Three-dimensional point cloud vehicle detection method in highway scene based on voting voxel fusion network

Abstract

The application discloses a kind of based on voting voxel fusion network's highway scene three-dimensional point cloud vehicle detection method, comprising the following steps: S1, constructs three-dimensional point cloud vehicle detection data set KITTI-Car-Weather in highway scene, including conventional, fog, rainy day and snow day;S2, for data set characteristics and detection task requirement, construct FN-DHV-VDHS network model for three-dimensional point cloud vehicle detection in highway scene;S3, model training and parameter optimization, construct the experimental result evaluation index of multiple weather conditions, respectively based on the detection average precision, detection speed and robustness of three-dimensional point cloud vehicle detection method in highway scene of voxel network fusion depth Hough voting are evaluated.The model performance and robustness of the application are superior to other commonly used vehicle detection network, and the average precision of vehicle detection in highway scene reaches (93.86, 84.49, 83.40), and the detection frame rate reaches 20.5Hz.Relative corrosion error under multiple weather conditions is only 73.29%.

Description

Technical Field

[0001] This invention belongs to the research field of intelligent highways and intelligent perception, specifically involving a method for 3D point cloud vehicle detection in highway scenarios based on deep learning fusion networks. Background Technology

[0002] The level of intelligent perception on highways is crucial for highway management, and reliable and effective 3D vehicle detection algorithms are the foundation and guarantee for this level. 3D vehicle detection in highway scenarios is one of the most important tasks in understanding and analyzing highway environmental information. Its purpose is to accurately locate vehicles and identify their type using 3D point cloud data. Because 3D point clouds provide accurate location and geometric information of targets and are insensitive to changes in lighting, vehicle detection based on 3D point cloud target detection can better acquire 3D spatial features to complete target detection, relying on the depth information provided by point clouds. However, vehicle detection in highway scenarios faces several challenges. On the one hand, vehicles move at high speeds; on the other hand, road conditions in spring and autumn, especially in mountainous and hilly areas, are often foggy at night. These characteristics necessitate that 3D vehicle detection models deployed in real-world systems possess both high detection accuracy and sufficient inference speed. Furthermore, frequent rain and snow in highway scenarios cause a rapid decline in the performance of LiDAR. Such weather changes require 3D vehicle detection models to have a certain degree of robustness. In view of this, the present invention designs a three-dimensional point cloud vehicle detection model based on deep Hough voting voxel fusion network to solve the problem of balancing accuracy, efficiency and robustness of three-dimensional point cloud vehicle detection under multiple weather conditions in highway scenarios. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based method for detecting 3D point cloud vehicles in highway scenarios. An FN-DHV-VDHS network model for 3D point cloud vehicle detection in highway scenarios is constructed. Through model design, training, and optimization, this method can effectively detect 3D point cloud vehicles in highway scenarios, solving the problem of balancing accuracy, efficiency, and robustness in 3D point cloud vehicle detection under various weather conditions in highway scenarios. This provides a foundation for the subsequent deployment of 3D point cloud vehicle detection in all-weather highway scenarios.

[0004] The technical solution adopted in this invention is: a method for 3D point cloud vehicle detection in highway scenes based on voxel network fusion and deep Hough voting. It includes the following steps:

[0005] S1. Construct a 3D point cloud vehicle detection dataset KITTI-Car-Weather for highway scenes, including normal, foggy, rainy, and snowy weather.

[0006] S2. Based on the characteristics of the dataset and the requirements of the detection task, construct an FN-DHV-VDHS network model for 3D point cloud vehicle detection in highway scenes;

[0007] S3. Model training and parameter optimization: Evaluation metrics for experimental results under various weather conditions are constructed. The average accuracy, detection speed, and robustness of the 3D point cloud vehicle detection method based on voxel network fusion and deep Hough voting in highway scenes are evaluated.

