A panoramic 3D measurement system for highly reflective aerospace components based on reference object and monocular / binocular fusion, employing the aforementioned panoramic 3D measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion, includes: The system configuration and calibration module is used to build a hardware system including a projector and at least two cameras, and logically divide it into a binocular structured photonics system and two monocular structured photonics modules, and complete the calibration of each subsystem. The high reflectivity suppression phase processing module is used to decode the image sequence acquired by the hardware system using a multi-step phase shift method that removes saturated frames pixel by pixel to obtain a robust absolute phase table in order to suppress the loss of phase information in high reflectivity areas. The monocular and binocular point cloud fusion and reconstruction module is used to perform binocular stereo matching and monocular reconstruction based on the absolute phase table, establish point-to-point mapping through a 3D point correspondence retrieval algorithm of pixel dimension, calculate the affine transformation from monocular reconstructed point cloud to binocular reference point cloud, perform monocular completion and fusion on the binocular matching failure area, and generate a local complete point cloud of the tested component. The reference-based global registration module is used to process three-dimensional reference objects that are rigidly fixed to the component and have a globally unique identifier. This module includes a reference object feature point calibration unit for reconstructing a high-precision global three-dimensional coordinate model of the reference object feature points; and a multi-view point cloud alignment unit for synchronously identifying and reconstructing reference object feature points in each scan, and calculating the rigid body transformation matrix from the current local point cloud to the global coordinate system through feature point ID matching and the Kabsch algorithm, so as to achieve accurate alignment and stitching of multi-view point clouds. The point cloud refinement and optimization module is used to perform quality assessment and repair on the fused point cloud. This module locates high-noise points through local normal vector consistency analysis and uses monocular reconstruction candidate points for competitive replacement. For areas that cannot be repaired, an interpolation algorithm based on radial basis functions is used for smooth completion, and finally outputs a complete and smooth 3D panoramic model of the components.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The method of this invention constructs a two-level anti-high-brightness mechanism of "preprocessing (pixel-by-pixel saturation frame culling) + core processing (mono-dual-eye fusion)". In the preprocessing stage, effective phase information is preserved to the maximum extent. On this basis, through pixel-level 3D point correspondence retrieval and affine registration algorithm, disordered point cloud registration is transformed into point-to-point registration, which effectively eliminates the scale error caused by multi-system calibration and achieves sub-pixel level fusion accuracy, thereby realizing high-fidelity reconstruction of highly reflective aerospace components.

[0017] 2. The method of the present invention uses local normal vector consistency as a quality index, and uses a data-driven adaptive optimization algorithm to competitively select and replace between binocular and monocular point clouds, and performs intelligent interpolation on regions that cannot be repaired. Combined with radial basis function (RBF) surface reconstruction technology, it effectively suppresses noise in the edge region of the fused point cloud and ensures that the surface is smooth and continuous.

[0018] 3. The method of this invention designs a stereo reference object with a globally unique ID identifier, establishes a mapping system of "feature point ID ↔ three-dimensional absolute coordinates", transforms the multi-view stitching problem into a rigid body transformation solution based on absolute coordinates, avoids the ambiguity and cumulative error of feature matching, and achieves high-precision panoramic three-dimensional reconstruction with consistent absolute coordinates without the need for a high-precision turntable.

[0019] 4. The system of this invention employs a two-stage anti-high-brightness mechanism of pixel-by-pixel saturation frame culling and monocular / dual-eye fusion, effectively suppressing phase loss and data holes on highly reflective surfaces. By introducing a stereo reference object with a globally unique ID, multi-view stitching is transformed into a rigid body transformation solution based on absolute coordinates, achieving high-precision, automated panoramic alignment without the need for a high-precision turntable or manual marking. Combined with local normal vector evaluation and competitive point cloud repair strategies, the smoothness and integrity of the point cloud at edges and in steep tilt areas are significantly improved. High-fidelity, high-precision, and fully automated 3D reconstruction is achieved in the precision measurement of aerospace components, overcoming the problems of traditional methods such as reliance on spraying, difficult stitching, and low efficiency. Attached Figure Description

