Point cloud data filtering method and system based on progressive triangulation network, and medium

Abstract

The invention relates to a point cloud data filtering method and system based on a progressive triangulation network, and a medium, and belongs to the technical field of point cloud data processing. The method comprises the following steps: projecting point cloud data to the same plane; moving the rectangular window to enable the rectangular window to traverse all the point cloud data so as to extract undetermined ground seed points; screening to-be-determined ground seed points through curved surface fitting, removing non-ground seed points, and obtaining final ground seed points; and establishing an initial triangulation network, calculating a repeated distance and a repeated angle between the unclassified points and the initial triangulation network, determining an encrypted triangulation network, and obtaining ground points. According to the method, more to-be-determined ground seed points are extracted by sliding the rectangular window, the ground points are prevented from being mistakenly rejected through multiple times of curved surface fitting and seed point screening, the noise point rejection effect is improved, the method adapts to complex terrains, and the filtering precision is improved by encrypting the initial triangulation network.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud data processing, and particularly relates to a point cloud data filtering method, system and medium based on a progressive triangular mesh. Background Art

[0002] Airborne Light Detection and Ranging (LiDAR) technology can quickly obtain the three-dimensional spatial information of ground objects and is widely used in landslide detection, forest inventory, Digital Elevation Model (DEM) generation, three-dimensional building reconstruction and other fields. However, since the three-dimensional point cloud data obtained by the system includes both ground points and object points such as buildings, vegetation and bridges, it is necessary to separate the ground points and object points, and this process is called filtering. Currently, the more widely used filtering algorithms include filtering algorithms based on mathematical morphology, filtering algorithms based on slope, filtering algorithms based on interpolation, etc. Among them, the filtering algorithm based on interpolation has a simple principle and good filtering effect. As a kind of interpolation filtering algorithm, the Progressive TIN Delaunay (PTD) algorithm is widely used due to its stability and high efficiency.

[0003] The traditional PTD algorithm depends on the simulation accuracy of the initial triangular mesh for the terrain, and the number and quality of ground seed points will greatly affect the filtering effect. To solve this problem, some scholars have improved it. For example, in 2020, Liu Yang et al. published a literature titled "An Improved Progressive Triangulation Point Cloud Filtering Algorithm" in the 45th volume of the journal "Science of Surveying and Mapping", proposed an extended local minimum method, and optimized the seed point selection strategy by combining thin plate spline interpolation; in 2019, Wang Jingxue et al. published a literature titled "Airborne LiDAR Point Cloud Filtering Algorithm Combining Morphology and TIN Triangulation" in the 44th volume of the journal "Science of Surveying and Mapping", first removed large objects through morphological filtering, and then used the filtering result as the input for subsequent triangular mesh filtering to reduce the negative impact caused by large objects; in 2015, Wang Shugen et al. published a literature titled "LiDAR Point Cloud Iterative Filtering Algorithm Based on Statistical Data to Select Seed Points" in the 38th volume of the journal "Geomatics & Spatial Information Technology", proposed a method for selecting seed points based on small grid elevation, mean square deviation and point density statistical data, and improved the filtering effect through two triangular mesh filterings.

[0004] However, most of the existing improved algorithms consider the influence of large-scale features on the selection of seed points. By filtering multiple times to reduce the grid size, more seed points can be obtained. However, the number of seed points obtained by the traditional square grid is still limited. Moreover, the traditional algorithm establishes a grid index for the original LiDAR point cloud data and selects the lowest point within the grid as the ground seed point. Since the grid used in this algorithm is a square grid larger than the maximum building size in the study area, the number of selected seed points is not only small, but also some low-level noise points will be misclassified as ground points and added to the initial TIN model. TIN is the English abbreviation of Triangle Irregular Network. As a result, there is a large difference between the initial TIN model and the real terrain, and the filtering accuracy still needs to be improved. In view of this, it is necessary to design a progressive triangular network point cloud data filtering method to solve the problems of large difference between the initial TIN model and the real terrain and insufficient filtering accuracy. Summary of the Invention

[0005] The purpose of the present invention is to provide a point cloud data filtering method, system and medium based on a progressive triangular network to solve the above problems.

