A refined modeling method for underground structures based on laser radar scanning

Abstract

The present invention provides a method for fine modeling of underground structures based on laser radar scanning, including: S1, collecting point cloud data and feature point data of underground structures; S2, preprocessing the data; S3, evenly selecting a number of feature check points in the feature point data to complete the precision analysis of the point cloud data; S4, voxelizing the three-dimensional point cloud and constructing multi-scale geometric features; S5, if the underground structure is an underground cave scene, removing internal dynamic objects; axial orientation and completing cross-section extraction and smoothing; completing triangulation and model generation; S6, if the underground structure is an underground artificial scene, extracting building faces; removing internal dynamic objects and face screening; completing triangulation and model generation. The present invention can effectively realize the accurate reconstruction of underground structure models, and provides a technical basis for solving the problems of low automation and complex procedures in three-dimensional reconstruction of underground scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional modeling, and in particular relates to a method for fine-grained modeling of underground structures based on laser radar scanning. Background Art

[0002] Urban underground space information is an important basic data that is indispensable for urban underground space planning, construction and management, among which underground structures are an important part of urban underground space planning and construction. Due to the limitations of computer software level and data collection methods, underground structures and their ancillary facilities were mostly used as basic data for management in the form of two-dimensional graphics in the past. However, this basic data has certain defects. For example, it can only express the position and relative relationship of underground structures on a two-dimensional plane, and cannot directly display their internal space and vertical morphology and changes.

[0003] Compared with the two-dimensional expression, the three-dimensional expression can better express the position and relative relationship of spatial objects for the internal structure of underground space. However, since the main structure of underground buildings is located underground, there are unfavorable conditions such as poor visibility and insufficient lighting. It is easy to miss or mismeasure during surveying and mapping, and it is difficult to express it through three-dimensional data. In particular, different types of underground buildings, including underground cave scenes (such as air-raid shelters and tunnels) and underground artificial scenes (such as underground parking lots), have different characteristics. The cluttered and irregular underground cave scenes make three-dimensional modeling difficult. The three-dimensional modeling of underground artificial scenes has problems such as large data volume, poor calculation efficiency, and the quality of multi-image modeling of dynamic objects (non-building components such as cars, pedestrians, and internal objects). In summary, there is a lack of methods in the prior art to achieve refined three-dimensional model reconstruction for underground buildings. Summary of the invention

[0004] The purpose of the present invention is to provide a method for fine-grained modeling of underground structures based on lidar scanning, which can effectively realize the accurate reconstruction of underground structure models and provide a technical basis for solving the problems of low automation and complex procedures in three-dimensional reconstruction of underground scenes.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] A method for fine-grained modeling of underground structures based on laser radar scanning, comprising:

[0007] S1. Use a laser radar scanner to collect point cloud data of underground structures, and use a total station to collect feature point data of underground structures;

[0008] S2, preprocessing the point cloud data to obtain preprocessed point cloud data;

[0009] S3, evenly selecting a number of feature check points from the feature point data collected by the total station, extracting the coordinate values ​​of the feature check points from the preprocessed point cloud data, and comparing them with the coordinate values ​​of the feature check points in the feature point data, calculating the relative accuracy of the point cloud, and completing the accuracy analysis of the point cloud data;

[0010] S4, organizing the preprocessed point cloud data through precision analysis into three-dimensional point cloud voxels, and constructing multi-scale geometric features for the point cloud data;

[0011] S5. If the underground structure is an underground cavern scene, the steps of fine reconstruction of the three-dimensional model include: removing dynamic objects inside the underground cavern according to the geometric features; orienting the axis of the underground cavern based on the L1 median algorithm; extracting and smoothing the cross section of the underground cavern; and constructing a triangulated network and generating a model.

[0012] S6. If the underground structure is an underground artificial scene, the steps of fine reconstruction of the three-dimensional model include: extracting building faces according to the geometric features; removing internal dynamic objects and screening faces in combination with prior knowledge; and completing triangulation network construction and model generation.

[0013] Furthermore, the pre-processing method in step S2 includes:

[0014] S201, importing point cloud data: importing point cloud data through point cloud processing software corresponding to the laser radar scanner;

[0015] S202, point cloud data check: check the layering and trajectory of the point cloud data to determine whether the data is normal;

[0016] S203, point cloud data stitching: completing point cloud data registration and point cloud data stitching according to different laser radar scanners;

[0017] S204, point cloud data optimization: point cloud noise reduction is performed by statistical filtering method; point cloud smoothing is performed based on moving least squares method;

[0018] S205, point cloud data simplification: plane fitting is performed using the RANSAC algorithm, and contour feature information is extracted by the intersection of the fitted planes; then the marked area is extracted by determining the color information; finally, the point cloud data of the non-contour and marked areas is simplified using the bounding box method.

[0019] Furthermore, in step S3, the relative accuracy calculation formula of the point cloud is as follows:

[0020] ;in, Indicates accuracy, represents the three-dimensional coordinates of the i-th feature check point in the point cloud, It represents the three-dimensional coordinates of the i-th feature check point measured by the total station, and n is the number of feature check points.

[0021] Furthermore, in step S4, the method for voxelizing the three-dimensional point cloud includes:

[0022] S401, voxelizing the preprocessed point cloud data, including: assuming that a voxel is a cube with a predefined side length, obtaining a lower left front point and an upper right rear point from a bounding box in the input point cloud data, dividing the bounding box of the point cloud into cubic units in the xyz three-axis direction according to the predefined voxel side length, and then calculating the number of cubic units in the xyz three-axis direction, comparing it with the coordinates of the lower left front point of the bounding box in the xyz three-axis direction, and obtaining a three-dimensional voxel index of each point;

[0023] S402 , performing voxel resampling and clustering processing on the voxelized point cloud data based on the octree structure.

[0024] Furthermore, the method for constructing multi-scale geometric features in step S4 includes:

[0025] S411, determining the dimension of each voxel cluster by using a PCA principal component analysis algorithm; making all voxel units in the voxel cluster have a characteristic value representing its dimension;

[0026] S412, calculating geometric features based on eigenvalues ​​to obtain geometric information of the preprocessed point cloud data;

[0027] S413. Geometric feature dimensions at different scales are obtained by setting different voxel sizes and octree scales.

[0028] Furthermore, the method for removing dynamic objects inside the underground cavern described in step S5 includes:

[0029] S501, analyzing the geometric feature dimensions of the preprocessed point cloud data at different scales to remove objects with obvious point features;

[0030] S502, using a progressive triangulation filtering algorithm to remove ground points of the underground cavern, so that the dynamic objects are independent of the remaining point cloud;

[0031] S503, using straight-through filtering combined with Euclidean clustering to determine dynamic objects and remove them.

[0032] Furthermore, the method for performing underground cavern axis orientation based on the L1 median algorithm in step S5 includes:

[0033] S511, extracting the point cloud skeleton based on the L1 median algorithm, and determining the one-dimensional local center of the point cloud after the dynamic objects inside the underground cavern are removed;

[0034] S512, iteratively clustering the center point of the one-dimensional local center, and introducing the PCA algorithm to judge the line features; when the distance between two cluster centers is less than the set threshold, it is considered that the local cluster center is obtained;

[0035] S513, reducing the cluster radius for reprocessing to obtain the pipeline axis.

