A method for removing lidar noise points under severe weather conditions based on graph neural networks

Abstract

This invention discloses a method for removing noise points from lidar data under adverse weather conditions based on graph neural networks, belonging to the field of image processing technology. The method first constructs a graph structure from point cloud data acquired by lidar, including vertex attributes such as spatial location, reflection intensity, local spatial density, local directional features, and edge relationships connecting vertices. Then, using a graph neural network, noise prediction is performed on each vertex through graph convolution. Next, the point cloud is restored based on the correspondence between vertices and the point cloud, and the noisy point cloud is removed. This method effectively achieves end-to-end noise removal without destroying the point cloud data structure, thereby improving the accuracy of lidar data. Experiments on the CADC dataset demonstrate that the method shows a significant performance improvement, possessing strong practicality and promotional value.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for removing noise points from lidar under severe weather conditions based on graph neural networks. Background Technology

[0002] LiDAR (Light Detection and Ranging) is a type of active sensor that emits a laser beam and detects the echo to achieve distance detection, generating three-dimensional point cloud data. It features high measurement accuracy and long detection range, and is widely used in autonomous driving, remote sensing, and mapping. In adverse weather conditions, such as rain and snow, environmental particles can scatter the laser beam, resulting in noise points. This noise can interfere with downstream applications. For example, in autonomous driving applications, it can prevent the identification of obstacles, and in surveying, it can affect mapping results.

[0003] Current research in point cloud denoising mainly falls into two categories: rule-based classifiers and image processing-based convolutional neural networks. Rule-based classifiers rely on low noise distribution density and weak reflection intensity. Radius Outlier Removal (ROR) methods identify noisy points based on the number of neighborhood points in the point cloud; if the number of neighborhood points is below a set threshold, they are removed. Statistical Outlier Removal (SOR) methods determine noisy points by calculating the average distance between a point and its k nearest neighbors. However, these methods require parameter adjustments based on different LiDAR systems, and their classification accuracy is not high.

[0004] Convolutional neural networks (CNNs) for image processing transform the point cloud denoising problem into an image processing problem by projecting 3D point clouds onto a 2D plane. For example, WeatherNet uses CNNs to denoise point clouds under weather conditions, effectively handling noisy point clouds affected by weather such as rain and fog. Furthermore, LiSnowNet employs unsupervised learning, evaluating the network output signal through Fourier transform and wavelet transform to achieve unsupervised point cloud denoising. However, projecting point clouds onto a 2D grid, due to the irregular distribution of the point clouds, leads to the loss of point cloud information in some grids, and the lack of information in other grids results in a waste of computational resources. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for removing noise points in lidar under severe weather conditions based on graph neural networks. It proposes to apply graph neural networks to lidar point cloud denoising theory to remove point cloud noise caused by severe weather and provide clean data for downstream applications.

[0006] The technical problem addressed by this invention is solved as follows:

[0007] A method for removing noise points from lidar under severe weather conditions based on graph neural networks includes the following steps:

[0008] Step 1: Preprocess the point cloud signal;

[0009] A frame of point cloud signal acquired by the lidar is represented as an M×4 array as the preprocessed point cloud signal, where M is the number of point clouds in a frame;

[0010] Step 2: Constructing the graph structure;

[0011] Vertices are constructed by downsampling the point cloud, and the relationship between edges between vertices is constructed by dynamic radius search. The normal vectors, point cloud quantity and spatial density within the vertex are extracted as local density features, and combined with the spatial coordinates and reflection intensity features of the vertex as the internal features of the graph structure.

[0012] Step 3: Inference of the Graph Neural Network

[0013] The graph structure data is input into the graph neural network; in the graph neural network, edge features are extracted by a multilayer perceptron (MLP); convolution operation is performed by a graph attention network (GAT) to aggregate information of adjacent vertices, and attention coefficients are calculated based on the extracted edge features and vertex features; the features of the vertices are updated by the MLP based on the attention coefficients; after passing through multiple layers of the network, the output is a prediction of whether a vertex is noise.

