Vehicle-mounted steel coil saddle point cloud segmentation method and crown block positioning system based on the method

Abstract

The application is a method for segmenting point cloud of a vehicle-mounted steel coil saddle and a crown positioning system based on the method. The method comprises the following steps: obtaining point cloud data of the vehicle-mounted steel coil saddle through a three-dimensional laser scanning holder, filtering the point cloud data in x, y and z coordinate directions, and segmenting a region of interest; setting an elevation threshold, using the elevation change in the x coordinate direction for clustering, separating the point cloud data of the vehicle-mounted plane and the saddle, fitting the vehicle-mounted plane using the RANSAC algorithm; using the Euclidean clustering to cluster the independently distributed saddle point cloud data in space, using an improved local convex connection growth algorithm to optimize the segmentation effect of the point cloud data for each saddle point cloud cluster, and refining the segmentation surface boundary; and calculating the x, y and z coordinates of the vehicle-mounted steel coil saddle using the segmented saddle support plane point cloud. The method can realize accurate positioning of the crown to the saddle and improve work efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional laser scanning positioning, in particular to a vehicle-mounted steel coil saddle point cloud segmentation method and a crown block positioning system based on the method. BACKGROUND

[0002] The traditional crown block performs steel coil loading operation on the vehicle mainly by relying on the command of ground personnel and the experience of operators. As a heavy equipment, the crown block is prone to cause safety accidents due to operation errors or visual obstacles in the actual production site. By using three-dimensional laser scanning positioning technology, a three-dimensional laser scanning pan is installed at the lower end of the crown block girder. When the vehicle enters the operation area, the pan rotates to scan, and the saddle center coordinates are obtained through segmentation processing and calculation. This fundamentally changes the operation mode of the intelligent crown block and realizes the effect of accurately positioning the vehicle-mounted steel coil saddle center coordinates.

[0003] At present, the segmentation of steel coil saddle point cloud mainly uses relatively traditional segmentation methods. For example, patent one (a vehicle saddle extraction method and system based on three-dimensional laser scanning, CN114399550A) proposes to project the point cloud data along the vertical plane of the vehicle parking direction, pair the sorted projection straight lines two by two, and extract the projection of the saddle. The disadvantage of this method is that the projection straight line matching is difficult, and it is easy to cause matching errors due to data noise errors. Literature one (TAN Yun-yue. Design and implementation of steel coil loading and unloading automatic positioning system based on point cloud data[D]. Chongqing University, 2017.) proposes to use RANSAC algorithm to fit the saddle support plane, and then use the Euclidean clustering space to separate the saddle point cloud. However, the disadvantage of this method is that the RANSAC algorithm has low segmentation completeness and accuracy for the saddle surface, and is prone to over-segmentation, which will affect the positioning results. Literature two (Wei S, Niu D, Li Q, et al. A 3D vehicle recognition system based on point cloud library[C]. 2019Chinese Control Conference (CCC). IEEE, 2019: 7023-7027.) proposes to use the architecture of edge detection and region growing method to segment the point cloud based on edge points, and identify the saddle with volume and normal features. However, the disadvantage of this method is that the accuracy of edge extraction is low, which will cause over-segmentation.

[0004] Through the above analysis, the problems and defects of the prior art are: the traditional crane positioning mainly relies on workers with operating experience, the production efficiency is low and the production cost is high, and the recognition accuracy of some working environments using photogrammetry technology is not as good as that of the three-dimensional laser scanner. When using a three-dimensional laser scanner for identification and positioning, for the segmentation of the saddle point cloud data, the traditional method has low accuracy and is prone to over-segmentation and under-segmentation problems, which directly affects the positioning accuracy of the center coordinates. SUMMARY

[0005] To solve the problems existing in the prior art, the present application provides a vehicle-mounted steel coil saddle point cloud segmentation method and a crane positioning system based on the method.

