Method and system for filtering point cloud in thick vegetation coverage area based on implicit neural representation

Abstract

The application discloses a kind of based on implicit neural representation's thick vegetation cover area point cloud filtering method and system, belong to point cloud filtering technical field, including obtaining point cloud data, according to point cloud data to establish adaptive irregular grid, based on adaptive irregular grid to point cloud data is segmented, obtain the point cloud data of multiple sub-regional;Each sub-regional point cloud data is input in the upsampling framework based on implicit neural network, obtain global point cloud data;Global point cloud data is filtered based on multi-scale feature point cloud filtering algorithm, and output global point cloud ground point set.Through setting adaptive irregular grid, the density of grid is automatically adjusted, to realize the fine segmentation of thick vegetation cover area point cloud data;Using the upsampling framework based on implicit neural network, the sparsity of point cloud is solved;Global point cloud data is filtered using multi-scale feature point cloud filtering algorithm, and the filtering effect of ground object and vegetation at all levels is improved.

Description

Technical Field

[0001] This invention belongs to the field of point cloud filtering technology, specifically relating to a point cloud filtering method and system for thick vegetation cover areas based on implicit neural representation. Background Technology

[0002] Point cloud data, as a set of discrete 3D points directly derived from real-world scenes, aims to accurately capture and reshape the 3D form of objects. However, due to the disordered nature and other characteristics of point cloud data, it poses significant challenges to the subsequent construction of Digital Elevation Models (DEMs) and the 3D reconstruction process. In processing raw point clouds, numerous scholars have conducted in-depth explorations of point cloud upsampling and filtering methods. How to generate dense and uniform point clouds without outliers from sparse point clouds through sampling and filtering has become a research focus in recent years. While various 3D point cloud processing devices have developed rapidly in recent years, achieving good results in sampling and filtering, their accuracy remains limited due to the sparseness of sampled point clouds and the less-than-ideal filtering performance in complex scene environments.

[0003] Traditional upsampling methods include interpolation upsampling and end-to-end upsampling. Interpolation upsampling includes nearest neighbor interpolation, bilinear interpolation, and bicubic interpolation. End-to-end upsampling includes self-supervised point cloud upsampling and arbitrary-scale upsampling. However, these methods also have their own problems. Nearest neighbor interpolation is fast, but it can result in jagged image edges. Bilinear interpolation is smoother than nearest neighbor interpolation, but it requires more computation. Bicubic interpolation produces smoother and more detailed images, but it also requires more computation. Self-supervised point cloud upsampling and arbitrary-scale upsampling are still limited to a fixed upsampling factor.

[0004] Commonly used filtering methods include difference-based filtering, slope-based filtering, and segmentation-based filtering. However, while these methods may perform well in specific scenarios, they are not ideal for handling point cloud data in complex scenarios.

[0005] Therefore, how to provide an effective technical solution to address the problems of sparse point clouds obtained from sampling and the less-than-ideal filtering effect in complex scene environments has become an urgent problem to be solved in existing technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a point cloud filtering method and system for thick vegetation cover areas based on implicit neural representation, so as to solve the above-mentioned problems existing in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a point cloud filtering method for thick vegetation cover areas based on implicit neural representation, comprising:

[0009] Acquire point cloud data, build an adaptive irregular grid based on the point cloud data, and segment the point cloud data based on the adaptive irregular grid to obtain point cloud data of multiple sub-regions;

[0010] The point cloud data of each sub-region is input into an upsampling framework based on an implicit neural network to obtain global point cloud data;

[0011] A multi-scale feature point cloud filtering algorithm is used to filter global point cloud data and output a global point cloud ground point set.

[0012] In one possible design, point cloud data is acquired, an adaptive irregular mesh is constructed based on the point cloud data, and the point cloud data is segmented based on the adaptive irregular mesh to obtain point cloud data for multiple sub-regions, including:

[0013] Acquire point cloud data, traverse the point cloud data, and obtain the spatial location and density of the point cloud data;

[0014] Construct an adaptive irregular mesh based on point cloud density, wherein the adaptive irregular mesh based on point cloud density includes multiple sub-region meshes;

[0015] The point cloud data is segmented using an adaptive irregular grid based on point cloud density, and the point cloud data is assigned to the corresponding sub-region grid according to the spatial location of the point cloud data, resulting in point cloud data of multiple sub-regions.

[0016] In one possible design, after obtaining point cloud data for multiple sub-regions, the following steps are also included:

[0017] Step 1: Traverse the point cloud data of each sub-region to obtain the distribution of the point cloud data in each sub-region. The distribution includes the point cloud data in each sub-region and the boundary point cloud data on the boundary of the sub-region. Filter out the outliers in the boundary point cloud data on the boundary of the sub-region based on the consistency of the normal vector to obtain the boundary point cloud.