[0036] Furthermore, the method for extracting and smoothing the underground cavern section in step S5 includes:

[0037] S521, performing step segmentation along the pipeline axis to obtain step local points; constructing two-dimensional voxels to sample the step local points, then traversing all sampling points, constructing plane normals through points before and after each sampling point, treating the sampling point as a point on the plane, and obtaining the plane equation where the sampling point is located; calculating the distance from the sampling point to the plane according to the plane equation to make a judgment, setting a reserved distance value, and retaining the sampling points whose distance is less than the reserved distance value;

[0038] S522, converting the retained sampling points into two-dimensional voxels to obtain two-dimensional coordinates of each point;

[0039] S523, continue to voxelize the points of the two-dimensional coordinates, calculate the average value of all points in each voxel, and use all the average values ​​as the cross-section points;

[0040] S524. Use interpolation method to complete the holes.

[0041] Furthermore, the method for extracting building faces in step S6 includes:

[0042] S601, decomposing the point cloud data of the underground artificial scene into a concise hybrid structure consisting of an inherent shape and a group of residual points based on an Efficient RANSAC segmentation algorithm;

[0043] S602, sampling the point cloud using a local sampling strategy, and using an octree to establish spatial proximity between samples;

[0044] S603, extract the first sample without restriction from all points in the point cloud, and then randomly select a unit C from any level of the octree, the first sample is included in the unit C, and other samples are only extracted from the unit C.

[0045] Furthermore, the method of removing internal dynamic objects and selecting facets in combination with prior knowledge in step S6 includes:

[0046] S611, according to the geometric features, screen out the faces with the largest and smallest Z values ​​in the Z-axis direction from the extracted building faces, and screen out the faces based on prior knowledge;

[0047] S612, slicing the remaining point clouds vertically along the horizontal direction to obtain vertical slices without top or bottom;

[0048] S613, constructing two-dimensional voxels for each vertical slice and performing cluster analysis;

[0049] S614: for each cluster of the vertical slice, based on the size of the outer bounding box and the topological relationship between the clusters, the clusters inside the outer bounding box are defined as internal dynamic objects and removed.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The present invention proposes a method for performing accuracy analysis before modeling. By comparing with the measurement data of the total station, the accuracy of the constructed three-dimensional model of underground buildings is analyzed to ensure that the accuracy requirements of underground space building surveying and mapping are met;

[0052] 2. Aiming at the problem that underground cave scenes are difficult to model due to their clutter and irregularity, the present invention proposes an automated modeling method based on axis orientation and iterative step smoothing. First, the original cluttered point cloud is organized and resampled through voxelization, octree data organization, and multi-scale geometric feature expression of complex spatial scenes, and a preliminary understanding of the point cloud attributes is obtained at the same time; secondly, the original three-dimensional point cloud is automatically generated into a Mesh three-dimensional model through the steps of internal dynamic object removal, axis orientation based on L1 median, section extraction and smoothing, triangulation construction, and Mesh structure generation, which can effectively solve the problem of inaccurate modeling caused by the irregular and cluttered underground cave scenes, and realize the full process automatic modeling from point cloud to Mesh structure generation;

[0053] 3. Aiming at the demand of 3D modeling of underground artificial scenes, the present invention first organizes and resamples the original messy point cloud through point cloud data organization, similar to cave scenes, to obtain a preliminary understanding of point cloud attributes; secondly, in view of the large amount of data, poor computational efficiency, and the quality of multi-image modeling of dynamic objects (cars, pedestrians, internal objects and other non-building components) in 3D modeling of artificial scenes, the geometric rules conform to the characteristics of the Manhattan world, and proposes building face extraction based on Efficient RANSAC to achieve rough face segmentation of buildings. Then, the outer frame and topological logic are used to cluster and remove the dynamic objects inside the building and screen the plane pieces to achieve effective face extraction of the building; finally, the effective face is used to construct a triangulated network to generate a model. In this process, the face extraction and internal dynamic object removal are optimized, providing a new solution for the construction of 3D models of underground artificial structures.

[0054] 4. The present invention can effectively realize the accurate reconstruction of the model, and provides a technical basis for solving the current problems of low automation level and complex process of three-dimensional reconstruction of underground scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the process of the present invention.

[0056] Figure 2 It is a display of the result of constructing the octree of the air-raid shelter in Example 1 of the present invention.

[0057] Figure 3 It is a demonstration of the removal of dynamic objects inside the air-raid shelter according to Example 1 of the present invention;

[0058] in Figure 3 (a) shows the original point cloud (colored by elevation), (b) shows the progressive triangulation with ground points removed; (c) shows the direct filtering + Euclidean clustering with dynamic object points removed (blue represents air-raid shelter points, and the other colors represent dynamic object clusters).

[0059] Figure 4 This is the result of the air-raid shelter axis orientation in Example 1 of the present invention (red indicates the center point and direction).

[0060] Figure 5 It is the air-raid shelter cross-section extraction and smoothing result display of Example 1 of the present invention;

[0061] in Figure 5 (a) is the original point cloud (top view), (b) is the extracted point cloud (top view), (c) is the original point cloud (free view), and (d) is the extracted point cloud (free view).

[0062] Figure 6 The three-dimensional reconstruction model of the air-raid shelter of Example 1 of the present invention is displayed;

[0063] in Figure 6 (a) is the air-raid shelter triangulated network model, (b) is the air-raid shelter mesh model display #1, (c) is the air-raid shelter mesh model display #2, (d) the interior display of the air-raid shelter mesh model, (e) is the air-raid shelter triangulated network model (detail display), and (f) is the air-raid shelter mesh model (detail display).

[0064] Figure 7 This is the underground parking lot patch detection and extraction result of Example 2 of the present invention.

[0065] Figure 8 This is a display of the dynamic object removal and facet screening results of the underground parking lot according to Example 2 of the present invention.

[0066] Fig. 9 It is a display of the reconstruction result of the three-dimensional model of the underground parking lot in Example 2 of the present invention;

[0067] in: Fig. 9 (a) is the point cloud after the dynamic objects inside the underground parking lot are removed and the facets are screened (top view), (b) is the triangulated network model of the underground parking lot (top view), and (c) is the Mesh model of the underground parking lot (top view). DETAILED DESCRIPTION

[0068] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0069] The method for fine-grained modeling of underground structures based on laser radar scanning proposed in the present invention is as follows: Figure 1 As shown, specifically including:

[0070] S1. Use a laser radar scanner to collect point cloud data of underground structures, and use a total station to collect feature point data of underground structures;

[0071] S2, preprocessing the point cloud data to obtain preprocessed point cloud data;

[0072] S3, evenly selecting a number of feature check points from the feature point data collected by the total station, extracting the coordinate values ​​of the feature check points from the preprocessed point cloud data, and comparing them with the coordinate values ​​of the feature check points in the feature point data, calculating the relative accuracy of the point cloud, and completing the accuracy analysis of the point cloud data;

[0073] S4, organizing the preprocessed point cloud data through precision analysis into three-dimensional point cloud voxels, and constructing multi-scale geometric features for the point cloud data;

[0074] S5. If the underground structure is an underground cavern scene, the steps of fine reconstruction of the three-dimensional model include: removing dynamic objects inside the underground cavern according to the geometric features; orienting the axis of the underground cavern based on the L1 median algorithm; extracting and smoothing the cross section of the underground cavern; and constructing a triangulated network and generating a model.