[0006] The present invention achieves the above purpose through the following technical solutions:

[0007] A point cloud data filtering method based on a progressive triangular network includes the following steps:

[0008] Step 1: Project the point cloud data onto the same plane;

[0009] Step 2: Move a rectangular window within the plane to make the rectangular window traverse all the point cloud data to extract the pending ground seed points;

[0010] Step 3: Fit the pending ground seed points using a cubic surface fitting function to obtain a surface, perform a height difference judgment on the surface, screen out non-ground seed points to update the pending seed points, and form the final ground seed points;

[0011] Step 4: Establish an initial triangular network based on the final ground seed points, calculate the repeated distance and repeated angle between the unclassified points and the initial triangular network, and determine the encrypted triangular network according to the relationship between the calculation results and the preset repeated distance threshold and repeated angle threshold, and obtain the ground points; the unclassified points include the remaining point cloud data and non-ground seed points after extracting the pending ground seed points.

[0012] As a further optimized solution of the present invention, in Step 2, the lowest point within the rectangular window is used as the initial ground seed point, and the initial ground seed points obtained during the traversal process are used as the pending ground seed points.

[0013] As a further optimization solution of the present invention, the movement of the rectangular window includes sliding and leaping. The rectangular window is divided into Window 1 and Window 2. Both Window 1 and Window 2 traverse all point cloud data. During the traversal process, the sliding direction of Window 2 is different from that of Window 1, and the leaping direction of Window 2 is different from that of Window 1.

[0014] As a further optimization solution of the present invention, Step 3 includes:

[0015] Sieving all the to-be-determined ground seed points by using a cubic surface fitting function, and finding n other to-be-determined ground seed points around each to-be-determined ground seed point through the KNN nearest neighbor method;

[0016] Taking the n other to-be-determined ground seed points as reference points, taking the to-be-determined ground seed points surrounded by the reference points as test points, and establishing a surface based on the n reference points and the cubic surface fitting function;

[0017] When the height difference between the test point and the surface is greater than or equal to a preset height difference threshold, sieve out the non-ground seed points to update the to-be-determined seed points and form the final ground seed points.

[0018] As a further optimization solution of the present invention, in Step 3, all the to-be-determined ground seed points are sieved m times by using a cubic surface fitting function, and the to-be-determined ground seed points retained in the last sieving are used as the final ground seed points. The height difference thresholds for each sieving are h1, h2,..., h m , and the height difference thresholds for each sieving decrease in turn.

[0019] As a further optimization solution of the present invention, in Step 3, all the to-be-determined ground seed points are sieved m times by using a cubic surface fitting function, and the to-be-determined ground seed points retained after the last sieving and the boundary points are used as the final ground seed points.

[0020] As a further optimization solution of the present invention, the plane for projecting the point cloud data is a rectangular plane, the boundary points are new points obtained by replacing the elevations of the four vertices of the rectangular plane, and the elevation of the new points is the elevation of the to-be-determined ground seed point closest to the vertex of the rectangular plane.

[0021] As a further optimization solution of the present invention, when determining the encrypted triangular network, if an unclassified point is newly added as a ground point, the obtained ground point is incorporated into the final ground seed points, and Step 4 is repeated until no new ground points are generated.

[0022] A point cloud data filtering system based on an incremental triangular network, comprising:

[0023] A point cloud data projection module, configured to project point cloud data onto the same plane;

[0024] A seed point extraction module, which is used to move a rectangular window within the plane to make the rectangular window traverse all point cloud data, so as to extract undetermined ground seed points;

[0025] A seed point screening module, which is used to fit the undetermined ground seed points by using a cubic surface fitting function to obtain a surface, judge the height difference of the surface, screen out non-ground seed points, so as to update the undetermined seed points and form final ground seed points;

[0026] A ground point acquisition module, which is used to establish an initial triangular network according to the final ground seed points, calculate the repeated distance and repeated angle between the unclassified points and the initial triangular network, and determine the encrypted triangular network according to the relationship between the calculation results and the preset repeated distance threshold and repeated angle threshold, and obtain ground points.

[0027] A storage medium, on which computer program instructions are stored, characterized in that when the computer program instructions are executed by a processor, each step of the above-mentioned point cloud data filtering method based on an incremental triangular network is realized.