[0018] Step 2: Establish a secondary subdivision grid. Divide the sub-regions with boundary point clouds into grids according to the secondary subdivision grid. Calculate the distance from each boundary point cloud to the centroid of the adjacent secondary subdivision grid. Perform inverse distance weighting on the adjacent secondary subdivision grids according to the distance. The adjacent secondary subdivision grids vote on each boundary point cloud according to the weight to obtain the first weight voting result. Assign each boundary point cloud to the secondary subdivision grid with the highest first weight voting result.

[0019] Step 3: Iterate through the point cloud data in the second-level subdivision grid again. If there are boundary point clouds in the second-level subdivision grid, establish a third-level subdivision grid. Divide the second-level subdivision grid with boundary point clouds according to the third-level subdivision grid. Calculate the distance from each boundary point cloud to the centroid of the adjacent third-level subdivision grid. Perform inverse distance weighting on the adjacent third-level subdivision grids according to the distance. The adjacent third-level subdivision grids vote on each boundary point cloud according to the weight to obtain the second weight voting result. Assign each boundary point cloud to the third-level subdivision grid with the highest second weight voting result.

[0020] Repeat the above steps until there are no boundary points on the grid boundary.

[0021] In one possible design, the implicit neural network-based upsampling framework includes an encoder and a decoder, the encoder including a position encoder, and the decoder including a feature extractor and an upsampling module; the point cloud data of each sub-region includes point cloud coordinate data; the point cloud data of each sub-region is input into the implicit neural network-based upsampling framework to obtain global point cloud data, including:

[0022] The point cloud data of each sub-region is input into the position encoder. The point cloud data of each sub-region is preprocessed to obtain the encoding information of each sub-region. The encoding information of each sub-region includes position attributes and color attributes. The point cloud coordinate data is mapped to the corresponding position attributes in a high-dimensional space using an implicit neural function, and the high-dimensional feature vector of each sub-region is output.

[0023] The feature extractor is used to learn and extract the high-dimensional feature vectors of each sub-region to obtain the terrain geometric features of each sub-region. The terrain geometric features of each sub-region are then input into the upsampling module.

[0024] In the upsampling module, the local features of the missing holes in the point cloud data of each sub-region are predicted based on the terrain geometric features of each sub-region to obtain the prediction results. Based on the prediction results, the missing holes in the point cloud are filled to obtain the filled point cloud of each sub-region. Based on the filled point cloud of each sub-region, the upsampling is performed by setting different upsampling ratios for different point cloud densities of different sub-regions, using the principle of stepped sampling number to generate high-resolution uniform density point clouds.

[0025] Global point cloud data is obtained by stitching together point clouds based on high-resolution, uniformly dense point clouds.

[0026] In one possible design, local features of missing holes in the point cloud data of each sub-region are predicted based on the terrain geometry features of each sub-region, resulting in prediction results. The missing holes in the point cloud are then filled in based on these prediction results, yielding the filled point clouds for each sub-region, including:

[0027] Based on point cloud coordinate data, implicit neural functions are used to predict the density distribution of missing holes in the point cloud data of each sub-region, and new point clouds are generated in the missing holes based on a preset density threshold.

[0028] For the new point cloud location offset, the new point cloud is corrected based on the location offset to obtain the completed point cloud for each sub-region.

[0029] In one possible design, based on point cloud coordinate data, the calculation expression for predicting the density distribution of missing holes in the point cloud data of each sub-region using an implicit neural function is as follows:

[0030] ;

[0031] In the formula, Point cloud coordinate data, This represents the position of a point in the point cloud along the x-axis. This represents the position of a point in the point cloud along the y-axis. This represents the position of a point in the point cloud along the z-axis. It is an implicit function. The density of point cloud data for each sub-region, This represents the distance between point clouds and adjacent point clouds in point cloud data where holes are missing in each sub-region along the x-axis. This represents the distance between point clouds and adjacent point clouds in point cloud data for point cloud data with missing holes in each sub-region along the y-axis direction. This represents the distance between point clouds and adjacent point clouds in the point cloud data of each sub-region with missing holes along the z-axis. For each sub-region, the density distribution value of the missing holes in the point cloud data is predicted. This is the activation function.

[0032] In one possible design, the global point cloud data includes a set of ground points; the global point cloud data is filtered based on a multi-scale, multi-feature point cloud filtering algorithm to output a global point cloud ground point set, including:

[0033] The spherical neighborhood method is used to extract multiple features from global point cloud data to obtain multi-feature data, which includes color information, roughness, and covariance-based features. The covariance-based features include linearity, verticality, flatness, and total variance features.

[0034] A multi-feature fusion method based on spherical neighborhood is used to fuse global point cloud data and multi-feature data to obtain a fused feature point set.