[0014] Step 4: Denoising and Reconstruction of Point Clouds

[0015] Based on the prediction results of the vertices, each vertex is determined to be either noise or an object, and the point cloud signal predicted to be an object is recovered based on the index relationship between the vertex and the point cloud.

[0016] Furthermore, in step 1, a frame of point cloud signal acquired by the lidar is represented as an M×4 array as the preprocessed point cloud signal P, where M is the number of point clouds in a frame, and 4 represents the four feature dimensions of the point cloud, including: the spatial coordinates X of the x-axis, y-axis and z-axis of the rectangular coordinate system centered on the lidar and the laser reflection intensity I.

[0017] Furthermore, the specific process of step 2 is as follows:

[0018] Downsampling M point clouds yields N vertices, represented as:

[0019] X←RandomSampling(P)(1)

[0020] Where RandomSampling represents the random downsampling operation, P represents the preprocessed point cloud signal, and X represents the vertex space coordinates obtained by downsampling;

[0021] Based on the spatial coordinates X of the vertex and the spatial coordinates X of the point cloud, each point cloud is divided to its nearest neighbor vertex, represented as:

[0022] Indices←Arg min(Distance(P,X)) (2)

[0023] Where Distance represents the distance between the point cloud and the vertex calculated based on P and X, and Indices is the index relationship established between the vertex and the point cloud;

[0024] Simultaneously, based on Indices, the center spatial position and average reflection intensity of each vertex are calculated. Local density features are extracted by the mean and number of point cloud spacings within the vertex. The normal vector within the point cloud is calculated as the local direction vector. The center spatial position, average reflection intensity, local density features, and local direction vector of the vertex are vector-linked to form the vertex features.

[0025] Within a defined search radius, if two vertices are spatially adjacent, then an edge exists connecting them; for the i-th vertex, this is represented as:

[0026]

[0027] Among them, R i X represents the search radius of the i-th vertex, γ is the scaling factor of the search radius, and X... i Let || represent the spatial position of the center of the i-th vertex, || represents the L2 norm, and NeighborsWithRadiusSearch(X) represents the position of the center of the center of the i-th vertex. i ,R i ) represents the search radius of the i-th vertex as R. i Neighbor search, E i Let be the set of edges connected to the i-th vertex.

[0028] Furthermore, in step 2, the features of the vertices are specifically represented as follows:

[0029]

[0030] Where Indices[j] represents the vertex index of the j-th point cloud partition, i is the vertex index, 1≤i≤N, and Index represents the set of all point cloud indices associated with the i-th vertex; i Let I[j] represent the reflection intensity characteristic of the i-th vertex, I[j] represent the laser reflection intensity of the j-th point cloud, and Mean represent the average value; X iLet X[j] represent the center spatial position of the i-th vertex, and let X[j] represent the spatial coordinates of the j-th point cloud; D i F represents the local density feature of the i-th vertex, Count represents the number of points in the current index, and Concat represents the concatenation operation on the vectors; i V represents the local direction vector of the i-th vertex, and NormalVector represents the normal vector calculated based on the spatial position; i This represents the characteristic of the i-th vertex.

[0031] Furthermore, in step 2, the scaling factor γ of the search radius is set to 0.01.

[0032] Furthermore, the specific process of step 3 is as follows:

[0033] Graph-structured data is represented as G = (V, E), where V represents vertex features and E represents the connections between vertices; in the t-th iteration of the graph neural network, the edge features between any two adjacent vertices i and k are... for:

[0034]

[0035] in, and This represents the features of the i-th and k-th vertices in the t-th iteration. and x represents the multilayer perceptron used to extract edge and vertex features to achieve nonlinear transformation. i and x k These represent the center spatial positions of the i-th vertex and the k-th vertex, respectively. This represents the translation correction amount corresponding to the i-th vertex in the t-th iteration;

[0036] In the t-th iteration, the attention coefficient between the i-th vertex and the k-th vertex. Represented as:

[0037]

[0038] in, This represents a multilayer perceptron that performs nonlinear transformations; LeakyReLU represents a class of nonlinear activation functions; and Softmax represents a normalized exponential function.

