Precise point cloud modeling method and system based on laser point cloud and deep learning algorithm

Abstract

The invention provides a point cloud accurate modeling method and system based on a laser point cloud and a deep learning algorithm, and the method comprises the steps: introducing an input transformation matrix and a maximum pooling layer into a Point-Net neural network, and extracting the global features of the laser point cloud; further processing the laser point cloud data by using a CSF algorithm; extracting the relative elevation of the laser point cloud data; and building a laser point cloud model, and training and testing the image. The invention provides a Point-Net deep learning classification algorithm based on PCA (Principal Component Analysis) dimension reduction. Firstly, in order to remove the influence of noise points on a classification result, denoising processing is carried out on point cloud data; in order to avoid the problem of insufficient feature information in point cloud classification, relative elevation features of the point cloud are extracted based on elevation information of the point cloud; thirdly, PCA data dimension reduction is carried out on the laser point cloud multi-dimensional data, and redundant data are removed; and finally, inputting the point cloud data into the Point-Net network, training a classification model, classifying experimental data, and comparing classification precision of different classification methods.

Description

Technical Field

[0001] The present invention belongs to the field of power grid radar laser point cloud modeling, and specifically relates to a point cloud precise modeling method based on laser point cloud and deep learning algorithm. Background Art

[0002] As an important channel for energy transmission, the safe and stable operation of power transmission lines is crucial to social development. Traditional power grid transmission line inspections often rely on manual inspections, but this method has the problems of high risk, low efficiency and difficulty in comprehensive coverage. At the same time, traditional inspection methods have certain limitations for inspections in complex terrain and vast areas. Therefore, efficient and accurate methods for collecting information on transmission lines are of great significance. Improving the modeling accuracy of laser point clouds plays an important role in fully perceiving the status and health of transmission lines.

[0003] The current methods of laser point cloud modeling include the following:

[0004] The point cloud data of the transmission line is collected by combining LiDAR and IMU.

[0005] The downsampling function in the Open3d library and the GaussianKDE function are used to calculate the point cloud density distribution in the two-dimensional plane and build a model.

[0006] PCA data dimensionality reduction and Point-Net neural network are used to establish the network model.

[0007] The existing methods for modeling laser point clouds have the following disadvantages:

[0008] 1. The method of using machine learning for identification has a complicated and difficult algorithm to construct, a low accuracy rate, and an extremely slow identification speed, making it impossible to achieve real-time detection.

[0009] 2. The modeling method that uses the downsampling function in the open3d library and the GaussianKDE function to calculate the point cloud density distribution in the two-dimensional plane has a fast modeling speed and high efficiency, but the accuracy is low and accurate modeling cannot be performed.

[0010] 3. The method of modeling the point cloud data of power transmission lines by combining LiDAR and IMU has a small collection error and high accuracy in data classification and processing, but the problems of uneven point cloud density and point cloud data registration are not solved.

[0011] 4. PCA data dimensionality reduction and Point-Net neural network are used to build network models. Information classification is accurate, which avoids the time-consuming and laborious point cloud multi-feature extraction of point cloud data. However, when selecting samples, only limited samples can be selected, which will affect the classification accuracy. There is also a problem of relatively few classification categories during classification. Summary of the invention

[0012] In order to solve the deficiencies in the prior art, the present invention provides a point cloud precise modeling method and system based on laser point cloud and deep learning algorithm.

[0013] The present invention adopts the following technical solution.

[0014] The present invention provides a point cloud accurate modeling method based on laser point cloud and deep learning algorithm, comprising:

[0015] The input transformation matrix and maximum pooling layer are introduced into the Point-Net neural network to extract the global features of the laser point cloud;

[0016] The laser point cloud data is further processed using the CSF algorithm;

[0017] Extract relative elevation from laser point cloud data;

[0018] Build a laser point cloud model to train and test images.

[0019] Preferably, the further processing of the laser point cloud data using the CSF algorithm includes:

[0020] The point cloud is inverted, and all laser points are projected onto the horizontal plane. The points on the plane are used as grid points. The height value of the laser point closest to the simulated point is obtained as IHV based on the KD tree nearest neighbor index, and all DTM digital ground model particles are traversed; the CSF algorithm is compiled using the programming language according to PCL, and the internal force between the simulated cloth particles is calculated. According to the rigidity parameters of the cloth, the relative position between the simulated cloth particles is determined; iterative calculations are performed, and the iteration is stopped after the preset conditions are met; the distance between the laser point cloud and the corresponding cloth particle is calculated, and when the distance is less than or equal to the maximum distance threshold, it is marked as a ground point, and when the distance is greater than the maximum distance threshold, it is marked as a non-ground point, and the maximum distance threshold = median + standard deviation multiple × standard deviation.

