Laser point cloud semantic segmentation method and system for automatic driving and medium

Abstract

The invention relates to a laser point cloud semantic segmentation method and system for automatic driving and a medium, and the method comprises the steps: M1, laser point cloud acquisition and downsampling: capturing three-dimensional point cloud data in an environment through a laser radar sensor, carrying out the preprocessing and downsampling, removing redundant points, and reserving representative information points; m2, performing voxelization modeling on a view cone: performing spatial division on the three-dimensional point cloud data through a voxelization model, so that the point cloud is expressed in a voxel form; and M3, deeply processing the Point Net network model: taking the point cloud data as input. According to the method, point-level and projection-level cross entropy loss is adopted to ensure the overall classification precision, boundary loss is innovatively introduced, the category boundary discrimination capability is effectively enhanced, misclassification of a segmentation result at the edge of an object is remarkably reduced, and the classification efficiency is improved. Therefore, the fine semantic understanding ability and robustness of the model to a complex static environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to a laser point cloud semantic segmentation method and system for automatic driving and a medium. BACKGROUND

[0002] Laser point cloud semantic segmentation, as a key field of three-dimensional data processing, focuses on classifying three-dimensional point clouds collected by devices such as LiDAR, i.e., assigning specific class labels to each point. This technology has shown important application value in multiple fields such as autonomous driving, robot navigation, and geographic information systems (GIS).

[0003] LiDAR technology LiDAR measures distance accurately by emitting and receiving reflected laser pulses, generating detailed three-dimensional point cloud data. Such data not only reveals object shape and position information, but also provides important clues about environmental structure, serving as the basis for three-dimensional spatial analysis and understanding.

[0004] Application of deep learning in point cloud processing With the development of deep learning, its powerful feature extraction capability and excellent generalization performance make it a powerful tool for point cloud processing. In view of the unique nature of point clouds, researchers have developed a series of specialized network architectures, such as point cloud segmentation methods based on projection maps and voxels, significantly improving processing results.

[0005] Multi-modal data fusion In some cases, using LiDAR data alone may not meet the ideal segmentation accuracy requirements. Therefore, combining other types of sensor data (such as RGB cameras, millimeter wave radars, etc.) has become an effective strategy to improve segmentation accuracy. This multi-source data fusion approach can fully utilize the advantages of different sensors, providing more comprehensive environmental perception capabilities.

[0006] Computational resources and optimization Point cloud data is large in volume, requiring high computational resources. In order to efficiently utilize existing hardware resources (such as GPU, TPU), and reduce computational complexity, research focuses on the application of model compression, sparsification, etc. techniques, aiming to improve processing efficiency while maintaining good performance.

[0007] Open source tools and frameworks The progress of technology has given rise to numerous open source tools and frameworks supporting point cloud processing, such as PCL (Point Cloud Library), Open3D, and PyTorch, etc. These resources greatly facilitate the exchange and cooperation between academia and industry, reducing the learning threshold and technical adoption difficulty for new entrants.

[0008] Point cloud semantic segmentation algorithm Point cloud segmentation, as one of the core tasks of three-dimensional computer vision, aims to divide point cloud data into different parts or regions, representing different objects, surfaces or scene elements. According to the processing mechanism, point cloud segmentation algorithms can be roughly divided into two categories: point-based methods and projection map-based methods. Each method has its unique advantages and application scenarios, and together they have promoted the progress and development of the field.

[0009] In the prior art, point-based point cloud segmentation algorithms, point-based point cloud semantic segmentation methods directly operate on the original, unordered, unstructured point cloud, avoiding information loss caused by voxelization or projection, and can preserve complete geometric details, making them particularly suitable for small objects, fine structures and boundary-sensitive scenes (such as pedestrians, bicycles, curbs, etc. in autonomous driving). The core challenge of this method is how to achieve efficient and permutation-invariant learning on unordered, non-uniform, high-dimensional sparse data. Although the point-based point cloud semantic segmentation method in this method can preserve the original geometric details and avoid quantization loss, its main disadvantage is the large computational and memory overhead: due to the need for dynamic search and feature aggregation (such as k-nearest neighbors or sphere queries) for each point and its neighborhood, it is inefficient when processing large-scale autonomous driving LiDAR point clouds (single frame up to hundreds of thousands to millions of points); at the same time, the unordered nature and non-uniform density make network design complex, making it difficult to ensure stability and real-time performance; in addition, the lack of regular structure makes it difficult to fully utilize the advantages of GPU parallel computing, limiting its deployment capabilities on vehicle embedded platforms. SUMMARY

