Voxel-oriented end-to-end 3D target detection method

Abstract

The invention discloses a voxel-oriented end-to-end 3D target detection method. The voxel-oriented end-to-end 3D target detection method comprises the following steps of A1, point cloud data input and data enhancement; a2, backbone network feature extraction: after the point cloud data is voxelized, the point cloud data is processed by two IDW-MSA modules, and rich feature information is extracted; and A3, detecting and classifying. According to the method, the IDW module is introduced, the large kernel convolution is used, the receptive field of the model is greatly increased, the MSA module is introduced, the attention mechanism is perceived by using neurons, additional parameters do not need to be added, the calculation efficiency is remarkably improved, and meanwhile high detection precision is kept; compared with the prior art, the method provided by the invention has the advantages that the target detection accuracy and speed under a complex background are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and artificial intelligence technology, and in particular to a voxel-oriented end-to-end 3D object detection method. Background Art

[0002] Three-dimensional object detection is an important research direction in the field of computer vision and is widely used in scenarios such as autonomous driving, robot navigation, and augmented reality. Compared with two-dimensional detection, three-dimensional object detection can more accurately describe the physical position and posture of the target in the spatial dimension, thereby providing a more comprehensive environmental perception capability. Thanks to the development of laser radar (LiDAR), depth cameras, and multi-sensor fusion technology, three-dimensional point cloud data provides researchers with rich spatial information and provides data support for accurate positioning and identification of targets.

[0003] Existing 3D object detection methods can be roughly divided into point cloud-based, voxel-based and multi-view projection-based methods. Point cloud-based methods directly process sparse point cloud data and can better preserve 3D geometric information, but the computational cost is high. Voxel-based methods simplify data representation and improve computational efficiency by dividing point clouds into voxel grids, but are prone to introduce quantization errors when processing large-scale sparse point clouds. In addition, multi-view projection methods convert 3D data into 2D views for processing. Although the features are retained to a certain extent, spatial structure information is easily lost.

[0004] voxelnext: The computational complexity is high. Since the voxelization process involves a large amount of data conversion, the computational complexity is relatively high, resulting in a slow processing speed. The IDW module of the present invention improves the feature extraction capability of sparse data by optimizing the convolution operation and reduces the computational complexity, making the model more efficient in real-time applications. PointNet++ improves the ability to capture local features through multi-level feature extraction, but leads to a significant increase in computational complexity and memory requirements. Therefore, how to find a balance between accuracy and computational complexity has always been a challenge facing three-dimensional object detection. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a voxel-oriented end-to-end 3D object detection method in view of the deficiencies in the prior art.

[0006] The technical solution of the present invention is as follows:

[0007] A voxel-oriented end-to-end 3D object detection method comprises the following steps:

[0008] A1 point cloud data input and data enhancement;

[0009] A2 backbone network feature extraction: After the point cloud data is voxelized, it is processed by two IDW-MSA modules to extract rich feature information; the introduction of the IDW module and the use of large kernel convolution greatly increase the receptive field of the model, and the introduction of the MSA module and the use of the neuron perception attention mechanism, without adding additional parameters, also significantly improves the computational efficiency while maintaining a high detection accuracy;

[0010] A3 highly compresses the voxels after feature extraction, projects them into a 2D bird's-eye view, directly places the sparse features on the BEV plane, and sums the features at the same position;

[0011] A4 Detection and classification: The detection head is used to predict the category and location of the target.

[0012] The method, in step A1, reads the KITTI data set, divides it into a training set and a validation set, and then performs data enhancement operations of random rotation, random scaling, random flipping, and random translation on the data set.

[0013] In the method, in step A2, the processed KITTI point cloud data is dynamically encoded with voxel features; this process divides the sparse point cloud data into three-dimensional voxel grids of different sizes, thereby simplifying the data representation; after voxelization, the result is a structured voxel grid, each voxel contains point cloud information within a certain range; by constructing a hybrid backbone feature extraction architecture, the voxel features are fully extracted for downstream tasks; first, the voxels in the sparse convolution tensor format are cascaded through two hybrid modules, the hybrid module includes a large kernel convolution module (IDW) and an attention mechanism neuron perception module (MSA) connected in sequence, and after being processed by the large kernel convolution module (IDW), a sparse convolution tensor is obtained, and then an attention mechanism neuron perception module (MSA) is connected; through the cascade of these two hybrid modules, the voxel features can be fully extracted to support subsequent operations.

