A method for filtering oblique photogrammetric point clouds based on cloth simulation algorithm

Abstract

This invention relates to a method for filtering oblique photogrammetric point clouds based on a cloth simulation algorithm, comprising: Step 1: establishing an optimal parameter library for cloth simulation filtering; Step 2: identifying ground targets using target detection technology; Step 3: defining the region of interest (ROI) for the ground targets; Step 4: dividing the ROI into a 3D grid for cloth simulation filtering based on the ROI of the ground targets; Step 5: analyzing terrain features and adaptively selecting optimal filtering parameters from the optimal parameter library based on the attribute labels of the 3D grid; Step 6: completing filtering and ground extraction based on the optimal filtering parameters. This invention pre-establishes an optimal parameter library for filtering corresponding to various terrains and ground features in a given scene, and then automatically selects the corresponding optimal filtering parameters from the parameter library through ground feature identification and terrain analysis. This achieves adaptive selection of filtering algorithm parameters in scenes with complex terrain or complex ground feature coverage, improving the automation level, efficiency, and ground extraction accuracy of filtering.

Description

Technical Field

[0001] This invention relates to the fields of civil engineering construction and surveying engineering technology, and in particular to an oblique photography point cloud filtering method based on a cloth simulation algorithm. Background Technology

[0002] In recent years, with the rapid development of hardware and algorithms, UAV oblique photogrammetry technology has made significant progress and has been widely applied in fields such as 3D city reconstruction and historical building digitization. By equipping a UAV platform with multi-lens sensors, ground images can be simultaneously acquired from multiple perspectives in one vertical direction and four oblique directions, quickly generating 3D ground models and obtaining 3D spatial data. Compared with other types of real-scene data acquisition methods such as 3D laser scanning, oblique photogrammetry technology offers higher data acquisition efficiency, lower cost, and wider applicability, thus becoming the mainstream data acquisition technology in many industries such as engineering construction and land surveying.

[0003] Oblique photogrammetry point clouds are one of the main forms of data output, generally used to generate digital terrain models (DTMs), earthwork calculations, and elevation control. However, compared to airborne LiDAR point clouds, oblique photogrammetry point clouds are characterized by their inability to penetrate vegetation, high density, and the inclusion of not only ground points but also non-ground points such as vegetation and buildings. These non-ground points significantly impact the accuracy of precise ground point cloud applications, thus often requiring filtering.

[0004] Existing point cloud filtering algorithms, such as filtering algorithms based on mathematical morphology, filtering algorithms based on slope, filtering algorithms based on surface fitting, and point cloud filtering algorithms based on irregular triangular networks, are mostly designed for LiDAR point clouds. They are poorly applicable to oblique photogrammetry point clouds and generally have drawbacks such as poor adaptability to complex terrain and complex parameter settings.

[0005] The cloth simulation filtering algorithm flips the surface point cloud and uses a "spring-mass" cloth model to cover the inverted point cloud from top to bottom, achieving point cloud filtering based on the elevation difference between the cloth model and the surface point cloud. The algorithm's parameter settings are simple, requiring only the selection of appropriate cloth stiffness coefficients, cloth grid resolution, iteration counts, and height threshold parameters, making it feasible for application on oblique photogrammetric point cloud data. However, the algorithm's key parameters are all global variables and cannot adaptively adjust according to terrain regions. For scenarios with large areas of complex terrain or feature coverage, its global variable parameters struggle to handle complex situations, resulting in poor filtering performance. Summary of the Invention

[0006] This invention provides a method for filtering oblique photography point clouds based on a cloth simulation algorithm to solve the above-mentioned technical problems.

[0007] To address the aforementioned technical problems, this invention provides a method for filtering oblique photogrammetric point clouds based on a cloth simulation algorithm, comprising the following steps:

[0008] Step 1: Establish a cloth simulation filtering optimal parameter library, which contains optimal filtering parameters for different terrain features obtained through manual experiments; the terrain includes at least flat land, gentle slope, and steep slope; the terrain features include at least no terrain features, trees, shrubs, buildings, and machinery; the filtering parameters include at least cloth stiffness coefficient, cloth grid resolution, number of cloth particle elevation correction iterations, and elevation threshold.

