Method for extracting and checking precision of three-dimensional point cloud feature data of cargo hold of container ship

Abstract

The application discloses a kind of container ship cargo hold three-dimensional point cloud feature data extraction and precision inspection method, it includes the following steps: (1) point cloud preprocessing;(2) the pre-segmentation of point cloud;(3) feature data segmentation and extraction;(4) the precision inspection of feature data: after completing feature data extraction, using RANSAC noise reduction, least square method detects guide rail flatness, and statistics envelope surface point proportion;Utilize the transverse and longitudinal distance of hausdorff distance to test guide rail spacing precision.The application uses statistical outlier rejection algorithm and RANSAC algorithm for preprocessing, significantly reduce the data volume while maintaining feature accuracy, and effectively segment and extract guide rail and bottom cone and other key features using DBSCAN and region growing algorithm, and the segmentation accuracy is more than 90%.In addition, the application innovatively proposes a three-dimensional point cloud flatness detection method and a method for calculating the transverse and longitudinal spacing based on hausdorff distance, which meets the container ship construction precision inspection standard.

Description

Technical Field

[0001] This invention relates to the field of simulation test chamber technology, specifically to a method for extracting and verifying the accuracy of three-dimensional point cloud feature data of container ship cargo holds. Background Technology

[0002] Traditional physical container testing: In the construction of container ships, the accuracy of container space dimensions directly affects the efficiency and safety of container loading and unloading during ship operation. Therefore, inspecting the container space dimensions before the container ship is put into operation is a crucial construction precision control process. Traditional container ship testing operations are usually conducted in a dry dock or pier. The testing relies on physical containers, using cranes at the dry dock or pier to lift standard containers one by one to each space to verify the installation errors of each space. This ensures that the container can smoothly slide down the guide rails to the hold during the testing process, and that the contact gap between the container base and the hold's adjusting pads and cones meets the precise requirements of the installation process. Traditional container testing operations have the following shortcomings: (1) Long trial period: Traditional trial operations require a lot of dock and wharf resources, resulting in a long trial period. Taking a 9400TEU container ship as an example, during the trial, the shipyard needs to use containers to inspect the positions of each container on the ship, and this process alone takes more than 30 days.

[0003] (2) Numerous safety hazards: The container trial process involves operations such as hoisting containers and personnel entering the cargo hold, which pose certain safety risks. At the same time, the container trial operation is carried out during the ship's final assembly stage, which is a cross-functional and three-dimensional operation. Other work cannot be completely stopped while the container trial is being conducted, and the container trial also requires the cooperation of a corresponding crane system, which is not conducive to the overall safety management of the shipyard.

[0004] (3) High cost of container testing: Container testing requires specialized container frames, hoisting equipment, professional operators and in-cabin testing personnel, resulting in high costs, which is not conducive to the shipyard's cost reduction and efficiency improvement.

[0005] Digital test chamber testing: Digital test chamber testing technology based on total station has been applied in the construction of container ships. By attaching reflective sheets to various points on the guide rails from the bottom to the top of the hold, the three-dimensional coordinates of each point on each guide rail are obtained. Based on these three-dimensional coordinates, data such as the horizontal and vertical spacing of the guide rails, the gap between the guide rails and the standard container walls, and the flatness of the guide rails are calculated. Digital test chamber testing has the following shortcomings: (1) The reflective sheet required for the digital test chamber was pasted during the assembly stage, ignoring the deformation of the guide rail during the welding process, resulting in a large deviation.

[0006] (2) The limited number of reflectors results in insufficient data base for the digital test chamber, which cannot fully reflect the accuracy of the chamber.

[0007] Point cloud recognition: Common point cloud recognition methods are based on deep learning. The constructed deep learning network requires a dataset as data samples for fitting to obtain a network structure with representational capabilities. The dataset required for training is often very large. However, the simulated test chamber technology based on 3D point clouds is still immature, and there are too few samples of cabin feature data for point cloud classification and recognition, resulting in insufficient datasets to support the training process of deep learning algorithms.

[0008] Existing patents related to physical test chamber technology suffer from problems such as long processing time, high safety risks, and high test chamber costs; existing patents related to digital test chamber technology suffer from problems such as large data deviations, insufficient data volume, and high degree of human intervention; existing patents related to point cloud recognition technology are limited by the scale of the dataset and cannot be effectively applied to the field of container ship simulated test chambers. Summary of the Invention

[0009] To address the aforementioned issues, this invention proposes a complete process for simulated container testing based on three-dimensional point cloud data of container ship cabins, and proposes a standard for verifying the accuracy of container ship construction using three-dimensional point cloud data.

[0010] First, this invention discloses a method for extracting and verifying the accuracy of three-dimensional point cloud feature data of container ship cargo holds, which includes the following steps: 1. Point cloud preprocessing steps; including: After acquiring the point cloud, a statistical outlier removal algorithm is used to remove noise based on the Gaussian distribution assumption and processed according to the "3 sigma" principle. After noise reduction, a voxelization downsampling method is used to perform 10% downsampling, replacing the points inside each voxel with the centroid, and optimizing the computational efficiency while preserving feature accuracy through hash function mapping.

