Tunnel section extraction method based on deep learning and PCL algorithm

Abstract

The invention discloses a tunnel section extraction method based on deep learning and a PCL algorithm, and belongs to the technical field of civil engineering. The method comprises the following steps: S1, carrying out point cloud noise reduction on tunnel point cloud data by adopting a PointClean Net deep learning algorithm; s2, semantic segmentation is carried out through a PointNet deep learning network algorithm to extract point cloud data of different positions in the tunnel; s3, extracting a lining boundary through coordinate transformation and projection, calculating and extracting a midline point according to a boundary point cloud, and extracting a three-dimensional central axis of the tunnel by adopting a least square method; and S4, based on the three-dimensional central axis of the tunnel, performing slice extraction on the lining point cloud, converting a three-dimensional slice into a two-dimensional coordinate, and fitting the section point cloud by adopting an RANSAC algorithm to obtain a tunnel section point cloud under the two-dimensional coordinate. According to the method, the tunnel section point cloud can be accurately extracted, and the extraction precision of the tunnel section information is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of civil engineering, and in particular relates to a tunnel section extraction method based on deep learning and PCL algorithm. Background Art

[0002] At present, traditional tunnel detection methods mainly rely on manual measurement and simple sensors. These methods not only consume a lot of manpower and time, but are also easily affected by environmental conditions, resulting in insufficient accuracy and reliability of the detection results. In addition, the complex geometry and environmental conditions of long-distance tunnels make it difficult for manual detection and traditional methods to fully and accurately obtain the structural information of the tunnel. With the development of technology, 3D laser scanning technology has gradually been applied to tunnel detection. 3D laser scanning technology is fast, simple and accurate. It can efficiently obtain complex geometric information in the tunnel and provides advanced technical support for tunnel deformation monitoring.

[0003] However, the amount of point cloud data is large and contains a large number of noise points. The point cloud data processing process is complex and cumbersome, especially in the point cloud processing of tunnels with longer spacing, these problems are more prominent. The existing point cloud processing technology has the following limitations: There are many noise points, which increases the difficulty of data cleaning and noise reduction. The processing process is complicated. Point cloud data processing includes multiple steps such as data preprocessing, noise reduction, alignment, segmentation, fitting, etc. Each step requires complex algorithms and fine parameter adjustments. The overall process is cumbersome and time-consuming. The cross-sectional accuracy is insufficient. The existing point cloud processing technology often finds it difficult to meet the high-precision requirements when processing complex tunnel geometries. Especially when extracting tunnel cross-sectional information, the accuracy problem is particularly prominent. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the above-mentioned related art at least to a certain extent.

[0005] To this end, the purpose of the present invention is to provide a tunnel section extraction method based on deep learning and PCL algorithm, which can accurately extract tunnel section point cloud and effectively improve the extraction accuracy of tunnel section information.

[0006] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0007] The embodiment of the present invention provides a tunnel section extraction method based on deep learning and PCL algorithm, the method comprising:

[0008] S1. Use PointCleanNet deep learning algorithm to perform point cloud denoising on the acquired tunnel point cloud data;

[0009] S2. Use the PointNet deep learning network algorithm to perform semantic segmentation on the tunnel point cloud data and extract point cloud data at different locations in the tunnel, including the lining;

[0010] S3. Accurately extract the lining boundary point cloud through coordinate transformation and projection, calculate and extract the centerline points based on the boundary point cloud, and use the least squares method to extract the three-dimensional centerline of the tunnel;

[0011] S4. Based on the three-dimensional central axis of the tunnel, the lining point cloud is sliced ​​and extracted, the three-dimensional slices are converted to two-dimensional coordinates, and then the RANSAC algorithm is used to fit the cross-sectional point cloud to obtain the tunnel cross-sectional point cloud in two-dimensional coordinates.

[0012] In addition, the tunnel section extraction method based on deep learning and PCL algorithm according to the present invention may also have the following additional technical features:

[0013] In some implementation modes, the specific contents of step S1 include:

[0014] S11, performing voxel downsampling processing on the acquired tunnel point cloud data to obtain low-density tunnel point cloud data;

[0015] S12, training point cloud denoising deep learning model;

[0016] S13. Use the trained point cloud denoising deep learning model to perform denoising on the point cloud data.

