Fusion of industrial laser sensor data and three-dimensional reconstruction method and system

Abstract

This invention belongs to the field of 3D measurement and reconstruction technology. It discloses a method and system for fusing and reconstructing 3D data from industrial laser sensors. By acquiring multi-point laser reflection images formed by a diffraction beam splitter, initial pixel coordinates are obtained through centroid extraction, and a convex hull topology is constructed. A unique topological coding index is assigned to each laser spot, achieving sub-pixel-level precise positioning. An intensity-weighted centroid iterative calculation and a nonlinear mapping model are introduced, combined with a boundary marking mechanism based on an adjacency graph, to accurately identify abrupt depth changes. A local weighted smoothing technique is used to generate a high-quality 3D point cloud while maintaining boundary discontinuities. Finally, an accurate 3D mesh model is constructed based on topological coding. This invention improves industrial measurement accuracy, enhances the ability to handle complex surfaces and discontinuous regions, and preserves key geometric features.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional measurement and reconstruction technology, and more specifically, to a method and system for fusing and reconstructing three-dimensional data from industrial laser sensor data. Background Technology

[0002] In the fields of industrial manufacturing and quality inspection, traditional laser sensor 3D reconstruction technology faces numerous technical bottlenecks and challenges. Existing methods generally employ simple gray-level weighting or threshold segmentation in the spot centroid extraction stage, which struggles to overcome pixel resolution limitations, leading to dimensional deviations in the inspection of high-precision parts. Furthermore, these technologies lack effective point cloud topology encoding mechanisms when dealing with complex curved surfaces or steeply edged industrial components, failing to accurately establish spatial correspondences between multiple laser points. This is particularly problematic when the surfaces of industrial products such as automotive body panels or aerospace structural parts have uneven or reflective areas, easily resulting in disordered spot spatial structures. Optical distortion compensation strategies for industrial cameras and laser projection systems are often overly simplified, failing to fully consider the impact of industrial environmental factors such as temperature changes and vibrations on the optical system, leading to systematic errors in 3D coordinate mapping. In the 3D data processing stage, existing algorithms perform poorly when handling product edges and depth discontinuities, struggling to achieve a balance between noise suppression and detail preservation, often resulting in edge deformation of precision molds or loss of small features in electronic components. Faced with the challenges of industrial scenarios such as reflective metal surfaces, complex internal cavities, and thin-walled structures, traditional methods suffer from poor mesh reconstruction quality and insufficient robustness, resulting in problems such as patch distortion, incorrect connections, and measurement blind spots.

[0003] In view of this, the present invention proposes a method and system for fusing and three-dimensional reconstruction of industrial laser sensor data to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a method for fusing and three-dimensional reconstruction of industrial laser sensor data, comprising: The reflection image of the multi-point laser beam formed after beam splitting by the diffraction beam splitter is obtained on the target surface. The reflection image contains multiple laser spots distributed in space. Centroids of multiple laser spots are extracted to obtain initial pixel coordinates, and the convex hull topology of the spot set is constructed based on the initial pixel coordinates. Based on the relative positional relationship of each spot in the convex hull topology, a unique topology coding index is assigned to each spot. The topology coding index represents the row and column coordinates of the spot in the preset grid array. Subpixel-level precise coordinates are obtained by performing intensity-weighted centroid iteration calculation on each light spot. The intensity-weighted centroid iteration calculation dynamically adjusts the weights based on the gradient direction of the edge pixels of the light spot. Substitute the subpixel-level precise coordinates into a nonlinear mapping model that includes lens radial distortion compensation terms and triangulation geometric parameters to calculate the three-dimensional spatial coordinates of the measurement points corresponding to each light spot. An adjacency graph between measurement points is established based on the topological coding index and three-dimensional spatial coordinates, and boundary marking is performed on adjacent measurement point pairs with abrupt depth changes. Three-dimensional point cloud data is generated by performing local weighted smoothing processing based on adjacency graph on three-dimensional spatial coordinates. The local weighted smoothing process maintains the depth discontinuity at the boundary markers. A 3D mesh model of the target surface is generated based on the mesh topology relationship between 3D point cloud data and topology coding index.

[0005] An industrial laser sensor data fusion and 3D reconstruction system, comprising methods for achieving industrial laser sensor data fusion and 3D reconstruction, including: A laser source, used to generate a laser beam; A diffraction beam splitter is placed in the output optical path of a laser source to split the laser beam into multiple points and project them onto the target surface. The camera module is installed at a preset triangulation angle with the optical axis of the laser source to acquire the reflection image of the multi-point laser beam formed after the beam is split by the diffraction beam splitter on the target surface. The control and processing electronic device is connected to the laser source and camera module respectively. It is used to control the synchronous working timing of the laser source and camera module, extract the centroid of multiple laser spots to obtain the initial pixel coordinates, and construct the convex hull topology of the spot set based on the initial pixel coordinates. Based on the relative positional relationship of each spot in the convex hull topology, a unique topology coding index is assigned to each spot. The topology coding index represents the row and column coordinates of the spot in the preset grid array. Subpixel-level precise coordinates are obtained by performing intensity-weighted centroid iteration calculation on each light spot. The intensity-weighted centroid iteration calculation dynamically adjusts the weights based on the gradient direction of the edge pixels of the light spot. Substitute the subpixel-level precise coordinates into a nonlinear mapping model that includes lens radial distortion compensation terms and triangulation geometric parameters to calculate the three-dimensional spatial coordinates of the measurement points corresponding to each light spot. An adjacency graph between measurement points is established based on the topological coding index and three-dimensional spatial coordinates, and boundary marking is performed on adjacent measurement point pairs with abrupt depth changes. Three-dimensional point cloud data is generated by performing local weighted smoothing processing based on adjacency graph on three-dimensional spatial coordinates. The local weighted smoothing process maintains the depth discontinuity at the boundary markers. A 3D mesh model of the target surface is generated based on the mesh topology relationship between 3D point cloud data and topology coding index. An optical housing is used to house and secure the laser source, diffraction beam splitter, camera module, and control and processing electronics to maintain their relative positions.

