Compressor crankshaft high-precision 3D shape detection method based on laser vision

Abstract

The invention discloses a compressor crankshaft high-precision 3D shape detection method based on laser vision, belongs to the technical field of precision measurement and intelligent manufacturing, and aims to solve the problems existing in crankshaft detection. The method comprises the following steps: step 1, scanning a compressor crankshaft in a non-contact manner by adopting a three-coordinate motion platform carrying a line laser contourgraph to obtain an original three-dimensional point cloud; 2, carrying out the preprocessing of the original three-dimensional point cloud through employing adaptive radius filtering and minimum point constraint voxel filtering; step 3, realizing pin hole position degree detection based on an improved RANSAC algorithm and two-dimensional boundary clustering, and fitting an end surface excircle and a pin hole circle to obtain the pin hole position degree; and step 4, realizing pin-end face perpendicularity detection through coarse registration and weighted ICP fine registration. The method is used for detecting the 3D morphology of the crankshaft.

Description

Technical Field

[0001] This invention relates to the field of precision measurement and intelligent manufacturing technology, specifically to a high-precision online detection technology for the position and perpendicularity of the crankshaft of an air conditioning compressor in a new energy vehicle. Background Technology

[0002] The crankshaft is the core power transmission component of the air conditioning compressor in new energy vehicles. Its geometric accuracy (position, perpendicularity, cylindricity, etc.) directly affects the compressor's operating efficiency, NVH performance, and the vehicle's range. The position accuracy of the pin hole and the perpendicularity accuracy of the pin not only directly affect the equipment's operational stability, energy efficiency, and service life, but are also closely related to key indicators such as overall machine vibration and noise, making them core elements for ensuring the compressor's efficient and reliable operation. According to the 2023 "Technical Roadmap for Thermal Management Systems of New Energy Vehicles" issued by the China Society of Automotive Engineers, the position accuracy of the compressor crankshaft main journal and the pin hole must be controlled within ±50 μm, and the perpendicularity of the pin after installation must be ≤0.05 radians; otherwise, it will lead to decreased energy efficiency and increased vibration and noise.

[0003] Existing detection methods can be divided into two main categories: contact and non-contact, but both have significant drawbacks:

[0004] The classic contact inspection method is coordinate measuring machine (CMM). This method has a high tolerance for surface roughness, but it suffers from problems such as micro-deformation caused by contact pressure (indentation deformation of 0.5-1μm), and the measurement speed is slow (approximately 30 minutes per piece). The accuracy and speed of manual ruler and gauge measurement depend heavily on the skill level of the measurer, making it difficult to meet the needs of large-scale, standardized industrial inspection.

[0005] The classic non-contact inspection method is structured light scanning. However, due to the influence of metal surface reflection, the point cloud loss rate in the crankshaft chamfer area is as high as 30%, resulting in low accuracy. It also requires multi-view stitching, leading to large cumulative errors. Chongqing University proposed an image inspection scheme based on a CCD area array camera in the field of non-contact measurement. By moving the camera to acquire crankshaft contour images and performing stitching processing, it can achieve multi-parameter measurement in low-precision scenarios such as forgings. However, its accuracy is limited by optical imaging resolution and image processing algorithms, and it is mainly used for preliminary inspection in the rough machining stage. Tokyo Seimitsu Corporation of Japan developed a mobile 3D optical measurement device that integrates three cameras and image analysis technology on the basis of a coordinate measuring machine. It converts the object's position coordinates into image information through scanning and extracts physical position data using a dedicated algorithm. It can complete crankshaft image processing within 40 seconds, compressing the overall inspection cycle to less than two minutes, making it suitable for large-scale production scenarios. However, its accuracy is relatively low, and it relies on complex image algorithm optimization to overcome noise interference.

[0006] In summary, crankshaft inspection has the following problems: First, the measurement accuracy of existing technologies is mostly limited to the millimeter level, which cannot achieve second-level detection with sub-micron level accuracy, and cannot meet the stringent requirements of micron-level accuracy for crankshafts in new energy vehicle compressors; Second, metal highlights and corrosion cause point cloud distortion, and traditional algorithms cannot adaptively compensate for this. Summary of the Invention

[0007] To address the aforementioned problems in crankshaft inspection, this invention provides a high-precision 3D topography inspection method for compressor crankshafts based on laser vision.

[0008] The present invention discloses a high-precision 3D topography detection method for compressor crankshafts based on laser vision, which includes the following steps:

[0009] Step 1: Use a three-coordinate motion platform equipped with a line laser profilometer to scan the compressor crankshaft in a non-contact manner to obtain the original three-dimensional point cloud;

[0010] Step 2: Preprocess the original 3D point cloud using adaptive radius filtering and minimum point number constraint voxel filtering;

[0011] Step 3: Detect pin hole position degree based on improved RANSAC algorithm and two-dimensional boundary clustering, and fit the outer circle of end face and the pin hole circle to obtain the pin hole position degree.

[0012] Step 4: Detect the perpendicularity of the pin-end face through coarse registration and weighted ICP fine registration.

[0013] Preferably, the line laser profilometer in step 1 is a Keyence LJ-S080 laser detector, which uses a blue laser light source.

[0014] Preferably, the search radius of the adaptive radius filtering in step 2 is dynamically adjusted according to the Euclidean distance.

[0015] Preferably, the process of obtaining the position in step 3 includes the following steps:

[0016] Step 31: Collect the point cloud of the crankshaft end face including the pin hole. After preprocessing, solve the crankshaft end face normal based on principal component analysis (PCA) and use KdTree to accelerate the solution of the normal.

[0017] Step 32: Fit the preprocessed end face point cloud and the normal obtained in step 31 using the improved RANSAC algorithm, and perform Rodriguez point cloud coordinate transformation based on the plane parameters.

[0018] Step 33: Extract the boundary features of the annular ring formed by the outer circle of the end face and the pin hole circle. The boundary angle threshold is dynamically adjusted based on the k nearest neighbors and the point cloud density.

[0019] Step 34: Based on the traditional Euclidean clustering, the extracted boundary features are clustered to extract the set of boundary feature points of the end face plane region and the pin hole circle region.

[0020] Step 35: Perform dynamic clustering and filtering on the extracted boundary feature point set, and output the optimal boundary feature point set;

[0021] Step 36: Use the improved RANSAC algorithm to fit the two-dimensional circle composed of the outer circle of the end face and the circle of the pin hole. The set distance between the two fitting results of the outer circle of the end face and the circle of the pin hole is used as the pin hole position degree.

