Method for removing laser radar noise points in severe weather based on graph neural network

Abstract

The invention discloses a method for removing laser radar noise points in severe weather based on a graph neural network, and belongs to the technical field of image processing. The method comprises the following steps: firstly, constructing a graph structure through point cloud data acquired by a laser radar, wherein the graph structure comprises vertex attributes of a spatial position, reflection intensity, local spatial density and local direction features and a side relation for connecting vertexes; and then, performing noise prediction on each vertex through graph convolution by using a graph neural network. Thirdly, the point cloud is recovered through the corresponding relation between the vertex and the point cloud, and the noise point cloud is removed; according to the method, end-to-end noise removal can be effectively realized on the premise that a point cloud data structure is not damaged, so that the accuracy of laser radar data is improved. Through experimental verification on a CADC data set, the method disclosed by the invention shows remarkable performance improvement, and has relatively high practicability and popularization value.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method for removing laser radar noise points in bad weather based on a graph neural network. Background Art

[0002] LiDAR (Light Detection and Ranging) is a type of active sensor that emits laser beams and detects echoes to achieve distance detection, generating point cloud data with three-dimensional coordinates. It has the characteristics of high measurement accuracy and long detection distance, and is widely used in fields such as autonomous driving and remote sensing mapping. In bad weather, such as rain and snow, particles in the environment will scatter the light beam emitted by the LiDAR, resulting in noise points. This type of noise point will interfere with downstream applications. For example, it will lead to the inability to identify obstacles in autonomous driving applications, and it will affect the mapping results in the field of surveying and mapping.

[0003] Current research in the field of point cloud denoising can be divided into two categories: rule-based classifiers and convolutional neural networks based on image processing. Rule-based classifiers are based on low noise distribution density and weak reflection intensity. The Radius Outlier Removal (ROR) method determines noise points based on the number of point cloud neighbors, and removes them if the number of points in the neighborhood is lower than the set threshold. The Statistical Outlier Removal (SOR) method determines noise points by calculating the average distance between a point and its k nearest neighbors. However, this type of method requires adjusting parameter settings according to different lidars, and the classification accuracy is not high.

[0004] Based on image processing, convolutional neural networks transform the point cloud denoising problem into an image processing problem by projecting the three-dimensional point cloud onto a two-dimensional plane. For example, Weathernet uses a convolutional neural network (CNN) to denoise point clouds under weather conditions, which can effectively process noisy point clouds affected by weather conditions such as rain and fog. In addition, LiSnowNet uses unsupervised learning to evaluate the network output signal through Fourier transform and wavelet transform to achieve unsupervised denoising of point clouds. However, when projecting the point cloud onto a two-dimensional grid, due to the irregular distribution of the point cloud, the point cloud information in some grids will be lost, and the lack of point cloud information in some grids will lead to a waste of computing resources. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for removing lidar noise points in severe weather based on graph neural networks. It proposes applying graph neural networks to the relevant theories of lidar point cloud denoising to achieve the removal of point cloud noise caused by severe weather and provide clean data for downstream applications.

[0006] The technical problem proposed by the present invention is solved in this way:

[0007] A method for removing laser radar noise points in bad weather based on graph neural network, comprising the following steps:

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

[0009] A frame of point cloud signal collected by the laser radar 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: Construction of graph structure;

[0011] Downsample the point cloud to construct vertices, and construct the edge relationship between vertices through dynamic radius search; extract the normal vector within the vertex, extract the number of point clouds within the vertex and the spatial density as local density features, and combine the spatial coordinates and reflection intensity features of the vertex as the internal features of the vertex of the graph structure;

[0012] Step 3: Reasoning of Graph Neural Network

[0013] Input the graph structure data into the graph neural network; in the graph neural network, extract edge features through the multi-layer perceptron MLP; perform convolution operations through the attention graph network GAT to achieve information aggregation of adjacent vertices, and calculate the attention coefficient based on the extracted edge features and vertex features; update the vertex features based on the attention coefficient through the multi-layer perceptron MLP; output the prediction of whether the vertex is noise through the multi-layer network;

[0014] Step 4: Denoising and reconstruction of point cloud

[0015] According to the prediction results of the vertices, each vertex is determined to be noise or an object, and the point cloud signal predicted to be an object is restored according to the index relationship between the vertex and the point cloud.