[0035] Elevation information of ground point set is obtained by using large-scale grid and then corrected by small-scale grid to obtain corrected elevation information. The fused feature point set is then corrected based on the corrected elevation information to obtain the corrected fused feature point set.

[0036] A low-resolution DSM model and a high-resolution DSM model are generated based on the corrected fused feature point set. An interpolation model is constructed based on the low-resolution DSM model and the high-resolution DSM model. The degree of elevation change between the low-resolution DSM model and the high-resolution DSM model is calculated based on the interpolation model. The mean and standard deviation of the degree of elevation change are calculated. The elevation variation coefficient of the DSM model is calculated based on the mean and standard deviation of the degree of elevation change. The elevation difference threshold is determined based on the elevation variation coefficient of the DSM model.

[0037] The global point cloud data is filtered out based on the elevation difference threshold, and the global point cloud ground point set is output.

[0038] In one possible design, the expressions for calculating the degree of elevation change, the coefficient of elevation variation, and the elevation difference threshold between the low-resolution DSM model and the high-resolution DSM model are as follows:

[0039] ;

[0040] In the formula, This represents the degree of elevation variation between the low-resolution and high-resolution DSM models. It is a high-resolution DSM model. This is a low-resolution DSM model. The elevation variation coefficient of the DSM model. The standard deviation is obtained based on the degree of elevation change. This is the average value obtained based on the degree of elevation change. The height difference threshold, and This is an empirical coefficient. The default value is 0.5. The default value is 0.1.

[0041] Secondly, the present invention provides a point cloud filtering system for thick vegetation cover areas based on implicit neural representation, comprising:

[0042] The acquisition and segmentation module is used to acquire point cloud data, build an adaptive irregular grid based on the point cloud data, and segment the point cloud data based on the adaptive irregular grid to obtain point cloud data of multiple sub-regions;

[0043] The implicit neural upsampling module is used to input the point cloud data of each sub-region into the upsampling framework based on the implicit neural network to obtain the global point cloud data.

[0044] The multi-scale feature point cloud filtering module is used to filter global point cloud data based on a multi-scale feature point cloud filtering algorithm and output a global point cloud ground point set.

[0045] Thirdly, the present invention provides a computer program product containing instructions that, when the instructions are executed on a computer, cause the computer to perform the point cloud filtering method for thick vegetation cover areas based on implicit neural representation as described in the first aspect above.

[0046] The beneficial effects of this invention are as follows:

[0047] This invention discloses a point cloud filtering method and system for thick vegetation cover areas based on implicit neural representation. The method includes acquiring point cloud data, establishing an adaptive irregular grid based on the point cloud data, segmenting the point cloud data based on the adaptive irregular grid to obtain point cloud data for multiple sub-regions, inputting the point cloud data of each sub-region into an upsampling framework based on an implicit neural network to obtain global point cloud data, and filtering the global point cloud data using a multi-scale feature point cloud filtering algorithm to output a global point cloud ground point set. This invention uses an adaptive irregular grid to process point cloud data from thick vegetation cover areas and automatically adjusts the grid density according to the scale and density characteristics of the input point cloud data, thereby achieving fine segmentation of point cloud data from thick vegetation cover areas. The upsampling framework based on an implicit neural network can reconstruct ground features and complete missing data in the point cloud, and differential upsampling addresses the sparsity of the point cloud. The multi-scale feature point cloud filtering algorithm filters the global point cloud data, improving the filtering effect on ground features and vegetation at various levels. Attached Figure Description

[0048] Figure 1 A flowchart illustrating a point cloud filtering method for thick vegetation cover areas based on implicit neural representation, provided as a first aspect of this embodiment;

[0049] Figure 2 This is a block diagram of a point cloud filtering system for thick vegetation cover areas based on implicit neural representation, provided as a second aspect of this embodiment. Detailed Implementation

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0051] It should be understood that although the terms "first" and "second" may be used herein to describe various objects, these objects should not be limited by these terms. These terms are only used to distinguish one object from another. For example, the first object may be referred to as the second object, and similarly, the second object may be referred to as the first object, without departing from the scope of the exemplary embodiments of the invention.

[0052] Example:

[0053] like Figure 1 As shown, the first aspect of this embodiment provides a point cloud filtering method for thick vegetation cover areas based on implicit neural representation. This method can be executed, but is not limited to, by a computer device or virtual machine with certain computing resources, such as a personal computer or smartphone, or by a virtual machine. Figure 1 As shown, the point cloud filtering method for thick vegetation cover areas based on implicit neural representation may include, but is not limited to, the following steps:

[0054] S1. Acquire point cloud data, establish an adaptive irregular grid based on the point cloud data, and segment the point cloud data based on the adaptive irregular grid to obtain point cloud data of multiple sub-regions;

[0055] Specifically, point cloud data can be obtained through airborne LiDAR (Light Detection and Ranging), which is a space measurement system that integrates modern laser scanning technology, computer technology, high dynamic carrier attitude measurement technology, and high-precision dynamic GPS (Global Positioning System) differential positioning technology.

