Grid-level semantic segmentation method and device based on laser radar point cloud, and medium

Abstract

The application relates to a grid-level semantic segmentation method based on a laser radar point cloud, which comprises the following steps: acquiring point cloud original data through the laser radar; extracting multiple features from the point cloud original data, including extracting voxel features, extracting point-level features and extracting look-around features; performing feature fusion on the extracted multiple features to obtain a fused feature map; performing grid-level feature coding on the fused feature map to obtain a grid feature map; and inputting the grid feature map into a semantic segmentation network for semantic prediction to obtain a grid-level semantic segmentation result. The application can improve the recognition accuracy of semantic segmentation and realize the recognition of more dynamic and static elements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving environment perception, and in particular to a grid-level semantic segmentation method based on laser radar point cloud, a device and a computer readable storage medium. BACKGROUND

[0002] In recent years, with the gradual deepening of the research on automatic driving, the application scenarios of automatic driving are becoming more and more rich, and showing a trend of expanding from closed road scenarios to open road scenarios. Under this background, the automatic driving system should have the ability to perceive the dynamic and static elements in the driving environment in a fine-grained manner.

[0003] Most of the classic automatic driving environment perception systems use target detection to obtain scene semantic information. Such methods are mostly composed of end-to-end deep neural networks, such as VoxelNet, PointPillars, SECOND, PointRCNN, PV-RCNN, etc. By using laser radar point cloud data as output, the end-to-end prediction of the spatial position, size, orientation, semantic category and other information of obstacles is realized. However, due to the characteristics of target detection, such methods can only detect obstacles with specific regular shapes, such as vehicles, pedestrians and cyclists.

[0004] In recent years, some works have proposed grid-level semantic segmentation methods, which are expected to achieve more fine-grained perception. Such methods use the method of encoding raw point cloud data into a top-down grid feature map to estimate the semantic category of each grid in an end-to-end manner, such as MotionNet, BE-STI, etc. However, the existing grid-level semantic segmentation methods have two problems: 1) Directly projecting point cloud data onto the expression of a top-down grid, the encoding ability for scene spatial structure information is weak, resulting in low segmentation accuracy; 2) The labels of the existing grid-level semantic segmentation are mostly from the annotation of bounding boxes, so they can only perform semantic segmentation on a small number of dynamic elements, and lack the identification of a large number of static elements. SUMMARY

[0005] To solve the above problems, the purpose of the present application is to provide a grid-level semantic segmentation method based on laser radar point cloud, a device and a computer readable storage medium, which can improve the identification accuracy of semantic segmentation and realize the identification of more dynamic and static elements.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a grid-level semantic segmentation method based on laser radar point cloud, which comprises:

[0008] acquiring point cloud raw data through the laser radar;

[0009] extracting a plurality of features from the point cloud raw data, including extracting voxel features, extracting point-level features, and extracting surround-view features;

[0010] performing feature fusion on the extracted plurality of features to obtain a fused feature map;

[0011] performing grid-level feature encoding on the fused feature map to obtain a grid feature map;

[0012] inputting the grid feature map into a semantic segmentation network for semantic prediction to obtain a grid-level semantic segmentation result.

[0013] In an implementation manner of the present application, the method further includes the step of pre-training the semantic segmentation network; and the target identified by the semantic segmentation network includes a dynamic target and a static target.

[0014] In an implementation manner of the present application, the point cloud raw data obtained by the lidar includes wherein p i = [x i , y i , z i ] T represents a point in the point cloud raw data represented by a three-dimensional space coordinate, and N represents the total number of points in the point cloud raw data.

[0015] In an implementation manner of the present application, the extracting of the voxel features includes:

[0016] performing voxelization operation on each point cloud raw data to obtain a corresponding one-dimensional vector voxel feature, and performing voxelization operation on all point cloud raw data to obtain a corresponding four-dimensional feature tensor;

[0017] encoding the four-dimensional feature tensor by using a sparse voxel feature encoding network to obtain an encoded voxel feature map as the voxel feature.

