Three-dimensional point cloud semantic segmentation method and system oriented to automatic driving deployment

Abstract

The invention relates to a three-dimensional point cloud semantic segmentation method and system for automatic driving deployment, and the method comprises the steps: K1, collecting the point cloud data information of a road in real time, processing the point cloud of the road through employing an improved normal vector space downsampling algorithm, and obtaining the point cloud data information of the road after downsampling processing; and K2, based on the point cloud data information of the road after downsampling processing, carrying out scattering point feature extraction on the point cloud of the road by adopting an improved scattering point feature extraction algorithm of frustum primitization to obtain data information of scattering point feature extraction of the point cloud of the road. According to the method, data quality and preprocessing are carried out to remove noise, complement missing data and improve the quality of input data, and efficient point cloud semantic segmentation is realized by adopting a multi-modal feature fusion strategy.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a three-dimensional point cloud semantic segmentation method and system for autonomous driving deployment. Background Art

[0002] Laser point cloud semantic segmentation is an important branch of 3D data processing, which aims to classify 3D point cloud data obtained from devices such as LiDAR, that is, to assign a specific category label to each point. This technology has important application value in the fields of autonomous driving, robot navigation, geographic information system (GIS), etc.

[0003] In the prior art, point-based point cloud segmentation algorithms operate directly on 3D point clouds, usually considering the local geometric features of each point and the spatial relationship between them. The following are some characteristics of point-based point cloud segmentation algorithms: Local feature analysis: The algorithm calculates the local features around each point and classifies the points based on these features. For example, the RANSAC (Random Sampling Consensus) algorithm is a commonly used technique for fitting model parameters, which can be used to extract planes or other geometric shapes from noise.

[0004] Clustering methods: Use clustering techniques such as DBSCAN, Mean Shift, etc. to group points with similar characteristics. These methods usually rely on distance metrics to determine which points should belong to the same cluster.

[0005] Deep learning models: In recent years, with the development of deep learning, neural network architectures specifically for point cloud processing have emerged, such as PointNet and its variants, which can directly learn features and perform segmentation from point cloud data.

[0006] However, due to the huge amount of point cloud file data, the point-based point cloud semantic segmentation algorithm has a huge amount of calculation. Although the accuracy of the point-based point cloud semantic segmentation algorithm is slightly higher than that of the projection-based point cloud segmentation algorithm, the real-time performance of the algorithm cannot meet the usage requirements of autonomous driving scenarios. Summary of the invention

[0007] In view of the above shortcomings of the prior art, the present invention provides a three-dimensional point cloud semantic segmentation method and system for autonomous driving deployment, which not only improves data quality and preprocessing to remove noise, complete missing data, and improve the quality of input data, but also adopts a multimodal feature fusion strategy to achieve efficient point cloud semantic segmentation.

[0008] In order to achieve the above-mentioned purpose and other related purposes, the technical solution provided by the present invention is as follows: a three-dimensional point cloud semantic segmentation method for autonomous driving deployment, the method comprising:

[0009] K1. Collect point cloud data information of the road in real time, and use an improved normal vector space downsampling algorithm to process the point cloud of the road to obtain point cloud data information of the road after downsampling;

[0010] K2. Based on the point cloud data information of the road after the downsampling process, the improved scattering point feature extraction algorithm of the view cone voxelization is used to extract the scattering point features of the road point cloud to obtain the data information of the scattering point feature extraction of the road point cloud;

[0011] K3. Based on the point cloud data information of the road after the downsampling process, the point cloud global feature extraction algorithm of the Aurora optimized PointNet network is used to extract features from the point cloud of the road to obtain data information of global feature extraction of the road point cloud;

[0012] K4. Based on the data information of global feature extraction of the road point cloud and the data information of scattered point feature extraction of the road point cloud, the features of the point cloud are fused using a multimodal feature fusion algorithm based on Fourier fitting of a one-dimensional probability density function to obtain fused point cloud feature data information of the road;

[0013] K5. Based on the point cloud feature data information of the fused road, the point cloud is semantically segmented using the ROI static environment semantic algorithm that integrates the point category confidence to obtain the point cloud data information after semantic segmentation.