[0075] S6. If the underground structure is an underground artificial scene, the steps of fine reconstruction of the three-dimensional model include: extracting building faces according to the geometric features; removing internal dynamic objects and screening faces in combination with prior knowledge; and completing triangulation network construction and model generation.

[0076] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0077] Embodiment 1:

[0078] In this embodiment 1, the underground structure to be modeled is an air-raid shelter, which belongs to an underground cave scene. The specific method of this embodiment includes:

[0079] 1. Use the ground-based lidar scanner TLS to collect point cloud data of the air-raid shelter: Use the total station to collect feature point data of the air-raid shelter.

[0080] 1. When collecting data using the ground-based LiDAR scanner TLS, a total of 10 stations are deployed, with an average interval of 5m between each station to ensure the overlap rate of point cloud data. Before scanning, a comprehensive inspection of the equipment is conducted, including checking whether the equipment accessories are complete, and checking the battery power and power-on status of the equipment; confirming that data can be obtained normally, and that the power capacity and memory capacity meet the operating time requirements.

[0081] After the inspection is completed, data collection begins:

[0082] (1) Scanning station setting:

[0083] According to the spatial environmental conditions of the air-raid shelter and the collection requirements, choose an area with a wide field of view and stable and safe ground to set up the scanning station.

[0084] (2) Target placement:

[0085] Target control points are used for point cloud stitching, registration or verification of point cloud data accuracy by making targets and measuring target coordinates.

[0086] (3) Point cloud collection:

[0087] After setting up the instrument, insert the battery, open the storage disk protective cover, insert the storage disk, then turn on the power switch, set the instrument parameters, press the start button, the scanner begins scanning and rotates by itself, and after rotating and scanning for one circle, the scanning stops, indicating that the work at the current station is completed. After turning off the instrument, move to the next station position for collection and scanning.

[0088] (4) Data export:

[0089] When all underground structures have been scanned, turn off the instrument, take out the storage disk and insert it into the computer, and export the scanning point cloud data to the computer.

[0090] 2. When using a total station to collect data on air-raid shelter feature points, select appropriate locations to lay out guide points through surveying the air-raid shelter scene.

[0091] (1) Preparation: Make sure the total station has sufficient power and check whether its functions are normal; select a suitable tripod and install the total station on it to ensure its stability.

[0092] (2) Setting up the total station: Set up the total station on a guide point at the entrance and exit of the underground parking lot. Use a leveler to adjust the level of the instrument until the bubble is in the center. Use a focuser to adjust the instrument's target mirror so that it is clearly visible.

[0093] (3) Set up the backsight point and forward point: Use a tripod and a reflecting prism as the guide backsight point and forward point; use a steel ruler to measure the height of the instrument and record it.

[0094] (4) Setting instrument parameters: The total station traverse data acquisition adopts a free coordinate system, with the traverse point at the entrance and exit of the air-raid shelter as the starting point for data acquisition;

[0095] (5) Traverse measurement: Aim the total station at the backsight point and the forward point respectively. Perform two rounds of measurement at each station, read the azimuth and distance. The total station automatically calculates the station coordinates and the forward point coordinates. After the measurement of each station is completed, move forward in sequence.

[0096] (6) Data processing: The total station converts the measured distance and azimuth into plane coordinate data in a free coordinate system.

[0097] (7) Feature point measurement: Set up the total station at the surveyed traverse point, and set up the tripod and reflecting prism at the adjacent traverse point. Use the reflecting prism as the backsight point and use the prism-free method to measure the feature points inside the air-raid shelter as verification points for model accuracy.

[0098] 2. Point cloud data preprocessing:

[0099] Due to factors such as the environment, equipment, and human operation, the raw data collected by 3D laser scanners usually contains a lot of noise, and there are problems such as data redundancy and missing spatial relationships, especially in outdoor scenes. Therefore, necessary preprocessing of the raw data can effectively guarantee subsequent applications and improve work efficiency. In this embodiment, TLS point cloud data preprocessing uses Trimble RealWorks software, as follows:

[0100] 1. Import point cloud data:

[0101] Open Trimble RealWorks and import the raw data collected by the ground-based LiDAR scanner TLS; generate a point cloud view;

[0102] 2. Point cloud data inspection:

[0103] (1) Check to make sure that the point cloud is not layered (layering: If an object is stationary during the entire scanning process, then if the same point cloud "shadow" appears on its edge, it may be layered. Or if the point cloud surface of a certain side is rough and blurred, it may also be layered.).

[0104] (2) Check the walking trajectory to ensure that the trajectory line does not "pass through the wall" or float in mid-air.

[0105] 3. Point cloud data stitching:

[0106] Point cloud data stitching is when a 3D laser scanner collects point cloud data of a target object. Due to reasons such as the large scanning area, the complex structure of the target object, the occlusion between the target objects and the limited field of view of the instrument itself, it is impossible to collect complete point cloud data of the target object by setting up a 3D laser scanner at a single station, and multi-site cloud data scanning is required.

[0107] Point cloud data stitching requires registration, which is similar to mathematical coordinate transformation. Through the corresponding points between the cloud data of different measuring stations, the transformation parameters between the two point cloud data are calculated. According to the transformation parameters, the point cloud data of one station is converted to the coordinate system of the cloud data of another station. Point cloud data registration is the process of determining the corresponding relationship and calculating the transformation parameters (rotation matrix and translation matrix).

[0108] Point cloud data stitching methods are mainly divided into feature-based point cloud data stitching and featureless point cloud data stitching. Feature-based point cloud data stitching refers to point cloud data stitching based on the feature parameters of point cloud data, mainly including target-based stitching, point cloud-based stitching, hybrid stitching, etc. Featureless point cloud data stitching refers to iterative calculation of point cloud data, searching for corresponding point pairs and solving transformation parameters to achieve point cloud data stitching. Among them, the ICP algorithm, as a classic point cloud iterative algorithm, is widely used in featureless point cloud data stitching.

[0109] (1) Target-based splicing;

[0110] The most widely used stitching method is to stitch point cloud data based on targets, which has a relatively good effect. The condition for target-based stitching is that three or more public targets are arranged in the public area between the scanning site sites. Targets are arranged at each station while performing 3D laser scanning. When laying out targets, they cannot be arranged on the same straight line. The coordinates of the target points are extracted from the collected point cloud data, and the coordinates of the public target points are used to complete the stitching of point cloud data from different measuring stations.

[0111] (2) Point cloud-based stitching;

[0112] Point cloud-based stitching is to use point cloud views to stitch point cloud data, and it is to use the common areas between different sites to stitch point cloud data. Virtually align the point cloud views of the common areas of the two survey stations, ensure that one site cloud is used as the basis and does not move, and through translation and rotation operations, make the common areas of the cloud data of the two sites overlap to the same position, thereby completing the stitching of the cloud data of the two sites. When the scanning area environment is complex, point cloud-based view stitching can achieve better results. For example, taking the cloud data of two survey sites as an example, point cloud view-based stitching is performed in Trimble RealWorks software.