[0010] In view of the above problems, the present application provides a laser point cloud semantic segmentation method, system and medium for autonomous driving, which not only adopts cross-entropy loss at the point level and projection map level to ensure overall classification accuracy, but also innovatively introduces boundary loss, effectively enhancing the discriminability of class boundaries and significantly reducing the misclassification phenomenon of segmentation results at object edges, thereby improving the fine semantic understanding ability and robustness of the model in complex static environments.

[0011] To achieve the above objects and other related objects, the technical solutions provided by the present application are as follows: A laser point cloud semantic segmentation method for autonomous driving, the method comprising: M1. Laser point cloud acquisition and downsampling: capturing three-dimensional point cloud data in the environment through a laser radar sensor, and preprocessing and downsampling to remove redundant points and retain representative information points; M2. Voxelization modeling: Through voxelization modeling, the three-dimensional point cloud data is spatially divided, so that the point cloud is represented in the form of voxels; M3. Deep processing of PointNet network model: Point cloud data is taken as input, and local and global features of each point are learned. PointNet learns independent features for each point, and then generates global descriptors through global feature aggregation operation; M4. Point cloud aggregation and 2D feature extraction: The voxelized point cloud data, i.e. distance map, is combined with other features to form aggregated point cloud data. Further feature extraction is performed based on the output of the yolov7Tiny model to capture more local and global information; M5. Scattering point feature extraction and fusion point feature: The voxelized point cloud is encoded into the feature layer, then projected to the 2D feature layer of the cone expansion, and then the 2D feature is convolved and encoded back to the point feature. The fusion of point feature and projected graph feature is completed, and the three-dimensional features extracted by PointNet and 2D features are fused; M6. ROI static environment semantics: According to the point class confidence, the region of interest is extracted, the ROI extracts specific class objects, and the extracted ROI region is further analyzed to obtain the class information of each point in the static environment, so that the model can recognize the objects and their classes in the scene, and form a complete semantic understanding of the environment.

[0012] Further, in step M2, the voxelization model uses the auxiliary coordinate system to perform cylindrical voxelization modeling on the point cloud data, converts the point cloud data from Cartesian coordinate system to cylindrical coordinate system, i.e. from coordinates (x, y, z) to (r, θ, Φ), and the conversion function is: , , , and the point cloud data is converted into a distance image through spherical projection.

[0013] Further, the aggregated point cloud data includes geometric features and texture features, as well as optical features extracted from images.

[0014] Further, in step M6, the calculation of the point class confidence is based on the fused features, and each point is predicted to be a certain class, and the model generates a confidence for each class prediction. The confidence represents the degree of trust of the model in the point classification, and a high confidence means that the point belongs to the same class.

[0015] Further, according to the fact that each point is predicted as a certain class, a loss function is constructed, first, for the semantic class of each point, a cross-entropy loss function is used to calculate the loss of each point, for the projection map, when constructing the network structure, the resolution of the semantic output is increased to the size of the projection map, and the cross-entropy loss is calculated for the true value projection map and the predicted projection map, at the same time, the segmentation boundary of the projection map also affects the final segmentation accuracy, so a boundary loss is introduced to represent the boundary between classes, and the boundary loss between the predicted projection map and the true value projection map is calculated.

[0016] Further, the boundary loss of the semantic boundary region of the distance image is calculated by extracting the boundary mask from the true value projection map through an edge detection operator or semantic discontinuity analysis, and in the boundary mask region, the difference between the predicted projection map and the true value projection map is weighted and punished by a loss function or a boundary-aware loss function form.

[0017] Further, in step M1, the preprocessing includes rotation, translation and splicing of the point cloud data.

