A deep learning-based 3D point cloud object segmentation method for multi-level highways

Abstract

The present invention discloses a method for three-dimensional point cloud target segmentation of a multi-level highway based on deep learning, and belongs to the technical field of three-dimensional point cloud target segmentation. The present invention first constructs a mapping relationship between preprocessed points and original point clouds through a point cloud data preprocessing operation based on a Map data structure and a K-D tree, and performs target segmentation processing on the preprocessed data based on the set multi-level highway point cloud target segmentation network, and then restores it to the original point cloud target segmentation result based on the mapping relationship. The data preprocessing operation of the present invention significantly reduces the amount of calculation. The target segmentation network of the present invention utilizes a feature aggregation module and an attention mechanism pooling to retain the geometric features of the point cloud data, and enhances the feature vector of the point, which is beneficial for the random sampling layer to extract global features. Establishing a residual network avoids gradient vanishing, and expands the receptive field of each point, which is beneficial for the random sampling layer to extract local geometric features of the point cloud, and greatly improves the accuracy of point cloud target segmentation.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional point cloud target segmentation technology, and in particular to a three-dimensional point cloud target segmentation method in a multi-level highway scene based on deep learning. Background Art

[0002] 3D point cloud target segmentation in multi-level highway scenes is to segment and identify different types of highway targets from the well-known highway scenes of different levels. It is the basis of fields such as assisted driving and autonomous driving. 3D point cloud target segmentation algorithms have gone through several stages over the years. From the traditional method proposed by Huang et al. to 3D point cloud target segmentation; later, some researchers proposed a point cloud target segmentation algorithm based on machine learning; foreign researchers Timo Hackel et al. proposed to extract neighboring point feature information based on KdTree, and then combined with the random forest method to realize 3D point cloud feature extraction, learning and fast classification; finally, with the proposal and development of deep learning methods, most scholars and researchers have focused on the field of deep learning for fast 3D point cloud target segmentation. The implementation of point cloud target segmentation based on deep learning methods includes different directions: voxelization-based methods, projection-based methods and point-based methods. (1) Point cloud voxelization is the process of simulating the geometric shape of a model or point cloud by using a spatially uniformly sized voxel grid. There are many ways to achieve model voxelization, such as octree-based 3D grid model voxelization. The voxelization method is to voxelize the point cloud and then use the convolutional neural network (CNN) to extract and learn the features of the voxels; (2) The projection-based method is to project the 3D point cloud into a two-dimensional image, including multi-view and spherical images, and then use the deep learning algorithm related to the two-dimensional image to extract and learn the features. The MVCNN proposed by Charles R. Qi et al. pioneered the use of two-dimensional views of three-dimensional data as the input of the convolutional neural network, and then used the pooling layer for aggregation to achieve better classification and target segmentation tasks; (3) Whether it is a voxel-based method or a projection-based method, there is a big drawback, that is, whether it is converting the three-dimensional point cloud into voxels or multi-views, it will occupy a lot of memory resources, and because of the huge amount of data, it requires a lot of computing resources, and it is impossible to achieve efficient point cloud recognition and segmentation. Therefore, direct point-based methods came into being. In 2017, PointNet was pioneered by CRQi, H.Su and others. The network learns the features of each point through several shared multi-layer perceptrons (Share MLP) and extracts global shape features through the maximum pooling layer (Maxpooling) to achieve point cloud target segmentation based on the point method.

[0003] LiDAR (Light Detection and Ranging) is an important instrument for collecting 3D point cloud data. Through LiDAR scanning, a lot of information about the target can be obtained, including 3D coordinate information, intensity information, color information, etc. Compared with 2D optical images, 3D point cloud data can better reflect the surrounding environment information. Therefore, 3D point cloud data has been applied to many fields including autonomous driving, robotics, 3D design, and medical treatment. However, due to the discreteness, sparsity, and massiveness of 3D point cloud data, especially in outdoor scenes such as highways, the data scale of point cloud data is huge and the density is uneven, resulting in low efficiency and low accuracy of target segmentation algorithms. Therefore, a multi-level highway 3D point cloud target segmentation method based on deep learning is needed to overcome the problems of existing methods. Summary of the invention

[0004] The purpose of the present invention is to provide a method for three-dimensional point cloud target segmentation of multi-level highways based on deep learning, so as to solve the problem of low efficiency and low accuracy of target segmentation of existing neural networks due to the large data scale and uneven density of point cloud data in multi-level highway scenes.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A method for segmenting a multi-level highway three-dimensional point cloud target based on deep learning comprises the following steps:

[0007] Step 1: Establish a data set: After performing regional segmentation, point cloud filtering, data thinning and classification labeling on the 3D point cloud data in the multi-level highway scene, a point cloud data set containing multiple types of objects in the multi-level highway scene is obtained;

[0008] Step 2: Data preprocessing: Based on Map (Map data structure) and KD tree (k-dimensional tree, a data structure of space partitioning, KD tree is actually a K-dimensional binary tree, each node contains a K-dimensional binary tree, and each node of the tree contains K-dimensional data), the data preprocessing operation of large-scale point cloud data is realized to reduce the scale of point cloud data and retain the mapping relationship between the original data and the preprocessed data. The mapping relationship is used to quickly restore the point cloud data:

[0009] Step 2.1: Determine the sampling grid size: According to the data volume of the point cloud data and the computing power of the computing resources, determine the size of the sampling grid during data preprocessing. The larger the grid, the smaller the data density after preprocessing. Conversely, the larger the data density after preprocessing. The grid size is set to K*K*K, which means that only one point is retained in the grid with a side length of K.

[0010] Step 2.2: Calculate the projection range of the point cloud data and the index position of the point: According to the grid size and the coordinate range of the input point cloud data, take the grid size as 1 unit, rebuild the coordinate system, calculate the projection range of the point cloud data in the new coordinate system, and traverse each point to calculate its index position in the new coordinate system;

[0011] Step 2.3: Construct a Map structure to store point cloud data and calculate the centroid position: Construct a Map structure to store the XYZ coordinates of the point cloud data, the label value, and count the number of points at each index position, and use the coordinate value and the number of points to calculate the centroid position of the point cloud data at the index position, and use the position point as the preprocessed point;

[0012] Step 2.4: Construct a KD tree to save the mapping relationship between the preprocessed points and the original point cloud: Use each preprocessed point as a node of the KD tree to build a KD tree, traverse all the points of the original point cloud, and project all the points of the original point cloud to the corresponding nodes of the KD tree by projection, so as to quickly restore the original point cloud data after target segmentation on the preprocessed point cloud data;

[0013] Step 3: Build and train a point cloud target segmentation network model: Build a target segmentation network model including an input layer, a random sampling layer (including a feature aggregation module and an attention mechanism pooling), an upsampling layer, and an output layer. Use the point cloud dataset to train and tune the above model, and obtain the optimal model based on the trained model.

[0014] Among them, the input layer extracts samples according to the point cloud data after input data preprocessing to obtain sampling points as input for subsequent feature extraction; the random sampling layer includes a feature aggregation module and an attention mechanism pooling, the feature aggregation module connects the feature information of the sampling point and its neighboring points in series, enhances the feature vector of the sampling point, and the attention mechanism pooling automatically learns the local feature information (including absolute position, relative position, Euclidean distance), summarizes the set of feature information of neighboring points, and finally generates the corresponding feature vector; the upsampling layer interpolates the point cloud data through the random sampling layer sharing the multi-layer perceptron; the output layer is composed of a fully connected layer for outputting the target segmentation result;

[0015] Step 4: Obtain target segmentation results based on the trained model: After performing data preprocessing on the segmented highway 3D point cloud data, input the trained point cloud target segmentation network model to obtain the target segmentation results. According to the target segmentation results, the mapping relationship established by the KD tree is used to quickly restore them to the original point cloud target segmentation results.

[0016] Preferably, step 1 comprises the following steps:

[0017] Step 1.1: Regional segmentation of the original highway point cloud data: segment the highway point cloud data of different lengths, set the length of each section to no more than 300m, the width to no more than 100m, and ensure that the types of targets in each section are rich;

[0018] Step 1.2: Filter the segmented point cloud data: According to the distribution range of the point cloud data in the three coordinate axes of XYZ, use a straight-through filter to remove outlier noise points within a specific range, and then use a statistical filter to remove outlier noise points with a density lower than a specified value (i.e., a relatively sparse density);

[0019] Step 1.3: Perform data thinning on the filtered point cloud data: Consider whether data thinning is needed based on the size of the 3D point cloud data after segmented filtering. If data thinning is needed, set the interval distance of randomly discarded points based on the degree of thinning, and use the random discarding method to complete a simple data thinning operation.

[0020] Step 1.4: Classify and label the point cloud data after data thinning: According to the ground TIN (triangular irregular network) model, the point cloud data is extracted into different target types using the elevation difference between different target types, and different labels are set.

