A steel truss extraction method based on point cloud segmentation and feature analysis

Abstract

The application discloses a steel truss extraction method based on point cloud segmentation and feature analysis, and belongs to the technical field of three-dimensional point cloud processing. The method mainly comprises the following steps: firstly, performing multi-scale filtering denoising on original point cloud and constructing an octree index; then, calculating local geometric features such as normal vectors, curvatures and first-order plane fitting errors of each point; based on adaptively selected seed points, combining the normal vector angle, curvature difference and adaptive edge criterion to realize region growing and preliminary segmentation; finally, adopting a multi-level cascade classifier from coarse to fine to screen the segmented regions, and sequentially passing through geometric size filtering, shape feature screening and multi-dimensional feature matching to accurately extract the steel truss point cloud. The application has the advantages of strong anti-interference ability, high extraction precision and high calculation efficiency without pre-training, and is suitable for engineering scenes such as bridge detection and building surveying.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud processing technology, and in particular to a method for extracting steel trusses based on point cloud segmentation and feature analysis. Background Technology

[0002] In the construction and operation of large-scale steel structure projects, such as bridges, large industrial plants, and stadiums, steel trusses serve as core load-bearing components. Accurate acquisition of their geometric shape is crucial for project quality control, safety monitoring, deformation analysis, and Building Information Modeling (BIM). In recent years, 3D laser scanning technology has become an important data acquisition method in this field due to its ability to quickly and non-contactly acquire high-precision, high-density 3D point cloud data of object surfaces.

[0003] However, automatically and accurately extracting the target steel truss structure from complex on-site scan point clouds remains a highly challenging technical problem. This is specifically manifested in the following aspects:

[0004] 1. Complex point cloud data quality: The construction site environment is complex. The acquired raw point cloud not only contains the target steel truss, but also inevitably mixes in a large number of non-target point clouds such as scaffolding, construction machinery, temporary facilities, vegetation, and even personnel, resulting in serious noise interference. At the same time, the point cloud density may be uneven, and there may be missing or occluded points, which directly affects the accuracy of subsequent processing.

[0005] 2. Limitations of Existing Processing Technologies: Currently, methods for extracting specific structures from point clouds can be mainly categorized as follows, but all have significant shortcomings: Methods based on traditional geometric rules: These methods typically rely on manually preset geometric rules (such as radius-based filtering, elevation-based segmentation, etc.) for screening. Their threshold settings heavily depend on expert experience and lack adaptability. In complex and varied scenarios, these rules are difficult to apply universally, easily leading to over-segmentation (erroneously deleting target objects) or under-segmentation (retaining too much noise), resulting in low automation and accuracy.

[0006] Deep learning-based point cloud segmentation methods: These methods are currently a research hotspot, achieving end-to-end semantic segmentation of point clouds by training deep neural networks (such as PointNet++, PointCNN, etc.). While achieving good results on some public datasets, they face significant bottlenecks in practical engineering applications: First, model performance is highly dependent on a large amount of high-quality, labeled training data, and data collection and labeling for specific steel truss scenarios requires enormous human and material resources, resulting in high data preparation costs; second, the generalization ability of well-trained models is limited, and their segmentation performance may drop sharply when faced with new scenarios that are inconsistent with the distribution of training data (such as different truss types, new lighting conditions, or different noise patterns), thus limiting their practicality; finally, deep learning models typically consume large amounts of computational resources, and their inference speed may be insufficient to meet the needs of real-time or near-real-time processing of large-scale point cloud data.

[0007] 3. The trade-off between efficiency and accuracy: Massive point cloud data poses a severe challenge to the computational efficiency of algorithms. Many complex processing algorithms, while achieving acceptable accuracy, are too time-consuming to handle large-scale point clouds suitable for engineering applications, failing to meet the urgent efficiency requirements of practical engineering. How to achieve efficient processing while maintaining high accuracy is a core problem that urgently needs to be solved.