[0016] Furthermore, in step 1, a frame of point cloud signals collected by the laser radar is represented as an M×4 array as a preprocessed point cloud signal P, where M is the number of point clouds in a frame, and 4 represents the four characteristic 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 laser radar and the laser reflection intensity I.

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

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

[0019] X←RandomSampling(P)(1)

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

[0021] According to the vertex spatial coordinates X and the point cloud spatial coordinates X, each point cloud is divided into the nearest neighbor vertices, expressed as:

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

[0023] Among them, Distance indicates 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] At the same time, according to the Indices, the central spatial position and average reflection intensity of each vertex will be calculated, the local density features will be extracted through the mean and number of point cloud spacing within the vertex, and the normal vector in the point cloud will be calculated as the local direction vector; the central spatial position, average reflection intensity, local density features and local direction vector of the vertex will be vector-linked as the features of the vertex.

[0025] Within the set search radius, if two vertices are adjacent in space, there is an edge connection; for the i-th vertex, it is expressed as:

[0026]

[0027] Among them, R i represents the search radius of the i-th vertex, γ is the scaling factor of the search radius, and X i represents the central spatial position of the i-th vertex, || || represents the search for the second norm, NeighborsWithRadiusSearch(X i ,R i ) indicates that the search radius for the i-th vertex is R i Neighbor search, E i is the set of edges connected to the i-th vertex.

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

[0029]

[0030] Among them, 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 serial numbers associated with the i-th vertex; I i represents the reflection intensity feature of the i-th vertex, I[j] represents the laser reflection intensity of the j-th point cloud, and Mean represents the average value; X irepresents the central spatial position of the i-th vertex, X[j] represents the spatial coordinates of the j-th point cloud; D i represents the local density feature of the i-th vertex, Count represents the number of point clouds in the current index, and Concat represents the concatenation operation of vectors; F i V represents the local direction vector of the i-th vertex, NormalVector represents the normal vector calculated according to the spatial position; i Represents the characteristics 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] The graph structure data is represented as G = (V, E), where V represents the vertex feature and E represents the connection relationship between vertices; in the tth iteration of the graph neural network, the edge feature between any two adjacent vertices i and k for:

[0034]

[0035] in, and represents the features of the i-th vertex and the k-th vertex in the t-th iteration, and Respectively represent the multi-layer perceptrons used to extract edge features and vertex features to achieve nonlinear transformation, x i and x k Respectively represent the central spatial positions of the i-th vertex and the k-th vertex, Represents the translation correction corresponding to the i-th vertex in the t-th iteration;

[0036] In the tth iteration, the attention coefficient between the i-th vertex and the k-th vertex It is expressed as:

[0037]

[0038] in, represents a multilayer perceptron that performs nonlinear transformation, LeakyReLU represents a type of nonlinear activation function, and Softmax represents a normalized exponential function;

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

[0040]

[0041] in, It is represented as a multi-layer perceptron of nonlinear operation, (i,k)∈E means obtaining the connections of all vertices connected to vertex i according to the edge connection relationship E;

[0042] At this point, one round of iteration for the i-th vertex is completed. One round of iteration requires updating the features of all vertices in the graph structure G. The iteration process will be repeated 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] Among them, p i represents the probability that the i-th vertex is predicted as noise, v i It represents the feature of the i-th vertex after the iteration is completed, Linear represents a single-layer linear transformation of the vector, and Sigmoid represents a type of activation function that normalizes the probability.

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

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

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

[0049]

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

[0051] Furthermore, in step 4, the threshold snow is set th Set to 0.3.

