A method for extracting terrain feature lines based on active learning

Abstract

The present application belongs to the technical field of terrain feature line extraction, and discloses a terrain feature line extraction method based on active learning. In view of the problems of inaccurate feature line extraction and inability to effectively capture terrain features in the current feature line extraction method, the following solutions are proposed: taking LiDAR ground point cloud data as input data, calculating the multi-feature information of each point cloud, and using an active learning model constructed by taking random forest as a base classifier and entropy sampling as a query function to extract candidate feature points in the point cloud data; the outlier points in the candidate feature point set are removed by using the DBSCAN method to retain the feature points at the terrain features; then the candidate features can be refined by using a skeleton line extraction method based on Laplace shrinkage, and the points in the original point cloud closest to the refined points are matched to obtain accurate terrain feature points; finally, all the refined terrain feature points are connected step by step by using the minimum spanning tree to obtain accurate terrain feature lines.

Description

Technical Field

[0001] This invention belongs to the field of terrain feature line extraction technology, and relates to a terrain feature line extraction method based on active learning. Background Technology

[0002] Existing methods for simplifying ground point clouds often result in severe terrain distortion in the generated DEMs at low simplification rates. Therefore, integrating terrain feature lines into the simplified point cloud to reduce feature loss caused by simplification is of significant research importance. Constructing feature lines requires obtaining accurate feature points. Numerous machine learning and deep learning methods have been widely applied in fields such as 3D modeling, achieving groundbreaking progress. However, these methods often require large amounts of labeled data, a process that is manually performed and very costly. Therefore, many researchers have begun to study how to reduce the need for manually labeled samples during data training. The basic idea of ​​active learning is to allow machine learning models to autonomously select some of the most representative or informative point cloud samples for labeling, thereby improving the model's learning efficiency and performance. However, the feature points obtained by machine learning models inevitably contain a large number of spurious feature points, thus failing to construct accurate terrain feature lines. Summary of the Invention

[0003] The purpose of this invention is to propose a terrain feature line extraction method based on active learning. This method uses multiple feature vectors of point cloud as input vectors, identifies candidate feature points in point cloud data based on active learning, and then refines the candidate feature points by denoising and refining them. Finally, it uses a minimum spanning tree to gradually connect all the refined terrain feature points to obtain accurate terrain feature lines.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A terrain feature line extraction method based on active learning includes the following steps:

[0006] Step 1. Using LiDAR ground point cloud data as input data, calculate the multi-feature information of each point cloud, and use an active learning model to extract candidate feature points from the point cloud data;

[0007] Among them, the active learning model uses random forest as the base classifier and entropy sampling as the query function;

[0008] Step 2. Use the DBSCAN method to remove outliers from the candidate feature point set and retain feature points at terrain features;

[0009] Step 3. The feature points obtained in Step 2 are refined using the Laplacian shrinkage method, and the nearest points in the original point cloud to the refined points are matched to obtain accurate terrain feature points.

[0010] Step 4. Use the minimum spanning tree to connect all the refined terrain feature points step by step to obtain the terrain feature lines.

[0011] The present invention has the following advantages:

[0012] As described above, this invention discloses a terrain feature line extraction method based on active learning. This method addresses the problems of excessive sample annotation costs in machine learning and information loss and low accuracy in traditional terrain break line extraction methods due to the need for rasterized point clouds. It proposes directly constructing feature multi-vectors from the point cloud while utilizing an active learning model to reduce the amount of manually labeled data. Then, a skeleton line extraction method based on Laplacian contraction is used to effectively simplify the acquisition of locally redundant candidate points by active learning. Furthermore, to construct accurate terrain feature lines, this invention employs a minimum spanning tree method, thereby ensuring the continuity and accuracy of terrain break line extraction from the point cloud data. In subsequent processes, this invention utilizes the obtained accurate terrain feature lines to assist in simplifying the point cloud to construct a high-quality DEM. Attached Figure Description

[0013] Figure 1 This is a flowchart of the terrain feature line extraction method based on active learning in an embodiment of the present invention.

[0014] Figure 2 This is a flowchart illustrating the extraction of terrain feature lines in an embodiment of the present invention.

[0015] Figure 3 This is an orthophoto of the study area in an embodiment of the present invention.

[0016] Figure 4 This is a reference DEM image of the study area in this embodiment of the invention.

[0017] Figure 5 This is a schematic diagram of feature line extraction (sample area 2) in an embodiment of the present invention.