[0021] Preferably, in step 3, the feature aggregation module and the attention mechanism pooling are connected together as a feature extraction operation, the two feature extraction operations are connected through a shared multi-layer perceptron, and stacked into a residual block, and finally N residual blocks are connected through a shared multi-layer perceptron to form a residual network, and the residual network is used as a random sampling layer, the output of the random sampling layer is interpolated by the upsampling layer, and then input into the fully connected layer to obtain the target segmentation result.

[0022] Preferably, in step 4, the target segmentation result of the original point cloud is restored according to the mapping relationship established by the KD tree based on the target segmentation result: through the mapping relationship established by the KD tree, each point of the preprocessed point cloud and its target segmentation result are traversed, and reversely mapped back to the original point cloud data to obtain the target segmentation result of the original point cloud data.

[0023] Preferably, the value range of K is recommended to be 0-1m.

[0024] Preferably, the value range of N is 1-4.

[0025] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0026] (1) The present invention reduces the data size of the original large-scale highway point cloud data through point cloud data preprocessing operations based on Map and KD tree, greatly reduces the amount of calculation for subsequent point cloud target segmentation in highway scenes using deep neural networks, saves a lot of computing resources, and improves the efficiency of target segmentation;

[0027] (2) The multi-level highway point cloud target segmentation network of the present invention retains the geometric features of the point cloud data by using the feature aggregation module and the attention mechanism pooling, and enhances the feature vector of the point, which is conducive to the random sampling layer to extract global features. The establishment of a residual network avoids the gradient disappearance and expands the receptive field of each point, which is conducive to the random sampling layer to extract the local geometric features of the point cloud, greatly improving the accuracy of point cloud target segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 is a flow chart of the method of the present invention;

[0030] Figure 2 A flowchart of establishing a data set for the present invention;

[0031] Figure 3 is a flow chart of data preprocessing of the present invention;

[0032] Figure 4 It is a Map structure diagram of data preprocessing of the present invention;

[0033] Figure 5 is a structural diagram of the network model of the present invention;

[0034] Figure 6 Schematic diagram of target segmentation result comparison of the present invention, wherein the (a) column represents the original point cloud data, the (b) column represents the true label result, and the (c) column represents the target segmentation result. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0036] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0037] like Figure 1As shown, an embodiment of the present invention provides a three-dimensional point cloud target segmentation method in a multi-level highway scene based on deep learning, comprising the following steps:

[0038] a: Establishing a data set: After performing regional segmentation, point cloud filtering, data thinning and classification labeling on the 3D point cloud data in the multi-level highway scene, a point cloud data set containing multiple types of objects in the multi-level highway scene is obtained (including training set and test set, the training set is used to train the network model, and the test set is used to test the processing performance), such as Figure 2 As shown, the specific processing includes:

[0039] a1: Regional segmentation: Use the segmentation function of Terrasolid (a laser radar point cloud data processing software) to first segment the large-scale point cloud data into blocks. The side length of each block is set to 1km. Based on the segmented data, the highway point cloud data of different lengths are manually segmented. The length of each road section is set to no more than 300m, and the width is set to no more than 100m. While segmenting, try to ensure that the types of targets in each road section are rich. The types of targets should include but not only billboards, roads, vegetation, pedestrian bridges, buildings and other types of targets. If the road section only contains roads and vegetation, you can consider abandoning the road section.

[0040] a2: Point cloud filtering: After the point cloud is segmented, the XYZ coordinate range of the point cloud data is roughly determined. According to the distribution range of the point cloud data in the three coordinate axes of XYZ, a straight-through filter is first used to remove outlier noise points within a specific range. The range size of each coordinate axis in the straight-through filter is set. The size is determined according to the segmentation. The setting range of X is 0-100, and the setting range of Y is 0-300. The noise points outside the range are filtered, and then the statistical filter is used to remove outlier noise points with relatively sparse density. The statistical filter will automatically calculate the regional density and filter the relatively sparse areas as noise points.

[0041] a3: Data thinning: Consider whether data thinning is needed based on the size of the 3D point cloud data after segmented filtering. If thinning is needed, set the interval distance of randomly discarded points according to the thinning degree. Use the random discarding method to complete simple data thinning operations. Use the Terrasolid tool to set the interval distance of randomly discarded points. The value is set based on the point cloud size and the thinning degree.