[0052] The beneficial effects of the present invention are:

[0053] (1) The present invention innovatively uses 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 a graph neural network to remove noise from lidar point clouds under severe weather conditions, and has high denoising accuracy, that is, accurately distinguishing between noise point clouds and object point clouds.

[0055] (3) The present invention provides a 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 vertex features of the graph neural network based on the differences between noise and objects, thereby achieving higher denoising accuracy.

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

[0058] (6) The present invention is not only simple to implement and easy to operate, but also has strong practicality and is suitable for popularization and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of the process of the method of the present invention;

[0060] Figure 2 The diagram shows the results of data / image processing in each step of the method described in the embodiment.

[0061] Figure 3 The figure shows the results of removing noise caused by different bad weather conditions in the method described in the embodiment. DETAILED DESCRIPTION

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

[0063] This embodiment provides a method for removing laser radar noise points in bad weather based on a graph neural network, and its flow chart is as follows: Figure 1 As shown in the figure, the graph neural network is used as a classifier, which can adapt to the irregularity of the point cloud and achieve efficient denoising. This method has the advantages of high denoising accuracy and does not destroy the point cloud structure. This graph neural network-based LiDAR point cloud denoising algorithm in bad weather 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 of the algorithm is: a frame of point cloud signal collected by the lidar.

[0067] In step 1, a frame of point cloud signals collected by the laser radar is sorted and deformed, which 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 characteristic 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 laser radar and the laser reflection intensity I.

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

[0069] Step 2: Construction of graph structure;

[0070] The point cloud is downsampled to construct vertices, and the edge relationship between vertices is constructed through dynamic radius search; the normal vector within the vertex, the number of point clouds within the vertex and the spatial density are extracted as local density features, and the spatial coordinates and reflection intensity features of the vertex are combined as the internal features of the vertex of the graph structure.

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

[0072] First, the M point clouds in step 1 are downsampled to N vertices:

[0073] X←RandomSampling(P)(1)

[0074] Among them, 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, each point cloud is divided into the nearest neighbor vertices according to the vertex spatial coordinates X and the point cloud spatial coordinates X, expressed as:

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

[0077] Wherein, Distance(·) indicates 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] At the same time, the central spatial position and average reflection intensity of each vertex are calculated according to the Indices, the local density features are extracted by the mean and number of the point cloud spacing within the vertex, and the normal vector in the point cloud is calculated as the local direction vector; the central spatial position, average reflection intensity, local density features and local direction vector of the vertex are vector-linked as the features of the vertex; specifically expressed as:

[0079]

[0080] Among them, 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 serial numbers associated with the i-th vertex; I i represents the reflection intensity feature of the i-th vertex, I[j] represents the laser reflection intensity of the j-th point cloud, and Mean represents the average value; X i represents the central spatial position of the i-th vertex, X[j] represents the spatial coordinates of the j-th point cloud; D i represents the local density feature of the i-th vertex, Count represents the number of point clouds in the current index, and Concat represents the concatenation operation of vectors; F i represents the local direction vector of the ith vertex, NormalVector represents the normal vector calculated according to the spatial position, and this embodiment uses the principal component analysis method to implement it; V i Represents the characteristics of the i-th vertex.

[0081] The graph structure includes vertices and edges. The feature construction of the vertices has been completed through the above steps. The edge connection relationship E can be expressed as a dynamic radius search problem, that is, within a given search radius, if two vertices are adjacent in space, there is an edge connection; for the i-th vertex, it is expressed as:

[0082]

[0083] Among them, R i NeighborsWithRadiusSearch(X i ,R i ) indicates that the search radius for the i-th vertex is R i Neighbor search, this embodiment uses binary tree search to implement, E i is the set of edges connected to the i-th vertex.