[0018] In an implementation manner of the present application, the extracting of the point-level features includes:

[0019] encoding the point cloud raw data by using a multi-layer perception (MLP) to obtain an encoded point-level feature map.

[0020] In an implementation manner of the present application, the extracting of the surround-view features includes:

[0021] The original point cloud data is projected to a surround view plane by using surround view projection; then a U-Net network structure is used for surround view image feature coding; and then a two-dimensional convolution operator is used for upsampling fusion to obtain a voxelized feature-coded surround view feature map.

[0022] In an implementation form of the present application, the feature fusion comprises: taking the voxel feature map, the point-level feature map and the surround view feature map as inputs, finding the voxel and the surround view image index corresponding to each point according to the projection mapping relationship between the original point cloud and the voxel and the surround view image, then taking out the voxel feature vectors and the image feature vectors at the corresponding positions and concatenating them to the point-level feature vectors corresponding to the points to obtain a point-level fusion feature map fused with the three kinds of feature expressions.

[0023] In an implementation form of the present application, the inputting of the grid feature map to the semantic segmentation network for semantic prediction to obtain a grid-level semantic segmentation result comprises:

[0024] The preset number of dynamic and static targets are sorted according to the set priority, and the grid-level semantic segmentation result is identified according to the order of priority.

[0025] In a second aspect, the present application provides a grid-level semantic segmentation device based on laser radar point cloud, the device comprises:

[0026] A data acquisition module is configured to acquire point cloud original data.

[0027] A feature extraction module is configured to extract multiple features from the point cloud original data, including extracting voxel features, extracting point-level features and extracting surround view features.

[0028] A feature fusion module is configured to fuse the extracted multiple features to obtain a fused feature map.

[0029] A feature coding module is configured to perform grid-level feature coding on the fused feature map to obtain a grid feature map.

[0030] A semantic segmentation module is configured to input the grid feature map to a semantic segmentation network for semantic prediction to obtain a grid-level semantic segmentation result.

[0031] In a third aspect, the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program runs to control the device where the computer readable storage medium executes the grid-level semantic segmentation method based on laser radar point cloud of the first aspect.

[0032] The present application has the following advantages: the present application obtains point cloud raw data through a laser radar, extracts various features from the point cloud raw data, fuses the various features, obtains a fused feature map, encodes the fused feature map at a grid level to obtain a grid feature map, performs semantic prediction based on the grid feature map, and obtains a grid-level semantic segmentation result, thereby improving the recognition accuracy of semantic segmentation and realizing identification of more types of dynamic and static elements. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a flow diagram of a grid-level semantic segmentation method based on laser radar point clouds provided by an embodiment of the present application;

[0034] Figure 2 is a data flow diagram of a grid-level semantic segmentation method based on laser radar point clouds provided by an embodiment of the present application. DETAILED DESCRIPTION

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are part of, rather than all of, the embodiments of the present application. Based on the described embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art belong to the scope of protection of the present application.

[0036] To solve the problems of low semantic segmentation accuracy and limited target recognition types in the prior art, the technical solutions of the present application correspondingly provide a grid-level semantic segmentation method based on laser radar point clouds, a device, and a computer-readable storage medium. The method comprises: obtaining point cloud raw data through a laser radar; extracting various features from the point cloud raw data, including extracting voxel features, extracting point-level features, and extracting surround-view features; fusing the various extracted features to obtain a fused feature map; encoding the fused feature map at a grid level to obtain a grid feature map; inputting the grid feature map into a semantic segmentation network to perform semantic prediction and obtain a grid-level semantic segmentation result. The present application can improve the recognition accuracy of semantic segmentation and realize identification of more types of dynamic and static elements.

[0037] For more details of the embodiments of the present application, please refer to the more drawings of the embodiments of the present application, which further illustrate the method, device, and medium provided by the present application.

