Denoising method of mechanical lidar point cloud data in snowy driving environment

Abstract

The present invention discloses a point cloud data denoising method suitable for mechanical laser radar in a snowy driving environment, belonging to the field of autonomous driving. The method processes point cloud data acquired by the laser radar, converting it into a point cloud depth information tensor; annotates the point cloud data to obtain label data; combines the point cloud depth information tensor with the label data to create a dataset; processes the dataset using convolution, residual attention, and discrete wavelet transform processing, optimizing weight parameters in the convolution and residual attention processing using a training set in the dataset; and evaluates processing methods containing trained weight parameters using a test set in the dataset. In actual application, the method input dimensions are adjusted to match the real-world data dimensions, and the method outputs a point cloud category prediction value, eliminating laser radar data points predicted as snowflakes from the original data.

Description

Technical Field

[0001] The present invention relates to a method for denoising mechanical laser radar point cloud data in a snowy driving environment, and belongs to the field of autonomous driving. Background Art

[0002] Autonomous vehicles use sensors such as lidar, cameras, and millimeter-wave radar to obtain information about surrounding vehicles, pedestrians, and other objects. The point cloud data generated by lidar offers long-range measurement, high position accuracy, and minimal influence from illumination. This makes it an effective source of information for autonomous driving environmental perception and can be used to develop various algorithms, including 3D object detection and scene reconstruction.

[0003] When driving in snowy conditions, falling snow affects LiDAR detection through absorption and scattering. Snowflakes can be considered diffuse solid objects, typically smaller than the laser beam cross-section. Small flakes produce significant backscattering. This interference can affect the point cloud data generated by LiDAR, compromising the accuracy of autonomous vehicle environmental perception algorithms such as obstacle detection in snowy conditions, and jeopardizing the safety of autonomous vehicles operating in snowy conditions. Denoising LiDAR point cloud data in snowy conditions can effectively remove snow noise from the point cloud data, maintaining the stability of autonomous vehicle environmental perception algorithms and improving the safety of autonomous vehicles operating in snowy conditions.

[0004] Existing research on unmanned driving environment perception rarely involves snowy environment conditions. CN202210632703.0 discloses a visual automatic driving control system, but its input source is an image rather than a three-dimensional point cloud generated by a lidar. Existing filtering-based radar point cloud denoising methods such as SOR, ROR, and DROR can filter out noise points in three-dimensional point clouds, but these methods are not specifically designed for snow point removal in snowy driving environments, and therefore are prone to failure in the case of a large number of unevenly distributed snowflakes. CN202210034793.3 discloses a lidar system and its ambient light denoising method, but does not involve the impact of snowflakes in the air on lidar data. DSOR (Kurup A, Bos J. Dsor: A scalable statistical filter for removing falling snow from lidar point clouds in severe winter weather [J]. arXiv preprint arXiv:2109.07078, 2021.) is a recently developed radar point cloud denoising method specifically for snowy environments. It is highly effective in removing snowflake noise, but as a filter-based method, it is typically deployed on CPUs, leaving room for improvement in detection speed. Compared to filter-based methods, deep learning-based methods can be accelerated using graphics processing units (GPUs), offering greater potential for both accuracy and speed. LiSnowNet (Yu MY, Vasudevan R, Johnson-Roberson M. LiSnowNet: Real-time Snow Removal for LiDAR Point Clouds [C] / / 2022 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022: 6820-6826.) is a method that uses a two-dimensional convolutional model to denoise snow-containing noisy point clouds. It surpasses DSOR in both speed and accuracy. However, it does not use annotated point cloud data and only optimizes the model by designing a loss function without labeled data. 4DenoiseNet ( A, OJALA R, TAMMI K. 4DenoiseNet: Adverse Weather Denoising from Adjacent Point Clouds[J / OL]. IEEE Robotics and Automation Letters, 2023, 8(1): 456-463.) also converts three-dimensional point clouds into two-dimensional data for processing, but it requires two frames of point cloud data before and after to run, which limits its operating conditions. Summary of the Invention