[0020] Figure 1 This is an overall technical roadmap of the panoramic measurement method for highly reflective aerospace components according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the hardware configuration and subsystem division of the measurement system according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a three-dimensional reference object with AprilTag codes of different IDs pasted on its surface, according to an embodiment of the present invention. Figure 4 This is a schematic diagram of a multi-view scanning implementation scheme according to an embodiment of the present invention; Figure 5 This is a flowchart of the 3D point correspondence retrieval and monocular / binocular fusion algorithm based on pixel dimension according to an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the noise generated in the edge region during single- and binocular point cloud fusion according to an embodiment of the present invention; Figure 7 This is a flowchart of the adaptive point cloud refinement process based on local geometric feature feedback, according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail 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 invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] Figure 1This is the overall technical roadmap of the panoramic three-dimensional topography measurement method for highly reflective aerospace components described in this invention. The scheme uses a measurement system composed of dual black-and-white cameras and a projector. During the measurement process, the high reflectivity problem is first solved by combining the proposed multi-step phase-shifting method of removing saturated frames pixel by pixel with a single- and binocular structured light fusion algorithm. Then, while maintaining relative stillness between the component and the reference object, the system extracts global features from the reference object while changing the observation perspective to acquire the component's point cloud. Based on the unique identifier and coordinate mapping algorithm of the reference object's feature points, the point clouds obtained from multiple measurements are aligned, thus completing the overall measurement of the aerospace component. This method effectively overcomes the problems of incomplete and inaccurate measurements caused by the high reflectivity of the aerospace component's surface material, as well as the difficulty in stitching due to its geometric features, ensuring the global accuracy of the measurement results. Specifically, it includes the following steps: S100: Build a binocular structured light measurement system, divide the system into one binocular structured light system and two monocular structured light systems, and calibrate them separately; like Figure 2 As shown, the measurement system hardware includes: a digital projector, two industrial cameras, a synchronization triggering device, and a computing unit. The two cameras are symmetrically arranged on either side of the projector, and adjustments are made to ensure full overlap of their fields of view. Logically, the system is divided into three subsystems: a monocular subsystem A consisting of camera 1 and the projector; a monocular subsystem B consisting of camera 2 and the projector; and a binocular subsystem C consisting of camera 1, the projector, and camera 2. The key to solving the high reflectivity problem in this invention is the use of a redundant phase-shift method to remove pixel saturation frames for initial high reflectivity removal, followed by the use of monocular and binocular structured light fusion to remove remaining highlights.

[0023] The premise of this invention for solving panoramic 3D measurement is to maintain the geometric consistency between the component and the reference object, and then change the relative spatial pose between the measurement system and the object being measured to perform multiple scans and extract the feature points of the reference object for alignment. Figure 3 This invention demonstrates a three-dimensional reference object embodiment with AprilTag codes of different IDs pasted on its surface. AprilTag codes of the same size but different IDs are pasted on the six faces of a cube. The size of the cube and the AprilTag codes are made according to the size of the component being measured. The figure shows a scheme in which AprilTag codes of size 50*50mm are pasted on the six faces of a 200*50*50mm cube, which is suitable for components with a length of about 200mm.

[0024] Regarding specific solutions for changing the relative spatial pose between the measurement system and the object being measured, the present invention provides the following two methods: using a robot end effector to move the measurement system; and fixing the object being measured and a reference object on a turntable to achieve rotation of the measured composition. Figure 4The diagrams show the two multi-view scanning schemes. Either the robot or the rotating platform can be selected, and the algorithm does not require any motion accuracy from either of them.

[0025] S200: The absolute phase table is preprocessed by a multi-step phase shift method that removes saturated frames pixel by pixel; Solving the problem of high reflectivity involves two steps: first, preprocessing is performed to remove some of the highlights; In the N-step PSP algorithm, let the coordinates of a point in the projector space be... The total number of steps is ,make Then the first Step the light intensity Represented as: , In the formula For ambient light intensity, Modulate the light intensity for the projector. Phase and light intensity at this point This can be confirmed by taking a picture with a camera. Given that there are three unknowns in the equation, at least three equations are needed to solve it. This corresponds to three unsaturated frames. When a portion of the image captured by the camera is overexposed, it manifests as... The grayscale value mapped to the camera is theoretically greater than 255, but it is truncated to 255, resulting in... The solution is incorrect, but if overexposed information (pixels with grayscale values ​​close to 255) is removed during decoding, and only the unexposed points are used for the solution, the solution can still be obtained. For ease of explanation, we simplify the light intensity equation by letting... And by performing a cosine expansion, the above equation becomes: , After removing overexposure information Zhang Image Construct the matrix equation: , Solve the equation using the least squares method. and The relative phase at that point can then be determined. : , The accuracy of phase recovery depends on having sufficient independent sampled data (at least three unsaturated sampled values) after removing saturated frames. The conventional four-step phase-shifting method exhibits weak robustness when dealing with highly reflective metallic surfaces: if a specific pixel is overexposed for two frames, it cannot be decoded due to insufficient effective information. To ensure phase recovery quality, observation redundancy needs to be introduced by increasing the number of phase-shifting steps; in practical measurements, at least a six-step phase-shifting sequence is used.