[0014] Furthermore, in step K1, the step of processing the point cloud of the road using the improved normal vector space downsampling algorithm includes:

[0015] K11. Based on the point cloud data information of the road, a normal vector function Q of the road point cloud is established.

[0016]

[0017] Among them, x i is the point cloud data information of the road, α is the differential factor of the point cloud of the road, the point cloud normal vector of the road is calculated, and the normal vector space of the road point cloud is constructed to obtain the data information of the normal vector space of the road point cloud;

[0018] K12. Based on the data information of the normal vector space of the road point cloud, a random sampling function W of the road point cloud is established.

[0019]

[0020] Among them, y is the data information of the normal vector space of the road point cloud, β1, β2 and β3 are the random sampling determining factors of the road point cloud;

[0021] K13. Based on the random sampling function W of the road point cloud, downsample the point cloud of the road to obtain point cloud data information of the road after downsampling.

[0022] Furthermore, the constraints of the random sampling determining factors β1, β2 and β3 of the road point cloud are:

[0023]

[0024] Furthermore, in step K2, the step of extracting scattered point features from the road point cloud using the improved scattering point feature extraction algorithm based on frustum voxelization includes:

[0025] K21. Based on the point cloud data information of the road after the downsampling process, a frustum voxel conversion function R of the road point cloud is established.

[0026]

[0027] Among them, z is the point cloud data information of the road after downsampling processing, δ1, δ2 and δ3 are the conversion factors of the road point cloud, and the point cloud of the road is converted into a frustum voxel to obtain the point cloud data information of the road after frustum voxelization;

[0028] K22. Projecting the point cloud data information of the road after the cone voxelization to the 2D feature layer of the cone expansion, extracting the 2D features of the road point cloud, and obtaining the 2D feature data information of the road point cloud;

[0029] K23. Based on the 2D feature data information of the road point cloud and the point cloud data information of the road after the view cone voxelization, construct a scattering point feature extraction function P of the road point cloud,

[0030]

[0031] Among them, a is the 2D feature data information of the road point cloud, b is the point cloud data information of the road after frustum voxelization, γ1, γ2 and γ3 are feature extraction factors of the road point cloud, and scattered point feature extraction is performed on the road point cloud to obtain the data information of the scattered point feature extraction of the road point cloud.

[0032] Furthermore, the feature extraction factors γ1, γ2 and γ3 of the road point cloud are:

[0033]

[0034] Among them, a is the 2D feature data information of the road point cloud, and b is the point cloud data information of the road after frustum voxelization.

[0035] Furthermore, in step K3, the point cloud global feature extraction algorithm using the Aurora optimized PointNet network is used to extract features from the point cloud of the road, including:

[0036] K31. Input the downsampled road point cloud data information into the PointNet network for training and learning, initialize the network parameters, and obtain the data information of the initialized network parameters;

[0037] K32. Based on the data information of the initialized network parameters, the aurora population is initialized, the population parameters and the maximum number of iterations of the population are determined, and the data information of the initialized aurora population is obtained;

[0038] K33. Based on the data information of the initialized aurora population, establish the fitness function S of the population individuals,

[0039]

[0040] Among them, c is the data information of the initialized aurora population, η1, η2 and η3 are the fitness determining factors of the aurora population, and the fitness values ​​of the individuals in the population are calculated to obtain the fitness value data information of the individuals in the population;

[0041] K34. Based on the fitness value data information of the population individuals, establish a target optimization function F,

[0042]

[0043] Among them, r is the fitness value data information of the population individual, λ1, λ2 and λ3 are the optimization factors of the population, and the initialized network parameters are optimized to obtain the optimized PointNet network;

[0044] K35. Based on the optimized PointNet network, the point cloud data information of the road after the downsampling process is input, and feature extraction is performed on the point cloud of the road to obtain data information of global feature extraction of the road point cloud.

[0045] Furthermore, in step K4, the multimodal feature fusion algorithm based on Fourier fitting one-dimensional probability density function is used to fuse the features of the point cloud, including:

[0046] K41. Based on the data information extracted from the global features of the road point cloud and the data information extracted from the scattered point features of the road point cloud, a Fourier fitting one-dimensional probability density function G of the road point cloud is established.