[0008] In summary, existing technologies struggle to achieve efficient, high-precision, and robust automatic extraction of steel trusses from complex scene point clouds without requiring extensive pre-training data. Therefore, developing a data-independent, adaptive, and efficient steel truss extraction method that balances processing efficiency and accuracy has significant theoretical value and broad engineering application prospects. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a steel truss extraction method based on point cloud segmentation and feature analysis. Compared with deep learning-based methods, this invention does not require a pre-training process, significantly reduces data preparation work, and improves engineering practicality and extraction efficiency.

[0010] The objective of this invention is achieved through the following technical solution: a method for extracting steel trusses based on point cloud segmentation and feature analysis, comprising the following steps:

[0011] Point cloud preprocessing and index construction stage: Obtain the original point cloud containing the steel truss, perform multi-scale filtering and denoising on the original point cloud, establish topological relationships for the denoised point cloud, and construct the point cloud index structure;

[0012] Local geometric feature calculation stage: Based on the point cloud index structure, calculate the local geometric features of each point in the point cloud. The local geometric features include at least the normal vector, curvature and first-order plane fitting error.

[0013] Adaptive seed point selection stage: Traverse all unsegmented points and select the point with the smallest first-order plane fitting error as the initial seed point;

[0014] The region growth stage based on multi-feature fusion and adaptive threshold is as follows: Centered on the initial seed point, each neighboring point is traversed, and it is determined whether it meets the similarity criterion and edge criterion. Neighboring points that simultaneously meet both the similarity criterion and edge criterion are included in the current growth region. Points with curvature less than the queue curvature threshold are added to the seed point queue as new growth points. After the current seed point has finished growing, the next seed point is taken from the seed point queue, and the seed point growth process is repeated until the seed point queue is empty, completing the growth of one region. The similarity criterion is configured according to the geometric characteristics of each component of the steel truss.

[0015] Iterative segmentation stage: After the growth of a region is completed, it is marked as an independent segmentation region. Then, the adaptive seed point selection stage and the region growth stage based on multi-feature fusion and adaptive threshold are repeated in the remaining unsegmented point cloud until all points are assigned to a certain region or are judged as noise points, thus obtaining multiple candidate segmentation regions.

[0016] The segmentation result screening stage based on a multi-level cascaded classifier: A multi-level cascaded classifier is used to perform cascaded screening of candidate segmentation regions, and the final output is the target steel truss point cloud that is determined to be a steel truss region; the cascaded screening is based on a computational cost hierarchical strategy, first performing a primary screening with low computational cost to exclude obvious non-target regions, and then performing intermediate screening and / or final screening with higher computational cost to accurately identify the regions after the initial screening; the screening criteria of the cascaded classifier are set based on prior knowledge of the geometric features of the steel truss structure, including the linear features, planar features and size range of each component of the standard steel truss.

[0017] Preferably, the point cloud preprocessing and index construction stage further includes the following steps:

[0018] A multi-scale filtering method based on octrees is used for denoising. An initial filtering radius is set to perform preliminary denoising on the original point cloud. Then, the filtering radius is gradually reduced for iterative filtering until outliers are removed while the main point cloud features are preserved. Finally, a method combining 3D raster and octree is used to establish topological relationships for the denoised point cloud and construct a point cloud index structure.

[0019] Preferably, the normal vector is calculated by fitting a local plane using principal component analysis (PCA); the curvature is obtained by fitting a local quadratic surface and calculating the Gaussian curvature and mean curvature of the points.

[0020] Preferably, the similarity criterion for the target neighborhood point is determined by the following steps: when the angle between the target neighborhood point and the normal vector of the current seed point / current region average normal vector is less than the angle threshold, and the absolute value of the average curvature difference between the target neighborhood and the current region is less than the curvature threshold, the target neighborhood point is determined to satisfy the similarity criterion.

[0021] Preferably, the following steps are used to determine whether the target neighborhood points satisfy the edge criteria:

[0022] Calculate the first-order plane fitting error Re_k of the target neighborhood points;

[0023] The adaptive threshold Re_th is calculated as mean(Re_k) + σ*std(Re_k), where mean(Re_k) represents the average value of the first-order plane fitting error Re_k, std(Re_k) represents the standard deviation of the first-order plane fitting error Re_k, and σ is an empirical coefficient.