[0084] In this real-time example, the number of downsampling N is set to 5000. Each point cloud is divided into the nearest neighbor vertices according to the spatial coordinates of the vertex and the point cloud. Indices indicates the index relationship between the vertex and the point cloud, which serves as the basis for restoring the point cloud in step 4. At the same time, the central spatial position and average reflection intensity of the vertex are calculated according to Indices as the spatial feature X and reflection intensity feature of the vertex The local density feature D is extracted by the mean and number of point cloud spacing within the vertex, and the normal vector in the point cloud is calculated as the local direction vector F. The features of the vertex can be obtained by vector connection. The graph structure includes vertices and edges, and the vertex feature construction has been completed through the above steps. The edge connection relationship E can be expressed as a dynamic radius search problem, that is, within a given radius, if two vertices are adjacent in space, there is an edge connection. In this embodiment, the scaling factor γ of the Velodyne 64-line laser radar is set to 0.01. At this point, step 2 completes the construction of the graph structure, and the graph structure data is as follows Figure 2 As shown in the “Graph Structure Data” in .

[0085] Step 3: Reasoning of Graph Neural Network

[0086] Input the graph structure data into the graph neural network; the application of graph neural network in the field of point cloud denoising is more suitable for the structure of point cloud, which will completely avoid the loss of information and waste of computing resources. In addition, the denoising method of graph neural network performs better in denoising accuracy and can more accurately identify the type of point cloud.

[0087] In the graph neural network, edge features are extracted through the multi-layer perceptron MLP; the convolution operation is performed through the attention graph network GAT to realize the information aggregation of adjacent vertices, and the attention coefficient is calculated based on the extracted edge features and vertex features; the multi-layer perceptron MLP is used to update the vertex features based on the attention coefficient; and the multi-layer network outputs the prediction of whether the vertex is noise.

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

[0089] Graph structure data can be represented as G = (V, E), where V represents vertex features and E represents the connection relationship between vertices. In the graph convolutional network, vertex features are refined through multi-layer perceptrons, and information aggregation is achieved through graph convolution with attention mechanism (GAT) on vertex features and edge features.

[0090] In the tth iteration of the graph neural network, the edge feature between any two adjacent vertices i and k is The calculation process is:

[0091]

[0092] in, and represents the features of the i-th vertex and the k-th vertex in the t-th iteration, and Respectively represent the multi-layer perceptrons used to extract edge features and vertex features to achieve nonlinear transformation, x i and x k Respectively represent the central spatial positions of the i-th vertex and the k-th vertex, Represents the translation correction corresponding to the i-th vertex in the t-th iteration.

[0093] According to the features of the extracted edges, by introducing the attention coefficient, it is possible to extract features of different weights for different adjacent nodes. In the tth iteration, the attention coefficient between the i-th vertex and the k-th vertex The calculation process can be expressed as:

[0094]

[0095] in, Represents a multilayer perceptron that performs nonlinear transformation, LeakyReLU represents a type of nonlinear activation function, and Softmax represents a normalized exponential function.

[0096] According to the extracted attention coefficient, the vertex features can be updated through convolution operation. The feature of the i-th vertex updated in the t+1th round of iteration is can be expressed as:

[0097]

[0098] in It is represented as a multi-layer perceptron of nonlinear operation, (i,k)∈E means obtaining the connections of all vertices connected to vertex i according to the edge connection relationship E;

[0099] At this point, one iteration of the i-th vertex i is completed; one iteration requires updating the features of all vertices in the graph structure G. In step 3, the above iterative process will be performed T times to refine the vertex features.

[0100] According to the refined vertex features, the prediction of whether the vertex is noise can be obtained through linear transformation. The probability prediction process of whether the i-th vertex is noise can be expressed as:

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

[0102] Among them, p i represents the probability that the i-th vertex is predicted as noise, v i It represents the feature of the i-th vertex after the iteration is completed, Linear represents a single-layer linear transformation of the vector, and Sigmoid represents a type of activation function that normalizes the probability.