[0113] Import the cloud data of the two measuring stations into the newly created point cloud project, set the point cloud data of one of the measuring stations as the basic point cloud data, enter the virtual alignment operation interface through the manual alignment tool, and find the common points of the cloud data of the two measuring stations, such as building corners and turning points, from the top view, and translate them to the same position by translation. Then make them overlap by rotating them, and judge whether each position completely overlaps by continuously zooming in and out the public point cloud view. Then, from the side view, translate up and down to make them overlap. Finally, make the common areas of the cloud data of the two measuring stations completely overlap in space, and complete the splicing of the cloud data of the two measuring stations.

[0114] (3) Mixed splicing

[0115] In the splicing of some point cloud data, there are situations where the scanning area environment is complex, there are many influencing factors, or multi-site cloud data splicing. Using one point cloud data splicing method cannot complete the point cloud data splicing, and two or more splicing methods need to be used to cooperate with each other to complete the point cloud data splicing. Hybrid splicing is mostly used for multi-site cloud data splicing and point cloud data collection quality is not ideal.

[0116] In this embodiment, Trimble RealWorks software is used, and a hybrid stitching method is adopted, mainly for automatic stitching of point cloud data. When an error occurs in the automatic stitching or the point cloud data cannot be stitched, target stitching and point cloud view stitching are used.

[0117] 4. Point cloud data optimization:

[0118] (1) Statistical filtering method for point cloud denoising:

[0119] Affected by factors such as scanner accuracy, environment, and the object being measured, the original point cloud data inevitably contains noise points and outliers far away from the target point cloud, which will seriously affect the accuracy of the local point cloud and thus affect the subsequent point cloud processing results. Considering that statistical filtering has a good effect on removing noise points and outliers, this embodiment selects a statistical filtering algorithm for point cloud denoising.

[0120] (2) Point cloud smoothing:

[0121] In this embodiment, the point cloud data is smoothed by the moving least squares (MLS) method. Statistical filtering is mainly used to remove outlier noise points. Usually, after denoising, there are still some small noise points or uneven point cloud density. Therefore, in this embodiment, the MLS method is used to smooth the point cloud after denoising to remove small irregular noise.

[0122] Moving least squares (MLS) is a method that approximates a curve or surface by establishing a local high-order polynomial. The basic idea is to fit a smooth surface based on local sampling points, move outlier noise points to the surface, and eliminate the noise.

[0123] In this embodiment, the roughness of the point cloud before and after smoothing is calculated with a radius of 0.03 m: the roughness of the laser point cloud is reduced from 0.0055 to 0.0014.

[0124] 5. Point cloud data simplification:

[0125] Point cloud data simplification, also known as point cloud data compression, is a method of collecting point cloud data of target objects using ground-based 3D laser scanners. The number of scanning stations is usually increased in order to obtain comprehensive information about the target object in actual collection, which makes the density of massive point cloud data even greater. Most of these point cloud data have little effect on model reconstruction and data collection requirements. If the massive point cloud data are stored, calculated, and processed, it will bring great difficulties to the calculation and processing, seriously affecting the efficiency of point cloud data calculation and processing. Massive point cloud data will also take up a lot of storage space, which will cause a huge burden on the storage of point cloud data. Therefore, it is necessary to simplify and compress the point cloud data. Under the premise of ensuring the structure of the target object, the point cloud data should be simplified and compressed as much as possible, and more point cloud data should be retained in the characteristic structure and important areas, and less point cloud data should be retained in the areas with simple structure and low importance, so as to improve the efficiency of point cloud data processing and modeling.

[0126] Different point cloud data reduction algorithms are constructed for different target objects and requirements, but no matter which point cloud data reduction method is used, the measurement criteria are: speed, simplicity, and accuracy.

[0127] (1) Accuracy: This refers to the accuracy of point cloud data simplification. The error between the point cloud data model after point cloud data simplification and compression and the original point cloud data model should be reduced as much as possible within the tolerance range, and sufficient point cloud data to express the characteristics of the target object and what we need should be retained. Low accuracy of point cloud data simplification will lead to difficulties in point cloud data modeling and incomplete expression of the model's structural information.

[0128] (2) Simplification degree: The ratio of the amount of point cloud data before and after simplification is called the simplification degree. When simplifying point cloud data, if the amount of thinned point cloud data is too small, the point cloud data simplification effect cannot be achieved. In this case, the point cloud data should be simplified to the greatest extent possible while ensuring that the features of the point cloud data are complete and accurate.

[0129] (3) Speed: While meeting the requirements of accuracy and simplicity, the speed of point cloud data reduction should be improved. The point cloud data reduction algorithm should be as simple as possible to improve the computational efficiency of the point cloud data reduction algorithm, thereby increasing the practicality of the point cloud data reduction algorithm.

[0130] It is difficult for the point cloud data simplification algorithm to meet the three criteria at the same time. When actually performing point cloud data simplification processing, it is necessary to select the criteria that are emphasized according to different needs, so as to select the point cloud data simplification algorithm, or improve the algorithm according to the needs, so that the point cloud data simplification can achieve the expected effect. The point cloud data of underground structures are mostly composed of regular rectangular structures, and the surface structure is relatively flat and simple. It is suitable for uniform point cloud data simplification algorithms such as the bounding box method to simplify point cloud data. In the three-dimensional data information collection, processing and model reconstruction of underground structures, indicator signs are important data information. Therefore, in the point cloud data simplification, it is necessary to retain a large amount of point cloud data at the indicator signs, and simplify the point cloud data of other non-indicator areas. When the traditional bounding box method simplifies the point cloud data, it will also simplify the number of point clouds in the indicator area, and it is impossible to retain a large amount of point cloud data. Therefore, this embodiment adopts the bounding box method based on RGB and contour preservation to retain the indicator area and contour structure, and uses the bounding box method to uniformly simplify the other point cloud data to achieve the purpose of simplifying the point cloud data.

[0131] Unlike feature information such as contours and boundaries, signposts cannot be extracted and identified through curvature and normal. However, due to the need for signposts to be eye-catching and clear, signposts and surrounding areas are usually markedly different in color to achieve the requirement of eye-catching and clear signposts. Therefore, based on the obvious color difference between the two areas, the signpost area can be identified based on the RGB attribute, achieving the purpose of retaining the point cloud data of the signpost area.

[0132] The basic idea of ​​the point cloud data simplification algorithm in this embodiment is to first use the RANSAC algorithm to perform plane fitting to estimate the plane of the air-raid shelter, and extract the contour feature information through plane intersection. Then, the area where the indicator mark is extracted is determined by color information (RGB). Finally, the bounding box method is used to simplify the point cloud data of the non-contour and indicator mark areas.

[0133] The point cloud data reduction algorithm can well retain a large amount of point cloud data in the sign area and reduce the non-sign area.

[0134] 3. Accuracy analysis:

[0135] In order to verify the point cloud accuracy of the ground-based LiDAR scanner TLS, 7 feature check points were evenly selected inside the air-raid shelter measurement area. The coordinates of these 7 feature check points were obtained using the prism-free method during total station data collection, and then the coordinate values ​​of these 7 feature check points were extracted from the point cloud data and compared with the data obtained by the total station. The comparison between the total station coordinates and the model-extracted coordinates is shown in the table below.