[0021] Preferably, the inverting the point cloud comprises:

[0022] Using the CSF simulation filtering algorithm, the terrain is flipped and the resulting point cloud data is converted from the DSM digital surface model to the DTM digital terrain model.

[0023] Preferably, the traversing all DTM digital terrain model particles comprises:

[0024] When the particle is affected by gravity, if the particle's height value CHV is less than IHV, the particle's height value is set to IHV, and the current cloth particle is set to an immovable particle.

[0025] Preferably, the extracting the relative elevation of the laser point cloud data comprises:

[0026] Get the ground points and non-ground points, build a KD tree for the ground points, search for k nearest neighbor points in the ground points for all point clouds, calculate the average of the k nearest neighbor points, and traverse to calculate the relative elevation of all point clouds. The formula is:

[0027] H 相对高程 =HH 均值 .

[0028] Preferably, the step of building a laser point cloud model and training and testing the image includes:

[0029] Obtain point cloud images containing transmission lines and transmission towers, perform manual or semi-automatic manual annotation through data annotation software to form an original data set, and randomly divide the data set into a training set and a test set in a ratio of 6:1; extract features of the images in the training set layer by layer, calculate the loss function of the anchor box and the annotated detection box formed by clustering, and perform error backpropagation to correct the network parameters.

[0030] Preferably, obtaining a loss function for the clustered anchor boxes and the labeled detection boxes includes:

[0031]

[0032] Among them, L IOU represents the overlap loss, L dis represents the center distance loss, L asp represents the width and height loss, w, h are the width and height of the prediction box, b is the image center distance, c is the aspect ratio of the image, C w and C h It is the width and height of the minimum bounding box covering the two Boxes, IOU = |A∩B| / |A∪B|, which refers to the intersection-over-union ratio of images A and B. g is the iteration step size, t is the number of iterations, and ρ represents the Euclidean distance between the two center points.

[0033] The present invention also provides a point cloud precision modeling system based on laser point cloud and deep learning algorithm, the system is a system used by the aforementioned point cloud precision modeling method based on laser point cloud and deep learning algorithm, comprising:

[0034] Feature extraction module, used to extract global features of laser point cloud;

[0035] CSF algorithm module, used to further process laser point cloud data using CSF algorithm;

[0036] The relative elevation extraction module is used to extract the relative elevation of laser point cloud data.

[0037] The present invention also provides a terminal, including a processor and a storage medium;

[0038] The storage medium is used to store instructions;

[0039] The processor is configured to operate according to the instructions to execute the steps according to the aforementioned method.

[0040] The present invention also provides a computer-readable storage medium on which a computer program is stored, characterized in that the program implements the steps of the above method when executed by a processor.

[0041] The beneficial effect of the present invention is that, compared with the prior art, the most popular direction of point cloud processing is to apply machine learning models to point cloud data classification. The most commonly used laser point cloud classification methods now include random forests, support vector machines, cluster analysis, deep learning and other methods. There are also many researchers currently committed to the research of laser point cloud classification. In the current research process of point cloud classification, machine learning-based methods often need to calculate a large number of point cloud features in order to obtain higher accuracy, which will increase the complexity of data processing and cause problems in feature selection.

[0042] Unlike the current algorithms, the algorithm of the present invention can well avoid the extraction of a large number of point cloud features, and can classify a variety of land object types. In order to solve the problems existing in the current point cloud classification process, the present invention proposes a Point-Net deep learning classification algorithm based on PCA dimensionality reduction. First, in order to remove the influence of noise points on the classification results, the point cloud data should be denoised; secondly, in order to avoid the problem of insufficient feature information in point cloud classification, the relative elevation features of the point cloud are extracted based on the elevation information of the point cloud; then, PCA data dimensionality reduction is performed on the multi-dimensional data of the laser point cloud to remove redundant data; finally, the point cloud data is input into the Point-Net network, the classification model is trained, and the experimental data is classified to compare the classification accuracy of different classification methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a schematic diagram of an exemplary application scenario in the present invention;

[0044] Figure 2 It is a schematic diagram of the Point-Net point cloud classification deep learning framework in the present invention;

[0045] Figure 3It is a CSF algorithm measurement principle diagram in the present invention;