[0018] Figure 6 This is a schematic diagram of feature line extraction (sample area 3) in an embodiment of the present invention.

[0019] Figure 7 This is a schematic diagram of active learning modeling in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0021] This embodiment describes a terrain feature line extraction method based on active learning. The method first extracts candidate terrain feature points from the point cloud based on multiple feature vectors and an active learning model. Then, considering that there are outliers and local redundant points among the candidate feature points, it uses point cloud denoising and thinning methods to extract accurate terrain feature points, and uses minimum spanning tree to construct vector feature lines that match the terrain features. These feature lines can help simplify the point cloud to construct a high-quality DEM.

[0022] like Figure 1 and Figure 2 As shown, a terrain feature line extraction method based on active learning includes the following steps:

[0023] Step 1. Using LiDAR ground point cloud data as input data, calculate the multi-feature information of each point cloud and use an active learning model to extract candidate feature points from the point cloud data.

[0024] like Figure 7 As shown, the active learning model uses random forest as the base classifier and entropy sampling as the query function.

[0025] Step 1 specifically involves:

[0026] Step 1.1 Point cloud feature vector construction.

[0027] Extracting multi-feature information from airborne LiDAR ground point cloud data. Since point clouds exhibit high spatial correlation, their correlation features can be calculated through the relative relationships between points and other points within their neighborhood.

[0028] There are several ways to define the local neighborhood of a point cloud, including KNN neighborhood, spherical neighborhood, and cylindrical neighborhood.

[0029] The KNN neighborhood defines the neighborhood size by searching for the k nearest neighbors to the current point, using a fixed number of points in the cloud. The spherical neighborhood determines the neighborhood range using a fixed nearest neighbor distance and is generally used in scenarios with relatively uniform point cloud density. The cylindrical neighborhood only considers the two-dimensional distance between the current point and its nearest neighbors projected onto a planar coordinate system. This neighborhood is typically used for airborne LiDAR point clouds, where there are no overlapping point sets on the same plane. Considering the characteristics of each neighborhood, we use the cylindrical neighborhood to construct elevation-related features and the KNN neighborhood to construct point cloud feature value-related features.

[0030] For airborne LiDAR point clouds, the elevation differences between points projected into a planar coordinate system contain rich feature information. Therefore, elevation-related feature information of the point cloud can be constructed based on cylindrical neighborhoods. In addition to elevation values, the difference between the current point and the mean of neighboring points, and the elevation difference between a point and the farthest point, the following features can also be calculated:

[0031] 1. Range of elevation variation: The difference between the highest and lowest points among neighboring points can reflect the maximum range of elevation fluctuation in the point cloud within the neighborhood.

[0032] 2. Elevation standard deviation: This can be statistically used to reflect the degree of deviation between the current point and the average elevation of neighboring points. The elevation standard deviation is significantly positively correlated with the topographic relief.

[0033] 3. Maximum positive elevation difference and maximum negative elevation difference: The maximum positive elevation difference is the difference between the highest point in the point cloud data and the current point; the maximum negative elevation difference is the difference between the lowest point in the neighborhood and the current point.

[0034] Maximum positive elevation difference and maximum negative elevation difference can be used to describe the degree of deviation between the current point and the nearby extreme elevation point cloud data, such as the obvious elevation difference between the top and bottom of a cliff edge.

[0035] 4. Elevation Variance: The variance of the elevation values ​​of all points in the point cloud data can be used to assess the dispersion of the elevation distribution of the point cloud data. When the local neighborhood is located in an area with large topographic relief, this value will be larger.

[0036] 5. Elevation skewness: The degree of skewness in the distribution of elevation values ​​of all points in point cloud data, which can be used to describe the shape of the elevation value distribution in point cloud data.

[0037]

[0038] Where X is the sample, i.e., the elevation values ​​of all points in the neighborhood, μ is the elevation mean, σ is the elevation standard deviation, and E is the expectation operator.

[0039] 6. Elevation kurtosis: The kurtosis of the elevation values ​​of all points in the point cloud data. It can be used to describe the peaks or flatness of the elevation value distribution in the point cloud data.

[0040]

[0041] Where n is the total number of points in the neighborhood, x i Let i be the elevation value of the i-th point. This is the average elevation of all points within the neighborhood.

[0042] 7. Normalized Elevation Value: The ratio of the difference between the elevation value of each point in the point cloud data and the average elevation value of all points. It can be used to eliminate the influence of the absolute magnitude of the elevation value and to compare the height distribution characteristics of different point cloud data.