[0006] The technical solution adopted by the present application to solve the technical problem is:

[0007] A vehicle-mounted steel coil saddle point cloud segmentation method, characterized in that the method comprises the following steps:

[0008] Firstly, the point cloud data of the vehicle-mounted steel coil saddle is obtained by a three-dimensional laser scanning pan-tilt and filtered in x, y, z three coordinate directions to segment out a region of interest (Region Of Interest, hereinafter referred to as ROI); wherein the x, y, z directions are respectively the vehicle driving direction, the direction perpendicular to the vehicle body, and the direction perpendicular to the ground;

[0009] Secondly, a height threshold is set, the point cloud data of the vehicle-mounted plane and the saddle is separated out by using the height change clustering in the x coordinate direction, and the RANSAC algorithm is used to fit the vehicle-mounted plane;

[0010] Thirdly, the independently distributed saddle point cloud data in space is clustered by using the Euclidean clustering, and for each saddle point cloud cluster, the improved locally convex connected patch (Locally Convex Connected Patches, hereinafter referred to as LCCP) algorithm is used to optimize the segmentation effect of the point cloud data and refine the segmentation surface boundary;

[0011] The improved LCCP algorithm comprises constructing a super voxel adjacency graph of the object surface block by using multi-scale super voxel fusion segmentation, then segmenting the super voxel adjacency graph by classifying whether the connection between two super voxels is convex or concave, dividing the same subgraph belonging to the effective convex connection edge by using a region growing algorithm, and finally optimizing the segmentation boundary by boundary marking to obtain the segmented saddle support plane point cloud;

[0012] The process of multi-scale super voxel fusion segmentation is: setting a distance threshold d T and the minimum size of super voxel division, dividing the three-dimensional point cloud into regular super voxels V iFor each super voxel V i , a plane P is fitted using PCA algorithm i , the distance d of each point in the super voxel to the fitted plane is calculated, if the maximum distance d max of the points in the super voxel to the plane is less than the distance threshold d T , the super voxel segmentation ends; otherwise, the super voxel is continuously divided into four sub-super voxels, a new fitted plane is fitted for each sub-super voxel, and it is judged whether the maximum distance d max of the points in the sub-super voxel to the plane is less than the distance threshold d T , if yes, the segmentation ends, if no, the sub-super voxel is re-divided in the above manner until the segmentation ends; if the current sub-super voxel reaches the minimum size of super voxel division but still cannot satisfy that the maximum distance d max is less than the distance threshold d T , all points belonging to the current sub-super voxel are filtered out as noise points;

[0013] The maximum optimization function W * used in the boundary marker optimization is shown in formulas (7)-(9):

[0014]

[0015] Wherein, W is the label of the point on the sub-graph after region growing; ω represents all effective segmentation planes obtained after region growing; a sub-graph after region growing represents a plane, denoted as Q k , k represents the label of the plane, and N is the number of points on all effective segmentation planes; G(W) is a boundary function; represents that the point x i belongs to the plane Q k , and |N j | represents the number of neighborhood points; is a judgment function;

[0016] The fourth step is to calculate the x, y and z coordinates of the vehicle-mounted steel coil saddle by using the segmented saddle support plane point cloud.

[0017] Thus, the segmentation and positioning of the vehicle-mounted steel coil saddle are completed.

[0018] In the first step, the filtering process is: using a pass-through filter to filter the point cloud data in x, y and z coordinate directions respectively, using voxel downsampling and density change filtering, a given point cloud density threshold is set, the average distance of each point to its nearest k points is calculated, which is defined as the point cloud density at the point, if it is less than the given point cloud density threshold, it is determined as a noise point cloud and deleted.

[0019] The process of calculating the x, y and z coordinates of the vehicle-mounted steel coil saddle in the fourth step is:

[0020] For the segmented saddle support plane point cloud, the segmented support plane point cloud is divided into two groups according to the elevation change along the y coordinate direction, the midpoint of the point cloud centroid connecting line of the two groups is calculated, and then the y coordinate of the steel coil saddle center point is obtained; the segmented support plane point cloud is divided into two groups according to the elevation change along the x coordinate direction, the two groups of point cloud data are respectively fitted as straight lines, and the x, z coordinates of the steel coil saddle center point are obtained according to the distance of the center point of the steel coil to the two straight lines = the radius of the steel coil, and then the position coordinates of the steel coil saddle center point are obtained, and positioning is realized.