[0046] Figure 4 It is a schematic diagram of the deep network in the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0048] Embodiment 1 of the present invention provides a point cloud accurate modeling method based on laser point cloud and deep learning algorithm, comprising:

[0049] Step 1: Introduce the input transformation matrix and the maximum pooling layer into the Point-Net neural network to extract the global features of the laser point cloud;

[0050] like Figure 1 As shown, the algorithm involved in the present invention identifies the target as a laser point cloud of the power grid;

[0051] like Figure 2 As shown, compared with the original version, the algorithm involved in the present invention introduces an input transformation matrix and a maximum pooling layer in the Point-Net neural network. The input transformation matrix network learns a rotation matrix through the position of the point cloud, adjusts the rotation matrix through the loss function, and rotates the input point cloud data to an angle that is more conducive to classification; the input transformation matrix is ​​added to both the input transformation and feature transformation modules. The first time is to rotate the point cloud to an angle that is more conducive to classification, avoiding the problem of low recognition due to stacking and overlap between point clouds of different features, and the second time is to align the extracted 64-dimensional features, that is, to achieve transformation of the point cloud at the feature level. The maximum pooling layer downsamples the input data after the convolution layer, retains the features of the point cloud while reducing the output dimension, and extracts the global features of the point cloud in the Point-Net network.

[0052] Step 2: Use CSF algorithm to further process the laser point cloud data;

[0053] like Figure 3As shown in the figure, it is the principle of the CSF algorithm adopted by the present invention. Assuming that a virtual piece of cloth is acted on the terrain surface by gravity, if the cloth is soft enough, it will be attached to the ground surface. At this time, the shape of the cloth is DSM (digital surface model); when the terrain is flipped, the shape of the cloth attached to the surface is DTM (digital terrain model). Compared with other filtering algorithms, the advantages of CSF simulation filtering algorithm are: 1) It can be applied to several different areas such as steep slopes, flat areas, urban areas, etc., and the corresponding filtering algorithm can be selected according to the terrain of the area; 2) Compared with the complex parameter adjustment process in the traditional filtering algorithm, the parameters of the cloth simulation filtering algorithm are easier to set and are relatively simple and easy to implement. Compared with the traditional filtering algorithm, the CSF algorithm can effectively reduce the amount of calculation and the amount of parameters. The filtering algorithm processing process is as follows:

[0054] Step 2.1, after the point cloud denoising process, the point cloud is inverted (i.e., the CSF analog filtering algorithm is used to flip the terrain. At this time, the point cloud data obtained is converted from the DSM digital surface model to the DTM digital terrain model. The parameters of this algorithm are easier to implement and can have better recognition in complex terrains such as steep slopes).

[0055] Step 2.2, project all laser points onto the horizontal plane, use the points on the plane as grid points, and obtain the height value of the laser point closest to the simulation point based on the KD tree nearest neighbor index as IHV.

[0056] Step 2.3, traverse all DTM digital ground model particles (i.e., particles that make up the DTM digital model as the ground reference). When the particles are affected by gravity, if the particle height value CHV is less than IHV, set the particle height value to IHV, and set the current cloth particle to an immovable particle.

[0057] Step 2.4, use the programming language according to PCL to complete the compilation of the CSF algorithm, calculate the internal force between the simulated cloth particles, and determine the relative positions between the simulated cloth particles according to the rigidity parameters of the cloth that have been set.

[0058] Step 2.5, repeat the calculation process of step 2.3 and step 2.4. After a large number of iterative calculations, it is found that when the number of iterations reaches about 1000 rounds, the accuracy can reach 90.60%, and the iteration can be stopped.

[0059] Step 2.6, calculate the distance between the laser point cloud and the corresponding cloth particle. When the distance is less than or equal to the maximum distance threshold, it is marked as a ground point. When the distance is greater than the maximum distance threshold, it is marked as a non-ground point (maximum distance threshold = median + standard deviation multiple × standard deviation).

[0060] Since point cloud data has elevation information, the relative elevation information of point clouds can well distinguish categories such as buildings, roads and vegetation. Therefore, it is necessary to extract the relative elevation of the point cloud (the difference between the target point and the ground point elevation) on the basis of filtering, so that the type of point cloud can be easily distinguished, which is conducive to the subsequent point cloud classification processing.

[0061] Step 3, extracting the relative elevation of the laser point cloud data;

[0062] Step 3.1, obtain ground points and non-ground points based on cloth simulation filtering.

[0063] Step 3.2: Build a KD tree for the ground points and search for k nearest neighbor points among all the point clouds.