[0005] The present invention aims to address the safety hazard caused by snowflakes affecting the laser radar (LiDAR) of autonomous vehicles in snowy driving environments. A method for denoising point cloud data from a mechanical LiDAR in a snowy driving environment is provided. The method processes point cloud data acquired by the LiDAR and converts it into a point cloud depth information tensor; labels the point cloud data to obtain label data; combines the point cloud depth information tensor with the label data to create a dataset; processes the dataset using convolution, residual attention, and discrete wavelet transform (DWT); optimizes weight parameters in the convolution and residual attention processes using a training set within the dataset; and evaluates the processing method containing the optimized weight parameters using a test set within the dataset. In practical application, the method's input dimensions are adjusted to match the dimensions of real-world data, and the method outputs a point cloud category prediction value, eliminating LiDAR data points predicted as snowflakes from the original data.

[0006] The objectives of the present invention are achieved through the following technical solutions.

[0007] The method for denoising mechanical lidar point cloud data in a snowy driving environment includes the following steps:

[0008] Step 1: Collect the point cloud data generated by the lidar in the snowy driving environment and construct the depth information tensor of the point cloud data; annotate the point cloud data and generate a dataset for the mechanical lidar point cloud denoising method;

[0009] 1) Process the point cloud data and construct the depth information tensor of the point cloud data; pre-process the point cloud data generated by the lidar in the snow driving environment, and assume that a frame of point cloud data generated by the lidar is a two-dimensional matrix in represents the real number domain space, n represents the number of points generated by one frame of laser radar point cloud data; (x i ,y i ,z i )| i∈n is the i-th row data in the matrix M, representing the i-th point in a frame of point cloud data, where x i ,y i, z i The coordinates of the point in the Cartesian coordinate system with the laser radar as the origin are x-axis, y-axis, and z-axis respectively; the two-dimensional matrix is ​​converted into a point cloud depth information tensor containing point cloud intensity information, and the point cloud data generated by the laser radar in the snow driving environment is collected to obtain a laser radar point cloud set with sufficient frames {M1,M2,M3,...,M n};

[0010] Step 1.1, for the lidar point cloud set {M1,M2,M3,...,M n} to process any point cloud two-dimensional matrix in the set Convert it from Cartesian to cylindrical coordinates:

[0011] θ i =arctan(-y i ,x i ) (1)

[0012]

[0013]

[0014] r i =f hash (φ i ) (4)

[0015] where θ i ,φ i are the azimuth and elevation angles of point i in the cylindrical coordinate system, △θ represents the scanning resolution of the mechanical laser radar in the horizontal and vertical directions, and f hash represents the lidar elevation angle resolution mapping function; x i ,y i , z i are the coordinates of the point in the Cartesian coordinate system with the lidar as the origin, i.e., the x-axis, y-axis, and z-axis. The cylindrical coordinate system is expanded along the z-axis with △θ as the interval to obtain a plane grid without filling value, r i and c i Respectively represent the horizontal and vertical correspondence of point i in the grid;

[0016] Step 1.2, r i with c i If there is a negative value, r i with c i After transformation, the row and column indices of the grid are mapped, and:

[0017] x min =-π / △θ (5)

[0018] y min=minf hash (φ i ) (6)

[0019] Among them, x min is the minimum angle value in the yaw direction, y min is the minimum angle value of the laser radar in the pitch direction;

[0020] Transforming equations (3) and (4) yields equations (7) and (8):

[0021]

[0022]

[0023] and are the row and column indices of the grid;

[0024] Step 1.3, get the grid and The value at

[0025]

[0026] All indexes of the raster map do not exist The value of is set to 99.0, indicating that the depth value at this location cannot be obtained; the grid of the obtained point cloud data is set to V, where V is a two-dimensional matrix;