[0028] The beneficial effects of the present invention are as follows:

[0029] 1) After the point cloud data is projected onto the same plane, the present invention moves a rectangular window on this plane through the IPTD algorithm. The rectangular window is divided into window one and window two. Both window one and window two traverse all point cloud data. During the traversal process, the sliding direction of window two is different from that of window one, and the jumping direction of window two is also different from that of window one. Through these two rectangular windows, more initial ground seed points can be extracted during the two traversals of the data, which helps to improve the filtering accuracy;

[0030] 2) The present invention uses a cubic surface fitting function to screen the undetermined ground seed points multiple times, and the height difference threshold decreases in turn, avoiding the risk of eliminating the undetermined ground seed points that can be used as the final ground seed points in the early stage, and improving the closeness of the fitting surface to the real ground in the later stage. During the later surface fitting process, since the fitting surface is close to the real ground and the height difference threshold is relatively small, the accuracy of seed point screening in the later stage can be improved, avoiding the wrong elimination of the final ground seed points, increasing the number of final ground seed points, reducing the risk of including noise points in the final ground seed points, and improving the adaptability to complex terrains;

[0031] 3) After the final ground seed points are generated, the present invention determines whether the unclassified points need to be added as ground points according to the repeated distance and repeated angle of the unclassified points. If new ground points are added, the initial triangular network is updated with the newly added ground points, and the above steps are repeated. Otherwise, the last updated initial triangular network is used as the encrypted triangular network, which can encrypt the initial triangular network, reintroduce the ground points that were wrongly eliminated before, and further improve the filtering accuracy. Description of the Drawings

[0032] Figure 1 is the flowchart of point cloud data processing of the present invention;

[0033] Figure 2 is the schematic diagram of the state when the initial ground seed points are extracted by the rectangular window of the present invention;

[0034] Figure 3 is the display diagram of the point cloud data of Dataset 2 of the present invention;

[0035] Figure 4 is the display diagram of the number and accuracy of seed points extracted by the PTD algorithm and the IPTD algorithm of the present invention;

[0036] Figure 5 is the true ground TIN model diagram of the present invention;

[0037] Figure 6 is the initial TIN model diagram of the PTD algorithm of the present invention;

[0038] Figure 7 is the initial TIN model diagram of the IPTD algorithm after 1 - time surface fitting of the present invention;

[0039] Figure 8 is the initial TIN model diagram of the IPTD algorithm after 2 - time surface fitting of the present invention;

[0040] Figure 9 is the classified point cloud diagram of the PTD algorithm of the present invention;

[0041] Figure 10 is the misclassified point cloud diagram of the PTD algorithm of the present invention;

[0042] Figure 11 is the classified point cloud diagram of the IPTD algorithm of the present invention;

[0043] Figure 12 is the misclassified point cloud diagram of the IPTD algorithm of the present invention;

[0044] Figure 13 is the display diagram of the filtering result of Dataset 2 of the present invention;

[0045] Figure 14 is the system block diagram of the present invention. Detailed Embodiments

[0046] The following further describes the present application in detail with reference to the drawings. It is necessary to point out here that the following detailed embodiments are only used to further illustrate the present application and cannot be understood as limiting the protection scope of the present application. Those skilled in the art can make some non - essential improvements and adjustments to the present application according to the above application content.

[0047] Example 1

[0048] As Figure 1 and Figure 2 shown, this embodiment relates to a point cloud data filtering method based on an incremental triangular mesh. This method uses an improved incremental triangular mesh filtering algorithm (hereinafter referred to as the IPTD algorithm) to filter the point cloud data collected in the study area. This point cloud data filtering method includes the following steps:

[0049] Step 1: Project the point cloud data onto the same plane. The point cloud data consists of multiple discrete points. This plane is a rectangular plane, which is set in the plane rectangular coordinate system. In the initial state, the vertex at the lower left corner of this rectangular plane coincides with the origin of the rectangular coordinate system, the lower side of the rectangular plane coincides with the positive semi-axis of the x-axis of the plane rectangular coordinate system, and the left side of the rectangular plane coincides with the positive semi-axis of the y-axis of the plane rectangular coordinate system.

[0050] Step 2: Slide and jump the rectangular window in the plane so that the rectangular window traverses all the point cloud data, and extract the pending ground seed points from the point cloud data within the rectangular window. During the extraction process, the lowest point within the rectangular window is used as the initial ground seed point, and the initial ground seed points obtained during the traversal process are summarized as the pending ground seed points.

[0051] It should be noted that this rectangular window is divided into Window 1 and Window 2. Both Window 1 and Window 2 traverse all the point cloud data, and during the traversal process, the sliding direction of Window 2 is different from that of Window 1, and the jumping direction of Window 2 is also different from that of Window 1. Through these two rectangular windows, more initial ground seed points can be extracted during the two traversals of the data, which helps to improve the filtering accuracy. It should be noted that in this embodiment, the sliding of the rectangular window means that there is an overlapping area between the rectangular window before movement and the rectangular window after movement, and the jumping of the rectangular window means that there is no overlapping area between the rectangular window before movement and the rectangular window after movement.