[0022] Preferably, the improved RANSAC algorithm changes the traditional RANSAC algorithm's method of determining interior points by assigning weights to the point cloud normals: points that satisfy the thresholds for distance from the point to the plane and the angle threshold between the point and the normal vector of the fitted plane are determined as interior points, forming a preferred set of interior points. The preferred set of interior points is then subjected to weighted singular value decomposition, and the plane equation is optimized with the consistency of the normal vector as the weight, and the centroid of the interior points is calculated.

[0023] Preferably, the boundary angle threshold in step 33 is dynamically adjusted based on the k-nearest neighbor and point cloud density. The threshold is lowered in dense point cloud areas to capture subtle geometric changes, and the threshold is raised in sparse areas to avoid noise misjudgment.

[0024] Preferably, step 34, which involves clustering the extracted boundary features using a two-dimensional boundary clustering method improved from traditional Euclidean clustering, is as follows:

[0025] The input point cloud is traversed point by point, skipping processed points. When an unprocessed point is found, a new clustering container is created and added to the queue to be processed as the starting point for growth.

[0026] Then, the points in the queue are processed in a loop. For each unprocessed point, a radius search is performed to obtain the neighborhood point set. If the neighborhood density is lower than the threshold or the normal exceeds the threshold, it is marked as noise. Otherwise, the current point is added to the cluster and the neighborhood points are traversed. Unprocessed adjacent points are added to the queue to recursively expand the clustering region. Finally, the current point is marked as processed.

[0027] After completing the region growing, the algorithm verifies the clustering effectiveness by using a minimum size threshold. Clusters that meet the threshold are output as valid clusters, while those that do not are discarded as noise.

[0028] Preferably, the process of obtaining the perpendicularity of the pin end face in step 4 is as follows:

[0029] Step 41: Insert the pin into the pin hole of the crankshaft, and collect and preprocess the original 3D point cloud from multiple angles;

[0030] Step 42: Perform point cloud registration, which consists of coarse registration using SAC-IA based on PFFH and fine registration using ICP, and output the registered point cloud and root mean square error.

[0031] Step 43: Perform downsampling processing;

[0032] Step 44: Extract the surface point cloud of the pin cylinder and crankshaft end face based on the improved Euclidean clustering algorithm;

[0033] Step 45: Use the improved RANSAC algorithm to fit the pin cylinder surface and the crankshaft end face, and obtain the deviation angle between the pin axis and the normal of the crankshaft end face as the crankshaft perpendicularity.

[0034] Preferably, step 42 introduces an adaptive weight adjustment mechanism during ICP fine registration, using the distance from the vertex to the viewpoint to weight the error metric function of ICP, making the algorithm focus more on vertices that are closer to the camera.

[0035] Preferably, step 44 improves the Euclidean clustering algorithm by introducing a z-axis height constraint condition into the traditional Euclidean clustering algorithm: Constraint 1: In the z-axis direction, the first constraint condition is established by measuring the perpendicular distance between the end face and the side of the pin; Constraint 2: The normal vector of the pin side point cloud always maintains a perpendicular relationship with the z-axis, and the normal vector of the end face point cloud maintains a parallel relationship with the z-axis.

[0036] The beneficial effects of this invention are as follows: This invention solves the over- or under-deletion problem of non-uniform point clouds through Euclidean distance mapping; it designs a dynamic adjustment strategy for angle thresholds to prevent false detection in sparse areas and improve sensitivity in dense areas; it designs an improved RANSAC algorithm with normal weighting to optimize the fitting accuracy of planes / cylinders with normal vector consistency; it uses a blue laser + Sham optics system to suppress metallic reflections, increasing the point cloud acquisition speed by 6 times; it constructs a three-coordinate motion control strategy (smoothing time 25ms), reducing scanning vibration error by 40%. It integrates FPFH-SAC-IA registration (bidirectional geometric constraints) and viewpoint-weighted ICP to solve the pin-end face registration deviation. The specific advantages are quantified as follows:

[0037] 1. Improved precision: The pin hole position error was reduced from ±0.10mm to ±0.05mm, and the pin perpendicularity error was reduced from ±0.9° to ±0.5°;

[0038] 2. Efficiency optimization: The single-item inspection time has been reduced from 300s to 1.2s, and the point cloud processing speed has been increased by 8 times (millions of point clouds < 1s).

[0039] 3. Enhanced robustness: It maintains a 90% detection rate even under interference from oil stains and rust, and can handle local occlusion with 40% point cloud loss. Attached Figure Description

[0040] Figure 1 This is a flowchart of a high-precision 3D topography detection method for compressor crankshaft based on laser vision, as described in this invention.

[0041] Figure 2 These are crankshaft exterior parts, among which Figure 2 (a) is an unpinned pin. Figure 2 (b) Insert the pin;

[0042] Figure 3 This is a flowchart of the position degree calculation process;

[0043] Figure 4 This is a flowchart for calculating verticality;

[0044] Figure 5 This is a flowchart of the improved RANSAC algorithm for plane and circle fitting in the position degree calculation process;

[0045] Figure 6 This is a flowchart of the improved RANSAC algorithm for pin cylinder fitting in the perpendicularity calculation process;

[0046] Figure 7 It is a positional instance point cloud map, where Figure 7 (a) is the original point cloud. Figure 7 (b) is the point cloud after adaptive radius filtering. Figure 7 (c) is the point cloud after voxel filtering with a small number of points constrained. Figure 7 (d) shows the clustering results;

[0047] Figure 8 It is a vertical instance point cloud collection. Figure 8 (a) to (d) are four images taken from a 45° overhead view and rotated 90° respectively;

[0048] Figure 9 It is to extract the effective point cloud image. Figure 9 (a) to (d) are respectively Figure 8 The effective point clouds from (a) to (d);

[0049] Figure 10 This is the result after three registrations. Figure 10 (a) to (d) are respectively Figure 9 (a) to (d) Results after three registrations;

[0050] Figure 11 It is the output result after point cloud filtering. Detailed Implementation

[0051] 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.

[0052] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0053] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0054] This invention focuses on the intelligent breakthrough of crankshaft position and perpendicularity detection. Based on the forefront of industrial inspection technology and guided by digitalization and intelligence, it aims to build a real-time online inspection system that integrates three-dimensional point cloud scanning and digital twin technology. Through multi-dimensional sensing collaboration and dynamic process feedback, it can achieve efficient and accurate measurement of crankshaft geometric position parameters and closed-loop quality control throughout the entire process.