[0017] In some of the implementation modes, the specific content of voxel downsampling in step S11 includes: dividing the tunnel point cloud data into a number of smaller voxels, and taking the center point of each voxel to form new point cloud data.

[0018] In some of the embodiments, in step S12, a training set is prepared by three-dimensionally annotating noise points, and the obtained training set is used to perform model training on the PointCleanNet deep learning algorithm to obtain a tunnel point cloud denoising model.

[0019] In some implementation modes, the specific contents of step S2 include:

[0020] S21. Use the PointNet point cloud object segmentation network to perform semantic segmentation on the point cloud data, and then label it to create a point cloud training set.

[0021] S22, training the model using the point cloud training set to obtain a tunnel point cloud segmentation model;

[0022] S23. Use the trained tunnel point cloud segmentation model to perform semantic segmentation on the point cloud data and extract the point cloud data of the tunnel lining.

[0023] In some of the embodiments, the content annotated in step S21 includes lining, ground, appendages and pipeline point clouds.

[0024] In some implementation modes, the specific contents of step S3 include:

[0025] S31, correcting the coordinate system of the point cloud data of the tunnel lining, and projecting the three-dimensional point cloud onto the XOY and YOZ planes, respectively; wherein X, Y, and Z represent the three axes of the coordinate system, respectively, and O represents the origin of the coordinate system;

[0026] S32, extracting the boundary of the tunnel lining by point cloud search to obtain boundary point cloud data of the tunnel lining;

[0027] S33, extracting lining YOZ surface two-dimensional centerline point cloud data and XOY surface two-dimensional centerline point cloud data;

[0028] S34. The linear model of the tunnel is obtained by fitting using the least square method, and the central axis of the tunnel is obtained by fitting the central axis of the joint lining projection.

[0029] In some of the embodiments, the criterion for the coordinate system correction in step S31 is that the direction of the tunnel is aligned with the coordinate axis;

[0030] Step S32 includes: along the axis direction of the tunnel, from the y min Starting from the position, set an appropriate step size and search for the maximum value x in the point cloud along the Y axis. max , minimum value x min , thereby extracting the boundary point cloud data of the lining; y min is the minimum value of the Y axis, x max is the maximum value of the X axis, x min is the maximum and minimum value;

[0031] Step S33 includes: calculating the mean of x as the midpoint of the horizontal midline, where The corresponding horizontal midline point coordinates are Traverse all point clouds along the Y axis to obtain the 2D midline point cloud of the lining YOZ surface, y i is the i-th value of the Y axis; the two-dimensional midline point cloud of the XOY point cloud is obtained in the same way.

[0032] In some implementation modes, the specific contents of step S4 include:

[0033] S41, setting the thickness of the point cloud slice and the spacing between adjacent slices along the central axis of the tunnel, and performing slice filtering processing;

[0034] S42, extracting slice point cloud using KD-Tree clustering segmentation algorithm;

[0035] S43, using PCA least squares method to perform plane fitting on the point cloud to achieve conversion from three-dimensional to two-dimensional slices;

[0036] S44. Extract tunnel section point cloud based on RANSAC algorithm.

[0037] In some embodiments, the sub-steps of step S44 include:

[0038] S441, randomly selecting two points from the point cloud data as sample points, which are used as initial parameters for generating a fitting model;

[0039] S442, draw a straight line using the two selected points, and calculate the geometric relationship between the points to obtain a preliminary fitting model;

[0040] S443, calculating the distances from the remaining points in the point cloud data to the straight line, and counting the number of points whose distances are less than the threshold according to a preset threshold;

[0041] S444, using the remaining point cloud data to verify the quality of the current model;

[0042] S445, repeat the above steps multiple times until the best ellipse model parameters are found.