[0056] Specifically, in step S1, point cloud data is acquired, an adaptive irregular mesh is established based on the point cloud data, and the point cloud data is segmented based on the adaptive irregular mesh to obtain point cloud data for multiple sub-regions, including:

[0057] S101. Acquire point cloud data, traverse the point cloud data, and obtain the spatial location and density of the point cloud data;

[0058] S102. Construct an adaptive irregular mesh based on point cloud density, wherein the adaptive irregular mesh based on point cloud density includes multiple sub-region meshes;

[0059] S103. The point cloud data is segmented using an adaptive irregular grid based on point cloud density, and the point cloud data is assigned to the corresponding sub-region grid according to the spatial location of the point cloud data, resulting in point cloud data of multiple sub-regions.

[0060] In this embodiment, a dictionary is also established to record point cloud data information of sub-regions.

[0061] Furthermore, after obtaining point cloud data for multiple sub-regions, the following steps are also included:

[0062] S104. Traverse the point cloud data of each sub-region to obtain the distribution of the point cloud data of each sub-region. The distribution includes the point cloud data in each sub-region and the boundary point cloud data on the boundary of the sub-region. Filter out the outliers in the boundary point cloud data on the boundary of the sub-region based on the consistency of the normal vector to obtain the boundary point cloud.

[0063] Specifically, the role of normal vector consistency is to identify and filter noise points or outliers. In three-dimensional space, each point cloud can be calculated to have a normal vector, which is perpendicular to the surface where the point is located. The basic idea is that for a smooth surface region, the normal vectors of neighboring points should be similar. If the normal vector of a point differs greatly from the normal vectors of its neighboring points, then this point may not be located on the same surface and can be identified as an outlier or noise. Therefore, outliers are filtered out to improve the accuracy of the data.

[0064] S105. Establish a two-level subdivision grid, divide the sub-region with boundary point clouds into grids according to the two-level subdivision grid, calculate the distance from each boundary point cloud to the centroid of the adjacent two-level subdivision grid, perform inverse distance weighting on the adjacent two-level subdivision grids according to the distance, and the adjacent two-level subdivision grids vote on each boundary point cloud according to the weight to obtain the first weight voting result, and assign each boundary point cloud to the two-level subdivision grid with the highest first weight voting result;

[0065] Specifically, inverse distance weighting reflects the inverse proportional relationship between weight and distance; voting decision is a voting method that incorporates weights, equivalent to the weighted average method. The number of votes obtained after voting is multiplied by the weight to obtain the weighted vote count. Then, all the weighted vote counts obtained from each boundary point cloud are summed to obtain the first weighted voting result.

[0066] S106. Iterate through the point cloud data in the second-level subdivision grid again. If there are boundary point clouds in the second-level subdivision grid, establish a third-level subdivision grid. Divide the second-level subdivision grid with boundary point clouds according to the third-level subdivision grid. Calculate the distance from each boundary point cloud to the centroid of the adjacent third-level subdivision grid. Perform inverse distance weighting on the adjacent third-level subdivision grids according to the distance. The adjacent third-level subdivision grids vote on each boundary point cloud according to the weight to obtain the second weight voting result. Assign each boundary point cloud to the third-level subdivision grid with the highest second weight voting result.

[0067] Specifically, both the first-weighted voting result and the second-weighted voting result are voting scores, and each boundary point cloud is assigned to the subdivision grid with the highest voting score.

[0068] S107. Repeat steps S104 to S106 until there are no boundary point clouds on the grid boundary.

[0069] Specifically, while performing mesh generation, a maximum recursion depth is set. When performing recursive mesh generation, the mesh level is less than or equal to the maximum recursion depth.

[0070] Furthermore, steps S105 to S107 only consider the case where the boundary point cloud is located in two adjacent grid regions. In this embodiment, preferably, when the boundary point cloud is located in a grid region where three or four adjacent grids intersect, the assignment of the boundary point cloud includes: calculating the distance from the boundary point cloud to the centroid of the adjacent grid, and performing inverse distance weighting on the distance to obtain a weight value, obtaining a multi-dimensional weight based on the weight value and feature similarity, normalizing the multi-dimensional weight, and selecting the grid with the largest multi-dimensional weight value as the assignment based on a preset judgment threshold.

[0071] S2. Input the point cloud data of each sub-region into an upsampling framework based on an implicit neural network to obtain global point cloud data;

[0072] Furthermore, the implicit neural network-based upsampling framework includes an encoder, a decoder, an MLP (Multi-Layer Perceptron) layer, and an output layer. The encoder includes a position encoder, and the decoder includes a Transformer-based feature extractor and an upsampling module. The point cloud data for each sub-region includes point cloud coordinate data.