[0026] S300: Uses a pre-processed phase table for monocular and binocular reconstruction, and establishes point-to-point mapping relationships through a 3D point correspondence retrieval algorithm in the pixel dimension; Due to the strong reflective properties of aerospace components, the aforementioned preprocessing methods cannot completely eliminate the influence of extremely overexposed areas, but they can resolve the decoding of slightly overexposed areas where blades overlap during observation by two cameras. Considering that the projector is a surface light source, the reflective characteristics of its generated highlight areas vary significantly under different viewing angles, resulting in a strong viewpoint dependence of overexposed areas in the images captured by each camera in a binocular system.

[0027] Based on the aforementioned physical characteristics, this invention deconstructs the measurement system into one binocular subsystem and two monocular subsystems. The two monocular subsystems act as different observation perspectives and perform binocular-monocular fusion to address the problem of reconstructing remaining overexposed areas. Given that the binocular subsystem has a longer measurement baseline and the highest point cloud reconstruction accuracy, this algorithm uses its generated point cloud as a registration benchmark and a bridge connecting the various monocular subsystems for point cloud fusion. It leverages the high-precision benchmark provided by the binocular subsystem while utilizing the advantages of the monocular subsystem's wide field of view and multiple observation perspectives. This effectively compensates for the view-dependent holes generated by the binocular subsystem when facing strong reflections. Furthermore, an edge smoothing algorithm is proposed to address the edge noise problem in binocular-monocular fusion scenarios.

[0028] To achieve efficient fusion of heterogeneous point clouds under a binocular structured light architecture, this study proposes a pixel-level 3D point correspondence retrieval algorithm. This algorithm aims to transform the conventional point cloud registration problem into a non-point-to-point registration problem to solve for the transformation parameters of spatial alignment between subsystems. By establishing rigorous geometric correspondence constraints at the pixel level, this method overcomes the limitations of traditional 3D spatially disordered point cloud registration, skipping complex feature extraction and descriptor matching calculations. It simplifies the registration process to the direct solution of linear algebraic equations. In scenarios such as aerospace components with smooth surfaces, sparse features, and severe reflection, the point-to-point mapping based on physical phase exhibits higher reliability compared to conventional feature matching algorithms. The flowchart of the algorithm is shown below. Figure 5 As shown.

[0029] The specific process is as follows: S310: The monocular and binocular subsystems are calibrated separately. The absolute phase table is obtained by decoding using a multi-step phase shift method with pixel-by-pixel saturation frame removal and stereo correction is performed to provide a geometrically consistent phase data basis for subsequent matching. The monocular and binocular subsystems were calibrated, with the monocular subsystem uniformly employing an inverse camera model. During measurement, a high-step fringe pattern was projected and simultaneously acquired by both cameras. A preprocessed absolute phase table was obtained by pixel-by-pixel removal of saturated frames and decoding using unexposed pixels. Subsequently, a correction mapping table was calculated using calibration parameters, and stereo correction was performed on the phase table to ensure the speed of subsequent matching.

[0030] S320: Perform stereo matching in the binocular subsystem to obtain the matching point set; use phase recovery and distortion correction to simultaneously derive the matching point sets corresponding to the two monocular subsystems and establish a one-to-one correspondence among the three. On the corrected phase map, stereo matching is performed on the binocular subsystem C. That is, for each pixel on camera 1, the sub-pixel coordinates corresponding to that pixel are matched in the same row on camera 2 to obtain the matching point set. Next, for the point set... For each coordinate on the screen, the original coordinates are obtained by inputting the calibration mapping table calculated in step S310, and the phase of the point is read from the uncorrected absolute phase table. The coordinates of the matching points on the projector are recovered from the phase. The calibration parameters are used to correct the distortion of these points, resulting in the matching point sets for monocular subsystems A and B respectively. and Both of these point sets are passed through Therefore and The matching points on are all with The matching points on the graph have a one-to-one correspondence.

[0031] S330: For pixel areas not covered by binocular matching, the unmatched points in the monocular subsystem are retrieved through a reverse matching strategy to fill in the blind spots and highlight areas of the binocular field of view. Pair set The coordinates of camera 1 are obtained by taking the complement of the entire image pixel set and repeating step S320. The operation resulted in the failure of camera 1 to complete the matching of the set of matching points on the binocular subsystem C. For point sets The calculation first requires obtaining The coordinates of camera 2 are in the complement of the entire image pixel set, but Since the coordinates of camera 2 are sub-pixel coordinates, it is difficult to directly perform complement operations on them. Therefore, this invention adopts a reverse matching strategy: for each pixel on camera 2, the corresponding sub-pixel coordinates of that pixel are matched in the same row on camera 1. The complement of the points that can be matched is then used to calculate the sub-pixel coordinates. .

