Radar point cloud sampling and semantic segmentation model system for autonomous driving scenarios

Abstract

The present invention relates to the field of autonomous driving scene perception technology, and in particular to a radar point cloud sampling and semantic segmentation model system for autonomous driving scenes. The system includes: a network input transformation module; a dual-branch autoencoder network model; and a semantic prediction module. The encoding layer in the upper branch autoencoder structure is composed of a local feature aggregation module and a random sampling operation, and the encoding layer of the lower branch is composed of a local feature aggregation module, a polar balanced random sampling module, and a random sampling operation; the upper and lower branch decoders each include four decoding layers, and the decoding layer is composed of a multi-layer perceptron and an upsampling operation. The polar balanced random sampling module proposed in the present invention helps the semantic segmentation model to fully learn point features at different distances and solve the long-tail problem of point distribution; the proposed consistency loss function can reduce the inconsistency of the model when using different sampling methods; the proposed model can efficiently perform complex outdoor scene parsing tasks.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving scene perception technology, and in particular to a radar point cloud sampling and semantic segmentation system for autonomous driving scenes. Background Art

[0002] Currently, methods for semantic segmentation tasks in autonomous driving scenarios mainly fall into three categories:

[0003] 1) Voxel-based methods: This method first discretizes 3D point cloud data into 3D voxel representations and then uses a 3D convolutional neural network model to predict semantic labels for these voxels. For example, the Cylinder3D method uses a cylindrical coordinate system to divide voxels.

[0004] 2) Point-based methods directly process unstructured, unordered 3D point cloud data and predict semantic labels for each point. For example, PointNet and PointNet++ are two pioneers of this type of method. They primarily use shared multi-layer perceptrons to learn point-by-point features, improving the performance of point-based network models. The RandLANet method uses random sampling to significantly improve the processing performance of large-scale point cloud data, and employs a local feature aggregation module to efficiently preserve geometric details and reduce the information loss caused by random sampling.

[0005] 3) Projection-based methods typically project 3D point cloud data into a 2D space (such as a depth map or bird's-eye view), then employ mature 2D convolutional neural network models for feature learning to improve inference speed. For example, SqueezeSeg is a groundbreaking method that uses spherical projection to convert irregular radar point cloud data into 2D image data, then employs the lightweight SqueezeNet and conditional random field methods to perform semantic segmentation.

[0006] With respect to the above prior art, the objective disadvantages thereof are as follows:

[0007] 1) Voxel-based methods: As the voxel resolution increases, the computational cost, model complexity, and memory consumption will increase significantly, limiting the execution efficiency of semantic segmentation.

[0008] 2) Projection-based methods cannot fully utilize 3D geometric information because they easily cause information loss during the projection process.

[0009] 3) Currently, most point-based methods use the farthest point sampling operation, which has high computational complexity and memory consumption. Therefore, most methods are only suitable for processing small-scale point clouds or indoor point clouds and cannot be extended to large-scale point clouds in autonomous driving scenarios.

[0010] 4) Although some methods use random sampling methods to achieve good performance of semantic segmentation models, they cannot solve the long-tail problem of radar point cloud distribution in autonomous driving tasks. That is, the points obtained by methods close to the radar sensor position are relatively dense, and as the distance increases, the density of points gradually becomes sparse. Summary of the Invention

[0011] The purpose of the present invention is to provide a radar point cloud sampling and semantic segmentation system for autonomous driving scenarios, which is used to solve the problems existing in the prior art in the above-mentioned background technology.

[0012] In order to achieve the above object, the present invention adopts the following technical solutions:

[0013] The present invention provides a radar point cloud sampling and semantic segmentation system for autonomous driving scenarios, characterized in that the system includes:

[0014] Network input transformation module: The input of the network is point cloud data with a dimension of N×dim, where N represents the total number of points in the input point cloud and dim refers to the feature dimension of each point. After the input point cloud is transformed through a fully connected layer, a feature representation of size N×8 is generated;

[0015] A dual-branch autoencoder network model, including an encoder and a decoder;

[0016] The encoder consists of four encoding layers, which are used to gradually reduce the size of the point cloud and increase the dimension of each point feature. The four encoding layers of the upper branch encoder are composed of a local feature aggregation module and a random sampling operation. The first encoding layer of the lower branch encoder is composed of a local feature aggregation module and a pole-balanced random sampling module, and the subsequent three encoding layers are composed of a local feature aggregation module and a random sampling operation. The decoders of the upper and lower branches have the same structure and include four decoding layers, each of which contains a multi-layer perceptron and an upsampling operation.