[0042] a4: Classification and labeling: According to the ground TIN model, the elevation difference between different target types is used to extract point cloud data into different target types and set different labels. Use the Terrasolid tool to establish a ground TIN model of point cloud data to separate ground points, and compare the elevation value of a point (center point) with the elevation value of each point within a given distance range. If the center point is significantly lower than other points, this point will be separated as a low point. The points above the ground are divided into three categories: low vegetation, medium-height vegetation, and high vegetation. Set the height range of each category, such as: low vegetation is set to points below 0.3m. This classification algorithm first uses the surface points separated above to establish a temporary TIN model, and then compares the height values ​​of other points with this TIN model. If it is lower than 0.3m, this point is classified into the low vegetation category. The other two classification methods are the same. This method can be used to set different elevation values ​​to separate billboards, buildings and other points. Finally, use manual classification to correct the misclassified points.

[0043] b: Data preprocessing: Based on Map and KD tree, large-scale point cloud data preprocessing operations are implemented to reduce the scale of point cloud data and retain the mapping relationship between the original data and the preprocessed data. The mapping relationship is used to quickly restore the point cloud data. The specific processing process is as follows: Figure 3 As shown, including:

[0044] b1: Set the grid size K: According to the data volume of the point cloud data and the computing power of the computing resources, determine the size of the sampling grid during data preprocessing. The larger the grid, the smaller the data density after preprocessing. Conversely, the larger the data density after preprocessing. The grid size is set to K*K*K, which means that only one point is retained in the grid with a side length of K. The larger the K, the more serious the loss of point cloud geometric features may be. Therefore, for highway point cloud data, the value of K is set in the range of 0-1m.

[0045] b2: Construct a new coordinate system with K as the unit, calculate the projection range of the original point cloud and the index position of each point projection: According to the grid size and the coordinate range of the input point cloud data, the coordinate system is reconstructed with the grid size K as 1 unit, and the projection range of the point cloud data in the new coordinate system is calculated. The calculation formula of the projection range X-axis range is as follows, and the Y-axis range is the same:

[0046]

[0047] In the formula, max x Indicates the maximum X-axis coordinate of the original point cloud data; ori x Indicates the starting coordinate of the X axis of the original point cloud data; size indicates the size of the voxel grid. Then, traverse each point to calculate its index position in the new coordinate system. The index position calculation formula is as follows:

[0048] mapIdx=index x +sampleX*index y +sampleX*sampleY*index z

[0049] Where mapIdx represents the final index value in the Map; index x 、index y 、index z They represent the index position in the grid calculated based on the X, Y, and Z coordinates of the current point. x 、index y 、index z The calculation formulas for these three values ​​are shown below.

[0050]

[0051] b3: Build a Map structure to store the point cloud data and calculate the center of gravity, using the center of gravity as the preprocessed point: Build a Map structure to store the XYZ coordinates of the point cloud data, label values, and count the number of points at each index position, such as Figure 4 , each Key (preprocessed index position) corresponds to a SampleData data structure, which contains the point cloud coordinates X, Y, Z, the label value lable and the variable count used for counting, and uses the coordinate value and the number of points to calculate the center of gravity of the point cloud data at the index position. The calculation formula is as follows:

[0052]

[0053] Finally, the centroid positions of all points in each virtual voxel grid are used to represent all points in the voxel grid, and the value of the label is the label value represented by the majority of points in the voxel grid.

[0054] b4: Construct a KD tree to save the mapping relationship between the preprocessed points and the original point cloud:

[0055] A KD tree is established for the preprocessed points, so that each point becomes a node of the KD tree. And through projection, all points of the original point cloud are projected to the corresponding nodes of the KD tree. In this way, when extracting samples later, you only need to extract the preprocessed points, and you can directly treat a node of the KD tree and its neighboring points as a point cloud as a sample based on the projection relationship. In high-dimensional space, this is more efficient and more accurate than the commonly used clustering algorithms, and is not affected by any parameters. At the same time, because a projection relationship is established between the original data and the preprocessed data, in subsequent reasoning, you only need to identify the preprocessed data, and then restore it to the original point cloud data through the projection relationship.