[0136]

[0137] As can be seen from the table, the difference between the coordinates of the feature check points collected by the total station and the coordinates extracted by the TLS point cloud is within 2 cm, and the maximum difference is 12 mm. According to the comparison results, the data collected by the ground-based lidar scanner TLS can correctly display the distribution pattern inside the air-raid shelter, and the relatively obvious buildings in the air-raid shelter can be accurately distinguished, and the details of the ground objects can be displayed, meeting the basic needs of surveying and mapping.

[0138] Fourth, organize the pre-processed point cloud data through precision analysis into three-dimensional point cloud voxels and construct multi-scale geometric features for it;

[0139] 1. 3D point cloud voxelization:

[0140] The collected pre-processed point cloud data of the air-raid shelter is voxelized and organized. A voxel is a cube with a predefined side length in a three-dimensional Euclidean space. The lower left front point and the upper right back point are obtained from the bounding box in the input point cloud data. The bounding box of the point cloud is divided into cubic units according to the predefined voxel side length in the xyz three-axis direction. Then the number of cubic units in the xyz three-axis direction is calculated respectively, and compared with the coordinates of the lower left front point of the bounding box in the xyz three-axis direction to obtain the three-dimensional voxel index of each point.

[0141] The above operations effectively reduce the amount of point cloud data and organize the point cloud data. At the same time, the point cloud data is initially smoothed and sampled consistently, reducing the uneven density of the point cloud caused by repeated scanning.

[0142] 2. Data organization based on octree:

[0143] Although neighborhood search in voxel space is simple and efficient, there is a lot of memory redundancy when using voxel resampling. Therefore, when processing massive point clouds, the octree structure is preferred.

[0144] The octree-based three-dimensional point cloud organization can improve the efficiency of point cloud indexing, and the voxelized point cloud data can be resampled and clustered based on the octree structure. Figure 2The following figure shows the result of constructing the octree for the air-raid shelter and the result at different octree scales. It can be seen that the neighborhood space corresponding to the octree is different at different scales, which helps to calculate the spatial dimension features at different scales.

[0145] 3. Multi-scale geometric feature expression of complex spatial scenes:

[0146] The dimension of each voxel cluster is determined by the PCA (Principal Component Analysis) algorithm.

[0147] Voxel units are divided into three types: linear, planar, and scattered. These three types are defined as follows: 1) If the eigenvalue of a voxel ,Keep , then the voxel is defined as a linear or one-dimensional voxel unit. For a linear voxel unit, the corresponding eigenvalue The eigenvector of is a valid eigenvector, indicating the effective direction of the point in the voxel unit. 2) If , then the voxel unit is defined as a planar or two-dimensional unit. In this case, the eigenvector is a valid eigenvector. 3) If , then this voxel unit is defined as a scattered or three-dimensional unit. Here, ">>" means much greater than, which is defined by a preset threshold.

[0148] All voxel cells in a cluster have a geometric tag representing their dimensions and, when used, a valid eigenvalue and eigenvector representing their characteristics.

[0149] The eigenvalue-based features can well realize the acquisition of geometric feature attributes of voxel clusters / point clusters based on octree and voxel organization. By setting different voxel sizes / octree scales, geometric feature dimensions at different scales can be obtained. By constructing multi-scale features, it is helpful to improve the scene understanding ability of 3D point clouds and subsequent processing applications.

[0150] At the same time, through the construction of the above geometric features, it can be seen that in the cave scene, due to the presence of natural objects such as grass, stones, and soil, its surface is irregular. Therefore, the point features (especially the messy point features) in the cave scene are more obvious, and they also need to be given priority in subsequent modeling. If it is an artificial scene building, it will be more regular, with obvious surface features and linear features. Therefore, in the subsequent modeling, it is necessary to consider how to retain the surface features while removing linear features and other point cloud data that affect modeling.

[0151] Five: Air-raid shelter modeling:

[0152] 1. Removal of dynamic objects inside the cave:

[0153] After voxelization and geometric feature extraction, the ground objects with obvious point features are first removed. Then the progressive triangulation filtering algorithm is used to split the cave into components and remove the ground points in the cave. When the parameters are set properly, the algorithm can completely separate the ground in the cave from other parts.

[0154] The progressive triangulation filtering algorithm used is a knowledge-guided triangulation progressive filtering method: first, the point cloud is regularly gridded (the grid interval is the average spacing of the point cloud), and the grid data is segmented object-oriented, and the ground points and non-ground points are roughly classified; secondly, a certain number of seed points are selected from the classified ground objects to construct the initial ground triangulation network, and the remaining ground points are gradually encrypted using a set threshold; finally, the non-ground point data is meshed by strong threshold interpolation, and the points participating in the final meshing are determined as ground points, and the points not participating in the meshing are determined as non-ground points. The filtering results are shown in the figure below. After excluding the ground point cloud, the human figures and gravel inside the cave will exist independently outside the remaining point cloud.

[0155] Next, the remaining point cloud after the ground point cloud is removed. Here, the straight-through filtering combined with Euclidean clustering is used for removal. Straight-through filtering means setting a threshold range on a specified dimension of the point cloud and dividing the data on this dimension into those within the threshold and those not within the threshold. Euclidean clustering is to analyze the Euclidean spatial distance between the neighborhoods of the point cloud and set the points that meet the neighborhood threshold as a candidate cluster.

[0156] First, the outer envelope size of the filtered point cloud is calculated. Then the air-raid shelter is redirected using projection to determine its size in the x, y, and z directions. The purpose of redirection is to translate the original point cloud (outer envelope) in the xoy plane so that it is as parallel to the coordinate axis as possible.

[0157] Then, according to the characteristics of the data, the neighborhood size is set (the cluster radius is set to 0.2 for the air-raid shelter scene, and 0.15 for the tunnel scene), and Euclidean clustering is performed to form candidate clusters. Finally, the straight-through filter is used to filter along the pipeline direction (xy plane), and the intersection with the cluster is calculated. If the cluster is within the straight-through filter threshold, it is determined to be a dynamic object. If it is outside the threshold, it is determined to be an air-raid shelter point.

[0158] The results are as follows Figure 3 As shown, (a) is the original point cloud (colored by elevation), (b) is the progressive triangulation filtering algorithm used to remove ground points, and (c) is the straight-through filtering combined with Euclidean clustering to remove dynamic object points (blue represents air-raid shelter points, and the other colors represent dynamic object clusters).

[0159] 2. Orient the air-raid shelter axis based on the L1 median algorithm:

[0160] After removing the internal dynamic objects, the next step is to use the L1 median algorithm to determine the axial direction of the air-raid shelter. Determining the axial direction of the air-raid shelter can extract the central axis of the air-raid shelter and the cross-sectional information of the air-raid shelter, providing an important data basis for modeling.

[0161] The L1 median algorithm is used to extract the skeleton of the point cloud and determine the local point cloud center. It extends the univariate median to a multivariate setting and represents the unique global center of a given set of points. At the same time, it has good robustness to outliers and noise. The L1-Median skeleton is equivalent to a spatially localized version of the L1-Median with conditional regularization. Assume that the three-dimensional point cloud of the air-raid shelter is , as the input point set, the output point set after L1-Median skeleton extraction algorithm is , the overall expression is:

[0162] ;

[0163] Among them, the first represents the local L1 median in Q, and the second term R(x) is the output point set Local point distribution regularization, I indexes the output point set X, J indexes the input point set Q, weight function is a smooth function that decays rapidly, and the radius h defines the size of the local neighborhood extracted by the L1-Median skeleton.