[0009] Step 2: Identify ground targets using target detection technology;

[0010] Step 3: Define the region of interest for the target ground feature;

[0011] Step 4: Divide the region of interest of the ground features into a 3D grid for fabric simulation filtering;

[0012] Step 5: Analyze the terrain features of the point cloud in the 3D grid and classify the type; based on the attribute labels of the 3D grid, adaptively select the optimal filtering parameters from the cloth simulation filtering optimal parameter library.

[0013] Step 6: Complete filtering and ground extraction based on the optimal filtering parameters.

[0014] Preferably, in step 1, the optimal filtering parameters are determined based at least on whether the target ground features are filtered out and the number of misclassified points.

[0015] Preferably, step 2 includes:

[0016] Step 2.1: Use image annotation software to annotate the site images acquired by oblique photography to obtain a dataset of typical ground feature detections;

[0017] Step 2.2: Construct a deep learning neural network model for target detection, and train the deep learning neural network model based on the dataset in Step 2.1 to obtain a neural network model for detecting typical ground features in the site that meets the requirements of accuracy, recall and detection precision;

[0018] Step 2.3: Input the site image acquired by oblique photography into the typical ground feature target detection neural network model trained in Step 2.2 to predict the location and category of typical ground features in the site image and output the prediction box attribute information.

[0019] Preferably, step 3 includes:

[0020] Step 3.1: Index and map the oblique photogrammetric 3D point cloud into 2D pixels in pixel space;

[0021] Step 3.2: Based on Euclidean distance, match the pixels of the two-dimensional pixels in Step 3.1 with the original image, mark the pixels within the prediction box described in Step 2.3 and index them back into the three-dimensional space to obtain the three-dimensional point cloud within the prediction box range;

[0022] Step 3.3: Based on the 3D point cloud within the predicted bounding box in Step 3.2, construct the first minimum bounding box and define the first minimum bounding box as the region of interest for ground objects in the 3D point cloud space.

[0023] Preferably, step 4 includes:

[0024] Step 4.1: Construct a second minimum bounding box based on the boundary of the oblique photogrammetric 3D point cloud, and divide the second minimum bounding box into an equidistant 3D grid in the horizontal direction;

[0025] Step 4.2: Traverse the regions of interest corresponding to each object in Step 3.3 and determine their overlap with the 3D grid.

[0026] If the entire region of interest of the ground object is located within a single 3D grid, the 3D grid layout remains unchanged.

[0027] If the region of interest of a ground feature spans multiple 3D grids, then the 3D grids that intersect with the region of interest of the ground feature are merged.

[0028] The 3D grid containing ground features is labeled with its ground feature attributes, while the grid that does not contain ground features is not labeled.

[0029] Step 4.3: Perform statistical analysis on the point clouds within each of the rearranged 3D grids in Step 4.2, and calculate the average elevation of the K nearest neighbors of each point. and variance The elevation of K's nearest neighbor exceeds Points within the specified range are discarded as gross errors.

[0030] Preferably, step 4.1 also includes determining whether there are point clouds in the 3D grid; if there are, the 3D grid is constructed; otherwise, the construction of the 3D grid is canceled.

[0031] Preferably, in the three-dimensional grid established in step 4.1, an overlapping buffer region is provided between two adjacent three-dimensional grids, and the length of the buffer region is 1 / 10 of the side length of the three-dimensional grid.

[0032] Preferably, step 5 includes:

[0033] Step 5.1: Draw contour maps of the point cloud models in each 3D grid set in Step 4, and randomly select several points to form an elevation point set. ;

[0034] Step 5.2: Calculate the slope based on the contour map and the points in the elevation point set P from Step 5.1;

[0035] Step 5.3: Calculate the slope angle of all slopes in Step 5.2 within each 3D grid. mean and variance And accordingly, the terrain of the site in each three-dimensional grid is divided;

[0036] Step 5.4: Based on the land feature labels predicted in Step 2.3 and the terrain types classified in Step 5.3, automatically select the filtering parameters corresponding to each 3D grid from the optimal parameter library for cloth simulation filtering in Step 1.