[0017] The semantic prediction module, consisting of three fully connected layers and one dropout layer, predicts a semantic label for each point.

[0018] Furthermore, the change in the number of points in the overall stage of the encoder is:

[0019] N→N / 4→N / 16→N / 64→N / 256;

[0020] The corresponding feature dimensions are:

[0021] 8→32→128→256→512;

[0022] The overall stage of the decoding layer, the number of points changes as follows:

[0023] N / 256 → N / 64 → N / 16 → N / 8 → N

[0024] The corresponding feature dimensions are:

[0025] 512→256→128→32→8.

[0026] Furthermore, the learning of the pole balanced random sampling deep semantic segmentation network model is determined by the loss function Conduct supervision, restraint and training;

[0027] Loss Function Expressed as:

[0028]

[0029] Among them, σ1 and σ2 are uncertainty weighting parameters, which are used to automatically adjust the optimization ratio of different loss terms;

[0030]

[0031] in, and They represent the predicted probabilities when using the polar balanced random sampling and random sampling methods respectively; C represents the number of categories in the data set, and c represents the cth category;

[0032]

[0033] Among them, C represents the number of categories in the data set, c represents the cth category, and w c represents the weight of category c, y c and represent the true value and the predicted probability respectively.

[0034] Furthermore, the establishment method of the pole-balanced random sampling module is as follows:

[0035] First, convert the Cartesian coordinate system where the 3D point cloud is located into a polar coordinate system:

[0036] Convert the point coordinates (x, y, z) to (ρ, θ, z), where ρ represents the radius, that is, the distance from the origin of the xy plane, θ represents the angle from the x-axis to the y-axis, and z represents the height;

[0037] Based on the constructed polar coordinate system, the input 3D point cloud is divided into different cylindrical blocks, the number of which is recorded as K. The farther the distance, the larger the cylindrical block. The final resolution of the obtained 3D polar representation is R×P×Z, where R, P, and Z represent radius, azimuth, and altitude, respectively.

[0038] Secondly, based on the number of target sampling points M and the number of divided cylindrical blocks K, the number of points to be downsampled in each cylindrical block is calculated and constructed into a list S with a dimension of K×1 n ; And during the downsampling process, try to balance the number of points in each cylindrical block to keep them as consistent as possible;

[0039] Finally, a sampling operation is performed in each cylindrical block: the points in each block are scrambled, and the previous S n [i], where i represents the i-th cylindrical block, forming a sub-point cloud; all the obtained sub-point clouds are combined together, and then a scrambling operation is performed to obtain the final sampling point cloud.

[0040] The present invention has at least the following beneficial effects:

[0041] The present invention proposes a new pole-balanced random sampling module, which can complete the sampling of three-dimensional point cloud data. The number of points in different distance ranges should be as uniformly distributed as possible, thereby helping the semantic segmentation model to fully learn the point features at different distances and solve the long-tail problem of point distribution.

[0042] This paper proposes a sampling consistency loss function that can learn the variability of different sampling methods, enhance the representation ability of the model by learning more diverse feature representations, and reduce the inconsistency of the model when using different sampling methods.

[0043] By combining the above two strategies, the present invention designs a new deep learning model for radar point cloud semantic segmentation for autonomous driving scenarios, and verifies it on a public dataset, showing that the model of the present invention has superior performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 Schematic diagram of the visualization and quantitative statistical analysis of the original input 3D point cloud data, random downsampling, and the pole-balanced random sampling results proposed by the present invention;

[0046] Figure 2 This is a diagram of the polar balanced random sampling network model;

[0047] Figure 3This is a visualization of the input point cloud, ground truth value, RandLA-Net semantic segmentation results, and the semantic segmentation results of the polar balanced random sampling network model (PCB-RandNet) of this patent. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] 1. Polar Cylinder Balanced Random Sampling (PCB-RS)

[0050] Please refer to the following for details: Figure 1 , Figure 1 The following are visualizations and quantitative statistical analysis of the original input 3D point cloud data, the random downsampling method, and the polar-equalized random sampling method proposed in this paper. Rows 1, 3, and 5 show the point cloud data visualization results; rows 2, 4, and 6 show the statistical histograms of the corresponding point counts relative to the distance from the radar sensor.