[0006] The technical effects and advantages of the present invention, which describes a method and system for fusing and reconstructing industrial laser sensor data in three dimensions: This invention improves the accuracy and reliability of 3D reconstruction, enabling accurate identification of minute defects and deformations in precision components, providing a solid guarantee for quality control. When dealing with complex industrial surfaces, this invention accurately captures key features, effectively avoiding edge blurring and structural deformation problems common in traditional methods. For industrial challenges such as reflective metallic surfaces, dark light-absorbing materials, and complex curved surfaces, this invention demonstrates adaptability and robustness, significantly expanding the application scope of 3D measurement technology. In real-time production line inspection, the efficient processing flow of this invention significantly shortens the inspection cycle, increases production line speed, and reduces quality inspection costs in manufacturing. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the industrial laser sensor data fusion and three-dimensional reconstruction method of the present invention; Figure 2 This is a schematic diagram of the industrial laser sensor data fusion and 3D reconstruction system of the present invention; Reference numerals: ① is the diffraction beam splitter, ② is the laser source, ③ is the camera module, ④ is the control and processing electronic equipment, and ⑤ is the optical housing. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] Please see Figure 1 In this embodiment of the invention, the specific implementation process of the method for fusing and three-dimensional reconstruction of industrial laser sensor data includes: The reflection images of the multi-point laser beam split by a diffraction beam splitter on the target surface are acquired. These reflection images contain multiple spatially distributed laser spots. The diffraction beam splitter decomposes a single laser beam into multiple regularly arranged laser points using the principle of optical diffraction, forming a structured array of spots projected onto the target surface. The reflection images are captured by a high-resolution industrial camera at a specific angle, recording the spatial distribution of the laser spots on the target surface. These distributions follow the principle of structured light encoding, and the spatial deformation of the spots contains three-dimensional morphological information of the target surface. During camera acquisition, a narrowband filter is used to allow only light near the laser wavelength to pass through, suppressing ambient light interference and improving the signal-to-noise ratio for spot recognition. The acquired images feature a high-contrast multi-point laser spot distribution, providing high-quality raw data for subsequent accurate centroid extraction and 3D reconstruction.

[0010] The algorithm extracts the centroids of multiple laser spots to obtain initial pixel coordinates, and then constructs the convex hull topology of the spot set based on these coordinates. Centroid extraction is the initial step in determining the precise location of the spots, and each independent spot is identified through threshold segmentation and connected component analysis. The algorithm first performs adaptive thresholding on the reflected image to separate the spot regions from the background; then, it applies a connected component labeling algorithm to identify and label each independent spot region; finally, it calculates the gray-level weighted centroid for each labeled region to obtain the initial pixel coordinates. The initial coordinates are represented by integers, laying the foundation for subsequent sub-pixel precision extraction. The convex hull topology is achieved by calculating the two-dimensional convex hull of the spot set. The Graham scan algorithm is used to construct the smallest convex polygon containing all the spots. This structure provides a global reference framework for subsequent topological coding, ensuring the uniqueness and spatial consistency of the encoding.

[0011] Based on the relative positional relationships of each light spot within the convex hull topology, a unique topological coding index is assigned to each light spot. This index represents the row and column coordinates of the light spot within a predefined mesh array. The topological coding index is a crucial step in establishing the spatial arrangement of the light spots, constructing a structured spatial reference system by assigning each spot a unique row and column identifier. The coding process first identifies the four corner points of the convex hull as the four reference points of the mesh array. Then, it uses the least squares method to fit the unit vectors of the row and column directions of the mesh array to establish a local coordinate system. Next, it projects the position vectors of each light spot relative to the origin and corner points onto the row and column directions, calculating the real-valued row and column indices. Finally, it rounds to the nearest integer to obtain the final topological coding index. This coding method effectively handles the nonlinear deformation of the light spot positions caused by surface undulations, ensuring the spatial correspondence of points during 3D reconstruction.

[0012] Subpixel-level precise coordinates are obtained by performing intensity-weighted centroid iterative calculations on each light spot. The weights are dynamically adjusted based on the gradient direction of the edge pixels. Subpixel-level precise coordinate extraction is a core technology for improving measurement accuracy. Iterative optimization overcomes pixel resolution limitations to achieve subpixel-level positioning. The algorithm first extracts the intensity distribution matrix of the light spot region; then calculates the two-dimensional gradient field to obtain the gradient direction information of the edge pixels; next, it dynamically adjusts the pixel weights based on the consistency between the gradient direction and the direction of the light spot center; finally, iterative weighted averaging gradually refines the centroid position. The iterative process continues until the change in centroid position is less than a preset convergence threshold (usually set to 0.01 pixels), ultimately obtaining subpixel-precision light spot coordinates, significantly improving the spatial resolution of the measurement.

[0013] Substituting subpixel-level precise coordinates into a nonlinear mapping model that includes lens radial distortion compensation terms and triangulation geometric parameters, the three-dimensional spatial coordinates of the measurement points corresponding to each light spot are calculated. Nonlinear mapping is a crucial step in converting two-dimensional image coordinates into three-dimensional spatial coordinates, taking into account the distortion characteristics of the optical system and the measurement geometry. The accuracy and stability of the three-dimensional coordinates are improved by iteratively minimizing the distance between the observation ray and the laser ray.

[0014] An adjacency graph is established between measurement points based on topological coding indexes and 3D spatial coordinates, and adjacent measurement point pairs with abrupt depth changes are marked at their boundaries. The adjacency graph is the foundation for constructing the point cloud topology, determining the spatial connections between points through the proximity relationships of the coding index. The construction process first determines the four-connected or eight-connected neighborhood of each measurement point based on the topological coding index; then, the 3D Euclidean distance between adjacent points is calculated; next, the statistical characteristics of the distance distribution are analyzed to calculate the depth abrupt change threshold; finally, point pairs with abnormal distances are identified and their boundaries are marked. This method effectively identifies object edges and discontinuous regions, providing important references for subsequent local smoothing and mesh generation.

[0015] 3D point cloud data is generated by locally weighted smoothing of 3D spatial coordinates based on an adjacency graph, preserving depth discontinuities at boundary markers. Locally weighted smoothing is a crucial step in improving point cloud quality, reducing measurement noise through weighted averaging of local regions while preserving the geometric features of objects. The smoothing process first determines the smoothing neighborhood for each measurement point, excluding adjacent points at boundary markers; then, weight coefficients are calculated based on 3D distance; finally, a weighted average is performed to generate the smoothed coordinates. This adaptive smoothing method suppresses noise while maintaining the geometric details and edge features of objects.

[0016] Based on the mesh topology relationships of 3D point cloud data and topological coding index, a 3D mesh model of the target surface is generated. The 3D mesh model is the final output of surface reconstruction, forming a complete surface representation by establishing facet connections between points. The generation process first determines four-connected 2×2 sub-mesh based on the topological coding index; then, it checks whether there are boundary markers between the four vertices of the sub-mesh; for sub-mesh without boundaries, it is divided into two triangular facets along the diagonal; for sub-mesh with boundaries, the triangle division method is adaptively adjusted according to the boundary position; finally, a normal vector consistency check is performed on all facets, and degenerate triangles are removed to generate the final 3D mesh model. The 3D mesh model has a regular topological structure and accurate geometry, fully expressing the 3D morphology of the target surface.