[0039] The feature of the i-th vertex updated in the (t+1)th iteration is: for:

[0040]

[0041] in, Represented as a multilayer perceptron with nonlinear operations, (i,k)∈E means obtaining the connections of all vertices connected to vertex i based on the edge connection relationship E;

[0042] This completes one iteration for the i-th vertex. One iteration requires updating the features of all vertices in the graph structure G. The iteration process will run for T rounds, where T is the set number of rounds.

[0043] The probability that the i-th vertex is predicted as noise is expressed as:

[0044] p i =Sigmoid(Linear(v i (8)

[0045] Where, p i v represents the probability that the i-th vertex is predicted as noise. i Let represent the feature of the i-th vertex after iteration, Linear represents a single-layer linear transformation of the vector, and Sigmoid represents a class of activation functions that normalize the probability.

[0046] Furthermore, the specific process of step 4 is as follows:

[0047] If the probability p of the i-th vertex being predicted as noise i The value is greater than or equal to the set threshold. th If the vertex is determined to be noise caused by snow, it will not be reconstructed during the reconstruction process; if the predicted probability p i The value is less than snow th If the vertex is identified as an object, then the point cloud represented by that vertex is reconstructed.

[0048] Reconstructing point cloud P out The process can be characterized as:

[0049]

[0050] Among them, P j This is the preprocessed point cloud signal associated with the j-th point cloud of the i-th vertex.

[0051] Furthermore, in step 4, a threshold value for snow is set. th Set it to 0.3.

[0052] The beneficial effects of this invention are:

[0053] (1) This invention is the first to use a graph neural network method to remove lidar noise caused by severe weather, thereby providing clean data for downstream applications.

[0054] (2) The present invention uses graph neural networks to remove noise from lidar point clouds under severe weather conditions, which has high noise removal accuracy, that is, it can accurately distinguish between noise point clouds and object point clouds.

[0055] (3) This invention provides the principle derivation process of the lidar point cloud denoising algorithm under severe weather conditions for graph neural networks, and establishes a theoretical foundation.

[0056] (4) The present invention extracts the vertex features of the graph neural network based on the difference between noise and objects, and achieves high noise reduction accuracy.

[0057] (5) The present invention can handle point cloud noise caused by a variety of severe weather conditions, such as rain, snow and fog.

[0058] (6) This invention is not only simple to implement and easy to operate, but also highly practical and suitable for widespread use. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the method described in this invention;

[0060] Figure 2 This is a diagram illustrating the results of data / image processing at each step of the method described in the embodiment.

[0061] Figure 3 The diagram illustrates the noise removal results caused by different severe weather conditions in the method described in the embodiments. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0063] This embodiment provides a method for removing noise points from lidar under severe weather conditions based on graph neural networks, and its flowchart is shown below. Figure 1 As shown, a graph neural network is used as the classifier, which can adapt to the irregularity of point clouds, thus achieving efficient denoising. This method has the advantages of high denoising accuracy and no damage to the point cloud structure. This graph neural network-based lidar point cloud denoising algorithm under severe weather conditions is simple to implement, has a shallow network structure, is easy to operate, and has strong practicality and scalability.

[0064] The method described in this embodiment specifically includes the following steps:

[0065] Step 1: Preprocess the point cloud signal;

[0066] The input to the algorithm is a frame of point cloud signal collected by the lidar.

[0067] In step 1, the point cloud signal of a frame acquired by the lidar is sorted and deformed, and can be represented as an M×4 array as the preprocessed point cloud signal P, where M is the number of point clouds in a frame, and 4 represents the four feature dimensions of the point cloud, including: the spatial coordinates X of the x-axis, y-axis and z-axis of the rectangular coordinate system centered on the lidar and the laser reflection intensity I.

[0068] In this embodiment, the processed point cloud signal is as follows: Figure 2 The “raw point cloud data” is shown in the image.

[0069] Step 2: Constructing the graph structure;

[0070] Vertices are constructed by downsampling the point cloud, and the relationship between edges between vertices is constructed by dynamic radius search. The normal vectors, point cloud quantity and spatial density within the vertices are extracted as local density features, and combined with the spatial coordinates and reflection intensity features of the vertices as the internal features of the vertices in the graph structure.