[0021] The specific process of obtaining the x, z coordinates of the steel coil saddle center point is as follows:

[0022] The segmented support plane point cloud is projected onto the xz plane to obtain two projection straight lines l1 and l2 of the support plane point cloud on the xz plane, the point O is the projection point of the steel coil placement position and is located on the median line s, the RANSAC algorithm is used for straight line fitting on the two groups of point clouds divided according to the elevation change along the x coordinate direction, and the general equation of the two projection straight lines l1 and l2 is obtained as formula (10):

[0023]

[0024] Since the steel coil is tangent to the two support planes, the x, z coordinates of the steel coil saddle center point are obtained by using formula (11):

[0025]

[0026] Formula (11) is the distance formula of the O point to the two straight lines, R is the radius of the steel coil, which is a known value, A1, B1, C1, A2, B2 and C2 are the corresponding coefficients of the general equation of the two straight lines.

[0027] The application also protects a crown block positioning system, which comprises a three-dimensional laser scanning pan-tilt, a host computer, a PLC programmable controller and a crown block; the three-dimensional laser scanning pan-tilt is installed at the lower end of the moving beam of the crown block, the laser scanning pan-tilt control software and the point cloud data processing visual operation interface are installed in the host computer, the host computer communicates with the laser scanning pan-tilt through TCP / IP, and the PLC programmable controller communicates with the host computer and the crown block through the RS485 interface; the vehicle-mounted steel coil saddle point cloud segmentation method is loaded in the host computer.

[0028] Compared with the prior art, the application has the following beneficial effects:

[0029] 1. Compared to traditional manual operation and photogrammetry methods for saddle positioning, this invention uses a 3D laser scanning gimbal to acquire point cloud data of the steel coil saddle on the transport vehicle. An improved Locally Convex Connected Patches (LCCP) algorithm is used to optimize the segmentation effect of the point cloud data, refining the segmentation surface boundaries. Compared to traditional segmentation methods, this reduces under-segmentation and over-segmentation errors, effectively improving the accuracy and efficiency of saddle positioning. This ensures that the overhead crane can accurately perform steel coil loading operations, providing the theoretical basis for automated and unmanned steel coil loading and unloading. 3D laser scanning technology is highly proactive, has low requirements for the scanning environment, and is easy to operate. It is suitable for unmanned overhead crane operations, acquiring point cloud data with 3D coordinate information. The data volume is large and high-density, containing optical feature information of the scanned object, enabling precise positioning of the saddle by the overhead crane and improving work efficiency.

[0030] 2. The steel coil saddle consists of a saddle panel, baffles, and a support base. Due to positioning requirements, the segmentation of the saddle primarily involves accurately segmenting the saddle panel. Concavity and convexity are the most significant characteristic information of the saddle plane, and the transition between concave and convex surface units indicates separation between objects. The improved LCCP algorithm segments the saddle panel based on this important geometric characteristic of point clouds, thus enabling accurate and effective segmentation. Using super-voxel multi-scale fusion and boundary marker optimization reduces under-segmentation and over-segmentation errors, thereby effectively improving the accuracy and efficiency of saddle positioning. Attached Figure Description

[0031] Figure 1 This is a control flowchart of one embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram of the point cloud obtained by a 3D laser scanning gimbal;

[0033] Figure 3 This is a schematic diagram of the point cloud of the working area after preprocessing;

[0034] Figure 4 This is a schematic diagram of the point cloud obtained from vehicle-mounted planar segmentation;

[0035] Figure 5 This is a schematic diagram of the point cloud obtained by segmentation using the improved local convex connection growth algorithm.

[0036] Figure 6 This is a schematic diagram of the linear projection method.

[0037] Figure 7 This is a schematic diagram of the overhead crane positioning system;

[0038] Figure 8 This is the control flowchart of the overhead crane positioning system. Detailed Implementation

[0039] The technical solution of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings, but this is not intended to limit the scope of protection of this application.