[0073] Specifically, in step S2, the point cloud data of each sub-region is input into an upsampling framework based on an implicit neural network to obtain global point cloud data, including:

[0074] S201. Input the point cloud data of each sub-region into the position encoder, preprocess the point cloud data of each sub-region to obtain the encoding information of each sub-region, wherein the encoding information of each sub-region includes position attributes; use implicit neural functions to map the point cloud coordinate data to the corresponding position attributes in high-dimensional space, and output the high-dimensional feature vector of each sub-region.

[0075] In this embodiment, the implicit neural function serves as a continuous representation of terrain geometry, enabling it to capture complex spatial relationships in sparse point clouds and output high-dimensional feature vectors for each sub-region, thereby enhancing the interpretability of the data and the accuracy and efficiency of the analysis. The encoded information for each sub-region includes, but is not limited to, location attributes and color attributes.

[0076] S202. Use a feature extractor to learn and extract the high-dimensional feature vectors of each sub-region to obtain the terrain geometric features of each sub-region, and input the terrain geometric features of each sub-region into the upsampling module;

[0077] Furthermore, the feature extractor is built based on Transformer.

[0078] S203. In the upsampling module, the local features of the missing holes in the point cloud data of each sub-region are predicted based on the terrain geometric features of each sub-region to obtain the prediction results. Based on the prediction results, the missing holes in the point cloud are filled to obtain the filled point cloud of each sub-region. Based on the filled point cloud of each sub-region, the upsampling ratio of different regions is set according to the principle of stepped sampling number for different point cloud densities and point cloud scales to generate high-resolution uniform density point clouds.

[0079] In this embodiment, the missing parts are predicted by using terrain geometric features, and the missing parts are filled in based on the prediction results, thus avoiding feature loss at a single scale.

[0080] Specifically, in step S203, based on the terrain geometric features of each sub-region, the local features of the missing holes in the point cloud data of each sub-region are predicted to obtain the prediction results. Based on the prediction results, the missing holes in the point cloud are filled in to obtain the filled point clouds of each sub-region, including:

[0081] S2031. Based on point cloud coordinate data, use implicit neural functions to predict the density distribution of missing holes in the point cloud data of each sub-region, and generate new point clouds in the missing holes based on a preset density threshold.

[0082] Furthermore, based on point cloud coordinate data, the calculation expression for predicting the density distribution of missing holes in the point cloud data of each sub-region using implicit neural functions is as follows:

[0083] ;

[0084] In the formula, Point cloud coordinate data, This represents the position of a point in the point cloud along the x-axis. This represents the position of a point in the point cloud along the y-axis. This represents the position of a point in the point cloud along the z-axis. It is an implicit function. The density of point cloud data for each sub-region, This represents the distance between point clouds and adjacent point clouds in point cloud data where holes are missing in each sub-region along the x-axis. This represents the distance between point clouds and adjacent point clouds in point cloud data for point cloud data with missing holes in each sub-region along the y-axis direction. This represents the distance between point clouds and adjacent point clouds in the point cloud data of each sub-region with missing holes along the z-axis. For each sub-region, the density distribution value of the missing holes in the point cloud data is predicted. This is the activation function.

[0085] S2032. Based on the location offset of the new point cloud, correct the offset of the new point cloud to obtain the completed point cloud of each sub-region.

[0086] S204. Perform point cloud stitching based on high-resolution density uniform point cloud to obtain global point cloud data.

[0087] Specifically, the global point cloud data includes all three-dimensional information such as ground point sets, objects, buildings, vegetation, and terrain.

[0088] Furthermore, after the point cloud upsampling is completed, a Kd-tree is constructed, and an upsampling quality evaluation mechanism is established. The global point cloud data is compared with the initially obtained point cloud data. The upsampling quality evaluation mechanism can be selected from methods such as point cloud density evaluation and DTM model visual discrimination evaluation. Point cloud density evaluation refers to the quantitative analysis of the distribution density of points in the point cloud data; DTM model visual discrimination evaluation refers to the evaluation of the quality and accuracy of DTM through visual inspection. The above evaluation mechanisms are all existing technologies and will not be elaborated here.

[0089] S3. The global point cloud data is filtered based on a multi-scale, multi-feature point cloud filtering algorithm to output a global point cloud ground point set.

[0090] Specifically, in step S3, the global point cloud data is filtered based on a multi-scale, multi-feature point cloud filtering algorithm to output a global point cloud ground point set, including:

[0091] S301. The spherical neighborhood method is used to extract multiple features from the global point cloud data to obtain multi-feature data. The multi-feature data includes color information, roughness, and covariance-based features. The covariance-based features include linearity, verticality, flatness, and total variance features.