[0037] Preferably, in step 5.3, the method for classifying the terrain is as follows: [The text abruptly ends here, so the translation stops.] For 3D grids with slopes less than a preset threshold, the average slope is used. Division; Slope variance Three-dimensional grids exceeding a preset threshold are identified as complex terrain.

[0038] Preferably, step 6 includes:

[0039] Step 6.1: Flip the point cloud within each 3D grid along the horizontal XOY plane to invert it;

[0040] Step 6.2: Initialize the simulated cloth using the filtering parameters obtained in Step 5.4, and place it above the highest point within the 3D grid;

[0041] Step 6.3: Project the point cloud and cloth particles onto the same horizontal plane, match the nearest neighbor point for each cloth particle and record its elevation H, which is regarded as the elevation threshold of the particle's movable state.

[0042] Step 6.4: Apply gravity to all movable cloth particles, calculate particle displacement and current elevation. If the particle elevation is lower than or equal to the elevation threshold H, fix the particle at elevation H and change its state to immovable.

[0043] Step 6.5: Calculate the internal forces between the cloth particles and correct the elevation of the movable particles;

[0044] Step 6.6: Repeat steps 6.4 and 6.5 until the maximum elevation change of the particles is small enough or the preset number of iterations is reached, then stop the cloth-laying simulation process;

[0045] Step 6.7: Calculate the elevation difference between each point in the 3D grid point cloud and its nearest neighboring cloth particles. If the elevation difference is less than or equal to the preset elevation threshold H, it is marked as a ground point and retained; otherwise, it is considered a non-ground point and filtered out.

[0046] Step 6.8: Merge the ground point clouds of all 3D grids, remove duplicate point clouds in the buffer area, and perform linear interpolation to encrypt and repair holes in the filtered areas. Ground extraction is now complete.

[0047] Compared with existing technologies, the oblique photography point cloud filtering method based on cloth simulation algorithm provided by this invention has the following advantages:

[0048] This invention achieves automated ground extraction from oblique photogrammetric point clouds in complex scenes through intelligent optimization parameter selection. While retaining the advantages of conventional algorithms such as simple parameters and ease of use, it saves steps such as manually dividing the filtering area, thus saving costs and improving efficiency. At the same time, it significantly improves the adaptability of the filtering algorithm to various terrains, better preserves the feature information of complex terrains, improves the ground extraction accuracy, and is beneficial to the subsequent engineering application of ground point clouds. Attached Figure Description

[0049] Figure 1 This is a flowchart of an oblique photography point cloud filtering method based on a cloth simulation algorithm in a specific embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram illustrating the mapping relationship between the image prediction region and the three-dimensional point cloud in a specific embodiment of the present invention;

[0051] Figure 3 This is a flowchart illustrating the mapping between an image prediction region and a 3D point cloud in a specific embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of three-dimensional grid division considering the region of interest of ground features in a specific embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram illustrating terrain analysis and slope calculation in a specific embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram of an oblique photography point cloud model according to a specific embodiment of the present invention;

[0055] Figure 7 This is a schematic diagram of the region of interest for a ground feature in a specific embodiment of the present invention;

[0056] Figure 8 This is a schematic diagram of the filtered result in a specific embodiment of the present invention. Detailed Implementation

[0057] To illustrate the technical solutions of the invention in more detail, specific embodiments are listed below to demonstrate the technical effects; it should be emphasized that these embodiments are used to illustrate the invention and are not intended to limit the scope of the invention.

[0058] This embodiment was implemented at a construction site.