[0018] In order to achieve the above-mentioned purpose and other related purposes, the application further provides a laser point cloud semantic segmentation system for automatic driving, comprising a computer device programmed or configured to execute the steps of the laser point cloud semantic segmentation method for automatic driving.

[0019] In order to achieve the above-mentioned purpose and other related purposes, the application further provides a computer readable storage medium having a computer program programmed or configured to execute the laser point cloud semantic segmentation method for automatic driving stored thereon.

[0020] The application has the following positive effects: 1. The application constructs an efficient and easy-to-deploy multi-modal point cloud semantic segmentation framework by fusing columnar voxelized point cloud, PointNet three-dimensional features and two-dimensional image features; one of its core advantages is the careful design of the loss function - not only cross-entropy loss at the point level and the projection map level is used to ensure the overall classification accuracy, but also boundary loss is innovatively introduced to effectively enhance the discrimination ability of the class boundary and significantly reduce the misclassification phenomenon of the segmentation result at the edge of the object, thereby improving the fine semantic understanding ability and robustness of the model to complex static environment.

[0021] 2.The method combines laser point cloud data with a deep learning model, adopts a multi-modal feature fusion strategy, and realizes efficient point cloud semantic segmentation. By extracting 2D features through a 2D encoder and extracting 3D features through PointNet, and combining voxel modeling and confidence calculation, the model can perform detailed semantic understanding on point cloud data, and is suitable for scenarios such as autonomous driving, robot navigation and intelligent monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A flowchart of the method of the present application is shown. Figure 2 A schematic diagram of the logic architecture of the present application is shown. Figure 3 A schematic diagram of the view cone voxel modeling of the present application is shown. DETAILED DESCRIPTION

[0023] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to assist in understanding, which should be considered only as exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Also, in order to be clear and concise, descriptions of well-known functions and structures are omitted in the following description.

[0024] Embodiment 1: As shown in Figure 1 or Figure 2 A laser point cloud semantic segmentation method for autonomous driving, the method comprising: M1.Laser point cloud acquisition and downsampling: through a laser radar sensor, three-dimensional point cloud data in the environment is captured, preprocessed and downsampled, redundant points are removed, and representative information points are retained; M2.View cone voxel modeling: through a voxel modeling, the three-dimensional point cloud data is spatially divided, so that the point cloud is represented in the form of voxels; M3.PointNet network model deep processing: taking the point cloud data as input, learning the local and global features of each point, PointNet independently learns the features of each point, and then generates a global descriptor through a global feature aggregation operation; M4.Point cloud aggregation and 2D feature extraction: merging the voxelized point cloud data, i.e., the distance map and other features, to form the aggregated point cloud data, and further extracting features based on the output of the yolov7Tiny model to capture more local and global information; M5. Scattering point feature extraction and fusion: The voxelized point cloud is encoded into a feature layer, then projected into a 2D feature layer of a view volume expansion, and then the 2D feature is convolved and encoded back to the point feature, completing the fusion of the point feature and the projected graph feature, and fusing the three-dimensional feature extracted by PointNet with the 2D feature; M6. ROI static environment semantics: According to the class confidence of the point, the region of interest is extracted, the ROI extracts the object of a specific class, and the ROI region extracted is further analyzed to obtain the class information of each point in the static environment, so that the model identifies the objects and their classes in the scene, and forms a complete semantic understanding of the environment.

[0025] In the present embodiment, as shown in Figure 3 M2, the voxelization model is a cylindrical voxelization model using an auxiliary coordinate system to model the point cloud data, which converts the point cloud data from a Cartesian coordinate system to a cylindrical coordinate system, i.e., from coordinates (x, y, z) to (r, θ, Φ), and the conversion function is: , , , and the point cloud data is converted into a distance image by spherical projection.

[0026] In the present embodiment, the aggregated point cloud data includes geometric features and texture features, as well as optical features extracted from images.

[0027] In the present embodiment, in step M6, the calculation of the class confidence of the point is based on the fused features, and each point is predicted to be a certain class, and the model generates a confidence for each class prediction, which represents the degree of trust of the model in the classification of the point, and a high confidence means that the point belongs to the same class.