[0043] Compared with the prior art, the present invention has at least the following beneficial effects:

[0044] In the embodiment of the present invention, the tunnel section extraction method based on deep learning and PCL algorithm is provided, which uses deep learning method to reduce noise and semantic segmentation of tunnel point cloud data, greatly improving the automation of data processing. Compared with traditional methods, this technology not only saves a lot of manpower and time, but also can efficiently and quickly complete complex tunnel detection tasks;

[0045] In the embodiment of the present invention, the tunnel section extraction method based on deep learning and PCL algorithm is provided, which can accurately remove noise points in point cloud data and accurately extract point cloud data of tunnel lining and ground through the training and optimization of deep learning model; the combined use of multiple fitting algorithms ensures the accuracy of boundary extraction of different types of tunnels;

[0046] In the embodiment of the present invention, the tunnel section extraction method based on deep learning and PCL algorithm is provided. In the processing process, the KD-Tree algorithm is combined to perform nearest neighbor search, thereby improving the efficiency and accuracy of point cloud processing. At the same time, the RANSAC algorithm is used for random sampling consistency fitting, thereby ensuring that the tunnel section can still be accurately extracted in the presence of noise points.

[0047] In the embodiment of the present invention, the tunnel section extraction method based on deep learning and PCL algorithm provided reduces the dependence on manual inspection, reduces labor costs, and improves the quality and efficiency of inspection work; this has significant economic benefits and practical value for the inspection and maintenance of long-distance tunnels.

[0048] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of a tunnel section extraction method based on deep learning and PCL algorithm disclosed in one embodiment of the present invention;

[0050] Figure 2 A schematic diagram of a point cloud segmentation network framework disclosed in an embodiment of the present invention;

[0051] Figure 3 A schematic diagram of a lining boundary projection disclosed in an embodiment of the present invention;

[0052] Figure 4 A schematic diagram of a least squares fitting midpoint line disclosed in an embodiment of the present invention;

[0053] Figure 5 A schematic diagram of a tunnel central axis disclosed in an embodiment of the present invention;

[0054] Figure 6 A schematic diagram of a tunnel slice disclosed in an embodiment of the present invention;

[0055] Figure 7 A schematic diagram of a tunnel plane fitting disclosed in an embodiment of the present invention;

[0056] Figure 8 A schematic diagram of tunnel section extraction disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] The embodiments of the present invention are described in detail below through specific embodiments and application scenarios in conjunction with the accompanying drawings.

[0059] See also Figure 1As shown, in some embodiments of the present invention, a tunnel section extraction method based on deep learning and PCL algorithm is provided, which uses the three-dimensional point cloud data of the tunnel to adopt a method combining deep learning network and point cloud processing technology to perform noise reduction processing on the tunnel point cloud and extract the tunnel lining point cloud; and based on this, the lining boundary is extracted, and the centerline point is calculated according to the boundary, the tunnel centerline is fitted, and the cluster segmentation algorithm is used to extract the tunnel section based on the centerline; the steps include:

[0060] Step 1: Use the PointCleanNet deep learning algorithm to perform point cloud denoising on the acquired tunnel point cloud data;

[0061] 1.1 Voxel downsampling

[0062] The original tunnel point cloud is processed by voxel downsampling to reduce the point cloud density while maintaining the overall geometric features, thereby reducing the data volume and algorithm complexity. This step is done by dividing the point cloud into smaller voxels and taking the center point of each voxel to form a new point cloud, thereby achieving noise reduction and simplified processing.

[0063] 1.2. Training point cloud denoising deep learning model

[0064] A large number of noise points will be generated in tunnel environments due to factors such as debris and equipment. The PointCleanNet point cloud cleaning network framework is used to create a training set by 3D annotating noise points, and then the model is trained to obtain a tunnel point cloud denoising model. PointCleanNet uses the powerful learning ability of deep learning to train on a large amount of data, so that the model can effectively identify and remove noise points.

[0065] 1.3 Point cloud denoising processing based on point cloud denoising model

[0066] The trained tunnel point cloud denoising model is used to denoise the point cloud data; the model can identify and distinguish noise points from actual tunnel points, and remove noise points while retaining the tunnel lining and ground features, thereby improving the quality of the point cloud data.

