A method for extracting multi-instance cylinders from unorganized point clouds

Abstract

This invention discloses a method for extracting multiple instances of cylinders from unorganized point clouds, belonging to the field of 3D point cloud processing technology. Addressing the problems of low accuracy, poor efficiency, and easy generation of pseudo-models in the extraction of multiple cylinders from unorganized point clouds containing severe noise, numerous outliers, and complex topological structures, the following scheme is proposed: A projection direction set is generated based on geometric priors, and irrelevant point clouds are filtered by combining point cloud normal orthogonal constraints; on the orthogonal plane of the current projection direction, the candidate interior points are sampled in interval groups using polar angles to construct a minimum point set for cylinder fitting; the minimum point set is used to fit the cylinder model parameters using iterative least squares; after calculating the cylinder model parameters, the candidate cylinder models are scored by combining the truncation quadratic cost and the cylinder angle coverage, and the cylinder model with the high quality score is selected, outputting the recognition result. This invention has high robustness and fast computational efficiency, and is mainly used for reverse engineering, 3D reconstruction, and completion measurement of large industrial facilities.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud processing technology, and in particular to a method for extracting multi-instance cylinders from unorganized point clouds. Background Technology

[0002] In recent years, high-precision 3D digitization technologies (such as terrestrial laser scanning and close-range photogrammetry) have been widely used in the manufacturing and measurement of large-scale industrial facilities. Utilizing non-contact, comprehensive data acquisition capabilities, these technologies have significantly reduced the manual workload of on-site measurements. However, acquiring high-density point clouds is only the initial stage; the fundamental bottleneck lies in how to transform these massive, unstructured, and unorganized discrete datasets into meaningful geometric models with semantic information.

[0003] In industrial reverse engineering, cylindrical structures (such as pipes and tanks) are the most common geometric primitives, and robust cylinder extraction is fundamental to completion measurement and facility management. Extracting individual cylinders directly from raw 3D point clouds presents significant challenges: variations in cylinder radius and length complicate the detection process; data is frequently plagued by severe noise, non-uniform sampling density, large-area occlusion, and a lack of topological context. Particularly in highly congested engineering scenarios, intersecting tubular components form complex topological networks, and surface penetration leads to severe degradation of local geometric features. Decoupling these highly connected regions to accurately segment and fit individual cylindrical primitives places extremely stringent demands on the robustness of detection algorithms.

[0004] Existing methods for extracting cylinders mainly face the following limitations: Hough transform-based methods rely on mapping spatial points to a parameter space for voting. The five-dimensional parameter space required for cylindrical representation leads to prohibitive computational overhead and memory consumption, making it difficult to handle massive industrial point clouds.

[0005] Region-growing-based methods heavily rely on accurate normal estimation and the selection of optimal seed points, and are prone to oversegmentation or undersegmentation when faced with noisy environments and complex topologies.

[0006] Traditional RANSAC-based methods often face problems such as a sharp drop in computational efficiency and the generation of pseudo-cylinders (incorrect models) when applied to highly polluted and complex scenarios containing multiple intersecting instances.

[0007] Deep learning methods suffer from a severe lack of large-scale, high-quality labeled datasets for complex industrial environments with intersecting cylindrical primitives, limiting their generalization capabilities and requiring expensive computing hardware. Summary of the Invention

[0008] To address the problems of existing technologies, this invention proposes a multi-instance cylinder extraction method for unorganized point clouds. This method solves the problems of low accuracy, poor efficiency, and easy generation of pseudo-models in the extraction of multi-cylinders when processing unorganized point clouds containing severe noise, a large number of outliers, and complex topology.

[0009] To achieve the above objectives, the present invention provides a method for extracting multi-instance cylinders from unorganized point clouds, comprising: The projection direction set is generated based on geometric priors, and irrelevant point clouds are filtered by combining point cloud normal orthogonal constraints, retaining the candidate set of interior points; On the orthogonal plane of the current projection direction, the candidate set of interior points is sampled in interval groups using the polar angle to construct the minimum point set for cylinder fitting; By using the minimum point set, the parameters of the cylindrical model are fitted through iterative least squares method to obtain candidate cylindrical models. The candidate cylinder models are scored by combining the truncation cost and the cylinder angle coverage, and the best multi-instance cylinder is output.

