A vehicle-mounted point cloud rod-shaped object instance segmentation method and computer-readable medium

Abstract

The present invention proposes a method for segmenting a vehicle-mounted point cloud rod-shaped object instance and a computer-readable medium. The present invention voxelizes the point cloud and determines the voxel attributes; uses voxels as nodes, the eight neighborhoods connecting the voxels as edges, and the Euclidean distance between the voxel centers as edge weights to construct a graph cut optimization model to obtain a coarse segmentation result when the energy function is optimized; uses a self-collected data set to train the PointNet++ network model, from which the 32-dimensional deep features of the original input point cloud are obtained; finds the connection area of ​​the adhesion rod-shaped objects in the coarse segmentation result and determines the buffer area; uses the points in the buffer area as nodes, the points within the k-nearest neighbors as neighbors to construct edges, and uses the high-dimensional deep learning feature distance as the edge weight to construct a graph cut optimization model to obtain the final fine segmentation result. The present invention retains the point cloud boundary information to the greatest extent, uses the energy function minimization to achieve global optimization of the segmentation result, and effectively realizes the refined segmentation of adhesion rod-shaped objects.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle-mounted laser radar point cloud data processing, and in particular relates to a vehicle-mounted point cloud rod-shaped object instance segmentation method and a computer-readable medium. Background Art

[0002] Pole-shaped features are important elements in urban roads, and as the construction of urban public facilities develops in the direction of "co-construction, sharing, and co-governance", pole-shaped features have become a carrier for collecting and interacting with various aspects of information such as roads, environment, and safety. Extracting pole-shaped features from vehicle-mounted three-dimensional laser scanning data is conducive to creating urban geographic spatial information databases, timely updating high-precision city maps, and promoting urban management and the construction of intelligent transportation systems. However, due to the large number and variety of pole-shaped structure targets in urban road scenes, and the phenomenon of mutual occlusion, it poses a huge challenge to the extraction and classification of pole-shaped features. However, urban road scenes are complex and there are many poles. The pole-shaped features containing pole-shaped structures are not only numerous, but also have complex and diverse types and forms. Pole-shaped feature category data and pole-shaped structure information are not sufficient to realize the interpretation and identification of feature targets in road scenes. Therefore, based on urban road pole-shaped feature data, this patent studies a single instance segmentation method for pole-shaped features under the premise of known pole-shaped structure coordinate information.

[0003] Pole-shaped features are objects that contain or are entirely pole-shaped, such as trees, streetlights, and traffic signs. Pole-shaped features are widely distributed man-made structures on urban roads, carrying a wealth of information about the road's conditions. However, due to the large number and variety of pole-shaped features in urban road scenes, as well as the mutual occlusion, pole-shaped features are difficult to extract independently. Studies of multiple datasets have revealed that among the numerous pole-shaped features in road scenes, street trees and man-made poles are more likely to be adjacent or adhered to each other. Segmenting pole-shaped features in 3D point clouds is a major challenge in current point cloud segmentation. While research on 3D point cloud segmentation is relatively mature, research on pole-shaped feature instance segmentation is relatively limited. Commonly used 3D point cloud segmentation methods can be categorized into the following: clustering-based, region-growing-based, model-based, graph-theory-based, and deep learning-based.

[0004] Clustering-based point cloud segmentation methods are mostly implemented in feature space. These algorithms offer good accuracy and stability for object-level model data. However, when processing point cloud data from large, complex scenes, the computation from point to feature space is time-consuming and susceptible to fluctuations in point cloud density. Region-growing-based point cloud segmentation methods are insensitive to noise points and are suitable for segmentation and classification of point cloud data in large, complex scenes. However, their effectiveness is overly dependent on the selection of seed surfaces and the growing strategy. Model-based point cloud segmentation methods have remained widely used in recent years. They rely on mathematical geometric prototypes, are insensitive to noise points, and offer fast processing speeds. However, they still have limitations for point cloud data without obvious regularities or large, complex scenes. Graph-based point cloud segmentation methods fully consider the global distribution and boundary characteristics of the data, are unaffected by point cloud density and noise, and are suitable for segmenting contiguous objects and large, complex scenes. However, their major drawback is the need to construct weights for each node. When dealing with large amounts of data, this requires simplifying the data while preserving point cloud boundary characteristics, resulting in high computational complexity and relatively low processing efficiency. In addition to the above types of segmentation algorithms, in recent years, many scholars have tried to use machine learning to extract features or directly apply end-to-end learning to point cloud segmentation problems. The results show that machine learning methods are highly robust and flexible, but they are overly dependent on data and usually take a lot of time.