[0052] Specifically, as Figure 2 shown, the solid-line rectangular window in the figure is Window 1 in the initial state, and the dashed-line rectangular window is Window 2 in the initial state. Window 1 and Window 2 have the same shape and are perpendicular to each other in the length direction. Project the point cloud data onto the same plane, make the long side of Window 1 parallel to the x-axis and the short side parallel to the y-axis, starting from the origin, move along the x-axis and y-axis directions with step sizes s and b respectively, and take the lowest point within the window as the pending seed point until the sliding window traverses all the point cloud data. Then make the long side of Window 2 parallel to the y-axis and the short side parallel to the x-axis, starting from the origin, move along the y-axis and x-axis directions with step sizes s and b respectively, and take the lowest point within the rectangular window as the pending seed point. Among them, b is the short-side size of the rectangular window.

[0053] The traditional point cloud data filtering algorithm is the Progressive TIN Density Filtering algorithm (hereinafter referred to as the PTD algorithm). In the initial ground seed point extraction method of the PTD algorithm, only the point cloud data is divided into rectangular grids, and the initial ground seed points are extracted from the grids. In the initial ground seed point extraction method of the IPTD algorithm of the present invention, more different regions can be delineated in two mobile rectangular planes, so as to obtain more pending seed points.

[0054] Step 3: Use the cubic surface fitting function to screen all pending ground seed points, and find n other pending ground seed points around each pending ground seed point by the KNN nearest neighbor method. Taking the n other pending ground seed points as reference points and the pending ground seed points surrounded by the reference points as test points, a surface is established based on the n reference points and the cubic surface fitting function. The height difference between the test point and the surface is denoted as ΔH, and the height difference threshold is denoted as h. When the height difference between the test point and the surface is less than the preset height difference threshold, that is, ΔH < h, then the test point is retained, and the test point is still a pending ground seed point; when the height difference between the test point and the surface is not less than the preset height difference threshold, that is, ΔH ≥ h, then it is determined that the test point is a non-ground seed point, and the non-ground seed point is removed.

[0055] In this step 3, all pending ground seed points are screened m times by using the cubic surface fitting function. The pending ground seed points and boundary points retained after the last screening are used as the final ground seed points. Here, the last time is the mth time. When the screening times reach the mth time, the surface fitting meets the requirements. The specific screening times can be adjusted according to actual needs. The height difference thresholds h for each screening are threshold h1, threshold h2,..., threshold h m , and the height difference thresholds for each screening decrease in turn, that is, threshold h1 > threshold h2 >... > threshold h m , in this embodiment, the seed points are constrained by means of multi-level surface fitting, that is, the number of seed points used for surface fitting is gradually reduced. By performing surface fitting on the seed points multiple times, the incorrect seed points are removed. Assuming that only one surface fitting is performed, that is, only one screening is performed, then the height difference threshold needs to be set small enough. However, due to the single surface fitting time, the surface fitted is quite different from the actual ground. The single small-scale height difference threshold will cause incorrect removal of the pending ground seed points that can be used as the final ground seed points, and it is also easy to cause the situation that noise points are included in the final ground seed points. It can be seen that the seed point screening effect of this fitting scheme is weaker than that of the fitting scheme of this embodiment.

[0056] In this embodiment, a relatively large height difference threshold is set during the initial surface fitting, which can avoid the risk of eliminating the pending ground seed points that can be used as the final ground seed points in the early stage, and improve the closeness between the fitting surface and the real ground in the later stage. During the later surface fitting process, since the fitting surface is close to the real ground and the height difference threshold is relatively small, the accuracy of seed point screening in the later stage can be improved, the risk of wrongly eliminating the final ground seed points can be avoided, the number of final ground seed points can be increased, the risk of including noise points in the final ground seed points can be reduced, and the adaptability to complex terrains can be improved. Since the plane for projecting the point cloud data is a rectangular plane, the four vertices of this rectangular plane can be transformed into boundary points.