[0038] As Figure 1 In one aspect of the embodiments of the present application, a grid-level semantic segmentation method based on laser radar point clouds is provided, which specifically comprises:

[0039] S01, acquiring point cloud raw data through the laser radar;

[0040] S02, extracting multiple features from the point cloud raw data, including extracting voxel features, extracting point level features, and extracting surround view features;

[0041] S03, performing feature fusion on the extracted multiple features to obtain a fused feature map;

[0042] S04, performing grid level feature coding on the fused feature map to obtain a grid feature map;

[0043] S05, inputting the grid feature map into a semantic segmentation network for semantic prediction to obtain a grid level semantic segmentation result.

[0044] The following will be combined with the data flow in Figure 2 In a more detailed embodiment of the present application, the method provided by the present application is described.

[0045] The method comprises:

[0046] S01, acquiring point cloud raw data through the laser radar;

[0047] Specifically, the point cloud raw data acquired by the laser radar is denoted as Where p i =[x i , y i , z i ] T represents a point in the point cloud, represented by its three-dimensional spatial coordinates. N represents the total number of points in the point cloud. Therefore, the raw data is It can be represented as a two-dimensional matrix with a shape of [N, 3].

[0048] S02, extracting multiple features from the point cloud raw data, including extracting voxel features, extracting point level features, and extracting surround view features;

[0049] Specifically, the raw point cloud data acquired by the laser radar is firstly organized in the manner of voxel, point level and surround view image respectively, and then is encoded by voxel feature encoder, point level feature encoder and surround view feature encoder respectively. The voxel feature encoder uses the point cloud data organized in the manner of voxel as input, uses a three-dimensional sparse convolution layer stacked in multiple layers for feature encoding, and outputs a feature tensor at the voxel level; the point level feature encoder uses the point cloud data organized in the manner of point level two-dimensional list as input, uses a multi-layer perceptron stacked in multiple layers for feature encoding, and outputs a feature vector at the point level; the surround view feature encoder uses the surround view image projected from the point cloud spherical coordinate as input, uses a two-dimensional convolution layer stacked in multiple layers for feature encoding, and outputs a two-dimensional surround view feature tensor in the spherical coordinate system.

[0050] The operation process of voxel feature encoding is as follows:

[0051] For the above point cloud data Let the maximum perception range in the three-dimensional space be [L0, W0, H0], and set the three-dimensional grid (voxel) size as [a x , a y , a z ], and let Then, the voxel set is obtained after voxel division of the three-dimensional space in the perception range:

[0052]

[0053] Each voxel v i，j，k is represented by the average value of the coordinates of the points located therein, that is:

[0054]

[0055] Where n i，j，k represents the number of points located in the voxel v i，j，k .

[0056] After the above voxelization operation, each voxel is represented by a one-dimensional vector with a length of 3, and it is noted that there are [LxWxH] voxels in the three-dimensional space in total. Therefore, after voxelization operation, the point cloud data of the whole scene can be represented as a four-dimensional feature tensor

[0057] Subsequently, the voxel data is encoded by a sparse voxel feature encoding network (SpVFE) to obtain the feature tensor

[0058] First, a sparse 3D convolution operator with a stride of 2 and a receptive field of 3 is used to perform three convolution downsampling operations on the voxel data V, resulting in three images of size V. Three feature maps

[0059] Then, an upsampling operator U with a stacking step size of 2 is used. s (·) and the sparse 3D convolution operator C with a stride of 1 and a receptive field of 3. s3 (·) For the above feature maps [V, V] s2 V s4 V s8 Upsampling and fusion are performed to obtain a unified feature map after voxelized feature encoding. The specific operation process is as follows:

[0060] F V4 =C s3 (V s4 U s (V s8 ))

[0061] F V2 =C s3 (V s2 U s (F V4 ))

[0062] F V =C s3 (V, U) s (F V2 ))

[0063] The operation process of a point-level feature encoder is as follows:

[0064] For raw point cloud data The point cloud data was encoded sequentially using multilayer perceptrons (MLPs) with output sizes of 64, 128, and 256 to obtain encoded point-level feature maps. The specific operation process is as follows:

[0065] F P64 =MLP 64 (P)

[0066] F P128 =MLP 128 (F P64 )

[0067] F P =MLP 256 (F P128 )

[0068] The operation process of the look-around feature encoder is as follows:

[0069] First, a panoramic projection is used to visualize the original point cloud data. Projecting the image onto the panoramic view plane yields a panoramic image. Where H r W r C r0 These represent the height, width, and number of channels of the panoramic image, respectively. For a given point p... i =[x i y i , z i ] T Through mapping ∏: p i →(u i v i ) T Project it onto the coordinates of the toroidal view image plane (u i v i ) T Projection mapping process ∏: p i →(u i v i ) T as follows:

[0070]

[0071] in, f = f up +f down This indicates the longitudinal scanning range of the lidar sensor. Then, point p... i Place the coordinates of the image plane on the (u) plane. i v i ) T In terms of location, obtain the corresponding panoramic image. Where C r0 =3.

[0072] For the panoramic image The U-Net network structure is used for feature encoding of the surrounding image. The specific process is as follows: First, a convolutional layer with a stride of 1, a receptive field of 3, and 32 channels is used to encode the features of the surrounding image I, resulting in a [H] dimension. r W r Feature map of

[32] Then, convolutional layers with stride of 2, receptive field of 3, and number of channels of 64, 128, and 256 were successively applied to I. s0 Perform feature encoding to obtain features of size I0. Multiple downsampled feature maps

[0073] Then, an upsampling operator U with a stacking step size of 2 is used.s (·) and a two-dimensional convolution operator C with a step size of 1 and a receptive field of 3 2d (·) up-sampling fusion is performed on the above feature maps [I s0 , I s2 , I s4 , I s8 ] to finally obtain a unified feature map after voxelization feature encoding The specific operation process is as follows:

[0074] F I128 = C 2d (I s4 , U s (I s8 ))

[0075] F I64 = C 2d (I s2 , U s (F I128 ))

[0076] F I = C 2d (I s0 , U s (F I64 ))

[0077] S03, the extracted multiple features are fused to obtain a fused feature map;

[0078] Specifically, the multi-view feature fusion module takes the voxel feature tensor, the point-level feature vector and the surround view feature tensor output by the previous stage as input, finds the corresponding voxel and surround view image index of each point according to the projection mapping relationship between the original point cloud and the voxel and the surround view image, and then takes out the voxel feature vector and the image feature vector at the corresponding position. The point-level fusion feature map fused with the three kinds of feature expressions is obtained by splicing the point-level feature vector corresponding to the point.

[0079] More specifically, for a given point p i = [x i , y i , z i ] T ∈ P, first find its corresponding feature vector f P = F P (p i ) in the point-level feature map , then find its corresponding voxel v ∈ V according to the voxelization process, and then index to the corresponding voxel feature vector f V = F V (u(p i )), and then according to the surround projection process∏: pi →(u i , v i ) T Find the corresponding look-around feature vector f I = F I (∏(p i ))。

[0080] On this basis, the above multi-view feature vectors are spliced together to obtain the spliced feature vector f i of the point p i =[p i , f P , f V , f I ], and then the fusion feature vectors of all points are spliced together to obtain the spliced feature map of the point cloud

[0081] S04, the raster-level feature encoding is performed on the fused feature map to obtain a raster feature map;

[0082] Specifically, the raster-level feature encoding module uses the point-level fusion feature map as input, divides the full space into overhead view rasters with a size of 0.25m*0.25m within the coverage range of the point cloud, and each raster contains different amounts of point clouds. For the point clouds inside each raster, a stacked multi-layer perception machine is used to further encode the point-level fusion feature map of the point clouds, and finally a unified one-dimensional feature vector is obtained through a maximum pooling operation, which is used as the feature expression of the raster. Arranging the feature vectors of all rasters under the overhead view together, a two-dimensional raster-level feature map under the overhead view is obtained.