[0011] 2. Point cloud pre-segmentation steps; including: The container is segmented based on its location features, and the point cloud of the hull and bulkhead is removed using the Random Sample Consensus Algorithm (RANSAC algorithm). Combined with the included angle constraint, the bottom cone and the container feet are accurately separated, while retaining key features.

[0012] 3. The steps for segmenting and extracting feature data include: The density-based noisy spatial clustering algorithm (hereinafter referred to as DBSCAN algorithm) and the region growing algorithm are used to perform multiple fine segmentations on the pre-segmented point cloud. Based on the DBSCAN algorithm, the region growing algorithm and the root mean square error of registration, the guide rail, the bottom cone and the planar feature point cloud are identified, and the non-feature parts are removed.

[0013] 4. The steps for verifying the accuracy of feature data include: After feature data extraction, RANSAC noise reduction and least squares method were used to detect the flatness of the guide rail, and the proportion of points in the envelope plane was statistically analyzed. Hausdorff distance was used to calculate the horizontal and vertical spacing to verify the accuracy of the guide rail spacing.

[0014] Based on the technical solution described above, a further preferred embodiment involves using a statistical outlier removal algorithm to remove noise and artifacts after acquiring the scanned point cloud. This algorithm assumes that the distances between each point in the sample and each point in its k-neighborhood follow a Gaussian distribution, and removes outliers whose difference from the sample mean exceeds three times the variance, according to the "3 sigma" principle. After point cloud denoising, a voxelization downsampling algorithm is used to achieve point cloud downsampling. This invention establishes a three-dimensional voxel grid based on the input point cloud data, using the centroids of all points contained in the voxel grid to replace other points.

[0015] Based on the technical solution described above, a further preferred approach is to use a hash function to map the original point cloud to a voxel grid with the same number of voxels as the downsampling target. In this example, setting the downsampling target to 10% of the original point cloud helps improve computational efficiency and reduce the computer's computational load while ensuring the accuracy of the feature data. The appropriate value for this parameter may vary depending on the application.

[0016] According to the technical solution described above, a further preferred embodiment of the pre-segmentation includes two parts: first, based on the characteristic that the container positions have the same geometric dimensions, the original point cloud is segmented into several subsets according to the container positions; then, the RANSAC algorithm is used to fit a three-point plane to remove the point clouds of the hull and bulkhead, and by controlling the distance threshold of points in the RANSAC algorithm plane, the bottom cone and the container foot plate are separated while deleting the ground point cloud, combined with the z-axis coordinate. The bottom cone, due to its fixed geometric shape and dimensions, plays an important role in the feature data extraction described later.

[0017] Based on the technical solution described above, a further preferred embodiment adds a constraint condition to the RANSAC algorithm, which specifies the angle θ between the fitting plane normal and the z-axis. In this specific embodiment, the angle θ is set to 15°. The appropriate value of this parameter may vary depending on the application. This constraint condition ensures that the plane accurately fits the point clouds of the hull and bulkhead, avoiding the loss of feature data.

[0018] According to the technical solution described above, a further preferred embodiment is to extract and identify the cabin feature point cloud data after completing the point cloud pre-segmentation; the present invention uses a multi-segmentation method that integrates a noisy DBSCAN algorithm and a region growing algorithm to extract cabin feature data.

[0019] According to the technical solution described above, a further preferred method is to perform initial segmentation based on the DBSCAN algorithm. The DBSCAN algorithm determines the final number of clusters through point cloud density information. Since the DBSCAN algorithm cannot segment closely connected guide rail point clouds and guide rail frame point clouds, the generated clusters are sorted according to the number of point clouds, and the cluster with the highest number of point clouds is selected for further segmentation.

[0020] Based on the technical solution described above, and further preferably, for the part of the point cloud that the DBSCAN algorithm cannot segment, this invention uses a region growing algorithm to perform secondary segmentation on the guide rail connected to the tying bridge. First, the curvature of each point is calculated. Points along sharp edges are removed by first using a curvature threshold. Then, the point with the smallest curvature is selected as the "seed point" by sorting by curvature in descending order. A neighborhood search is performed starting from the "seed point". When the difference between the normal vector of the points in its neighborhood and the normal vector of the seed point is less than a set threshold, it is included in the growth region. Clustering is formed through neighborhood growth.

[0021] According to the technical solution described above, in a further preferred embodiment, after extracting all the guide rails, the same region growing algorithm as the above-mentioned guide rail and tying bridge segmentation is used to extract the four planes in contact with the container; in order to obtain the complete guide rail plane under the premise of high-precision segmentation, this invention introduces a clustering point cloud point number threshold. By controlling the upper and lower limits of the point cloud point number of the clustering result, the segmentation accuracy is guaranteed and the number of segmented surfaces is reduced.