[0055] Specific Implementation Method 1: The following is combined with... Figures 1 to 11 This embodiment describes a high-precision 3D topography detection method for compressor crankshafts based on laser vision. The method includes the following steps:

[0056] Step 1: Use a three-coordinate motion platform equipped with a line laser profilometer to scan the compressor crankshaft in a non-contact manner to obtain the original three-dimensional point cloud;

[0057] Step 2: Preprocess the original 3D point cloud using adaptive radius filtering and minimum point number constraint voxel filtering;

[0058] Step 3: Detect pin hole position degree based on improved RANSAC algorithm and two-dimensional boundary clustering, and fit the outer circle of end face and the pin hole circle to obtain the pin hole position degree.

[0059] Step 4: Detect the perpendicularity of the pin-end face by using FPFH-SAC-IA coarse registration and weighted ICP fine registration.

[0060] The line laser profilometer in step 1 uses a Keyence LJ-S080 laser detector, which uses a high-brightness blue laser light source.

[0061] The platform body and columns are made of 00 grade granite to ensure accuracy. The Z-axis uses a lead screw module to control the scanner's lifting and lowering, while the two planar axes use linear motors to control the X and Y axis movement.

[0062] The Keyence LJ-S080 laser inspection instrument is used, which integrates motor scanning light cutting technology and Schahm's law optical system to achieve focus-free large depth-of-field scanning; the high-concentration blue laser light source effectively suppresses the interference of metal surface reflection and improves the stability of point cloud acquisition on complex crankshaft surfaces.

[0063] The LJ-S080 employs a fan-shaped diffused laser and optical light-receiving system, capturing images by rotating and scanning this optical system. However, blind spots can occur due to obstruction of the laser illuminating the object and the laser reflected back from the object into the light-receiving component. The XYZ axis measurement range layout has been optimized to avoid optical obstruction.

[0064] In step 2, the original 3D point cloud is preprocessed, namely, adaptive radius filtering and minimum point number constraint voxel filtering. The search radius of the adaptive radius filtering is dynamically adjusted according to the Euclidean distance.

[0065] Adaptive radius filtering: Original 3D point cloud , where i is the sampling point number.

[0066] Point cloud radius filtering removes noise based on neighbor density analysis, applying it to each sampling point in the point cloud data. A spherical neighborhood with radius r is constructed with the point as the center, and the number of neighboring points m contained in the neighborhood is counted. If m exceeds the preset threshold M, the point is determined to be a valid data point and retained; otherwise, it is discarded as a noise point.

[0067] To accelerate neighborhood search, dynamic Euclidean adaptive radius filtering is performed on the original 3D point cloud for each point. Set the search radius R, and optimize the dynamic parameters of the search radius R adaptively: Set the minimum search radius. When less than the threshold When using the minimum protective radius value, the search radius R is set to... To avoid overfitting in the near-field region, if the search radius R is higher than the threshold... Then, the search radius R is dynamically adjusted according to the Euclidean distance by expanding the radius value according to the preset proportional coefficient.

[0068] Minimum number of points constrained voxel filtering: A dynamic voxel adjustment mechanism that adaptively adjusts voxel size to ensure that each effective voxel contains at least [number of points]. This allows for a balance between noise reduction and feature preservation.

[0069] Physical preparation for position monitoring: crankshaft exposed pin holes, pins not inserted; see [link / reference]. Figure 2 As shown in (a).

[0070] See Figure 3 Step 3: Calculation process for pin hole position:

[0071] Step 31: Collect the point cloud of the crankshaft end face including the pin hole. After preprocessing, solve the crankshaft end face normal based on principal component analysis (PCA) and use KdTree to accelerate the solution of the normal.

[0072] For the data collection method, please refer to step 1; for the preprocessing process, please refer to step 2.

[0073] The crankshaft end face normals are solved using Principal Component Analysis (PCA) on the preprocessed end face point cloud data, with KdTree used to accelerate the solution. PCA is an unsupervised dimensionality reduction method based on eigenvalue decomposition of the covariance matrix, its core being the identification of the principal direction of the data distribution. In 3D point cloud processing, normal estimation is achieved by calculating the minimum variance direction of points in the local neighborhood; this direction corresponds to the eigenvector corresponding to the minimum eigenvalue of the covariance matrix. KdTree (k-dimensional tree), as a spatial partitioning data structure, accelerates the nearest neighbor search process and reduces time complexity.

[0074] Step 32: Fit the preprocessed end face point cloud and the normal obtained in step 31 using the improved RANSAC algorithm, and perform Rodriguez point cloud coordinate transformation based on the plane parameters.

[0075] The principle of the traditional RANSAC algorithm can be summarized as follows: randomly select the smallest subset of samples that meets the requirements for solving the model parameters from the original data and use this subset to perform initial parameter estimation; then construct an error evaluation system, use a set spatial distance threshold to determine the validity of all data points, and classify the points that meet the error constraints into the valid sample set; after multiple rounds of parameter estimation and validity verification, finally select the model that contains the most interior points or meets the minimum residual condition and use it as the optimal solution.

[0076] This invention introduces an improved RANSAC algorithm. For details on the improved RANSAC algorithm, please refer to [link / reference needed]. Figure 5 As shown, the improved RANSAC algorithm changes the traditional RANSAC algorithm's method of determining interior points by assigning weights to the point cloud normals: points that satisfy the thresholds for distance from the point to the plane and the angle threshold between the point and the normal vector of the fitted plane are determined as interior points, forming a preferred set of interior points. The preferred set of interior points is then subjected to weighted singular value decomposition, and the plane equation is optimized with the consistency of the normal vector as the weight, and the centroid of the interior points is calculated.

[0077] After preprocessing the point cloud, Principal Component Analysis (PCA) is used to calculate the normals. During sampling, the angle between the normal vectors of the three points is verified to avoid collinearity. Interior points are screened using a dual threshold of distance and normal vector. Weighted SVD of the normal vectors is used to optimize the plane equations, and the centroids of the interior points are calculated to improve robustness. This effectively improves the plane recognition accuracy under noise and missing section conditions. Compared with the traditional RANSAC algorithm, this algorithm has a higher accuracy in fitting planes. The results output plane parameters and the points involved in the fitting. Rodriguez point cloud coordinate transformation is performed based on the plane parameters to facilitate subsequent processing.

[0078] Step 33: Extract the boundary features of the annular ring formed by the outer circle of the end face and the pin hole circle. The boundary angle threshold is dynamically adjusted based on the k nearest neighbors and the point cloud density.