[0071] Furthermore, the specific process of step 2 is as follows:

[0072] First, downsample the M point clouds from step 1 into N vertices:

[0073] X←RandomSampling(P)(1)

[0074] Where RandomSampling(·) represents the random downsampling operation, P represents the preprocessed point cloud signal, and X represents the vertex space coordinates obtained by downsampling.

[0075] Next, based on the vertex spatial coordinates X and the point cloud spatial coordinates X, each point cloud is divided to its nearest neighbor vertex, represented as:

[0076] Indices←Arg min(Distance(P,X)) (2)

[0077] Where Distance(·) represents the distance between the point cloud and the vertex calculated based on P and X, and Indices is the index relationship established between the vertex and the point cloud, which serves as the basis for restoring the point cloud in step 4.

[0078] Simultaneously, based on Indices, the center spatial position and average reflection intensity of each vertex are calculated. Local density features are extracted using the mean and number of point cloud spacings within a vertex, and the normal vector within the point cloud is used as the local direction vector. The center spatial position, average reflection intensity, local density features, and local direction vector of a vertex are then linked to form the vertex's features; specifically, this is expressed as follows:

[0079]

[0080] Where Indices[j] represents the vertex index of the j-th point cloud partition, i is the vertex index, 1≤i≤N, and Index represents the set of all point cloud indices associated with the i-th vertex; i Let I[j] represent the reflection intensity characteristic of the i-th vertex, I[j] represent the laser reflection intensity of the j-th point cloud, and Mean represent the average value; X i Let X[j] represent the center spatial position of the i-th vertex, and let X[j] represent the spatial coordinates of the j-th point cloud; D i F represents the local density feature of the i-th vertex, Count represents the number of points in the current index, and Concat represents the concatenation operation on the vectors; i V represents the local direction vector of the i-th vertex, and NormalVector represents the normal vector calculated based on the spatial position. This embodiment uses principal component analysis to implement it; i This represents the characteristic of the i-th vertex.

[0081] The graph structure includes vertices and edges. The vertex features have been constructed through the steps described above. The edge connection E can be represented as a dynamic radius search problem, meaning that within a given search radius, if two vertices are spatially adjacent, an edge connection exists. For the i-th vertex, it is represented as:

[0082]

[0083] Among them, R i Let represent the search radius of the i-th vertex, γ be the scaling factor of the search radius (constrained by the lidar parameters), and || represent the L2 norm. NeighborsWithRadiusSearch(X i ,R i ) represents the search radius of the i-th vertex as R. i For neighbor search, this embodiment uses a binary tree search method. i Let be the set of edges connected to the i-th vertex.

[0084] In this real-time example, the number of downsampled samples, N, is set to 5000. Each point cloud is assigned to its nearest neighbor based on the spatial coordinates of the vertices and the point cloud. Indices represent the index relationship established between vertices and the point cloud, serving as the basis for point cloud recovery in step 4. Simultaneously, based on Indices, the center spatial position and average reflection intensity of the vertex are calculated, serving as the vertex's spatial feature X and reflection intensity feature. Local density features D are extracted by the mean and number of point cloud spacings within vertices, and local direction vectors F are calculated using the normal vectors within the point cloud. Vertex features are obtained by connecting these vectors. The graph structure includes vertices and edges; vertex feature construction has been completed through the above steps. The edge connection relationship E can be represented as a dynamic radius search problem, i.e., if two vertices are spatially adjacent within a given radius, an edge connection exists. In this embodiment, the scaling factor γ of the Velodyne 64-line LiDAR is set to 0.01. Thus, step 2 completes the graph structure construction, and the graph structure data is as follows: Figure 2 As shown in the “Graph Structure Data” section.

[0085] Step 3: Inference of the Graph Neural Network

[0086] Inputting graph-structured data into a graph neural network (GNN) allows the application of GNNs in point cloud denoising to be better suited to the structure of point clouds, completely avoiding information loss and wasted computational resources. Furthermore, GNN denoising methods exhibit better denoising accuracy and can more accurately identify the type of point cloud.