[0092] Specifically, the spherical neighborhood method is a method for defining neighborhoods in a metric space, primarily used to analyze the relationship between a point in the space and its surrounding points.

[0093] Furthermore, features such as verticality and flatness are added to the building feature extraction to improve the completeness of the extraction; features such as roughness and total variance are added to the classification of vegetation at different levels in complex mountainous areas to improve the accuracy of the classification.

[0094] S302. Use a multi-feature fusion method based on spherical neighborhood to fuse global point cloud data and multi-feature data to obtain a fused feature point set;

[0095] S303. Use a large-scale grid to obtain elevation information from the ground point set, and use a small-scale grid to correct the elevation information to obtain corrected elevation information. Correct the fused feature point set based on the corrected elevation information to obtain the corrected fused feature point set.

[0096] Specifically, after obtaining the fused point set, the fused point set is input into two-dimensional and three-dimensional grids and presented in multiple forms. Then, a large-scale grid is used to obtain the elevation information in the ground point set, and a small-scale grid is used to correct the elevation information to improve the authenticity and reliability of the obtained local details.

[0097] S304. Generate a low-resolution DSM model (Digital Surface Model) and a high-resolution DSM model based on the corrected fused feature point set. Construct a difference model based on the low-resolution DSM model and the high-resolution DSM model. Calculate the degree of elevation change between the low-resolution DSM model and the high-resolution DSM model based on the difference model. Calculate the average and standard deviation of the degree of elevation change. Calculate the elevation variation coefficient of the DSM model based on the average and standard deviation of the degree of elevation change. Calculate and determine the elevation difference threshold based on the elevation variation coefficient of the DSM model.

[0098] Furthermore, the calculation expressions for the degree of elevation change, the coefficient of elevation variation, and the elevation difference threshold between the low-resolution DSM model and the high-resolution DSM model are as follows:

[0099] ;

[0100] In the formula, This represents the degree of elevation variation between the low-resolution and high-resolution DSM models. It is a high-resolution DSM model. This is a low-resolution DSM model. The elevation variation coefficient of the DSM model. The standard deviation is obtained based on the degree of elevation change. This is the average value obtained based on the degree of elevation change. The height difference threshold, and This is an empirical coefficient. The default value is 0.5. The default value is 0.1.

[0101] S305. Filter out ground features in the global point cloud data based on the elevation difference threshold and output the global point cloud ground point set.

[0102] Furthermore, after outputting the global point cloud ground point set, a filtering performance evaluation system is established to assess the effectiveness of the filtering algorithm.

[0103] In this embodiment, a multi-scale feature point cloud filtering algorithm is preferably selected for filtering, but other commonly used filtering algorithms can also be selected. Its application scenarios are not limited to the thick vegetation cover area described in this article, but can be extended to more complex scenarios.

[0104] This embodiment discloses a point cloud filtering method for thick vegetation cover areas based on implicit neural representation, including acquiring point cloud data, establishing an adaptive irregular grid based on the point cloud data, segmenting the point cloud data based on the adaptive irregular grid to obtain point cloud data for multiple sub-regions, inputting the point cloud data of each sub-region into an upsampling framework based on an implicit neural network to obtain global point cloud data, and filtering the global point cloud data based on a multi-scale feature point cloud filtering algorithm to output a global point cloud ground point set. This embodiment employs an adaptive irregular grid to segment point cloud data, dynamically adjusting the density of the point cloud data to handle point cloud data from various complex scenarios, thereby achieving fine segmentation of the point cloud data. Within the implicit neural network-based upsampling framework, improved encoder-decoder results are used to extract ground feature data and complete the point cloud data. Furthermore, a preset upsampling coefficient is employed to better address density differences between different blocks, resolving the sparsity issue of the point cloud. Simultaneously, features such as verticality and flatness are added during the filtering process, improving the completeness of building extraction. The addition of roughness and total variance features enhances the classification effect of vegetation at different levels in mountainous areas.

[0105] like Figure 2As shown, the second aspect of this embodiment provides a point cloud filtering system for thick vegetation cover areas based on implicit neural representation, comprising:

[0106] The acquisition and segmentation module is used to acquire point cloud data, build an adaptive irregular grid based on the point cloud data, and segment the point cloud data based on the adaptive irregular grid to obtain point cloud data of multiple sub-regions;

[0107] The implicit neural upsampling module is used to input the point cloud data of each sub-region into the upsampling framework based on the implicit neural network to obtain the global point cloud data.

[0108] The multi-scale feature point cloud filtering module is used to filter global point cloud data based on a multi-scale feature point cloud filtering algorithm and output a global point cloud ground point set.