[0017] In this embodiment of the invention, the detailed implementation steps for assigning a unique topology coding index to each light spot include: The four corner points of the convex hull topology are identified and marked as the coordinates of the four corner points of the mesh array. Corner point identification is the foundation for establishing the mesh coordinate system, defining the mesh boundary by determining the positions of the four poles. The identification process first calculates the centroid and principal axis directions of the convex hull to establish a local coordinate system; then, in the principal axis coordinate system, the convex hull vertices with the largest and smallest x and y coordinates are selected to determine the four corner point spots; finally, based on the relative positions of the corner points, they are marked as upper left, upper right, lower left, and lower right corner points, respectively. To improve the robustness of corner point identification, projection contour analysis and angle selection strategies are adopted to ensure accurate identification of corner points even when the spot distribution is irregular. These four corner points serve as reference points for the mesh array, defining the range and orientation of the entire mesh and providing a benchmark framework for subsequent row and column coordinate calculations.

[0018] Using the four corner light spots as references, the row and column unit vectors of the grid array are fitted using the least squares method. Direction vector fitting is crucial for establishing the local coordinate system of the grid, determining the principal axis directions through a data-driven approach. The fitting process first assumes an ideal arrangement of light spots on a regular grid, establishing the correspondence between the actual and ideal positions of the spots; then, it constructs a least squares objective function to minimize the projection error from the actual positions to the ideal grid; finally, it obtains the optimal row and column unit vectors by solving the normal equations. The mathematical model for fitting is: ; in, This represents the actual location of the light spot. The origin and corner positions are... and These are the unit vectors in the row and column directions, respectively. and These are the row and column indices in the ideal mesh. The least-squares problem is solved using Singular Value Decomposition (SVD) to obtain the optimal direction vectors. These vectors are typically not parallel to the image coordinate system, but are adaptively determined based on the actual light spot distribution, thus handling mesh distortion caused by camera tilt or target surface tilt.

[0019] Calculate the position vector of each non-corner spot relative to the preset origin corner point, and project these position vectors onto the row and column unit vectors respectively. Projection calculation is the core step in determining the spot grid coordinates, converting the two-dimensional position into grid row and column parameters through vector algebra. The calculation process first selects the upper left corner point as the preset origin; then, for each spot, calculate its position difference vector relative to the origin; next, project this vector onto the row and column unit vectors respectively to obtain the projection length in the row and column directions; finally, normalize the projection length according to the preset grid spacing to obtain the real-valued row and column coordinates. The projection calculation formula is: ; ; in, and They are light spots The row and column projection coordinates, The location of the light spot. The origin position, and These are the unit vectors in the row and column directions, respectively. and The preset grid spacing is typically determined by the average distance between adjacent light spots or specified through system calibration parameters to ensure the scale consistency of the projected coordinates.

[0020] The row and column indices of each light spot are calculated based on the ratio of the projection length to the preset grid spacing. These real values ​​are then rounded to obtain integer row and column coordinates, which are used as the topology coding index. Coordinate rounding is a crucial step in generating the final topology code, quantizing continuous projection coordinates into discrete grid indices. The process first rounds the projection coordinates to integer values; then, it checks the distribution of the rounded results to ensure the grid indices are continuous and free of gaps; finally, it performs a global consistency check on all points in the light spot set to ensure the integrity and coherence of the grid structure. The mathematical expression for the rounding process is: ; ; in, and For light spots The integer row and column indices are used as the final topological encoding index. This rounding-based quantization method is simple and effective, accurately reflecting the relative position of the light spot in the grid, and providing a structured point correspondence for subsequent 3D reconstruction.

[0021] The system detects conflicts where multiple light spots map to the same topology coding index. If a conflict exists, the real values ​​of the conflicting light spots are fine-tuned and corrected. Conflict detection and correction are mechanisms to ensure the uniqueness of topology coding, maintaining the consistency of the mesh structure by checking and repairing coding conflicts. The detection process first constructs a coding index hash table to record the list of light spots corresponding to each index; then, it identifies the index positions containing multiple light spots and marks them as conflict points; for conflict points, its original real coordinates are analyzed to determine the optimal fine-tuning scheme. The correction strategy is based on the nearest neighbor principle, fine-tuning the real coordinates of conflicting light spots towards nearby empty mesh positions, minimizing the adjustment amount while ensuring no new conflicts are generated. The correction formula is: ; ; in, and These are the fine-tuned real coordinates. and Original coordinates and For adjustment amounts, the total adjustment amount is usually chosen to be the one that makes the total adjustment amount... The minimum acceptable solution involves fine-tuning and rounding to obtain an updated topology coding index, ensuring that all light spots have a unique mesh location identifier, thus providing a reliable topology for subsequent 3D reconstruction and mesh generation.

[0022] In this embodiment of the invention, the detailed implementation steps of the light intensity weighted centroid iterative calculation include: A spot intensity distribution matrix is ​​constructed by extracting the grayscale values ​​of all pixels within the coverage area of ​​a single spot. The intensity distribution matrix is ​​a digital representation of the spot's characteristics, recording its shape, size, and intensity distribution. The extraction process first determines the region of interest (ROI) based on the initial centroid position. The ROI size is typically set to 1.5-2 times the spot diameter to ensure complete coverage of the spot and its surroundings. Then, the grayscale value of each pixel within the ROI is read to construct a two-dimensional intensity matrix. Simultaneously, the relative coordinates of the pixels are recorded to establish a mapping relationship between grayscale values ​​and positions. For images with high dynamic range, grayscale normalization is also required to map grayscale values ​​to a uniform range, eliminating the influence of brightness differences between different spots. The intensity distribution matrix provides a complete data foundation for subsequent background fitting and gradient analysis and is the core input for accurate centroid calculation.

[0023] The net spot intensity distribution is obtained by performing Gaussian background fitting on the spot intensity distribution matrix and subtracting background noise intensity. Background subtraction is a key step in improving the signal-to-noise ratio. It involves separating the spot signal from the background noise to extract a clean spot distribution. The fitting process first selects pixels in the area surrounding the spot, typically those with intensities below the average. Then, a Gaussian plane model is used to fit the background distribution, capturing gradients and non-uniform backgrounds. Finally, the fitted background model is subtracted from the original intensity distribution to obtain the net spot intensity distribution. The mathematical expression of the Gaussian background model is: ; in, For the background model in position The intensity value at that location, arrive The fitting parameters were determined using the least squares method. After background subtraction, all negative values ​​were set to zero to ensure the physical validity of the net intensity distribution, and the results were normalized to facilitate subsequent weight calculations. This background subtraction method effectively handles non-uniform lighting and ambient light interference, improving the accuracy and stability of centroid calculation.