[0010] Optionally, the step of generating a projection direction set based on geometric priors and filtering irrelevant point clouds by combining point cloud normal orthogonal constraints to retain the candidate set of interior points includes: Set the main direction of the point cloud, and starting from this main direction, diverge outwards in all directions according to the set angle tolerance to generate a complete hemispherical distribution projection direction set. For each point in the point cloud, calculate the dot product of its normal vector and the principal direction vector, and filter out irrelevant points whose normal vector and principal direction vector are not orthogonal.

[0011] Optionally, the step of sampling the candidate interior points at intervals using polar angles on the orthogonal plane of the current projection direction to construct a minimum point set for cylinder fitting includes: Construct an orthogonal basis perpendicular to the current projection direction; Calculate the polar angle of the normal of the filtered point cloud on the projection plane formed by the orthogonal basis, and group it into the corresponding angular sectors; Extract the midpoint within each corner sector, detect whether the midpoint can be fitted into a two-dimensional circle that meets physical constraints, and record the corner sectors that pass the detection. Point clouds are randomly extracted from the detected corner sectors to construct a minimum point set for cylinder fitting.

[0012] Optionally, the specific conditions for meeting the physical constraints include: (1) the total number of points inside all corner sectors is greater than or equal to the minimum number of points required to fit the cylinder; (2) the radius of the fitted two-dimensional circle is within a preset range of the point cloud bounding box; and (3) the distance from the midpoint of the fitted two-dimensional circle that meets conditions (1) and (2) is less than a set distance threshold. m The ratio is greater than or equal to the set ratio threshold.n .

[0013] Optionally, the step of using the minimum point set to fit the cylindrical model parameters through iterative least squares method to obtain a candidate cylindrical model includes: Let the direction vector of the central axis of the cylinder be... a =( a x , a y , a z Points on the central axis of the cylinder c =( c x , c y , c z ), cylinder radius r ; For the set of minimum points, the first i points [ x i , y i , z i ] T The residual distance from the cylindrical surface e i The partial derivatives with respect to the seven parameters constitute the Jacobian row vector. J i : , The Jacobian matrix is ​​updated iteratively using the Gauss-Newton method or the Levberg-Marquardt algorithm. J i This yields high-precision local cylinder parameters.

[0014] Optionally, the scoring of candidate cylindrical models by combining the truncation quadratic cost and the cylinder angle coverage includes: Calculate the squared residuals from all interior points in the minimum point set to the candidate cylindrical surface. The residual will be less than the preset residual threshold τ 2 The interior points form an interior point set I, and the truncation quadratic cost is calculated. The formula is as follows: , Calculate the polar angle θ of the interior point set projected onto the orthogonal two-dimensional plane and arrange them in ascending order; Calculate the encirclement closure difference between adjacent polar angles and the first and last polar angles, and extract the maximum angular interval Δ. max The effective coverage radian on the cylindrical surface is calculated, and the effective coverage radian is normalized as the weight γ of the cylindrical angle coverage rate, as shown in the following formula: , Cut off the second cost Multiplying the score by the cylinder angle coverage weight γ yields the final score for the candidate cylinder model.

[0015] By adopting the above technical solution, the present invention has at least the following beneficial effects: 1. Extremely high robustness: By using projection direction constraints and connectivity polar angle interval sampling, the problem that traditional RANSAC is prone to getting trapped in local suboptimal solutions in dense point clouds is effectively avoided, and primitives can be accurately extracted in severe noise and massive outliers.

[0016] 2. Significantly improved recognition efficiency: The sequential direction extraction strategy avoids the huge computational burden of high-dimensional parameter space; the strict two-dimensional circle pre-detection mechanism greatly filters out invalid random samples and accelerates convergence.

[0017] 3. Strong multi-instance and anti-interference capabilities: The specially customized angle coverage scoring function severely punishes models with incomplete spatial distribution, and completely solves the problem of "pseudo-cylinders" generated frequently in pipeline intersection and interpenetration areas. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a method for extracting multi-instance cylinders from unorganized point clouds provided in an embodiment of the present invention; Figure 2 This is the original point cloud image; Figure 3 This is a diagram showing the results of extracting cylinders from multiple instances. Detailed Implementation

[0020] 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.

[0021] refer to Figure 1 This invention provides a method for extracting multi-instance cylinders from unorganized point clouds. Figure 2The point cloud shown is subjected to cylindrical extraction, including: I. Projection direction generation and filtering based on geometric priors.

[0022] In this embodiment, a projection direction set is generated based on geometric priors, and irrelevant point clouds are filtered out by combining point cloud normal orthogonal constraints, retaining the candidate set of interior points. Since industrial structures have obvious main direction features, sequential extraction is performed, which not only speeds up the calculation but also reduces interference from subsequent direction recognition.