[0022] Based on the technical solution described above, a further preferred method is to sort all the extracted clusters of multiple guide rail planes, bottom cones, and sub-blocks in descending order by calculating the root mean square error of the distance between corresponding points after registration with the standard bottom cone point cloud. The root mean square error of the registration between the bottom cone point cloud in the cabin and the standard bottom cone point cloud is much lower than that of other clusters. Based on this, all bottom cone point clouds in the cabin can be identified. Then, through the clear positional relationship between each component, the feature point clouds of each guide rail plane and sub-block are identified, and irrelevant point clouds are eliminated.

[0023] Based on the technical solution described above, a further preferred embodiment involves performing accuracy verification after extracting and identifying the cabin point cloud feature data. This invention employs a guide rail flatness detection method and a guide rail longitudinal and transverse spacing calculation method suitable for 3D point clouds. First, flatness detection is performed: the input guide rail point cloud is denoised using the RANSAC algorithm, a reference plane is calculated using the least squares method, and a histogram of the distance distribution from each point to the reference plane is generated. Two envelope planes parallel to the reference plane are generated at a distance of ±2mm from the reference plane. Point clouds outside the two envelope planes are marked, and the percentage of points within the envelope planes is calculated. The guide rail flatness is calculated segment by segment in 1m units, with a 0.5m overlap between segments to ensure that this flatness detection method can adequately describe the condition of the complete guide rail. Second, guide rail longitudinal and transverse spacing is detected: this invention calculates the guide rail longitudinal and transverse spacing based on Hausdorff distance. The Hausdorff distance between two clusters of point clouds is calculated, and the larger of the two values ​​is taken as the guide rail longitudinal and transverse spacing. The accuracy of the guide rail spacing is judged based on the detection results, and the accuracy verification conclusion for the container ship is obtained.

[0024] Compared with the prior art, the present invention has the following beneficial effects: (1) A complete simulation test chamber process based on 3D point cloud is proposed, which significantly reduces the test chamber cycle and cost, reduces safety risks, and the calculation process does not require manual intervention. Among them, the guide rail segmentation algorithm has an accuracy of over 93.5%.

[0025] (2) A method for inspecting the accuracy of container ship cargo holds based on three-dimensional point cloud is proposed. The calculated result is less than 2 mm from the actual measured value, which can fully reflect the accuracy of the hold and incorporate welding deformation into the calculation.

[0026] (3) In order to solve the problem that there are too few cabin feature data samples for point cloud classification and recognition, and there is not enough dataset to support the training process of deep learning algorithm, this invention proposes a method to identify the bottom cone point cloud by calculating the root mean square error of the distance between corresponding points after each cluster is registered with the standard bottom cone point cloud, and then identify the feature point cloud of each guide plane and sub-block through the clear positional relationship between each component, and remove irrelevant point clouds. Attached Figure Description

[0027] Figure 1 The images show the on-site scanning operation: (a) a 3D laser scanner and (b) the cabin operation site.

[0028] Figure 2 The data are box point cloud data: (a) 3D point cloud, (b) front view, and (c) top view.

[0029] Figure 3 The bottom cone and the base plate are segmented using a distance threshold combined with the z-coordinate, as shown in (a) top view and (b) side view. Arrow 1 represents the bottom cone, arrow 2 represents the box foot plate, and arrow 3 represents the bottom of the hold.

[0030] Figure 4 The images show the pre-segmentation results of the original point cloud based on the RANSAC algorithm: (a) the original point cloud, (b) the point cloud after segmentation of the hull, and (c) the point cloud after segmentation of the bulkhead.

[0031] Figure 5 This is a schematic diagram illustrating the principle of the DBSCAN algorithm.

[0032] Figure 6 The image shows the initial segmentation effect of feature data based on the DBSCAN algorithm: (a) DBSCAN clustering of box point clouds, (b) clustering after noise removal.

[0033] Figure 7 This is a schematic diagram of the connection between the guide rail and the binding bridge.

[0034] Figure 8 The images show the results of secondary segmentation of feature data based on the region growing algorithm: (a) the selected point cloud to be segmented, and (b) the clustering of the segmented point cloud. Arrow 1 represents the guide rail, and arrow 2 represents the tying bridge.

[0035] Figure 9 The images show a comparison of the guide rail segmentation results: (a) the original guide rail point cloud, (b) the segmented point cloud, and (c) over-segmentation.

[0036] Figure 10 The diagram shows the four-plane segmentation of a guide rail based on a region growing algorithm. (a) shows the guide rail to be segmented, and (b) shows the result of the four-plane segmentation. Arrow 1 represents the original guide rail, and arrow 2 represents the guide rail plane.

[0037] Figure 11 This is a diagram of a container ship's hold bottom loading system. It includes: 1. Guide rails, 2. Bottom cones, and 3. Container dividers.