[0064] Step 3.3, calculate the mean of the k nearest neighbor points obtained.

[0065] Step 3.4, traverse and obtain the relative elevation of all point clouds, the formula is:

[0066] H 相对高程 =HH 均值

[0067] (1) Algorithm description:

[0068] Input: Dataset Ground = {(x1, y1, z1), (x2, y2, z2), ..., (x m ,y m , z m )}, All={(x1, y1, z1), (x2, y2, z2),..., (x, n y n ,z n )}; where x, y, and z refer to the three-dimensional coordinate axes of the point cloud. This algorithm can well extract the relative elevation features of the point cloud and can be applied to point cloud classification.

[0069] Step 4: Build a laser point cloud model. The overall image processing process includes two parts: training and testing.

[0070] (1) We use LiDAR and drone aerial photography to collect point cloud images of transmission lines and towers.

[0071] (2) Then, data annotation software is used to perform manual or semi-automatic manual annotation to form the original data set;

[0072] (3) The dataset is randomly divided into training set and test set in a ratio of 6:1, and the training data is expanded by Open3d, maximum pooling, photometric distortion, geometric distortion and motion compensation.

[0073] (4) Again, the prepared training set is passed through the backbone, neck, and detection head of the built object detection network in turn. In the process of feature extraction layer by layer, the feature maps of the image are fused at different scales. Finally, the clustered anchor boxes and the labeled detection boxes are used to obtain the loss function, and the error is back-propagated to correct the network parameters. The specific method is as follows:

[0074] The present invention adopts a new loss function: EIOU. The penalty term of EIOU is based on the penalty term of CIOU, which separates the influencing factor of aspect ratio and calculates the length and width of the target box and anchor box respectively. The loss function contains three parts: overlap loss, center distance loss, and width and height loss. The first two parts continue the method in CIOU, but the width and height loss directly minimizes the difference between the width and height of the target box and the anchor box, making the convergence speed faster. The calculation formula is as follows:

[0075]

[0076] Among them, L IOU represents the overlap loss, L dis represents the center distance loss, L asp represents the width and height loss, w, h are the width and height of the prediction box, b is the image center distance, c is the aspect ratio of the image, C w and C h It is the width and height of the minimum bounding box covering the two Boxes, IOU = |A∩B| / |A∪B|, which refers to the intersection-over-union ratio of images A and B. g is the iteration step size, t is the number of iterations, and ρ represents the Euclidean distance between the two center points.

[0077] EIOU takes into account the overlapping area, center point distance, and true difference in length, width, and side length. It solves the fuzzy definition of aspect ratio based on CIOU and can explain the sample imbalance problem in prediction box regression.

[0078] (5) The test set is passed into the network to test the accuracy of the network model to determine whether the convergence conditions are met.

[0079] (6) Figure 4 As shown in the figure, after steps 4 and 5 are completed, the network weights obtained from the training are brought into the deep network again, and repeated training and verification are performed for 750 rounds. After the loss function and the average accuracy converge, the training is stopped.

[0080] (7) The trained weight file is extracted and deployed on the edge, and the model can automatically establish a more accurate laser point cloud model and calibrate it.

[0081] Embodiment 2 of the present invention provides a point cloud precision modeling system based on laser point cloud and deep learning algorithm, and the system is a system used by the aforementioned point cloud precision modeling method based on laser point cloud and deep learning algorithm, comprising:

[0082] Feature extraction module, used to extract global features of laser point cloud;

[0083] CSF algorithm module, used to further process laser point cloud data using CSF algorithm;

[0084] The relative elevation extraction module is used to extract the relative elevation of laser point cloud data.

[0085] The beneficial effect of the present invention is that, compared with the prior art, the most popular direction of point cloud processing is to apply machine learning models to point cloud data classification. The most commonly used laser point cloud classification methods now include random forests, support vector machines, cluster analysis, deep learning and other methods. There are also many researchers currently committed to the research of laser point cloud classification. In the current research process of point cloud classification, machine learning-based methods often need to calculate a large number of point cloud features in order to obtain higher accuracy, which will increase the complexity of data processing and cause problems in feature selection.