[0079] The accuracy and efficiency of boundary feature point extraction directly affect the point cloud reconstruction and fitting results. This invention designs an adaptive angle threshold optimization and integrates it into a two-dimensional boundary clustering based on an improvement of traditional Euclidean clustering. The limitation of a fixed angle threshold is that a globally uniform threshold is difficult to adapt to dynamic changes in regional geometric features, leading to over-segmentation of high-curvature regions and missed detection of smooth regions. The core of the improved algorithm is to dynamically adjust the angle threshold through k-nearest neighbor density: lower the threshold in dense point cloud regions to capture subtle geometric changes, and raise the threshold in sparse regions to avoid noise-induced misjudgments, achieving adaptive optimization of boundary extraction.

[0080] Set the average neighborhood spacing of the point cloud The neighborhood density is Edge, boundary angle threshold The solution must satisfy the following constraints:

[0081]

[0082] Among them The threshold value is the average neighborhood spacing of the point cloud. After a few hours, Take threshold Core calculation function It consists of the following model.

[0083] Establish the three-dimensional parametric space equations:

[0084]

[0085] Rules for constructing the coefficient matrix and right-hand side terms:

[0086]

[0087] Solve using the column pivoting QR decomposition method:

[0088]

[0089] This process ensures that the parameters are optimal under typical operating conditions. and Reflecting the logarithmic scaling property of density, the exponent It is determined by the distance attenuation effect in engineering modeling.

[0090] Establish a bimodal piecewise function to characterize the fundamental angle threshold. :

[0091]

[0092] In the formula:

[0093] The term achieves density logarithmic compensation, with a compensation coefficient of . . Construct an exponential decay function, with the decay rate controlled by the parameter 0.02. It provides nonlinear distance saturation characteristics to avoid overestimation under large spacing.

[0094] Neighborhood search radius A dual-modal power-law model is adopted:

[0095]

[0096] The proportion factor was obtained by optimizing the Levenberg-Marquardt algorithm. Adjusting the base radius and power exponent Control density sensitivity. Radius constraint is... To avoid overfitting.

[0097] The local noise level assessment uses second-order statistics:

[0098]

[0099] in For the first The number of point neighbors, This represents the local distance mean.

[0100] Angle noise correction factor:

[0101]

[0102] Convert the noise standard deviation to the angular domain:

[0103]

[0104] Based on engineering practice experience, a threshold of 15° is adopted.

[0105] Curvature estimation and curvature compensation model based on principal component analysis (PCA):

[0106]

[0107] The weighting function is designed as follows:

[0108]

[0109] The final compensation result is as follows:

[0110]

[0111] Final boundary angle threshold adaptively adjusts the output (after double limiting):

[0112]

[0113] This constraint ensures: lower limit Curves prevent oversensitivity upper limit Avoid ineffective large angles.

[0114] This improved algorithm for adaptively adjusting angle thresholds addresses the inherent limitations of static thresholds in traditional point cloud boundary detection. Through deep fusion of problem modeling and constraints, it transforms dynamic geometric features and point cloud density distribution into mathematical constraints in a multi-dimensional parameter space, constructing a nonlinear mapping relationship based on density logarithmic scaling and distance decay. In the dynamic parameter initialization phase, the algorithm achieves autonomous calibration of threshold parameters through column-pivot QR decomposition and Levenberg-Marquardt optimization, eliminating reliance on manual parameter tuning. The innovative design of the multimodal nonlinear computation module (including a bimodal piecewise function and a power-law neighborhood model) effectively balances the need for pseudo-boundary suppression in sparse regions and geometrical abrupt change capture in dense regions, avoiding the overfitting risk of traditional methods through exponential decay and distance saturation characteristics. The adaptive neighborhood analysis mechanism, through a dynamic adjustment strategy of K-nearest neighbors combined with majority voting and directional consistency constraints, achieves accurate characterization of local geometric features: using a large neighborhood range in sparse regions to improve fault tolerance, and shrinking the neighborhood scale in dense regions to enhance resolution.

[0115] To address noise interference, the algorithm constructs a second-order statistical model and an angle-domain noise propagation model through a noise suppression module. This model converts the local distance mean and neighborhood distribution discrete measures into angle correction coefficients, significantly reducing noise interference with threshold determination. Furthermore, curvature compensation technology extracts local curvature features through principal component analysis and dynamically adjusts boundary sensitivity using a weighting function. This enhances resistance to oversegmentation at sharp edges and improves the recognition probability of weak boundaries in smooth regions. By establishing a collaborative analysis mechanism of multi-scale geometric features and local statistical properties, the algorithm overcomes the adaptability bottleneck of fixed thresholds in complex scenes: in sharp edge regions, the dynamically adjusted threshold effectively suppresses false boundary misjudgments caused by point cloud sparsity; in smooth surface regions, adaptive sensitivity control accurately captures subtle geometric changes. By introducing a noise propagation model and curvature compensation mechanism, the algorithm significantly improves anti-interference capabilities while maintaining boundary integrity, freeing automated processing systems from dependence on manual parameter tuning.

[0116] This algorithm constructs an adaptive angle threshold framework by dynamically sensing geometric features and point cloud density distribution. Combined with noise suppression, curvature compensation, and multimodal neighborhood analysis, it achieves synergistic optimization of boundary detection accuracy and anti-interference capability in complex scenarios, overcoming the inherent shortcomings of traditional static thresholds.

[0117] This method breaks through the limitations of traditional static thresholds by dynamically sensing the geometric features and density distribution of point clouds. It constructs a nonlinear mapping relationship based on logarithmic density scaling and distance attenuation, and combines noise suppression, curvature compensation, and multimodal neighborhood analysis techniques to achieve synergistic optimization of boundary detection accuracy and anti-interference capability in complex scenes. The algorithm ultimately outputs all boundary feature points that meet the requirements.

[0118] Step 34: Based on the traditional Euclidean clustering, the extracted boundary features are clustered to extract the set of boundary feature points of the end face plane region and the pin hole circle region.

[0119] This algorithm achieves efficient point cloud clustering by using both point cloud distance and normal vectors as criteria. Its core process is as follows:

[0120] The input point cloud is traversed point by point, skipping processed points. When an unprocessed point is found, a new clustering container is created and added to the queue to be processed as the starting point for growth.

[0121] Then, the points in the queue are processed in a loop. For each unprocessed point, a radius search is performed to obtain the neighborhood point set. If the neighborhood density is lower than the threshold or the normal exceeds the threshold, it is marked as noise. Otherwise, the current point is added to the cluster and the neighborhood points are traversed. Unprocessed adjacent points are added to the queue to recursively expand the clustering region. Finally, the current point is marked as processed.