[0023] Specifically, the thinnest dimension in the point cloud is set as the main direction. Starting from this main direction, the points diverge outwards according to a set angular tolerance to generate a complete hemispherical projection direction set. For each point in the point cloud, the inner product of its normal vector and the main direction vector is calculated, and irrelevant points whose normal vector and the main direction vector are not orthogonal are filtered out.

[0024] II. High-quality sampling based on polar angle connectivity.

[0025] In this embodiment, on the orthogonal plane of the current projection direction, the candidate set of interior points is sampled in interval groups using the polar angle to construct the minimum point set for cylinder fitting; Specifically, an orthogonal basis perpendicular to the current projection direction is constructed; the polar angle of the normal of the filtered point cloud on the projection plane formed by the orthogonal basis is calculated, and the point cloud is divided into corresponding corner sectors for grouping; the midpoints in each corner sector are extracted, and it is detected whether the midpoints can be fitted into a two-dimensional circle that meets the physical constraints, and the corner sectors that pass the detection are recorded; point clouds are randomly extracted from the corner sectors that pass the detection to construct a minimum point set for cylinder fitting.

[0026] The specific conditions for meeting physical constraints include: (1) The total number of points inside all corner sectors is greater than or equal to the minimum number of points required to fit a cylinder. Although fitting the parameters of a cylinder only requires a few points, the cylindrical part in the actual point cloud data has many points. Therefore, a minimum number of points required to fit a cylinder is set, such as 1000 points.

[0027] (2) The radius of the fitted 2D circle is within the preset range of the point cloud bounding box. First, record the maximum and minimum values ​​of the point cloud on the x, y, and z axes, for a total of 6 values. These six values ​​can be used to generate a cube that can enclose all points; this cube is the point cloud bounding box. In order to prevent some point clouds from fitting a cylinder with a very large radius (for example, it may accidentally fit a plane), first calculate the bounding box of this point cloud, and limit the radius of the fitted cylinder by the dimensions (length, width, height, or diagonal) of this bounding box.

[0028] (3) The proportion of the distance from the midpoint that meets conditions (1) and (2) to the fitted two-dimensional circle that is less than the set distance threshold m is greater than or equal to the set proportion threshold n. First, calculate the distance from the midpoint that passes through the previous constraints to the fitted two-dimensional circle, and count the distances that are less than the set distance threshold. The number of median points is calculated. Finally, the proportion of these points to all median points that passed the previous constraints is calculated, and it is checked whether this proportion is greater than or equal to a set proportion threshold. The formula is as follows: .

[0029] in, For the median site that meets conditions (1) and (2), To fit the center of a two-dimensional circle, To fit the radius of a two-dimensional circle.

[0030] (3) Iterative least squares cylinder fitting.

[0031] In this embodiment, the minimum point set is used to fit the parameters of the cylindrical model through iterative least squares method to obtain the candidate cylindrical model.

[0032] Specifically, let the direction vector of the cylinder's central axis be... a =( a x , a y , a z Points on the central axis of the cylinder c =( c x , c y , c z ), cylinder radius r For the set of minimum points, the first i points [ x i , y i , z i ] T The residual distance from the cylindrical surface e i The partial derivatives with respect to the seven parameters constitute the Jacobian row vector. J i : , The Jacobian matrix is ​​updated iteratively using the Gauss-Newton method or the Levberg-Marquardt algorithm. J i This yields high-precision local cylinder parameters.

[0033] (4) Model scoring based on angle coverage.

[0034] In this embodiment, in order to accurately measure the integrity of the extracted cylinder and prevent local arcs from being incorrectly fitted as whole circles due to dense point clouds, the candidate cylinder model is scored by combining the truncation secondary cost and the cylinder angle coverage, and the best multi-instance cylinder is output.

[0035] Specifically, calculate the squared residuals from all interior points in the minimum point set to the candidate cylindrical surface. The residual will be less than the preset residual threshold τ 2 The interior points form an interior point set I, and the truncation quadratic cost is calculated. The formula is as follows: , Calculate the polar angle θ of the interior point set projected onto the orthogonal two-dimensional plane and arrange them in ascending order; Calculate the encirclement closure difference between adjacent polar angles and the first and last polar angles, and extract the maximum angular interval Δ. max The effective coverage radian on the cylindrical surface is calculated, and the effective coverage radian is normalized as the weight γ of the cylindrical angle coverage rate, as shown in the following formula: , Cut off the second cost Multiplying the score by the cylinder angle coverage weight γ yields the final score for the candidate cylinder model.