[0024] The two-dimensional gradient field of the net spot intensity distribution is calculated to obtain the gradient direction of each pixel. Gradient field calculation is fundamental to determining the edge characteristics of the spot; by analyzing the direction and magnitude of intensity changes, the geometric structure of the spot is captured. The calculation process first applies the Sobel or Scharr operator to the net intensity distribution to calculate the first derivatives in the x and y directions, respectively; then, based on the derivatives in the two directions, the gradient magnitude and direction angle are calculated; finally, a threshold is used to retain significant gradient information and filter out noise. The gradient direction calculation formula is: ; in, For pixels gradient direction angle at that location, and They are respectively direction and The gradient component of the direction. The range of the direction angle is... The gradient direction represents the direction of the fastest increase in intensity. For an ideal light spot, the gradient direction usually points towards the center of the light spot. The gradient field not only provides directional information of edge pixels but also reflects the shape characteristics of the light spot, providing an important basis for subsequent directional consistency analysis and weight calculation.

[0025] The orientation consistency coefficient of each pixel is calculated based on the relationship between the gradient direction and the center direction of the light spot. The product of the net light spot intensity distribution and the orientation consistency coefficient is used as the correction weight. Orientation consistency analysis is an innovative technique to improve centroid accuracy. It dynamically adjusts pixel weights by evaluating the degree of consistency between the gradient direction and the ideal direction. The analysis process first calculates the pointing vector from each pixel to the currently estimated centroid; then it calculates the angle between this pointing vector and the gradient direction; finally, based on the size of the angle, the orientation consistency coefficient is determined. For an ideal circular light spot, the gradient direction should be opposite to the center pointing direction, and the angle should be close to... The formula for calculating the directional consistency coefficient is: ; in, For pixels The directional consistency coefficient, The gradient direction angle, From The angle pointing towards the current centroid. The coefficient ranges from [0,1], with values ​​closer to 1 indicating a more consistent direction. The correction weight combines net intensity and directional consistency, calculated using the following formula: ; in, For the final adjusted weights, Net light spot intensity, This is a weighting factor (usually 1-2), controlling the degree of influence of the orientation factor. This orientation-consistency-based weight correction can effectively suppress the influence of asymmetric spot and background interference, significantly improving the accuracy and robustness of centroid localization.

[0026] The first iteration's centroid coordinates are obtained by weighted averaging of the pixel coordinates using corrected weights. Weighted averaging is the core operation in centroid calculation; it determines the geometric center of the spot by comprehensively considering the weight distribution of each pixel. The calculation process uses the classic centroid formula, but the weights no longer depend solely on pixel intensity; instead, they comprehensively consider both intensity and orientation consistency. The formula for calculating the first iteration's centroid coordinates is: ; ; in, and The centroid coordinates for the first iteration. For pixels The corrected weights are summed over all pixels within the area covered by the light spot. This weighted averaging method fully utilizes the intensity distribution and geometric characteristics of the light spot, and compared to the traditional intensity-weighted method, it can more accurately locate the center of the light spot, especially for non-ideal light spots and noisy environments, thus improving the robustness and accuracy of the positioning.

[0027] The area covered by the light spot is reduced from the centroid coordinates of the first iteration, and the weight calculation and weighted averaging process is repeated until the change in centroid coordinates is less than a preset convergence threshold. Iterative optimization is the key to overcoming pixel resolution limitations. By refining the centroid position multiple times, sub-pixel-level positioning accuracy is achieved. The optimization process first reduces the ROI range from the centroid of the first iteration, typically to 50-70% of the original ROI; then, the directional consistency and corrected weights within this area are recalculated; next, a new round of weighted averaging is performed to obtain updated centroid coordinates; finally, the change in centroid coordinates is calculated to determine whether the convergence condition is met. The formula for calculating the change in centroid during iterations is: ; in, The change in the center of mass, and For the first The centroid coordinates of each iteration are determined. The iteration terminates when the change is less than a preset threshold (typically 0.01 pixels) or when the maximum number of iterations (typically 10) is reached. The final output sub-pixel coordinates are accurate to the order of 0.01 pixels, far exceeding the resolution of the original image, significantly improving the accuracy of subsequent 3D reconstruction. This iterative optimization strategy combines region reduction and dynamic weight adjustment, improving computational efficiency while ensuring positioning accuracy, making it a core technology for high-precision spot localization.

[0028] In this embodiment of the invention, the detailed implementation steps for constructing the nonlinear mapping model include: The camera's intrinsic parameter matrix and radial distortion coefficients are obtained, and distortion correction is performed on the sub-pixel level precise coordinates to obtain the distortion-corrected image coordinates. Intrinsic parameter acquisition and distortion correction are fundamental to optical system calibration, restoring the geometric correctness of the image by eliminating the effects of lens distortion. The acquisition process first uses a standard calibration board to acquire calibration images from different viewpoints; then, the camera's intrinsic parameter matrix and distortion coefficients are calculated using Zhang's calibration method or OpenCV calibration tools; finally, the calibration parameters are applied to the sub-pixel coordinates for distortion correction transformation. The mathematical representation of the intrinsic parameter matrix is: ; in, and Focal length and The principal point coordinates are used. The radial distortion correction formula is: ; ; in, and For the corrected coordinates, and Original coordinates The normalized distance to the image center. , , The radial distortion coefficient is used. The corrected image coordinates eliminate the nonlinear effects of lens distortion, providing geometrically correct input for subsequent ray calculations, which is a prerequisite for accurate 3D reconstruction.

[0029] The observation ray equations for each spot are calculated based on the distortion-corrected image coordinates and the intrinsic parameter matrix. Observation ray calculation is a core step in 3D reconstruction. Through the inverse transformation of the intrinsic parameter matrix, image plane points are mapped to rays in the camera coordinate system. The calculation process first converts the distortion-corrected pixel coordinates to normalized image coordinates; then, using the inverse of the intrinsic parameter matrix, the ray direction vector is calculated; finally, combined with the camera's pose in the world coordinate system, the equations are converted to ray equations in the world coordinate system. The parameters of the observation rays are expressed as follows: ; in, To observe points on the ray, Let this be the position of the camera's optical center in the world coordinate system. Let be the ray direction vector. Here are the ray parameters. The formula for calculating the direction vector is: ; in, For the camera rotation matrix, The inverse of the intrinsic parameter matrix. These are normalized pixel coordinates. The observation ray represents the three-dimensional straight line from the camera's optical center to the corresponding image point. It is a fundamental geometric element of triangulation and provides a constraint for subsequent ray intersection calculations.