[0109] This embodiment discloses a point cloud filtering system for thick vegetation cover areas based on implicit neural representation. A segmentation module acquires point cloud data, establishes an adaptive irregular grid based on the point cloud data, and segments the point cloud data into multiple sub-regions. An implicit neural upsampling module inputs the point cloud data of each sub-region into an upsampling framework based on an implicit neural network to obtain global point cloud data. A multi-scale feature point cloud filtering module uses a multi-scale feature point cloud filtering algorithm to filter the global point cloud data, outputting a global point cloud ground point set. The adaptive irregular grid can automatically adjust its density according to the scale and density characteristics of the input point cloud data, thereby achieving fine segmentation of point cloud data in thick vegetation cover areas. The use of a multi-scale feature point cloud filtering algorithm to filter the global point cloud data improves the filtering effect on ground features and vegetation at various levels.

[0110] The third aspect of this embodiment provides a computer program product, including a computer program or instructions, which, when executed by a computer, implement the point cloud filtering method for thick vegetation cover areas based on implicit neural representation as described in the first aspect.

[0111] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A point cloud filtering method for thick vegetation cover areas based on implicit neural representation, characterized in that, include: Acquire point cloud data, build an adaptive irregular grid based on the point cloud data, and segment the point cloud data based on the adaptive irregular grid to obtain point cloud data of multiple sub-regions; Point cloud data for each sub-region is input into an upsampling framework based on an implicit neural network to obtain global point cloud data. The upsampling framework includes an encoder and a decoder; the encoder includes a position encoder, and the decoder includes a feature extractor and an upsampling module. The point cloud data for each sub-region includes point cloud coordinate data. The process of inputting point cloud data for each sub-region into the upsampling framework to obtain global point cloud data includes: The point cloud data of each sub-region is input into the position encoder. The point cloud data of each sub-region is preprocessed to obtain the encoding information of each sub-region, which includes the position attribute. The implicit neural function is used to map the point cloud coordinate data to the corresponding position attribute in the high-dimensional space and output the high-dimensional feature vector of each sub-region. The feature extractor is used to learn and extract the high-dimensional feature vectors of each sub-region to obtain the terrain geometric features of each sub-region. The terrain geometric features of each sub-region are then input into the upsampling module. In the upsampling module, local features of missing holes in the point cloud data of each sub-region are predicted based on the terrain geometry features of each sub-region, obtaining prediction results. Based on the prediction results, the missing holes in the point cloud are filled in, resulting in the filled point clouds of each sub-region. Based on the filled point clouds of each sub-region, a stepped sampling principle is adopted, setting different upsampling ratios for different sub-region point cloud densities to generate high-resolution, uniformly dense point clouds. Specifically, the local features of missing holes in the point cloud data of each sub-region are predicted based on the terrain geometry features of each sub-region, obtaining prediction results. Based on the prediction results, the missing holes in the point cloud are filled in, resulting in the filled point clouds of each sub-region, including: Based on point cloud coordinate data, implicit neural functions are used to predict the density distribution of missing holes in the point cloud data of each sub-region, and new point clouds are generated in the missing holes based on a preset density threshold. For the new point cloud location offset, the new point cloud is corrected based on the location offset to obtain the completed point cloud for each sub-region; Global point cloud data is obtained by stitching together point clouds based on high-resolution, uniformly dense point clouds. A multi-scale feature point cloud filtering algorithm is used to filter global point cloud data and output a global point cloud ground point set. The multi-scale features include color information, roughness, and covariance-based features.

2. The point cloud filtering method for thick vegetation cover areas based on implicit neural representation according to claim 1, characterized in that, Acquire point cloud data, construct an adaptive irregular mesh based on the point cloud data, and segment the point cloud data based on the adaptive irregular mesh to obtain point cloud data of multiple sub-regions, including: Acquire point cloud data, traverse the point cloud data, and obtain the spatial location and point density of the point cloud data; Construct an adaptive irregular mesh based on point cloud density, wherein the adaptive irregular mesh based on point cloud density includes multiple sub-region meshes; The point cloud data is segmented using an adaptive irregular grid based on point cloud density, and the point cloud data is assigned to the corresponding sub-region grid according to the spatial location of the point cloud data, resulting in point cloud data of multiple sub-regions.

