A method and system for filtering height estimates of residual ground feature point clouds

Abstract

The application discloses a height estimation method and system for filtering residual ground feature point clouds, and belongs to the technical field of laser point cloud data processing. The method comprises the following steps: randomly selecting a sampling point in the filtered laser point cloud, recording the spatial coordinates and relative coordinates in the neighborhood of the sampling point, and forming a sampling point cloud block; extracting a feature vector from the sampling point cloud block, and using the feature vector to extract the height gradient scalar of each point in the sampling point cloud block; mapping the extracted feature vector to a point cloud semantic probability after flattening, and classifying the point cloud by a preset threshold to obtain point cloud information; constructing a point cloud height above ground regression model, and training based on the constructed semantic guided height above ground prediction loss function; and using the trained point cloud height above ground regression model to estimate the height above ground of each point in the test data. The application not only realizes the consistency of point cloud semantic information and height above ground regression logic, but also realizes accurate prediction of the height above ground of the point cloud.

Description

Technical Field

[0001] This invention belongs to the field of laser point cloud data processing technology, specifically relating to a method and system for estimating the height of filtered residual ground feature point clouds. Background Technology

[0002] Point cloud ground filtering algorithms (especially deep learning-based methods) have demonstrated strong adaptability in complex terrain scenes and can usually achieve high classification accuracy in single scenes. However, due to the high similarity of local geometric features between low-lying ground features and rugged bare ground in the actual surface environment, existing filtering methods based on pure geometric features are difficult to effectively distinguish between these two types of points, resulting in low-lying ground features often being mixed into the initial ground point extraction results.

[0003] While fusing full-waveform data or multispectral imagery can effectively identify significant non-ground points, such auxiliary data is often difficult to obtain in practical engineering applications. In contrast, using only the XYZ coordinate information of the point cloud, accurately estimating the ground clearance of initial ground points through a progressive regression method, and achieving the detection and correction of all types of pseudo-ground points based on graph cut optimization has broader applicability. However, current research lacks information on estimating the height of residual ground features after ground filtering of airborne laser point clouds. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for estimating the height of filtered residual ground feature point clouds. This method and system can ensure consistency between the semantic information of the point cloud and the ground height regression logic, and can also provide robust, efficient, and precise estimation of the ground height of filtered residual low-lying ground features.

[0005] This invention provides the following technical solution:

[0006] Firstly, a height estimation method for filtered residual ground feature point clouds is provided, including the following steps:

[0007] S1. Randomly select a sampling point s from the filtered laser point cloud. i Record its spatial coordinates and its relative coordinates within its neighborhood to form a sampling point cloud.

[0008] S2. Extract feature vectors from the sampling point cloud S that can characterize the differences between low-lying features and the ground.

[0009] S3. Utilize the extracted feature vectors Extract the height gradient scalar g of each point in the sampled point cloud block S. i ;

[0010] S4. Extract the feature vector After flattening, the data is mapped to point cloud semantic probabilities, and point cloud classification is performed using a preset threshold to obtain point cloud information. i ∈{0,1};

[0011] S5. Construct a point cloud altitude regression model and train it based on the constructed semantically guided altitude prediction loss function.

[0012] S6. Use the trained point cloud ground height regression model to estimate the ground height of each point in the test data.

[0013] Optionally, step S2 includes the following steps:

[0014] S2.1 For the input sampled point cloud block S, use the k-NN algorithm to calculate the indices of the k nearest neighbor points for each point;

[0015] S2.2 Input the nearest neighbor point index and the coordinates of the point cloud into the edge feature extraction layer to extract the relative features of each point;

[0016] S2.3. Input the relative features into a dense edge convolutional layer composed of several fully connected layers to extract the intermediate feature h of each point. i ;

[0017] S2.4, Based on the current intermediate feature h i The k-NN algorithm is used to recalculate the indices of the k nearest neighbors of each point, generating a dynamically updated adjacency index.