[0047]

[0048] Among them, h1 is the data information of global feature extraction of road point cloud, h2 is the data information of scattering point feature extraction of road point cloud, μ1, μ2 and μ3 are Fourier fitting factors of road point cloud, which characterize the probability density of road point cloud and obtain the data information of probability density of road point cloud;

[0049] K42. Based on the data information of the probability density of the road point cloud, a multimodal fusion function H of the road point cloud is established.

[0050]

[0051] Among them, g is the data information of the probability density of the road point cloud, θ1, θ2 and θ3 are the multimodal fusion factors of the road point cloud;

[0052] K43. Based on the multimodal fusion function H of the road point cloud, the features of the point cloud are fused to obtain point cloud feature data information of the fused road.

[0053] Furthermore, in step K5, the step of performing semantic segmentation on the point cloud using the ROI static environment semantic algorithm integrating point category confidence includes:

[0054] K51. Based on the fused road point cloud feature data information, establish a category confidence function L of the road point cloud,

[0055]

[0056] Among them, v is the point cloud feature data information of the fused road, ρ1, ρ2 and ρ3 are the weight coefficients of the road point cloud, which characterize the confidence of the road point cloud, and filter the road point cloud to remove the road point cloud with low confidence, and obtain the filtered road point cloud data information;

[0057] K52. Based on the point cloud data information of the filtered road, a ROI static environment semantic function J of the road point cloud is established.

[0058]

[0059] Among them, q is the point cloud data information of the screened road, ω1, ω2 and ω3 are the environmental semantic segmentation factors of the road point cloud;

[0060] K53. Based on the ROI static environment semantic function J of the road point cloud, perform semantic segmentation on the point cloud to obtain point cloud data information after semantic segmentation.

[0061] Furthermore, the constraint condition of the semantic segmentation factor of the road point cloud is:

[0062]

[0063] In order to achieve the above-mentioned objectives and other related objectives, the present invention also provides a three-dimensional point cloud semantic segmentation system for autonomous driving deployment, including a computer device, which is programmed or configured to execute any one of the steps of the three-dimensional point cloud semantic segmentation method for autonomous driving deployment.

[0064] The present invention has the following positive effects:

[0065] 1. The present invention processes the point cloud of the road by adopting an improved normal vector space downsampling algorithm, and extracts scattered point features from the point cloud of the road in combination with an improved scattering point feature extraction algorithm using frustum voxelization. It not only optimizes the network structure and algorithm deployment through voxel division, network backbone replacement, operator optimization, model quantization and other aspects, but also improves the running speed of the algorithm on the autonomous driving computing terminal and reduces the difficulty of algorithm deployment.

[0066] 2. The present invention extracts features from the point cloud of the road by using the point cloud global feature extraction algorithm of the Aurora optimized PointNet network, and combines the multimodal feature fusion algorithm based on Fourier fitting one-dimensional probability density function to fuse the features of the point cloud and the ROI static environment semantic algorithm based on the fusion point category confidence to perform semantic segmentation on the point cloud. This not only improves data quality and preprocessing to remove noise, complete missing data, and improve the quality of input data, but also the robustness of the model to ensure that the model can still work stably under different lighting and weather conditions, which is often achieved through data enhancement and multi-sensor fusion, and can greatly improve the practicability and reliability of the point cloud semantic segmentation model in autonomous driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0068] Figure 2 It is a schematic diagram of the flow chart of the improved normal vector space downsampling algorithm of the present invention;

[0069] Figure 3 It is a flow chart of the improved scattering point feature extraction algorithm for cone voxelization of the present invention;

[0070] Figure 4 It is a flow chart of the point cloud global feature extraction algorithm of the Aurora optimized PointNet network of the present invention;

[0071] Figure 5 It is a flow chart of a multimodal feature fusion algorithm based on Fourier fitting of a one-dimensional probability density function of the present invention;

[0072] Figure 6It is a flow chart of the ROI static environment semantic algorithm integrating point category confidence of the present invention. DETAILED DESCRIPTION

[0073] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0074] Example 1: Figure 1 As shown, a 3D point cloud semantic segmentation method for autonomous driving deployment includes:

[0075] K1. Collect point cloud data information of the road in real time, and use an improved normal vector space downsampling algorithm to process the point cloud of the road to obtain point cloud data information of the road after downsampling;