[0040] This invention relates to a method for segmenting point clouds of a vehicle-mounted steel coil saddle (hereinafter referred to as the method, see [link]). Figures 1-5 The process includes the following steps:

[0041] The first step is to acquire point cloud data of the vehicle-mounted steel coil saddle using a 3D laser scanning gimbal, such as... Figure 2 As shown; the parking area of ​​the vehicle is set as the Region of Interest (ROI). First, the point cloud is transferred from the scanner coordinate system to the scene coordinate system, with the centroid of the calibration object as the origin, and the vehicle driving direction, the direction perpendicular to the vehicle body, and the direction perpendicular to the ground as the x, y, and z axes, respectively. Pass-through filtering is used to filter the point cloud data in the x, y, and z coordinate directions, respectively, through voxel downsampling and density change-based filtering. The main shape features of the point cloud are preserved while reducing the amount of point cloud data, resulting in the preprocessed point cloud data of the working area, as shown. Figure 3 As shown.

[0042] Because the 3D laser scanning gimbal in this application has a visual blind zone when acquiring the target, the point cloud height within the blind zone varies abruptly, the point density is sparse, and this interferes with point cloud segmentation. A density variation filtering-based noise reduction process is adopted: The principle of density variation filtering for noisy point clouds is as follows: the 3D laser scanner has a visual blind zone when acquiring the target, and the noisy point cloud within the blind zone has the characteristics of height abrupt changes and sparse point density, which interferes with point cloud segmentation. Given a point cloud density threshold, the average distance from each point to its k nearest points is calculated and defined as the point cloud density at that point. If the distance is less than the given point cloud density threshold, it is determined to be a noisy point cloud and deleted.

[0043] The second step involves setting an elevation threshold and clustering based on elevation changes along the x-axis. The elevation changes along the x-axis are significant for the vehicle front and vehicle-mounted planes. Points exceeding the set elevation threshold are discarded, while those below are retained. This removes the point cloud data from the vehicle front. Therefore, the point cloud data for the vehicle-mounted plane and saddle are separated. The RANSAC algorithm is then used to fit the vehicle-mounted plane data. Figure 4 As shown; in this embodiment, it is only necessary to know the elevation change of the point cloud along the x-axis direction, and the point cloud of the vehicle plane and the saddle can be separated by a height difference.

[0044] The third step is to use Euclidean clustering to cluster the independently distributed saddle point cloud data in space. For each saddle point cloud cluster, multi-scale supervoxel fusion segmentation is used to construct the adjacency graph of the object surface block. Then, the supervoxel adjacency graph is segmented by classifying whether the connection between two supervoxels is convex or concave. The region growing algorithm is used to divide the same subgraph belonging to the effective convex connection edge. Finally, the segmentation boundary is optimized by boundary marking to obtain the segmented saddle support plane point cloud.

[0045] 3-1. Constructing a supervoxel adjacency graph of an object's surface blocks using multi-scale supervoxel fusion segmentation:

[0046] Set distance threshold d T The minimum size for dividing the supervoxel is used to divide the 3D point cloud into regular supervoxels V of size S. i ,use To indicate, among which Center of mass, Let N be the normal vector. i Let N be the number of point clouds in a supervoxel, and let e∈N be the edge connecting adjacent supervoxels. i For each supervoxel V i PCA was used to fit the plane P i Calculate the distance d from each point within the supervoxel to the fitted plane. If the maximum distance d from a point within the supervoxel to the plane is... max Less than the distance threshold d T If the segmentation of the supervoxel is successful, the segmentation ends; otherwise, the supervoxel will be divided into four sub-supervoxels. A new fitting plane will be fitted for each sub-supervoxel, and the distance from each point within the sub-supervoxel to the fitting plane will be calculated. The maximum distance d from the point within the sub-supervoxel to the plane will be determined. max If the distance is less than the threshold, segmentation ends; if not, sub-hypervoxes are continued until segmentation ends. If the current sub-hypervoxe reaches its minimum size but segmentation is still unsuccessful (maximum distance d), the process continues. max If the distance threshold is still not less than the current sub-supervoxel, then all points belonging to the current sub-supervoxel will be considered noise points and filtered out.

[0047] The retained supervoxels and all sub-supervoxels are called segmented supervoxels, hereinafter collectively referred to as supervoxels. All segmented supervoxels form a supervoxel adjacency graph of the object surface block. At this time, the segmented supervoxels are of different sizes. Multi-scale supervoxel fusion segmentation can automatically adjust the segmentation under any given supervoxel size, achieve fission, and obtain segmented supervoxels of different sizes. It has a short running time, high efficiency, and can achieve good segmentation results.