[0018] S2.5. Repeat steps S2.2-S2.4 to form multiple sequentially stacked multi-layer dense edge convolutions to extract multiple intermediate features h. i Max pooling is used to fuse the data into a form that can represent each point s. i Feature vectors of multi-scale features

[0019] Optionally, step S3 includes the following steps:

[0020] S3.1, Combine the coordinates of the sampling point cloud block S with each point s i Corresponding feature vector splicing into fusion feature C xyz ;

[0021] S3.2, fuse features C through 1D initial convolution. xyz Encoding for each point s i Higher-order features;

[0022] S3.3, Connect several ResNet blocks to process each point s i The higher-order features output a latent space vector with consistent dimensions;

[0023] S3.4. By batch normalization and ReLU activation, the latent space vector is decoded into a height gradient scalar g. i .

[0024] Optionally, step S4 includes the following steps:

[0025] S4.1, The extracted feature vectors are processed through a feature flattening layer. Flattened into a one-dimensional vector;

[0026] S4.2. Using a classification layer composed of several fully connected layers and a sigmoid activation function, a one-dimensional vector is mapped to a point cloud semantic probability P. i ;

[0027] S4.3. Through the threshold comparison layer, the semantic probability P of the point cloud is calculated. i Threshold comparison is performed to obtain point cloud semantic information. i ∈{0,1}.

[0028] Optionally, the semantically guided altitude prediction loss function L in step S5 all for:

[0029]

[0030] Among them, L cls Represents the cross-entropy loss in binary classification. This represents the regression loss based on the ground point's height above the ground. The value represents the regression loss for the height above ground of low-lying features; B represents the batch size, n represents the number of points in each batch, and y represents the height above ground. pq The true label of the q-th point in the p-th batch, z pq σ(z) represents the original output of the q-th point in the p-th batch. pq () is the sigmoid function; This represents the total number of all ground points. This represents the predicted height of the point above the ground. h represents the number of all low-lying feature points. pq Represents the actual height of a point above the ground, where α, β, and δ are respectively L cls , and The corresponding weights, θ are The weight.

[0031] Optionally, step S5 specifically includes:

[0032] S5.1 Set the initial ground clearance of all points in the sampling point cloud block S to 0;

[0033] S5.2 Extract the feature vector of each point in the sampling point cloud S.

[0034] S5.3 Utilizing eigenvectors Extract the height gradient scalar g of each point in the sampled point cloud block S. i ;

[0035] S5.4, Sample each point s in the point cloud block S. i The height gradient scalar g of n neighborhood points i The aggregated gradient is obtained by accumulating gradients through gradient aggregation layers;

[0036] S5.5 Obtain the ground elevation of the intermediate point cloud by superimposing the product of the dynamic step size and the aggregate gradient. And based on the feature vector Recalculate each point s in the sampled point cloud block S i The nearest neighbor index;

[0037] S5.6 Repeat steps S5.3-S5.5 t times, gradually adjusting the point cloud height above the ground by aggregating gradient ascent and dynamic step size decay;

[0038] S6.7. The point cloud ground height regression model is trained under supervision using a semantically guided ground height prediction loss function and a labeled training set.

[0039] Optionally, the specific formula for gradually adjusting the point cloud's height above the ground is as follows:

[0040]

[0041] η t =η0γ t

[0042] in, s represents the time point t of the iteration. i The predicted ground clearance, η t η is the dynamic step size, η0 is the initial step size, γ is the decay coefficient, t is the number of iterations, and ΔHAG is the dynamic step size. i This is a semantically weighted gradient.

[0043] Secondly, a height estimation system for filtered residual ground feature point clouds is provided, including:

[0044] The sampling point cloud block module randomly selects a sampling point s from the filtered laser point cloud. i Record its spatial coordinates and its relative coordinates within its neighborhood to form a sampling point cloud.