[0028] In the present embodiment, according to the fact that each point is predicted to be a certain class, a loss function is constructed, first, for the semantic class of each point, a cross-entropy loss function is used to calculate the loss for each point, for the projection graph, the resolution of the semantic output is increased to the size of the projection graph when constructing the network structure, and the cross-entropy loss is calculated for the true value projection graph and the predicted projection graph, at the same time, the segmentation boundary of the projection graph also affects the final segmentation accuracy, so a boundary loss is introduced, and the boundary loss between the predicted projection graph and the true value projection graph is calculated to represent the boundary between the classes.

[0029] In this embodiment, the boundary loss of the semantic boundary region calculation of the distance image is that a boundary mask is extracted from the ground truth projection map by an edge detection operator or semantic discontinuity analysis, and a weighted loss penalty or a boundary-aware loss function form is applied to the difference between the predicted projection map and the ground truth projection map in the boundary mask region.

[0030] In this embodiment, in step M1, the preprocessing includes rotation, translation and splicing of the point cloud data.

[0031] Embodiment 2: Based on the laser point cloud semantic segmentation method for automatic driving in embodiment 1, the present application is further described and explained as follows.

[0032] As shown in Figure 1 or Figure 2 (1) Laser point cloud acquisition and downsampling Three-dimensional point cloud data in the environment is captured by a laser radar sensor (such as LiDAR). Each point cloud contains XYZ coordinates and reflection intensity values, and the point cloud can represent the position information of objects in the environment in three-dimensional space. Since the original point cloud data is usually very dense and contains a large number of points, direct processing will consume a large amount of computing resources. Downsampling reduces the number of points by removing redundant points and retaining representative information points, thereby reducing the computational burden. A series of data enhancement steps are also performed on the point cloud data during the point cloud data processing process, such as rotation, translation, splicing, etc.

[0033] (2) Voxelization modeling Through voxelization modeling, three-dimensional point cloud data is spatially divided so that the point cloud is represented in the form of voxels (similar to 3D pixels). The voxelization process can structure the point cloud data in three-dimensional space, facilitating subsequent feature extraction and processing. The output of this step is a voxel grid formed, in which the point cloud data is discretized. The data within each voxel grid represents the local point cloud density and features. Traditional voxelization methods use a Cartesian coordinate system, and this method uses an auxiliary coordinate system for cylindrical voxelization modeling of point cloud data, converting point cloud data from a Cartesian coordinate system to a cylindrical coordinate system, i.e., from coordinates (x, y, z) to (r, θ, Φ), with the conversion formula being: , , , The point distribution obtained by subdividing the data using columnar voxels instead of cubic voxels is more uniform, with a higher proportion of non-empty areas, and this becomes more pronounced with increasing distance. Furthermore, the point cloud data is transformed into a distance image through spherical projection. The point cloud obtained by the LiDAR rotating and scanning for one revolution is equivalent to a hollow cylinder centered on itself. If this cylinder is unfolded, the point cloud can be projected onto an image plane, which is the distance image.

[0034] The projection process requires the use of the torch scatter operator. This operator has high requirements for the version of tensorRT during model deployment and cannot be deployed on devices with lower versions. Therefore, a native implementation of pytorch is used as an alternative, simplifying the scatter operator (scatter_max) and making the algorithm easier to deploy.

[0035] (3) PointNet point cloud encoder PointNet is a deep learning model particularly well-suited for processing unstructured point cloud data. It directly takes point cloud data as input and learns the local and global features of each point. PointNet learns features independently for each point and then generates a global descriptor through global feature aggregation operations (such as max pooling). This process preserves important spatial geometric information from the point cloud.

[0036] (4) Point cloud aggregation The voxelized point cloud data, i.e., the distance map, is merged with other features to form a richer input representation. The purpose of aggregation is to fuse information from different perspectives or perception methods so that subsequent feature extraction can obtain a more comprehensive environmental description. The aggregated point cloud data can include geometric features, texture features, and optical features extracted from images, etc.

[0037] (5) 2D feature extraction Based on the output of the YOLOv7Tiny model, further feature extraction is performed to capture more local and global information. The 2D feature extraction process can be understood as extracting detailed features of the point cloud in a two-dimensional plane. These features can be used to capture details lost during projection, while enhancing the semantic expressiveness of the point cloud.