[0027] Step 1.4, for any lidar point (x i ,y i ,z i )| i∈n , each has an echo intensity value p i Correspondingly, construct a matrix P of the same size as V, with the index in P and The value at p i ; Set the index in P that does not exist i The value of is set to 0, indicating no echo intensity; V and P are stacked to form the point cloud depth information tensor Where w and h are the sizes of T in the first two dimensions respectively; each M corresponds to a T, and the point cloud depth information tensor set {T1, T2, T3, ..., T n};

[0028] 2) Annotate the point cloud data,

[0029] For the lidar point cloud set {M1,M2,M3,...,M n} to mark, Add a column based on the data, where the element 1 represents snowflakes and 0 represents normal data, forming a matrix containing labeling information.

[0030] To S i The first three columns of data in the table are used to generate the index of the point cloud data annotation grid using equations (1) to (8), and the value at the index is set to S i The fourth column value at the corresponding index is used to get the label matrix For {M1,M2,M3,...,M n}, we can get the label set {Ya1, Ya2, Ya3,..., Ya n};

[0031] 3) T i and Ya i Combine them and get the data set {(T1,Ya1),(T2,Ya2),...,(T n ,Ya n )}; Generate the training data set required for the radar point cloud denoising method and divide it into a training set and a test set;

[0032] Step 2: The training set obtained in step 1 is processed using convolution, discrete wavelet transform, residual attention, and discrete inverse wavelet transform; the weight parameters in convolution, residual attention, and discrete inverse wavelet transform are optimized through forward and backward propagation processes.

[0033] Split the training set into several subsets, where each subset contains bs Ts and Yas. Let the set of Ts in each subset be X and the set of Yas be Y. Then perform the following processing on any X:

[0034] 1) Assume that the convolution process is:

[0035] Mid conv =conv(In conv ) (10)

[0036] Mid bn =BN(Mid conv ) (11)

[0037] Out conv =max(0,Mid bn ) (12)

[0038] The conv in (10) represents the two-dimensional convolution kernel operation. conv Represents the input of the convolution process, Mid conv represents the intermediate output of (10), BN in (11) represents the batch normalization operation, Mid bnIndicates the intermediate output of (11), and max in (12) indicates the output of Mid bn All elements in the value greater than 0, Out conv Represents the output of the convolution process.

[0039] Take X as the input of the convolution process, perform two convolution processes, and get the output X1.

[0040] 2) Perform discrete wavelet transform on X1 to obtain output X2.

[0041] 3) Assume residual attention processing:

[0042]

[0043]

[0044]

[0045] (13)-(15), In att Represents the input of residual attention processing, Out att represents the output of residual attention, and is an intermediate variable, avgpool represents average pooling, fc represents a fully connected neural network operation, softmax represents a normalized exponential operation, and cba represents a convolution operation. X2 is used as input for residual attention processing to obtain X3.

[0046] 4) Perform discrete wavelet transform and residual attention processing on X3 to obtain X4.

[0047] 5) Perform discrete wavelet transform and two residual attention processes on X4 to obtain X5.

[0048] 6) Perform two residual attention processes and discrete inverse wavelet transform on X5 and then add it to X4 to obtain X6.

[0049] 7) Perform residual attention processing and discrete inverse wavelet transform on X6 and add it to X3 to obtain X7.

[0050] 8) Perform residual attention processing and discrete inverse wavelet transform on X7 and add it to X1 to obtain X8.

[0051] 9) Add X8 to X and use the sigmoid function for activation:

[0052]

[0053] In (16), e is a natural constant, is the output.

[0054] 10) And the corresponding Y is substituted into the loss function:

[0055]

[0056] In (17), F represents the F-norm.

[0057] 11) According to the calculation The gradient descent method is used to optimize the weight parameters in all convolution operations and residual attention operations in 1) to 10).