[0045] The feature extraction module extracts feature vectors from the sampled point cloud S that can characterize the differences between low-lying features and the ground.

[0046] The height gradient prediction module utilizes the extracted feature vectors. Extract the height gradient scalar g of each point in the sampled point cloud block S. i ;

[0047] The point cloud semantic classification module extracts the feature vectors. After flattening, the data is mapped to point cloud semantic probabilities, and point cloud classification is performed using a preset threshold to obtain point cloud information. i ∈{0,1};

[0048] The model building and training module constructs a point cloud ground height regression model and trains it based on the constructed semantically guided ground height prediction loss function.

[0049] The prediction module uses a trained point cloud ground height regression model to estimate the ground height of each point in the test data.

[0050] Thirdly, a computer device is provided, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the height estimation method for filtered residual ground feature point clouds as described in any one of the first aspects.

[0051] Fourthly, a computer-readable storage medium is provided for storing a computer program; when the computer program is executed by a processor, it implements the steps of the height estimation method for the filtered residual ground feature point cloud as described in any one of the first aspects.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] (1) This invention decomposes the problem of predicting the ground clearance of residual low-lying features by ground filtering into three main steps: feature extraction, height gradient prediction, and ground clearance regression. This effectively reduces the overall complexity of the problem and significantly enhances its adaptability to rugged surfaces or complex terrain. In scenarios where low-lying features and rugged ground have similar geometric features and traditional filtering algorithms are prone to misjudgment, this application can achieve more robust and precise identification and ground clearance prediction.

[0054] (2) This invention relies on a differential regression strategy based on high gradient scalars and semantic information to complete the prediction. This strategy can select different optimization paths according to the category information of the points, making the prediction results more accurate and having good semantic consistency. Attached Figure Description

[0055] Figure 1 This is a flowchart of the height estimation method for filtered residual ground feature point clouds of the present invention;

[0056] Figure 2 This is a flowchart of the process of extracting feature vectors that can characterize the differences of low-lying land features according to the present invention.

[0057] Figure 3 This is a flowchart of the height gradient prediction process of the present invention;

[0058] Figure 4 This is a flowchart of the point cloud semantic classification process of the present invention;

[0059] Figure 5 This is a schematic diagram of the point cloud altitude regression network of the present invention. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the scope of protection of the present invention. It should be noted that the term "comprising" and any variations thereof in the specification, claims and the above-mentioned drawings of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or devices.

[0061] Example 1

[0062] like Figure 1 As shown, a height estimation method for filtered residual ground feature point clouds includes the following steps:

[0063] S1: Randomly select a sampling point s from the filtered laser point cloud. i Record its spatial coordinates and its relative coordinates within its neighborhood to form a sampling point cloud.

[0064] Alternatively, a large amount of laser point cloud data covering an area of ​​500m × 500m can be selected. A sampling point can be randomly selected from the entire point cloud, and 1000 neighboring points can be selected centered on this sampling point to form the sampled point cloud data.

[0065] S2: Extract feature vectors from the sampled point cloud S that can characterize the differences in low-lying land cover. like Figure 2 As shown, step S2 includes the following sub-steps:

[0066] S2.1 For the input sampled point cloud block S, use the k-NN algorithm to calculate the indexes of the k nearest neighbor points for each point; alternatively, k can be 16.

[0067] S2.2 Input the nearest neighbor point index and the coordinates of the point cloud into the edge feature extraction layer to extract the relative features of each point.

[0068] S2.3. Input the relative features into a dense edge convolutional layer composed of several fully connected layers to extract the intermediate feature h of each point. i .

[0069] S2.4, Based on the current intermediate feature h i The k-NN algorithm is used to recalculate the indices of the k nearest neighbors of each point, generating a dynamically updated adjacency index.

[0070] S2.5. Repeat steps S2.2-S2.4 to form multiple sequentially stacked multi-layer dense edge convolutions to extract multiple intermediate features h. i Max pooling is used to fuse the data into a form that can represent each point s. i Feature vectors of multi-scale features Alternatively, there can be four layers of dense edge convolutions.