[0048] 3-2. Utilizing the edges connecting the segmented supervoxels The concavity-convexity relationship is used to segment the supervoxel adjacency graph. For two adjacent supervoxels, the concavity-convexity relationship of the supervoxels is determined using equation (1):

[0049]

[0050] In equation (1), These are the centroids of adjacent supervoxels. Their normal vectors are respectively. Let α1 and α2 be the vectors connecting the two centroids, and α1 and α2 be the normal vectors, respectively. with vector The included angle;

[0051] To reduce errors caused by noise, concave surfaces with low curvature are treated as convex surfaces, and a concavity tolerance threshold β is introduced. Thresh Establish basic convexity criteria For equation (2):

[0052]

[0053] When two hypervoxels are connected by only one point, their surfaces are discontinuous and considered invalid connections. Equation (3) is used to determine whether the surfaces of two adjacent hypervoxels are invalidly connected:

[0054]

[0055] In equation (3), Let γ be the cross product of the two normal vectors. sum vector The included angle;

[0056] To reduce the error caused by noise, an angle threshold γ is introduced. Thresh :

[0057]

[0058] Equation (5) is derived from equations (3) and (4), and equation (5) is the soundness criterion:

[0059]

[0060] Combining equations (2) and (5), we obtain the final discriminant formula, where conv represents the discriminant formula. We then obtain the intersection of equations (5) and (2):

[0061]

[0062] 3-3. Use region growing to partition the same subgraph belonging to valid convex edges. The same subgraph forms a plane, and all points on the plane have the same label. Obtain all valid partitioned planes through region growing. Select and label any seed supervoxel. This label is propagated on the graph through depth-first search. Depth-first search is only allowed to grow on convex edges. Once no new supervoxel can be assigned to a line segment, select a new unlabeled seed supervoxel and propagate the new label. Repeat this process until all supervoxels have been labeled.

[0063] 3-4. To solve the boundary labeling optimization problem, this application uses the maximization function W. * :

[0064]

[0065] Where W is the label of the points on the subgraph after region growing, ω represents all planes of effective segmentation, and G(W) is the boundary function used to evaluate the regularity of the planes; W xi =k represents point x i Belongs to plane Q k Q k For any subgraph, N j Representing point x i The neighborhood points, |N j | represents the number of neighboring points, and N is the number of points on all effectively partitioned planes. As can be seen from equation (9), the goal of boundary label optimization is to optimize each point x... i By specifying the optimal label k, and in order to maximize the boundary function G(W), we can address the issue that due to the varying sizes of the supervoxels, multiple planes merging together can create jagged edges. The optimization function process reduces the number of sharp points at the boundary, forces a smooth boundary, and eliminates the effects of over-segmented small planes and jagged boundaries. Figure 5 A schematic diagram of the point cloud obtained by segmentation using the improved local convex connection growth algorithm.

[0066] Step 4: Divide the segmented support plane point cloud into two groups according to the elevation change along the y-coordinate direction. Calculate the midpoint of the line connecting the centroids of the two groups to obtain the y-coordinate of the steel coil saddle center point. Divide the segmented support plane point cloud into two groups according to the elevation change along the x-coordinate direction. Fit the two groups of point cloud data to straight lines respectively. Use the surface projection straight line method (based on the distance from the center point of the steel coil to these two straight lines = the radius of the steel coil) to obtain the x and z coordinates of the steel coil saddle center point, thereby obtaining the position coordinates of the steel coil saddle center point and achieving positioning.

[0067] This completes the segmentation and positioning of the vehicle-mounted steel coil saddle.

[0068] The surface projection straight line method projects the segmented support plane point cloud onto the xz plane, obtaining two projection straight lines l1 and l2 of the support plane point cloud onto the xz plane. Point O is the projection point of the steel coil placement position and lies on the perpendicular bisector s. Figure 6 This is a schematic diagram of the surface projection straight line method. Using the RANSAC algorithm, straight lines are fitted to two groups of point clouds divided according to elevation changes along the x-coordinate direction, and the general equations of the two projected straight lines l1 and l2 are obtained as equation (10):

[0069]

[0070] Since the steel coil is tangent to the two supporting planes, the x and z coordinates of the center point of the steel coil saddle can be obtained using equation (11):

[0071]

[0072] Equation (11) is the formula for the distance from point O to the two lines, R is the radius of the steel coil, which is a known value, and A1, B1, C1, A2, B2, and C2 are the corresponding coefficients of the general equations of the two lines.