[0024] If the first-order plane fitting error Re_k < the adaptive threshold Re_th, then the target neighborhood point is determined as an internal point satisfying the edge criterion; otherwise, it is an edge point.

[0025] Preferably, the empirical coefficient σ is in the range of 0.5-1.5.

[0026] Preferably, the cascaded screening includes primary screening based on geometric dimensions, intermediate screening based on shape features, and final screening based on multidimensional feature matching, with progressively increasing computational costs.

[0027] Preferably, the primary screening based on geometric dimensions includes the following steps:

[0028] Calculate the length, width, height, and volume of the basic bounding box for each candidate segmentation region, and compare it with the preset steel truss component size threshold range, eliminating regions that do not match the size.

[0029] Preferably, the intermediate screening based on shape features includes the following steps:

[0030] For the candidate segmentation regions that pass the initial screening, the shape features describing their overall morphology are calculated. The shape features include at least linearity, flatness, and scattering. Based on the preset steel truss component shape threshold, regions with inconsistent shapes are eliminated to obtain the final candidate regions.

[0031] Preferably, the final screening based on multidimensional feature matching includes the following steps:

[0032] Extract the multidimensional feature vectors of the final candidate regions that have passed the intermediate screening, including the principal direction and axial dimension ratio based on PCA, and combine them with the linearity and flatness calculated during the intermediate screening to form a complete feature description;

[0033] The multidimensional feature vector is matched with a predefined standard feature template for steel trusses, and a comprehensive similarity score is calculated.

[0034] If the overall similarity score is higher than the preset threshold, the target area is determined to be the steel truss area.

[0035] The beneficial effects of this invention are:

[0036] 1) High computational efficiency, fast processing speed, and strong practicality: This invention significantly shortens the overall processing time by constructing an efficient octree index structure, adopting a region growth algorithm that avoids global computation, and innovatively applying a cascading screening framework.

[0037] 2) The cascaded filtering framework prioritizes simple features with low computational cost (such as bounding box size) to quickly eliminate a large number of non-target regions, ensuring that subsequent complex feature matching is performed only on a very small number of candidate regions. This greatly reduces unnecessary computational overhead, thereby achieving a leap in computational speed. This method effectively overcomes the bottleneck of large computational load and long processing time when directly processing massive point cloud data, and can meet the needs of efficiency and high-frequency data processing in practical engineering.

[0038] 3) Strong anti-interference capability and high extraction accuracy: This invention effectively overcomes noise interference in complex field environments by using multi-scale filtering and a region growing algorithm based on multi-feature fusion (normal vector, curvature) and adaptive edge criteria. Compared with simple segmentation methods using a single threshold, this invention can more precisely identify the true boundaries of the steel truss. Even when the point cloud density is uneven or there are local gaps, it can ensure the integrity and accuracy of the segmented region, thereby achieving high-precision extraction of the target structure from a noisy background.

[0039] 4) High practicality and adaptability: This invention is based entirely on the geometric features and prior knowledge of steel trusses for segmentation and recognition, without relying on large amounts of labeled data for model pre-training. This fundamentally overcomes the bottlenecks of deep learning methods, such as difficult data collection, high labeling costs, and weak model generalization ability. The method of this invention has stronger universality and can be quickly applied to different types of steel truss structures or new engineering scenarios, greatly improving the engineering practical value and ease of use of the method.

[0040] 5) High degree of automation and good robustness: From the adaptive selection of seed points to the adaptive threshold judgment of region growing, and then to the automated screening process of cascaded classifiers, this invention greatly reduces the need for manual intervention and parameter adjustment. The method is systematic and exhibits good robustness to different data scenarios and point cloud quality, ensuring the consistency and reliability of the results. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0042] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0043] See Figure 1 This invention provides a technical solution: a method for extracting steel trusses based on point cloud segmentation and feature analysis, comprising the following steps:

[0044] Point cloud preprocessing and index construction stage: Obtain the original point cloud containing the steel truss, perform multi-scale filtering and denoising on the original point cloud, establish topological relationships for the denoised point cloud, and construct the point cloud index structure;

[0045] Local geometric feature calculation stage: Based on the point cloud index structure, calculate the local geometric features of each point in the point cloud. The local geometric features include at least the normal vector, curvature and first-order plane fitting error.