[0071] Step 3: Utilize the extracted feature vectors Extracting the point cloud height gradient scalar g from the sampled point cloud block S i .like Figure 3 As shown, step S3 includes the following sub-steps:

[0072] S3.1, Combine the sampled point cloud S with each point s i Corresponding feature vector splicing into fusion feature C xyz ;

[0073] S3.2, fuse features C through 1D initial convolution. xyz Encoding for each point s i Higher-order features;

[0074] S3.3. Process the higher-order features from step S3.2 using four sequentially connected ResNet blocks. Within each ResNet block, features C are fused through residual connections and injection. xyz Enhance expressive power and output latent space vectors with consistent dimensions.

[0075] S3.4. Through batch normalization and ReLU activation, the latent space vector is finally decoded into a height gradient scalar g. i .

[0076] S4: Utilizing the extracted feature vectors By combining the sampled point cloud block S with point cloud semantic classification, a fine classification of ground points and low-lying feature points is achieved, yielding point cloud semantic information. For example... Figure 4 As shown, it includes the following sub-steps:

[0077] S4.1, The feature vectors extracted in step S2 are processed through a feature flattening layer. Flattened into a one-dimensional vector;

[0078] S4.2. Using a classification layer composed of several fully connected layers and a sigmoid activation function, a one-dimensional vector is mapped to a point cloud semantic probability P. i ;

[0079] S4.3. Through the threshold comparison layer, the semantic probability P of the point cloud is calculated. i Threshold comparison is performed to obtain point cloud semantic information. i ∈{0,1}.

[0080] Step S5: Construct a point cloud altitude regression model and train it based on the constructed semantically guided altitude prediction loss function.

[0081] A point cloud altitude regression model is constructed using steps S2-S4, and each point s in the sampled point cloud data S is used. i The model is trained under supervised instruction based on its ground clearance to obtain a well-trained model. For example... Figure 5 As shown, it includes the following sub-steps:

[0082] S5.1 Set the initial ground clearance of all points in the sampling point cloud block S to 0;

[0083] S5.2. Extract each point s from the sampling point cloud S using step S2. i eigenvectors

[0084] S5.3 Utilizing eigenvectors In step S3, the height gradient scalar g of each point in the sampled cloud block is extracted. i .

[0085] S5.4, Calculate the height gradient scalar g of the n neighboring points of each point in the sampled point cloud block. i The aggregated gradient is obtained by accumulating gradient aggregation layers.

[0086] S5.5 Obtain the ground elevation of the intermediate point cloud by superimposing the product of the dynamic step size and the aggregate gradient. And based on the feature vector Recalculate the nearest neighbor index of each point in the local point cloud block.

[0087] S5.6 Repeat steps S5.2 and S5.5 30 times, gradually adjusting the point cloud height HAG by aggregating gradient ascent and dynamic step size decay.

[0088] The iterative optimization formula for ground clearance in the point cloud ground clearance regression model is defined as follows:

[0089]

[0090] η t =η0γ t

[0091] In the formula, s represents the time point t of the iteration. i The predicted ground clearance, η t η is the dynamic step size, η0 is the initial step size, γ is the decay coefficient, t is the number of iterations, and ΔHAG is the dynamic step size. i Semantic weighted gradient;

[0092] Using semantically guided loss function L for ground height prediction all The labeled training set is used to train the point cloud altitude regression model.

[0093] Semantic-guided loss function for predicting ground altitude L all for:

[0094]

[0095] Among them, L cls Represents the cross-entropy loss in binary classification. This represents the regression loss based on the ground point's height above the ground. The value represents the regression loss for the height above ground of low-lying features; B represents the batch size, n represents the number of points in each batch, and y represents the height above ground. pq The true label of the q-th point in the p-th batch, z pq σ(z) represents the original output of the q-th point in the p-th batch. pq () is the sigmoid function; This represents the total number of all ground points. This represents the predicted height of the point above the ground. h represents the number of all low-lying feature points. pq Represents the actual height of a point above the ground, where α, β, and δ are respectively L cls , and The corresponding weights, θ are The weight.