[0030] The laser emission ray equations for each spot are calculated based on the beam-splitting angle parameters of the diffraction beam splitter and the topological coding index. Laser ray calculation is another key constraint in completing the 3D reconstruction. The emission direction of each spot is determined through the geometric relationship of the diffraction beam splitter. The calculation process first determines the beam-splitting angle of the ideal grid points based on the design parameters of the diffraction beam splitter; then, the corresponding ideal beam-splitting angle is found according to the topological coding index of the spot; finally, the emission ray equations are calculated by combining the laser's pose in the world coordinate system. The parameters of the laser ray are expressed as follows: ; in, For a point on the laser beam, This represents the position of the laser emission point in the world coordinate system. Let be the ray direction vector. Here are the ray parameters. The formula for calculating the direction vector is: ; in, For the laser rotation matrix, and These are the horizontal and vertical beam splitting angles, indexed by topological encoding. and diffraction angle interval The calculation yielded: ; ; and The maximum beam splitting angle is typically half the array size multiplied by the angular interval. This laser ray calculation method based on topological coding establishes a mapping relationship between the image position of the laser spot and the spatial ray, providing a second geometric constraint for triangulation.

[0031] Find the closest point pair between the equations of the observed ray and the laser emitted ray, using the midpoint as the initial 3D spatial coordinates. Ray intersection is the core algorithm of triangulation; by calculating the closest point between two non-intersecting rays, the position of the 3D point is determined. Due to measurement and calibration errors, the two rays usually do not intersect precisely, but rather have a minimum distance. The solution process uses an analytical method, minimizing the distance between the two rays by calculating the optimal values ​​of the ray parameters. The formula for calculating the closest point pair is: ; ; in, and For optimal ray parameters, Let be the modulus of the cross product of the ray directions, and represent the perpendicularity of the two rays. The closest point pairs are respectively... and The midpoint is calculated as follows: ; As the initial three-dimensional spatial coordinates, it provides the approximate position of the three-dimensional point, providing a good initial value for subsequent nonlinear optimization, and ensuring the rapid convergence and stability of the optimization process.

[0032] A nonlinear residual function is established between the initial 3D spatial coordinates and the image coordinates after distortion correction. This function is then optimized by minimizing the nonlinear residual function. Nonlinear optimization is a key technique for improving the accuracy of 3D reconstruction. Through an iterative approach, multiple constraints are comprehensively considered to accurately estimate the coordinates of 3D points. The optimization process first establishes a reprojection error model to describe the geometric constraints of 3D points projected onto the image plane; then, a comprehensive residual function is constructed, including reprojection error and geometric consistency constraints; finally, the Levenberg-Marquardt (LM) algorithm is used for iterative optimization, gradually adjusting the positions of the 3D points to minimize the residual function. The nonlinear residual function is defined as: ; in, For the residual function, For the 3D points to be optimized, for The coordinates of the point projected onto the image plane. The coordinates of the observed image, The distance function from a point to a line. , , These are the weighting coefficients.

[0033] Through repeated iterations until the residual change is less than a preset threshold or the maximum number of iterations is reached, the optimized three-dimensional spatial coordinates are finally output, which significantly improves the accuracy and consistency of point positions and lays a high-quality data foundation for subsequent point cloud processing and mesh generation.

[0034] In this embodiment of the invention, the detailed implementation steps for establishing an adjacency graph and marking boundaries include: The set of neighboring measurement points for each measurement point is determined based on the row and column relationships of the topological coding index. Neighborhood determination is fundamental to constructing the point cloud topology, defining the connectivity between points through the adjacency relationships of the coding index. The determination process first analyzes the topological coding index (r, c) of each measurement point; then, based on the 4-connectivity or 8-connectivity rules, the set of indexes for neighboring points is determined; finally, the existence of points corresponding to these indices is verified, constructing a valid neighborhood set. The set of indexes for a 4-connectivity neighborhood is { }, an 8-connected network also needs to include { In practical applications, due to insufficient occlusion or reflection conditions, some light spots may be missing. Therefore, it is necessary to verify the validity of neighboring points, remove non-existent points, and ensure the correctness of adjacency relationships. This neighborhood definition method based on topological coding fully utilizes the regular arrangement characteristics of structured light, and is more efficient and topologically correct than distance-based neighborhood search, providing structured spatial relationships for subsequent boundary detection and smoothing processing.

[0035] The 3D Euclidean distance between each measurement point and its neighboring measurement points is calculated as the adjacent edge length. Edge length calculation is a fundamental method for quantifying the spatial relationship between points, using Euclidean distance to measure the degree of spatial separation between adjacent points. The calculation process calculates the distance in 3D space for each effective neighboring point of each measurement point. The distance value directly reflects the local geometric characteristics of the object's surface; the distance value is relatively uniform in smooth regions, while abrupt changes occur at edges and corners. These distance values, as the length attribute of adjacent edges, provide important geometric information for subsequent boundary detection, effectively identifying discontinuous regions and boundary features of objects.

[0036] The median and standard deviation of all adjacent edge lengths are calculated to determine the depth abrupt change threshold. Threshold calculation is the core of adaptive boundary detection; a suitable judgment criterion is determined through statistical analysis. The calculation process first collects the length values ​​of all valid adjacent edges; then, the median of these values ​​is calculated. ) and standard deviation ( Finally, based on statistical characteristics, the threshold for determining deep mutations is determined. The formula for calculating the threshold is: ; in, The threshold for determining deep mutations. The adjustment factor is typically 2-3. The median represents a typical value for adjacent edge lengths, reflecting the overall smoothness of the object's surface; the standard deviation quantifies the dispersion of the length distribution, characterizing the complexity and variability of the surface. This statistically-based adaptive thresholding method can automatically adjust the judgment criteria according to the geometric characteristics of specific objects, avoiding both over-detection leading to misjudgment and missed detection of obvious boundaries, exhibiting strong adaptability and robustness.

[0037] Pairs of adjacent measurement points whose adjacent edge length exceeds the depth abrupt change threshold are marked as candidate boundary edges. Candidate boundary detection is a preliminary step in screening potential boundaries, identifying regions where abrupt changes may occur based on the length threshold. The detection process traverses all adjacent edges, marking edges whose length is significantly greater than the normal level as candidate boundary edges. The judgment criteria are: ; in, For point and points Candidate boundary markers between, This is the corresponding Euclidean distance. The threshold value calculated earlier is used. Candidate boundary markers provide an initial indication of potentially discontinuous regions, but due to measurement noise and local geometric variations, they may contain some false positives, requiring further verification and screening to ensure the accuracy and reliability of boundary detection.