[0038] (6) Scattering point feature extraction This step encodes the voxelized point cloud into the feature layer, then projects it into the 2D feature layer of the view volume expansion, and then convolves and encodes the 2D features to scatter back to the point features, completing the fusion of point features and projected graph features. The above steps extract useful features from the point cloud and can effectively convert between different representations, so that the machine learning model can better understand and classify point cloud data. This method combines the advantages of three-dimensional data processing and the powerful capabilities of two-dimensional image processing to achieve better performance.

[0039] (7) Fusion of point features Multi-modal feature fusion combines the three-dimensional features extracted by PointNet with 2D features. In this way, two-dimensional image features and three-dimensional geometric features can be combined to achieve a more comprehensive understanding of the scene. Feature fusion can be done through weighted averaging, concatenation, or adaptive learning. The fused features retain the spatial structure information of the point cloud and contain semantic information of the image, which helps to enhance the classification and segmentation capabilities of the model.

[0040] (8) Point class confidence calculation Classification prediction: Based on the fused features, point-level classification prediction is performed. Each point is predicted to belong to a certain class (such as vehicle, pedestrian, road, etc.), and the model generates a confidence score for each class prediction.

[0041] Confidence evaluation: Confidence represents the degree of trust that the model has in the classification of the point. A higher confidence means that the point is more likely to belong to that class. This step helps to remove low-confidence points in classification, thereby improving the accuracy of segmentation.

[0042] (9) Loss function For methods based on projected graphs, loss functions can be used in multiple ways. First, for the semantic class of each point in the final output, a cross-entropy loss function is used to calculate the loss for each point. For the projected graph, a semantic output with a resolution equal to the size of the projected graph is added to the network structure, and a cross-entropy loss is calculated between the true projected graph and the predicted projected graph. At the same time, the segmentation boundary of the projected graph also has a great influence on the final segmentation accuracy. This method innovatively introduces a boundary loss to calculate the boundary loss between the predicted projected graph and the true projected graph, making the boundaries between classes more accurate and effectively reducing the phenomenon of misclassification at the boundaries.

[0043] (10) ROI static environment semantics Region of Interest (ROI) extraction: Based on the point class confidence, the region of interest (ROI) is extracted, such as a specific object or scene area. ROI extraction can help focus on specific class objects, ignore irrelevant information in the background, and further analyze the extracted ROI area to obtain the class information of each point in the static environment. This allows the model to identify objects and their classes in the scene, forming a complete semantic understanding of the environment.

[0044] In this embodiment, the present application provides a laser point cloud semantic segmentation system for automatic driving, comprising a computer device programmed or configured to perform the steps of the laser point cloud semantic segmentation method for automatic driving.

[0045] In this embodiment, the present application provides a computer readable storage medium having stored thereon a computer program programmed or configured to perform the laser point cloud semantic segmentation method for automatic driving.

[0046] Any reference to memory, storage, databases, or other media in the embodiments herein can include non-volatile and / or volatile storage. Non-volatile storage can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile storage can include random-access memory (RAM), or external cache memory. By way of illustration, and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct Rambus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM). The computer program can be stored in any form of a computer program product, such as a computer program storage medium.

[0047] In summary, the present application not only uses cross-entropy loss at the point level and projection map level to ensure overall classification accuracy, but also innovatively introduces boundary loss, effectively enhancing the discriminability of class boundaries, significantly reducing misclassification of segmentation results at object edges, and thus improving the model's fine semantic understanding ability and robustness in complex static environments.

[0048] The above detailed description does not limit the scope of the disclosure. Various modifications, combinations, sub-combinations and alternatives can be made to the detailed description. Any modification, equivalent replacement and improvement etc. made within the spirit and principle of the disclosure shall be included in the scope of the disclosure.