[0043] For the eigenvalues ​​of point clouds, the relevant features can be obtained by calculating the covariance matrix of all point cloud coordinates within the KNN neighborhood. The covariance matrix reflects the relationship between the point cloud coordinates within the neighborhood. Local point cloud fitting based on Principal Component Analysis (PCA) involves eigenvalue decomposition of the covariance matrix C, generating three eigenvalues ​​λ0, λ1, and λ2. These eigenvalues ​​represent the variance of the point cloud within the neighborhood in three different directions.

[0044] Since the eigenvalues ​​satisfy λ0≤λ1≤λ2, they can be used to describe the spatial distribution of point clouds in the neighborhood, and a series of derived features can be calculated based on the three eigenvalues.

[0045]

[0046] Where, p i For the i-th point in the KNN neighborhood, It is the mean of all points in the neighborhood.

[0047] 1. Eigenvalue λ2: The curvature eigenvalue λ2 refers to the maximum curvature value on the point cloud data, which can be used to describe the flatness of the point cloud data surface.

[0048] 2. Normal Vector: The formula for the local fitting plane of the neighborhood point cloud is: Ax + By + Cz = 1. Where A, B, and C are parameters, x, y, and z are variables, and the normal vector of the point cloud is the eigenvector n(ABC) corresponding to the minimum eigenvalue vector λ0.

[0049] 3. Anisotropy coefficient: Used to describe the anisotropy of the point cloud data surface, that is, how the surface shape changes in different directions. Its expression is as follows:

[0050]

[0051] 4. Linearity exponent: Used to describe linear changes in the surface of point cloud data, such as edges and corners. Its expression is as follows:

[0052]

[0053] 5. Planar Index: Used to describe planar regions on the surface of point cloud data, such as smooth surface regions. Its expression is as follows:

[0054]

[0055] 6. Sphere index: Used to describe the spherical region on the surface of point cloud data, its expression is as follows:

[0056]

[0057] 7. Total Change Index of Structure Tensor: This feature reflects the 3D structural characteristics of the neighborhood point cloud, and its expression is as follows:

[0058]

[0059] 8. The sum of eigenvalues ​​is expressed as follows:

[0060] λ SUM =λ0+λ1+λ2

[0061] 9. Curvature Variation: Curvature variation refers to the difference between curvature characteristic values. It can be used to describe local morphological changes on the surface of point cloud data, such as peaks and valleys. Its expression is as follows:

[0062]

[0063] Step 1.2. Data preprocessing.

[0064] Data preprocessing includes two steps: feature vector normalization and dataset partitioning.

[0065] The purpose of feature vector normalization is to scale the input feature vectors to the same range or distribution, avoid feature weight imbalance, and thus better train the active learning model.

[0066] The dataset partitioning step randomly divides the dataset into 70% training data and 30% test data. The training data is used for model prediction, and the test data is used to verify the accuracy of the prediction labels.

[0067] Step 1.3. Base classifier selection.

[0068] Random forests were chosen for training. Random forests are an ensemble learning method whose basic idea is to use multiple decision trees for classification or regression. In a random forest, each decision tree randomly selects data features and generates a prediction result. The final classification or regression result is the average of the predictions from all decision trees or the majority vote result.

[0069] This method can reduce the variance of the sample set and improve the robustness of the model, effectively avoiding overfitting.

[0070] Step 1.4. Selecting the query function.

[0071] Entropy sampling, used as a query function, is a query function based on information entropy. Its principle is to calculate the entropy value for each unlabeled sample and then select samples with higher entropy values ​​for labeling. Higher entropy values ​​indicate greater uncertainty in the sample; therefore, labeling samples with higher uncertainty can achieve better performance improvements.

[0072] For example, in binary classification, the uncertainty sampling strategy simply queries the instance whose posterior positive probability is closest to 0.5.

[0073] Step 1.5. Model training and testing.

[0074] The active learning model is trained using the multi-feature vectors and feature labels of 70% of the training data, namely point cloud data; and the multi-feature vectors of 30% of the test data are used as input data to predict whether the test point cloud data is a feature point.

[0075] Step 2. Use the DBSCAN method to remove outliers from the candidate feature point set and retain the feature points at the terrain features.