[0038] Figure 12 The precise extraction effect of compartment feature data: (a) Overall, (b) Part 1, (c) Part 2, (d) Part 3, (e) Part 4. Among them, arrow 1 represents the bottom cone, arrow 2 represents the guide rail, and arrow 3 represents the compartment block.

[0039] Figure 13 This is a flowchart of the guide rail flatness detection process based on 3D point cloud.

[0040] Figure 14 The results of the guide rail flatness test are shown in the following diagrams: (a) Histogram of the distance from the point cloud to the reference plane, and (b) Marked points outside the envelope.

[0041] Figure 15 This is a line graph showing the results of the guide rail flatness calculation.

[0042] Figure 16A line graph showing the comparison of the calculated longitudinal spacing of the guide rails. Detailed Implementation

[0043] The embodiments of the present invention are described in full below, providing detailed implementation methods and specific calculation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0044] 1. Point cloud data acquisition and preprocessing 1.1 Point Cloud Data Acquisition The on-site scanning operation diagram of hold 7F of a 16,000 TEU container ship is as follows: Figure 1 As shown. The original scan point cloud of this section contains approximately 2.01 × 10⁻⁶ points. 9 Point cloud data is collected using a laser 3D scanner and stitched together using accompanying software.

[0045] 1.2 Point Cloud Data Acquisition and Preprocessing The original point cloud dataset contains over 2 billion points. Directly computing such a massive dataset would consume a large amount of memory and storage space, reducing computational efficiency. Therefore, preprocessing of the original point cloud data is necessary. First, a statistical outlier removal algorithm is used to remove noise and artifacts. Given a sample... Calculate the distance between each point in the sample and each point in its k-neighborhood. ,in ; .assumed The terms in the equation follow a Gaussian distribution. :: (1) (2) According to the "3 Sigma" principle, for or Points that are outliers can be defined as outliers and removed from the sample.

[0046] Secondly, voxel grid downsampling is used to reduce point cloud data while preserving the shape characteristics of the point cloud. The voxel grid downsampling algorithm constructs a 3D voxel grid from the input point cloud data, using the centroids of all points contained within a voxel to replace other points. Given a 3D point set... and voxel grid size r ,but x, y, z The dimension of the voxel raster is: (3) in , ,

[0047] Voxelization index at each point h for: (4) (5) in This is for rounding down.

[0048] Voxelization index h Sorting is performed so that points with the same voxelization index fall within the same voxel raster. These points are then replaced with the centroids of all points within the voxel, thus completing the voxelization downsampling. For cabin point cloud data, a large number of voxels have a point count of 0. To reduce computational complexity, a hash function is used to map the original point cloud into a container containing the target number of points for downsampling. A hash function defined as follows can be used:

[0049] Where % represents the modulo function. containersize The number of target containers.

[0050] Finally, taking advantage of the fact that each bin has the same geometric dimensions, the original point cloud is divided into several subsets according to the bins. Figure 2 The image shows one of the box point cloud locations, which contains a total of 8,978,507 points.

[0051] 1.3 Pre-segmentation of point clouds in bulkheads and bilges The hull and bulkhead planar point clouds account for over 70% of the box point cloud. Pre-segmenting and removing these point clouds can significantly reduce the computational load. For such large-scale planar point clouds, using the RANSAC algorithm for three-point plane fitting is an effective segmentation method. Adding the angle between the plane normal and the z-axis to the RANSAC algorithm... The constraints ensure accurate planar fitting of the hull and bulkhead point clouds, avoiding the loss of feature data. By controlling the distance threshold of points within the plane in the RANSAC algorithm, the bottom cone and box foot plate can be separated by combining the z-axis coordinates while deleting the hull point cloud, such as... Figure 3 As shown. The effect of segmenting the point cloud of the bilge and bulkhead based on the RANSAC algorithm is as follows. Figure 4 As shown, the original point cloud has 8,978,507 points, and the point cloud after pre-segmentation by the hull and bulkhead has 1,938,202 points. While preserving the accuracy of feature data, the number of point clouds is reduced by 78.4%, proving that pre-segmentation of container point clouds effectively reduces the computational load and improves computational efficiency.

[0052] 2. Segmentation and Extraction of Feature Data The pre-segmented container point clouds can be divided into two categories based on the types of components they contain: point clouds of bilge containers that simultaneously include bottom cones, sub-blocks, and guide rails, and point clouds of other containers that only contain guide rails. This paper proposes an algorithm based on the DBSCAN algorithm and the region growing algorithm to extract guide rail and bottom cone point clouds through multiple segmentations. Data verification and accuracy testing are performed using the point clouds of four bilge containers (No. 12, 10, 8, and 6) in hold 7F of a 16,000 TEU container ship as examples.