[0087] In a graph neural network, edge features are extracted using a multilayer perceptron (MLP); convolution operations are performed using a graph attention network (GAT) to aggregate information from adjacent vertices, and attention coefficients are calculated based on the extracted edge features and vertex features; the MLP updates the vertex features based on the attention coefficients; and the network outputs a prediction of whether a vertex is noise.

[0088] Furthermore, the specific process of step 3 is as follows:

[0089] Graph-structured data can be represented as G = (V, E), where V represents vertex features and E represents the connections between vertices. In graph convolutional networks, vertex features are refined using a multilayer perceptron, and information is aggregated by performing graph convolution with attention mechanism (GAT) on vertex and edge features.

[0090] In the t-th iteration of the graph neural network, the edge features between any two adjacent vertices i and k are... The calculation process is as follows:

[0091]

[0092] in, and This represents the features of the i-th and k-th vertices in the t-th iteration. and x represents the multilayer perceptron used to extract edge and vertex features to achieve nonlinear transformation. i and x k These represent the center spatial positions of the i-th vertex and the k-th vertex, respectively. It represents the translation correction amount corresponding to the i-th vertex in the t-th iteration.

[0093] Based on the extracted edge features, by introducing an attention coefficient, features with different weights can be extracted for different neighboring nodes. In the t-th iteration, the attention coefficient between the i-th vertex and the k-th vertex... The calculation process can be represented as:

[0094]

[0095] in, represents a multilayer perceptron undergoing nonlinear transformation, LeakyReLU represents a class of nonlinear activation functions, and Softmax represents a normalized exponential function.

[0096] Based on the extracted attention coefficients, vertex features can be updated through convolution operations. The features of the i-th vertex updated in the (t+1)-th iteration are: It can be represented as:

[0097]

[0098] in Represented as a multilayer perceptron with nonlinear operations, (i,k)∈E means obtaining the connections of all vertices connected to vertex i based on the edge connection relationship E;

[0099] This completes one iteration for the i-th vertex; each iteration requires updating the features of all vertices in the graph structure G. In step 3, the above iteration process will be performed T times to refine the vertex features.

[0100] Based on the refined vertex features, a linear transformation can be used to predict whether a vertex is noise. The probability prediction process for whether the i-th vertex is noise can be expressed as:

[0101] p i =Sigmoid(Linear(v i (8)

[0102] Where, p i v represents the probability that the i-th vertex is predicted as noise. i Let represent the feature of the i-th vertex after iteration, Linear represents a single-layer linear transformation of the vector, and Sigmoid represents a class of activation functions that normalize the probability.

[0103] In this step, the features of adjacent edges are first calculated based on the spatial coordinates of the vertex. Then, based on the extracted edge features, attention coefficients are introduced to extract features with different weights for different adjacent nodes. Afterwards, based on the extracted attention coefficients, the vertex features are updated through convolution operations with an attention mechanism. In this real-time example, MLP... h MLP g MLP f The number of channels set are (8,3), (8,32,32), and (8,32,32). The above steps complete one iteration for vertex i. One iteration requires updating all vertices in the graph structure G. In this real-time example, the above iteration process will be performed three times to refine the vertex features. Based on the refined vertex features, a linear transformation can be used to predict whether a vertex is noise. The predicted vertex result can be obtained from... Figure 2 The term "predicted vertex data" is used in this context.

[0104] Step 4: Denoising and Reconstruction of Point Clouds

[0105] Based on the prediction results of the vertex, determine whether the vertex is noise, and recover the point cloud signal predicted to be an object based on the index relationship between the vertex and the point cloud.

[0106] Furthermore, the specific process of step 4 is as follows:

[0107] If the probability p of the i-th vertex being predicted as noise i The value is greater than or equal to the set threshold. th If the vertex is determined to be noise caused by snow, it will not be reconstructed during the reconstruction process; conversely, if the predicted probability p i The value is less than snow th If the vertex is an object, then the point cloud represented by that vertex is reconstructed.

[0108] Reconstructing point cloud P out The process can be characterized as:

[0109]

[0110] Among them, P j This is the preprocessed point cloud signal associated with the j-th point cloud of the i-th vertex.