[0008] Preferably, in step S1, the KITTI dataset is used as the original point cloud, and a weather corrosion simulation algorithm based on physical principles is used to simulate the scene containing vehicles, thereby constructing a 3D point cloud vehicle detection dataset KITTI-Car-Weather for highway scenes. The specific steps are as follows:

[0009] S11. Collect the KITTI dataset, remove the data with the scenes Campus and Person in the dataset, and save the remaining point cloud data containing vehicles as the dataset KITTI-Car. Then, use LabelCloud software to re-label the data.

[0010] S12. Divide the re-annotated KITTI-Car dataset into a 3D point cloud vehicle detection training set and a validation set, with 80% as the training set and 20% as the validation set.

[0011] S13. Input the KITTI-Car dataset into the weather corrosion simulation algorithm to simulate point cloud data for the corresponding weather conditions;

[0012] S14. The point cloud data obtained by the weather corrosion simulation algorithm is divided into "Light", "Moderate" and "Heavy" according to the degree of corrosion. The data is collected, organized and a three-dimensional point cloud vehicle detection dataset KITTI-Car-Weather in the highway scene is constructed.

[0013] In step S11, the 3D point cloud vehicle detection dataset KITTI-Car includes three scenes with a total of 6684 point cloud images, while in step S14, the 3D point cloud vehicle detection dataset KITTI-Car-Weather includes four weather conditions, each with three degrees of corrosion, and a total of 66840 point cloud images.

[0014] Preferably, in step S1, a weather erosion simulation algorithm is used to simulate point cloud data for the corresponding weather conditions. For the weather erosion simulation algorithm, a physical fog simulation method is used to create fog data, a LISA rain simulation method is used to create rain data, and a LISA snow simulation method is used to create snow data. The attenuation coefficient α and backscattering coefficient β for fog are (0, 0.005, 0.01, 0.02, 0.03, 0.06) and (0.008, 0.05, 0.2), respectively, and the rainfall rate r for rain is... rain The snowfall rate r on snowy days is (0.20, 1.56, 7.29). snow The value is (0.73, 1.563.12).

[0015] Preferably, in step S2, an FN-DHV-VDHS network model for 3D point cloud vehicle detection in highway scenes is constructed. The network is built using a feature extraction module that fuses points and voxels. The specific construction steps are as follows:

[0016] S21 Constructs a voxel-based feature extraction proposal box generation network:

[0017] The first part voxelizes the point cloud and extracts the original voxel features using a voxel feature encoding layer; the second part combines spatial sparse convolution and submanifold sparse convolution into a 3D sparse convolution layer to further extract features from the voxels, and then uses a sparse-to-dense layer to form a bird's-eye view, transforming the 3D sparse features into dense 2D features; the third part uses a region generation network containing a 2D backbone network to extract bird's-eye view features and generate candidate regions.

[0018] S22 constructs a voting point perception network based on deep Hough voting:

[0019] Using the original point cloud as input data, a novel multi-range sampling method—a segmentation-based sampling method—is employed and fused with farthest point sampling and feature sampling methods to sample seed point sets and key point sets, respectively. Bilinear interpolation is then used to fuse bird's-eye view features onto the seed points and key points. Finally, a voting point generation network generates a new set of voting points by multiplying the number of voting points generated for each seed point.

[0020] S23 constructs a vehicle detection network that fuses point and voxel features:

[0021] The RoI-gridpooling module is used to divide the proposal boxes generated by the voxel-based feature extraction proposal box generation network into a grid. The feature aggregation is performed on the point set composed of voting points and key points with the grid points as the center by setting a variable radius. Finally, the aggregated features are used to perform confidence prediction and proposal box optimization respectively.