[0096] S6. Use the model trained in step S5 to estimate the ground clearance of each point in the test data.

[0097] Example 2

[0098] A height estimation system for filtered residual ground feature point clouds includes:

[0099] The sampling point cloud block module randomly selects a sampling point s from the filtered laser point cloud. i Record its spatial coordinates and its relative coordinates within its neighborhood to form a sampling point cloud.

[0100] The feature extraction module extracts feature vectors from the sampled point cloud S that can characterize the differences between low-lying features and the ground.

[0101] The height gradient prediction module utilizes the extracted feature vectors. Extract the height gradient scalar g of each point in the sampled point cloud block S. i ;

[0102] The point cloud semantic classification module extracts the feature vectors. After flattening, the data is mapped to point cloud semantic probabilities, and point cloud classification is performed using a preset threshold to obtain point cloud information. i ∈{0,1};

[0103] The model building and training module constructs a point cloud ground height regression model and trains it based on the constructed semantically guided ground height prediction loss function.

[0104] The prediction module uses a trained point cloud ground height regression model to estimate the ground height of each point in the test data.

[0105] Example 3

[0106] The present invention provides a computer device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the above-described method for estimating the height of filtered residual ground feature point clouds.

[0107] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0108] Example 4

[0109] The present invention provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the above-described method for estimating the height of filtered residual ground feature point clouds.

[0110] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0111] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The systems, devices, and storage media disclosed in the embodiments are described simply because they correspond to the methods disclosed in the embodiments; relevant details can be found in the method section.

[0112] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0113] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for estimating the height of filtered residual ground feature point clouds, characterized in that, Includes the following steps: S1. Randomly select a sampling point from the filtered laser point cloud. Record its spatial coordinates and its relative coordinates within its neighborhood to form a sampling point cloud. ; S2, from sampling point cloud blocks Extract feature vectors that can characterize the differences between low-lying features and the ground surface. ; S3. Utilize the extracted feature vectors Extract sampling point cloud blocks The height gradient scalar at each point in the middle ; S4. Extract the feature vector After being flattened, the data is mapped to point cloud semantic probabilities, and then the point clouds are classified using a preset threshold to obtain point cloud information. ; S5. Construct a point cloud altitude regression model and train it based on the constructed semantically guided altitude prediction loss function. S6. Use the trained point cloud ground height regression model to estimate the ground height of each point in the test data; Step S4 includes the following steps: S4.1, The extracted feature vectors are processed through a feature flattening layer. Flattened into a one-dimensional vector; S4.

2. Using a classification layer composed of several fully connected layers and a sigmoid activation function, a one-dimensional vector is mapped to point cloud semantic probabilities. ; S4.

3. Using a threshold comparison layer, the semantic probability of the point cloud is evaluated. Point cloud semantic information is obtained by threshold comparison. ; Step S5 specifically includes: S5.1, Set sampling point cloud blocks All points have an initial ground clearance of 0. S5.2 Extracting sampling point cloud blocks The feature vector of each point ; S5.3 Utilizing eigenvectors Extract sampling point cloud blocks The height gradient scalar at each point in the middle ; S5.4, Sample point cloud blocks Each point in of The height gradient scalar of each neighboring point The aggregated gradient is obtained by accumulating gradients through gradient aggregation layers; S5.5 Obtain the ground elevation of the intermediate point cloud by superimposing the product of the dynamic step size and the aggregate gradient. And based on the feature vector Recalculate the sampling point cloud block Each point in The nearest neighbor index; S5.6, Repeat In steps S5.3-S5.5, the point cloud height above the ground is gradually adjusted by aggregating gradient ascent and dynamic step size decay. S6.