[0083] More specifically, the range [L0, W0] of the overhead view raster, the two-dimensional raster size [a x , a y ] are set, and the space is divided into two-dimensional rasters under the overhead view of the relative vehicle coordinate system According to the coordinate information under the overhead view plane contained in the two-dimensional features of the feature map F C Determine the feature vector corresponding to the point located in each raster, and the feature of the raster is represented by the average of the feature vectors of these points:

[0084]

[0085] Finally, a two-dimensional raster feature map

[0086] S05, input the raster feature map into a semantic segmentation network for semantic prediction to obtain a raster-level semantic segmentation result. ​

[0087] In order to realize the function of semantic segmentation, it is necessary to generate a grid-level semantic segmentation label in advance according to the point-level semantic true value label for effective training of the entire network. In the process of gridding, it is inevitable that multiple points with different semantic categories fall inside the same grid. The determination rule of the grid semantics in this case is as follows: 1) The point cloud inside the grid is sorted according to the category priority shown in Table 1 below. If there is a high-priority category, all low-priority semantic points are ignored; 2) If there are multiple categories of the same level semantics, the semantic with the most points is taken as the semantic label of the grid.

[0088]

[0089] Table 1

[0090] The grid-level semantic segmentation head takes the two-dimensional grid-level feature map generated in the previous stage as input, and performs semantic information regression on the feature map through stacked multi-layer two-dimensional convolution operations, and finally outputs the grid-level semantic category information, thereby realizing the output of the grid-level semantic segmentation result.

[0091] More specifically, the grid-level semantic segmentation head is composed of two stacked two-dimensional convolution layers C 2d with a step size of 1, a receptive field of 3, and channel numbers of 64 and 23 respectively, and finally outputs a semantic probability distribution map F pred with a size of [L0xW0x23].

[0092] F pred = C 2d (C 2d (V BEV )

[0093] Finally, the argmax operation is performed on F pred to obtain the final semantic prediction result S pred .

[0094] S pred = argmax(F pred )

[0095] Another aspect of the embodiments of the present application also provides a corresponding device, comprising:

[0096] A data acquisition module for acquiring point cloud raw data;

[0097] A feature extraction module for extracting multiple features from the point cloud raw data, including extracting voxel features, extracting point-level features, and extracting look-around features;

[0098] A feature fusion module for fusing the extracted multiple features to obtain a fused feature map;

[0099] a feature encoding module, configured to perform grid-level feature encoding on the fused feature map to obtain a grid feature map;

[0100] a semantic segmentation module, configured to input the grid feature map into a semantic segmentation network to perform semantic prediction and obtain a grid-level semantic segmentation result.

[0101] In another aspect of the embodiments of the present application, a computer storage medium is also provided.

[0102] The embodiments of the present application also provide a computer readable storage medium, which includes a stored program. When the program is run, the computer readable storage medium controls a device where the computer readable storage medium is located to perform the method described above. The specific implementation process is not repeated here.

[0103] The embodiments of the present application also provide a computer device. The computer device of the embodiments includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the foregoing method in the embodiments is implemented. To avoid repetition, details are not repeated here. Alternatively, when the computer program is executed by the processor, the functions of each model / unit in the apparatus in the embodiments are implemented. To avoid repetition, details are not repeated here.

[0104] The computer device can be a desktop computer, a notebook, a palm computer, a server, and a cloud server, etc. The computer device can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that more or fewer components than those shown can be included, or some components can be combined or different components can be included, for example, the computer device can also include an input / output device, a network access device, a bus, etc.

[0105] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0106] The memory can be an internal storage unit of the computer device, for example, a hard disk or a memory of the computer device. The memory can also be an external storage device of the computer device, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card and the like equipped on the computer device. Further, the memory can include both the internal storage unit and the external storage device of the computer device. The memory is used to store the computer program and other programs and data required by the computer device. The memory can also be used to temporarily store the data that has been output or will be output.