[0058] 12) For all subsets in the training set, loop 1) to 11), set a threshold v for the loss function, and when the loss function value is less than the threshold v, terminate the loop and save the weight parameters in all convolution operations and residual attention operations in 1) to 10).

[0059] Step 3: Use the test set obtained in step 1 to evaluate the weight parameters in the convolution processing and residual attention processing optimized in step 2 to obtain the weight parameters that meet the accuracy requirements;

[0060] Use the point cloud data denoising method with the optimized weight parameters in step 2 to process the lidar point cloud set of the test set, compare the output with the labeled data in the test set, and calculate the average precision (AP) of the denoising method on the test set. If the AP requirement is met, save the optimized weight parameters, otherwise repeat steps 2 and 3.

[0061] Step 4: Use the denoising method that meets the accuracy requirements in step 3 to denoise the snowy environment;

[0062] During the denoising process of the surrounding point cloud snowflakes, the bs value of the input data is set to 1 to ensure that the input dimension of the denoising method matches the input dimension of the real information; for the output of the denoising method The value range of each element should be between 0 and 1. If the element value is greater than the threshold k, the corresponding lidar point is considered to be a snowflake and is removed; if it is less than the threshold k, it is considered to be a normal environmental obstacle and is retained.

[0063] Beneficial effects:

[0064] 1. The technical solution of this invention utilizes a deep learning-based denoising method to denoise the snow noise generated by LiDAR data in snowy driving environments. By preprocessing the point cloud data generated by the LiDAR, the 3D point cloud information is mapped onto a 2D depth map. This preserves the complete point cloud information while avoiding the significant computational overhead of applying 3D convolution in the denoising process. Deep learning methods are used to accurately identify and filter out snow noise from the 2D depth map, achieving point cloud denoising in snowy driving environments.

[0065] 2. The snow denoising method disclosed in this invention is based on deep learning. By constructing a fully symmetrical convolutional neural network and rationally arranging the order of convolution processing and residual attention processing, it can effectively extract surrounding obstacles and snow features while ensuring the consistency of input and output. The entire method processes data conveniently and quickly. Accelerating the denoising method with a GPU can further increase the computational speed of the denoising method, meeting the real-time requirements of point cloud denoising.

[0066] 3. The snow denoising method disclosed in the present invention utilizes residual attention processing and discrete wavelet transform processing, which can improve the ability to extract snowflake features to a certain extent and avoid the denoising method from incorrectly eliminating non-snowflake point clouds when the snowflake distribution is atypical. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic diagram of the direction of a Cartesian coordinate system with the laser radar as the origin in step 1 of the denoising method disclosed in the present invention;

[0068] Figure 2 A schematic diagram of the coordinate directions of the radar point cloud depth map constructed in step 1 of the denoising method disclosed in the present invention;

[0069] Figure 3 Schematic diagram of the point cloud denoising method in step 2 of the denoising method disclosed in the present invention;

[0070] Figure 4 Schematic diagram of residual attention processing in the point cloud denoising method in step 2 of the denoising method disclosed by the present invention;

[0071] Figure 5 A comparison chart of denoising point cloud data containing snowflakes using the denoising method disclosed in the present invention. DETAILED DESCRIPTION

[0072] In order to better illustrate the purpose, content and advantages of the present invention, the invention is further described below with reference to the accompanying drawings and examples.

[0073] The present invention provides a point cloud denoising method suitable for a mechanical laser radar in a snowy driving environment. In the snowy environment, an unmanned vehicle equipped with a mechanical laser radar on its roof travels at medium or low speeds on the road.