[0038] The gradient values ​​in the depth direction are calculated at the two ends of the candidate boundary edge. Candidate boundary edges with gradient values ​​greater than a preset gradient threshold are marked as boundaries. Gradient verification is a crucial step in determining the true boundary. By analyzing the significance of depth changes, it distinguishes between true boundaries and fluctuations caused by noise. The verification process first calculates the difference in the depth direction (usually the z-axis) between the two ends of the candidate boundary; then it compares this difference with the distance between the two points in the xy plane to calculate the depth gradient; finally, it compares this with a preset gradient threshold to determine the final boundary marker. This dual-judgment method, combining distance and gradient, considers both the spatial distribution of points and the characteristics of depth variation. It can accurately identify the true discontinuous boundaries of the object surface, providing important structural information for subsequent smoothing and mesh generation, ensuring that geometric details are preserved at the boundaries and that they are not over-smoothed.

[0039] In this embodiment of the invention, the detailed implementation steps of local weighted smoothing processing include: For each measurement point, a smooth neighborhood is formed by extracting its unmarked neighboring measurement points. Smooth neighborhood extraction is a crucial step in preserving boundary features, avoiding smoothing across boundaries by excluding boundary points. The extraction process first checks all adjacency relationships of the measurement points; then excludes adjacent edges marked as boundaries; and finally collects the remaining valid neighboring points to form a set of smooth neighborhoods. For points located on object boundaries, their smooth neighborhoods are typically smaller than those of interior points, sometimes even containing only 1-2 neighboring points. The definition of the smooth neighborhood ensures that the smoothing operation is performed only within the same continuous surface and does not cross depth discontinuities, thus preserving the object's edge features and geometric details and avoiding the edge blurring problem common in traditional smoothing methods. This boundary-aware neighborhood definition method is a key technology for high-quality surface reconstruction, providing a structurally sound set of points for subsequent weighted averaging.

[0040] Based on the spatial distances between the measurement point and its neighboring measurement points in the smoothing neighborhood, a Gaussian kernel function is used to calculate the smoothing weights. Weight calculation is the core of controlling the smoothing intensity, ensuring the locality and adaptability of the smoothing through distance decay. The calculation process first measures the 3D Euclidean distances between the measurement point and each point in its neighborhood; then, the Gaussian kernel function maps these distances to weight values, with closer points having larger weights and farther points having smaller weights. This distance-based weight calculation method ensures that the smoothing effect decays smoothly with distance, with closer points having a greater impact and farther points having a smaller impact. This guarantees the locality of the smoothing while avoiding the loss of detail caused by over-smoothing, providing a flexible control mechanism for obtaining high-quality point cloud data.

[0041] The smoothing weights are normalized, and then the normalized weights are used to calculate a weighted average of the 3D spatial coordinates of each measurement point within the smoothing neighborhood to obtain the smoothed 3D coordinates. Weighted averaging is the core algorithm of the smoothing operation; by comprehensively considering the positions and weights of neighboring points, it calculates smoother coordinate values. The smoothing process first normalizes the weights of each point to ensure that the sum of the weights is 1; then, the normalized weights are used to calculate a weighted average of the coordinates of the neighboring points to obtain the new smoothed coordinates.

[0042] This weighted averaging method comprehensively considers the information of all points in the neighborhood, weighting them according to their distance. This effectively suppresses random measurement noise and local fluctuations. At the same time, since boundary points have been excluded when defining the neighborhood, it does not blur the edges and geometric features of objects, thus preserving important structural information and significantly improving the quality and visual effect of point clouds.

[0043] The process involves determining whether the measurement point is located at the endpoint of a boundary marker. If so, the weight of the smoothed 3D coordinates is reduced and then weighted and fused with the original 3D spatial coordinates. Boundary point processing is a key technique for maintaining sharp edges, ensuring that geometric features at the boundary are not over-smoothed through special processing. The process first determines whether the measurement point is a boundary point, i.e., whether there are adjacent edges marked as boundaries; then, based on the characteristics of boundary points, the smoothing intensity is adjusted; finally, the smoothed result is weighted and fused with the original coordinates to obtain the final 3D coordinates.

[0044] For boundary points, a strategy of reducing smoothing intensity is adopted, and the calculation formula is as follows: ; in, For the final coordinates, For smoothed coordinates, Original coordinates The smoothing intensity coefficient is typically set to 0.3-0.5 for boundary points and 0.8-1.0 for non-boundary points. This adaptive smoothing strategy ensures that boundary points retain more of their original features, avoiding edge blurring, while interior points receive stronger smoothing to suppress noise. The resulting 3D point cloud data is smooth and noise-free, while maintaining clear edges and detailed features, providing a high-quality geometric foundation for subsequent mesh generation.

[0045] In this embodiment of the invention, the detailed implementation steps for generating the three-dimensional mesh model include: The four vertex measurement points of each 2×2 sub-grid in the mesh array are determined based on the row and column coordinates of the topology-coded index. Sub-grid extraction is the foundation of mesh model generation, defining quadrilateral mesh cells through topology coding. The extraction process first determines the valid range of row and column indices; then it iterates through all possible sub-grid locations. For each position, check the four vertices. , , , The method checks whether all valid measurement points exist; finally, it collects all complete sub-mesh with four vertices as the basic unit for triangulation. This sub-mesh extraction method based on topological encoding does not require complex spatial search and topological reconstruction. It directly utilizes the regular arrangement characteristics of structured light to efficiently establish the connection relationship between points, providing a structured topological framework for subsequent triangular patch generation.

[0046] The process checks whether there are boundary edges marked with boundary markers between the four vertex measurement points of each 2×2 submesh. Boundary checking is a prerequisite for adaptive triangulation, adjusting the patch division strategy by identifying boundary conditions within the submesh. The checking process determines whether any of the six possible edges (four edges and two diagonals) of each submesh are marked as boundaries. Boundary determination is achieved by querying the previously established boundary markers B(i,j). If any of the six edges is marked as a boundary, its position and direction are recorded, providing a basis for subsequent triangulation. This boundary-aware submesh analysis method can identify discontinuous regions and boundary features on the object's surface, ensuring that the generated mesh model maintains geometric accuracy in these regions and avoids erroneous connections across discontinuous boundaries, providing topological guarantees for high-quality surface reconstruction.

[0047] If no boundary edge exists, the four vertices of the sub-mesh are divided into two triangular patches along the diagonal. If a boundary edge exists, the division method of the triangular patches is adaptively adjusted according to the position of the boundary edge. Triangulation is the core step in generating the mesh surface, maintaining the geometric correctness of the surface through a boundary-aware adaptive division strategy. The division process first determines whether a boundary edge exists within the sub-mesh; for sub-meshes without boundaries, a standard diagonal division is used to form two triangular patches; for sub-meshes with boundaries, the division method is adjusted according to the position of the boundary edge to ensure that the triangular edges do not cross the boundary. The standard division for sub-meshes without boundaries is as follows: ; ; Adaptive triangulation with boundaries selects diagonals that do not intersect the boundary as dividing lines based on the location of the boundary edges, or adjusts to a more suitable dividing pattern in complex cases, and may even abandon some connections to ensure the correctness of the geometric topology. This adaptive triangulation method considers the continuity characteristics of the object surface, avoids the incorrect connections and distortions caused by traditional regular division at the boundaries, and greatly improves the visual quality and geometric accuracy of the mesh model.