[0076] Due to the complexity of real terrain and the correlation of feature points in local space, the properties of feature points and their nearby pseudo-feature points are quite similar in three-dimensional space. Therefore, the candidate feature point set identified based on active learning exhibits obvious strip-like clustering at feature lines. This invention considers using the candidate feature point set as the processing object to further extract its skeleton line points as refined feature points of the land surface. Radar point clouds are prone to sparse distribution or significant elevation differences in some areas due to factors such as ground cover obstruction. The feature point set predicted by active learning will inevitably contain isolated non-feature line points. Therefore, noise in the candidate feature point set needs to be removed before extracting skeleton feature points.

[0077] This invention employs an outlier point cloud removal method based on the DBSCAN algorithm. DBSCAN is a density-based clustering algorithm that divides a dataset into several density-connected clusters. In DBSCAN, by setting a certain radius ε and a minimum number of samples MinPts, points with lower density can be identified as outliers and removed.

[0078] After processing with the DBSCAN algorithm, outlier data points with a density lower than the cluster density will be removed. The specific process is as follows:

[0079] Step 2.1. Set the search radius ε and the minimum number of samples MinPts.

[0080] Select a random point P1 from the set of candidate feature points that has at least MinPts points within its search radius.

[0081] Step 2.2. Evaluate each point in the neighborhood of point P1 to determine if point P1 has MinPts points within the epsilon distance. If point P1 has at least MinPts points in its neighborhood, point P1 will become another core point, and the core point set will be expanded. If a point does not meet the MinPts criterion, it will be marked as a boundary point. This process continues until the cluster is surrounded by boundary points, at which point the cluster search has stopped. If a point is neither a core point nor a boundary point, it will be considered a candidate feature point.

[0082] Step 2.3. Select a new random point from the unlabeled candidate feature point set and repeat the process of Step 2.2 to identify the next cluster until all points are divided into core point set, boundary point set and discrete point set.

[0083] By removing outliers from the point cloud, candidate feature points composed of core points and boundary points are obtained.

[0084] Step 3. The feature points obtained in Step 2 are refined using a skeleton line extraction method based on Laplacian contraction. The nearest points in the original point cloud to the refined points are then matched to obtain accurate terrain feature points.

[0085] After removing noise points from the candidate feature points, the candidate points need to be refined. Considering that the point set is mainly distributed near the feature lines, this invention uses a point cloud skeleton line extraction method based on Laplacian contraction. Local connectivity relationships of the point set are established through triangulation, allowing the point cloud data to be contracted using a weighted Laplacian operator. Then, the topological structure is refined to finally obtain the curve skeleton points of the point cloud. The points obtained by this method are not the true points in the original point cloud data; therefore, a one-to-one correspondence between these skeleton points and the original point cloud is required. The final terrain feature points are matched by KNN projection of the identified skeleton points onto a two-dimensional plane using the original point cloud. The point cloud skeleton line extraction process is as follows:

[0086] Step 3.1. Using the skeleton line extraction method based on Laplacian contraction, perform local Delaunay triangulation on each point cloud obtained in Step 2.

[0087] Step 3.2. By continuously reducing the vertices of the triangular profile through topological thinning, the candidate feature point set is gradually refined.

[0088] Step 3.3. Using the KNN nearest neighbor search algorithm, with each refined feature point as the origin, calculate the Euclidean distance between the origin and the original point cloud point by point, and select the original point closest to the origin as the terrain feature point. Finally, the refined terrain feature points are all points in the original point cloud.

[0089] Step 4. Use the minimum spanning tree to connect all the refined terrain feature points step by step to obtain the terrain feature lines.

[0090] The terrain feature points obtained from the above steps are not line features. A minimum spanning tree is used to connect the detected feature point data to generate smooth terrain feature lines. Step 4 specifically involves:

[0091] Step 4.1. Calculate the distance between each pair of adjacent terrain feature points obtained in Step 3, and use it as the edge weight of the point.

[0092] Step 4.2. Mark the point cloud obtained in Step 3 as a set of candidate points and create an empty set of selected points in advance; randomly select a point from the set of candidate points and add it to the set of selected points. Each time, select the edge with the shortest distance between the set of selected points and the set of candidate points, and add the points in the set of unselected points connected by the edge to the set of selected points.

[0093] Step 4.3. The construction of the minimum spanning tree, i.e. the construction of terrain feature lines, is completed when all points are in the selected set.

[0094] In addition, to verify the effectiveness of the active learning-based terrain feature line extraction method proposed in this invention, experimental data from the OpenTopography website (https: / / portal.opentopography.org) were selected.