[0074] Step 1: Collect the point cloud data generated by the lidar in the snowy driving environment and construct the depth information tensor of the point cloud data; annotate the point cloud data and generate a dataset for the mechanical lidar point cloud denoising method;

[0075] 1) Process the point cloud data and construct the depth information tensor of the point cloud data; pre-process the point cloud data generated by the lidar in the snow driving environment, and assume that a frame of point cloud data generated by the lidar is a two-dimensional matrix in represents the real number domain space, n represents the number of points generated by one frame of laser radar point cloud data; (x i ,y i ,z i )| i∈n is the i-th row data in the matrix M, representing the i-th point in a frame of point cloud data, where x i ,y i , z i They are the coordinates of the point in the x-axis, y-axis, and z-axis in the Cartesian coordinate system with the laser radar as the origin, and their specific directions are as follows Figure 1 As shown in the figure, the two-dimensional matrix is ​​converted into a point cloud depth information tensor containing point cloud intensity information, and the point cloud data generated by the lidar in the snow driving environment is collected to obtain a 3000-frame lidar point cloud set {M1, M2, M3, ..., M 3000};

[0076] Step 1.1, for the lidar point cloud set {M1,M2,M3,...,M 3000} to process any point cloud two-dimensional matrix in the set Convert it from Cartesian to cylindrical coordinates:

[0077] θ i =arctan(-y i ,x i ) (1)

[0078]

[0079]

[0080] r i =f hash (φ i ) (4)

[0081] where θ i ,φi are the azimuth and elevation angles of point i in the cylindrical coordinate system, △θ represents the scanning resolution of the mechanical laser radar in the horizontal and vertical directions, and f hash represents the lidar elevation angle resolution mapping function; x i ,y i , z i are the coordinates of the point in the Cartesian coordinate system with the lidar as the origin, i.e., the x-axis, y-axis, and z-axis. The cylindrical coordinate system is expanded along the z-axis with △θ as the interval to obtain a plane grid without filling value, r i and c i Respectively represent the horizontal and vertical correspondence of point i in the grid;

[0082] Step 1.2, r i with c i If there is a negative value, r i with c i After transformation, the row and column indices of the grid are mapped, and:

[0083] x min =-π / △θ (5)

[0084] y min =minf hash (φ i ) (6)

[0085] Among them, x min is the minimum angle value in the yaw direction, y min is the minimum angle value of the laser radar in the pitch direction;

[0086] Transforming equations (3) and (4) yields equations (7) and (8):

[0087]

[0088]

[0089] and are the row and column indices of the grid;

[0090] Step 1.3, get the grid and The value at

[0091]

[0092] Because the laser radar ray is too far away or absorbed, there is no value at a certain index on the raster map. Therefore, it does not exist at all indexes of the raster map. The value of the position is set to 99.0, indicating that the depth value at this position cannot be obtained; after assigning values ​​to all index positions of the grid map, the grid depth map V of a frame of point cloud data is obtained. V can be regarded as a one-dimensional matrix. Its coordinate direction is shown as follows Figure 2 shown.

[0093] Step 1.4, for any lidar point (x i ,y i ,z i )| i∈n , each has an echo intensity value p i Correspondingly, construct a matrix P of the same size as V, with the index in P and The value at p i ; Set the index in P that does not exist i The value of is set to 0, indicating no echo intensity; V and P are stacked to form the point cloud depth information tensor Where w and h are the sizes of T in the first two dimensions respectively; each M corresponds to a T, and the point cloud depth information tensor set {T1, T2, T3, ..., T 3000};

[0094] 2) Annotate the point cloud data,

[0095] For the lidar point cloud set {M1,M2,M3,...,M 3000} to mark, Add a column based on the data, where the element 1 represents snowflakes and 0 represents normal data, forming a matrix containing labeling information.