[0073] The aforementioned elevation threshold, distance threshold, and point cloud density threshold are set as follows: Based on the actual height difference between the vehicle's front and the vehicle's horizontal plane, the elevation threshold is set to two-thirds of the highest height of the points on the elevation map; based on the resolution of the point cloud data from the 3D laser scanning gimbal, the distance threshold is set to 0.05m; the point cloud density of all calculated points is statistically analyzed according to a normal distribution, and the mean μ and variance σ of this normal distribution are calculated. Points falling within the interval (-∞, μ-3σ) in the normal distribution are identified as noise points and deleted. Therefore, the point cloud density threshold is set to the point cloud density value corresponding to the point at the horizontal coordinate μ-3σ in this normal distribution.

[0074] The present invention also provides a crane positioning system based on the above method (hereinafter referred to as the system, see below). Figures 7-8 The system includes a 3D laser scanning gimbal, a host computer, a PLC programmable controller, and an overhead crane. The 3D laser scanning gimbal is installed at the lower end of the overhead crane's main beam. The host computer is equipped with laser scanning gimbal control software and a point cloud data processing visualization interface. The host computer communicates bidirectionally with the laser scanning gimbal via TCP / IP. The PLC programmable controller communicates with the host computer and the overhead crane via an RS485 interface.

[0075] The positioning process of the overhead crane system is as follows: a 3D laser scanning gimbal scans the onboard steel coil saddle and transmits the acquired point cloud data of the steel coil saddle to a host computer. The point cloud data is then segmented by the visualization point cloud processing software in the host computer to obtain the position coordinates of the center point of the steel coil saddle. The PLC programmable controller controls the overhead crane to pick up the steel coil from the logistics warehouse and move it to the onboard area. It then determines whether the crane has moved above the target saddle. If so, the crane boom is lowered to the target saddle position to complete the steel coil loading. Otherwise, the host computer replans the crane's movement path based on the current boom position and the center position of the saddle to move the crane to the target position. The visualization point cloud processing software includes the above-mentioned onboard steel coil saddle point cloud segmentation method.

[0076] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for segmenting point clouds of a vehicle-mounted steel coil saddle, characterized in that, The method includes the following steps: The first step is to acquire point cloud data of the vehicle-mounted steel coil saddle using a 3D laser scanning gimbal and filter the point cloud data in the x, y, and z coordinate directions to segment the region of interest; where x, y, and z are the vehicle driving direction, the direction perpendicular to the vehicle body, and the direction perpendicular to the ground, respectively. The second step is to set an elevation threshold, use elevation changes in the x-coordinate direction to cluster the point cloud data of the vehicle plane and the saddle, and use the RANSAC algorithm to fit the vehicle plane. The third step is to use Euclidean clustering to cluster the independently distributed saddle point cloud data in space. For each saddle point cloud cluster, an improved local convex connection growth algorithm is used to optimize the segmentation effect of the point cloud data and refine the segmentation surface boundary. The improved local convex connection growth algorithm includes constructing a supervoxel adjacency graph of the object surface block by multi-scale supervoxel fusion segmentation, segmenting the supervoxel adjacency graph by classifying whether the connection between two supervoxels is convex or concave, dividing the same subgraph belonging to the effective convex connection edge by using a region growth algorithm, and finally optimizing the segmentation boundary by boundary marking to obtain the segmented saddle support plane point cloud. The multi-scale supervoxel fusion segmentation process is as follows: A distance threshold d is set. T The minimum size for dividing the supervoxel is used to divide the 3D point cloud into regular supervoxels V of size S. i For each supervoxel V i The PCA algorithm is used to fit the plane P. i Calculate the distance d from each point within the supervoxel to the fitted plane. If the maximum distance d from a point within the supervoxel to the plane is... max Less than the distance threshold d T If the segmentation of the supervoxel is successful, the segmentation ends; otherwise, the supervoxel will be divided into four sub-supervoxels. A new fitting plane will be fitted for each sub-supervoxel, and the maximum distance d from the interior point of the sub-supervoxel to the plane will be determined. max Is it less than the distance threshold d? T If the value is less than the minimum size, the segmentation ends; if it is not less than the minimum size, the sub-hypervoxes are re-divided in the above manner until the segmentation ends. If the current sub-hypervoxe reaches the minimum size of the hypervoxe division but still cannot satisfy the maximum distance d, the segmentation ends. max Less than the distance threshold d T If so, all points belonging to the current sub-hypervoxel will be considered noise points and filtered out; The boundary marker optimization uses a maximum optimization function W. * As shown in formulas (7)-(9): Where W is the label of the point on the subgraph after region growing; ω represents all the effective partitioning planes obtained after region growing; a subgraph after region growing represents a plane, denoted by Q. k Let k represent the label of the plane, N be the number of points on all validly segmented planes, and G(W) be the boundary function. Representing point x i Belongs to plane Q k , |N j | indicates the number of neighboring points; For evaluation functions; Step 4: Using the segmented saddle support plane points, calculate the x, y, and z coordinates of the vehicle-mounted steel coil saddle. This completes the segmentation and positioning of the vehicle-mounted steel coil saddle.