[0059] Prior to implementation, technicians pre-deployed five image control points and operated a DJI Mavic 2 Enterprise drone platform to perform oblique photography along a predetermined flight path. After the field data collection was completed, the Context Capture software was used to perform aerial triangulation and 3D reconstruction of the image sequence, generating an oblique photography point cloud model of the construction site, as shown in the attached image. Figure 6 As shown.

[0060] Please refer to Figure 1 The oblique photography point cloud model of the construction site is processed using the oblique photography point cloud filtering method based on the cloth simulation algorithm provided by this invention, including the following steps:

[0061] Step 1: Establish a cloth simulation filtering optimal parameter library, which contains the optimal filtering parameters for different terrain features obtained through artificial experiments; the terrain can include at least flat land, gentle slope and steep slope; the terrain features can include at least no terrain features, trees, shrubs, buildings and machinery; the filtering parameters can include at least cloth stiffness coefficient, cloth grid resolution, cloth particle elevation correction iteration number and elevation threshold.

[0062] In some embodiments, the optimal filtering parameters can be determined at least based on whether the target ground features are filtered out and the number of misclassified points.

[0063] In some embodiments, the adjustment range of the filtering parameters during manual experiments can be as shown in Table 1, and the optimal parameter library for the construction site material placement simulation filtering is shown in Table 2.

[0064] Table 1 Filter parameter adjustment range

[0065]

[0066] Table 2. Optimal Parameter Library for Material Placement Simulation Filtering at Construction Sites

[0067]

[0068] Step 2: Identify ground targets using target detection technology, specifically including:

[0069] Step 2.1: Use image annotation software (e.g. LabelImg) to annotate the site images acquired by oblique photography, obtain a typical site feature target detection dataset, and randomly divide it into training set, validation set and test set in a ratio of 8:1:1.

[0070] Step 2.2: Construct a deep learning neural network model for target detection based on the single-stage target detection algorithm framework YOLO v5, and train, verify, and test the deep learning neural network model based on the dataset in Step 2.1 to obtain a typical site feature target detection neural network model with satisfactory accuracy, recall, and detection precision, which can be referred to as the target detection model.

[0071] In one embodiment, the initial training parameters were set as follows: model size s, batch size 64, image resolution 640, learning rate 0.01, and number of training epochs 300. After multiple iterations of training, the detection results of the object detection model are shown in Table 3.

[0072] Table 3 Target detection model detection results

[0073]

[0074] Step 2.3: Input the site image obtained by oblique photography into the typical ground feature target detection neural network model trained in Step 2.2 to predict the location and category of typical ground features appearing in the site image. The category of ground feature is the ground feature label, and the prediction box attribute information is output.

[0075] Taking a site image from one embodiment as an example, the output result is as follows: 0 0.457031 0.239258 0.90625 0.326172 0.785468 2 0.480469 0.709961 0.960938 0.576172 0.86284

[0078] The first column represents the feature labels: 0 for machinery, 1 for temporary structures, and 2 for building materials. The second and third columns represent the horizontal and vertical coordinates of the center point of the prediction box. The fourth and fifth columns represent the width and height information of the prediction box. The sixth column represents the confidence level of the prediction box. Therefore, machinery and building materials were detected in the site image, located on the upper left and lower left sides of the image, respectively.

[0079] Step 3: Define the Region of Interest (ROI) for the target ground feature. Please refer to [link / reference]. Figure 2 and Figure 3 Specifically, it includes:

[0080] Step 3.1: Index and map the oblique photogrammetric 3D point cloud into 2D pixels in pixel space. The mapping matrix expression is:

[0081]

[0082] In the formula, I It is the camera intrinsic parameter matrix, used to describe the transformation relationship from camera coordinates to pixel coordinates; E It is the camera extrinsic matrix, used to describe the transformation relationship from real-world coordinates to camera coordinates; f x and f y They are focal length and x shaft and y The ratio of pixel size along the axial direction; x 0 and y 0 These are the actual coordinates of the camera's principal point. s It is the coordinate axis tilt parameter; R It is a camera rotation matrix. T It is the camera translation matrix.