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

[0104] Step 4: Denoising and reconstruction of point cloud

[0105] Whether the vertex is noise is determined based on the prediction result of the vertex, and the point cloud signal predicted to be an object is restored 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 that the i-th vertex is predicted as noise is p i The value is greater than or equal to the set threshold snow th The vertex is judged to be the noise caused by snow, and the point cloud represented by the vertex will not be restored in the reconstruction process; on the contrary, when the predicted probability p i The value is less than snow th , then the vertex is determined to be an object, and the point cloud represented by the vertex is reconstructed.

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

[0109]

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

[0111] Through the above steps, an end-to-end severe weather point cloud denoising process based on graph neural network is realized.

[0112] In this embodiment, snow th is set to 0.3. The reconstructed result can be obtained by Figure 2 “Reconstructed point cloud data” is represented in the figure.

[0113] In summary, the present invention innovatively uses a method based on a graph neural network to denoise the point cloud signals of a lidar in severe weather. Different from the existing denoising methods of rule-based and image-based convolutional neural networks, the present invention can achieve a high-precision point cloud denoising effect without destroying the point cloud structure, that is, accurately distinguishing between noise point clouds and object point clouds. In addition, the present invention extracts the vertex features of the graph neural network based on the difference between noise and objects, thereby further improving the denoising effect. Not only that, the present invention can handle point cloud noise caused by more types of severe weather, such as rain, snow, fog, etc. Figure 3 As shown, from left to right, the denoising effect of the laser radar point cloud noise brought by the present invention in a snowy environment, a rainy environment, and a foggy environment is shown. The present invention is simple to implement, does not have a deep network structure, is easy to operate, and has strong practicality and scalability.

[0114] The structures listed above are only basic structures of the present invention as examples and should not be used to limit the scope of protection of the present invention. Any changes or modifications that are made to the main design concept and spirit of the present invention and have no substantive significance, and the technical problems they solve are still consistent with the present invention, should be included in the scope of protection of the present invention.

Claims

1. A method for removing laser radar noise points in bad weather based on graph neural network, characterized in that: The following steps are involved: Step 1: pre-process the point cloud signal; A frame of point cloud signal collected by the laser radar 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: Construction of graph structure; Downsample the point cloud to construct vertices, and construct the edge relationship between vertices through dynamic radius search; extract the normal vector within the vertex, extract the number of point clouds within the vertex and the spatial density as local density features, and combine the spatial coordinates and reflection intensity features of the vertex as the internal features of the vertex of the graph structure; Step 3: Reasoning of Graph Neural Network Input the graph structure data into the graph neural network; in the graph neural network, extract edge features through the multi-layer perceptron MLP; perform convolution operations through the attention graph network GAT to achieve information aggregation of adjacent vertices, and calculate the attention coefficient based on the extracted edge features and vertex features; update the vertex features based on the attention coefficient through the multi-layer perceptron MLP; output the prediction of whether the vertex is noise through the multi-layer network; Step 4: Denoising and reconstruction of point cloud According to the prediction results of the vertices, each vertex is determined to be noise or an object, and the point cloud signal predicted to be an object is restored according to the index relationship between the vertex and the point cloud.

2. The method for removing laser radar noise points in bad weather based on graph neural network according to claim 1 is characterized in that: In step 1, a frame of point cloud signal collected by the laser radar 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 characteristic 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 laser radar and the laser reflection intensity I.