[0076] K2. Based on the point cloud data information of the road after the downsampling process, the improved scattering point feature extraction algorithm of the view cone voxelization is used to extract the scattering point features of the road point cloud to obtain the data information of the scattering point feature extraction of the road point cloud;

[0077] K3. Based on the point cloud data information of the road after the downsampling process, the point cloud global feature extraction algorithm of the Aurora optimized PointNet network is used to extract features from the point cloud of the road to obtain data information of global feature extraction of the road point cloud;

[0078] K4. Based on the data information of global feature extraction of the road point cloud and the data information of scattered point feature extraction of the road point cloud, the features of the point cloud are fused using a multimodal feature fusion algorithm based on Fourier fitting of a one-dimensional probability density function to obtain fused point cloud feature data information of the road;

[0079] K5. Based on the point cloud feature data information of the fused road, the point cloud is semantically segmented using the ROI static environment semantic algorithm that integrates the point category confidence to obtain the point cloud data information after semantic segmentation.

[0080] In this embodiment, if Figure 1 As shown, in step K1, the process of processing the point cloud of the road using the improved normal vector space downsampling algorithm includes:

[0081] K11. Based on the point cloud data information of the road, a normal vector function Q of the road point cloud is established.

[0082]

[0083] Among them, x i is the point cloud data information of the road, α is the differential factor of the point cloud of the road, the point cloud normal vector of the road is calculated, and the normal vector space of the road point cloud is constructed to obtain the data information of the normal vector space of the road point cloud;

[0084] K12. Based on the data information of the normal vector space of the road point cloud, a random sampling function W of the road point cloud is established.

[0085]

[0086] Among them, y is the data information of the normal vector space of the road point cloud, β1, β2 and β3 are the random sampling determining factors of the road point cloud;

[0087] K13. Based on the random sampling function W of the road point cloud, downsample the point cloud of the road to obtain point cloud data information of the road after downsampling.

[0088] In this embodiment, the constraints of the random sampling determining factors β1, β2 and β3 of the road point cloud are:

[0089]

[0090] In this embodiment, if Figure 3 As shown, in step K2, the improved scattering point feature extraction algorithm using frustum voxelization is used to extract scattering point features from the point cloud of the road, including:

[0091] K21. Based on the point cloud data information of the road after the downsampling process, a frustum voxel conversion function R of the road point cloud is established.

[0092]

[0093] Among them, z is the point cloud data information of the road after downsampling processing, δ1, δ2 and δ3 are the conversion factors of the road point cloud, and the point cloud of the road is converted into a frustum voxel to obtain the point cloud data information of the road after frustum voxelization;

[0094] K22. Projecting the point cloud data information of the road after the cone voxelization to the 2D feature layer of the cone expansion, extracting the 2D features of the road point cloud, and obtaining the 2D feature data information of the road point cloud;

[0095] K23. Based on the 2D feature data information of the road point cloud and the point cloud data information of the road after the view cone voxelization, construct a scattering point feature extraction function P of the road point cloud,

[0096]

[0097] Among them, a is the 2D feature data information of the road point cloud, b is the point cloud data information of the road after frustum voxelization, γ1, γ2 and γ3 are feature extraction factors of the road point cloud, and scattered point feature extraction is performed on the road point cloud to obtain the data information of the scattered point feature extraction of the road point cloud.

[0098] In this embodiment, the feature extraction factors γ1, γ2 and γ3 of the road point cloud are:

[0099] Among them, a is the 2D feature data information of the road point cloud, and b is the point cloud data information of the road after frustum voxelization.

[0100] Example 2: Based on the three-dimensional point cloud semantic segmentation method for autonomous driving deployment in Example 1, the present invention is further illustrated and described below.