2. The point cloud segmentation method for vehicle-mounted steel coil saddle according to claim 1, characterized in that, In the first step, the filtering process is as follows: the point cloud data is filtered in the x, y, and z coordinate directions using pass-through filtering, which is implemented by voxel downsampling and density change-based filtering. Given a point cloud density threshold, the average distance from each point to its k nearest points is calculated and defined as the point cloud density at that point. If the distance is less than the given point cloud density threshold, it is determined to be a noisy point cloud and deleted.

3. The point cloud segmentation method for vehicle-mounted steel coil saddle according to claim 1, characterized in that, The fourth step, calculating the x, y, and z coordinates of the vehicle-mounted steel coil saddle, is as follows: The segmented point cloud of the saddle support plane is divided into two groups according to the elevation change along the y-coordinate direction. The midpoint of the line connecting the centroids of the two groups of point clouds is calculated, thus obtaining the y-coordinate of the center point of the steel coil saddle. The segmented point cloud of the support plane is also divided into two groups according to the elevation change along the x-coordinate direction. The two groups of point cloud data are fitted to straight lines respectively. The x and z coordinates of the center point of the steel coil saddle are obtained according to the distance from the center point of the steel coil to these two straight lines = the radius of the steel coil. Thus, the position coordinates of the center point of the steel coil saddle are obtained, achieving positioning.

4. The point cloud segmentation method for vehicle-mounted steel coil saddle according to claim 3, characterized in that, The specific process for obtaining the x and z coordinates of the center point of the steel coil saddle is as follows: The segmented support plane point cloud is projected onto the xz plane to obtain two projection lines l1 and l2 of the support plane point cloud on the xz plane. Point O is the projection point of the steel coil placement position and is located on the perpendicular bisector s. The RANSAC algorithm is used to perform line fitting on the two groups of point clouds divided according to the elevation change along the x-coordinate direction to obtain the general equation of the two projection lines l1 and l2 as equation (10): Since the steel coil is tangent to the two supporting planes, the x and z coordinates of the center point of the steel coil saddle can be obtained using equation (11): Equation (11) is the formula for the distance from point O to the two lines, R is the radius of the steel coil, which is a known value, and A1, B1, C1, A2, B2, and C2 are the corresponding coefficients of the general equations of the two lines.

5. A crane positioning system, characterized in that, The system includes a 3D laser scanning gimbal, a host computer, a PLC programmable controller, and an overhead crane. The 3D laser scanning gimbal is installed at the lower end of the overhead crane's main beam. The host computer contains laser scanning gimbal control software and a point cloud data processing visualization interface. The host computer communicates bidirectionally with the laser scanning gimbal via TCP / IP. The PLC programmable controller communicates with the host computer and the overhead crane via an RS485 interface. The host computer is loaded with the point cloud segmentation method for vehicle-mounted steel coil saddle as described in any one of claims 1-4.

Images