[0122] After completing the region growing, the algorithm verifies the clustering effectiveness by using a minimum size threshold. Clusters that meet the threshold are output as valid clusters, while those that do not are discarded as noise.

[0123] This process combines distance constraints with a dual filtering mechanism (single-point density verification and overall scale verification) to effectively suppress noise interference while ensuring spatial continuity. Its time complexity is linearly related to the point cloud size and the number of adjacency relationships, making it suitable for robust extraction of continuous structures in complex scenarios.

[0124] Step 35: Perform dynamic clustering and filtering on the extracted boundary feature point set, and output the optimal boundary feature point set;

[0125] A maximum cluster extraction paradigm based on statistical significance testing and divide-and-conquer optimization is adopted. The cluster size distribution is modeled as a power-law process. The optimal truncation criterion is derived through maximum likelihood estimation, and information entropy is introduced as a quantitative indicator of cluster structural integrity, thus constructing a dual-constraint optimization objective function. To balance size and structural compactness, a comprehensive scoring function is constructed:

[0126]

[0127] in These are the weighting coefficients. and This represents the extreme value of the entropy for all clusters. , For the i,j-th cluster, For clusters The cluster entropy. For the evaluation function... Solve the problem to obtain the optimal cluster and output the optimal set of boundary feature points.

[0128] The post-processing of 3D point cloud clustering employs a divide-and-conquer strategy and statistical optimization to extract the main cluster structure. Power-law distribution modeling and information entropy mechanisms are used to filter out spatially consistent main clusters from noise, ensuring high-fidelity representation of the geometric morphology. This technique utilizes an energy dissipation model to dynamically sort the main clusters, enabling subsequent fitting algorithms to focus on representative features, effectively improving the robustness of geometric modeling and avoiding noise interference. The final output is a stable and optimal set of boundary points.

[0129] Suppose the clustering results include candidate clusters Its size follows a power-law distribution:

[0130]

[0131] This distribution indicates that the probability of large-scale clusters occurring increases with... It increases and then decays exponentially. This is for estimating parameters. Define the likelihood function:

[0132]

[0133] right Taking the derivative and setting it to zero, we obtain the maximum likelihood estimate:

[0134]

[0135] in As the minimum effective cluster size threshold, the system exhibits significant long-tail characteristics, requiring priority extraction of the top two clusters.

[0136] Perform information entropy-driven cluster optimization and define clusters. The structural entropy of the junction is:

[0137]

[0138] in Let be the cluster centroid. A smaller structural entropy indicates stronger spatial consistency of the point cloud. To balance scale and structural compactness, a comprehensive scoring function is constructed:

[0139]

[0140] in These are the weighting coefficients. and Let be the extreme value of the entropy of all clusters. Optimal cluster selection is equivalent to solving:

[0141]

[0142] The cluster sorting process is modeled as an energy minimization problem. The system Hamiltonian is defined, and the process is simulated using the Metropolis-Hastings algorithm.

[0143]

[0144] in For clusters The temperature modulus parameters reflect its scale stability; For Lagrange multipliers; The total number of points. Based on probability. Accepting the exchange, as T approaches zero, the system converges to the lowest energy state, which is the optimal solution arranged in descending order of size.

[0145] The post-processing of 3D point cloud clustering, through a divide-and-conquer strategy and statistical mechanics optimization, extracts the main cluster structure, providing a physically meaningful input foundation for subsequent point cloud fitting, geometric modeling, and scene reconstruction. By employing power-law distribution modeling and an information entropy-driven cluster selection mechanism, the algorithm accurately extracts spatially consistent and significantly sized main clusters from the initial clusters plagued by noise. The structural integrity of these main clusters ensures a high-fidelity representation of the target object's geometry, significantly improving the robustness of the fitting. The dynamic sorting process of the main clusters achieves global stability through an energy dissipation model, enabling subsequent fitting algorithms to focus on the most representative geometric features and avoiding model distortion caused by noisy clusters or outliers.

[0146] Step 36: Use the improved RANSAC algorithm to fit the two-dimensional circle composed of the outer circle of the end face and the circle of the pin hole. The set distance between the two fitting results of the outer circle of the end face and the circle of the pin hole is used as the pin hole position degree.

[0147] The crankshaft end face point cloud data is represented as two-dimensional circles: the outer circle of the end face and the pin hole circle. The positional deviation of the two circles is the pin hole position degree.

[0148] The improved RANSAC algorithm is used to fit a two-dimensional circle. The improved RANSAC algorithm flowchart can be found here. Figure 5 As shown, a distance threshold D and a radius deviation threshold R are first set as dual screening conditions. These serve as the thresholds for the perpendicular distance between the feature point and the plane containing the spatial circle during the iterative calculation process, and the difference between the distance from each feature point to the center of the fitted circle and the calculated radius. A candidate plane model is established by exhaustively combining three points. The perpendicular distance from each data point to the plane is calculated using the spatial projection method. Data points that meet the distance thresholds are recorded as primary valid points, and the number of times they pass through is counted.

[0149] After completing all three-point combination traversals, probability screening and model optimization are performed, and initial data cleaning is conducted based on the statistical results. High-frequency valid points conforming to the planar model are retained to form a primary dataset W1, which is used to reconstruct the optimal planar equation. These planar parameters are then used as geometric constraints for subsequent spatial circle fitting.

[0150] After screening and optimization, spatial circle fitting and secondary verification are required. In the optimized dataset W1, candidate circle models are established by enumerating four-point combinations. The center coordinates and radius parameters are calculated using spatial geometric analytical methods. Data points that meet the radius threshold are selected through radial deviation analysis, and the verification count of each point is recorded. Finally, secondary data cleaning is performed based on statistical frequency to obtain a high-confidence dataset W2. These points are considered to have passed the second verification. After traversing all combinations in the point set W1 or after a finite number of iterations, feature points with significantly fewer satisfying counts are deleted, and the feature points with the most satisfying counts are selected as the final qualified point set W2. The parameters of the final fitted spatial plane and the center coordinates and radius of the fitted spatial circle are recalculated.

[0151] The RANSAC algorithm counts the number of cases that meet the threshold in each iteration. , The dataset containing the inliers and their quantities is determined. After traversing all combinations or after a finite number of iterations, the dataset with the largest number of inliers is selected as the final optimization result. The distance between these two fitted results is used as the final result for pin hole position.