[0164] The output point set X of the algorithm is a curve skeleton, which represents the one-dimensional local center of the shape of the input point set Q. At the same time, in order to iterate the cluster points, it is also necessary to derive the center point iteration formula. Due to the large changes in the diameter of the air-raid shelter, the PCA algorithm is introduced to judge the line features, thereby reducing the axis point offset caused by the change in diameter. In this embodiment, the threshold is set to 0.00001. When the distance between two cluster centers is less than 0.0001, it is considered that the local cluster center is obtained.

[0165] After the first iteration of the point cloud center position, a fixed search radius may cause the axis to deviate due to the significant differences in the diameter of the air-raid shelter. In thicker pipes, if the search radius is too small, the cluster center may be disturbed by the pipe wall, resulting in the cluster point not accurately reflecting the center of the pipe segment, but the local center. On the other hand, if a larger cluster radius is used, the central cluster point may be affected by the point cloud of another part of the pipe segment at the thin bend, resulting in a shift in the cluster center, and the cluster center may be too sparse, making it difficult to accurately reflect the trend of the air-raid shelter.

[0166] In order to effectively deal with this problem, a two-step strategy is adopted. First, a larger radius is selected for preliminary clustering in order to establish a preliminary cluster center. Afterwards, these preliminary results are reprocessed using more stringent parameters (smaller radius and fewer initial points). Due to the use of a smaller clustering radius, those points that have successfully approached the central axis will be retained. As for those points that have not yet approached the central axis, since there are other preliminary clustering points in their cluster center, they will move closer to the central axis again. Thus, the pipeline axis is finally obtained. The final air-raid shelter axis orientation result is as follows: Figure 4 shown.

[0167] 3. Extraction and smoothing of cavern sections:

[0168] After obtaining the axis, the cross section of the air-raid shelter can be extracted according to the axis. The process includes:

[0169] (1) Divide the air-raid shelter along the axis. First, the air-raid shelter point cloud is step-by-step segmented according to the axis to obtain step-by-step local air-raid shelter points. Here, the step length is set to 0.5m. Then, a two-dimensional voxel is constructed for the step-by-step local air-raid shelter points for sampling, and then all the sampling points are traversed. The plane normal is constructed through the points before and after each sampling point, and the sampling point is regarded as a point on the plane. The equation of the plane where the sampling point is located can be obtained through these two data. Then, the distance from the point to the plane is calculated according to the equation of the plane to make a judgment. In order to improve the calculation speed, a kd tree is constructed with the sampling point as the center, and points within the threshold are retained. The threshold is slightly larger than the radius of the widest part of the cave. Then, the distance from each point to the plane is calculated, and points within 0.25 meters from the plane are retained.

[0170] (2) 2D voxelization. Then project these points onto the plane to obtain the projection point of each point on the cross section. The current projection point still belongs to the three-dimensional coordinate. For the convenience of subsequent processing, the point cloud is reduced in dimension and two unit vectors are defined in space. .

[0171] ;

[0172] The projection of the sampling point in the z-axis direction on the plane is defined as vector, and then define the axis direction of this position in the air-raid shelter as vector, ;

[0173] Once the definition is completed, the coordinates of each point on the plane can be converted to (A, B).

[0174] (3) Voxel smoothing. The two-dimensional coordinate points are voxelized according to their angles with the u-axis. The point cloud is divided into four voxels according to the angle. Then the average of all points in each voxel is calculated and used as the section points.

[0175] (4) Hole completion. Considering factors such as scanning or occlusion, there may be holes inside the air-raid shelter. In order to complete the holes, the interpolation method in statistics is used here to fill the missing voxels. That is, by querying the voxel points before and after the missing point, the angle difference between the missing position and the front and back voxel points is used as the weight, and the position of the front and back voxel points is combined. The interpolation method is used to estimate the unknown point through the known position points. Assume that the voxel at the i-th position is missing, and the point to be supplemented for the voxel is X, then the voxels before and after the position are searched. Find the voxel with an angle less than the i-th position. If there is a point at this position, the coordinate of the point is A, and the serial number of this position is a. Find the voxel with an angle greater than the i-th position. If there is a point at this position, the coordinate of the point is B, and the serial number of this position is b.

[0176] Then use the formula ; It is possible to estimate unknown points through known points. As the number of sampled voxels increases, the angle corresponding to each voxel gradually decreases, so that the missing parts can be supplemented more accurately. This makes the subsequent model construction more realistic.

[0177] Through the above process, we can obtain Figure 5 The cross section of the air-raid shelter and the smoothing result are shown. It can be seen that the point cloud is effectively extracted, and the extracted point cloud is smoother and more orderly, which is helpful for the subsequent reconstruction of the model.

[0178] 4. Triangulation network construction and model generation:

[0179] After tunnel section extraction and smoothing, in order to construct a three-dimensional Mesh model, the two-dimensional point cloud needs to be recalculated into three-dimensional space based on the unit vector. , Vector, trace the section information back to three-dimensional space to form a continuous cross-section point cloud. After obtaining the complete cross-section point cloud, it is necessary to establish the topological relationship between points. Maintain the topological structure of the cross section. When connecting between surfaces, perform Delaunay triangulation.

[0180] There are two topological relationships within the current 3D section point set:

[0181] (1) Inside each section, since the points are voxelized in two dimensions according to the angle, their arrangement follows a certain rule. Arrange the points according to the angle of each section. Next, connect the adjacent section points in a clockwise order, so that the collective shape of each cross section can be constructed. In this way, the tunnel structure can be accurately described.

[0182] (2) Since the same segmentation interval angle is used between each section, the section points of each section can correspond one-to-one with the section points of the next section. The section points are sorted using the sequence of axis sampling points to obtain the front and back information of each section. In this way, the front section and the back section form a one-to-one correspondence. By utilizing this relationship between multiple sections, a topological link between sections can be constructed.

[0183] Through the above process, all the section points and triangulated networks are connected to construct the final 3D model. The results of the triangulated network model and Mesh model of the air-raid shelter scene are as follows: Figure 6 As shown, Figure 6 (a) is the air-raid shelter triangulated network model, (b) is the air-raid shelter mesh model display #1, (c) is the air-raid shelter mesh model display #2, (d) the interior display of the air-raid shelter mesh model, (e) is the air-raid shelter triangulated network model (detail display), and (f) is the air-raid shelter mesh model (detail display).

[0184] Embodiment 2:

[0185] In this embodiment 2, the underground structure to be modeled is an underground parking lot, which belongs to an underground artificial scene. In this embodiment, the specific method includes:

[0186] Step 1: Use the ground-based LiDAR scanner TLS to collect point cloud data of the underground parking lot: Use the total station to collect feature point data of the underground parking lot;

[0187] 1. Point cloud data collection:

[0188] The TLS data collection method for underground parking lots is the same as the data collection method for air-raid shelters, and also follows the requirements of pre-operation preparation, station design, reducing the impact of environmental factors, and ensuring data overlap. First, stations should be arranged at the four corners of the survey area, and then the stations should be reasonably arranged between the corners according to the distance and the required data precision. The scanning distance is about 10m-20m, and the distance between the station and the target needs to be kept between 3 and 8m to facilitate the extraction of target information. In this embodiment, a total of 12 stations are arranged for TLS collection in underground parking lots, with an average interval of 10m between each station to ensure the overlap rate of point cloud data. After checking the equipment, select an area with a wide field of view and a stable and safe ground to set up a scanning station, start scanning and rotate with itself, and after rotating and scanning for one week, the scanning stops and stops, indicating that the current station work is completed. After turning off the instrument, move to the next station position for collection and scanning. When the underground parking lot is completely scanned, turn off the instrument, unplug the storage disk, plug it into the computer, and export the scanned point cloud results data to the computer.