[0111] Through the above steps, a one-time end-to-end denoising process for severe weather point clouds based on graph neural networks was achieved.

[0112] In this embodiment, snow th Set to 0.3. The reconstructed result can be obtained from... Figure 2 The term "reconstructed point cloud data" is used in this context.

[0113] In summary, this invention innovatively employs a graph neural network-based method for denoising lidar point cloud signals in severe weather. Unlike existing rule-based and image-based convolutional neural network denoising methods, this invention achieves high-precision point cloud denoising without damaging the point cloud structure, accurately distinguishing between noisy and object point clouds. Furthermore, this invention extracts vertex features from the graph neural network based on the differences between noise and objects, further enhancing the denoising effect. Moreover, this invention can handle point cloud noise from various types of severe weather, such as rain, snow, and fog. Figure 3 As shown, from left to right, the noise reduction effects of this invention on lidar point cloud noise in snow, rain, and fog environments are illustrated. This invention is simple to implement, lacks a deep network structure, is easy to operate, and has strong practicality and scalability.

[0114] The structures listed above are merely basic structures for illustrative purposes and should not be used to limit the scope of protection of this invention. Any modifications or refinements made to the main design concept and spirit of this invention that are not of substantial significance but still solve the same technical problem as this invention should be included within the scope of protection of this invention.

Claims

1. A method for removing noise points from lidar under severe weather conditions based on graph neural networks, characterized in that, Includes the following steps: Step 1: Preprocess the point cloud signal; A frame of point cloud signal acquired by the lidar is represented as an M×4 array as the preprocessed point cloud signal, where M is the number of point clouds in a frame; Step 2: Constructing the graph structure; Vertices are constructed by downsampling the point cloud, and the relationship between edges between vertices is constructed by dynamic radius search. The normal vectors, point cloud quantity and spatial density within the vertex are extracted as local density features, and combined with the spatial coordinates and reflection intensity features of the vertex as the internal features of the graph structure. Step 3: Inference of the Graph Neural Network The graph structure data is input into the graph neural network; in the graph neural network, edge features are extracted by a multilayer perceptron (MLP); convolution operation is performed by a graph attention network (GAT) to aggregate information of adjacent vertices, and attention coefficients are calculated based on the extracted edge features and vertex features; the features of the vertices are updated by the MLP based on the attention coefficients; after passing through multiple layers of the network, the output is a prediction of whether a vertex is noise. Step 4: Denoising and Reconstruction of Point Clouds Based on the prediction results of the vertices, each vertex is determined to be either noise or an object, and the point cloud signal predicted to be an object is recovered based on the index relationship between the vertex and the point cloud.

2. The method for removing lidar noise points under severe weather conditions based on graph neural networks according to claim 1, characterized in that, In step 1, a frame of point cloud signal acquired by the lidar is represented as an M×4 array as the preprocessed point cloud signal P, where M is the number of point clouds in a frame and 4 represents the four feature dimensions of the point cloud, including: the spatial coordinates X of the x-axis, y-axis and z-axis of the rectangular coordinate system centered on the lidar and the laser reflection intensity I.

3. The method for removing lidar noise points under severe weather conditions based on graph neural networks according to claim 1, characterized in that, The specific process of step 2 is as follows: Downsampling M point clouds yields N vertices, represented as: X←RandomSampling(P)(1) where RandomSampling represents the random downsampling operation, P represents the preprocessed point cloud signal, and X represents the vertex space coordinates obtained by downsampling; Based on the spatial coordinates X of the vertex and the spatial coordinates X of the point cloud, each point cloud is divided to its nearest neighbor vertex, represented as: Indices←Argmin(Distance(P,X)) (2) Where Distance represents the distance between the point cloud and the vertex calculated based on P and X, and Indices is the index relationship established between the vertex and the point cloud; Simultaneously, based on Indices, the center spatial position and average reflection intensity of each vertex are calculated. Local density features are extracted by the mean and number of point cloud spacings within the vertex. The normal vector within the point cloud is calculated as the local direction vector. The center spatial position, average reflection intensity, local density features, and local direction vector of the vertex are vector-linked to form the vertex features. Within a defined search radius, if two vertices are spatially adjacent, then an edge exists connecting them; for the i-th vertex, this is represented as: Among them, R i X represents the search radius of the i-th vertex, γ is the scaling factor of the search radius, and X... i Let || represent the spatial position of the center of the i-th vertex, || represents the L2 norm, and NeighborsWithRadiusSearch(X) represents the position of the center of the center of the i-th vertex. i ,R i ) represents the search radius of the i-th vertex as R. i Neighbor search, E i Let be the set of edges connected to the i-th vertex.