[0032] S340: Based on the above matching point set, five sets of point clouds are reconstructed using the disparity-depth formula (binocular) and the projection equation (monocular), respectively, including the binocular reference point cloud, two corresponding monocular point clouds, and two completed point clouds; The five sets of matching points calculated in steps S320 and S330 are reconstructed. The calculation formulas are differentiated based on whether stereo correction has been performed on the matching points: the binocular subsystem uses the disparity-depth formula for reconstruction; the stereo-corrected monocular subsystem uses the projection equation for reconstruction. Five sets of point clouds are obtained, among which... and The generated point clouds are respectively and ,and Generated point cloud There is a one-to-one correspondence. and Generated point cloud and This involves reconstructing the blind spots and highlight areas of the binocular subsystem.

[0033] S350: To eliminate the scale error caused by multi-system calibration, the progressive threshold iteration method is used to calculate the affine transformation matrix from monocular point cloud to binocular reference point cloud, and the transformation accuracy is optimized by gradually eliminating noise points. To eliminate scale differences in multi-system calibration, the calculation was performed from... and arrive The affine transformation matrix is ​​used instead of the rigid body transformation matrix. To suppress the influence of noise on the calculation results, a progressive threshold iteration method is adopted: first, the initial transformation and residuals are calculated using all point pairs; a threshold is set and noise points with residuals exceeding the threshold are removed; the transformation relationship is recalculated using the remaining points; the residual threshold is lowered and the above operation is repeated until the calculated transformation makes the residuals of all points lower than the minimum threshold. The result is obtained by calculating the affine transformation matrix instead of the rigid body transformation matrix. and arrive Affine transformation and .

[0034] S360: The affine transformation obtained by completing the point cloud is fused with the binocular reference point cloud to form a complete and seamless local point cloud, effectively filling the point cloud gaps caused by high reflectivity and blind spots.

[0035] right and Apply respectively and Afterwards, with By mixing, a complete local point cloud of the object under test is obtained, which fills in the high-light holes and the blind spots of the C field of view of the binocular subsystem.

[0036] S400: Calculates affine transformation through point-to-point mapping relationship, searches for points that failed to match during stereo matching of binocular structured light system, performs monocular reconstruction and applies the affine transformation to perform mono- and binocular fusion to remove residual specular highlights, and denoises the edge noise of the fused point cloud. Considering that the binocular subsystem can only reconstruct the overlapping area of ​​the two cameras, while the monocular subsystem can reconstruct the entire successfully decoded area in the field of view, resulting in a larger reconstruction area for the latter, binocular and monocular structured light fusion, while solving the specular highlight problem, also further completes the point cloud not appearing in the shared field of view of the two cameras. However, because the binocular structured light system experiences a significant decrease in the signal-to-noise ratio of the projected fringes captured by one camera at the edge of its field of view or in areas with large surface tilt angles, leading to a significant increase in noise in these areas, it will... Figure 6 The edge region shown exhibits an issue of uneven point cloud merging.

[0037] To address the issues of point cloud unsmoothness and noise surge at the edges of the binocular subsystem C's field of view, this paper proposes an adaptive reconstruction strategy based on field-of-view region division. The binocular reconstruction results are retained in the stable central region of the binocular common field of view, while point cloud reconstruction is performed using a monocular subsystem in quality-degraded areas such as the edges of the binocular field of view or areas with large surface tilt angles. To this end, an adaptive optimization method based on local geometric feature feedback is proposed. The core of this method lies in using the normal vector deviation as a noise measure to dynamically evaluate and correct the point cloud quality. The flowchart is shown below. Figure 7 As shown.

[0038] The specific process is as follows: S410: Based on the preliminary fused point cloud, the surface normal vector of each point is estimated through local neighborhood principal component analysis, and the standard deviation of the angle between the normal vectors in the neighborhood is calculated. Points exceeding a preset threshold are marked as high-noise points, thereby locating areas with poor reconstruction quality. First, the k-neighborhood search algorithm is used to construct the preliminary spatial topological relationships of the fused point cloud, and then the surface normal vector of each sampling point is estimated based on the principal component analysis (PCA) method. Subsequently, the point cloud is traversed and the standard deviation of the angle between each sampling point and the normal vector in its neighborhood is calculated. If the indicator exceeds the preset maximum noise threshold, mark the point.

[0039] S420: For marked high-noise points, retrieve the original reconstructed points of their corresponding coordinates in the two monocular subsystems as candidates, and re-evaluate the consistency of the normal vectors of the local region after replacement; if any candidate point can significantly reduce the noise index, then replace the original noise point with that monocular point to achieve data optimization based on viewpoint advantage. For the marked high-noise points, the algorithm retrieves the original reconstruction points generated by the two monocular subsystems at the corresponding coordinates as correction candidate points. These points already have a one-to-one correspondence during monocular and binocular fusion. Then, the monocular reconstruction points are substituted into the original local neighborhood structure, and the standard deviation of the angle between the local normal vectors is recalculated. If any monocular point can make If the value drops significantly below the preset maximum threshold, then three-dimensional point replacement is performed, using the visual advantage of the monocular system to compensate for the accuracy disadvantage of the binocular system.