[0046] Adaptive seed point selection stage: Traverse all unsegmented points and select the point with the smallest first-order plane fitting error as the initial seed point;

[0047] The region growth stage based on multi-feature fusion and adaptive threshold is as follows: Centered on the initial seed point, each neighboring point is traversed, and it is determined whether it meets the similarity criterion and edge criterion. Neighboring points that simultaneously meet both the similarity criterion and edge criterion are included in the current growth region. Points with curvature less than the queue curvature threshold are added to the seed point queue as new growth points. After the current seed point has finished growing, the next seed point is taken from the seed point queue, and the seed point growth process is repeated until the seed point queue is empty, completing the growth of one region. The similarity criterion is configured according to the geometric characteristics of each component of the steel truss.

[0048] Iterative segmentation stage: After the growth of a region is completed, it is marked as an independent segmentation region. Then, the adaptive seed point selection stage and the region growth stage based on multi-feature fusion and adaptive threshold are repeated in the remaining unsegmented point cloud until all points are assigned to a certain region or are judged as noise points, thus obtaining multiple candidate segmentation regions.

[0049] The segmentation result screening stage based on a multi-level cascaded classifier: A multi-level cascaded classifier is used to perform cascaded screening of candidate segmentation regions, and the final output is the target steel truss point cloud that is determined to be a steel truss region; the cascaded screening is based on a computational cost hierarchical strategy, first performing a primary screening with low computational cost to exclude obvious non-target regions, and then performing intermediate screening and / or final screening with higher computational cost to accurately identify the regions after the initial screening; the screening criteria of the cascaded classifier are set based on prior knowledge of the geometric features of the steel truss structure, including the linear features, planar features and size range of each component of the standard steel truss.

[0050] In this embodiment, the present invention achieves a leap in processing efficiency through coordinated optimization using multiple methods. First, by constructing an octree index structure, rapid lookup of point cloud neighborhood relationships is achieved. Second, the region growing algorithm avoids complex global calculations, limiting the computational load to local regions. Most importantly, it innovatively adopts a multi-layer cascaded classifier framework. Through a "coarse-to-fine" strategy, it first uses computationally inexpensive geometric size features (primary filtering) to quickly eliminate most non-target regions (such as ground, vegetation, etc.), and then performs more complex shape feature analysis (intermediate filtering) and fine feature matching (final filtering) on ​​only a very small number of candidate regions. This design ensures efficient use of computing resources, avoids unnecessary and expensive calculations on irrelevant point clouds, and thus significantly shortens the overall processing time of massive point cloud data. It effectively overcomes the problems of large computational load and long processing time in existing technologies, and can meet the needs of high-efficiency and high-frequency data processing in practical engineering.

[0051] In the adaptive seed point selection phase, all undivided points are traversed, and the point with the smallest first-order plane fitting error is selected as the initial seed point. This is because points with small first-order plane fitting errors are usually located within flat regions, and starting growth from there is beneficial for the stable expansion of the region. In practice, all points can be sorted according to the magnitude of their first-order plane fitting errors, or points on the steel truss can be manually selected as seed points.

[0052] In some embodiments, the point cloud preprocessing and index building stage further includes the following steps:

[0053] A multi-scale filtering method based on octrees is used for denoising. An initial filtering radius is set to perform preliminary denoising on the original point cloud. Then, the filtering radius is gradually reduced for iterative filtering until outliers are removed while the main point cloud features are preserved. Finally, a method combining 3D raster and octree is used to establish topological relationships for the denoised point cloud and construct a point cloud index structure.

[0054] In this embodiment, after acquiring the original point cloud, preprocessing is performed first. A multi-scale filtering method based on octrees is employed: first, an initial filtering radius is set (typically 3-4 times the average density of the point cloud) to perform preliminary denoising. Then, the filtering radius is gradually reduced, and the filtering process is repeated iteratively until outliers are effectively removed while the main point cloud features are preserved. Next, a method combining 3D raster and octrees is used to establish topological relationships for the denoised point cloud, constructing an efficient point cloud index structure to facilitate subsequent rapid neighborhood search.