[0048] Perform a normal vector consistency check and unify the normal vector direction for all triangular faces. Normal vector unification is a crucial step in generating an effective mesh model, establishing correct internal and external surface relationships by ensuring all faces face the same direction. The unification process first calculates the normal vector for each triangular face; then, it establishes an adjacency graph of the faces; next, starting from a initial face, it checks the normal vector directions of adjacent faces using breadth-first search; if the directions are inconsistent (angle greater than 90 degrees), the face is flipped; finally, this process is repeated for all connected regions to ensure the normal vector direction of the entire mesh is consistent. The normal vector calculation formula is: ; in, For the normal vector of the surface, , , Here are the coordinates of the three vertices of the triangle. This normal vector unification process ensures the consistency of the surface orientation of the mesh model, providing a correct geometric representation for subsequent rendering and analysis, avoiding topological errors such as mixed inside and outside surfaces in the model, and improving the quality and usability of the mesh model.

[0049] Degenerate triangular facets with areas smaller than a preset minimum area threshold or aspect ratios greater than a preset aspect ratio threshold are removed. The remaining triangular facets are then organized into a 3D mesh model data structure. Facet selection and organization are the final steps in generating the final mesh model. Through quality control and data organization, a complete and efficient 3D surface representation is constructed. The selection process first calculates the area and aspect ratio of each triangular facet; then compares them with preset thresholds to remove degenerate, low-quality facets; finally, the remaining high-quality facets are organized into a standard mesh data structure. Triangle quality is evaluated using two indicators: area and aspect ratio. An excessively small area indicates that the triangle may be a pseudo-structure generated by noise, while an excessively large aspect ratio indicates that the triangle is overly elongated and has poor geometric approximation. The aspect ratio calculation formula is: ; in, Aspect ratio, The longest side of the triangle. The area is the triangle area. The elimination criteria typically set an area threshold of 5% of the average area and an aspect ratio threshold of 10. The final retained triangle faces are organized according to a standard mesh data structure, including a list of vertex coordinates, a list of face indices, and optional additional information such as normal vectors and texture coordinates, forming a complete 3D mesh model that provides a high-quality geometric representation for subsequent visualization, analysis, and applications.

[0050] The above describes the industrial laser sensor data fusion and 3D reconstruction system in the embodiments of this application. The following describes the industrial laser sensor data fusion and 3D reconstruction system in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 2 ① is a diffraction beam splitter, ② is a laser source, ③ is a camera module, ④ is a control and processing electronic device, and ⑤ is an optical housing; one embodiment of the industrial laser sensor data fusion and 3D reconstruction system in this application includes: A laser source, used to generate a laser beam; A diffraction beam splitter is placed in the output optical path of a laser source to split the laser beam into multiple points and project them onto the target surface. The camera module is installed at a preset triangulation angle with the optical axis of the laser source to acquire the reflection image of the multi-point laser beam formed after the beam is split by the diffraction beam splitter on the target surface. The control and processing electronic device is connected to the laser source and camera module respectively. It is used to control the synchronous working timing of the laser source and camera module, extract the centroid of multiple laser spots to obtain the initial pixel coordinates, and construct the convex hull topology of the spot set based on the initial pixel coordinates. Based on the relative positional relationship of each spot in the convex hull topology, a unique topology coding index is assigned to each spot. The topology coding index represents the row and column coordinates of the spot in the preset grid array. Subpixel-level precise coordinates are obtained by performing intensity-weighted centroid iteration calculation on each light spot. The intensity-weighted centroid iteration calculation dynamically adjusts the weights based on the gradient direction of the edge pixels of the light spot. Substitute the subpixel-level precise coordinates into a nonlinear mapping model that includes lens radial distortion compensation terms and triangulation geometric parameters to calculate the three-dimensional spatial coordinates of the measurement points corresponding to each light spot. An adjacency graph between measurement points is established based on the topological coding index and three-dimensional spatial coordinates, and boundary marking is performed on adjacent measurement point pairs with abrupt depth changes. Three-dimensional point cloud data is generated by performing local weighted smoothing processing based on adjacency graph on three-dimensional spatial coordinates. The local weighted smoothing process maintains the depth discontinuity at the boundary markers. A 3D mesh model of the target surface is generated based on the mesh topology relationship between 3D point cloud data and topology coding index. An optical housing is used to house and secure the laser source, diffraction beam splitter, camera module, and control and processing electronics to maintain their relative positions.

[0051] This invention achieves high-precision fusion and 3D reconstruction of industrial laser sensor data through diffraction-splitting laser reflection image processing, subpixel-level centroid extraction, topological coding, nonlinear mapping, and local smoothing. The adaptive boundary processing method of this invention can accurately preserve the geometric features and edge details of objects, effectively improving the accuracy and visual quality of 3D reconstruction, and providing a systematic solution for industrial inspection and precision measurement.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0053] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0054] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for fusing and 3D reconstruction of industrial laser sensor data, characterized in that, include: The reflection image of the multi-point laser beam formed after beam splitting by the diffraction beam splitter is obtained on the target surface, and the reflection image contains multiple laser spots distributed in space. Centroids are extracted from the multiple laser spots to obtain initial pixel coordinates, and a convex hull topology of the spot set is constructed based on the initial pixel coordinates. Based on the relative positional relationship of each spot in the convex hull topology, a unique topology coding index is assigned to each spot, and the topology coding index represents the row and column coordinates of the spot in the preset grid array; Subpixel-level precise coordinates are obtained by performing intensity-weighted centroid iteration calculation on each light spot. The intensity-weighted centroid iteration calculation dynamically adjusts the weights based on the gradient direction of the edge pixels of the light spot. Substitute the subpixel-level precise coordinates into a nonlinear mapping model that includes lens radial distortion compensation terms and triangulation geometric parameters to calculate the three-dimensional spatial coordinates of the measurement points corresponding to each light spot. An adjacency graph between measurement points is established based on the topology coding index and the three-dimensional spatial coordinates, and boundary marking is performed on adjacent measurement point pairs with abrupt depth changes. The three-dimensional spatial coordinates are subjected to local weighted smoothing based on the adjacency graph to generate three-dimensional point cloud data, and the local weighted smoothing process maintains depth discontinuity at the boundary markers; Based on the mesh topology relationship between the three-dimensional point cloud data and the topology encoding index, a three-dimensional mesh model of the target surface is generated.