[0022] Preferably, in step S3, an evaluation index for experimental results under multiple weather conditions is constructed to assess the average accuracy, detection speed, and robustness of the 3D point cloud vehicle detection method based on voxel network fusion deep Hough voting in highway scenes, and to analyze the effectiveness of the model improvement strategy. The specific construction steps are as follows:

[0023] S31. Construct evaluation indicators for experimental results under multiple weather conditions:

[0024] In the evaluation indicators for vehicle inspection, the average precision AP40@0.7 is selected for inspection accuracy, the frame rate (FPS) is selected for inspection speed, and the relative corrosion error (RCE) is selected for robustness.

[0025] S32. Comparative Analysis of the Effects of Improved Strategies:

[0026] The effectiveness of the voting algorithm fusion detection and the effectiveness of the improved voting algorithm were compared and analyzed under various weather conditions.

[0027] As a preferred method, the evaluation indicators for vehicle inspection include average precision (AP40@0.7), inspection speed (frames per second, FPS), and robustness (relative corrosion error, RCE). The calculation formulas for each evaluation indicator are as follows:

[0028] Calculate the average accuracy of detection, AP40@0.7 (IoU=0.7):

[0029]

[0030]

[0031] Where IoU represents the intersection-union ratio, B pred B represents the predicted bounding box. GT This represents the true bounding box. AP (Average Precision) represents the average precision, where P represents the probability that a predicted positive sample is actually a positive sample, and R represents the probability that a sample that is actually a positive sample is predicted to be a positive sample.

[0032] Calculate the detection speed frame rate (FPS):

[0033]

[0034] FPS (Frames per second) represents the number of frames per second. per_pointcloud This represents the time required to detect one frame of point cloud;

[0035] Calculate the robust relative corrosion error (RCE):

[0036]

[0037]

[0038] Among them, AP cor AP represents the average accuracy of detection under corrosive conditions. w,s This indicates the average accuracy of detection for the severity of corrosion under weather conditions (w). RCE (Relative Corruption Error) represents the relative corrosion error. clean This represents the average accuracy of detection under corrosive conditions.

[0039] The beneficial effects of this invention are as follows: The FN-DHV-VDHS 3D point cloud vehicle detection model constructed by this invention for highway scenes can effectively detect 3D point cloud vehicles in highway scenes. The average detection accuracy of the model is (93.86, 84.49, 83.40), which is higher than other common 3D point cloud vehicle detection models. The model's detection frame rate is 20.5Hz, which is at a leading level among common 3D point cloud vehicle detection models and exceeds the baseline of real-time processing. The relative erosion error of the model is 73.29%, which is lower than other common 3D point cloud vehicle detection models. The model of this invention effectively solves the problem of balancing accuracy, efficiency, and robustness of 3D vehicle detection under various weather conditions in highway scenes, providing a foundation for the subsequent deployment of 3D point cloud vehicle detection in all-weather highway scenes. Attached Figure Description

[0040] Figure 1 A partial visualization of the KITTI-Car-Weather 3D point cloud vehicle detection dataset in a highway scene.

[0041] In the figure, (a) is a point cloud under normal conditions, (b) is a point cloud under foggy conditions, (c) is a point cloud under rainy conditions, and (d) is a point cloud under snowy conditions. The blue box represents the real frame.

[0042] Figure 2 The structure diagram of the FN-DHV-VDHS model;

[0043] Figure 3 The image shows the vehicle detection performance of the FN-DHV-VDHS model on the validation set. (a) represents the vehicle detection performance under normal weather conditions, (b) represents the vehicle detection performance under foggy conditions, (c) represents the vehicle detection performance under rainy conditions, and (d) represents the vehicle detection performance under snowy conditions. The green boxes represent prediction boxes. Detailed Implementation

[0044] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0045] like Figure 1-3 As shown, a method for 3D point cloud vehicle detection in a highway scene based on voxel network fusion and deep Hough voting includes the following steps:

[0046] Step 1: Collect the KITTI dataset. Remove data from the Campus and Person scenes, saving the remaining point cloud data containing vehicles as the KITTI-Car dataset, consisting of 6684 point cloud images. Re-annotate the KITTI-Car dataset using LabelCloud software. Divide the re-annotated KITTI-Car dataset into a 3D point cloud vehicle detection training set and a validation set, with 80% for training and 20% for validation. Input the KITTI-Car dataset into a weather erosion simulation algorithm to simulate point cloud data under corresponding weather conditions. Use a physical fog simulation method to create fog data, a LISA rain simulation method to create rain data, and a LISA snow simulation method to create snow data. The attenuation coefficient α and backscattering coefficient β for fog are (0, 0.005, 0.01, 0.02, 0.03, 0.06) and (0.008, 0.05, 0.2), respectively, and the rainfall rate r for rainy days is... rain The snowfall rate r on snowy days is (0.20, 1.56, 7.29). snow The values ​​are (0.73, 1.563.12). Point cloud data obtained through weather erosion simulation algorithm were divided into "Light", "Moderate" and "Heavy" according to the degree of erosion. The data were collected, organized and constructed into a 3D point cloud vehicle detection dataset KITTI-Car-Weather for highway scenes, consisting of 66,840 point cloud images.

[0047] Step 2: Construct the FN-DHV-VDHS network model for 3D point cloud vehicle detection in highway scenes. The network is built using a feature extraction module that fuses point and voxel features. This includes a voxel-based feature extraction proposal box generation network, a deep Hough voting-based voting point perception network, and a vehicle detection network that fuses point and voxel features. The structure is as follows: Figure 2 As shown, the specific construction steps are as follows:

[0048] (1) First, the space containing the original point cloud is divided into several three-dimensional grids along the XYZ dimensions, retaining the non-empty voxels containing points, and keeping the number of points in each non-empty voxel the same. The disordered point cloud is transformed into an ordered representation, and features are extracted from the non-empty voxels using a voxel feature encoder. The feature of each non-empty voxel is the mean of the point-by-point features within it. The range of the point cloud is set to [0, -40, -3, 70.4, 40, 1], where the X-axis is 0 to 70.4 m, the Y-axis is -40 to 40 m, and the Z-axis is -3 to 1 m. The unit voxel size is set to [0.05, 0.05, 0.1], in meters, and the maximum value of a point in a unit voxel is set to 5.

[0049] (2) The obtained voxels are input into the 3D backbone network. Using a feature extraction block combining submanifold sparse convolution and 3D sparse convolution, features are extracted from the voxelized point clouds at four different downsampling scales, transforming them into 3D feature volumes to complete 3D scene encoding. The Z-axis height dimension of the 8x 4D feature map is merged with the feature dimension and projected into a pseudo-top view from the BEV perspective. The 3D backbone network uses submanifold sparse convolution and 3D sparse convolution to generate feature dimensions of 16, 32, 64, and 64 at downsampling scales of 1x, 2x, 4x, and 8x, respectively. Then, the 3D features are stacked along the Z-axis to form a 2D BEV feature map, with the feature dimension becoming 128 (64*2).

[0050] (3) In the region proposal network (RPN) part, a pseudo-top-down view feature is first extracted using a backbone network composed of 2D convolutions. Then, multi-scale features are concatenated using deconvolution layers. In the 2D backbone network, two 6-layer 2D convolutional blocks generate (128, 256) dimensional features at 1x and 1 / 2x resolution, respectively. Then, upsampling generates two 256-dimensional features, which are concatenated to form a 512-dimensional feature map. Finally, a high-quality 3D proposal box is generated using a detector head. The RPN loss function uses region proposal loss:

[0051] L1 = L CLS +γ reg L reg +γ dir L dir

[0052] Among them, the classification loss function L CLS Using Focal Loss, the regression loss function L reg Using smooth L1, the loss of the direction angle is calculated using the function L. dir Use Cross Entropy. γ reg and γ dirThe loss weights were set to 2.0 and 0.2 respectively.