[0067] Step 2: Use the PointNet deep learning network algorithm to perform semantic segmentation on the tunnel lining and the ground to extract point cloud data such as the tunnel lining, ground, appendages, and pipelines. The point cloud segmentation network framework is as follows: Figure 2As shown in the figure, the PointNet model extracts local and global features of the input point cloud data through a multi-layer perceptron (MLP), and then maps these features to corresponding categories through a classification layer, thereby achieving effective classification of the point cloud data; it retains the characteristics of each point by point-by-point operations, and captures the overall geometric information of the point cloud through global feature aggregation, and finally outputs the classification results through a fully connected layer; this structure enables PointNet to process irregular point cloud data while having high computational efficiency and flexibility.

[0068] The specific steps of this step include:

[0069] 2.1. Create a point cloud training set

[0070] The PointNet point cloud object segmentation network framework is used to manually annotate the 3D lining and ground point clouds to create a training set. PointNet is a classic deep learning model that can perform semantic segmentation on point cloud data. By annotating the lining, ground, appendages, and pipeline point clouds, a model that can accurately identify these structures is trained.

[0071] 2.2. Training point cloud semantic segmentation model

[0072] The point cloud and the corresponding label file are input into the instance segmentation network for training to obtain the tunnel point cloud segmentation model. Through training with a large amount of data, the model can learn and identify the features of the tunnel lining, ground, appendages, and pipeline point clouds, and achieve high-precision semantic segmentation.

[0073] 2.3 Tunnel lining extraction based on point cloud segmentation model

[0074] The trained tunnel point cloud segmentation model is used to perform semantic segmentation on the point cloud data. The model can identify and mark each point as a tunnel lining or other part; based on the semantic label, the point cloud data of the tunnel lining is extracted to provide accurate input data for subsequent processing.

[0075] Step 3: Accurately extract the lining boundary point cloud through coordinate transformation and projection, calculate and extract the centerline points based on the boundary point cloud, and extract the three-dimensional centerline using the least squares method;

[0076] 3.1 Lining coordinate transformation and double projection

[0077] The coordinate system of the tunnel lining point cloud is transformed so that the direction of the tunnel is basically aligned with the coordinate axis; then, the three-dimensional point cloud is projected onto the XOY and YOZ planes respectively, which can simplify the calculation of the midline points in the subsequent processing; the lining boundary projection is as follows Figure 3 As shown, it is calculated by calculating the mean of the point cloud in the x direction Get the midpoint of the horizontal midline Then traverse all points along the y-axis to generate a two-dimensional midline point cloud of the XOY plane;

[0078] 3.2 Point cloud search and boundary extraction

[0079] Along the axis of the tunnel, from the y min Starting from the position, set an appropriate step size and search for the maximum value x in the point cloud along the y-axis direction max , minimum value x min ;Through this method, the boundary point cloud data of the lining can be extracted;

[0080] 3.3. Calculation and extraction of midline points

[0081] Calculate the mean of x As the midpoint of the horizontal midline, According to the mean The corresponding horizontal midline point coordinates are Traverse all point clouds along the y-axis to obtain the two-dimensional centerline point cloud of the lining YOZ surface. Similarly, obtain the two-dimensional centerline point cloud of the XOY point cloud.

[0082] 3.4. 3D central axis fitting

[0083] For the straight tunnel part, the least squares method is used for fitting. The best straight line model can be fitted through the principle of "minimum residual sum of squares" of the least squares method. The tunnel centerline is obtained by fitting the equation of the lining projection centerline. Its three-dimensional space equation is:

[0084]

[0085] The "minimum residual sum of squares" principle of the least squares method can be used to fit the best straight line model and further extract the central axis of the tunnel. The results of the least squares method fitting the midpoint line are as follows: Figure 4 As shown in the figure, the tunnel centerline is obtained by fitting the equation as Figure 5 shown.