7. The point cloud ground height regression model is trained under supervision using a semantically guided ground height prediction loss function and a labeled training set.

2. The height estimation method for filtered residual ground feature point clouds according to claim 1, characterized in that, Step S2 includes the following steps: S2.1, Input sampling point cloud blocks The k-NN algorithm is used to calculate the indices of the k nearest neighbors for each point; S2.2 Input the nearest neighbor point index and the coordinates of the point cloud into the edge feature extraction layer to extract the relative features of each point; S2.

3. Input the relative features into a dense edge convolutional layer composed of several fully connected layers to extract the intermediate features of each point. ; S2.4, Based on the current intermediate features The k-NN algorithm is used to recalculate the indices of the k nearest neighbors of each point, generating a dynamically updated adjacency index. S2.

5. Repeat steps S2.2-S2.4 to form multiple sequentially stacked multi-layer dense edge convolutions to extract multiple intermediate features. Max pooling is used to fuse the data into a form that can represent each point. Feature vectors of multi-scale features .

3. The height estimation method for filtered residual ground feature point clouds according to claim 1, characterized in that, Step S3 includes the following steps: S3.1, Sample point cloud blocks coordinates of each point corresponding feature vector splicing as a fusion feature ; S3.2, Fusing features through 1D initial convolution. Encoding for each point Higher-order features; S3.3 Connect several ResNet blocks to process each point The higher-order features output a latent space vector with consistent dimensions; S3.

4. By batch normalization and ReLU activation, the latent space vector is decoded into a high-gradient scalar. .

4. The height estimation method for filtered residual ground feature point clouds according to claim 1, characterized in that, The semantically guided altitude prediction loss function described in step S5 for: ； ； ； ； in, Represents the cross-entropy loss in binary classification. This represents the regression loss based on the ground point's height above the ground. This indicates the regression loss due to the height of low-lying features above the ground. Indicates batch size, This indicates the number of points in each batch. No. The first batch The real label of each point Indicates the first The first batch The original output of each point, It is the sigmoid function; This represents the total number of all ground points. This represents the predicted height of the point above the ground. This indicates the number of all low-lying features. This represents the actual height of the point above the ground. , and They are respectively , and The corresponding weights for The weight.

5. The height estimation method for filtered residual ground feature point clouds according to claim 1, characterized in that, The specific formula for gradually adjusting the point cloud's altitude above the ground is: in, Indicates the first Next iteration time point The predicted height above the ground, For dynamic step size, The initial step size, The attenuation coefficient is... For the number of iterations, This is a semantically weighted gradient.

6. A height estimation system for filtered residual point clouds, comprising the steps of the height estimation method for filtered residual point clouds according to any one of claims 1-5, characterized in that, include: The sampling point cloud block module randomly selects a sampling point from the filtered laser point cloud. Record its spatial coordinates and its relative coordinates within its neighborhood to form a sampling point cloud. ; The feature extraction module extracts features from the sampled point cloud. Extract feature vectors that can characterize the differences between low-lying features and the ground surface. ; The height gradient prediction module utilizes the extracted feature vectors. Extract sampling point cloud blocks The height gradient scalar at each point in the middle ; The point cloud semantic classification module extracts the feature vectors. After being flattened, the data is mapped to point cloud semantic probabilities, and then the point clouds are classified using a preset threshold to obtain point cloud information. ; The model building and training module constructs a point cloud ground height regression model and trains it based on the constructed semantically guided ground height prediction loss function. The prediction module uses a trained point cloud ground height regression model to estimate the ground height of each point in the test data.

7. A computer device, characterized in that, It includes a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the height estimation method for filtered residual ground feature point clouds as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Used to store computer programs; when the computer programs are executed by a processor, they implement the steps of the height estimation method for filtered residual ground feature point clouds as described in any one of claims 1-5.

Images