[0040] S430: If the noise still exceeds the threshold after replacement, the measurement data at that location is deemed invalid, and the noise point is removed to form a local void. An interpolation algorithm based on radial basis functions is used to smoothly interpolate based on the geometric features of the surrounding reliable area, thereby achieving continuous surface reconstruction of the void area.

[0041] The algorithm follows the principle of "prioritizing measurement data fidelity." If, after the trial replacement in step S420, the standard deviation of the normal vector angle still exceeds the threshold, it is determined that the original measurement feature at that location has completely failed due to occlusion or system noise. In this case, the noisy point is removed, forming a local topological hole. For these removed areas, a point cloud completion algorithm based on radial basis function (RBF) interpolation is invoked to smoothly fill the hole using the curvature features of the surrounding high signal-to-noise ratio region. This mechanism ensures the global geometric continuity of the aerospace component while achieving a smooth transition and refined representation of the edge region.

[0042] S500: Create a reference object for the size of the object being measured, calibrate the reference object, and simultaneously photograph the object being measured and the reference object during measurement. Change the relative spatial pose between the measurement system and the object being measured, measure again, and align the obtained point cloud to the calibrated point cloud to achieve complete point cloud reconstruction.

[0043] This invention employs a three-dimensional reference object with a globally unique ID, AprilTag, pasted on its surface to assist in measurement. AprilTag, with its internal asymmetric encoding characteristics, achieves highly reliable feature point identification and accurate pose estimation, effectively solving the problem of erroneous matching with traditional symmetrical markers.

[0044] Measurement using a reference object first requires calibration of the reference object. Considering the large number of AprilTag codes used by the reference object, printing high-precision AprilTag codes would significantly increase production costs. Therefore, this invention provides a scheme for calibrating reference objects made with low-precision codes using a pre-calibrated binocular system. The specific process for reference object calibration and subsequent alignment steps is as follows: S510: Uses a binocular system to capture stereo reference objects with AprilTag codes from multiple perspectives, extracts and reconstructs the three-dimensional coordinates of feature points from each perspective, and simultaneously establishes a stable association between feature points and their unique IDs. A binocular system is used to photograph reference objects in different poses to ensure that the binocular system can simultaneously capture different surfaces from multiple perspectives. Then, feature points are extracted from the images of the two cameras and 3D reconstruction is performed to obtain multiple sets of local point clouds of the reference objects. Utilizing the asymmetric encoding characteristics of AprilTag, the association record between feature points and their unique IDs is established simultaneously in this process.

[0045] S520: Select a point cloud from a certain viewpoint as the global coordinate system reference, use the common ID feature points to solve the rigid body transformation through the Kabsch algorithm, gradually align all viewpoint point clouds, and form an initial global 3D model of the reference object feature points. In the acquired multiple sets of local point clouds, one set is designated as the global reference coordinate system, and the point clouds from the remaining views are used as the point clouds to be registered. By searching for common feature point pairs with the same ID as those in the already registered set, the optimal rigid body transformation matrix is ​​solved using the Kabsch algorithm, and the point clouds to be registered are incorporated into the global base coordinate system. The above steps are repeated until the point clouds from all views are spatially aligned. Finally, for the same feature point observed repeatedly from multiple views, the arithmetic mean of its spatial coordinate observations is calculated and used as the initial data for the global 3D model of the reference feature point.

[0046] S530: Using the initial model as the initial value, a weighted distance residual optimization function is constructed. Nonlinear optimization is performed by combining the observation confidence of feature points in different images to improve the accuracy and consistency of the global coordinates of feature points and obtain a high-precision reference model. The initial data obtained in step S520 still has room for improvement in accuracy. It is used as the initial value for global nonlinear optimization to improve reconstruction accuracy. The optimization function for point cloud registration is defined as follows: , The optimization function is the sum of the geometric distance residuals between the local observation point and the corresponding point in the global optimization model after spatial transformation. The optimization function also incorporates a weighted approach for each point, balancing the confidence level of corner points with varying sharpness. In the formula, , The total number of images, The total number of feature points on the reference object. Indicates the first If you see the first one on the picture If there are 1 points, the value is 1; otherwise, it is 0. The confidence level for this point is determined by the sharpness of the top corner of the tag. , Indicates the first The rotation matrix and translation vector corresponding to the image. In the The first image observed The original 3D coordinates of each point In the reference coordinate system, the first... The coordinates of the corner points to be optimized are used. The point cloud data optimized using this optimization function is used as the final data of the global 3D model of the reference object feature points, denoted as . .