[0086] Step 4: Based on the three-dimensional central axis of the tunnel, slice the lining point cloud, convert the three-dimensional slices into two-dimensional coordinates, and finally use the RANSAC algorithm to fit the cross-section point cloud to obtain the tunnel cross-section point cloud in two-dimensional coordinates;

[0087] 4.1. Tunnel point cloud slicing processing according to the 3D central axis

[0088] Based on the tunnel centerline, the direction vector P is obtained; the thickness of the point cloud slice and the spacing between adjacent slices are set along the P direction to perform slice filtering; in this way, the three-dimensional point cloud data can be sliced ​​at a certain interval for subsequent processing; tunnel slices are as follows Figure 6 As shown;

[0089] 4.2. Extracting slice point cloud based on KD-Tree clustering segmentation algorithm

[0090] Use the KD-Tree algorithm to search for neighbors on the tunnel point cloud; by setting a distance threshold, aggregate the point clouds with distances less than the threshold to complete the extraction of 3D cross-section slices; the KD-Tree algorithm can efficiently perform spatial searches and is a common tool in point cloud processing;

[0091] 4.3. Plane fitting of point cloud based on PCA least squares method

[0092] PCA performs least squares plane fitting on the local point cloud and uses the obtained plane normal vector as the normal vector of the sampling point. By calculating the rotation parameter between the plane normal vector and the XOY plane direction vector, the point cloud is rotated around the centroid to complete the conversion from 3D to 2D slices. This step uses PCA to reduce the dimension and convert the 3D slices into a 2D coordinate system, thereby obtaining the 2D cross-sectional data of the tunnel. The tunnel plane fitting results are shown in Figure 2. Figure 7 As shown;

[0093] 4.4 Tunnel section point cloud extraction based on RANSAC algorithm

[0094] The main steps of the random sampling consistency fitting algorithm include random sampling, model fitting, evaluation and iteration; the algorithm iteration process is as follows:

[0095] 1) Randomly select 2 points from the point cloud data as sample points; these two points are used to generate the initial parameters of the fitting model; since the definition of an ellipse requires at least two points, these two points are selected as the starting points;

[0096] 2) Draw a straight line using the two selected points and calculate the geometric relationship of the selected points to obtain a preliminary fitting model;

[0097] 3) Calculate the distance from the ellipse point cloud except the selected point in the point cloud data to the straight line, and count the number of points whose distance is less than the threshold value according to the pre-set threshold value t; the calculation process is as follows:

[0098] (1-(1-e) s ) N =1-p (3)

[0099] Among them, e is the outlier rate, s is the number of sampling points each time, and p is the correct probability;

[0100] 4) Use the point cloud with a distance less than the threshold to verify the quality of the current model;

[0101] 5) Repeat the above steps several times until the optimal ellipse model parameters are found;

[0102] The parametric equation of the ellipse is:

[0103]

[0104] Among them, u, v, and n satisfy perpendicular to each other, a is the radius of the long side of the ellipse, b is the radius of the short side of the ellipse, (C x , C y , C z )) is the coordinate of the center point of the ellipse; the equation is solved by the random sampling consistency fitting algorithm to obtain the ellipse parameters a, b, and the coordinate of the center point; by calculating Figure 7 The rotation parameter between the plane normal vector and the XOY plane direction vector rotates the point cloud around the center of mass to complete the conversion from 3D to 2D slices, and the 2D tunnel section extraction result is obtained as follows: Figure 8 shown.

[0105] Parts of the present invention that are not described in detail are well known to those skilled in the art.

[0106] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A tunnel section extraction method based on deep learning and PCL algorithm, characterized in that: The method comprises: S1. Use PointCleanNet deep learning algorithm to perform point cloud denoising on the acquired tunnel point cloud data; S2. Use the PointNet deep learning network algorithm to perform semantic segmentation on the tunnel point cloud data and extract point cloud data at different locations in the tunnel, including the lining; S3. Accurately extract the lining boundary point cloud through coordinate transformation and projection, calculate and extract the centerline points based on the boundary point cloud, and use the least squares method to extract the three-dimensional centerline of the tunnel; S4. Based on the three-dimensional central axis of the tunnel, the lining point cloud is sliced ​​and extracted, the three-dimensional slices are converted to two-dimensional coordinates, and then the RANSAC algorithm is used to fit the cross-sectional point cloud to obtain the tunnel cross-sectional point cloud in two-dimensional coordinates.

2. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 1 is characterized in that: The specific contents of step S1 include: S11, performing voxel downsampling processing on the acquired tunnel point cloud data to obtain low-density tunnel point cloud data; S12, training point cloud denoising deep learning model; S13. Use the trained point cloud denoising deep learning model to perform denoising on the point cloud data.

3. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 2 is characterized in that: The specific content of voxel downsampling in step S11 includes: dividing the tunnel point cloud data into a number of smaller voxels, and taking the center point of each voxel to form new point cloud data.

4. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 2 is characterized in that: In step S12, a training set is prepared by three-dimensionally annotating noise points, and the obtained training set is used to perform model training on the PointCleanNet deep learning algorithm to obtain a tunnel point cloud denoising model.

5. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 1 is characterized in that: The specific contents of step S2 include: S21. Use the PointNet point cloud object segmentation network to perform semantic segmentation on the point cloud data, and then label it to create a point cloud training set. S22, training the model using the point cloud training set to obtain a tunnel point cloud segmentation model; S23. Use the trained tunnel point cloud segmentation model to perform semantic segmentation on the point cloud data and extract the point cloud data of the tunnel lining.

6. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 5 is characterized in that: The content annotated in step S21 includes lining, ground, appendages and pipeline point clouds.

7. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 1 is characterized in that: The specific contents of step S3 include: S31, correcting the coordinate system of the point cloud data of the tunnel lining, and projecting the three-dimensional point cloud onto the XOY and YOZ planes, respectively; wherein X, Y, and Z represent the three axes of the coordinate system, respectively, and O represents the origin of the coordinate system; S32, extracting the boundary of the tunnel lining by point cloud search to obtain boundary point cloud data of the tunnel lining; S33, extracting lining YOZ surface two-dimensional centerline point cloud data and XOY surface two-dimensional centerline point cloud data; S34. The linear model of the tunnel is obtained by fitting using the least square method, and the central axis of the tunnel is obtained by fitting the central axis of the joint lining projection.

8. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 7 is characterized in that: The standard for coordinate system correction in step S31 is that the direction of the tunnel is aligned with the coordinate axis; Step S32 includes: along the axis direction of the tunnel, from the y min Starting from the position, set an appropriate step size and search for the maximum value x in the point cloud along the Y axis. max , minimum value x min , thereby extracting the boundary point cloud data of the lining; y min is the minimum value of the Y axis, x max is the maximum value of the X axis, x min is the maximum and minimum value; Step S33 includes: calculating the mean of x as the midpoint of the horizontal midline, where The corresponding horizontal midline point coordinates are Traverse all point clouds along the Y axis to obtain the 2D midline point cloud of the lining YOZ surface, y i is the i-th value of the Y axis; the two-dimensional midline point cloud of the XOY point cloud is obtained in the same way.

9. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 1 is characterized in that: The specific contents of step S4 include: S41, setting the thickness of the point cloud slice and the spacing between adjacent slices along the central axis of the tunnel, and performing slice filtering processing; S42, extracting slice point cloud using KD-Tree clustering segmentation algorithm; S43, using PCA least squares method to perform plane fitting on the point cloud to achieve conversion from three-dimensional to two-dimensional slices; S44. Extract tunnel section point cloud based on RANSAC algorithm.

10. The tunnel section extraction method based on deep learning and PCL algorithm according to claim 9 is characterized in that: The sub-steps of step S44 include: S441, randomly selecting two points from the point cloud data as sample points, which are used as initial parameters for generating a fitting model; S442, draw a straight line using the two selected points, and calculate the geometric relationship between the points to obtain a preliminary fitting model; S443, calculating the distances from the remaining points in the point cloud data to the straight line, and counting the number of points whose distances are less than the threshold according to a preset threshold; S444, using the remaining point cloud data to verify the quality of the current model; S445, repeat the above steps multiple times until the best ellipse model parameters are found.

Images