[0014] In the method described, in step A2, the IDW module decomposes the large convolution kernel into multiple small convolution kernels, and performs feature transformation through a 11×11×11 3D convolution operation. In this process, the feature dimension is not changed. Therefore, IDW divides the features of the sparse convolution tensor into four different parts along the channel dimension, denoted as {x hw ,x w ,x h ,x id}, each group uses a different convolution kernel for training, and finally merges them together; it is expressed as: x'=Concat(x hw ,x w ,x h ,x id); it not only obtains a larger receptive field, but also improves the convolution efficiency, and also improves the anti-interference ability and robustness of the 3D target detector.

[0015] In the method described, in step A2, in MSA, the attention mechanism is used to infer the 3D attention weight for each neuron in the feature map. MSA calculates the mean and variance fusion of the input data and outputs it using the sigmoid activation function. The whole process does not require adding additional parameters to the original network.

[0016] In the method described, in MSA, the input feature map x is marked as (b, c, d, h, w), where b represents the batch size, c represents the number of channels, d represents the depth, h represents the height, and w represents the width; the mean and variance of the input are calculated respectively, and then an adaptive adjustment factor e is proposed t The calculation formula is as follows:

[0017] e t =(x-mean)^2 / (4*(var+lamda))+0.5

[0018] Among them, mean is the average value of the feature map in the spatial dimension (height and width); var is the variance of the feature map in the spatial dimension, which is calculated using the mean; lamda is a variable parameter to prevent numerical instability and ensure the stability of the calculation process; finally, the output obtained by multiplying the original input x by the sigmoid activation function means that the model has made an adaptive adjustment to the features, thereby enhancing the sensitivity of the model to the features.

[0019] In the method, in step A3, the detection head part will be subjected to sparse maximum pooling and then fully connected operations to predict the location and category of the target; the sparse maximum pooling operation can effectively reduce the amount of calculation while retaining key feature information, and the fully connected layer is responsible for converting this information into the final prediction result; in this process, the detection head part not only needs to accurately locate the target, but also accurately identify the category of the target, which is crucial to ensure the safety and reliability of the autonomous driving system. In this way, the system can respond quickly and make decisions, thereby ensuring the safe driving of the vehicle in a complex traffic environment.

[0020] The method, in step A1, the random rotation: the point cloud Vm in the marked m-th target frame is selected to rotate along the Z axis, the rotation angle is set to θ, and the original coordinates of the point are (x m ,y m ,z m ), the rotated coordinates (x', y', z') are calculated using the following formula:

[0021]

[0022] After the rotation transformation, the new point cloud set is recorded as V1.

[0023] In the method, in step A1, the random scaling is as follows: the point cloud Vm in the marked mth target frame is scaled, and the scaling factors are set to 0.5 and 1.5 respectively, to simulate the situation of vehicles of different sizes on the road. After the scaling transformation, the new point cloud data is recorded as V2.

[0024] The method described, in the step A1, the random translation: randomly generate an offset offset(x,y) for the marked point cloud Vm in the mth target frame, set the offset range x, y∈[-d,d], d is the maximum absolute value of the offset, so that the point cloud in the target frame plus the randomly generated displacement value offset(x,y), the point cloud after adding offset(x,y) is still within the river range, simulating the target point cloud position of vehicles in different positions; after the translation transformation, the new point cloud data is obtained and recorded as V3.

[0025] The method described, in the step A1, the random cropping: randomly downsample the point cloud Vm in the marked m-th target box, determine the number of target points to be N, use a uniformly distributed random number generator to generate n different random indexes from the range [0, N-1]; select corresponding data points from the original data set according to the generated random indexes, and the number of target points in the m-th target box changes from N to n; after the cropping transformation, the new point cloud data is obtained and recorded as V4.

[0026] The voxel-oriented end-to-end 3D object detection method proposed in the present invention aims to solve the shortcomings of existing high-performance detection methods such as complex architecture and large number of parameters. Current three-dimensional object detection methods are mainly divided into point cloud-based, voxel-based and multi-view projection-based methods. The point cloud-based method can effectively retain three-dimensional geometric information, but the computational cost is high and the processing efficiency is low. The voxel-based method improves the computational efficiency by dividing the point cloud into voxel grids, but it is easy to introduce quantization errors when processing large-scale sparse point clouds, affecting the accuracy. Although the multi-view projection method can retain certain features, it is easy to lose important spatial structure information in the process of converting to two-dimensional views.