[0096] To S i The first three columns of data in the table are used to generate the index of the point cloud data annotation grid using equations (1) to (8), and the value at the index is set to S i The fourth column value at the corresponding index is used to get the label matrix For {M1,M2,M3,...,M 3000}, we can get the label set {Ya1, Ya2, Ya3,..., Ya 3000};

[0097] 3) T i and Ya i Combine them and get the data set {(T1,Ya1),(T2,Ya2),...,(T 3000 ,Ya 3000 )}; Generate the training data set required for the radar point cloud denoising method and divide it into a training set and a test set;

[0098] Step 2,

[0099] The training set is split into 100 subsets, each of which contains 30 Ts and Yas. Let the set of Ts in each subset be X and the set of Yas be Y. Then, for any X, perform the following processing:

[0100] 1) Assume that the convolution process is:

[0101] Mid conv =conv(In conv ) (10)

[0102] Mid bn =BN(Mid conv ) (11)

[0103] Out conv =max(0,Mid bn ) (12)

[0104] The conv in (10) represents the two-dimensional convolution kernel operation. conv Represents the input of the convolution process, Mid conv represents the intermediate output of (10), BN in (11) represents the batch normalization operation, Mid bn Indicates the intermediate output of (11), and max in (12) indicates the output of Mid bn All elements in the value greater than 0, Out conv Represents the output of the convolution process.

[0105] Take X as the input of the convolution process, perform two convolution processes, and get the output X1.

[0106] 2) Perform discrete wavelet transform on X1 to obtain output X2.

[0107] 3) Assume residual attention processing:

[0108]

[0109]

[0110]

[0111] (13)-(15), In att Represents the input of residual attention processing, Out att represents the output of residual attention, and is an intermediate variable, avgpool represents average pooling, fc represents fully connected neural network operation, softmax represents normalized exponential operation, and cba represents convolution processing. X2 is used as input for residual attention processing to obtain X3. The specific operation process is as follows Figure 4 shown.

[0112] 4) Perform discrete wavelet transform and residual attention processing on X3 to obtain X4.

[0113] 5) Perform discrete wavelet transform and two residual attention processes on X4 to obtain X5.

[0114] 6) Perform two residual attention processes and discrete inverse wavelet transform on X5 and then add it to X4 to obtain X6.

[0115] 7) Perform residual attention processing and discrete inverse wavelet transform on X6 and add it to X3 to obtain X7.

[0116] 8) Perform residual attention processing and discrete inverse wavelet transform on X7 and add it to X1 to obtain X8.

[0117] 9) Add X8 to X and use the sigmoid function for activation:

[0118]

[0119] In (16), e is a natural constant, is the output.

[0120] 10) And the corresponding Y is substituted into the loss function:

[0121]

[0122] In (17), F represents the F-norm and bs is 30.

[0123] 11) According to the calculation The gradient descent method is used to optimize the weight parameters in all convolution operations and residual attention operations in 1) to 10).

[0124] 12) For all subsets in the training set, loop 1) to 11), set a threshold v for the loss function, and when the loss function value is less than the threshold v, terminate the loop and save the weight parameters in all convolution operations and residual attention operations in 1) to 10).

[0125] Step 3: Use the point cloud data denoising method with the optimized weight parameters in step 2 to process the lidar point cloud set of the test set, compare the output with the labeled data in the test set, and calculate the average precision (AP) of the denoising method on the test set. If the AP requirement is met, save the weight parameters, otherwise repeat step 2.

[0126] Step 4,

[0127] The snowy driving environment LiDAR point cloud data denoising method that has been trained and meets the evaluation indicators of the test set is used to denoise the snowy environment. During the denoising process of the surrounding environment point cloud snow, the bs value of the input data is set to 1 to ensure that the input dimension of the denoising method matches the real information input dimension; for the output of the denoising method The value range of each element should be between 0 and 1. If the element value is greater than the threshold value of 0.5, the corresponding lidar point is considered to be a snowflake and is removed; if it is less than 0.5, it is considered to be a normal environmental obstacle and is retained.

[0128] By using the laser radar point cloud data denoising method of this embodiment, the laser radar carried by the unmanned vehicle is used to collect data and perform denoising tests in a snowy environment. The test results are as follows: Figure 5 As shown in the figure, the left half is the original LiDAR point cloud data, and the right half is the point cloud data after denoising. It can be seen that the snowflake noise in the LiDAR point cloud data after denoising is effectively removed.