3. The method for removing laser radar noise points in bad weather based on graph neural network according to claim 1 is characterized in that: The specific process of step 2 is: Downsampling the M point clouds yields N vertices, expressed 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; According to the vertex spatial coordinates X and the point cloud spatial coordinates X, each point cloud is divided into the nearest neighbor vertices, expressed as: Indices←Argmin(Distance(P,X)) (2) Among them, Distance indicates 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; At the same time, according to the Indices, the central spatial position and average reflection intensity of each vertex will be calculated, the local density features will be extracted through the mean and number of point cloud spacing within the vertex, and the normal vector in the point cloud will be calculated as the local direction vector; the central spatial position, average reflection intensity, local density features and local direction vector of the vertex will be vector-linked as the features of the vertex. Within the set search radius, if two vertices are adjacent in space, there is an edge connection; for the i-th vertex, it is expressed as: Among them, R i represents the search radius of the i-th vertex, γ is the scaling factor of the search radius, and X i represents the central spatial position of the i-th vertex, || || represents the search for the second norm, NeighborsWithRadiusSearch(X i ,R i ) indicates that the search radius for the i-th vertex is R i Neighbor search, E i is the set of edges connected to the i-th vertex.

4. The method for removing laser radar noise points in bad weather based on graph neural network according to claim 3 is characterized in that: In step 2, the features of the vertices are specifically expressed as: Among them, 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 serial numbers associated with the i-th vertex; I i represents the reflection intensity feature of the i-th vertex, I[j] represents the laser reflection intensity of the j-th point cloud, and Mean represents the average value; X i represents the central spatial position of the i-th vertex, X[j] represents the spatial coordinates of the j-th point cloud; D i represents the local density feature of the i-th vertex, Count represents the number of point clouds in the current index, and Concat represents the concatenation operation of vectors; F i V represents the local direction vector of the i-th vertex, NormalVector represents the normal vector calculated according to the spatial position; i Represents the characteristics of the i-th vertex.

5. The method for removing laser radar noise points in bad weather based on graph neural network according to claim 3 is characterized in that: In step 2, the scaling factor γ of the search radius is set to 0.

01.

6. The method for removing laser radar noise points in bad weather based on graph neural network according to claim 4 is characterized in that: The specific process of step 3 is: The graph structure data is represented as G = (V, E), where V represents the vertex feature and E represents the connection relationship between vertices; in the tth iteration of the graph neural network, the edge feature between any two adjacent vertices i and k for: in, and represents the features of the i-th vertex and the k-th vertex in the t-th iteration, and Respectively represent the multi-layer perceptrons used to extract edge features and vertex features, x i and x k Respectively represent the central spatial positions of the i-th vertex and the k-th vertex, Represents the translation correction corresponding to the i-th vertex in the t-th iteration; In the tth iteration, the attention coefficient between the i-th vertex and the k-th vertex It is expressed as: in, represents a multilayer perceptron that performs nonlinear transformation, LeakyReLU represents a type of nonlinear activation function, and Softmax represents a normalized exponential function; The feature of the i-th vertex updated in the t+1th round of iteration is for: in, It is represented as a multi-layer perceptron of nonlinear operation, (i,k)∈E means obtaining the connections of all vertices connected to vertex i according to the edge connection relationship E; At this point, one round of iteration for the i-th vertex is completed. One round of iteration requires updating the features of all vertices in the graph structure G. The iteration process will be repeated 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) Among them, p i represents the probability that the i-th vertex is predicted as noise, v i It represents the feature of the i-th vertex after the iteration is completed, Linear represents a single-layer linear transformation of the vector, and Sigmoid represents a type of activation function that normalizes the probability.

7. The method for removing laser radar noise points in bad weather based on graph neural network according to claim 1 is characterized in that: The specific process of step 4 is: If the probability that the i-th vertex is predicted as noise is p i The value is greater than or equal to the set threshold snow th , then the vertex is judged to be the noise caused by snow, and the point cloud represented by the vertex will not be restored during the reconstruction process; if the predicted probability p i The value is less than snow th , then the vertex is determined to be an object, and the point cloud represented by the vertex is reconstructed; Reconstruct point cloud P out The process can be characterized as: Among them, P j is the preprocessed point cloud signal of the j-th point cloud associated with the i-th vertex.

8. The method for removing laser radar noise points in bad weather based on graph neural network according to claim 7 is characterized in that: In step 4, set the threshold snow th Set to 0.3.

Images