[0053] (4) For the input point cloud P0 (number of points N), the farthest point sampling is used first to downsample to obtain a point set of 4096 points. Point features are then learned at a scale of (0.2, 0.4, 0.8) to obtain point set P1. A second downsampling is performed on point set P1 containing feature information using a feature sampling method, resulting in a point set P2 with 1024 points. A third downsampling is performed using both feature sampling and segmentation sampling methods. The segmentation sampling method utilizes 3D segmentation labels obtained free of charge from ground-truth 3D bounding boxes, calculates the probability that points containing high-dimensional features belong to foreground points using MLP1, and then performs binary classification using sigmoid. The feature sampling method outputs a point set P3 containing 512 points, and the segmentation sampling method outputs a point set P4 containing 512 points. P4 is referred to as the seeds in this method. The segmentation sampling method uses Focal Loss as the loss function.

[0054] L seg (p i )=-α(1-p t ) γ log(p t )

[0055] (5) Regarding the obtained Bilinear interpolation is used to project seed points onto the BEV (Body Object Model), and the obtained high-level BEV information is concatenated into the feature information of the original point cloud. Using seeds containing both parts of information as input, MLP2 is used to regress the pointing vector of each seed point from its corresponding object center. and eigenvectors This generates votes(v) i =[x i +Δx i ;f i +Δf i The resulting votes maintain the same feature dimensions as the seeds. The loss function for votes uses smooth L1:

[0056]

[0057] (6) The obtained votes can be used to optimize the high-quality proposal boxes obtained in the first stage. The model selection utilizes the RoI-grid pooling module to implement vote-aware RoI pooling. For a given proposal box, it is divided into a 6×6×6 grid, and 6×6×6 grid points are selected. With the grid point as the center, a variable radius of (0.8, 1.6) is set. PointNet is used to aggregate the features of nearby votes, and multi-scale feature fusion is used to obtain the point features of the grid points. Then, two layers of MLP3 are used to obtain feature vectors for confidence prediction and proposal box optimization, respectively. The candidate box optimization loss function includes confidence prediction loss and proposal box optimization loss, using Cross Entropy and smooth L1 respectively:

[0058] L refine =L confidence +L box

[0059] Step 3: Construct evaluation metrics for experimental results under multiple weather conditions. Evaluate the average accuracy, detection speed, and robustness of the 3D point cloud vehicle detection method based on voxel network fusion and deep Hough voting in highway scenes. For vehicle detection evaluation metrics, the average accuracy (AP40@0.7) is used for detection accuracy, the frame rate (FPS) is used for detection speed, and the relative erosion error (RCE) is used for robustness. The calculation formulas for each evaluation metric are as follows:

[0060] Calculate the average accuracy of detection, AP40@0.7 (IoU=0.7):

[0061]

[0062]

[0063] Calculate the detection speed frame rate (FPS):

[0064]

[0065] Calculate the robust relative corrosion error (RCE):

[0066]

[0067]

[0068] The FN-DHV-VDHS 3D point cloud vehicle detection model constructed in this invention can effectively detect 3D point cloud vehicles in highway scenes. The model's average detection accuracy is (93.86, 84.49, 83.40), all higher than other common 3D point cloud vehicle detection models. The model's detection frame rate is 20.5Hz, which is among the leading levels of common 3D point cloud vehicle detection models and exceeds the baseline of real-time processing. The model's relative erosion error is 73.29%, lower than other common 3D point cloud vehicle detection models.