[0152] Example of positional degree: See Figure 7 The original point cloud data was acquired using the LJ-S080 high-precision 3D laser scanning system, with a total of 462,063 discrete sampling points. (See original point cloud data). Figure 7 (a) In the geometrically complex chamfered transition regions and the boundaries and interiors of pin hole structures, multiple physical effects during the optical scanning process, including but not limited to laser incident angle deviation, occlusion effects, and multiple reflection interference, cause significant geometric distortion in the point cloud data in local areas. This distorted data not only introduces systematic errors, affecting the accuracy of surface fitting based on the least squares method, but also causes topological damage to subsequent feature extraction algorithms, thereby reducing the reliability of 3D reconstruction and geometric analysis. Therefore, a hybrid filtering strategy based on statistical analysis and geometric constraints must be adopted, combining point cloud density distribution and curvature continuity detection, to achieve robust extraction of effective point clouds, ensuring that the accuracy and stability of subsequent data processing meet the requirements of engineering applications.

[0153] In the processing of 3D point cloud data, distorted point clouds and end-face point clouds exhibit significant statistical differences in local spatial density distribution. To address the distortion at point cloud boundaries, pin hole boundaries, and the internal regions of pin holes, this study employs a point cloud radius filtering algorithm based on neighborhood density feature analysis. This algorithm calculates the point cloud distribution density within the spherical neighborhood of each sampling point and removes outliers based on a preset density threshold. Traditional radius filtering algorithms exhibit good robustness in uniform sampling scenarios; however, under non-uniform sampling conditions, the limitation of their fixed search radius leads to over-filtering or distortion retention of point clouds in edge regions. Therefore, a dynamic parameter optimization strategy is proposed. This strategy establishes a mapping function between the search radius R and the local Euclidean distance distribution of the point cloud, enabling adaptive adjustment of the search radius. The minimum search radius is determined by the vertical distance between the detection plane and the sensor, while the dynamic search radius for edge regions is nonlinearly expanded based on the spatial distribution characteristics of the point cloud, thus effectively preserving key geometric features while maintaining filtering accuracy. (See the image for the point cloud extracted by dynamic radius filtering.) Figure 7 As shown in (b).

[0154] The extracted point cloud contains 599,301 points. Distortion points at point cloud boundaries, pin hole boundaries, and inside pin holes were filtered out to ensure high-quality input point cloud for subsequent fitting and extraction. A VoxelGrid voxel mesh filter with a minimum point count constraint was used. This filter spatially downsampled the point cloud data while maintaining accuracy, reducing the amount of data to be reduced in subsequent steps. See the image for the voxel-filtered point cloud. Figure 7 As shown in (c).

[0155] The number of points after filtering is 462,063, a 22.9% reduction compared to before filtering, thus optimizing the speed of subsequent fitting operations. Principal Component Analysis (PCA) is used to calculate the point cloud normals. PCA is an unsupervised dimensionality reduction method based on the eigenvalue decomposition of the covariance matrix. Normal estimation is achieved by calculating the minimum variance direction of points in the local neighborhood, which corresponds to the eigenvector corresponding to the minimum eigenvalue of the covariance matrix. Due to the large size of the point cloud data, the KdTree acceleration mechanism needs to be enabled to reduce the time complexity of traditional nearest neighbor search.

[0156] The modified RANSAC method was used to estimate the features of the planar point cloud. The fitted plane was obtained as -0.012x + 0.063y - 0.998z + -9.463 = 0. After filtering, the number of interior points was 448,910, with a filtering rate of 97.15%. This high filtering rate is attributed to effective point cloud extraction and early filtering using adaptive radius filtering. After fitting, point cloud boundary extraction was performed. The boundary extraction results of the fitted planar point cloud are a series of clusters of varying sizes; the clustering results are shown in [reference needed]. Figure 7 As shown in (d).

[0157] After 3D point cloud feature extraction, 3422 valid points were identified in the outer circular boundary region, while the pin hole boundary region contained 1136 discrete sampling points. A robust plane fitting algorithm, Random Sample Consensus (RANSAC), was used to fit the outer circular boundary point set. This algorithm effectively overcomes outlier interference and obtains optimal plane parameter estimates through iterative random sampling and model validation. Subsequently, based on the least squares projection principle, all end-face boundary point sets were orthogonally projected onto the fitting plane, achieving dimensionality reduction mapping from the 3D spatial points of the end face to the 2D parameter domain, providing a standardized data foundation for subsequent geometric feature analysis. Similarly, the same processing method was applied to all pin hole boundary points. This process not only eliminated spatial bias caused by measurement noise but also preserved the topological relationships of key geometric features.

[0158] The fitting plane is -0.012x + 0.064y - 0.997z - 9.461 = 0. After completing the orthogonal projection of the outer circle's point set and the inner circle of the pin hole onto the plane, this study employs a constrained Random Sample Consensus Algorithm (RANSAC) to perform robust geometric fitting of the two two-dimensional circles. This improved algorithm introduces a geometric constraint of center distance, simultaneously optimizing model parameters and constraint satisfaction during iterative sampling. By minimizing the reprojection error function, the constrained nonlinear least squares optimization problem is solved, and the algorithm ultimately converges to obtain the optimal radius solution and center coordinates of the two circles. The fitting results include the coordinates of the eccentric hole and the radii of the two circles, yielding the positional degree.

[0159] For physical preparation of perpendicularity monitoring, insert a pin into the crankshaft pin hole, see [link / reference]. Figure 2 As shown in (b).

[0160] See Figure 4 Step 4: Calculation process of perpendicularity:

[0161] Step 41: Insert the pin into the pin hole of the crankshaft, and collect and preprocess the original 3D point cloud from multiple angles;

[0162] The raw point cloud collected in step 1 includes the crankshaft end face point cloud and the pin cylinder surface point cloud, and then preprocessing is performed in step 2.

[0163] Step 42: Perform point cloud registration, which consists of coarse registration using SAC-IA based on PFFH and fine registration using ICP, and output the registered point cloud and root mean square error.

[0164] We construct local feature descriptions of the point cloud using FPFH (Fast Point Feature Histograms), and perform feature matching and transformation matrix estimation using the SAC-IA algorithm to achieve initial coarse registration under rotation and translation invariance. Through keypoint selection and feature dimension weighting mechanisms, we significantly improve robustness and convergence speed. The coarse registration provides initial alignment for the fine registration.