[0005] In summary, traditional point cloud segmentation methods have good generalization capabilities but face the bottleneck of insufficient artificial feature representation. Deep learning, on the other hand, is data-driven, and the learned features can fully characterize the differences between samples, but it suffers from low generalization. Therefore, it is necessary to conduct research on point cloud segmentation methods that integrate deep learning features to improve the accuracy of instance segmentation of contiguous rod-shaped objects. To address these issues, this paper proposes an instance segmentation method that integrates deep learning and a graph cut model to extract rod-shaped objects. Experimental verification shows that this method has high accuracy and robustness. Summary of the Invention

[0006] In order to achieve high-precision instance segmentation of pole-shaped objects in urban street scenes, the present invention proposes a vehicle-mounted point cloud pole-shaped object instance segmentation method and a computer-readable medium.

[0007] The technical solution adopted by the method of the present invention is a method for segmenting rod-shaped objects in a vehicle-mounted point cloud, which specifically includes the following steps:

[0008] Step 1: Iteratively divide the vehicle point cloud into multiple voxels according to the spatial position, obtain the center point coordinates of each voxel, and calculate the Euclidean distance between the center point coordinates of any two voxels to represent the Euclidean distance between any two voxels;

[0009] Step 2: Construct a voxel-based graph cut model by combining multiple voxels and the Euclidean distance between voxels;

[0010] Step 3: Construct an energy function model, take the minimization of the energy function model as the optimization solution goal, and optimize and solve it through an exhaustive algorithm to obtain the first rod-shaped feature roughly segmented after the optimization solution and the second rod-shaped feature roughly segmented after the optimization solution;

[0011] Step 4: Combine multiple voxels to build a buffer;

[0012] Step 5: Use the PointNet++ network to perform fine segmentation of stick-shaped objects;

[0013] Preferably, the step 2 is as follows:

[0014] Define the first rod-shaped feature to be segmented , the second rod-shaped feature to be segmented ;

[0015] The voxel set of the rod-shaped structure of the first rod-shaped feature to be segmented is defined as ;

[0016] The voxel set of the rod-shaped structure of the second rod-shaped feature to be segmented is defined as ;

[0017] definition for The seed area;

[0018] definition for The seed area;

[0019] 、 The attribute values ​​of are marked as N;

[0020] except The remaining voxel set of except The remaining voxel sets of the remaining voxel sets together constitute the remaining voxel set to be segmented, which is defined as ;

[0021] The attribute value of is marked as 0;

[0022]

[0023] in, For all voxel sets of the rod-shaped object to be segmented, obtain multiple three-dimensional coordinates of each voxel in all voxel sets of the rod-shaped object to be segmented;

[0024] Each voxel to be segmented in the remaining set of voxels to be segmented is sequentially segmented into the first rod-shaped object to be segmented Or the second rod-shaped feature to be segmented , that is, assign a category label to each voxel to be segmented in the remaining set of voxels to be segmented, and determine whether it belongs to 、 , as follows:

[0025] Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in;

[0026] Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in;

[0027] Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in;

[0028] Will The mean of the three-dimensional coordinates of the center point of each voxel is calculated The three-dimensional coordinates of the center point are defined as ;

[0029] Will The mean of the three-dimensional coordinates of the center point of each voxel is calculated The three-dimensional coordinates of the center point are defined as ;

[0030] The node set for building the graph model is defined as follows:

[0031]