[0101] like Figure 1 As shown, a 3D point cloud semantic segmentation method for autonomous driving deployment includes:

[0102] K1. Collect point cloud data information of the road in real time, and use an improved normal vector space downsampling algorithm to process the point cloud of the road to obtain point cloud data information of the road after downsampling;

[0103] K2. Based on the point cloud data information of the road after the downsampling process, the improved scattering point feature extraction algorithm of the view cone voxelization is used to extract the scattering point features of the road point cloud to obtain the data information of the scattering point feature extraction of the road point cloud;

[0104] K3. Based on the point cloud data information of the road after the downsampling process, the point cloud global feature extraction algorithm of the Aurora optimized PointNet network is used to extract features from the point cloud of the road to obtain data information of global feature extraction of the road point cloud;

[0105] K4. Based on the data information of global feature extraction of the road point cloud and the data information of scattered point feature extraction of the road point cloud, the features of the point cloud are fused using a multimodal feature fusion algorithm based on Fourier fitting of a one-dimensional probability density function to obtain fused point cloud feature data information of the road;

[0106] K5. Based on the point cloud feature data information of the fused road, the point cloud is semantically segmented using the ROI static environment semantic algorithm that integrates the point category confidence to obtain the point cloud data information after semantic segmentation.

[0107] In this embodiment, if Figure 4As shown, in step K3, the point cloud global feature extraction algorithm using the Aurora optimized PointNet network is used to extract features from the point cloud of the road, including:

[0108] K31. Input the downsampled road point cloud data information into the PointNet network for training and learning, initialize the network parameters, and obtain the data information of the initialized network parameters;

[0109] K32. Based on the data information of the initialized network parameters, the aurora population is initialized, the population parameters and the maximum number of iterations of the population are determined, and the data information of the initialized aurora population is obtained;

[0110] K33. Based on the data information of the initialized aurora population, establish the fitness function S of the population individuals,

[0111]

[0112] Among them, c is the data information of the initialized aurora population, η1, η2 and η3 are the fitness determining factors of the aurora population, and the fitness values ​​of the individuals in the population are calculated to obtain the fitness value data information of the individuals in the population;

[0113] K34. Based on the fitness value data information of the population individuals, establish a target optimization function F,

[0114]

[0115] Among them, r is the fitness value data information of the population individual, λ1, λ2 and λ3 are the optimization factors of the population, and the initialized network parameters are optimized to obtain the optimized PointNet network;

[0116] K35. Based on the optimized PointNet network, the point cloud data information of the road after the downsampling process is input, and feature extraction is performed on the point cloud of the road to obtain data information of global feature extraction of the road point cloud.

[0117] In this embodiment, if Figure 5 As shown, in step K4, the multimodal feature fusion algorithm based on Fourier fitting one-dimensional probability density function is used to fuse the features of the point cloud, including:

[0118] K41. Based on the data information extracted from the global features of the road point cloud and the data information extracted from the scattered point features of the road point cloud, a Fourier fitting one-dimensional probability density function G of the road point cloud is established.

[0119]

[0120] Among them, h1 is the data information of global feature extraction of road point cloud, h2 is the data information of scattering point feature extraction of road point cloud, μ1, μ2 and μ3 are Fourier fitting factors of road point cloud, which characterize the probability density of road point cloud and obtain the data information of probability density of road point cloud;

[0121] K42. Based on the data information of the probability density of the road point cloud, a multimodal fusion function H of the road point cloud is established.

[0122]

[0123] Among them, g is the data information of the probability density of the road point cloud, θ1, θ2 and θ3 are the multimodal fusion factors of the road point cloud;

[0124] K43. Based on the multimodal fusion function H of the road point cloud, the features of the point cloud are fused to obtain point cloud feature data information of the fused road.

[0125] In this embodiment, if Figure 6 As shown, in step K5, the use of the ROI static environment semantic algorithm fused with point category confidence to perform semantic segmentation on the point cloud includes:

[0126] K51. Based on the fused road point cloud feature data information, establish a category confidence function L of the road point cloud,

[0127]

[0128] Among them, v is the point cloud feature data information of the fused road, ρ1, ρ2 and ρ3 are the weight coefficients of the road point cloud, which characterize the confidence of the road point cloud, and filter the road point cloud to remove the road point cloud with low confidence, and obtain the filtered road point cloud data information;

[0129] K52. Based on the point cloud data information of the filtered road, a ROI static environment semantic function J of the road point cloud is established.

[0130]

[0131] Among them, q is the point cloud data information of the screened road, ω1, ω2 and ω3 are the environmental semantic segmentation factors of the road point cloud;

[0132] K53. Based on the ROI static environment semantic function J of the road point cloud, perform semantic segmentation on the point cloud to obtain point cloud data information after semantic segmentation.