[0055] In some embodiments, the normal vector is calculated by fitting a local plane using principal component analysis (PCA); the curvature is obtained by fitting a local quadratic surface and calculating the Gaussian curvature and mean curvature of the points.

[0056] In this embodiment, a K-neighborhood is calculated for each point using the established index. The normal vector is then calculated based on this neighborhood: typically, a local plane is fitted using Principal Component Analysis (PCA), and its normal vector is the normal vector of that point.

[0057] Curvature calculation: By fitting a local quadratic surface, the Gaussian curvature and mean curvature of the points are calculated.

[0058] First-order plane fitting error: This value reflects the degree of fitting of a point to the local plane formed by the point and its neighboring points. The larger the value, the more likely the point is to be located at the edge or where the features change drastically.

[0059] In some embodiments, the following steps are used to determine whether the target neighborhood point meets the similarity criterion: when the angle between the target neighborhood point and the normal vector of the current seed point / current region average normal vector is less than the angle threshold, and the absolute value of the average curvature difference between the target neighborhood and the current region is less than the curvature threshold, the target neighborhood point is determined to meet the similarity criterion.

[0060] In some embodiments, the following steps are used to determine whether the target neighborhood points satisfy the edge criteria:

[0061] Calculate the first-order plane fitting error Re_k of the target neighborhood points;

[0062] Calculate the adaptive threshold Re_th = mean(Re_k) + σ * std(Re_k), where mean(Re_k) represents the average value of calculating the first-order plane fitting error Re_k, std(Re_k) represents the standard deviation of calculating the first-order plane fitting error Re_k, and σ is an empirical coefficient;

[0063] If the first-order plane fitting error Re_k < the adaptive threshold Re_th, the target neighborhood point is determined to be an interior point satisfying the edge criterion; otherwise, it is an edge point.

[0064] In some embodiments, the value range of the empirical coefficient σ is 0.5 - 1.5.

[0065] In this embodiment, with the selected initial seed point as the center, traverse its K-nearest neighbor points. For each neighborhood point, determine whether it satisfies the following two criteria:

[0066] (1) Similarity criterion: Normal vector angle: The normal vector angle θ_i,k between the neighborhood point and the current seed point (or the average normal vector of the current region) needs to be less than the set threshold θ_threshold. Curvature difference: The absolute value of the difference between the average curvature of the neighborhood point and the current region needs to be less than the set threshold ΔH_threshold.

[0067] (2) Edge criterion: Calculate the first-order plane fitting error Re_k of this neighborhood point. Calculate the adaptive threshold Re_th = mean(Re_k) + σ * std(Re_k), where mean(Re_k) represents the average value (Mean) of calculating the first-order plane fitting error Re_k, std(Re_k) represents the standard deviation (Standard Deviation) of calculating the first-order plane fitting error Re_k, and σ is an empirical coefficient (usually adjusted between 0.5 - 1.5).

[0068] Judgment: If Re_k < Re_th, this point is determined to be an "interior point" and satisfies the edge criterion; otherwise, it is an edge point.

[0069] Only the neighborhood points that satisfy both the similarity criterion and the edge criterion will be added to the current growing region. If the newly added point has a small curvature (for example, less than the set threshold C_threshold), it will be added to the seed point queue as a new growing point. After the current seed point has finished growing, take out the next seed point from the queue and repeat the process until the seed point queue is empty, and a region growth is completed.

[0070] In some embodiments, the cascaded screening includes primary screening based on geometric size with gradually increasing calculation cost, intermediate screening based on shape features, and final screening based on multi-dimensional feature matching.

[0071] In this embodiment, to significantly improve computational efficiency while ensuring accuracy, a multi-layer cascaded classifier is employed to filter candidate segmentation regions from coarse to fine, step-by-step. This cascaded structure systematizes the screening process, ensuring that only genuine steel truss candidate regions proceed to the next stage, which has a higher computational cost.