[0083] Measurements were taken from the visible light camera mounted on the flight platform in some embodiments, and the intrinsic parameter matrix (in pixels) was obtained as follows:

[0084]

[0085] The camera extrinsic parameter matrix corresponding to a certain aerial photograph is:

[0086]

[0087] When mapping a 3D point cloud, multiplying the mapping matrix above by the real-world coordinates of each point in the point cloud yields its 2D coordinates in pixel space, thus generating a 2D image. The specific calculation formula is as follows:

[0088]

[0089] In the formula, ( , , ) are the coordinates of a point in the real-world coordinate system. u,v () represents the coordinates of a point in the pixel coordinate system.

[0090] Step 3.2: Based on Euclidean distance, match the pixels of the two-dimensional pixels in Step 3.1 with the original image, mark the pixels within the prediction box described in Step 2.3, and index them back into the three-dimensional space to obtain the three-dimensional point cloud within the prediction box range.

[0091] Step 3.3: Based on the 3D point cloud within the predicted bounding box in Step 3.2, construct the first minimum bounding box, defining this first minimum bounding box as the region of interest for ground features in the 3D point cloud space, such as... Figure 7 As shown.

[0092] Step 4: Divide the region of interest of the ground features into a 3D grid for cloth simulation filtering. Please refer to [reference needed]. Figure 4 Specifically, it includes:

[0093] Step 4.1: Construct a second minimum bounding box based on the boundaries of the oblique photogrammetric 3D point cloud (the maximum and minimum values ​​of the point cloud in the X, Y, and Z directions), and divide the second minimum bounding box into an equidistant 3D grid in the horizontal plane. In some embodiments, the 3D grid can be uniformly divided on the top of the minimum bounding box using multiple squares with equal side lengths, and the side length of the 3D grid can be 50m.

[0094] In some embodiments, this step further includes determining whether there are point clouds in the 3D grid. If there are, the 3D grid is constructed; otherwise, the construction of the 3D grid is canceled to avoid an invalid state in which the constructed 3D grid has no point clouds.

[0095] In some embodiments, considering the possibility of drastic terrain changes or large-area cover between two three-dimensional grids, when establishing a three-dimensional grid, an overlapping buffer area can be set between two adjacent three-dimensional grids. The length of the buffer area can be 1 / 10 of the side length of the three-dimensional grid, for example, 5m.

[0096] Step 4.2: Traverse the regions of interest (ROIs) corresponding to each feature target in Step 3.3, determine their overlap with the 3D grid, and if the entire ROI of the feature target is located within a single 3D grid, keep the 3D grid layout unchanged; if the ROI of the feature target spans multiple 3D grids, merge the 3D grids that intersect with the ROI of the feature target so that the merged 3D grid can contain the complete feature ROI; mark the feature attributes of the 3D grids that contain the feature target, and do not mark the feature attributes of the grids that do not contain the feature target.

[0097] Step 4.3: Perform statistical analysis on the point clouds within each of the rearranged 3D grids in Step 4.2, and calculate the average elevation of the K nearest neighbors of each point. and variance The elevation of K's nearest neighbor exceeds Points within the specified range are discarded as gross errors.

[0098] Taking the removal of a coarse-looking point P in a point cloud from a certain embodiment as an example:

[0099] The elevations of the 100 nearest neighbors (K = 100) of a point Q within a 3D grid are statistically analyzed, and the mean elevation is calculated. and variance The measurements are 9.875m and 0.062m respectively, and we take... k =3, then the elevation threshold of the nearest neighbors of this point is

[0100]

[0101] Since point P is one of the 100 nearest neighbors of point Q and its elevation is 10.685m, point P is determined to be an outlier and is deleted.

[0102] Step 5: Analyze the terrain features of the point cloud in the 3D grid and classify its type. Based on the attribute labels of the 3D grid, adaptively select the optimal filtering parameters from the cloth simulation filtering optimal parameter library. Please refer to [link / reference]. Figure 5 Specifically, it includes:

[0103] Step 5.1: Draw contour maps of the point cloud models in each 3D grid set in Step 4, and randomly select several points to form an elevation point set. .