[0069] The average accuracy of 3D point cloud vehicle detection in highway scenarios is shown in the table below:

[0070] Table 1 Evaluation Table of 3D Point Cloud Vehicle Detection in Highway Scenes

[0071]

[0072] Table 1 shows that FN-DHV-VDHS achieves a good balance between detection performance and efficiency. Regarding model performance, compared to models using only one encoding method, both PV-RCNN-3D-VDHS (using a fusion of original point cloud and voxel encoding) and the proposed FN-DHV-VDHS model demonstrate good performance. At different difficulty levels, both PV-RCNN-3D-VDHS and FN-DHV-VDHS achieved AP scores above 80.00. Compared to PV-RCNN-3D-VDHS, FN-DHV-VDHS achieved performance improvements of 2.51%, 1.77%, and 2.94% at easy, medium, and hard difficulty levels, respectively. In terms of real-time model processing speed, apart from the single-stage models 3DSSD-3D-VDHS, SECOND-3D-VDHS and the two-stage model FN-DHV-VDHS of this method, which reached over 20 FPS, the remaining two-stage methods failed to reach the baseline of 15 FPS for real-time processing.

[0073] The corrosion robustness and relative corrosion error of 3D point cloud vehicle detection in highway scenes are shown in the table below:

[0074] Table 2. Corrosion Robustness and Relative Corrosion Error of 3D Point Cloud Vehicle Detection in Highway Scenarios

[0075]

[0076] Table 2 shows that under foggy conditions, except for 3DSSD-3D-VDHS, the AP of all other models remained above 60, while the performance of PV-RCNN-3D-VDHS and the proposed model showed little change. Only the proposed model maintained an AP above 80. Rainy and snowy conditions significantly impacted model performance. The APs of Point RCNN-3D-VDHS, PV-RCNN-3D-VDHS, and FN-DHV-VDHS all remained above 50, with FN-DHV-VDHS leading the other two models with APs of 54.26 and 56.52. Considering the overall corrosion robustness of the models, calculating their relative corrosion errors reveals that only Point RCNN-3D-VDHS, PV-RCNN-3D-VDHS, and FN-DHV-VDHS had relative corrosion errors below 80%, while the proposed FN-DHV-VDHS achieved the lowest relative corrosion error of 73.29%. This indicates that the proposed model exhibits superior robustness under adverse weather conditions compared to the other five models.

[0077] The technical means disclosed in the invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the invention, and these improvements and modifications are also considered within the scope of protection of the invention.

Claims

1. A method for 3D point cloud vehicle detection in highway scenes based on voxel network fusion and deep Hough voting, characterized in that, Includes the following steps: S1. Construct a 3D point cloud vehicle detection dataset KITTI-Car-Weather for highway scenes, including normal, foggy, rainy, and snowy weather. S2. Based on the characteristics of the dataset and the requirements of the detection task, construct an FN-DHV-VDHS network model for 3D point cloud vehicle detection in highway scenes; S3. Model training and parameter optimization: Construct evaluation indicators for experimental results under multiple weather conditions, and evaluate the average accuracy, detection speed and robustness of the 3D point cloud vehicle detection method based on voxel network fusion deep Hough voting in highway scenes. In step S2, an FN-DHV-VDHS network model for 3D point cloud vehicle detection in highway scenarios is constructed. The network is built using a feature extraction module that fuses points and voxels. The specific construction steps are as follows: S21. Construct a voxel-based feature extraction proposal box generation network: The first part voxelizes the point cloud and extracts the original voxel features using a voxel feature encoding layer; the second part combines spatial sparse convolution and submanifold sparse convolution into a 3D sparse convolution layer to further extract features from the voxels, and then uses a sparse-to-dense layer to form a bird's-eye view, transforming the 3D sparse features into dense 2D features; the third part uses a region generation network containing a 2D backbone network to extract bird's-eye view features and generate candidate regions. S22. Construct a voting point perception network based on deep Hough voting: Using the original point cloud as input data, a novel multi-range sampling method is employed for point sampling, including segmentation-based sampling, farthest point sampling, and feature sampling. First, farthest point sampling is used for the first downsampling to generate two sets of point clouds. Then, feature sampling is used for the second downsampling of these two sets. Next, segmentation-based sampling is performed on one set of point clouds, while feature sampling is performed on the other for the third downsampling. Finally, seed point sets and key point sets are sampled separately. Bilinear interpolation is used to fuse the bird's-eye view features into the seed points and key points. Then, a voting point generation network is used to generate a new set of voting points, with a ratio of one voting point generated for each seed point. S23. Construct a vehicle detection network that fuses point and voxel features: The RoI-grid pooling module is used to divide the proposal boxes generated by the voxel-based feature extraction proposal box generation network into a grid. The feature aggregation is performed on the point set composed of voting points and key points with the grid points as the center by setting a variable radius. Finally, the aggregated features are used to perform confidence prediction and proposal box optimization respectively.