2. The method according to claim 1, characterized in that, Assigning a unique topology coding index to each light spot includes: Identify the four corner spotlights of the convex hull topology vertex and mark them as the coordinates of the four corner points of the mesh array; Using the four corner light spots as a reference, the row direction unit vector and column direction unit vector of the grid array are fitted by the least squares method; Calculate the position vector of each non-corner spot relative to the preset origin corner point, and project the position vector onto the row direction unit vector and the column direction unit vector respectively; The row and column indices of each spot are calculated based on the ratio of the projection length to the preset grid spacing. The real values ​​are then rounded to obtain integer row and column coordinates as the topology coding index. Detect conflicts where multiple light spots are mapped to the same topology coding index. If a conflict exists, fine-tune and correct the real value of the conflicting light spot.

3. The method according to claim 1, characterized in that, The intensity-weighted centroid iterative calculation includes: Extract the grayscale values ​​of all pixels within the coverage area of ​​a single light spot to construct a light spot intensity distribution matrix; Gaussian background fitting is performed on the light spot intensity distribution matrix and the background noise intensity is subtracted to obtain the net light spot intensity distribution; Calculate the two-dimensional gradient field of the net spot intensity distribution to obtain the gradient direction of each pixel; The directional consistency coefficient of the pixel is calculated based on the relationship between the gradient direction and the center of the light spot, and the product of the net light spot intensity distribution and the directional consistency coefficient is used as the correction weight. The first iteration centroid coordinates are obtained by weighting the pixel coordinates using the corrected weights. The area covered by the light spot is reduced by using the centroid coordinates of the first iteration as the center, and the weight calculation and weighted average process is repeated until the change in the centroid coordinates is less than the preset convergence threshold.

4. The method according to claim 1, characterized in that, The construction of the nonlinear mapping model includes: Obtain the camera's intrinsic parameter matrix and radial distortion coefficients, and perform distortion correction processing on the sub-pixel level precise coordinates to obtain the distortion-corrected image coordinates; The observation ray equations corresponding to each spot are calculated based on the distortion-corrected image coordinates and the intrinsic parameter matrix. The laser emission ray equations corresponding to each spot are calculated based on the beam splitting angle parameters of the diffraction beam splitter and the topological coding index. Solve for the nearest point pair between the observed ray equation and the laser emitted ray equation, and use the midpoint as the initial three-dimensional spatial coordinates; A nonlinear residual function is established between the initial three-dimensional spatial coordinates and the distortion-corrected image coordinates, and optimization is performed by minimizing the nonlinear residual function.

5. The method according to claim 1, characterized in that, The process of establishing an adjacency graph and marking its boundaries includes: The set of neighboring measurement points for each measurement point is determined based on the grid row and column relationship of the topology coding index; Calculate the three-dimensional Euclidean distance between each measurement point and its adjacent measurement points as the adjacent edge length; Calculate the depth abrupt change threshold by calculating the median and standard deviation of all adjacent edge lengths. The adjacent measurement point pairs whose adjacent edge length is greater than the depth mutation determination threshold are marked as candidate boundary edges; Calculate the gradient values ​​in the depth direction at the measurement points at both ends of the candidate boundary edge, and mark the candidate boundary edges whose gradient values ​​are greater than a preset gradient threshold.

6. The method according to claim 1, characterized in that, The local weighted smoothing process includes: For each measurement point, extract its neighboring measurement points that are not marked by the boundary to form a smooth neighborhood; Based on the spatial distance between the measurement point and each adjacent measurement point in the smooth neighborhood, the smoothing weight is calculated using the Gaussian kernel function. The smoothing weights are normalized, and the normalized smoothing weights are used to perform a weighted average of the three-dimensional spatial coordinates of each measurement point in the smoothing neighborhood to obtain the smoothed three-dimensional coordinates. Determine whether the measurement point is located at the endpoint of the boundary mark. If so, reduce the weight of the smoothed 3D coordinates and perform a weighted fusion with the original 3D spatial coordinates.

7. The method according to claim 1, characterized in that, The generation of the 3D mesh model includes: The four vertex measurement points of each 2×2 subgrid in the grid array are determined based on the row and column coordinates of the topology coding index; Check whether there is a boundary edge marked by the boundary marker between the four vertex measurement points of each of the 2×2 subgrids; If there is no boundary edge, the four vertices of the sub-mesh are divided into two triangular patches along the diagonal. If there is a boundary edge, the division method of the triangular patches is adaptively adjusted according to the position of the boundary edge. Perform a normal vector consistency check on all triangular facets and unify the normal vector direction; Degenerate triangular facets with an area smaller than a preset minimum area threshold or an aspect ratio greater than a preset aspect ratio threshold are removed, and the remaining triangular facets are organized into a three-dimensional mesh model data structure.

8. A system for fusing and reconstructing three-dimensional data from industrial laser sensors, used to implement the method described in any one of claims 2 to 7, characterized in that, include: A laser source, used to generate a laser beam; A diffraction beam splitter is disposed in the output optical path of the laser source to split the laser beam into multiple laser points and project them onto the target surface. The camera module is installed at a preset triangulation angle with the optical axis of the laser source to acquire the reflection image of the multi-point laser formed after the beam splitter splits the laser beam on the target surface. A control and processing electronic device is communicatively connected to the laser source and the camera module, respectively, for controlling the synchronous working timing of the laser source and the camera module, extracting the centroid of the multiple laser spots to obtain the initial pixel coordinates, and constructing the convex hull topology of the spot set based on the initial pixel coordinates; Based on the relative positional relationship of each spot in the convex hull topology, a unique topology coding index is assigned to each spot, and the topology coding index represents the row and column coordinates of the spot in the preset grid array; Subpixel-level precise coordinates are obtained by performing intensity-weighted centroid iteration calculation on each light spot. The intensity-weighted centroid iteration calculation dynamically adjusts the weights based on the gradient direction of the edge pixels of the light spot. Substitute the subpixel-level precise coordinates into a nonlinear mapping model that includes lens radial distortion compensation terms and triangulation geometric parameters to calculate the three-dimensional spatial coordinates of the measurement points corresponding to each light spot. An adjacency graph between measurement points is established based on the topology coding index and the three-dimensional spatial coordinates, and boundary marking is performed on adjacent measurement point pairs with abrupt depth changes. The three-dimensional spatial coordinates are subjected to local weighted smoothing based on the adjacency graph to generate three-dimensional point cloud data, and the local weighted smoothing process maintains depth discontinuity at the boundary markers; Based on the mesh topology relationship between the three-dimensional point cloud data and the topology coding index, a three-dimensional mesh model of the target surface is generated. An optical housing is used to house and fix the laser source, the diffraction beam splitter, the camera module, and the control and processing electronics to maintain their relative positions.

Images