[0129] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for denoising mechanical lidar point cloud data in a snowy driving environment, characterized by: The following steps are involved: Step 1: Collect the point cloud data generated by the lidar in the snowy driving environment and construct the depth information tensor of the point cloud data; annotate the point cloud data and generate a dataset for the mechanical lidar point cloud denoising method; 1) Process the point cloud data and construct the depth information tensor of the point cloud data; pre-process the point cloud data generated by the lidar in the snow driving environment, and assume that a frame of point cloud data generated by the lidar is a two-dimensional matrix in represents the real number domain space, n represents the number of points generated by one frame of laser radar point cloud data; (x i ,y i ,z i )| i∈n is the i-th row data in the matrix M, representing the i-th point in a frame of point cloud data, where x i ,y i , z i The coordinates of the point in the Cartesian coordinate system with the laser radar as the origin are x-axis, y-axis, and z-axis respectively; the two-dimensional matrix is ​​converted into a point cloud depth information tensor containing point cloud intensity information, and the point cloud data generated by the laser radar in the snow driving environment is collected to obtain a laser radar point cloud set with sufficient frames {M1,M2,M3,...,M n }; Step 1.1, for the lidar point cloud set {M1,M2,M3,...,M n } to process any point cloud two-dimensional matrix in the set Convert it from Cartesian to cylindrical coordinates: θ i =arctan(-y i ,x i ) (1) r i =f hash (f i ) (4) where θ i ,φ i are the azimuth and elevation angles of point i in the cylindrical coordinate system, Δθ represents the scanning resolution of the mechanical laser radar in the horizontal and vertical directions, and f hash represents the lidar elevation angle resolution mapping function; x i ,y i , z i are the coordinates of the point in the Cartesian coordinate system with the lidar as the origin, i.e., the x-axis, y-axis, and z-axis. The cylindrical coordinate system is expanded along the z-axis with Δθ as the interval to obtain a plane grid without filling value, r i and c i Respectively represent the horizontal and vertical correspondence of point i in the grid; Step 1.2, r i with c i If there is a negative value, r i with c i After transformation, the row and column indices of the grid are mapped, and: x min =-π / Δθ (5) y min =min f hash (f i ) (6) Among them, x min is the minimum angle value in the yaw direction, y min is the minimum angle value of the laser radar in the pitch direction; Transforming equations (3) and (4) yields equations (7) and (8): r i t and are the row and column indices of the grid; Step 1.3, get r in the grid i t and The value at All indexes of the raster map do not exist The value of is set to 99.0, indicating that the depth value at this location cannot be obtained; the grid of the obtained point cloud data is set to V, where V is a two-dimensional matrix; Step 1.4, for any lidar point (x i ,y i ,z i )| i∈n , each has an echo intensity value p i Correspondingly, construct a matrix P of the same size as V, with the index in P and The value at p i ; Set the index in P that does not exist i The value of is set to 0, indicating no echo intensity; Stack V and P to form the point cloud depth information tensor Where w and h are the sizes of T in the first two dimensions respectively; each M corresponds to a T, and the point cloud depth information tensor set {T1, T2, T3, ..., T n }; 2) Annotate the point cloud data, For the lidar point cloud set {M1,M2,M3,...,M n } to mark, Add a column based on the data, where the element 1 represents snowflakes and 0 represents normal data, forming a matrix containing labeling information. To S i The first three columns of data in the table are used to generate the index of the point cloud data annotation grid using equations (1) to (8), and the value at the index is set to S i The fourth column value at the corresponding index is used to get the label matrix For {M1,M2,M3,...,M n }, we can get the label set {Ya1, Ya2, Ya3,..., Ya n }; 3) T i and Ya i Combine them and get the data set {(T1,Ya1),(T2,Ya2),...,(T n ,Ya n )}; Generate the training data set required for the radar point cloud denoising method and divide it into a training set and a test set; Step 2: The training set obtained in step 1 is processed using convolution, discrete wavelet transform, residual attention, and discrete inverse wavelet transform. The weight parameters in convolution and residual attention are optimized through forward and backward propagation. Step 3: Use the test set obtained in step 1 to evaluate the weight parameters in the convolution processing and residual attention processing optimized in step 2 to obtain the weight parameters that meet the accuracy requirements; Step 4: Use the denoising method that meets the accuracy requirements in step 3 to denoise the snowy environment.