[0107] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0108] In several embodiments provided by the present application, it should be understood that the disclosed system, device and method can be implemented by other manners. For example, the above-described device embodiments are only schematic, for example, the division of the above-described units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0109] The integrated unit in the form of the software function unit described above can be stored in a computer readable storage medium. The software function unit described above is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute part of the steps of the method described above in the embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk and various program code storage media.

[0110] The above is only a preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A grid-level semantic segmentation method based on lidar point clouds, characterized in that, The method includes: The raw point cloud data is acquired using the lidar. Multiple features are extracted from the raw point cloud data, including voxel features, point-level features, and look-around features. The extracted features are fused to obtain a fused feature map; The fused feature map is subjected to raster-level feature encoding to obtain a raster feature map; The raster feature map is input into a semantic segmentation network for semantic prediction to obtain raster-level semantic segmentation results. The extracted voxel features include: Voxelization is performed on each point cloud raw data to obtain the voxel features of the corresponding one-dimensional vector. After voxelization of all point cloud raw data, the corresponding four-dimensional feature tensor is obtained. The four-dimensional feature tensor is encoded using a sparse voxel feature encoding network to obtain the encoded voxel feature map as the voxel feature; The feature fusion includes: taking voxel feature map, point-level feature map and surround view feature map as input, finding the voxel and surround view image index corresponding to each point according to the projection mapping relationship between the original point cloud and voxels and surround view images, and then extracting the voxel feature vector and image feature vector at the corresponding position and concatenating them to the point-level feature vector corresponding to that point to obtain a point-level fused feature map that fuses the three feature expressions.

2. The grid-level semantic segmentation method based on lidar point clouds according to claim 1, characterized in that, The method further includes: a step of pre-training the semantic segmentation network; the targets identified by the semantic segmentation network include dynamic targets and static targets.

3. The grid-level semantic segmentation method based on lidar point clouds according to claim 2, characterized in that, The raw point cloud data acquired by the lidar is ,in This represents a point in the original point cloud data, represented by its three-dimensional spatial coordinates. This represents the total number of points in the original point cloud data.

4. The grid-level semantic segmentation method based on lidar point clouds according to claim 1, characterized in that, The extracted point-level features include: For the original point cloud data, a multilayer perceptron (MLP) is used for encoding to obtain the encoded point-level feature map.

5. The grid-level semantic segmentation method based on lidar point clouds according to claim 4, characterized in that, The extraction of surround view features includes: The original point cloud data is projected onto the surrounding view plane using surround projection; then the U-Net network structure is used to encode the surrounding image features; and finally, a two-dimensional convolution operator is used for upsampling and fusion to obtain the surrounding feature map after voxelization feature encoding.

6. The grid-level semantic segmentation method based on lidar point clouds according to claim 1, characterized in that, The step of inputting the raster feature map into a semantic segmentation network for semantic prediction to obtain raster-level semantic segmentation results includes: A preset number of dynamic and static targets are sorted according to a set priority, and the raster-level semantic segmentation results are identified according to the priority order.

7. A grid-level semantic segmentation device based on lidar point clouds, characterized in that, The device includes: The data acquisition module is used to acquire raw point cloud data; The feature extraction module is used to extract various features from the original point cloud data, including voxel features, point-level features, and look-around features. The feature fusion module is used to fuse multiple extracted features to obtain a fused feature map. The feature encoding module is used to perform grid-level feature encoding on the fused feature map to obtain a grid feature map; The semantic segmentation module is used to input the raster feature map into the semantic segmentation network for semantic prediction and obtain the raster-level semantic segmentation result. The apparatus implements the method of claim 1.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed, controls the device containing the computer-readable storage medium to perform the grid-level semantic segmentation method based on lidar point clouds as described in any one of claims 1 to 6.

Images