[0027] Feature extraction enhancement: The IDW module is designed to enhance the model's ability to extract 3D data features. By decomposing a large convolution kernel into multiple small convolution kernels, the module not only obtains a larger receptive field but also improves the computational efficiency of convolution. In this way, IDW can effectively capture global information, reduce information loss, and thus improve detection accuracy. In addition, due to its structural design, IDW has excellent anti-interference ability and robustness, and can accurately identify targets in complex environments.

[0028] Lightweight and efficient: The MSA module is a new attention mechanism proposed for convolutional neural networks. Compared with traditional channel attention and spatial attention modules, MSA does not need to add additional parameters to the original network, greatly reducing the complexity of the model. The module infers the importance of each neuron by optimizing an energy function and uses a fast closed-form solution to calculate the 3D attention weights of the feature map. In this way, MSA can flexibly and efficiently improve the expressive power of convolutional neural networks and significantly enhance the model's sensitivity to key features.

[0029] Combining IDW and MSA modules, the 3D feature extraction network of the present invention starts with dynamic voxel feature encoding of point cloud data, compresses it into a 2D bird's-eye view after sufficient feature extraction, and finally uses it for 3D object detection. This process not only improves the efficiency of feature extraction, but also ensures stability and accuracy in different environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a structural diagram of the voxel-oriented end-to-end 3D object detection method of the present invention;

[0031] Figure 2 It is the structure diagram of the MSA attention mechanism of the present invention;

[0032] Figure 3 Detection results of IDW-MSA mixed backbone; (a) original image; (b) detection result of Complex-YOLO; (c) detection result of Ours. DETAILED DESCRIPTION

[0033] The present invention is described in detail below in conjunction with specific embodiments.

[0034] The voxel-oriented end-to-end 3D object detection method of the present invention includes a data enhancement step, a backbone network feature extraction step, and a detection and classification step. The structure diagram of the 3D object detection method is as follows: Figure 1 As shown in the figure: Data enhancement enhances data diversity through data transformation, the backbone network is responsible for extracting feature information of the input image, and the detection head is responsible for detecting vehicles and pedestrians. The specific detection process of the voxel-oriented end-to-end 3D object detection method is as follows:

[0035] A1 Data enhancement: Read the KITTI dataset, divide it into a training set and a validation set, and then perform data enhancement operations such as random rotation, random scaling, random flipping, and random translation on the dataset.

[0036] The point cloud data is rotated along the Z axis by a certain angle, scaled along the XYZ axis by a certain proportion, translated along the XYZ axis by a certain amount, and randomly cropped to obtain data that is several times more than the data set, thus expanding the diversity of the data. Point cloud data enhancement specifically includes:

[0037] Random rotation: Select the point cloud Vm in the marked mth target frame and rotate it along the Z axis. Set the rotation angle to θ and the original coordinates of the point to (x m ,y m ,z m ), the rotated coordinates (x', y', z') are calculated using the following formula:

[0038]

[0039] After the rotation transformation, the new point cloud set is recorded as V1.

[0040] Random scaling: The point cloud Vm in the marked mth target box is scaled, and the scaling factors are set to 0.5 and 1.5 respectively to simulate the situation of vehicles of different sizes on the road. After the scaling transformation, the new point cloud data is recorded as V2.

[0041] Random translation: randomly generate an offset offset(x,y) for the marked point cloud Vm in the mth target frame, set the offset range x,y∈[-d,d], d is the maximum absolute value of the offset, so that the point cloud in the target frame plus the randomly generated displacement value offset(x,y) is still within the river range after adding offset(x,y), simulating the target point cloud position of vehicles at different positions. After the translation transformation, the new point cloud data is recorded as V3.

[0042] Random cropping: Randomly downsample the point cloud Vm in the marked mth target box, determine the number of target points to be N, and use a uniformly distributed random number generator to generate n different random indexes from the range [0, N-1]. Select the corresponding data points from the original data set according to the generated random index, and the number of target points in the mth target box changes from N to n. After the cropping transformation, the new point cloud data is recorded as V4.

[0043] The point clouds obtained in the above four steps and the original annotated target point cloud together constitute a new training set Strain, that is, Strain = V1+V2+V3+V4.

[0044] A2 backbone network: Dynamic voxel feature encoding of processed KITTI point cloud data. This process divides the sparse point cloud data into 3D grids (voxels) of different sizes to simplify data representation. After voxelization, the result is a structured voxel grid, where each voxel contains point cloud information within a certain range.