3. The point cloud filtering method for thick vegetation cover areas based on implicit neural representation according to claim 2, characterized in that, After obtaining point cloud data for multiple sub-regions, the following steps are also included: Step 1: Traverse the point cloud data of each sub-region to obtain the distribution of the point cloud data in each sub-region. The distribution includes the point cloud data in each sub-region and the boundary point cloud data on the boundary of the sub-region. Filter out the outliers in the boundary point cloud data on the boundary of the sub-region based on the consistency of the normal vector to obtain the boundary point cloud. Step 2: Establish a secondary subdivision grid. Divide the sub-regions with boundary point clouds into grids according to the secondary subdivision grid. Calculate the distance from each boundary point cloud to the centroid of the adjacent secondary subdivision grid. Perform inverse distance weighting on the adjacent secondary subdivision grids according to the distance. The adjacent secondary subdivision grids vote on each boundary point cloud according to the weight to obtain the first weight voting result. Assign each boundary point cloud to the secondary subdivision grid with the highest first weight voting result. Step 3: Iterate through the point cloud data in the second-level subdivision grid again. If there are boundary point clouds in the second-level subdivision grid, establish a third-level subdivision grid. Divide the second-level subdivision grid with boundary point clouds according to the third-level subdivision grid. Calculate the distance from each boundary point cloud to the centroid of the adjacent third-level subdivision grid. Perform inverse distance weighting on the adjacent third-level subdivision grids according to the distance. The adjacent third-level subdivision grids vote on each boundary point cloud according to the weight to obtain the second weight voting result. Assign each boundary point cloud to the third-level subdivision grid with the highest second weight voting result. Repeat the above steps until there are no boundary point clouds on the grid boundary.

4. The point cloud filtering method for thick vegetation cover areas based on implicit neural representation according to claim 1, characterized in that, Based on point cloud coordinate data, the calculation expression for predicting the density distribution of missing holes in the point cloud data of each sub-region using implicit neural functions is as follows: ； In the formula, Point cloud coordinate data, This represents the position of a point in the point cloud along the x-axis. This represents the position of a point in the point cloud along the y-axis. This represents the position of a point in the point cloud along the z-axis. It is an implicit function. The density of point cloud data for each sub-region, This represents the distance between point clouds and adjacent point clouds in point cloud data where holes are missing in each sub-region along the x-axis. This represents the distance between point clouds and adjacent point clouds in point cloud data for point cloud data with missing holes in each sub-region along the y-axis direction. This represents the distance between point clouds and adjacent point clouds in the point cloud data of each sub-region with missing holes along the z-axis. For each sub-region, the density distribution value of the missing holes in the point cloud data is predicted. This is the activation function.

5. The point cloud filtering method for thick vegetation cover areas based on implicit neural representation according to claim 1, characterized in that, The global point cloud data includes a ground point set; the global point cloud data is filtered based on a multi-scale, multi-feature point cloud filtering algorithm to output a global point cloud ground point set, including: The spherical neighborhood method is used to extract multiple features from global point cloud data to obtain multi-feature data, which includes color information, roughness, and covariance-based features. The covariance-based features include linearity, verticality, flatness, and total variance features. A multi-feature fusion method based on spherical neighborhood is used to fuse global point cloud data and multi-feature data to obtain a fused feature point set. Elevation information of ground point set is obtained by using large-scale grid and then corrected by small-scale grid to obtain corrected elevation information. The fused feature point set is then corrected based on the corrected elevation information to obtain the corrected fused feature point set. A low-resolution DSM model and a high-resolution DSM model are generated based on the corrected fused feature point set. An interpolation model is constructed based on the low-resolution DSM model and the high-resolution DSM model. The degree of elevation change between the low-resolution DSM model and the high-resolution DSM model is calculated based on the interpolation model. The mean and standard deviation of the degree of elevation change are calculated. The elevation variation coefficient of the DSM model is calculated based on the mean and standard deviation of the degree of elevation change. The elevation difference threshold is determined based on the elevation variation coefficient of the DSM model. The global point cloud data is filtered out based on the elevation difference threshold, and the global point cloud ground point set is output.

6. The point cloud filtering method for thick vegetation cover areas based on implicit neural representation according to claim 5, characterized in that, The calculation expressions for the degree of elevation change, coefficient of variation of elevation, and elevation difference threshold between the low-resolution DSM model and the high-resolution DSM model are as follows: ； In the formula, This represents the degree of elevation variation between the low-resolution and high-resolution DSM models. It is a high-resolution DSM model. This is a low-resolution DSM model. The elevation variation coefficient of the DSM model. The standard deviation is obtained based on the degree of elevation change. This is the average value obtained based on the degree of elevation change. The height difference threshold, and This is an empirical coefficient. The default value is 0.

5. The default value is 0.

1.

7. A point cloud filtering system for thick vegetation cover areas based on implicit neural representation, used to implement the method according to any one of claims 1 to 6, characterized in that, include: The acquisition and segmentation module is used to acquire point cloud data, build an adaptive irregular grid based on the point cloud data, and segment the point cloud data based on the adaptive irregular grid to obtain point cloud data of multiple sub-regions; The implicit neural upsampling module is used to input the point cloud data of each sub-region into the upsampling framework based on the implicit neural network to obtain the global point cloud data. The multi-scale feature point cloud filtering module is used to filter global point cloud data based on a multi-scale feature point cloud filtering algorithm and output a global point cloud ground point set.

8. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or the instructions are executed by the computer, they implement the point cloud filtering method for thick vegetation cover areas based on implicit neural representation as described in any one of claims 1 to 6.

Images