[0165] After coarse registration using SAC-IA, the source and target point clouds are initially aligned to similar spatial positions and orientations, but local misalignment issues still exist. In the pin perpendicularity detection scenario, due to the limitations of the 3D scanner's scanning perspective (data acquisition from top to bottom), the side of the pin cylinder only covers sparse point clouds, while the upper surface of the crankshaft, due to its planar structure, occupies the majority of the data. To improve 3D registration accuracy, the ICP (Iterative Closest Point) precise registration algorithm needs to be introduced to optimize the initial results. The ICP algorithm performs point-by-point matching calculations. For the high-density and complex surface features of the crankshaft scanning point cloud, an improved ICP framework is proposed: An adaptive weight adjustment mechanism is introduced, using the distance from the vertex to the viewpoint to weight the ICP error metric function, making the algorithm more focused on vertices closer to the camera. A weight function based on viewpoint distance is constructed, quantifying the scanning perspective difference into point-pair matching confidence weights, thus strengthening the registration contribution of key areas on the pin cylinder surface. After coarse registration using SAC-IA and fine registration using modified ICP, the registered point cloud and root mean square error are output. Since the overall point cloud after registration contains a large number of overlapping or closely spaced duplicate points, downsampling is performed to ensure the weights of these points return to normal and to reduce the computational burden, and the processed point cloud is then output.

[0166] An adaptive weight adjustment mechanism is introduced during ICP fine registration, using the distance from the vertex to the viewpoint to weight the error metric function of ICP, making the algorithm focus more on vertices that are closer to the camera.

[0167] Step 43: Perform downsampling processing;

[0168] The point cloud downsampling process consists of the following steps: First, input point cloud data containing normal vectors; second, configure the running parameters; then, calculate the vertex curvature based on the normal vectors; then, construct a curvature histogram according to a set number of intervals; subsequently, uniformly and randomly select a specified number of vertices from the histogram; finally, output the downsampling result.

[0169] Step 44: Extract the surface point cloud of the pin cylinder and crankshaft end face based on the improved Euclidean clustering algorithm;

[0170] The processed point cloud is extracted to obtain the point cloud of the pin surface. Traditional Euclidean clustering uses Euclidean distance as the core to determine point cloud inclusion or exclusion, but the point cloud at the pin root is sparse and distorted due to occlusion from the shooting angle and the gap in the pin hole. For pin-type parts with z-axis height symmetry, an improved Euclidean clustering algorithm introduces a z-axis height distance threshold constraint and a normal dual-condition constraint mechanism: the normal vector of the pin side point cloud is always perpendicular to the z-axis, while the normal vector of the end face point cloud is parallel to the z-axis. Based on these two complementary geometric feature constraints, a dual-condition Euclidean clustering algorithm is designed. By simultaneously satisfying the distance constraint and the normal direction constraint, this algorithm can more accurately achieve feature recognition and segmentation of the end face and pin side, significantly improving the accuracy and efficiency of point cloud segmentation for industrial parts, and effectively solving the misclassification problem of traditional algorithms in axial symmetry feature recognition.

[0171] The improved Euclidean clustering algorithm introduces a z-axis height constraint into the traditional Euclidean clustering algorithm: Constraint 1: In the z-axis direction, the first constraint is established by measuring the perpendicular distance between the end face and the side of the pin; Constraint 2: The normal vector of the pin side point cloud always maintains a perpendicular relationship with the z-axis, and the normal vector of the end face point cloud maintains a parallel relationship with the z-axis.

[0172] Step 45: Use the improved RANSAC algorithm to fit the pin cylinder surface and the crankshaft end face, and obtain the deviation angle between the pin axis and the normal of the crankshaft end face as the crankshaft perpendicularity.

[0173] For the extracted pin point cloud, an improvement to the traditional RANSAC cylinder fitting method is proposed: The random sampling consensus algorithm constructs an initial model through random sampling, selects interior points based on geometric residuals, and optimizes parameters using the least squares method, ultimately outputting the model with the highest interior point support rate. However, when the point cloud density is uneven or there is a lot of noise, relying solely on geometric distance to determine interior points leads to insufficient utilization of normal vector information, affecting fitting accuracy. The improved scheme introduces a normal weight mechanism, using curvature or neighborhood consistency-weighted sampling to prioritize high-weight normal vector points, effectively suppressing noise interference. Its probabilistic framework can effectively suppress noise interference, especially suitable for data scenarios with a high proportion of outliers. This improves the accuracy of the cylinder axis direction and the stability of radius estimation, making it particularly suitable for complex scenarios such as industrial point clouds.

[0174] For the improved RANSAC algorithm flowchart in this step, please refer to [link / reference]. Figure 6As shown, the improved scheme introduces a normal weighting mechanism into the original process. Normal vectors are weighted based on the local curvature of the point cloud or neighborhood consistency. During the RANSAC random sampling stage, points corresponding to high-weighted normal vectors are prioritized, thereby reducing the influence of noise points. This method, by strengthening the guiding role of normal vector information, can more accurately identify points on the real cylindrical surface. It is suitable for point cloud data with noise interference or industrial data. The weighted fitting results show significant improvements in both axial direction accuracy and radius estimation stability.

[0175] Verticality example: A 3D point cloud scanner captures data at a 45° overhead angle and rotates 90° to collect raw data, such as... Figure 8 As shown, the effective point cloud was extracted using a unified coordinate system, with the number of points being 337466, 293794, 267127, and 285111 respectively. The results are as follows. Figure 9 As shown, the results of the three registrations can be found in [link to documentation]. Figure 10 As shown, the final registration is performed with the upper end face of the pin, and the number of registration points and the corresponding RMS (root mean square error) are output. The output RMS is at the sub-millimeter level. After analyzing multiple sets of experimental data, the RMS (root mean square value) of the output displacement is stable in the range of 0.04–0.07 mm. This result is significantly better than the upper limit of the tolerance required by the design.

[0176] Downsampling and returning the results, processing with point cloud filtering, dividing the curvature histogram into 16 intervals, and using a sampling rate of 20% of the original point cloud, the output results are as follows. Figure 11 As shown, the number of points in the point cloud after filtering is 591,852, a reduction of 54.71% compared to before weight reduction. This point cloud will be used for fitting and analysis in subsequent steps.

[0177] When performing point cloud fitting, since the sampling principle of the 3D scanner is based on the triangular reflection principle, it will produce large distortion when collecting data near the chamfer. This invention introduces distance constraints to filter out point clouds with a z-axis distance less than a certain value, thus filtering out point clouds near the chamfer.

[0178] The parameter estimation method is set to RANSAC (Random Sample Consensus), which estimates model parameters through random sampling and consistency verification. RANSAC requires additional parameters for fitting the cylindrical model. The radius range of the cylindrical model is set to (5, 7), and the algorithm will only search for cylinders with radii within this range. This constraint accelerates the search process and improves the accuracy of the results. Model coefficient optimization is enabled; after segmentation, the obtained cylinder parameters will be optimized, improving the estimation accuracy of parameters such as the cylinder's axial direction and radius.