[0045] A 3D feature extraction network is constructed, and the voxel data format obtained after voxelization is a sparse convolution tensor format. By constructing a hybrid backbone feature extraction architecture, the voxel features are fully extracted for downstream tasks. The present invention cascades two hybrid modules, each of which includes a large kernel convolution module (IDW) and a neuron perception module (MSA) connected in sequence. After the convolution module (IDW) is processed, a sparse convolution tensor is obtained, and then a convolution operation (MSA) is connected. Through the cascade of two hybrid modules, voxel features can be fully extracted to support subsequent operations. The large kernel convolution module (IDW) and the neuron perception module (MSA) effectively combine the advantages of 3D convolution and attention mechanism. In IDW, the large convolution kernel is decomposed into multiple small convolution kernels, which not only obtains a larger receptive field, but also improves the convolution efficiency, and also improves the anti-interference ability and robustness of the 3D target detector. In MSA, neuron-aware attention is used to infer 3D attention weights for each neuron in the feature map without adding parameters to the original network, making it more efficient and natural in capturing key information in vision tasks.

[0046] After feature extraction, the voxels are highly compressed and projected into a 2D bird's-eye view. The sparse features are directly placed on the BEV plane and the features are summed at the same position.

[0047] The present invention designs an IDW module mainly to obtain a larger receptive field, thereby reducing information loss and improving detection accuracy. Compared with other voxel-based methods, the method of the present invention not only has advantages in extracting global information, but also has more accurate activation of key parts. First, for the voxels in the sparse convolution tensor format obtained after voxelization, the present invention performs feature transformation through 11×11×11 3D convolution operations. In this process, the present invention does not change its feature dimension. Therefore, the present invention divides the features of the sparse convolution tensor into four different parts along the channel dimension, denoted as {x hw ,x w ,x h ,x id}. Each group is trained with different convolution kernels and then merged together. It can be expressed as: x' = Concat(x hw ,x w ,x h ,x id ).

[0048] In MSA, we propose an attention mechanism module for convolutional neural networks. Compared with the existing channel attention module and spatial attention module, our module does not need to add parameters to the original network, but infers the 3D attention weights of the feature map in one layer. Specifically, we propose to optimize an energy function to discover the importance of each neuron. We further derive a fast closed-form solution to the energy function and show that the solution can be achieved in very simple steps. Another advantage of this module is that most operators are selected according to the solution of the defined energy function, avoiding too much effort on structural adjustment. Through quantitative evaluation on various visual tasks, we prove the flexibility and effectiveness of this module, which improves the expressiveness of many convolutional neural networks.

[0049] In MSA, the input feature map x is labeled (b, c, d, h, w), where b represents the batch size, c represents the number of channels, d represents the depth, h represents the height, and w represents the width. The mean and variance of the input are calculated respectively, and then the adaptive adjustment factor e is proposed t The calculation formula is as follows:

[0050] e t =(x-mean)^2 / (4*(var+lamda))+0.5

[0051] Where mean is the average value of the feature map in the spatial dimensions (height and width). var is the variance of the feature map in the spatial dimensions, which is calculated using the mean. lamda is a variable parameter to prevent numerical instability and ensure the stability of the calculation process.

[0052] Finally, the output obtained by applying the sigmoid activation function to the original input x means that the model has made an adaptive adjustment to the features, thereby enhancing the model's sensitivity to the features.

[0053] A3 detection head: In the detection head part, sparse maximum pooling and then fully connected operations are performed to predict the location and category of the target. The sparse maximum pooling operation can effectively reduce the amount of calculation while retaining key feature information, and the fully connected layer is responsible for converting this information into the final prediction result. In this process, the detection head part not only needs to accurately locate the target, but also accurately identify the category of the target, which is crucial to ensuring the safety and reliability of the autonomous driving system. In this way, the system can respond and make decisions quickly, thereby ensuring the safe driving of the vehicle in a complex traffic environment. The 3D target detection box is drawn in the point cloud scene based on the obtained 3D target position and category parameters.

[0054] A4 outputs results: Feeds the detection results back to the driving system in real time, accurately providing the traffic conditions around the vehicle.

[0055] 1 Experimental platform and parameters

[0056] In order to make the experimental results more uniform, the models all use the same hyperparameters, use the Adam optimizer, combine the OneCycle learning rate scheduling strategy, the initial learning rate is 0.001, the batch-size is set to 8, and the training rounds are 200. The experimental environment of the present invention is shown in Table 1.