[0057] In this embodiment, the boundary points are new points obtained by replacing the elevations of the four vertices of the rectangular plane. The elevation of the new points is the elevation of the pending ground seed point closest to the vertex of the rectangular plane, so as to ensure that the TIN model established by the final ground seed points and the new points can cover all the point cloud data and further improve the filtering accuracy. The long side dimension of the rectangular window in this embodiment should be greater than the maximum overlooking dimension of the largest building in the study area, and the short side dimension can be adjusted appropriately. In this way, a part of the area in the length direction of the rectangular window will not be within the coverage of the building, and the selected seed points will not be the points on the building. Moreover, since the rectangular window is a rectangular window with a short side and a relatively small overall area, more initial ground seed points can be obtained.

[0058] In addition, in some other embodiments, it is not excluded to use only the pending ground seed points retained in the last screening as the final ground seed points. Under this scheme, the initial TIN model established using these seed points may be difficult to cover all the point cloud data, and the final filtering accuracy is insufficient compared with this embodiment.

[0059] Step 4: Establish an initial triangular network based on the final ground seed points, calculate the repeated distance and repeated angle between the unclassified points and the initial triangular network, and determine the encrypted triangular network based on the calculation results of the unclassified points and the preset repeated distance threshold and repeated angle threshold, and obtain the ground points; the unclassified points include the remaining point cloud data after extracting the pending ground seed points and the non-ground seed points.

[0060] In this step, during the process of determining the encrypted triangular network, it is determined whether the unclassified points need to be added as ground points according to the repeated distance and repeated angle of the unclassified points. If new ground points are added, the added ground points are incorporated into the final ground seed points, and the above steps are repeated until no new ground points are generated. The initial triangular network updated last time is used as the encrypted triangular network. Each node in the encrypted triangular network is a final ground seed point, and the final ground seed points are the ground points.

[0061] Specifically, the unclassified points include the discrete points excluded when extracting the initial ground seed points and the discrete points removed during the fitting process. The calculation methods of the repeated distance and repeated angle of the unclassified points are as follows. Each triangle of the initial triangular mesh encloses a triangular plane. Calculate the repeated distance and repeated angle formed by the unclassified points in each triangle of the initial triangular mesh and the corresponding triangular plane. Compare the calculated repeated distance and repeated angle with the preset repeated distance threshold and repeated angle threshold respectively. When the repeated distance is less than the repeated distance threshold and the repeated angle is less than the repeated angle threshold, add the unclassified point as a ground point. Then incorporate the newly added ground points into the final ground seed points, and obtain an updated initial triangular mesh based on the updated final ground seed points. Then repeat the above steps until no new ground points are generated. To achieve the purpose of accurately classifying all discrete points in the point cloud data.

[0062] To more clearly illustrate the present invention and its advantages, the following will further explain the method provided in Embodiment 1 of the present invention in combination with specific examples and their related drawings.

[0063] This example uses Dataset 1 and Dataset 2 for filtering experiments: (1) Dataset 1 is a dataset publicly released by the International Society for Photogrammetry and Remote Sensing (ISPRS). This Dataset 1 has a total of 15 samples, which contain various terrain features. The specific terrain features are shown in Table 1. (2) Dataset 2 is a part of the campus of Anhui University of Science and Technology. Dataset 2 was obtained by a DJI M350 drone equipped with an Easy Scan X10 camera system. The terrain in this area is relatively flat, including large buildings, vegetation, and small ground objects such as cars. The area of the region is 200m×250m, and the point cloud density is 11.2 points / m². The real point cloud map is as Figure 3 shown. To facilitate the subsequent evaluation of the experimental results, the Terrasolid software is used to manually classify Dataset 2, and all point clouds are divided into two categories: ground points and object points. The Terrasolid software is the first set of commercial lidar data processing software.

[0064] Table 1 Sample Terrain Features of Dataset 1

[0065]

[0066] This example uses the filtering error evaluation criteria provided by the International Society for Photogrammetry and Remote Sensing to quantitatively evaluate the filtering results of Dataset 1 and Dataset 2, including the first type of error (T1), the second type of error (T2), the total error (T), and the Kappa coefficient. The calculations of the three errors and the Kappa coefficient are as shown in Equations (1)-(4):

[0067]

[0068] Among them, a represents the correctly classified ground points, b represents the misclassified points where ground points are classified as non-ground points, c represents the misclassified points where non-ground points are classified as ground points, and d represents the correctly classified non-ground points. The specific parameters are shown in Table 2.