[0133] In this embodiment, the constraint condition of the semantic segmentation factor of the road point cloud is:

[0134]

[0135] In this embodiment, the present invention provides a three-dimensional point cloud semantic segmentation system for autonomous driving deployment, including a computer device that is programmed or configured to execute any one of the steps of the three-dimensional point cloud semantic segmentation method for autonomous driving deployment.

[0136] The present invention provides a computer-readable storage medium, on which is stored a computer program programmed or configured to execute any one of the three-dimensional point cloud semantic segmentation methods for autonomous driving deployment.

[0137] Any reference to memory, storage, database or other media used in the embodiments provided in the present application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM).

[0138] In summary, the present invention not only improves data quality and preprocessing to remove noise, complete missing data, and improve the quality of input data, but also adopts a multimodal feature fusion strategy to achieve efficient point cloud semantic segmentation.

[0139] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A 3D point cloud semantic segmentation method for autonomous driving deployment, characterized in that: The method comprises: K1. Collect point cloud data information of the road in real time, and use an improved normal vector space downsampling algorithm to process the point cloud of the road to obtain point cloud data information of the road after downsampling; K2. Based on the point cloud data information of the road after the downsampling process, the improved scattering point feature extraction algorithm of the view cone voxelization is used to extract the scattering point features of the road point cloud to obtain the data information of the scattering point feature extraction of the road point cloud; K3. Based on the point cloud data information of the road after the downsampling process, the point cloud global feature extraction algorithm of the Aurora optimized PointNet network is used to extract features from the point cloud of the road to obtain data information of global feature extraction of the road point cloud; K4. Based on the data information of global feature extraction of the road point cloud and the data information of scattered point feature extraction of the road point cloud, the features of the point cloud are fused using a multimodal feature fusion algorithm based on Fourier fitting of a one-dimensional probability density function to obtain fused point cloud feature data information of the road; K5. Based on the point cloud feature data information of the fused road, the point cloud is semantically segmented using the ROI static environment semantic algorithm that integrates the point category confidence to obtain the point cloud data information after semantic segmentation.

2. The three-dimensional point cloud semantic segmentation method for autonomous driving deployment according to claim 1, characterized in that: In step K1, the process of processing the point cloud of the road using the improved normal vector space downsampling algorithm includes: K11. Based on the point cloud data information of the road, a normal vector function Q of the road point cloud is established. Among them, x i is the point cloud data information of the road, α is the differential factor of the point cloud of the road, the point cloud normal vector of the road is calculated, and the normal vector space of the road point cloud is constructed to obtain the data information of the normal vector space of the road point cloud; K12. Based on the data information of the normal vector space of the road point cloud, a random sampling function W of the road point cloud is established. Among them, y is the data information of the normal vector space of the road point cloud, β1, β2 and β3 are the random sampling determining factors of the road point cloud; K13. Based on the random sampling function W of the road point cloud, down-sample the point cloud of the road to obtain point cloud data information of the road after down-sampling.

3. The 3D point cloud semantic segmentation method for autonomous driving deployment according to claim 2, characterized in that: The constraints of the random sampling determining factors β1, β2 and β3 of the road point cloud are:

4. The three-dimensional point cloud semantic segmentation method for autonomous driving deployment according to claim 1, characterized in that: In step K2, the step of extracting scattered point features from the road point cloud using the improved scattering point feature extraction algorithm based on frustum voxelization includes: K21. Based on the point cloud data information of the road after the downsampling process, a frustum voxel conversion function R of the road point cloud is established. Among them, z is the point cloud data information of the road after downsampling processing, δ1, δ2 and δ3 are the conversion factors of the road point cloud, and the point cloud of the road is converted into a frustum voxel to obtain the point cloud data information of the road after frustum voxelization; K22. Projecting the point cloud data information of the road after the cone voxelization to the 2D feature layer of the cone expansion, extracting the 2D features of the road point cloud, and obtaining the 2D feature data information of the road point cloud; K23. Based on the 2D feature data information of the road point cloud and the point cloud data information of the road after the view cone voxelization, construct a scattering point feature extraction function P of the road point cloud, Among them, a is the 2D feature data information of the road point cloud, b is the point cloud data information of the road after frustum voxelization, γ1, γ2 and γ3 are feature extraction factors of the road point cloud, and scattered point feature extraction is performed on the road point cloud to obtain the data information of the scattered point feature extraction of the road point cloud.