[0053] 2.1 Initial Segmentation of Feature Data Based on DBSCAN Algorithm The DBSCAN algorithm (Density-Based Spatial Clustering of Applications with Noise) is a typical density-based clustering algorithm. This type of density clustering algorithm generally assumes that clusters can be determined by the density of sample distributions. The core idea of ​​the DBSCAN algorithm is that two clusters of high-density point clouds will be separated by low-density points. The DBSCAN algorithm is based on a set of "neighborhood" parameters (…). r , MinPts To characterize the density of samples, given a sample Define the following concepts as follows: Figure 5 As shown: r- Neighborhood: ,That r - Neighborhood sample D and The distance is no greater than r The sample, i.e. ; Core Point: If of r- The neighborhood contains at least MinPts One sample, i.e. ,but It is a core point; Border point: if of r- The number of neighborhood samples is less than MinPts However, if there is a core point, then It is a boundary point; Noise: If of r- The number of neighborhood samples is less than MinPts If there is no core point, then It's noise.

[0054] Starting from any core point in the sample, clusters are generated from adjacent core points and boundary points. The calculation terminates when all core points have been visited. The DBSCAN algorithm has the advantages of not requiring specific cluster shapes, automatically determining the number of clusters, and identifying and removing noise in the scene, making it suitable for segmenting and extracting complex cabin point clouds. The box point cloud processed by the DBSCAN algorithm is shown below. Figure 6 As shown, after noise removal, the box point cloud is automatically segmented into 24 clusters. This algorithm has a good clustering effect on separated guide rails and bottom cones, but it cannot segment guide rails connected to the tie-down bridge. It is necessary to filter the clustering results by the number of points in the point cloud and extract the cluster with the highest number of points for further segmentation.

[0055] 2.2 Secondary Segmentation of Feature Data Based on Region Growing Algorithm Density-based clustering cannot segment two closely connected clusters of point clouds. To segment the guide rails connected to the ties bridge and perform four-plane segmentation on each guide rail, a normal-based region growing algorithm is used for secondary segmentation.

[0056] 2.2.1 Division of guide rails connected to the tying bridge The principle of the region growing algorithm is to start with a "seed point" in the sample and search its neighborhood. When the points in its neighborhood meet certain conditions, clusters are formed through neighborhood growth. The segmentation quality of the region growing algorithm is highly dependent on the classification conditions and the selection of the "seed point," which needs to be determined according to the target to be segmented. Starting from the characteristics of the connection points of the point cloud to be segmented, it can be found that the guide rail and the binding bridge connection plate are perpendicular to each other, such as... Figure 7 As shown, the included angle of the normal is therefore chosen as the classification criterion. The selection of seed points is another factor affecting segmentation quality. To segment the plane to the maximum extent, the curvature of each point needs to be calculated. First, points along sharp edges are removed using a curvature threshold, and then the points with the minimum curvature are selected as "seed points" by descending curvature order. This method of selecting "seed points" ensures that the algorithm grows from the flattest area, minimizing the number of segments in the segmented plane and obtaining a complete plane.

[0057] Using a curvature threshold of 0.5 to remove edge points and a normal angle less than 15° as a classification criterion, the results of segmenting the guide rail and the ties bridge using a normal-based region growing algorithm are as follows: Figure 8 As shown in Table 1, to accurately measure segmentation accuracy, the ratio of the number of guide rail point cloud points segmented by the region growing algorithm to the number of guide rail point cloud points segmented manually using Cloud Compare software is used as the segmentation accuracy. The numbers are as follows: Figure 8 (b) The guide rails are arranged from left to right. The segmentation accuracy of the three guide rails all exceeds 90%, proving that the region growing algorithm can accurately segment the guide rails and the tying bridge, and the number of segments is also well controlled.

[0058] Table 1 Algorithm Accuracy Table

[0059] 2.2.2 Four-plane division of the guide rail After obtaining the point cloud of all guide rails within the designated box location, the guide rails need to be segmented into four planes. The four planes of the guide rail are mutually perpendicular, with circular transitions at the connections. While the normal-based region growing algorithm is still applicable, the segmentation process requires a trade-off between under-segmentation and over-segmentation. To prevent local under-segmentation, more aggressive segmentation parameters are needed to obtain high-precision segmentation results; however, aggressive segmentation parameters can lead to over-segmentation, making it impossible to obtain the complete guide rail plane.

[0060] To obtain the complete guide rail plane while achieving high-precision segmentation, this paper introduces a point count threshold for clustering point clouds. By controlling the upper and lower limits of the point cloud count in the clustering results, segmentation accuracy is ensured while reducing the number of segmented faces. A comparison of segmentation effects with and without setting the point count threshold is shown in the figure. Figure 9 As shown, the region growing algorithm with a set point threshold results in 6 clusters, while the conventional region growing segmentation algorithm results in 24 clusters. This demonstrates that setting a point threshold effectively controls the number of clusters and ensures the integrity of the guide rail plane. The four-plane segmentation effect of the guide rail is shown below. Figure 10 As shown.