[0069] Table 2 Definition of filtering error

[0070]

[0071] The analysis of the above experimental results is as follows:

[0072] (1) Analysis of dataset 1

[0073] Considering the terrain features of each sample and through a large number of experiments, the parameters involved in the IPTD algorithm and the PTD algorithm are optimized. As shown in Table 3, for the IPTD algorithm, the length of the rectangular window is selected in the same way as the size of the square grid in the PTD algorithm. The width of the rectangular window is about 1 / 10 of the grid length, and the sliding step s is taken as 1 / 3 of the grid length of each sample. When removing noise points by fitting adjacent seed points, if only 1 fitting is performed, when the terrain fluctuates greatly, some noise points cannot be removed; if 3 fittings are performed, a large number of real ground seed points will be removed while removing noise points. Considering comprehensively, 2 surface fittings are performed in this example. The number of seed points used for fitting twice is 24 and 12 respectively. Due to the different terrain features of each sample, the height difference threshold h varies slightly.

[0074] The two algorithms are respectively used to filter the 15 sample data in dataset 1, and the number of extracted seed points and the accuracy of seed points are as Figure 4 shown. The column features in the figure are used to represent the number of seed points, and the line features are used to represent the extraction accuracy of seed points.

[0075] It can be seen from the figure that the number of seed points extracted by the IPTD algorithm in the 15 sample data has been greatly improved compared with the PTD algorithm. Among them, the number of seed points in samples s51, s52, s53, and s61 all exceed 2000, and the extraction accuracy of seed points in the 15 samples has also been improved to varying degrees, all reaching more than 97%. Among them, samples s12, s31, and s41 have obvious improvements. Referring to the terrain features of the samples in Table 1, it shows that the algorithm proposed in Example 1 can well remove the low-value noise points and some small-scale ground objects existing in the seed points, and can more effectively extract ground seed points in some areas with simple ground object types and flat terrain, and ensure the accuracy of seed points while increasing the number of seed points.

[0076] To more intuitively compare the filtering effects of the two algorithms, taking sample s23 as an example, the filtering results of the two algorithms are analyzed.

[0077] Figures 5 - 8 The initial TIN models of the two filtering algorithms are shown. It can be seen from the figure that the number of seed points obtained by the PTD algorithm is small, resulting in a large difference between the initial TIN model and the real ground. For the IPTD algorithm, due to the addition of a large number of ground seed points, the initial TIN models obtained after 1 - time fitting and 2 - time fitting have high simulation accuracy for the real ground. However, there are obvious noise points in the initial TIN model established after 1 - time surface fitting. In contrast, the initial TIN model established after 2 - time surface fitting has higher simulation accuracy for the real ground.

[0078] The filtering results of sample s23 are as Figures 9 - 12 shown. From the two groups of misclassified point clouds Figure 10 and 12 it can be seen that there are a large number of error points at the building edges for both algorithms. However, the number of error points in Figure 10 is significantly more than that in Figure 12 . On the one hand, it is because the PTD algorithm is difficult to correctly classify the point clouds not included in the initial TIN model. On the other hand, the increase in seed points enables the IPTD algorithm to better classify some ground objects. Combining Figure 5 it can be seen that the initial TIN model established by the IPTD algorithm can better simulate the terrain, and thus can improve the filtering accuracy of the algorithm.

[0079] According to the error evaluation criteria given by the International Society for Photogrammetry, the statistical results of the filtering errors of 15 sample data in dataset 1 under the two filtering methods are shown in Table 3.

[0080] Table 3 Parameters involved in the IPTD algorithm and filtering results of dataset 1

[0081]

[0082]

[0083] As can be seen from Table 3, the total error of the IPTD algorithm proposed in Example 1 in all samples is less than that of the traditional PTD algorithm. The average value is reduced from 11.7% to 4.52%. The improvement effects of the filtering errors in different regions are different, which is related to the complexity of the terrain. Among them, the regions of samples s11, s24, and s53 are relatively complex and there are steep slopes. Although a large number of seed points are added, it is still impossible to correctly determine the slopes and the low vegetation on the slopes, and there are large errors in the filtering results. The terrain of samples s21, s31, and s42 is relatively flat compared with other regions, and the ground objects are relatively simple. After adding seed points, the filtering accuracy of the algorithm is improved to a certain extent, and the total error is less than 2%. In addition, the type-I errors of all samples have decreased to varying degrees. Although the type-II errors of some samples have increased, generally speaking, the IPTD algorithm proposed in Example 1 is superior to the classical PTD algorithm.