[0051] from Figure 1 In the visualization results, we can clearly see that the point density of radar point cloud data for autonomous driving scenarios is different at different locations, and the density distribution of the 3D point cloud and the distance between the point and the radar sensor show a long-tail distribution, that is, the closer to the sensor, the denser the point cloud data obtained, and as the distance increases, the point cloud density gradually decreases. In addition, by analyzing Figure 1 We further discovered that a long-tail distribution tends to result in closer points being more likely to be retained, while points at medium or long distances are more likely to be discarded. This results in the network model being unable to learn sufficient and effective features from points at different distances. Random sampling methods cannot address this imbalanced distribution problem.

[0052] Therefore, in order to solve the above problems, the present invention proposes a new pole-balanced random sampling strategy. The core idea is that the number of points in different distance ranges of the three-dimensional point cloud data obtained after downsampling should be as uniformly distributed as possible.

[0053] See Algorithm 1 in the table below, which describes the polar-balanced random sampling process. Specifically, the polar-balanced random sampling strategy first converts the Cartesian coordinate system of the 3D point cloud into a polar coordinate system. The point coordinates (x, y, z) are converted into (ρ, θ, z), where ρ represents the radius (the distance from the origin of the xy plane), θ represents the angle from the x-axis to the y-axis, and z represents the height. Based on the constructed polar coordinate system, the input 3D point cloud is divided into different cylindrical blocks. The greater the distance, the larger the cylindrical blocks. The resulting 3D polar representation has a resolution of R×P×Z, where R, P, and Z represent the radius, azimuth, and height, respectively.

[0054] Secondly, based on the number of target sampling points M and the number of divided cylindrical blocks K, the number of points to be downsampled in each cylindrical block is calculated and constructed into a list of dimension K×1, S n During the downsampling process, the number of points in each cylindrical block is balanced as much as possible to keep them as consistent as possible.

[0055] Finally, a sampling operation similar to random sampling is performed in each cylindrical block. Specifically, the points in each block are scrambled and the previous S n [i] (i represents the i-th cylindrical block) constitutes a sub-point cloud. Then all the obtained sub-point clouds are combined together and a scrambling operation is performed to obtain the final sampling point cloud.

[0056] The use of the pole-balanced random sampling strategy can ensure that the distribution of sampling points in different distance ranges is as uniform as possible. Figure 1 The visualization results in columns 4 and 5 also demonstrate that the proposed sampling method can change the point cloud distribution and retain as many points as possible at medium and long distances. This further ensures that the deep learning network model can learn more comprehensive feature information suitable for semantic segmentation from points at different distance ranges.

[0057]

[0058] 2. Sampling Consistency Loss (SCL)

[0059] Radar semantic segmentation datasets often suffer from class imbalance. For example, in the SemanticKITTI dataset, the ratio of roads and sidewalks is hundreds of times higher than that of people and motorcycles.

[0060] To solve this problem, we choose to use the weighted cross entropy loss function To better optimize the category distribution, the formula is as follows:

[0061]

[0062] Where C represents the number of categories in the data set, c represents the cth category, ω c represents the weight of category c, y c and represent the true value and the predicted probability respectively.

[0063] In addition, in order to enable the model to learn the variability of different sampling methods, the present invention designs a sampling consistency loss function By learning more diverse feature representations, the model's representational capabilities are enhanced. At the same time, it can constrain the probability distribution of the model's output to remain consistent as much as possible under different sampling methods. The specific formula is defined as follows:

[0064]

[0065] and They represent the predicted probabilities when using the extreme column balanced random sampling and random sampling methods, respectively. C represents the number of categories in the dataset, and c represents the cth category.

[0066] The overall loss function for supervising the network model learning proposed by this invention is the weighted sum of the above two loss functions:

[0067]

[0068] However, this loss function leads to a tedious process of searching for optimal weight parameters.

[0069] To address this issue, the present invention introduces an uncertainty weighting method, which automatically adjusts the optimization ratio of different loss terms by introducing two learnable parameters σ1 and σ2. To maintain stability, the present invention also adds two additional logarithmic regularization terms. Therefore, the final loss function is defined as follows:

[0070]

[0071] 3. Polar Cylinder Balanced Random Sampling Network (PCB-RandNet)

[0072] Figure 2 The overall structure of the pole-balanced random sampling deep semantic segmentation network model proposed in this invention for autonomous driving scenarios is given.