[0189] 2. Total station feature point data collection:

[0190] (1) Preparation: Make sure the total station has sufficient power and check whether its functions are normal; select a suitable tripod and install the total station on it to ensure its stability.

[0191] (2) Setting up the total station: Set up the total station on a guide point at the entrance and exit of the underground parking lot. Use a leveler to adjust the level of the instrument until the bubble is in the center. Use a focuser to adjust the instrument's target mirror so that it is clearly visible.

[0192] (3) Set up the backsight point and forward point: Use a tripod and a reflecting prism as the guide backsight point and forward point; use a steel ruler to measure the height of the instrument and record it.

[0193] (4) Setting instrument parameters: The total station traverse data collection adopts a free coordinate system, with the two traverse points at the entrance and exit of the underground parking lot as the starting points for data collection; the total station automatically measures the instrument height and inputs the rearview reflector prism height into the instrument.

[0194] (5) Traverse measurement: Aim the total station at the backsight point and the forward point respectively. Perform two rounds of measurement at each station, read the azimuth and distance. The total station automatically calculates the station coordinates and the forward point coordinates. After the measurement of each station is completed, move forward in sequence.

[0195] (6) Data processing: The total station converts the measured distance and azimuth into plane coordinate data in a free coordinate system.

[0196] (7) Feature point measurement: Set up the total station at the surveyed traverse point, and set up the tripod and reflective prism at the adjacent traverse point. Use the reflective prism as the backsight point and use the prism-free method to measure the feature points inside the underground parking lot as verification points for model accuracy.

[0197] 2. Point cloud data preprocessing:

[0198] The point cloud data preprocessing method is the same as that in Example 1, including 1. importing point cloud data; 2. checking point cloud data; 3. stitching point cloud data; 4. optimizing point cloud data; 5. streamlining point cloud data; the specific process will not be repeated here.

[0199] 3. Accuracy analysis:

[0200] In order to verify the accuracy of the 3D reconstructed model, 15 feature points were evenly selected as check points within the survey area. The coordinates of the check points were obtained using a prism-free total station during data collection. The coordinate values ​​of the check points were extracted from the TLS point cloud data and compared with the total station data. The comparison between the total station coordinates and the model-extracted coordinates is shown in the table below.

[0201]

[0202] From the above table, we can see that the difference between the coordinates collected by the total station and the coordinates extracted by the TLS point cloud is within 1 cm, and the maximum difference is 8 mm. The TLS point cloud can correctly display the distribution pattern of the underground parking lot and the details of the ground objects, meeting the basic requirements for reconstructing the 3D model of the underground parking lot.

[0203] Fourth, organize the pre-processed point cloud data through precision analysis into three-dimensional point cloud voxels and construct multi-scale geometric features for it;

[0204] This step is the same as the processing method of Example 1, which also involves voxelization of the three-dimensional point cloud: data organization based on an octree; and multi-scale geometric feature expression of complex spatial scenes. The specific process will not be repeated here.

[0205] 5. Underground parking lot modeling:

[0206] 1. Building face extraction based on Efficient RANSAC:

[0207] Efficient RANSAC is an algorithm for detecting basic shapes in unordered point clouds. It decomposes the point cloud into a concise hybrid structure consisting of an intrinsic shape and a set of residual points. Each detected shape represents a set of desired points. The algorithm is based on random sampling and can detect planes, spheres, cylinders, etc. At the same time, the algorithm is based on good robustness and noise resistance.

[0208] The algorithm is defined as: given a point cloud and the corresponding normal , output a set of basic shapes As well as the corresponding point sets and the remaining points.

[0209] Here, the point cloud is sampled using a local sampling strategy, and the smaller the distance between two points, the higher the prior probability that they belong to the same shape. Therefore, an octree is used to establish the spatial proximity between samples. When selecting a point for a new candidate point, the first sample p1 is extracted without restriction from all points. Then a cell C is randomly selected from any level of the octree, so that p1 is included in C, and then the other k-1 samples are extracted only from cell C.

[0210] In this step, C needs to be selected at an appropriate level. C can be selected from a level according to a uniform distribution that reflects the possible new levels containing good units, thereby further improving the sampling efficiency.

[0211] like Figure 7The figure shows the results of using Efficient RANSAC to detect and extract patches in an underground parking lot. It can be seen that the surface objects (walls, floors, ceilings, etc.) in the underground parking lot are well detected and extracted. However, due to the existence of internal objects, not all patches are valid.

[0212] 2. Internal dynamic object removal and facet screening:

[0213] Dynamic objects inside underground parking lots include cars, pedestrians and other objects. Removing dynamic objects inside parking lots is relatively simple compared to removing them in tunnel scenes. Here, the project adopts a strategy of removing internal dynamic objects and selecting patches combined with prior knowledge.

[0214] (1) First, the facets with the largest and smallest Z values ​​are selected based on the size of the point cloud bounding box. As we know from prior knowledge that the floor is at the bottom of the parking lot and the ceiling is at the top of the parking lot, the facet candidates of the floor and ceiling can be eliminated.

[0215] (2) Then, the remaining point cloud is sliced ​​vertically along the horizontal direction to obtain each “topless and bottomless” vertical slice segment.

[0216] (3) Next, construct a two-dimensional voxel for each vertical slice and perform cluster analysis (Euclidean clustering or clustering methods such as DBSACN can be used). In order to maintain processing efficiency, Euclidean clustering is used here and KD-Tree is used for neighborhood search.

[0217] (4) For each slice cluster, based on the size of the outer bounding box and the topological relationship between clusters, the clusters inside the outer bounding box are defined as internal dynamic objects and removed. This process is repeated until the underground parking lot point cloud is obtained, as shown in Figure 8 shown.

[0218] 3. Triangulation network construction and model generation:

[0219] After the facets are screened and the dynamic objects are removed, a valid facet point cloud of the underground parking lot is obtained. Then the triangulation network can be constructed and the model can be generated. This step is similar to the modeling of the air-raid shelter in Example 1. First, a Delaunay triangulation network is constructed. It should be noted that if the data used has large holes (such as data obtained by a ground-based scanner), the holes need to be filled before the triangulation network is constructed. Since the valid facets of the building have been obtained, it is only necessary to perform plane linear interpolation on the holes in the valid facets, which will not be described in detail here.

[0220] At the same time, the construction strategy of the triangulated network is similar to that of the air-raid shelter. The triangulated network construction results and Mesh model are shown in Figure 2. Fig. 9 As shown, Fig. 9 is the result of the top-down modeling of the underground parking lot, where Fig. 9 (a) is the point cloud (top view) after the dynamic objects inside the underground parking lot are removed and the facets are screened, (b) is the triangulated network model of the underground parking lot (top view), and (c) is the Mesh model of the underground parking lot (top view). The modeling results in other directions and positions are similar and will not be repeated here.