[0057] Table 1 Experimental environment

[0058]

[0059]

[0060] 2 Datasets

[0061] The KITTI dataset was jointly released by Karlsruhe Institute of Technology (KIT) and Toyota Technological Institute at Chicago (TTI-C) in 2012. It has become one of the most widely used datasets for evaluating computer vision algorithms in autonomous driving scenarios. The dataset is mainly used to evaluate the performance of various vehicle-mounted computer vision tasks, including stereo, optical flow, visual odometry, 3D object detection, and 3D tracking.

[0062] The KITTI dataset is collected in real road environments, covering a variety of scenes such as urban areas, rural areas, and highways. Each image contains up to 15 cars and 30 pedestrians, with varying degrees of occlusion and truncation. For the 3D object detection task, KITTI provides 14,999 images and corresponding point cloud data, of which 7,481 groups are used for training and 7,518 groups are used for testing. The dataset has detailed annotations for three types of targets: cars, pedestrians, and bicycles, and contains a total of 80,256 labeled objects.

[0063] 3 Evaluation Metrics

[0064] The experiment tests the car at three levels: easy, medium, and difficult. All test results are calculated using the 40 recall positions on the KITTI official test server to calculate the mean average precision (PAP), and the RIoU of the car is set to 0.7. The specific calculation process of AP is as follows:

[0065]

[0066] Where: S1 = {1 / 40, 2 / 40, ..., 1}. P inter (r) is the interpolation function, and the calculation process is:

[0067]

[0068] 4 Comparative Experiments

[0069] Table 1 is a comparison of the AP of the algorithm of the present invention with different algorithms. The detection accuracy results of other algorithms are derived from the official KITTI ranking list data. The compared algorithms include VoxelNet, F-PointNet, AVOD, SECOND, DVFENet, PointPillars, and ComplexYOLO. As can be seen from Table 1, the detection accuracy of the algorithm of the present invention is much higher than that of classic three-dimensional target detection algorithms such as VoxelNet, FPointNet, and Complex-YOLO, and the real-time detection of the algorithm is realized. The experiment shows that the detection accuracy of this method is high and the real-time performance is strong.

[0070] Table 1 Comparison with mainstream algorithms on the KITTI dataset

[0071]

[0072] 5 Ablation Experiment

[0073] In order to verify the effectiveness of the algorithm and each improved module of the present invention, each improved part is gradually added and an ablation experiment is performed on the KITTI dataset to verify the impact of each improved part on the detection performance of the algorithm. The results are shown in Table 2.

[0074] Table 2 Ablation test results of the proposed algorithm on the KITTI validation set

[0075]

[0076]

[0077] Table 2 shows in detail the impact of each improved part on the 3D object detection performance. For point cloud 3D object detection, IDW convolution and MSA attention mechanism are used, which increases the network complexity and affects the detection speed of the network model. In experimental networks 2 and 3, the network adds IDW convolution layer and MSA attention mechanism respectively, and the detection speed of the model is improved. Therefore, the experiment shows that the use of IDW-MSA hybrid backbone for 3D object detection can meet the requirements of real-time detection, and the detection accuracy is significantly improved.

[0078] 6 Comparison of 3D object detection results

[0079] To demonstrate the detection capability of the IDW-MSA hybrid backbone, visualization results are provided. Figure 3 (a) The original image, (b) the detection results of Complex-YOLO, and (c) the detection results of Ours are shown. The detection results are magnified for easier observation. The method of the present invention achieves a significant improvement in detection speed, which is better than Complex-YOLO. This improvement is of great value for practical applications in 3D object detection and intelligent driving systems.

[0080] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should be protected by the claims attached to the present invention.

Claims

1. A voxel-oriented end-to-end 3D object detection method, characterized in that: The following steps are involved: A1 point cloud data input and data enhancement; A2 backbone network feature extraction: After the point cloud data is voxelized, it is processed by two IDW-MSA modules to extract rich feature information; the introduction of the IDW module and the use of large kernel convolution greatly increase the receptive field of the model, and the introduction of the MSA module and the use of the neuron perception attention mechanism, without adding additional parameters, also significantly improves the computational efficiency while maintaining a high detection accuracy; A3 highly compresses the voxels after feature extraction, projects them into a 2D bird's-eye view, directly places the sparse features on the BEV plane, and sums the features at the same position; A4 Detection and classification: The detection head is used to predict the category and location of the target.

2. The method according to claim 1, characterized in that: In the step A1, the KITTI dataset is read and divided into a training set and a validation set, and then the dataset is subjected to data enhancement operations of random rotation, random scaling, random flipping, and random translation.