[0095] The study area of ​​this invention is located in the western region of Yuba, California. The main terrain of the sample data is mountainous; orthophotos and original DEMs of the 6 groups of areas are shown below. Figure 3 and Figure 4 As shown. Sample areas 1 and 2 were both rectangular regions of 500m × 500m, with a point density ranging from 0.76 to 1.13 pts / m². 2 The average terrain slope is 28.44.0-32.45°.

[0096] To accurately evaluate the performance of the filtering algorithm, the raw data was first automatically filtered using Terrasolid. Then, by comparing the manually edited misclassified points with orthophotos, data containing only airborne LiDAR ground point clouds was obtained.

[0097] The results generated at each step are presented step by step, and the accuracy of feature line extraction is verified by visual inspection. In sample areas 1 and 2, terrain candidate points identified based on active learning are shown below. Figure 5 (a) and Figure 6As shown in (a), a large number of terrain points were identified in the terrain features, which proves the accuracy of active learning in extracting feature points. However, there are also a large number of outliers in the sample area caused by local small bumps or depressions, which need to be removed to avoid interfering with the construction of terrain feature lines.

[0098] Depend on Figure 5 (b) and Figure 6 (b) It can be seen that the candidate feature points obtained by the active learning method contain a large number of redundant feature points at the terrain features. The redundant points are iteratively refined by the Laplacian shrinkage method until more accurate feature points are obtained. Figure 5 (c) and Figure 6 (c) The feature points extracted in this step are distributed as a whole in the features of the terrain.

[0099] It is worth noting that, due to the relatively gentle terrain in some areas, there are local discontinuities at points along the ridge. Figure 5 (c) shown in the lower left and Figure 6 (c) The upper left part. However, by establishing the proximity topology between point clouds through minimum spanning trees, the final constructed terrain feature lines are relatively continuous and basically match the terrain, such as Figure 5 (d) and Figure 6 As shown in (d).

[0100] This invention directly utilizes multiple features of point clouds for robust feature point recognition, while simultaneously constructing relatively accurate terrain feature lines.

[0101] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for extracting terrain feature lines based on active learning, characterized in that, Includes the following steps: Step 1. Using LiDAR ground point cloud data as input data, calculate the multi-feature information of each point cloud, and use an active learning model to extract candidate feature points from the point cloud data; Among them, the active learning model uses random forest as the base classifier and entropy sampling as the query function; Step 2. Use the DBSCAN method to remove outliers from the candidate feature point set and retain feature points at terrain features; Step 3. The feature points obtained in Step 2 are refined using the skeleton line extraction method based on Laplacian contraction, and the nearest points in the original point cloud to the refined points are matched to obtain accurate terrain feature points. Step 4. Use the minimum spanning tree to connect all the refined terrain feature points step by step to obtain the terrain feature lines.

2. The terrain feature line extraction method based on active learning according to claim 1, characterized in that, Step 1 specifically involves: Step 1.1 Point cloud feature vector construction; Extracting multi-feature information of point clouds from airborne LiDAR ground point cloud data; Step 1.

2. Data preprocessing; Data preprocessing includes two steps: feature vector normalization and dataset partitioning. The dataset partitioning step randomly divides the dataset into 70% training data and 30% test data. The training data is used for model prediction, and the test data is used to verify the accuracy of the predicted labels. Step 1.

3. Base classifier selection; Random forest is selected for training; in random forest, each decision tree randomly selects data features and generates prediction results. The final classification or regression result is the average of the prediction results of all decision trees or the majority vote result. Step 1.

4. Selecting the query function; Entropy sampling as a query function is a query function based on information entropy. Step 1.

5. Model training and testing; The active learning model is trained using the multi-feature vectors and feature labels of 70% of the training data, namely point cloud data; and the multi-feature vectors of 30% of the test data are used as input data to predict whether the test point cloud data is a feature point.

3. The terrain feature line extraction method based on active learning according to claim 2, characterized in that, The point cloud feature vector constructed in step 1.1 includes elevation-related features and feature value-related features of the point cloud. Among them, cylindrical neighborhoods are used to construct point cloud elevation-related features; For the eigenvalue-related features of the point cloud, the covariance matrix of all point cloud coordinates within the KNN neighborhood is obtained.