[0086] Unlike the current algorithms, the algorithm of the present invention can well avoid the extraction of a large number of point cloud features, and can classify a variety of land object types. In order to solve the problems existing in the current point cloud classification process, the present invention proposes a Point-Net deep learning classification algorithm based on PCA dimensionality reduction. First, in order to remove the influence of noise points on the classification results, the point cloud data should be denoised; secondly, in order to avoid the problem of insufficient feature information in point cloud classification, the relative elevation features of the point cloud are extracted based on the elevation information of the point cloud; then, PCA data dimensionality reduction is performed on the multi-dimensional data of the laser point cloud to remove redundant data; finally, the point cloud data is input into the Point-Net network, the classification model is trained, and the experimental data is classified to compare the classification accuracy of different classification methods.

[0087] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0088] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0089] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0090] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A point cloud accurate modeling method based on laser point cloud and deep learning algorithm, characterized in that: include: The input transformation matrix and maximum pooling layer are introduced into the Point-Net neural network to extract the global features of the laser point cloud; The laser point cloud data is further processed using the CSF algorithm; Extract relative elevation from laser point cloud data; Build a laser point cloud model to train and test images.

2. The point cloud precise modeling method based on laser point cloud and deep learning algorithm according to claim 1, characterized in that: The further processing of the laser point cloud data using the CSF algorithm includes: The point cloud is inverted, and all laser points are projected onto the horizontal plane. The points on the plane are used as grid points. The height value of the laser point closest to the simulated point is obtained as IHV based on the KD tree nearest neighbor index, and all DTM digital ground model particles are traversed; the CSF algorithm is compiled using the programming language according to PCL, and the internal force between the simulated cloth particles is calculated. According to the rigidity parameters of the cloth, the relative position between the simulated cloth particles is determined; iterative calculations are performed, and the iteration is stopped after the preset conditions are met; the distance between the laser point cloud and the corresponding cloth particle is calculated, and when the distance is less than or equal to the maximum distance threshold, it is marked as a ground point, and when the distance is greater than the maximum distance threshold, it is marked as a non-ground point, and the maximum distance threshold = median + standard deviation multiple × standard deviation.

3. The point cloud precise modeling method based on laser point cloud and deep learning algorithm according to claim 2, characterized in that: The inverting process of the point cloud comprises: Using the CSF simulation filtering algorithm, the terrain is flipped and the resulting point cloud data is converted from the DSM digital surface model to the DTM digital terrain model.

4. The point cloud precise modeling method based on laser point cloud and deep learning algorithm according to claim 3, characterized in that: The traversal of all DTM digital terrain model particles includes: When the particle is affected by gravity, if the particle's height value CHV is less than IHV, the particle's height value is set to IHV, and the current cloth particle is set to an immovable particle.

5. The point cloud precise modeling method based on laser point cloud and deep learning algorithm according to claim 4, characterized in that: The relative elevation of the laser point cloud data is extracted as follows: Get the ground points and non-ground points, build a KD tree for the ground points, search for k nearest neighbor points in the ground points for all point clouds, calculate the average of the k nearest neighbor points, and traverse to calculate the relative elevation of all point clouds. The formula is: H 相对高程 =H-H 均值 。 6. The point cloud precise modeling method based on laser point cloud and deep learning algorithm according to claim 5, characterized in that: The construction of the laser point cloud model and the training and testing of the image include: Obtain point cloud images containing transmission lines and transmission towers, perform manual or semi-automatic manual annotation through data annotation software to form an original data set, and randomly divide the data set into a training set and a test set in a ratio of 6:1; extract features of the images in the training set layer by layer, calculate the loss function of the anchor box and the annotated detection box formed by clustering, and perform error backpropagation to correct the network parameters.

7. The point cloud precise modeling method based on laser point cloud and deep learning algorithm according to claim 6, characterized in that: The method of obtaining a loss function for the anchor frame formed by clustering and the labeled detection frame includes: Among them, L IOU represents the overlap loss, L dis represents the center distance loss, L asp represents the width and height loss, w, h are the width and height of the prediction box, b is the image center distance, c is the aspect ratio of the image, C w and C h It is the width and height of the minimum bounding box covering the two Boxes, IOU = |A∩B| / |A∪B|, which refers to the intersection-over-union ratio of images A and B. g is the iteration step size, t is the number of iterations, and ρ represents the Euclidean distance between the two center points.

8. A point cloud precision modeling system based on laser point cloud and deep learning algorithm, the system being a system used in a point cloud precision modeling method based on laser point cloud and deep learning algorithm as claimed in any one of claims 1 to 7, comprising: Feature extraction module, used to extract global features of laser point cloud; CSF algorithm module, used to further process laser point cloud data using CSF algorithm; The relative elevation extraction module is used to extract the relative elevation of laser point cloud data.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images