[0104] Step 5.2: Based on the contour map and the points in the elevation point set P from Step 5.1, calculate the slope as follows:

[0105]

[0106] In the formula, It's the slope. It's the slope angle. It is an elevation point , The three-dimensional coordinates.

[0107] Step 5.3: Calculate the slope angle of all slopes in Step 5.2 within each 3D grid. mean and variance And so on, the terrain of the site in each three-dimensional grid is divided.

[0108] In some embodiments, the method for classifying terrain can be: [analyzing] the slope variance. For 3D grids with an angle less than 5° below a preset threshold, the slope is averaged. The slopes are classified into three categories: flat land, gentle slopes, and steep slopes; the slope variance is also considered. Three-dimensional grids with an angle greater than a preset threshold of 5° are directly identified as complex terrain, as shown in Table 4.

[0109] Table 4. Criteria for Classification of Terrain Types

[0110]

[0111] Step 5.4: Based on the land feature labels predicted in Step 2.3 and the terrain types classified in Step 5.3, automatically select the filtering parameters corresponding to each 3D grid from the optimal parameter library for cloth simulation filtering in Step 1.

[0112] Step 6: Complete filtering and ground extraction based on the optimal filtering parameters, specifically including:

[0113] Step 6.1: Flip the point cloud in each 3D grid along the horizontal plane XOY to invert it.

[0114] Step 6.2: Initialize the simulated cloth using the filtering parameters obtained in Step 5.4, and place it above the highest point within the 3D grid.

[0115] Taking a steep slope area A containing mechanical equipment and an area B with complex terrain as examples in a certain embodiment, the preset parameters of the cloth simulation filtering process are shown in Table 5.

[0116] Table 5. Example of preset parameter selection for a certain area in a certain embodiment.

[0117]

[0118] Step 6.3: Project the point cloud and cloth particles onto the same horizontal plane, match the nearest neighbor point for each cloth particle and record its elevation H, which is regarded as the elevation threshold of the particle's movable state.

[0119] Step 6.4: Apply gravity to all movable cloth particles, calculate particle displacement and current elevation. If the particle elevation is lower than or equal to the elevation threshold H, fix the particle at elevation H and change its state to immovable.

[0120] The equations for determining the displacement of each particle under gravity are as follows:

[0121]

[0122] In the formula, It is the particle in time t The position at that moment; This is the time step, set to 0.65; m is the particle mass, usually set to 1; G is the gravitational force acting on the particle.

[0123] Step 6.5: Calculate the internal forces between the cloth particles and correct the elevation of the movable particles.

[0124] The formula for calculating the displacement of a particle under the action of internal forces is as follows:

[0125]

[0126] In the formula, d Represents the displacement vector of the particle; b It is a Boolean value, which is 1 when the particle is in a movable state, and 0 otherwise. p i and p 0 These are the positions of the particle that is currently about to move and the positions of its adjacent particles; n It is a unit vector in the vertical direction, that is... (0,0,1) T .

[0127] Step 6.6: Repeat steps 6.4 and 6.5 until the maximum elevation change of the particles is small enough or the preset number of iterations is reached, then stop the cloth simulation process.

[0128] Step 6.7: Calculate the elevation difference between each point in the 3D grid point cloud and its nearest neighboring cloth particles. If the elevation difference is less than or equal to the preset elevation threshold H, it is marked as a ground point and retained; otherwise, it is considered a non-ground point and filtered out.

[0129] Step 6.8: Merge all 3D grid ground point clouds, remove duplicate point clouds in the buffer area, and perform linear interpolation to refine the filtered areas and repair holes. Finally, the oblique photogrammetry point cloud cloth simulation filtering and ground extraction are completed, as shown below. Figure 8 As shown.