[0047] S540: In actual measurement, each scan synchronously acquires the component point cloud and the reference object feature points. The current feature point set is matched with the calibrated global model by the feature point ID. The optimal rigid body transformation matrix is ​​calculated, the component point cloud is transformed to a unified global coordinate system, and finally all viewpoint point clouds are fused to obtain a complete 3D model.

[0048] Each time a component is scanned, the object under test is photographed simultaneously, and the point cloud of the component is thus obtained. The feature points on the object under test are identified and reconstructed to obtain the feature point set. ,search and For points with the same ID, use the Kabsch algorithm to calculate the difference between them. arrive Optimal rigid body transformation matrix ,Will Applied to This will align the point cloud to the desired shape. Under the global coordinate system formed by the above process, after converting all point clouds from all perspectives to the same global coordinate system, the points are merged and duplicate points are deleted to obtain a complete 3D global point cloud model.

[0049] The method of this invention constructs a two-level anti-highlight strategy. In the preprocessing stage, effective phase information is extracted through a pixel-by-pixel saturation frame culling algorithm to achieve preliminary repair of overexposed areas. Subsequently, a pixel-dimensional three-dimensional point retrieval algorithm is introduced, and residual exposure noise is solved through monocular and binocular fusion, achieving high-fidelity reconstruction of highly reflective aerospace components. Principal component analysis (PCA) is used to dynamically evaluate the local normal vector consistency index, and radial basis function (RBF) interpolation algorithm is combined to intelligently repair noisy areas; this specifically solves the problems of edge noise surge and surface unevenness in monocular and binocular fusion scenarios. Using globally unique ID-identified AprilTag stereo reference objects, a "feature point ID ↔ three-dimensional absolute coordinates" mapping system is established; combined with a confidence-weighted global bundle adjustment optimization model, the spatial geometric consistency of the panoramic three-dimensional model is ensured without the need for a high-precision turntable.

[0050] In another embodiment of the present invention, a panoramic 3D measurement system for highly reflective aerospace components based on reference object and monocular / binocular fusion is provided. The method for panoramic 3D measurement of highly reflective aerospace components based on reference object and monocular / binocular fusion includes: The system configuration and calibration module is used to build a hardware system including a projector and at least two cameras, and logically divide it into a binocular structured photonics system and two monocular structured photonics modules, and complete the calibration of each subsystem. The high reflectivity suppression phase processing module is used to decode the image sequence acquired by the hardware system using a multi-step phase shift method that removes saturated frames pixel by pixel to obtain a robust absolute phase table in order to suppress the loss of phase information in high reflectivity areas. The monocular and binocular point cloud fusion and reconstruction module is used to perform binocular stereo matching and monocular reconstruction based on the absolute phase table, establish point-to-point mapping through a 3D point correspondence retrieval algorithm of pixel dimension, calculate the affine transformation from monocular reconstructed point cloud to binocular reference point cloud, perform monocular completion and fusion on the binocular matching failure area, and generate a local complete point cloud of the tested component. The reference-based global registration module is used to process three-dimensional reference objects that are rigidly fixed to the component and have a globally unique identifier. This module includes a reference object feature point calibration unit for reconstructing a high-precision global three-dimensional coordinate model of the reference object feature points; and a multi-view point cloud alignment unit for synchronously identifying and reconstructing reference object feature points in each scan, and calculating the rigid body transformation matrix from the current local point cloud to the global coordinate system through feature point ID matching and the Kabsch algorithm, so as to achieve accurate alignment and stitching of multi-view point clouds. The point cloud refinement and optimization module is used to perform quality assessment and repair on the fused point cloud. This module locates high-noise points through local normal vector consistency analysis and uses monocular reconstruction candidate points for competitive replacement. For areas that cannot be repaired, an interpolation algorithm based on radial basis functions is used for smooth completion, and finally outputs a complete and smooth 3D panoramic model of the components.

[0051] The system of this invention employs a two-stage anti-high-brightness mechanism of pixel-by-pixel saturation frame culling and monocular / dual-eye fusion, effectively suppressing phase loss and data holes on highly reflective surfaces. By introducing a stereo reference object with a globally unique ID, multi-view stitching is transformed into a rigid body transformation solution based on absolute coordinates, achieving high-precision, automated panoramic alignment without the need for a high-precision turntable or manual marking. Combined with local normal vector evaluation and a competitive point cloud repair strategy, the smoothness and integrity of the point cloud at edges and in steep tilt areas are significantly improved. High-fidelity, high-precision, and fully automated 3D reconstruction is achieved in the precision measurement of aerospace components, overcoming the problems of traditional methods such as reliance on spraying, difficult stitching, and low efficiency.