[0032] in, Represents the node set of the graph model, C represents The three-dimensional coordinates of the center point of each voxel in;

[0033] The set of edges that construct the graph model is defined as follows:

[0034] ={t-links,n-links}

[0035] in, Represents the set of edges of the graph model, and t-links represents the center point of any voxel in C and or The set of connected edges, n-links represents the set of edges between the center point of any voxel in C and the center points of its neighboring voxels;

[0036] Construct the weight set of the edges of the graph model. The weight of each edge in the weight set of the edges of the graph model is The Euclidean distance between any voxels in ;

[0037] Preferably, the energy function model in step 3 is defined as follows:

[0038]

[0039]

[0040]

[0041] in, E Represents the energy function model, which contains data items and energy terms , is the Euclidean distance calculation formula.

[0042] Indicates that the label Assign to voxel The penalty cost is calculated for each voxel to be segmented. Center point To the terminal node and The differences between Represents the set of all voxels of the rod-shaped features to be segmented;

[0043]

[0044] Where, is the constant term empirical threshold, ln() represents logarithmic operation, exp represents exponential operation, For any voxel to be segmented Distribute separately Tags and The penalty cost for labeling is:

[0045]

[0046] in, Any voxel representing the center point to be segmented The center point to The Euclidean distance, Any voxel representing the center point to be segmented The center point to The Euclidean distance, Indicates the assigned label, Used to measure the continuity between data. Represents the voxels to be segmented at the center point , the voxel to be segmented at the center point The center point of

[0047] The calculation method is shown in the following formula:

[0048]

[0049] in, Represents the voxel at the center point to be segmented To the center point of the voxel to be segmented The Euclidean distance of the center point;

[0050] The first rod-shaped feature is roughly segmented after the optimization solution is obtained through optimization solution , The second rod-shaped feature roughly segmented after optimization solution ;

[0051] Preferably, the buffer is constructed by combining multiple voxels in step 4, specifically as follows:

[0052] If any voxel in step 1 and multiple 26 neighborhoods belong to the first rod-shaped object to be segmented defined in step 2 Or the second rod-shaped feature to be segmented , then the arbitrary voxel exists in the first rod-shaped object to be segmented Or the second rod-shaped feature to be segmented the interior;

[0053] If the multiple 26 neighborhoods of any voxel in step 1 contain both the first rod-shaped object to be segmented The voxel in also contains the second rod-shaped feature to be segmented If the voxel in the first rod-shaped feature is Or the second rod-shaped feature to be segmented The adhesion area between them is defined as the voxel set:

[0054]

[0055] in, The set of voxels representing the adhesion area, The voxel set representing the adhesion area i voxels in the adhesion region, The number of voxels in the adhesion region in the voxel set representing the adhesion region;

[0056] buffer The voxel set defined as the adhesion region and the twenty-six neighboring sets of the voxel set of the adhesion region;

[0057] The twenty-six neighborhood sets of the voxel set in the adhesion area are defined as follows:

[0058] M

[0059] Where M represents the twenty-six neighborhood sets of the voxel set in the adhesion area, The voxel in the twenty-six neighborhood sets representing the adhesion area is the voxel in the twenty-six neighborhood sets representing the adhesion area. i Twenty-six neighborhoods;

[0060] In the region where the voxels are located, repeated voxels are not counted repeatedly, i.e. ;

[0061] Preferably, the step 5 is as follows:

[0062] Step 5.1: High-dimensional deep learning feature calculation;

[0063] Introduce the PointNet++ network and use multiple groups of sticky rod-shaped objects with coordinates and category labels as training data. Input each group of sticky rod-shaped objects in the training data into the PointNet++ network for two abstraction collection layers, feature transfer layers, and loss function calculations to obtain the buffer described in step 4. Multi-dimensional 32-dimensional high-dimensional deep learning features for each point in;

[0064] Among them, the loss function of PointNet++ is defined as:

[0065]

[0066] Among them, min means minimization, Indicates the i Points in the feature space Features in j Points in the feature space The Euclidean distance of the features in ;