3. The method according to claim 1, characterized in that In the step A2, the processed KITTI point cloud data is subjected to dynamic voxel feature encoding; this process divides the sparse point cloud data into three-dimensional voxel grids of different sizes, thereby simplifying the data representation; after voxelization, the result obtained is a structured voxel grid, each voxel containing point cloud information within a certain range; By constructing a hybrid backbone feature extraction architecture, voxel features are fully extracted for downstream tasks; first, the voxels in the sparse convolution tensor format are cascaded through two hybrid modules, and the hybrid module includes a large kernel convolution module (IDW) and an attention mechanism neuron perception module (MSA) connected in sequence. After being processed by the large kernel convolution module (IDW), a sparse convolution tensor is obtained, and then an attention mechanism neuron perception module (MSA) is connected; through the cascade of these two hybrid modules, voxel features can be fully extracted to support subsequent operations.

4. The method according to claim 3, characterized in that: In step A2, the IDW module decomposes the large convolution kernel into multiple small convolution kernels and performs feature transformation through 11×11×11 3D convolution operation. In this process, the feature dimension is not changed. Therefore, IDW divides the features of the sparse convolution tensor into four different parts along the channel dimension, denoted as {x hw ,x w ,x h ,x id }, each group uses a different convolution kernel for training, and finally merges them together; it is expressed as: x'=Concat(x hw ,x w ,x h ,x id ); it not only obtains a larger receptive field, but also improves the convolution efficiency, and also improves the anti-interference ability and robustness of the 3D target detector.

5. The method according to claim 3, characterized in that: In step A2, in MSA, the attention mechanism is used to infer the 3D attention weight for each neuron in the feature map. MSA calculates the mean and variance fusion of the input data and outputs it using the sigmoid activation function. The whole process does not require adding additional parameters to the original network.

6. The method according to claim 5, characterized in that In MSA, the input feature map x is marked as (b, c, d, h, w), where b represents the batch size, c represents the number of channels, d represents the depth, h represents the height, and w represents the width; the mean and variance of the input are calculated respectively, and then the adaptive adjustment factor e is proposed t The calculation formula is as follows: it is t =(x-mean)^2 / (4*(var+lamda))+0.5 Among them, mean is the average value of the feature map in the spatial dimension (height and width); var is the variance of the feature map in the spatial dimension, which is calculated using the mean; lamda is a variable parameter to prevent numerical instability and ensure the stability of the calculation process; finally, the output obtained by multiplying the original input x by the sigmoid activation function means that the model has made an adaptive adjustment to the features, thereby enhancing the sensitivity of the model to the features.

7. The method according to claim 1, characterized in that In step A3, in the detection head part, sparse maximum pooling and then full connection operations are performed to predict the location and category of the target; The sparse maximum pooling operation can effectively reduce the amount of calculation while retaining key feature information, while the fully connected layer is responsible for converting this information into the final prediction result; in this process, the detection head part not only needs to accurately locate the target, but also accurately identify the category of the target, which is crucial to ensure the safety and reliability of the autonomous driving system. In this way, the system can respond quickly and make decisions, thereby ensuring the safe driving of the vehicle in a complex traffic environment.

8. The method according to claim 2, characterized in that: In the step A1, the random rotation is as follows: the point cloud Vm in the marked mth target frame is selected to be rotated along the Z axis, and the rotation angle is set to θ. The original coordinates of the point are (x m ,y m ,z m ), the rotated coordinates (x', y', z') are calculated using the following formula: After the rotation transformation, the new point cloud set is recorded as V1.

9. The method according to claim 2, characterized in that: In the step A1, the random scaling is as follows: the point cloud Vm in the marked mth target frame is scaled, and the scaling factors are set to 0.5 and 1.5 respectively, to simulate the situation of vehicles of different sizes on the road. After the scaling transformation, the new point cloud data is recorded as V2.

10. The method according to claim 2, characterized in that In the step A1, the random translation is as follows: a point cloud Vm in the marked mth target frame is randomly generated with an offset offset(x,y), and the offset range x,y∈[-d,d] is set, where d is the maximum absolute value of the offset, so that the point cloud in the target frame is added with the randomly generated displacement value offset(x,y), and the point cloud after adding offset(x,y) is still within the river channel, simulating the target point cloud position of vehicles at different positions; after the translation transformation, the new point cloud data is obtained and recorded as V3.

Images