[0036] Finally, the optimal multi-instance cylinder is output, such as Figure 3 As shown.

[0037] The present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for extracting multi-instance cylinders from unorganized point clouds, characterized in that, include: The projection direction set is generated based on geometric priors, and irrelevant point clouds are filtered by combining point cloud normal orthogonal constraints, retaining the candidate set of interior points; On the orthogonal plane of the current projection direction, the candidate set of interior points is sampled in interval groups using the polar angle to construct the minimum point set for cylinder fitting; By using the minimum point set, the parameters of the cylindrical model are fitted through iterative least squares method to obtain candidate cylindrical models. The candidate cylinder models are scored by combining the truncation cost and the cylinder angle coverage, and the best multi-instance cylinder is output.

2. The method for extracting multi-instance cylinders from unorganized point clouds according to claim 1, characterized in that, The process of generating a projection direction set based on geometric priors and filtering irrelevant point clouds by combining point cloud normal orthogonal constraints to retain a candidate set of interior points includes: Set the main direction of the point cloud, and starting from this main direction, diverge outwards in all directions according to the set angle tolerance to generate a complete hemispherical distribution projection direction set. For each point in the point cloud, calculate the dot product of its normal vector and the principal direction vector, and filter out irrelevant points whose normal vector and principal direction vector are not orthogonal.

3. The method for extracting multi-instance cylinders from unorganized point clouds according to claim 1, characterized in that, The step of sampling the candidate interior points at intervals using polar angles on the orthogonal plane of the current projection direction to construct a minimum point set for cylinder fitting includes: Construct an orthogonal basis perpendicular to the current projection direction; Calculate the polar angle of the normal of the filtered point cloud on the projection plane formed by the orthogonal basis, and group it into the corresponding angular sectors; Extract the midpoint within each corner sector, detect whether the midpoint can be fitted into a two-dimensional circle that meets physical constraints, and record the corner sectors that pass the detection. Point clouds are randomly extracted from the detected corner sectors to construct a minimum point set for cylinder fitting.

4. The method for extracting multi-instance cylinders from unorganized point clouds according to claim 3, characterized in that, The specific conditions for meeting the physical constraints include: (1) the total number of points inside all corner sectors is greater than or equal to the minimum number of points required to fit the cylinder; (2) the radius of the fitted two-dimensional circle is within the preset range of the point cloud bounding box; and (3) the distance from the midpoint of the fitted two-dimensional circle that meets conditions (1) and (2) is less than a set distance threshold. m The ratio is greater than or equal to the set ratio threshold. n .

5. The method for extracting multi-instance cylinders from unorganized point clouds according to claim 1, characterized in that, The process of using a minimum point set to fit the parameters of a cylindrical model through iterative least squares method to obtain candidate cylindrical models includes: Let the direction vector of the central axis of the cylinder be... a =( a x , a y , a z Points on the central axis of the cylinder c =( c x , c y , c z ), cylinder radius r ; For the set of minimum points, the first i points [ x i , y i , z i ] T The residual distance from the cylindrical surface e i The partial derivatives with respect to the seven parameters constitute the Jacobian row vector. J i : ， The Jacobian matrix is ​​updated iteratively using the Gauss-Newton method or the Levberg-Marquardt algorithm. J i This yields high-precision local cylinder parameters.

6. The method for extracting multi-instance cylinders from unorganized point clouds according to claim 1, characterized in that, The scoring of candidate cylindrical models by combining truncation quadratic cost and cylinder angle coverage includes: Calculate the squared residuals from all interior points in the minimum point set to the candidate cylindrical surface. The residual will be less than the preset residual threshold τ 2 The interior points form an interior point set I, and the truncation quadratic cost is calculated. The formula is as follows: ， Calculate the polar angle θ of the interior point set projected onto the orthogonal two-dimensional plane and arrange them in ascending order; Calculate the encirclement closure difference between adjacent polar angles and the first and last polar angles, and extract the maximum angular interval Δ. max The effective coverage radian on the cylindrical surface is calculated, and the effective coverage radian is normalized as the weight γ of the cylindrical angle coverage rate, as shown in the following formula: ， Cut off the second cost Multiplying the score by the cylinder angle coverage weight γ yields the final score for the candidate cylinder model.

Images