[0179] The normal weight parameter plays a crucial role in model fitting. The larger the normal weight, the greater the influence of the normal direction on the result. In engineering practice, 0.2 to 0.4 is an empirical value that balances the influence of point position and normal direction well, but the specific value still needs to be determined. The maximum number of iterations represents the number of iterations required to find the correct model. The more iterations, the higher the probability of finding the correct model, but the longer the computation time. For now, it is set to 1000 iterations to ensure fitting of cylinders in complex scenes.

[0180] Calculate the perpendicularity between the fitted cylinder normal and the plane normal using the vector angle formula:

[0181]

[0182] , These are the normal to the cylinder and the normal to the plane, respectively. The angle between the cylinder normal and the plane normal.

[0183] The perpendicularity of the crankshaft fitting under different normal weights w and normal constraints n is obtained when the normal weight w is 0.30 and the normal constraint n is 0.96. The perpendicularity obtained by fitting is closest to the reference part. Therefore, this value is selected as the normal weight w and normal constraint n for perpendicularity detection.

[0184] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A high-precision 3D topography detection method for compressor crankshafts based on laser vision, characterized in that, The method includes the following steps: Step 1: Use a three-coordinate motion platform equipped with a line laser profilometer to scan the compressor crankshaft in a non-contact manner to obtain the original three-dimensional point cloud; Step 2: Preprocess the original 3D point cloud using adaptive radius filtering and minimum point number constraint voxel filtering; Step 3: Detect pin hole position degree based on improved RANSAC algorithm and two-dimensional boundary clustering, and fit the outer circle of end face and the pin hole circle to obtain the pin hole position degree. Step 4: Detect the perpendicularity of the pin-end face through coarse registration and weighted ICP fine registration.

2. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 1, characterized in that, The line laser profilometer in step 1 uses a Keyence LJ-S080 laser detector, which uses a blue laser light source.

3. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 2, characterized in that, The search radius of the adaptive radius filter described in step 2 is dynamically adjusted according to the Euclidean distance.

4. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 1, characterized in that, Step 3, obtaining the positional accuracy, includes the following steps: Step 31: Collect the point cloud of the crankshaft end face including the pin hole. After preprocessing, solve the crankshaft end face normal based on principal component analysis (PCA) and use KdTree to accelerate the solution of the normal. Step 32: Fit the preprocessed end face point cloud and the normal obtained in step 31 using the improved RANSAC algorithm, and perform Rodriguez point cloud coordinate transformation based on the plane parameters. Step 33: Extract the boundary features of the annular ring formed by the outer circle of the end face and the pin hole circle. The boundary angle threshold is dynamically adjusted based on the k nearest neighbors and the point cloud density. Step 34: Based on the traditional Euclidean clustering, the extracted boundary features are clustered to extract the set of boundary feature points of the end face plane region and the pin hole circle region. Step 35: Perform dynamic clustering and filtering on the extracted boundary feature point set, and output the optimal boundary feature point set; Step 36: Use the improved RANSAC algorithm to fit the two-dimensional circle composed of the outer circle of the end face and the circle of the pin hole. The set distance between the two fitting results of the outer circle of the end face and the circle of the pin hole is used as the pin hole position degree.

5. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 4, characterized in that, The improved RANSAC algorithm changes the traditional RANSAC algorithm's method of determining interior points by assigning weights to the point cloud normals: points that satisfy the thresholds for distance from the point to the plane and the angle threshold between the point and the normal vector of the fitted plane are determined as interior points, forming a preferred set of interior points. Weighted singular value decomposition is applied to the preferred set of interior points, and the plane equation is optimized and the centroid of the interior points is calculated using the consistency of the normal vector as the weight.

6. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 4, characterized in that, Step 33 sets the boundary angle threshold dynamically based on the k-nearest neighbor and point cloud density. The threshold is lowered in dense point cloud areas to capture subtle geometric changes, and the threshold is increased in sparse areas to avoid noise misjudgment.

7. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 4, characterized in that, Step 34 involves clustering the extracted boundary features using a two-dimensional boundary clustering method improved from traditional Euclidean clustering: The input point cloud is traversed point by point, skipping processed points. When an unprocessed point is found, a new clustering container is created and added to the queue to be processed as the starting point for growth. Then, the points in the queue are processed in a loop. For each unprocessed point, a radius search is performed to obtain the neighborhood point set. If the neighborhood density is lower than the threshold or the normal exceeds the threshold, it is marked as noise. Otherwise, the current point is added to the cluster and the neighborhood points are traversed. Unprocessed adjacent points are added to the queue to recursively expand the clustering region. Finally, the current point is marked as processed. After completing the region growing, the algorithm verifies the clustering effectiveness by using a minimum size threshold. Clusters that meet the threshold are output as valid clusters, while those that do not are discarded as noise.

8. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 1, characterized in that, Step 4: Obtaining the perpendicularity of the pin end face. Step 41: Insert the pin into the pin hole of the crankshaft, and collect and preprocess the original 3D point cloud from multiple angles; Step 42: Perform point cloud registration, which consists of coarse registration using SAC-IA based on PFFH and fine registration using ICP, and output the registered point cloud and root mean square error. Step 43: Perform downsampling processing; Step 44: Extract the surface point cloud of the pin cylinder and crankshaft end face based on the improved Euclidean clustering algorithm; Step 45: Use the improved RANSAC algorithm to fit the pin cylinder surface and the crankshaft end face, and obtain the deviation angle between the pin axis and the normal of the crankshaft end face as the crankshaft perpendicularity.

9. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 1, characterized in that, Step 42 introduces an adaptive weight adjustment mechanism during ICP fine registration, using the distance from the vertex to the viewpoint to weight the error metric function of ICP, making the algorithm focus more on vertices that are closer to the camera.

10. The high-precision 3D topography detection method for compressor crankshaft based on laser vision according to claim 1, characterized in that, Step 44 improves the Euclidean clustering algorithm by introducing z-axis height constraints into the traditional Euclidean clustering algorithm: Constraint 1: In the z-axis direction, the first constraint is established by measuring the perpendicular distance between the end face and the side of the pin; Constraint 2: The normal vector of the pin side point cloud always maintains a perpendicular relationship with the z-axis, and the normal vector of the end face point cloud maintains a parallel relationship with the z-axis.

Images