[0067] Where:

[0068]

[0069] in, Respectively represent i Points in the feature space Features in j Points in the feature space Features in

[0070] Construct a single-point graph cut model in the buffer zone through steps 2 and 3;

[0071] Defining a buffer zone All points in are:

[0072]

[0073] in, represents a set of buffer points, Indicates the first i points;

[0074] The first rod-shaped feature after rough segmentation Except in the buffer zone The three-dimensional coordinates of all points in the mean are calculated to get the new center point as ;

[0075] The second rod-shaped feature after rough segmentation Except in the buffer zone The three-dimensional coordinates of all points in the mean are calculated to get the new center point as ;

[0076] Construct a node set based on the point graph model, which is defined as follows:

[0077]

[0078] in, Represents the node set of the graph model, C represents The three-dimensional coordinates of the center point of each voxel in;

[0079] Construct a set of edges for a vertex-based graph model, defined as follows:

[0080] ={t-links,n-links}

[0081] in, Represents the set of edges of the graph model, t-links represents Any point in and or The set of connected edges, represented by n-links Any point in Points within its k nearest neighbors The set of edges between the edges;

[0082] Construct the weight set of the edges of the graph model. The weight of each edge in the weight set of the edges of the graph model is The Euclidean distance of high-dimensional deep learning features between any two points in ;

[0083] Energy function model calculation;

[0084] Taking the minimization of the energy function model as the optimization solution goal, the exhaustive algorithm is used to optimize the solution and obtain the first rod-shaped object with fine segmentation after optimization solution. , the second rod-shaped feature after fine segmentation after optimization solution ;

[0085] .

[0086] in, E represents the energy function model, Represents a data item, represents the energy term;

[0087] Indicates that the label Give points The penalty cost is calculated as follows:

[0088]

[0089] Where, is the constant term empirical threshold, for Distribute separately Tags and The penalty cost when labeling;

[0090]

[0091] in, Indicates the first i Point to The high-dimensional deep learning feature distance between Indicates the first i Point to High-dimensional deep learning feature distance between them;

[0092] It is used to measure the continuity between data and is calculated as follows:

[0093]

[0094] in, Indicates the first i point to the buffer j The Euclidean distance of points. exp represents exponential calculation.

[0095] The present invention also provides a computer-readable medium, which stores a computer program executed by an electronic device. When the computer program runs on the electronic device, the steps of the vehicle-mounted point cloud rod-shaped object instance segmentation method are executed.

[0096] This method integrates artificial and deep features to effectively and fully extract local information between connected rod-shaped features, preserving the boundary information of the point cloud to a large extent. Furthermore, the method constructs graph cut models based on voxels and then on individual points, and achieves global optimization of the segmentation results by minimizing the energy function. This coarse-to-fine segmentation strategy effectively achieves refined segmentation of connected rod-shaped features. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 : A flow chart of a method according to an embodiment of the present invention;

[0098] Figure 2 : The effect diagram of the present invention implementation. DETAILED DESCRIPTION

[0099] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0100] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0101] The following combination Figure 1-2 The technical solution of the embodiment of the present invention is a method for segmenting rod-shaped objects in a vehicle-mounted point cloud, which is specifically as follows:

[0102] like Figure 1 FIG. 1 is a flow chart of a method according to an embodiment of the present invention.

[0103] Step 1: Iteratively divide the vehicle point cloud into multiple voxels according to the spatial position, obtain the center point coordinates of each voxel, and calculate the Euclidean distance between the center point coordinates of any two voxels to represent the Euclidean distance between any two voxels;

[0104] Step 2: Construct a voxel-based graph cut model by combining multiple voxels and the Euclidean distance between voxels;

[0105] Define the first rod-shaped feature to be segmented , the second rod-shaped feature to be segmented ;

[0106] The voxel set of the rod-shaped structure of the first rod-shaped feature to be segmented is defined as ;

[0107] The voxel set of the rod-shaped structure of the second rod-shaped feature to be segmented is defined as ;

[0108] definition for The seed area;