[0130] In summary, the oblique photogrammetry point cloud filtering method based on cloth simulation algorithm provided by this invention includes the following steps: Step 1: Establishing an optimal parameter library for cloth simulation filtering, which contains optimal filtering parameters for different terrain features obtained through manual experiments; the terrain includes at least flat land, gentle slope, and steep slope; the terrain features include at least no terrain features, trees, shrubs, buildings, and machinery; the filtering parameters include at least cloth rigidity coefficient, cloth grid resolution, cloth particle elevation correction iteration number, and elevation threshold; Step 2: Identifying terrain features using target detection technology; Step 3: Setting the region of interest (ROI) for the terrain features; Step 4: Dividing the three-dimensional grid for cloth simulation filtering based on the ROI of the terrain features; Step 5: Analyzing the terrain features of the point cloud in the three-dimensional grid and classifying the type, and adaptively selecting the optimal filtering parameters from the optimal parameter library for cloth simulation filtering based on the attribute labels of the three-dimensional grid; Step 6: Completing filtering and ground extraction based on the optimal filtering parameters. This invention proposes an improved algorithm for oblique photogrammetry point cloud cloth simulation filtering that combines ground feature recognition and terrain analysis. It can improve conventional cloth simulation filtering algorithms by using image recognition target detection and terrain analysis, and realize automatic grid division and adaptive adjustment of filtering parameters for oblique photogrammetry point cloud filtering in complex terrain or complex ground feature coverage scenarios, thereby improving the accuracy and efficiency of ground point cloud extraction.

[0131] Obviously, those skilled in the art can make various modifications and variations to the invention without departing from the spirit and scope of the invention. Therefore, if these modifications and variations fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A method for filtering oblique photogrammetric point clouds based on a cloth simulation algorithm, characterized in that, Includes the following steps: Step 1: Establish a cloth simulation filtering optimal parameter library, which contains optimal filtering parameters for different terrain features obtained through manual experiments; the terrain includes at least flat land, gentle slope, and steep slope; the terrain features include at least no terrain features, trees, shrubs, buildings, and machinery; the filtering parameters include at least cloth stiffness coefficient, cloth grid resolution, number of cloth particle elevation correction iterations, and elevation threshold. Step 2: Identify ground targets using target detection technology; Step 2 includes: Step 2.1: Use image annotation software to annotate the site images acquired by oblique photography to obtain a dataset of typical ground feature detections; Step 2.2: Construct a deep learning neural network model for target detection, and train the deep learning neural network model based on the dataset in Step 2.1 to obtain a neural network model for detecting typical ground features in the site that meets the requirements of accuracy, recall and detection precision; Step 2.3: Input the site image obtained by oblique photography into the typical ground feature target detection neural network model trained in Step 2.2 to predict the location and category of typical ground features in the site image and output the prediction box attribute information; Step 3: Define the region of interest for the target feature; Step 3 includes: Step 3.1: Index and map the oblique photogrammetric 3D point cloud into 2D pixels in pixel space; Step 3.2: Based on Euclidean distance, match the pixels of the two-dimensional pixels in Step 3.1 with the original image, mark the pixels within the prediction box described in Step 2.3 and index them back into the three-dimensional space to obtain the three-dimensional point cloud within the prediction box range; Step 3.3: Based on the 3D point cloud within the predicted bounding box in Step 3.2, construct the first minimum bounding box and define the first minimum bounding box as the region of interest for ground objects in the 3D point cloud space; Step 4: Divide the region of interest of the ground features into a 3D grid for cloth simulation filtering; Step 4 includes: Step 4.1: Construct a second minimum bounding box based on the boundary of the oblique photogrammetric 3D point cloud, and divide the second minimum bounding box into an equidistant 3D grid in the horizontal direction; Step 4.2: Traverse the regions of interest corresponding to each object in Step 3.3 and determine their overlap with the 3D grid. If the entire region of interest of the ground object is located within a single 3D grid, the 3D grid layout remains unchanged. If the region of interest of a ground feature spans multiple 3D grids, then the 3D grids that intersect with the region of interest of the ground feature are merged. The 3D grid containing ground features is labeled with its ground feature attributes, while the grid that does not contain ground features is not labeled. Step 4.3: Perform statistical analysis on the point clouds within each of the rearranged 3D grids in Step 4.2, and calculate the average elevation of the K nearest neighbors of each point. and variance The elevation of K's nearest neighbor exceeds Points within the specified range are discarded as gross errors. k Greater than 0; Step 5: Analyze the terrain features of the point cloud in the 3D grid and classify the type; based on the attribute labels of the 3D grid, adaptively select the optimal filtering parameters from the cloth simulation filtering optimal parameter library. Step 6: Complete filtering and ground extraction based on the optimal filtering parameters.