[0056] c: Build and train the point cloud target segmentation network: Build a target segmentation network model including input layer, random sampling layer (including feature aggregation module and attention mechanism pooling), upsampling layer, and output layer. Use the training set of the point cloud dataset to train and tune the above model to obtain the optimal model, such as Figure 5 As shown:

[0057] c1: Define the network model: The input layer extracts samples based on the pre-processed point cloud data as the input for subsequent feature extraction. The feature aggregation module and the attention mechanism pooling combine to form a random sampling layer. The purpose of the random sampling layer is to downsample the point cloud data of the input layer while ensuring that the geometric features of the point cloud are not destroyed as much as possible. The feature aggregation module takes the i-th point p after sampling i As the center point, use the KNN (K-Nearest Neighbor) algorithm to find the K points near the center point based on the Euclidean distance calculation Then, for the points near the center point, the relative positions are explicitly encoded to obtain the encoded feature vector And the transmission of feature information is completed through the multi-layer perceptron MLP, as shown in the following formula:

[0058]

[0059] Among them, MLP() represents the output of the multi-layer perceptron. Indicates p i The kth neighboring point of .

[0060] At the same time, the feature vector of the neighboring point is connected in series with the feature of the point to increase its feature vector. The enhanced feature vector Among them, f i k express This allows the network to explicitly observe local geometric patterns during feature extraction. The attention mechanism pooling first calculates the attention score. For a set of local feature vectors obtained The network learns a unique attention mechanism score for each feature through a shared function g(·). The shared function is implemented by an MLP and connected to the softmax function. The network automatically learns to obtain the ratio of the features of each neighboring point and allocates the position according to the importance of the feature. The more important the feature, the more positions it occupies. The definition of this function is as follows:

[0061]

[0062] Where W is the learnable weight. Then, based on the learned attention mechanism score, each feature is weighted and summed to obtain a new feature vector The weighted summation formula is as follows:

[0063]

[0064] The upsampling layer interpolates the point cloud data through the shared MLP of the random sampling layer to restore the point cloud data, and finally connects the fully connected layer as the output layer.

[0065] c2: Establish a residual network: connect the feature aggregation module and the attention mechanism pooling together as a feature extraction operation, connect the two feature extraction operations through a shared MLP, and stack them into a residual block, such as Figure 5 As shown in the figure, N residual blocks are finally connected through a shared MLP to form a residual network, which can not only avoid gradient vanishing, but also expand the network's receptive field, making it easier for the network to extract the local geometric structure of the point cloud. Since the more residual blocks are stacked, the slower the network training and reasoning speed will be, so the recommended setting range of N is 1-4.

[0066] d: Get the segmentation result. After data preprocessing, the test data is input into the trained target segmentation retention process to obtain the target segmentation result. According to the target segmentation result, the mapping relationship established by the KD tree is used to quickly restore it to the original point cloud target segmentation result.

[0067] This example compares the effect of setting different grid sizes K on target segmentation time and accuracy during data preprocessing, as shown in Table 1, and compares the effect of the number of residual blocks N on target segmentation efficiency and accuracy, as shown in Table 2:

[0068] Table 1 The influence of grid size K on target segmentation results

[0069] The value of K (m) mIoU(%) OA (%) Training time (h) Reasoning time(s) 0 88.49 96.7 18 14.52 0.3 87.25 95.8 14 10.28 0.5 86.62 95.2 9 6.72 0.7 83.36 93.6 6 5.74 1 79.14 92.1 4 4.43

[0070] Table 2 The impact of the number of residual blocks N on the target segmentation results

[0071] The value of N mIoU(%) OA (%) Training time (h) Reasoning time (s) 1 86.62 95.2 9 6.72 2 92.04 97.9 20 14.17