2. The method according to claim 1, characterized in that, In step S1, the KITTI dataset is used as the original point cloud, and a weather corrosion simulation algorithm based on physical principles is used to simulate the scene containing vehicles, thereby constructing a three-dimensional point cloud vehicle detection dataset KITTI-Car-Weather in the highway scene.

3. The method according to claim 2, characterized in that, The specific method for constructing the 3D point cloud vehicle detection dataset KITTI-Car-Weather in a highway scene is as follows: S11. Collect the KITTI dataset, remove the data with the scenes Campus and Person in the dataset, and save the remaining point cloud data containing vehicles as the dataset KITTI-Car. Then, use LabelCloud software to re-label the data. S12. Divide the re-annotated KITTI-Car dataset into a 3D point cloud vehicle detection training set and a validation set, with 80% as the training set and 20% as the validation set. S13. Input the KITTI-Car dataset into the weather corrosion simulation algorithm to simulate point cloud data for the corresponding weather conditions; S14. The point cloud data obtained by the weather corrosion simulation algorithm is divided into "Light", "Moderate" and "Heavy" according to the degree of corrosion. The data is collected, organized and a three-dimensional point cloud vehicle detection dataset KITTI-Car-Weather in the highway scene is constructed.

4. The method according to claim 3, characterized in that, For the weather corrosion simulation algorithm, fog data is created using the physical fog simulation method, rain data is created using the LISA rain simulation method, and snow data is created using the LISA snow simulation method.

5. The method according to claim 4, characterized in that, Attenuation coefficient in foggy weather and backscattering coefficient They are respectively and Rainfall rate on rainy days for Snowfall rate on snowy days for .

6. The method according to claim 1, characterized in that, The S3 section constructs evaluation metrics for experimental results under multiple weather conditions to assess the average accuracy, detection speed, and robustness of the 3D point cloud vehicle detection method based on voxel network fusion and deep Hough voting in highway scenes. The effectiveness of model improvement strategies is also analyzed. The specific construction steps are as follows: S31. Construct evaluation indicators for experimental results under multiple weather conditions: In the evaluation indicators for vehicle inspection, the average precision AP40@0.7 is selected for inspection accuracy, the frame rate FPS is selected for inspection speed, and the relative corrosion error RCE is selected for robustness. S32. Comparative Analysis of the Effects of Improved Strategies: The effectiveness of the voting algorithm fusion detection and the effectiveness of the improved voting algorithm were compared and analyzed under various weather conditions.

7. The method according to claim 6, characterized in that: The specific content of the evaluation index for the experimental results of 3D point cloud vehicle detection in a highway scene under multiple weather conditions in S3 is as follows: Calculate the average accuracy of detection, AP40@0.7 (IoU=0.7): in, Indicates intersection, union, and ratio. Represents the predicted bounding box. Represents the actual bounding box; Indicates average accuracy. This represents the probability that a sample predicted to be positive is actually a positive sample. This represents the probability of predicting a positive sample when the actual sample is positive. Calculate the detection speed frame rate (FPS): in, (Frames per second) indicates the number of frames per second. This represents the time required to detect one frame of point cloud; Calculate the robust relative corrosion error (RCE): in, This indicates the average accuracy of detection under corrosive conditions. Indicates in In this weather environment Average accuracy of severity detection; Indicates relative corrosion error. This represents the average accuracy of detection under corrosive conditions.

Images