5. The three-dimensional point cloud semantic segmentation method for autonomous driving deployment according to claim 4, characterized in that: The feature extraction factors γ1, γ2 and γ3 of the road point cloud are: Among them, a is the 2D feature data information of the road point cloud, and b is the point cloud data information of the road after frustum voxelization.

6. The three-dimensional point cloud semantic segmentation method for autonomous driving deployment according to claim 1, characterized in that: In step K3, the point cloud global feature extraction algorithm using the Aurora optimized PointNet network is used to extract features from the point cloud of the road, including: K31. Input the downsampled road point cloud data information into the PointNet network for training and learning, initialize the network parameters, and obtain the data information of the initialized network parameters; K32. Based on the data information of the initialized network parameters, the aurora population is initialized, the population parameters and the maximum number of iterations of the population are determined, and the data information of the initialized aurora population is obtained; K33. Based on the data information of the initialized aurora population, establish the fitness function S of the population individuals, Among them, c is the data information of the initialized aurora population, η1, η2 and η3 are the fitness determining factors of the aurora population, and the fitness values ​​of the individuals in the population are calculated to obtain the fitness value data information of the individuals in the population; K34. Based on the fitness value data information of the population individuals, establish a target optimization function F, Among them, r is the fitness value data information of the population individual, λ1, λ2 and λ3 are the optimization factors of the population, and the initialized network parameters are optimized to obtain the optimized PointNet network; K35. Based on the optimized PointNet network, the point cloud data information of the road after the downsampling process is input, and feature extraction is performed on the point cloud of the road to obtain data information of global feature extraction of the road point cloud.

7. The three-dimensional point cloud semantic segmentation method for autonomous driving deployment according to claim 1, characterized in that: In step K4, the multimodal feature fusion algorithm based on Fourier fitting one-dimensional probability density function is used to fuse the features of the point cloud, including: K41. Based on the data information extracted from the global features of the road point cloud and the data information extracted from the scattered point features of the road point cloud, a Fourier fitting one-dimensional probability density function G of the road point cloud is established. Among them, h1 is the data information of global feature extraction of road point cloud, h2 is the data information of scattering point feature extraction of road point cloud, μ1, μ2 and μ3 are Fourier fitting factors of road point cloud, which characterize the probability density of road point cloud and obtain the data information of probability density of road point cloud; K42. Based on the data information of the probability density of the road point cloud, a multimodal fusion function H of the road point cloud is established. Among them, g is the data information of the probability density of the road point cloud, θ1, θ2 and θ3 are the multimodal fusion factors of the road point cloud; K43. Based on the multimodal fusion function H of the road point cloud, the features of the point cloud are fused to obtain point cloud feature data information of the fused road.

8. The three-dimensional point cloud semantic segmentation method for autonomous driving deployment according to claim 1, characterized in that: In step K5, the step of performing semantic segmentation on the point cloud using the ROI static environment semantic algorithm integrating point category confidence includes: K51. Based on the fused road point cloud feature data information, establish a category confidence function L of the road point cloud, Among them, v is the point cloud feature data information of the fused road, ρ1, ρ2 and ρ3 are the weight coefficients of the road point cloud, which characterize the confidence of the road point cloud, and filter the road point cloud to remove the road point cloud with low confidence, and obtain the filtered road point cloud data information; K52. Based on the point cloud data information of the filtered road, a ROI static environment semantic function J of the road point cloud is established. Among them, q is the point cloud data information of the screened road, ω1, ω2 and ω3 are the environmental semantic segmentation factors of the road point cloud; K53. Based on the ROI static environment semantic function J of the road point cloud, perform semantic segmentation on the point cloud to obtain point cloud data information after semantic segmentation.

9. The three-dimensional point cloud semantic segmentation method for autonomous driving deployment according to claim 8, characterized in that: The constraint condition of the semantic segmentation factor of the road point cloud is:

10. A 3D point cloud semantic segmentation system for autonomous driving deployment, comprising a computer device, characterized in that: The computer device is programmed or configured to execute the steps of the three-dimensional point cloud semantic segmentation method for autonomous driving deployment as described in any one of claims 1 to 9.

Images