[0072] According to Table 1 and Table 2, the multi-level highway 3D point cloud target segmentation method of the present application has improved recognition efficiency and accuracy after adopting data preprocessing and residual structure; Figure 6 As shown in Figure 2, the effects before and after target segmentation are compared. Figure 6 The first column (a) in is the original point cloud data. Figure 6 The (b) column in the figure shows the actual label results, including roads, ground, billboards, overpasses, green vegetation, etc. Figure 6 The (c)th column in the figure is the result of target segmentation, and the black circled parts are the points that are segmented incorrectly. In summary, the three-dimensional point cloud target segmentation method for graded highways provided by the present invention realizes the rapid and accurate segmentation and recognition of point cloud data in highway scenes, and establishes a data set through operations such as regional segmentation, point cloud filtering, and point cloud thinning, thereby reducing the amount of calculation and improving the quality of network training. At the same time, the data preprocessing algorithm greatly reduces the scale of point cloud data and improves the efficiency of target segmentation. By establishing a target segmentation network with a residual structure, feature extraction and learning of point cloud data are performed to improve the accuracy of target segmentation.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0074] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for segmenting multi-level highway three-dimensional point cloud objects based on deep learning, characterized in that: The following steps are involved: Step 1: Create a dataset: After performing regional segmentation, point cloud filtering, data thinning and classification labeling on the three-dimensional point cloud data in the multi-level highway scene, a point cloud dataset containing multiple types of objects in the multi-level highway scene is obtained; Step 2: Data Preprocessing: Step 2.1: Set the grid size to K*K*K, that is, only one point is retained in the grid with a side length of K; Step 2.2: According to the sampling grid size and the coordinate range of the input point cloud data, the new coordinate system is reconstructed with the grid size as 1 unit, the projection range of the point cloud data in the new coordinate system is calculated, and each point is traversed to calculate its index position in the new coordinate system; Step 2.3: Construct a Map data structure to store the three-dimensional coordinates (X, Y, Z) of the point cloud data, the label value, and count the number of points at each index position, and use the coordinate value and the number of points to calculate the centroid position of the point cloud data at the index position, and use the centroid position as the preprocessed point; Step 2.4: Construct a KD tree to save the mapping relationship between the preprocessed points and the original point cloud: Use each preprocessed point as a node of the KD tree to build a KD tree, traverse all the points of the original point cloud, and project all the points of the original point cloud to the corresponding nodes of the KD tree by projection; Step 3: Build and train the point cloud object segmentation network model: Constructing a target segmentation network model including an input layer, a random sampling layer, an upsampling layer, and an output layer, training and tuning the target segmentation network model based on a training set of a point cloud data set, and obtaining a trained target segmentation network model; Among them, the input layer extracts samples based on the point cloud data after input data preprocessing to obtain sampling points as input for subsequent feature extraction; The random sampling layer includes a feature aggregation module and an attention mechanism pooling module. The feature aggregation module concatenates the feature information of the sampling point and its neighboring points to enhance the feature vector of the sampling point. The attention mechanism pooling automatically learns the local feature information, summarizes the set of feature information of the neighboring points, and finally generates the corresponding feature vector. The upsampling layer interpolates the point cloud data by sharing a multi-layer perceptron through a random sampling layer; The output layer is composed of a fully connected layer and is used to output the target segmentation result; Step 4: Get the final target segmentation result: Data preprocessing is performed on the segmented highway 3D point cloud data, and then the trained point cloud target segmentation network model is input to obtain the target segmentation result. According to the target segmentation result, the mapping relationship between the preprocessed points and the original point cloud constructed in step 2 is used to restore the original point cloud target segmentation result.

2. The method according to claim 1, characterized in that The step 1 comprises the following steps: Step 1.1: Regional segmentation of the original highway point cloud data: segment the highway point cloud data of different lengths, with the length of each section set to no more than 300m and the width set to no more than 100m; Step 1.2: Filter the segmented point cloud data: According to the distribution range of the point cloud data on the three coordinate axes of X, Y, and Z, use a straight-through filter to remove outlier noise points within a specific range, and then use a statistical filter to remove outlier noise points with a density lower than a specified value; Step 1.3: Perform data thinning on the filtered point cloud data: When the scale of the three-dimensional point cloud data after segmented filtering reaches the specified value, the interval distance of the randomly discarded points is set according to the thinning degree, and the random discarding method is used to perform data thinning operation; Step 1.4: Classify and label the point cloud data processed in step 1.3: According to the ground triangular irregular model, use the elevation difference between different target types to extract the point cloud data into different target types and set different labels.

3. The method according to claim 1, characterized in that In step 2, the value range of K is 0-1m.

4. The method according to claim 1, characterized in that In the step 3, the feature aggregation module and the attention mechanism pooling are connected together as a feature extraction operation, the two feature extraction operations are connected through a shared multi-layer perceptron, and stacked into a residual block, and finally N residual blocks are connected through a shared multi-layer perceptron to form a residual network, and the residual network is used as a random sampling layer, the output of the random sampling layer is interpolated by the upsampling layer, and then input into the fully connected layer to obtain the target segmentation result.

5. The method according to claim 4, characterized in that The value range of N is 1-4.

6. The method according to claim 1, characterized in that In step 4, the mapping relationship between the preprocessed points and the original point cloud constructed in step 2 is restored to the target segmentation result of the original point cloud: based on the mapping relationship between the preprocessed points and the original point cloud, each point of the preprocessed point cloud and its target segmentation result are traversed, and reversely mapped back to the original point cloud data to obtain the target segmentation result of the original point cloud data.

Images