[0073] The model is mainly composed of two autoencoder models with similar structures and shared parameters, namely a specially designed dual-branch autoencoder network model. The details are as follows:

[0074] Network Input Transformation Module: The network input is radar point cloud data from large-scale outdoor scenes for autonomous driving scenarios. The point cloud has dimensions of N × dim, where N represents the total number of points in the input point cloud and dim refers to the feature dimension of each point (such as x, y, z coordinates and R, G, B colors). After the input point cloud is transformed through a fully connected layer, a feature representation of size N × 8 is generated.

[0075] Encoder: Such as Figure 2 As shown in the figure, the deep network model proposed in the present invention is composed of a specially designed dual-branch autoencoder network model, which includes an encoder and a decoder.

[0076] The encoder uses four encoding layers to gradually reduce the size of the point cloud and increase the dimensionality of each point feature. Each encoder layer consists of a local feature aggregation module derived from the RandLANet model and a random sampling operation. It is important to note that, unlike the encoder structure in the upper branch of the model, the first encoder layer in the lower branch of the model uses the proposed pole-balanced random sampling method instead of random sampling. This is due to the following reasons: the point cloud after pole-balanced random sampling is more evenly distributed across different distances; this balanced distribution is well maintained in subsequent stages of the deep learning network, even when random sampling is used for downsampling. Therefore, the proposed sampling strategy is only used in the first encoder layer. Furthermore, since pole-balanced random sampling can be viewed as a collection of several sub-random sampling processes, this sampling strategy configuration also enables efficient network model operation with low computational cost. Overall, during the encoder stage, the number of points changes from (N→N / 4→N / 16→N / 64→N / 256), and the corresponding feature dimensions change from (8→32→128→256→512).

[0077] Decoder: The upper and lower branches of the network use the same decoder structure. Each decoder layer contains a multi-layer perceptron and an upsampling operation. The upsampling operation uses the nearest neighbor interpolation strategy. The corresponding number of points changes from (N / 256 → N / 64 → N / 16 → N / 8 → N), and the corresponding feature dimensions are (512 → 256 → 128 → 32 → 8).

[0078] Semantic prediction: At the back end of the autoencoder structure, a semantic label is predicted for each point by adding three fully connected layers and a dropout layer, with the following dimensions: (N, 8) → (N, 64) → (N, 32) → (N, n class ), where n class The prediction outputs of the upper and lower network models are supervised and constrained by the sampling consistency loss function proposed by the present invention.

[0079] Finally, the learning of the entire semantic segmentation network model will be supervised by the loss function designed in Section 2.

[0080] 4. Technical Effect

[0081] Dataset: SemanticKITTI and SmanticPOSS datasets are used to verify the effectiveness and performance of the present invention. The SemanticKITTI dataset is the largest dataset for point cloud semantic segmentation tasks in autonomous driving tasks. It contains 22 sequences of 43,552 radar scans collected from a city in Germany and densely annotated based on points. Sequences 00 to 10 (19,130 scans) are used for training, and sequence 08 (4,071 scans) is used for verification. Sequences 11 to 21 (20,351) are used for testing. SemanticPOSS is a relatively small, sparse and challenging dataset collected on campus by Peking University using the Hesai Panda radar sensor. It contains 2,988 different complex scenes. The data is divided into 6 data sequences, of which sequence 03 is used for testing and the remaining sequences are used for training.

[0082] Implementation details: The present invention is implemented in Python language combined with the PyTorch deep learning framework. All verification experiments are performed on the NVIDIA RTX 3090 GPU. The network model proposed in the present invention uses Adam as the optimizer, trained for 100 epochs, and the initial learning rate is set to 0.01, which is reduced by 5% after each epoch. For the two datasets, the resolution of the polar representation used in the polar balanced random sampling is 64x64x16. For the SemanticKITTI dataset, the point cloud is kept in the range of [distance: 3~max(·)m, z: [-3.0~1.5m]]. For the SemanticPOSS dataset, the point cloud is kept in the range of [distance: 3~80m, z: -3.0~3m].