[0109] definition for The seed area;

[0110] 、 The attribute values ​​of are marked as N;

[0111] except The remaining voxel set of except The remaining voxel sets of the remaining voxel sets together constitute the remaining voxel set to be segmented, which is defined as ;

[0112] The attribute value of is marked as 0;

[0113]

[0114] in, For all voxel sets of the rod-shaped object to be segmented, obtain multiple three-dimensional coordinates of each voxel in all voxel sets of the rod-shaped object to be segmented;

[0115] Each voxel to be segmented in the remaining set of voxels to be segmented is sequentially segmented into the first rod-shaped object to be segmented Or the second rod-shaped feature to be segmented , that is, assign a category label to each voxel to be segmented in the remaining set of voxels to be segmented, and determine whether it belongs to 、 , as follows:

[0116] Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in;

[0117] Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in;

[0118] Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in;

[0119] Will The mean of the three-dimensional coordinates of the center point of each voxel is calculated The three-dimensional coordinates of the center point are defined as ;

[0120] Will The mean of the three-dimensional coordinates of the center point of each voxel is calculated The three-dimensional coordinates of the center point are defined as ;

[0121] The node set for building the graph model is defined as follows:

[0122]

[0123] in, Represents the node set of the graph model, C represents The three-dimensional coordinates of the center point of each voxel in;

[0124] The set of edges that construct the graph model is defined as follows:

[0125] ={t-links,n-links}

[0126] in, Represents the set of edges of the graph model, and t-links represents the center point of any voxel in C and or The set of connected edges, n-links represents the set of edges between the center point of any voxel in C and the center points of its neighboring voxels;

[0127] Construct the weight set of the edges of the graph model. The weight of each edge in the weight set of the edges of the graph model is The Euclidean distance between any voxels in ;

[0128] Step 3: Construct an energy function model, take the minimization of the energy function model as the optimization solution goal, and optimize and solve it through an exhaustive algorithm to obtain the first rod-shaped feature roughly segmented after the optimization solution and the second rod-shaped feature roughly segmented after the optimization solution;

[0129] The energy function model described in step 3 is defined as follows:

[0130]

[0131]

[0132]

[0133] in, E Represents the energy function model, which contains data items and energy terms , is the Euclidean distance calculation formula.

[0134] Indicates that the label Assign to voxel The penalty cost is calculated for each voxel to be segmented. Center point To the terminal node and The differences between Represents the set of all voxels of the rod-shaped features to be segmented;

[0135]

[0136] Where, is the constant term empirical threshold, ln() represents logarithmic operation, exp represents exponential operation, For any voxel to be segmented Distribute separately Tags and The penalty cost for labeling is:

[0137]

[0138] in, Any voxel representing the center point to be segmented The center point to The Euclidean distance, Any voxel representing the center point to be segmented The center point to The Euclidean distance, Indicates the assigned label, Used to measure the continuity between data. Represents the voxels to be segmented at the center point , the voxel to be segmented at the center point The center point of

[0139] The calculation method is shown in the following formula:

[0140]

[0141] in, Represents the voxel at the center point to be segmented To the center point of the voxel to be segmented The Euclidean distance of the center point;

[0142] The first rod-shaped feature is roughly segmented after the optimization solution is obtained through optimization solution , The second rod-shaped feature roughly segmented after optimization solution ;

[0143] Step 4: Combine multiple voxels to build a buffer;

[0144] If any voxel in step 1 and multiple 26 neighborhoods belong to the first rod-shaped object to be segmented defined in step 2 Or the second rod-shaped feature to be segmented , then the arbitrary voxel exists in the first rod-shaped object to be segmented Or the second rod-shaped feature to be segmented the interior;

[0145] If the multiple 26 neighborhoods of any voxel in step 1 contain both the first rod-shaped object to be segmented The voxel in also contains the second rod-shaped feature to be segmented If the voxel in the first rod-shaped feature is Or the second rod-shaped feature to be segmented The adhesion area between them is defined as the voxel set:

[0146]