4. The method for removing lidar noise points under severe weather conditions based on graph neural networks according to claim 3, characterized in that, In step 2, the features of the vertices are specifically represented as follows: Where Indices[j] represents the vertex index of the j-th point cloud partition, i is the vertex index, 1≤i≤N, and Index represents the set of all point cloud indices associated with the i-th vertex; i Let I[j] represent the reflection intensity characteristic of the i-th vertex, I[j] represent the laser reflection intensity of the j-th point cloud, and Mean represent the average value; X i Let X[j] represent the center spatial position of the i-th vertex, and let X[j] represent the spatial coordinates of the j-th point cloud; D i F represents the local density feature of the i-th vertex, Count represents the number of points in the current index, and Concat represents the concatenation operation on the vectors; i V represents the local direction vector of the i-th vertex, and NormalVector represents the normal vector calculated based on the spatial position; i This represents the characteristic of the i-th vertex.

5. The method for removing lidar noise points under severe weather conditions based on graph neural networks according to claim 3, characterized in that, In step 2, the scaling factor γ of the search radius is set to 0.

01.

6. The method for removing lidar noise points under severe weather conditions based on graph neural networks according to claim 4, characterized in that, The specific process of step 3 is as follows: Graph-structured data is represented as G = (V, E), where V represents vertex features and E represents the connections between vertices; in the t-th iteration of the graph neural network, the edge features between any two adjacent vertices i and k are... for: in, and This represents the features of the i-th and k-th vertices in the t-th iteration. and Let x represent the multilayer perceptron used to extract edge features and vertex features, respectively. i and x k These represent the center spatial positions of the i-th vertex and the k-th vertex, respectively. This represents the translation correction amount corresponding to the i-th vertex in the t-th iteration; In the t-th iteration, the attention coefficient between the i-th vertex and the k-th vertex. Represented as: in, This represents a multilayer perceptron that performs nonlinear transformations; LeakyReLU represents a class of nonlinear activation functions; and Softmax represents a normalized exponential function. The feature of the i-th vertex updated in the (t+1)th iteration is: for: in, Represented as a multilayer perceptron with nonlinear operations, (i,k)∈E means obtaining the connections of all vertices connected to vertex i based on the edge connection relationship E; This completes one iteration for the i-th vertex. One iteration requires updating the features of all vertices in the graph structure G. The iteration process will run for T rounds, where T is the set number of rounds. The probability that the i-th vertex is predicted as noise is expressed as: p i =Sigmoid(Linear(v i )) (8) Where, p i v represents the probability that the i-th vertex is predicted as noise. i Let represent the feature of the i-th vertex after iteration, Linear represents a single-layer linear transformation of the vector, and Sigmoid represents a class of activation functions that normalize the probability.

7. The method for removing lidar noise points under severe weather conditions based on graph neural networks according to claim 1, characterized in that, The specific process of step 4 is as follows: If the probability p of the i-th vertex being predicted as noise i The value is greater than or equal to the set threshold. th If the vertex is determined to be noise caused by snow, it will not be reconstructed during the reconstruction process; if the predicted probability p i The value is less than snow th If the vertex is identified as an object, then the point cloud represented by that vertex is reconstructed. Reconstructing point cloud P out The process can be characterized as: Among them, P j This is the preprocessed point cloud signal associated with the j-th point cloud of the i-th vertex.

8. The method for removing lidar noise points under severe weather conditions based on graph neural networks according to claim 7, characterized in that, In step 4, a threshold value for snow is set. th Set it to 0.3.

Images