Claims

1. A semantic segmentation method for laser point clouds used in autonomous driving, characterized in that, The method includes: M1. Laser point cloud acquisition and downsampling: Capture 3D point cloud data in the environment through a lidar sensor, and perform preprocessing and downsampling to remove redundant points and retain representative information points; M2. View frustum voxelization modeling: Through voxelization modeling, the 3D point cloud data is spatially divided so that the point cloud is represented in the form of voxels; M3.PointNet network model deep processing: Taking point cloud data as input, it learns the local and global features of each point. PointNet performs independent feature learning for each point, and then generates a global descriptor through global feature aggregation. M4. Point Cloud Aggregation and 2D Feature Extraction: The voxelized point cloud data, i.e. the distance map, is merged with other features to form aggregated point cloud data. Further feature extraction is performed based on the output of the yolov7Tiny model to capture more local and global information. M5. Scattering point feature extraction and fusion: The voxelized point cloud is encoded into the feature layer, then projected onto the 2D feature layer of the view frustum expansion, and then the 2D features are convolved and encoded to scatter back to the point features, thus completing the fusion of point features and projection map features, and fusing the 3D features extracted by PointNet with the 2D features. M6.ROI Static Environment Semantics: Based on the category confidence of points, regions of interest are extracted. ROIs extract objects of specific categories. Further semantic analysis is performed on the extracted ROI regions to obtain category information for each point in the static environment. This enables the model to recognize objects and their categories in the scene, forming a complete semantic understanding of the environment.

2. The laser point cloud semantic segmentation method for autonomous driving according to claim 1, characterized in that, In step M2, the voxelization model is a cylindrical voxelization model of the point cloud data using an auxiliary coordinate system, converting the point cloud data from a Cartesian coordinate system to a cylindrical coordinate system, that is, from coordinates (x, y, z) to (r, θ, Φ). The transformation function is: ， ， ， The point cloud data is then converted into a distance image using spherical projection.

3. The laser point cloud semantic segmentation method for autonomous driving according to claim 1, characterized in that: The aggregated point cloud data includes geometric and texture features, as well as optical features extracted from the image.

4. The laser point cloud semantic segmentation method for autonomous driving according to claim 1, characterized in that, In step M6, the category confidence of the point is calculated by performing point-level classification prediction based on the fused features. Each point will be predicted to be a certain category, and the model will generate a confidence score for each category prediction. The confidence score represents the degree of confidence of the model in the point classification. A high confidence score means that the point belongs to the same category.

5. The laser point cloud semantic segmentation method for autonomous driving according to claim 4, characterized in that: Based on the premise that each point will be predicted to belong to a certain category, a loss function is constructed. First, for the semantic category of each point, the cross-entropy loss function is used to calculate the loss for each point. For the projection map, a semantic output with a resolution of the projection map size is added when constructing the network structure. The cross-entropy loss is calculated for the ground truth projection map and the predicted projection map. At the same time, the segmentation boundary of the projection map also affects the final segmentation accuracy, thus introducing a boundary loss. The boundary loss between the predicted projection map and the ground truth projection map is calculated to represent the boundary between categories.

6. The laser point cloud semantic segmentation method for autonomous driving according to claim 1, characterized in that, The method also includes a point-level cross-entropy loss calculated between the predicted category and the real label for each 3D point; an image-level cross-entropy loss calculated between the predicted semantics and the real projected label for each pixel on the distance image generated by the view frustum projection; and a boundary loss calculated based on the semantic boundary region of the distance image to minimize the geometric or semantic difference between the predicted boundary and the real boundary.

7. The laser point cloud semantic segmentation method for autonomous driving according to claim 6, characterized in that: The boundary loss calculated for the semantic boundary region of the distance image is obtained by extracting a boundary mask from the ground truth projection map through edge detection operators or semantic discontinuity analysis, and applying a weighted loss penalty to the difference between the predicted projection map and the ground truth projection map within the boundary mask region, or by adopting a boundary-aware loss function.

8. The laser point cloud semantic segmentation method for autonomous driving according to claim 1, characterized in that, In step M1, the preprocessing includes rotating, translating, and stitching the point cloud data.

9. A laser point cloud semantic segmentation system for autonomous driving, comprising a computer device, characterized in that, The computer device is programmed or configured to perform the steps of the laser point cloud semantic segmentation method for autonomous driving as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is programmed or configured to perform the laser point cloud semantic segmentation method for autonomous driving as described in any one of claims 1 to 8.

Images