[0052] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A panoramic three-dimensional measurement method for highly reflective aerospace components based on reference objects and monocular / binocular fusion, characterized in that, Includes the following steps: S100: Build a binocular structured light measurement system, divide the system into one binocular structured light system and two monocular structured light systems, and calibrate them separately; S200: The absolute phase table is preprocessed by a multi-step phase shift method that removes saturated frames pixel by pixel; S300: Uses a pre-processed phase table for monocular and binocular reconstruction, and establishes point-to-point mapping relationships through a 3D point correspondence retrieval algorithm in the pixel dimension; S400: Calculates affine transformation through point-to-point mapping relationship, searches for points that failed to match during stereo matching of binocular structured light system, performs monocular reconstruction and applies the affine transformation to perform mono- and binocular fusion to remove residual specular highlights, and denoises the edge noise of the fused point cloud. S500: Create a reference object for the size of the object being measured, calibrate the reference object, and simultaneously photograph the object being measured and the reference object during measurement. Change the relative spatial pose between the measurement system and the object being measured, measure again, and align the obtained point cloud to the calibrated point cloud to achieve complete point cloud reconstruction.

2. The panoramic three-dimensional measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion as described in claim 1, characterized in that, The binocular structured light measurement system in step S100 includes: a digital projector, an industrial camera, a synchronous triggering device, and a computing unit. The industrial camera is symmetrically arranged on both sides of the digital projector, and the fields of view of the three are fully overlapped by adjustment. The binocular structured light measurement system is logically divided into three subsystems, including: a monocular subsystem A consisting of camera 1 and projector, a monocular subsystem B consisting of camera 2 and projector, and a binocular subsystem C consisting of camera 1, projector, and camera 2.

3. The panoramic three-dimensional measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion according to any one of claims 1-2, characterized in that, The multi-step phase shifting method in step S200 employs at least a six-step phase shifting sequence; the pixel-by-pixel removal of saturated frames specifically involves: for each pixel, identifying and removing image frames whose grayscale values ​​have reached saturation, and using only the remaining unsaturated frame data for phase calculation to obtain the absolute phase of the pixel.

4. The panoramic three-dimensional measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion according to any one of claims 1-3, characterized in that, The specific process of the 3D point correspondence algorithm in step S300 includes: S310: The monocular and binocular subsystems are calibrated separately. The absolute phase table is obtained by decoding using a multi-step phase shift method with pixel-by-pixel saturation frame removal and stereo correction is performed to provide a geometrically consistent phase data basis for subsequent matching. S320: Perform stereo matching in the binocular subsystem to obtain the matching point set; use phase recovery and distortion correction to simultaneously derive the matching point sets corresponding to the two monocular subsystems and establish a one-to-one correspondence among the three. S330: For pixel areas not covered by binocular matching, the unmatched points in the monocular subsystem are retrieved through a reverse matching strategy to fill in the blind spots and highlight areas of the binocular field of view. S340: Based on the above matching point set, five sets of point clouds are reconstructed using the disparity-depth formula (binocular) and the projection equation (monocular), respectively, including the binocular reference point cloud, two corresponding monocular point clouds, and two completed point clouds; S350: To eliminate the scale error caused by multi-system calibration, the progressive threshold iteration method is used to calculate the affine transformation matrix from monocular point cloud to binocular reference point cloud, and the transformation accuracy is optimized by gradually eliminating noise points. S360: The affine transformation obtained by completing the point cloud is fused with the binocular reference point cloud to form a complete and seamless local point cloud, effectively filling the point cloud gaps caused by high reflectivity and blind spots.

5. The panoramic three-dimensional measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion according to any one of claims 1-4, characterized in that, The affine transformation described in step S400 is calculated based on the point-to-point mapping relationship established in step S300, and is used to unify the scale of the monocular reconstructed point cloud and the binocular reference point cloud. The inverse matching strategy is used to retrieve pixel areas not covered by the stereo matching, and monocular reconstruction is performed on these areas to obtain a complete point cloud. The complete point cloud is then fused with the stereo reference point cloud after undergoing the affine transformation to remove remaining highlights and fill in blind spots.

6. The panoramic three-dimensional measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion as described in claim 5, is characterized in that, The step of denoising the edge noise of the fused point cloud in step S400 includes: S410: Based on the preliminary fused point cloud, the surface normal vector of each point is estimated through local neighborhood principal component analysis, and the standard deviation of the angle between the normal vectors in the neighborhood is calculated. Points exceeding a preset threshold are marked as high-noise points, thereby locating the region C with poor reconstruction quality. S420: For marked high-noise points, retrieve the original reconstructed points of their corresponding coordinates in the two monocular subsystems as candidates, and re-evaluate the consistency of the normal vectors of the local region after replacement; if any candidate point can significantly reduce the noise index, then replace the original noise point with that monocular point to achieve data optimization based on viewpoint advantage. S430: If the noise still exceeds the threshold after replacement, the measurement data at that point is deemed invalid, and the noise point is removed to form a local void. An interpolation algorithm based on radial basis functions is used to perform smooth interpolation based on the geometric features of the surrounding reliable region, thereby realizing the reconstruction of continuous curved surfaces in the cavity region.