2. The oblique photography point cloud filtering method based on cloth simulation algorithm as described in claim 1, characterized in that, In step 1, the optimal filtering parameters are determined based at least on whether the target features are filtered out and the number of misclassified points.

3. The oblique photography point cloud filtering method based on cloth simulation algorithm as described in claim 1, characterized in that, Step 4.1 also includes determining whether there are point clouds in the 3D grid. If there are, the 3D grid is constructed; otherwise, the construction of the 3D grid is canceled.

4. The oblique photography point cloud filtering method based on cloth simulation algorithm as described in claim 1, characterized in that, In the three-dimensional grid established in step 4.1, an overlapping buffer region is set between two adjacent three-dimensional grids, and the length of the buffer region is 1 / 10 of the side length of the three-dimensional grid.

5. The oblique photography point cloud filtering method based on cloth simulation algorithm as described in claim 4, characterized in that, Step 5 includes: Step 5.1: Draw contour maps of the point cloud models in each 3D grid set in Step 4, and randomly select several points to form an elevation point set. ; Step 5.2: Based on the contour map and elevation point set in Step 5.1 P Calculate the slope at each point in the diagram; Step 5.3: Calculate the slope angle of all slopes in Step 5.2 within each 3D grid. mean and variance And accordingly, the terrain of the site in each three-dimensional grid is divided; Step 5.4: Based on the land feature labels predicted in Step 2.3 and the terrain types classified in Step 5.3, automatically select the filtering parameters corresponding to each 3D grid from the optimal parameter library for cloth simulation filtering in Step 1.

6. The oblique photography point cloud filtering method based on cloth simulation algorithm as described in claim 5, characterized in that, In step 5.3, the method for classifying terrain is as follows: [The text abruptly shifts to a different topic] ... slope variance For 3D grids with slopes less than a preset threshold, the average slope is used. Division; Slope variance Three-dimensional grids exceeding a preset threshold are identified as complex terrain.

7. The oblique photography point cloud filtering method based on cloth simulation algorithm as described in claim 6, characterized in that, Step 6 includes: Step 6.1: Flip the point cloud within each 3D grid along the horizontal XOY plane to invert it; Step 6.2: Initialize the simulated cloth using the filtering parameters obtained in Step 5.4, and place it above the highest point within the 3D grid; Step 6.3: Project the point cloud and cloth particles onto the same horizontal plane, match the nearest neighbor point for each cloth particle and record its elevation H, which is regarded as the elevation threshold of the particle's movable state. Step 6.4: Apply gravity to all movable cloth particles, calculate particle displacement and current elevation. If the particle elevation is lower than or equal to the elevation threshold H, fix the particle at elevation H and change its state to immovable. Step 6.5: Calculate the internal forces between the cloth particles and correct the elevation of the movable particles; Step 6.6: Repeat steps 6.4 and 6.5 until the maximum elevation change of the particles reaches the preset number of iterations, then stop the cloth-laying simulation process; Step 6.7: Calculate the elevation difference between each point in the 3D grid point cloud and its nearest neighboring cloth particles. If the elevation difference is less than or equal to the preset elevation threshold H, it is marked as a ground point and retained; otherwise, it is considered a non-ground point and filtered out. Step 6.8: Merge the ground point clouds of all 3D grids, remove duplicate point clouds in the buffer area, and perform linear interpolation to encrypt and repair holes in the filtered areas. Ground extraction is now complete.

Images