[0084] (2) Analysis of Dataset 2

[0085] Table 4 Statistical Results of Filtering for Dataset 2

[0086]

[0087] Compared with Dataset 1, Dataset 2 has a greater point cloud density and more complex ground objects, which can better reflect the filtering effect. The two algorithms are respectively used to filter Dataset 2, and the results are as Figure 13 shown. It can be seen from the figure that both algorithms can filter out most of the ground objects in the non-building area, and the filtering effects are good. However, in the building area, due to the influence of low vegetation and steps, the PTD algorithm does not filter out some low eaves, as shown in the area marked by the red frame in the figure, while the IPTD algorithm of Example 1 can filter it out better. In addition, by calculating the filtering error, a quantitative analysis of the filtering results is carried out. As shown in Table 4, compared with the PTD algorithm, the improved algorithm can extract more seed points, and the number increases from 46 to 425, and it can ensure the accuracy of the seed points. On this basis, both types of errors are reduced, and the total error is also reduced from the original 1.39% to 0.93%. Generally speaking, the IPTD algorithm proposed in Example 1 is superior to the classical PTD algorithm.

[0088] To better illustrate the filtering effect of the IPTD algorithm, 15 samples in dataset 1 are classified into four terrains according to the terrain, namely flat ground, flat ground with gentle slope, flat ground with steep slope, and steep slope, and into two scenes according to the ground objects, namely urban and forest. The filtering results of the IPTD algorithm are compared with those of several classical algorithms, including a multi-scale adaptive slope filtering algorithm proposed in Document 1, a fast progressive triangulation network filtering algorithm using adjacent surface information proposed in Document 2, and the Wack algorithm, Elmqvist algorithm, Sohn algorithm, and Pfeifer algorithm announced by the International Society for Photogrammetry and Remote Sensing involved in Document 1. Among them, Document 1 is a document titled "A multi-scale adaptive point cloud slope filtering algorithm" published by Wang Wenqi et al. in Volume 47 of the Journal of Wuhan University in 2022, and Document 2 is a document titled "A fast progressive TIN densification filtering algorithm for airborne LiDAR data using adjacent surface information" published by Li H et al. in the journal IEEE J-STARS in 2021. The total filtering errors of several filtering algorithms are shown in Table 5.

[0089]

[0090]

[0091] Note: The samples marked with * are forests.

[0092] As can be seen from Table 5: The Wack algorithm and Elmqvist algorithm have better filtering effects in urban areas, but poorer filtering effects in areas with large terrain slopes and forest areas. The Sohn algorithm and Pfeifer algorithm both start from the whole and use the overall information of the point cloud for filtering, with good filtering effects and good algorithm stability in various terrains and scenes. Document 1 uses multi-scale grids and adaptively determines the slope threshold for filtering, which can achieve adaptive filtering and has achieved good filtering effects in each sample area. The average filtering error is only lower than that of triangulation network filtering. Especially in the steep slope area, the filtering accuracy is higher than that of other algorithms. Document 2 uses adjacent surface information to block-filter the point cloud, which can greatly improve the filtering speed and has achieved good filtering effects in various terrains and scenes, but ignores the overall information of the point cloud, resulting in a lower filtering accuracy than the IPTD algorithm. The average filtering error of the IPTD algorithm is less than that of the other several algorithms, and in various terrains and scenes, the filtering accuracy of most samples is higher than that of other algorithms, indicating that the IPTD algorithm has good applicability and can achieve good filtering effects in various terrains and scenes.

[0093] In summary, aiming at the problem that the classic progressive triangular mesh filtering algorithm has a small number of initial seed points and noise points when selecting initial seed points, the IPTD algorithm proposes an improved method for selecting seed points by moving windows, and after obtaining the initial seed points, uses a cubic surface function to remove noise points through multi-scale surface fitting. Experiments show that this method can greatly increase the number of initial seed points, improve the extraction accuracy of seed points, enable the initial TIN model to better simulate the terrain, and can improve the filtering accuracy of point clouds to a certain extent in different terrain areas during subsequent filtering. By comparing with several typical filtering algorithms, the stability of the IPTD algorithm is further illustrated.

[0094] Embodiment 2

[0095] This embodiment relates to a point cloud data filtering system based on a progressive triangular mesh. This system is used to implement the point cloud data filtering method disclosed in Embodiment 1. The point cloud data filtering system includes:

[0096] A point cloud data projection module, used to project point cloud data onto the same plane;

[0097] A seed point extraction module, used to move a rectangular window within the plane to make the rectangular window traverse all point cloud data to extract pending ground seed points;

[0098] A seed point screening module, used to fit the pending ground seed points with a cubic surface fitting function to obtain a surface, perform height difference judgment on the surface, screen out non-ground seed points to update the pending seed points, and form final ground seed points;

[0099] A ground point acquisition module, used to establish an initial triangular mesh based on the final ground seed points, calculate the repeated distance and repeated angle between the unclassified points and the initial triangular mesh, and determine the encrypted triangular mesh according to the relationship between the calculation results and the preset repeated distance threshold and repeated angle threshold, and obtain ground points.