[0061] 2.3 Feature Data Identification of Guide Rails and Base Cones After segmenting the feature point clouds of the bottom cone, guide rails, and compartment blocks, it is necessary to identify each feature point cloud and remove irrelevant points. A common method is point cloud classification and recognition based on deep learning. However, due to the limited sample size of the compartment feature data, there is insufficient dataset to support the training process of deep learning algorithms, necessitating the search for alternative methods.

[0062] Among the point clouds of guide rails, bottom cones, and container blocks to be identified, the bottom cone point cloud has a unique geometric shape, uniform specifications, and small volume, making it a good identification target. Furthermore, the container ship hold's bottom loading system has clear installation benchmarks, and the positions of each component are clearly defined, such as... Figure 11 As shown. Therefore, a point cloud registration algorithm can be used first to sort the point clouds of the base cone by the registration accuracy and overlap of each component with the standard base cone point cloud, and then the remaining components can be identified by their positional relationships.

[0063] The Iterative Closest Point (ICP) algorithm is a commonly used point cloud registration algorithm. The core of this algorithm lies in calculating the rotation matrix that makes the point cloud to be matched coincide with the target point cloud. R Translation vector t The calculation process of the rotation matrix and translation vector can be transformed into... R, t Optimization problem: Given a set of points to be matched and target point set ,have: (7) Solution process: ① Find the point set A and B Center of mass: (8) ②Pair set A and B Centralization: (9) ③ Perform singular value decomposition: (10) in, U and V It is an orthogonal matrix. Σ It is a diagonal matrix.

[0064] Solved rotation matrix R Translation vector t for: (11) The rotation matrix obtained by solving R Translation vector t Calculate the set of points to be matched A Repeat the above operation on the transformed coordinates, when Small enough or Δ R Δ t When the value is sufficiently small, the iteration terminates. At this point, the calculated... R, t That is, the rotation matrix and translation vector that make the point cloud to be matched and the target point cloud coincide.

[0065] To measure the registration accuracy, the root mean square error (RMSE) of the distance between corresponding points in the two registered point clouds is calculated: (12) in, m This represents the number of corresponding point pairs. This is the Euclidean distance between corresponding points after registration. The root mean square error (RMSE) represents the true distance between corresponding points after registration. Ideally, the distance between corresponding points after registration should be 0. A smaller RMSE indicates higher point cloud registration accuracy. The RMSE obtained by registering the point cloud clusters of each hull in Section 2.1 with the standard hull cone point cloud is shown in Table 2.

[0066] Table 2 shows the root mean square error of the registration between each point cloud cluster and the standard base cone point cloud.

[0067] The point cloud clusters are arranged in ascending order of root mean square error; the first 12 point clouds constitute the cabin floor cone point cloud. After eliminating irrelevant point clouds based on positional relationships, the resulting cabin feature point cloud set is as follows: Figure 12 As shown.

[0068] 3. Accuracy verification of cabin characteristic data for simulated test chambers After obtaining the cabin feature data, it is necessary to complete the simulation test accuracy inspection based on the feature data. This paper proposes a simulation test accuracy inspection method based on three-dimensional point cloud data. The acceptance accuracy index of a certain 16,000 TEU container ship is shown in Table 3.

[0069]

[0070] 3.1 Inspection of guide rail flatness Table 3. Acceptance Accuracy Indicators for Cargo Holds of Container Ships (Unit: mm) Currently, the guide rail flatness inspection in container ship acceptance relies on a total station. The specific procedure involves marking the ends of the guide rail connecting plates on the lashing bridge from top to bottom before installation, typically a dozen or so marks. The coordinates of each mark point along the ship's length are measured using a total station, and the guide rail flatness is determined by comparing the coordinate values. This inspection method ignores the deformation of the guide rail after welding, resulting in significant errors. To accurately inspect guide rail flatness, this invention proposes a flatness inspection method based on three-dimensional point cloud data. The inspection process is as follows: Figure 13 As shown, the envelope plane generated by the ±2mm spacing is calculated based on the construction accuracy standard of ≤4mm coplanarity of the container ship guide surface.

[0071] Setting a distance threshold using the RANSAC algorithm can effectively control noise on the guide rail point cloud surface, but noise cannot be completely removed. Paint buildup on the guide rail surface also interferes with flatness detection results. Therefore, the percentage of points within the envelope plane is used to measure the guide rail flatness. The flatness detection results are as follows: Figure 14 As shown, the green solid line marks the envelope plane, and the red solid line marks the point farthest from the reference plane. The distances of each point in the guide rail point cloud to the reference plane approximately follow a Gaussian distribution. If the proportion of interior points exceeds 95%, then under a confidence level of 0.95, the flatness of the guide rail can be considered to meet the construction accuracy requirements.