2. The method for denoising mechanical lidar point cloud data in a snowy driving environment as claimed in claim 1, characterized in that: The specific implementation of step 2 is: Split the training set into several subsets, where each subset contains bs Ts and Yas. In each subset, the set of Ts is X and the set of Yas is Y. Then, perform the following processing on any X: 1) Convolution processing is: Mid conv =conv(In conv ) (10) One bn =BN(One conv ) (11) Out conv =max(0,Mid bn ) (12) The conv in formula (10) represents the two-dimensional convolution kernel operation. conv Represents the input of the convolution process, Mid conv represents the intermediate output of (10); BN in formula (11) represents the batch normalization operation, Mid bn Indicates the intermediate output of (11); max in (12) indicates the output of Mid bn All elements in the value greater than 0, Out conv Represents the output of convolution processing; Take X as the input of the convolution process, perform two convolution processes, and get the output X1; 2) Perform discrete wavelet transform on X1 to obtain output X2; 3) Residual attention processing: In formulas (13)-(15), In att Represents the input of residual attention processing, Out att represents the output of residual attention, and is an intermediate variable, avgpool represents average pooling, fc represents fully connected neural network operation, softmax represents normalized exponential operation, and cba represents convolution processing; X2 is used as input for residual attention processing to obtain X3; 4) Perform discrete wavelet transform and residual attention processing on X3 to obtain X4; 5) Perform discrete wavelet transform and two residual attention processes on X4 to obtain X5; 6) Perform two residual attention processes and discrete inverse wavelet transform on X5 and add it to X4 to obtain X6; 7) Perform residual attention processing and discrete inverse wavelet transform on X6 and add it to X3 to obtain X7; 8) Perform residual attention processing and discrete inverse wavelet transform on X7 and add it to X1 to obtain X8; 9) Add X8 to X and use the sigmoid function for activation: In formula (16), e is a natural constant, is the output; 10) And the corresponding Y is substituted into the loss function: In formula (17), F represents the F-norm; 11) According to 10) Gradient descent is used to optimize the weight parameters of all convolution operations and residual attention operations in 1) to 10); 12) For all subsets in the training set, loop 1) to 11). When the loss function value is less than the preset threshold v, terminate the loop and save the weight parameters in all convolution operations and residual attention operations in 1) to 10).

3. The method for denoising mechanical lidar point cloud data in a snowy driving environment as claimed in claim 1, characterized in that: Step 3 is implemented as follows: Use the point cloud data denoising method with the optimized weight parameters in step 2 to process the lidar point cloud set of the test set, compare the output with the labeled data in the test set, and calculate the average precision (AP) of the denoising method on the test set. If the AP requirement is met, save the optimized weight parameters, otherwise repeat steps 2 and 3.

4. The method for denoising mechanical lidar point cloud data in a snowy driving environment as claimed in claim 1, characterized in that: Step 4 is implemented as follows: Denoising of snowy driving environment is performed using a well-trained LiDAR point cloud data denoising method that meets the evaluation criteria of the test set; In the denoising process of the surrounding point cloud snowflakes, the bs value of the input data is set to 1 to ensure that the input dimension of the denoising method matches the input dimension of the real information; for the output of the denoising method The value range of each element should be between 0 and 1. If the element value is greater than the threshold k, the corresponding lidar point is considered to be a snowflake and is removed; if it is less than the threshold k, it is considered to be a normal environmental obstacle and is retained.

Images