[0083] Results: Table 1 shows the segmentation performance of the method proposed in the present invention in different distance ranges. The results in the table show that the pole-balanced random sampling method, sampling consistency loss and pole-balanced random sampling network model proposed in the present invention can significantly improve the semantic segmentation performance of the network model in different distance ranges. Specifically, in three medium and long distance ranges (30m~40m, 40m~50m and greater than 50m), compared with the random sampling method, the pole-balanced random sampling strategy of the present invention improves the average intersection-over-union (mIoU) performance by 1.9%, 2.2% and 1.9% respectively. In addition, after introducing the sampling consistency loss, the performance of the network model achieved an average intersection-over-union (mIoU) of 2.3%, 2.3%, 3.1%, 5.5%, 4.2% and 4.6% in the six distance ranges respectively. Figure 3 The visual display of the input point cloud, ground truth value, RandLA-Net semantic segmentation result and the semantic segmentation result of the polar balanced random sampling network model (PCB-RandNet) of the present invention are respectively given, which can clearly give the qualitative analysis results.

[0084] Table 1: Evaluation results at different distance ranges on the SemanticKITTI validation set. RS indicates random sampling is used in the method. PCB-RS indicates the method incorporates pole-balanced random sampling. PCB-RS+SCL indicates that both pole-balanced random sampling and sampling consistency loss are used in the network model.

[0085]

[0086] Table 2 compares the validation results of the proposed method and other related work on the SemanticPOSS dataset. BAF-LAC, BAAR, and RandLA-Net are three mainstream methods for radar point cloud semantic segmentation. In the experiments, the proposed polar-balanced random sampling strategy and sampling consistency loss function are introduced into each of these three methods to further demonstrate the effectiveness of the proposed sampling strategy and loss function. The results in the table show that the proposed polar-balanced random sampling and sampling consistency loss function indeed bring significant performance improvements on more challenging data containing smaller objects. Specifically, for the BAF-LAC method, the use of polar-balanced random sampling and sampling consistency loss improves mean intersection over union (mIoU) by 2.9% and 2.4%, respectively. Their combined use improves performance from 51.2% to 56.5%. For the BAAF method, polar-balanced random sampling and sampling consistency improve mIoU by 2.4% and 1.0%, respectively. Combining the two improves segmentation performance from 54.4% to 58.0%. For RandLA-Net, the introduction of pole-balanced random sampling and sampling consistency achieved performance improvements of 2.7% and 1.3% respectively. The combination of the two improved the performance from 52.8% to 56.8% mean intersection over union (mIoU).

[0087] Table 2: Evaluation results on the SemanticPOSS dataset, where BAF-LAC, BAAF, and RandLA-Net are the three mainstream methods in the field of radar semantic segmentation. For fair comparison, the present invention uses Python and PyTorch frameworks to reproduce the above methods, and introduces the pole-balanced random sampling strategy and sampling consistency loss of the present invention into these methods.

[0088]

[0089] The effect of polar balanced random sampling is further verified: Table 3 shows the semantic segmentation performance of the deep network model proposed in the present invention in different distance ranges when different sampling methods are used. Taking into account the randomness of the entire network structure, in order to make a fair comparison, the results under three different random seeds are given. The experimental results show that: 1) As the distance to the radar sensor increases, the segmentation performance drops sharply. 2) Compared with random sampling, the polar balanced random sampling proposed in the present invention achieves significant performance improvement at medium and long distances. 3) In the close range (<10m), polar balanced random sampling also achieves a certain performance improvement. The reason why random sampling fails to achieve better performance compared with the sampling strategy proposed in the present invention is that random sampling mainly aggregates areas at closer distances, and the network overfits to these closer features, which leads to a relative performance degradation.

[0090] Table 3 Comparison of mean intersection and union of random sampling and pole-balanced random sampling at different distance ranges on the SemanticKITTI validation set under different random seed settings

[0091]

[0092] The effect of sampling consistency loss is further verified: Table 4 shows the verification results of further verifying the effectiveness of the sampling consistency loss function and uncertainty weighting method proposed in the present invention. It can be seen from the results that the fixed weight value only brings insignificant performance improvement, and may even cause performance degradation. However, by adopting a learnable weighting method, better performance can be obviously obtained. In addition, the verification results on the sub-sampling set also show that the introduction of the sampling consistency loss function can reduce the performance variability of the model when using different random sampling methods, and the learnable weighting method can also reduce this variability.

[0093] Table 4 Evaluation of sampling consistency loss and different loss weights on the SemanticKITTI validation set (valset)

[0094]

[0095] The SemanticKITTI dataset was selected to further validate the effectiveness of the proposed polar-balanced random sampling and sampling consistency loss introduced in the BAF-LAC and BAAR methods. The experimental results in Table 5 demonstrate that the introduction of balanced random sampling and sampling consistency loss leads to sustained performance improvements for the BAF-LAC and BAAR methods. Furthermore, as expected, the performance improvements at medium and long distances are significant.