[0072] Table 4. Calculation results of guide rail flatness (unit: mm)

[0073] Table 4 above shows the flatness calculation results for each guide rail from 2400mm to 7000mm. The calculation process is as follows: Since the height of a complete guide rail from the bottom to the top of the cabin exceeds 20m, it is obviously unreasonable to measure the flatness of the guide rail using the complete guide rail. This invention calculates the flatness of the guide rail segment by segment in 1m units, with a 0.5m overlap between two segments to ensure that this flatness detection method can better describe the condition of the complete guide rail. The flatness calculation results for each guide rail from 2400mm to 7000mm are shown in Table 4. The calculation results are plotted as a line graph as shown below. Figure 15 As shown, of the 24 guide rail surfaces tested, 19 met the construction accuracy requirements for flatness.

[0074] 3.2 Inspection of the longitudinal and transverse spacing of the guide rails For a pair of guide rail point clouds, these two point cloud clusters are not parallel and corresponding, therefore the distance between the two guide rails cannot be directly measured. This paper proposes a method for calculating the horizontal and vertical spacing of the guide rails based on Hausdorff distance. Hausdorff distance measures the distance between two sets in space, given two point cloud clusters in space. A and B From the set A To set B The Hausdorff distance is a maximal-minimal function, defined as: (13) Similarly, the Hausdorff distance from set B to set A is: (14) in, For any measure between points a and b, in this paper, simply use Defined as the Euclidean distance between two points a and b. The Hausdorff distance is asymmetric, meaning that in most cases... Therefore, two clusters are defined. A and B The Hausdorff distance between them is: (15) Factors such as unavoidable noise on the guide rail point cloud surface and paint film buildup on the guide rail surface can lead to... Since the distance is smaller than the actual distance, the RANSAC algorithm needs to be used to filter noise from the two clusters of point clouds first.

[0075] To verify the accuracy of the Hausdorff distance in describing the guide rail spacing, the longitudinal spacing of the six guide rails extracted earlier was calculated, and the results are shown in Table 5. Table 5 Comparison of Calculation Results for Longitudinal Spacing of Guide Rails (Unit: mm)

[0076] The calculated data is plotted as a line graph, as shown below. Figure 16 The error between the guide rail spacing obtained by Hausdorf distance and the actual guide rail spacing is within 2mm, and the trend of change is consistent with the actual measurement data, proving that Hausdorf distance can accurately describe the guide rail spacing.

[0077] The embodiments of this application demonstrate that: 1. Improved computational efficiency: By using statistical outlier removal and the RANSAC algorithm, the amount of point cloud data was effectively reduced by 78.4%, the preprocessing process of cabin feature data was optimized, and the subsequent computation speed was improved.

[0078] 2. Improved point cloud segmentation accuracy: The proposed multi-segmentation algorithm, combined with DBSCAN and region growing algorithms, achieves an accuracy of over 90% in the identification of guide rails and bottom cones, significantly enhancing the ability to accurately extract key cabin structures.

[0079] 3. Accuracy inspection method: The three-dimensional point cloud flatness detection technology and the horizontal and vertical spacing calculation method based on Hausdorf distance are incorporated into the deformation of the guide rail after welding. The error compared with the measured data is less than 2mm, which meets the accuracy inspection standards for container ships and reflects the technological innovation and practicality.

[0080] 4. Systematized simulation test chamber technology: A complete simulation test chamber process has been constructed, from data acquisition and preprocessing to accuracy verification, providing powerful tools and technical support for engineering practices in related fields.

[0081] 5. Significantly Enhanced Future Application Potential: With the continuous development of unsupervised semantic segmentation technology in the field of 3D point clouds, this technology is expected to be applied to simulation test chamber technology. Unsupervised algorithms do not require manual supervision signals or pre-trained models, which can simplify the calculation process, greatly improve computational efficiency, and are widely applicable to various 3D point cloud scenarios. This provides strong technical support for the further optimization and expansion of simulation test chamber technology, demonstrating broad application prospects and significant beneficial effects.

[0082] It should be understood that this invention is not limited to the specific embodiments and examples listed above. In fact, the scope of protection of this invention should be defined by its claims and is not limited to the details described in the specific embodiments or examples. Any equivalent transformations, improvements to the technical solutions, or modifications, combinations, and applications of the above embodiments based on the concept of this invention, without departing from the spirit and essence of this invention, shall fall within the scope of protection of this invention.

Claims

1. A method for extracting and verifying the accuracy of 3D point cloud feature data from container ship cargo holds, characterized in that, Includes the following steps: (1) Point cloud preprocessing: After acquiring the point cloud, a statistical outlier removal algorithm is used to remove noise based on the Gaussian distribution assumption and processed according to the "3 sigma" principle. After noise reduction, a voxelization downsampling method is used to perform 10% downsampling, replacing the points within each voxel with the centroid, and optimizing the computational efficiency while preserving feature accuracy through hash function mapping. (2) Pre-segmentation of point clouds: The original point cloud was segmented according to the container location features and the RANSAC algorithm was used to remove the point cloud of the hull and bulkhead. The hull and bulkhead were accurately separated by combining the included angle constraints. The bottom cone and the hull foot plate were separated by combining the ordinate coordinate, while retaining key features. (3) Segmentation and extraction of feature data: Density-based noisy spatial clustering and region growing algorithms are used to perform multiple fine segmentations on the pre-segmented point cloud. Based on the DBSCAN algorithm, region growing algorithm and registration root mean square error, the feature point clouds of guide rail, bottom cone and box block are identified, and non-feature parts are removed. (4) Accuracy verification of feature data: After feature data extraction, RANSAC noise reduction and least squares method were used to detect the flatness of the guide rail, and the proportion of points in the envelope plane was statistically analyzed. Hausdorff distance was used to calculate the horizontal and vertical spacing to verify the accuracy of the guide rail spacing.