4. The terrain feature line extraction method based on active learning according to claim 3, characterized in that, The elevation-related features include: Elevation value, mean difference between current point and neighboring points, elevation difference between point and farthest point, elevation variation range, elevation standard deviation, maximum positive elevation difference and maximum negative elevation difference, elevation variance, elevation skewness, elevation kurtosis, normalized elevation value; The elevation variation range reflects the maximum range of elevation fluctuations in the point cloud within the neighborhood. The standard deviation of elevation reflects the degree of deviation between the current point and the average elevation of neighboring points; The maximum positive elevation difference is the difference between the highest point in the point cloud data and the current point, while the maximum negative elevation difference is the difference between the lowest point in the neighborhood and the current point. The maximum positive elevation difference and the maximum negative elevation difference are used to describe the degree of deviation between the current point and the nearby extreme elevation point cloud data; Elevation variance refers to the variance of the elevation values ​​of all points in a point cloud data, and is used to assess the dispersion of the elevation distribution in point cloud data. Elevation skewness refers to the degree of skewness in the distribution of elevation values ​​of all points in point cloud data, and is used to describe the shape of the elevation value distribution in point cloud data; Elevation kurtosis refers to the kurtosis of the distribution of elevation values ​​of all points in point cloud data, and is used to describe the peak or flatness of the elevation value distribution of point cloud data. Normalized elevation value is the ratio of the difference between the elevation value of each point in point cloud data and the average elevation value of all points. It is used to eliminate the influence of the absolute magnitude of the elevation value and to compare the height distribution characteristics of different point cloud data.

5. The terrain feature line extraction method based on active learning according to claim 3, characterized in that, The eigenvalue-related features of the point cloud include eigenvalue λ2, normal vector, anisotropy coefficient, linear exponent, plane exponent, sphere exponent, total change exponent of structure tensor, and curvature change; The curvature eigenvalue λ2 refers to the maximum curvature value on the point cloud data, which is used to describe the flatness of the point cloud data surface; Anisotropy coefficients are used to describe the anisotropy of the surface of point cloud data, that is, how the surface shape changes in different directions; Linear exponents are used to describe linear changes on the surface of point cloud data; Planar indices are used to describe planar regions on the surface of point cloud data; The sphericity index is used to describe spherical regions on the surface of point cloud data; The total change index of the structure tensor is used to reflect the three-dimensional structural features of the neighborhood point cloud. Curvature variation is the difference between curvature feature values, used to describe local morphological changes on the surface of point cloud data.

6. The terrain feature line extraction method based on active learning according to claim 1, characterized in that, Step 2 specifically involves: Step 2.

1. Set the search radius ε and the minimum number of samples MinPts; Select a random point P1 from the candidate feature point set that has at least MinPts points within its search radius; Step 2.

2. Evaluate each point in the neighborhood of point P1 to determine whether point P1 has MinPts points within the epsilon distance; If there are at least MinPts points in the neighborhood of point P1, point P1 will become another core point, and the core point set will be expanded; if a point does not meet the MinPts criterion, it will be marked as a boundary point; and so on. When the cluster is surrounded by boundary points, the cluster search has stopped. If a point is neither a core point nor a boundary point, it will be a candidate feature point. Step 2.

3. Select a new random point from the unlabeled candidate feature point set and repeat the process of Step 2.2 to identify the next cluster until all points are divided into core point set, boundary point set and discrete points; By removing outliers from the point cloud, candidate feature points composed of core points and boundary points are obtained.

7. The terrain feature line extraction method based on active learning according to claim 1, characterized in that, Step 3 specifically involves: Step 3.

1. Using the skeleton line extraction method based on Laplacian contraction, perform local Delaunay triangulation on each point cloud obtained in Step 2; Step 3.

2. By continuously reducing the vertices of the triangular profile through topological thinning, the candidate feature point set is gradually refined; Step 3.

3. Using the KNN nearest neighbor search algorithm, with each refined feature point as the origin, calculate the Euclidean distance between the origin and the original point cloud point by point, and select the original point closest to the origin as the terrain feature point. Finally, the refined terrain feature points are all points in the original point cloud.

8. The terrain feature line extraction method based on active learning according to claim 1, characterized in that, Step 4 specifically involves: Step 4.

1. Calculate the distance between each pair of adjacent terrain feature points obtained in Step 3, and use it as the edge weight of the point; Step 4.

2. Mark the point cloud obtained in Step 3 as a set of candidate points and create an empty set of selected points in advance; randomly select a point from the set of candidate points and add it to the set of selected points. Each time, select the edge with the shortest distance between the set of selected points and the set of candidate points, and add the points in the set of unselected points connected by the edge to the set of selected points. Step 4.

3. The construction of the minimum spanning tree, i.e. the construction of terrain feature lines, is completed when all points are in the selected set.

Images