[0100] Embodiment 3

[0101] This embodiment relates to a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, each step in the point cloud data filtering method of Embodiment 1 can be implemented.

[0102] The above embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it cannot be understood as a limitation to the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A point cloud data filtering method based on progressive triangulation, characterized in that: The following steps are involved: Step 1: Project the point cloud data onto the same plane; Step 2: moving the rectangular window in the plane so that the rectangular window traverses all point cloud data to extract undetermined ground seed points; Step 3: Fit the undetermined ground seed points using a cubic surface fitting function to obtain a surface, perform height difference judgment on the surface, screen out non-ground seed points, update the undetermined seed points, and form final ground seed points; Step 4: Establish an initial triangulated network based on the final ground seed points, calculate the repetition distance and repetition angle between the unclassified points and the initial triangulated network, determine the encrypted triangulated network based on the relationship between the calculation results and the preset repetition distance threshold and repetition angle threshold, and obtain the ground points; the unclassified points include the point cloud data remaining after extracting the pending ground seed points and the non-ground seed points.

2. The point cloud data filtering method according to claim 1, characterized in that: In the step 2, the lowest point in the rectangular window is used as the initial ground seed point, and the initial ground seed point obtained during the traversal process is used as the pending ground seed point.

3. The point cloud data filtering method according to claim 1, characterized in that: The movement of the rectangular window includes sliding and jumping. The rectangular window is divided into window one and window two. Both window one and window two traverse all point cloud data. During the traversal process, the sliding direction of window two is different from that of window one, and the jumping direction of window two is different from that of window one.

4. The point cloud data filtering method according to claim 1, characterized in that: The step three comprises: All undetermined ground seed points are screened using the cubic surface fitting function, and the n remaining undetermined ground seed points around each undetermined ground seed point are found using the KNN nearest neighbor method; Taking the remaining n undetermined ground seed points as reference points, taking the undetermined ground seed points surrounded by the reference points as test points, and establishing a surface based on the n reference points and the cubic surface fitting function; When the height difference between the test point and the curved surface is greater than or equal to a preset height difference threshold, non-ground seed points are screened out to update the pending seed points to form final ground seed points.

5. The point cloud data filtering method according to claim 4, characterized in that: In step 3, all pending ground seed points are screened m times using the cubic surface fitting function, and the pending ground seed points retained in the last screening are used as the final ground seed points. The height difference thresholds of each screening are h1, h2, ..., h m , and the height difference threshold of each screening decreases successively.

6. The point cloud data filtering method according to claim 4, characterized in that: In the step three, all pending ground seed points are screened m times using a cubic surface fitting function, and the pending ground seed points and boundary points retained after the last screening are used as final ground seed points.

7. The point cloud data filtering method according to claim 6, characterized in that: The plane for point cloud data projection is a rectangular plane, the boundary points are new points obtained by replacing the four vertices of the rectangular plane with their elevations, and the elevations of the new points are the elevations of the undetermined ground seed points closest to the vertices of the rectangular plane.

8. The point cloud data filtering method according to claim 1, characterized in that: When determining the encrypted triangulated network, if an unclassified point is added as a ground point, the acquired ground point is included in the final ground seed point, and step 4 is repeated until no new ground point is generated.

9. A point cloud data filtering system for implementing the method according to any one of claims 1 to 8, characterized in that: include: Point cloud data projection module, used to project point cloud data into the same plane; A seed point extraction module is used to slide a rectangular window in the plane so that the rectangular window traverses all point cloud data and extracts initial ground seed points from the point cloud data in the rectangular window; A seed point screening module is used to screen the initial ground seed points using a cubic surface fitting function, remove non-ground seed points, and obtain final ground seed points; The ground point acquisition module is used to establish an initial triangulated network based on the final ground seed point, calculate the repetition distance and repetition angle between the unclassified point and the initial triangulated network, determine the encrypted triangulated network according to the relationship between the calculation result and the preset repetition distance threshold and repetition angle threshold, and obtain the ground point.

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

Images