2. The method according to claim 1, characterized in that, In step (1), after acquiring the scanned point cloud, a statistical outlier removal algorithm is used to remove noise and artifacts. The statistical outlier removal algorithm assumes that the distance set between each point in the sample and each point in its k-neighborhood follows a Gaussian distribution. According to the "3 sigma" principle, outliers that differ from the sample mean by more than three times the variance are removed. After the point cloud is denoised, a voxelization downsampling algorithm is used to downsample the point cloud. In step (1), a hash function is used to map the original point cloud to a voxel grid with the same number of downsampled targets.

3. The method according to claim 1, characterized in that, In step (2), the pre-segmentation includes two parts. First, based on the characteristic that the container positions have the same geometric size, the original point cloud is divided into several subsets according to the container positions. Then, the RANSAC algorithm is used to fit a three-point plane to remove the point cloud of the hull and the bulkhead. By controlling the distance threshold of the points in the RANSAC algorithm plane, the bottom cone and the container foot plate are separated while deleting the hull point cloud, combined with the z-axis coordinate.

4. The method according to claim 3, characterized in that, In step (2), a constraint condition is added to the RANSAC algorithm to fit the angle θ between the normal of the fitting plane and the z-axis. The setting of this constraint condition ensures that the plane accurately fits the point cloud of the hull and the bulkhead.

5. The method according to claim 1, characterized in that, In step (3), after the point cloud pre-segmentation is completed, the cabin feature point cloud data is extracted and identified; the cabin feature data is extracted by a multi-segmentation method that combines the noisy DBSCAN algorithm and the region growing algorithm.

6. The method according to claim 5, characterized in that, In step (3), the initial segmentation is performed based on the DBSCAN algorithm. The DBSCAN algorithm determines the final number of clusters through point cloud density information. The DBSCAN algorithm cannot segment closely connected guide rail point clouds and guide rail frame point clouds. By sorting the generated clusters according to the number of point clouds, the cluster with the highest number of point clouds is selected and extracted for further segmentation.

7. The method according to claim 6, characterized in that, In step (3), for the part of the point cloud that the DBSCAN algorithm cannot segment, the region growth algorithm is used to perform secondary segmentation on the guide rail connected to the tying bridge. First, the curvature of each point is calculated. Then, the points along the sharp edges are removed by setting a curvature threshold. The point with the smallest curvature is selected as the "seed point" by curvature descending order. The neighborhood search is performed starting from the "seed point". When the difference between the normal vector of the points in its neighborhood and the seed point is less than the set threshold, it is included in the growth region. Clustering is formed through neighborhood growth.

8. The method according to claim 7, characterized in that, In step (3), after all the guide rails are extracted, the same region growing algorithm as the guide rail and tie bridge segmentation above is used to extract the four planes in contact with the container; a clustering point cloud point number threshold is introduced, and the segmentation accuracy is guaranteed and the number of segmented surfaces is reduced by controlling the upper and lower limits of the point cloud point number of the clustering result.

9. The method according to claim 8, characterized in that, In step (3), for all the extracted guide rail planes, bottom cones and sub-blocks, the root mean square error of the distance between corresponding points after registration with the standard bottom cone point cloud is calculated and sorted in descending order. The root mean square error of the registration between the bottom cone point cloud in the cabin and the standard bottom cone point cloud is much lower than that of other clusters. Based on this, all bottom cone point clouds in the cabin are identified. Then, through the clear positional relationship between each component, the feature point clouds of each guide rail plane and sub-block are identified, and irrelevant point clouds are eliminated.

10. The method according to claim 1, characterized in that, In step (4), firstly, flatness detection is performed: the input guide rail point cloud is denoised using the RANSAC algorithm, the least squares method is used to calculate the reference plane and generate a histogram of the distance distribution from each point to the reference plane, two envelope planes parallel to the reference plane are generated at a position ±2mm from the reference plane, the point cloud outside the two envelope planes is marked and the percentage of points inside the envelope plane is calculated; the flatness of the guide rail is calculated segment by segment in units of 1m, and the overlap between the two segments is 0.5m; secondly, the horizontal and vertical spacing of the guide rail is detected: the horizontal and vertical spacing of the guide rail is calculated based on the Hausdorf distance, by calculating the relative Hausdorf distance between the two clusters of point clouds, and taking the larger value of the two as the horizontal and vertical spacing of the guide rail.

Images