[0096] Table 5 Evaluation results of BAF-LAC and BAAF methods introducing polar balanced random sampling and sampling consistency loss in different distance ranges of SemanticKITTI dataset

[0097]

[0098] In summary:

[0099] (1) The present invention proposes a new polar cylinder balanced random sampling (PCB-RS) method to complete the sampling of three-dimensional point cloud data, which helps the semantic segmentation model to fully learn point features at different distances and solve the long-tail problem of point distribution.

[0100] (2) This paper proposes a sampling consistency loss function (SCL) to reduce the inconsistency of the model when using different sampling methods.

[0101] (3) Combining the above two strategies, a new deep learning model for radar point cloud semantic segmentation in autonomous driving scenarios was designed, named the Polar Cylinder Balanced Random Sampling Network (PCB-RandNet) model, which achieved superior performance on public datasets.

[0102] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the invention as claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A radar point cloud sampling and semantic segmentation system for autonomous driving scenarios, characterized by: The system comprises: Network input transformation module: The input of the network is point cloud data with a dimension of N×dim, where N represents the total number of points in the input point cloud and dim refers to the feature dimension of each point. After the input point cloud is transformed through a fully connected layer, a feature representation of size N×8 is generated; A dual-branch autoencoder network model, including an encoder and a decoder; The encoder consists of four encoding layers, which are used to gradually reduce the size of the point cloud and increase the dimension of each point feature. The four encoding layers of the upper branch encoder are composed of a local feature aggregation module and a random sampling operation. The first encoding layer of the lower branch encoder is composed of a local feature aggregation module and a pole-balanced random sampling module, and the subsequent three encoding layers are composed of a local feature aggregation module and a random sampling operation. The decoders of the upper and lower branches have the same structure and both include four decoding layers, each of which contains a multi-layer perceptron and an upsampling operation. The semantic prediction module consists of three fully connected layers and one dropout layer to predict a semantic label for each point; The establishment method of the pole-balanced random sampling module is as follows: First, convert the Cartesian coordinate system where the 3D point cloud is located into a polar coordinate system: Convert the point coordinates (x, y, z) to (ρ, θ, z), where ρ represents the radius, that is, the distance from the origin of the xy plane, θ represents the angle from the x-axis to the y-axis, and z represents the height; Based on the constructed polar coordinate system, the input 3D point cloud is divided into different cylindrical blocks, the number of which is recorded as K. The farther the distance, the larger the cylindrical block. The final resolution of the obtained 3D polar representation is R×P×Z, where R, P, and Z represent radius, azimuth, and altitude, respectively. Secondly, based on the number of target sampling points M and the number of divided cylindrical blocks K, the number of points to be downsampled in each cylindrical block is calculated and constructed into a list S with a dimension of K×1 n ; And during the downsampling process, try to balance the number of points in each cylindrical block to keep them as consistent as possible; Finally, a sampling operation is performed in each cylindrical block: the points in each block are scrambled, and the previous S n [i], where i represents the i-th cylindrical block, forming a sub-point cloud; all the obtained sub-point clouds are combined together, and then a scrambling operation is performed to obtain the final sampling point cloud.

2. The radar point cloud sampling and semantic segmentation system for autonomous driving scenarios according to claim 1, characterized in that: The overall stage of the encoder, the change of its point number is: N→N / 4→N / 16→N / 64→N / 256; The corresponding feature dimensions are: 8→32→128→256→512； The overall stage of the decoding layer, the number of points changes as follows: N / 256 → N / 64 → N / 16 → N / 8 → N The corresponding feature dimensions are: 512→256→128→32→8。 3. The radar point cloud sampling and semantic segmentation system for autonomous driving scenarios according to claim 1, characterized in that: The learning of the polar balanced random sampling deep semantic segmentation network model is determined by the loss function Conduct supervision, restraint and training; Loss Function Expressed as: Among them, σ1 and σ2 are uncertainty weighting parameters, which are used to automatically adjust the optimization ratio of different loss terms; in, and They represent the predicted probabilities when using the polar balanced random sampling and random sampling methods respectively; C represents the number of categories in the data set, and c represents the cth category; Among them, C represents the number of categories in the data set, c represents the cth category, ω c represents the weight of category c, y c and represent the true value and the predicted probability respectively.

Images