[0072] In some embodiments, the geometry-based primary screening includes the following steps:

[0073] Calculate the length, width, height, and volume of the basic bounding box for each candidate segmentation region, and compare it with the preset steel truss component size threshold range, eliminating regions that do not match the size.

[0074] In this embodiment, the basic bounding box geometry of each candidate region is calculated, including the length, width, height, and volume of the smallest enclosing cuboid. Based on prior knowledge, reasonable size threshold ranges for typical steel truss components (such as members and gusset plates) are set. Judgment mechanism: Regions whose size or volume in any dimension significantly exceeds the threshold range (e.g., too large or too small) will be immediately rejected at this stage, thus quickly eliminating significantly non-target objects such as ground, vegetation, and equipment.

[0075] In some embodiments, the shape-feature-based intermediate filtering includes the following steps:

[0076] For the candidate segmentation regions that pass the initial screening, the shape features describing their overall morphology are calculated. The shape features include at least linearity, flatness, and scattering. Based on the preset steel truss component shape threshold, regions with inconsistent shapes are eliminated to obtain the final candidate regions.

[0077] In this embodiment, for the regions that pass the initial screening, shape features describing their overall morphology are calculated, mainly including: linearity: used to screen regions with rod-like characteristics; flatness: used to identify possible node plates or connecting surfaces; and scattering: used to exclude irregular, disorganized noise point clouds.

[0078] Judgment Mechanism: Based on the prior shape of each component of the steel truss, a combined threshold for linearity and flatness is set. Areas that do not conform to the typical shape characteristics of members or gusset plates (such as clumps with excessive scattering) will be rejected at this stage.

[0079] In some embodiments, the final screening based on multidimensional feature matching includes the following steps:

[0080] Extract the multidimensional feature vectors of the final candidate regions that have passed the intermediate screening, including the principal direction and axial dimension ratio based on PCA, and combine them with the linearity and flatness calculated during the intermediate screening to form a complete feature description;

[0081] The multidimensional feature vector is matched with a predefined standard feature template for steel trusses, and a comprehensive similarity score is calculated.

[0082] If the overall similarity score is higher than the preset threshold, the target area is determined to be the steel truss area.

[0083] In this embodiment, for the final candidate regions that have passed the first two levels of screening, a more precise discrimination process is initiated: Feature extraction: Calculate its refined multi-dimensional feature vector, including the principal direction based on PCA (used to determine spatial orientation), axial dimension ratio, and combine it with the linearity and flatness calculated during the intermediate screening to form a complete feature description. Similarity calculation: Match the feature vector of this region with a predefined standard feature template for steel trusses to calculate a comprehensive similarity score.

[0084] Final determination: If the overall similarity score is higher than the preset strict threshold (e.g., ≥95%), then the region is determined to be the target steel truss point cloud, and the extraction operation is performed.

[0085] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for extracting steel trusses based on point cloud segmentation and feature analysis, characterized in that: Includes the following steps: Point cloud preprocessing and index construction stage: Obtain the original point cloud containing the steel truss, perform multi-scale filtering and denoising on the original point cloud, establish topological relationships for the denoised point cloud, and construct the point cloud index structure; Local geometric feature calculation stage: Based on the point cloud index structure, calculate the local geometric features of each point in the point cloud. The local geometric features include at least the normal vector, curvature and first-order plane fitting error. Adaptive seed point selection stage: Traverse all unsegmented points and select the point with the smallest first-order plane fitting error as the initial seed point; The region growth stage based on multi-feature fusion and adaptive threshold is as follows: Centered on the initial seed point, each neighboring point is traversed, and it is determined whether it meets the similarity criterion and edge criterion. Neighboring points that simultaneously meet both the similarity criterion and edge criterion are included in the current growth region. Points with curvature less than the queue curvature threshold are added to the seed point queue as new growth points. After the current seed point has finished growing, the next seed point is taken from the seed point queue, and the seed point growth process is repeated until the seed point queue is empty, completing the growth of one region. The similarity criterion is configured according to the geometric characteristics of each component of the steel truss. Iterative segmentation stage: After the growth of a region is completed, it is marked as an independent segmentation region. Then, the adaptive seed point selection stage and the region growth stage based on multi-feature fusion and adaptive threshold are repeated in the remaining unsegmented point cloud until all points are assigned to a certain region or are judged as noise points, thus obtaining multiple candidate segmentation regions. The segmentation result screening stage based on multi-level cascaded classifier: The candidate segmentation regions are screened in a cascaded manner using a multi-level cascaded classifier, and the final output is the target steel truss point cloud that is determined to be the steel truss region. The cascaded screening is based on a computational cost-based hierarchical strategy. It first performs a primary screening with low computational cost to exclude obvious non-target areas, and then performs intermediate screening and / or final screening with higher computational cost to accurately identify the areas after the initial screening. The screening criteria of the cascaded classifier are set based on prior knowledge of the geometric features of the steel truss structure, including the linear features, planar features and size range of each component of the standard steel truss.

2. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 1, characterized in that: The point cloud preprocessing and index building stage also includes the following steps: A multi-scale filtering method based on octrees is used for denoising. An initial filtering radius is set to perform preliminary denoising on the original point cloud. Then, the filtering radius is gradually reduced for iterative filtering until outliers are removed while the main point cloud features are preserved. Finally, a method combining 3D raster and octree is used to establish topological relationships for the denoised point cloud and construct a point cloud index structure.

3. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 1, characterized in that: The normal vector is calculated by fitting a local plane using principal component analysis (PCA); the curvature is obtained by fitting a local quadratic surface and calculating the Gaussian curvature and mean curvature of the points.

4. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 1, characterized in that: The following steps are used to determine if a target neighbor point meets the similarity criterion: when the angle between the target neighbor point and the normal vector of the current seed point / current region average normal vector is less than the angle threshold, and the absolute value of the average curvature difference between the target neighbor point and the current region is less than the curvature threshold, then the target neighbor point is determined to meet the similarity criterion.

5. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 1, characterized in that: The following steps are used to determine whether the target neighborhood points meet the edge criteria: Calculate the first-order plane fitting error Re_k of the target neighborhood points; The adaptive threshold Re_th is calculated as mean(Re_k) + σ*std(Re_k), where mean(Re_k) represents the average value of the first-order plane fitting error Re_k, std(Re_k) represents the standard deviation of the first-order plane fitting error Re_k, and σ is an empirical coefficient. If the first-order plane fitting error Re_k < the adaptive threshold Re_th, then the target neighborhood point is determined as an internal point satisfying the edge criterion; otherwise, it is an edge point.

6. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 5, characterized in that: The empirical coefficient σ ranges from 0.5 to 1.

5.

7. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 1, characterized in that: The cascaded screening includes primary screening based on geometric dimensions, intermediate screening based on shape features, and final screening based on multidimensional feature matching, with progressively increasing computational costs.

8. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 7, characterized in that: The aforementioned geometry-based primary screening includes the following steps: Calculate the length, width, height, and volume of the basic bounding box for each candidate segmentation region, and compare it with the preset steel truss component size threshold range, eliminating regions that do not match the size.

9. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 7, characterized in that: The intermediate-level screening based on shape features includes the following steps: For the candidate segmentation regions that pass the initial screening, the shape features describing their overall morphology are calculated. The shape features include at least linearity, flatness, and scattering. Based on the preset steel truss component shape threshold, regions with inconsistent shapes are eliminated to obtain the final candidate regions.

10. The steel truss extraction method based on point cloud segmentation and feature analysis according to claim 7, characterized in that: The final screening based on multidimensional feature matching includes the following steps: Extract the multidimensional feature vectors of the final candidate regions that have passed the intermediate screening, including the principal direction and axial dimension ratio based on PCA, and combine them with the linearity and flatness calculated during the intermediate screening to form a complete feature description; The multidimensional feature vector is matched with a predefined standard feature template for steel trusses, and a comprehensive similarity score is calculated. If the overall similarity score is higher than the preset threshold, the target area is determined to be the steel truss area.

Images