[0221] It can be seen from the results of the above three-dimensional model construction that the density of the triangulated network formed by artificial scene buildings such as underground parking lots is greater than that of underground cave scenes. This is because, during the processing, it is necessary to restore the detailed features of the building as much as possible. Therefore, not much smoothing is done in the sampling and smoothing links to maintain the original properties of the building. As for the underground cave scene described in Example 1, since it is an irregular and messy environment, it needs to be smoothed to achieve the effective construction of the triangulated network.

[0222] The above-mentioned embodiments are only preferred implementations of the present invention, and are only used to help understand the method and core ideas of the present application. The protection scope of the present invention is not limited to the above-mentioned embodiments. All technical solutions under the idea of ​​the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for fine-grained modeling of underground structures based on laser radar scanning, characterized in that: include: S1. Use a laser radar scanner to collect point cloud data of underground structures, and use a total station to collect feature point data of underground structures; S2, preprocessing the point cloud data to obtain preprocessed point cloud data; S3, evenly selecting a number of feature check points from the feature point data collected by the total station, extracting the coordinate values ​​of the feature check points from the preprocessed point cloud data, and comparing them with the coordinate values ​​of the feature check points in the feature point data, calculating the relative accuracy of the point cloud, and completing the accuracy analysis of the point cloud data; S4, organizing the preprocessed point cloud data through precision analysis into three-dimensional point cloud voxels, and constructing multi-scale geometric features for the point cloud data; S5. If the underground structure is an underground cavern scene, the steps of fine reconstruction of the three-dimensional model include: removing dynamic objects inside the underground cavern according to the geometric features; orienting the axis of the underground cavern based on the L1 median algorithm; extracting and smoothing the cross section of the underground cavern; and constructing a triangulated network and generating a model. S6. If the underground structure is an underground artificial scene, the steps of fine reconstruction of the three-dimensional model include: extracting building faces according to the geometric features; removing internal dynamic objects and screening faces in combination with prior knowledge; and completing triangulation network construction and model generation.

2. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 1 is characterized in that: The pre-processing method in step S2 includes: S201, importing point cloud data: importing point cloud data through point cloud processing software corresponding to the laser radar scanner; S202, point cloud data check: check the layering and trajectory of the point cloud data to determine whether the data is normal; S203, point cloud data stitching: completing point cloud data registration and point cloud data stitching according to different laser radar scanners; S204, point cloud data optimization: point cloud noise reduction is performed by statistical filtering method; point cloud smoothing is performed based on moving least squares method; S205, point cloud data simplification: plane fitting is performed using the RANSAC algorithm, and contour feature information is extracted by the intersection of the fitted planes; then the marked area is extracted by determining the color information; finally, the point cloud data of the non-contour and marked areas is simplified using the bounding box method.

3. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 1 is characterized in that: In step S3, the relative accuracy calculation formula of the point cloud is as follows: ;in, Indicates accuracy, represents the three-dimensional coordinates of the i-th feature check point in the point cloud, It represents the three-dimensional coordinates of the i-th feature check point measured by the total station, and n is the number of feature check points.

4. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 1 is characterized in that: In step S4, the method for voxelizing the three-dimensional point cloud comprises: S401, voxelizing the preprocessed point cloud data, including: assuming that a voxel is a cube with a predefined side length, obtaining a lower left front point and an upper right rear point from a bounding box in the input point cloud data, dividing the bounding box of the point cloud into cubic units in the xyz three-axis direction according to the predefined voxel side length, and then calculating the number of cubic units in the xyz three-axis direction, comparing it with the coordinates of the lower left front point of the bounding box in the xyz three-axis direction, and obtaining a three-dimensional voxel index of each point; S402 , performing voxel resampling and clustering processing on the voxelized point cloud data based on the octree structure.

5. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 4 is characterized in that: The method for constructing multi-scale geometric features in step S4 includes: S411, determining the dimension of each voxel cluster by using a PCA principal component analysis algorithm; making all voxel units in the voxel cluster have a characteristic value representing its dimension; S412, calculating geometric features based on eigenvalues ​​to obtain geometric information of the preprocessed point cloud data; S413. Geometric feature dimensions at different scales are obtained by setting different voxel sizes and octree scales.

6. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 1 is characterized in that: The method for removing dynamic objects inside the underground cavern described in step S5 includes: S501, analyzing the geometric feature dimensions of the preprocessed point cloud data at different scales to remove objects with obvious point features; S502, using a progressive triangulation filtering algorithm to remove ground points of the underground cavern, so that the dynamic objects are independent of the remaining point cloud; S503, using straight-through filtering combined with Euclidean clustering to determine dynamic objects and remove them.

7. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 1 is characterized in that: The method for performing underground cavern axis orientation based on the L1 median algorithm in step S5 includes: S511, extracting the point cloud skeleton based on the L1 median algorithm, and determining the one-dimensional local center of the point cloud after the dynamic objects inside the underground cavern are removed; S512, iteratively clustering the center point of the one-dimensional local center, and introducing the PCA algorithm to judge the line features; when the distance between two cluster centers is less than the set threshold, it is considered that the local cluster center is obtained; S513, reducing the cluster radius for reprocessing to obtain the pipeline axis.

8. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 7 is characterized in that: The method for extracting and smoothing the underground cavern section in step S5 includes: S521, performing step segmentation along the pipeline axis to obtain step local points; constructing two-dimensional voxels to sample the step local points, then traversing all sampling points, constructing plane normals through points before and after each sampling point, treating the sampling point as a point on the plane, and obtaining the plane equation where the sampling point is located; calculating the distance from the sampling point to the plane according to the plane equation to make a judgment, setting a reserved distance value, and retaining the sampling points whose distance is less than the reserved distance value; S522, converting the retained sampling points into two-dimensional voxels to obtain two-dimensional coordinates of each point; S523, continue to voxelize the points of the two-dimensional coordinates, calculate the average value of all points in each voxel, and use all the average values ​​as the cross-section points; S524. Use interpolation method to complete the holes.

9. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 1 is characterized in that: The method for extracting building facets in step S6 includes: S601, decomposing the point cloud data of the underground artificial scene into a concise hybrid structure consisting of an inherent shape and a group of residual points based on an Efficient RANSAC segmentation algorithm; S602, sampling the point cloud using a local sampling strategy, and using an octree to establish spatial proximity between samples; S603, extract the first sample without restriction from all points in the point cloud, and then randomly select a unit C from any level of the octree, the first sample is included in the unit C, and other samples are only extracted from the unit C.

10. The method for fine-grained modeling of underground structures based on laser radar scanning according to claim 9 is characterized in that: The method of removing internal dynamic objects and selecting facets in combination with prior knowledge in step S6 includes: S611, according to the geometric features, screen out the faces with the largest and smallest Z values ​​in the Z-axis direction from the extracted building faces, and screen out the faces based on prior knowledge; S612, slicing the remaining point clouds vertically along the horizontal direction to obtain vertical slices without top or bottom; S613, constructing two-dimensional voxels for each vertical slice and performing cluster analysis; S614: for each cluster of the vertical slice, based on the size of the outer bounding box and the topological relationship between the clusters, the clusters inside the outer bounding box are defined as internal dynamic objects and removed.

Images