7. The panoramic three-dimensional measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion according to any one of claims 1-6, characterized in that, The reference object mentioned in step S500 is a three-dimensional structure with an AprilTag code bearing a globally unique ID pasted on its surface. The size of the AprilTag code is adapted to the size of the three-dimensional structure of the reference object, and the overall size of the reference object is determined according to the size of the aircraft component being measured. During the measurement process, the reference object and the aircraft component being measured are rigidly fixed to ensure that their relative pose remains unchanged.

8. The panoramic three-dimensional measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion as described in claim 7, is characterized in that, The specific process for the reference object calibration and alignment steps described in step S500 is as follows: S510: Uses a binocular system to capture stereo reference objects with AprilTag codes from multiple perspectives, extracts and reconstructs the 3D coordinates of feature points from each perspective, and simultaneously establishes a stable association between feature points and their unique IDs; S520: Select a point cloud from a certain viewpoint as the global coordinate system reference, use the Kabsch algorithm to solve the rigid body transformation using common ID feature points, and gradually align all viewpoint point clouds to form an initial global 3D model of the reference object feature points; S530: Using the initial model as the initial value, a weighted distance residual optimization function is constructed. Nonlinear optimization is performed by combining the observation confidence of feature points in different images to improve the accuracy and consistency of the global coordinates of feature points and obtain a high-precision reference model. S540: In actual measurement, each scan synchronously acquires the component point cloud and reference object feature points. The current feature point set is matched with the calibrated global model by the feature point ID, the optimal rigid body transformation matrix is ​​calculated, the component point cloud is transformed to a unified global coordinate system, and finally all viewpoint point clouds are fused to obtain a complete 3D model.

9. The panoramic three-dimensional measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion as described in claim 8, characterized in that, The formula for the weighted distance residual optimization function mentioned in step S530 is: ， In the formula, , The total number of images, This represents the total number of feature points on the reference object. Indicates the first If you see the first one on the picture If there are 1 points, the value is 1; otherwise, it is 0. The confidence level for this point is determined by the sharpness of the top corner of the tag. , Indicates the first The rotation matrix and translation vector corresponding to the image. In the The first image observed The original 3D coordinates of each point In the reference coordinate system, the first... The coordinates of the corner points to be optimized; the point cloud data optimized using this optimization function is used as the final data of the global 3D model of the reference object feature points, denoted as . .

10. A panoramic 3D measurement system for highly reflective aerospace components based on reference object and monocular / binocular fusion, employing the panoramic 3D measurement method for highly reflective aerospace components based on reference object and monocular / binocular fusion as described in claims 1-9, comprising: The system configuration and calibration module is used to build a hardware system including a projector and at least two cameras, and logically divide it into a binocular structured photonics system and two monocular structured photonics modules, and complete the calibration of each subsystem. The high reflectivity suppression phase processing module is used to decode the image sequence acquired by the hardware system using a multi-step phase shift method that removes saturated frames pixel by pixel to obtain a robust absolute phase table in order to suppress the loss of phase information in high reflectivity areas. The monocular and binocular point cloud fusion and reconstruction module is used to perform binocular stereo matching and monocular reconstruction based on the absolute phase table, establish point-to-point mapping through a 3D point correspondence retrieval algorithm of pixel dimension, calculate the affine transformation from monocular reconstructed point cloud to binocular reference point cloud, perform monocular completion and fusion on the binocular matching failure area, and generate a local complete point cloud of the tested component. The reference-based global registration module is used to process three-dimensional reference objects that are rigidly fixed to the component and have a globally unique identifier. This module includes a reference object feature point calibration unit, which is used to reconstruct a high-precision global three-dimensional coordinate model of the reference object feature points; and a multi-view point cloud alignment unit, which is used to synchronously identify and reconstruct the reference object feature points in each scan, and calculate the rigid body transformation matrix from the current local point cloud to the global coordinate system through feature point ID matching and the Kabsch algorithm, so as to achieve accurate alignment and stitching of multi-view point clouds. The point cloud refinement and optimization module is used to perform quality assessment and repair on the fused point cloud. This module locates high-noise points through local normal vector consistency analysis and uses monocular reconstruction candidate points for competitive replacement. For areas that cannot be repaired, an interpolation algorithm based on radial basis functions is used for smooth completion, and finally outputs a complete and smooth 3D panoramic model of the components.

Images