[0147] in, The set of voxels representing the adhesion area, The voxel set representing the adhesion area i voxels in the adhesion region, =1024 represents the number of voxels in the adhesion region in the voxel set of the adhesion region;

[0148] buffer The voxel set defined as the adhesion region and the twenty-six neighboring sets of the voxel set of the adhesion region;

[0149] The twenty-six neighborhood sets of the voxel set in the adhesion area are defined as follows:

[0150] M

[0151] Where M represents the twenty-six neighborhood sets of the voxel set in the adhesion area, The voxel in the twenty-six neighborhood sets representing the adhesion area is the voxel in the twenty-six neighborhood sets representing the adhesion area. i Twenty-six neighborhoods;

[0152] In the region where the voxels are located, repeated voxels are not counted repeatedly, i.e. ;

[0153] Step 5: Use the PointNet++ network to perform fine segmentation of stick-shaped objects;

[0154] Step 5.1: High-dimensional deep learning feature calculation;

[0155] Introduce the PointNet++ network and use multiple groups of sticky rod-shaped objects with coordinates and category labels as training data. Input each group of sticky rod-shaped objects in the training data into the PointNet++ network for two abstraction collection layers, feature transfer layers, and loss function calculations to obtain the buffer described in step 4. Multi-dimensional 32-dimensional high-dimensional deep learning features for each point in;

[0156] Among them, the loss function of PointNet++ is defined as:

[0157]

[0158] Among them, min means minimization, Indicates the i Points in the feature space Features in j Points in the feature space The Euclidean distance of the features in ;

[0159] Where:

[0160]

[0161] in, Respectively represent i Points in the feature space Features in j Points in the feature space Features in

[0162] Construct a single-point graph cut model in the buffer zone through steps 2 and 3;

[0163] Defining a buffer zone All points in are:

[0164]

[0165] in, represents a set of buffer points, Indicates the first i points;

[0166] The first rod-shaped feature after rough segmentation Except in the buffer zone The three-dimensional coordinates of all points in the mean are calculated to get the new center point as ;

[0167] The second rod-shaped feature after rough segmentation Except in the buffer zone The three-dimensional coordinates of all points in the mean are calculated to get the new center point as ;

[0168] Construct a node set based on the point graph model, which is defined as follows:

[0169]

[0170] in, Represents the node set of the graph model, C represents The three-dimensional coordinates of the center point of each voxel in;

[0171] Construct a set of edges for a vertex-based graph model, defined as follows:

[0172] ={t-links,n-links}

[0173] in, Represents the set of edges of the graph model, t-links represents Any point in and or The set of connected edges, represented by n-links Any point in Points within its k nearest neighbors The set of edges between the edges;

[0174] Construct the weight set of the edges of the graph model. The weight of each edge in the weight set of the edges of the graph model is The Euclidean distance of high-dimensional deep learning features between any two points in ;

[0175] Energy function model calculation;

[0176] Taking the minimization of the energy function model as the optimization solution goal, the exhaustive algorithm is used to optimize the solution and obtain the first rod-shaped object with fine segmentation after optimization solution. , the second rod-shaped feature after fine segmentation after optimization solution ;

[0177] .

[0178] in, E represents the energy function model, Represents a data item, represents the energy term;

[0179] Indicates that the label Give points The penalty cost is calculated as follows:

[0180]

[0181] Where, is the constant term empirical threshold, for Distribute separately Tags and The penalty cost when labeling;

[0182]

[0183] in, Indicates the first i Point to The high-dimensional deep learning feature distance between Indicates the first i Point to High-dimensional deep learning feature distance between them;

[0184] It is used to measure the continuity between data and is calculated as follows:

[0185]

[0186] in, Indicates the first i point to the buffer j The Euclidean distance of points. exp represents exponential calculation.

[0187] like Figure 2 As shown in the figure, (a) is the input original data, (b) is the manually labeled true value, (c) is the coarse segmentation result after step 3 of the present invention, and (d) is the fine segmentation result after step 5 of the present invention, which is the final result of the present invention. By comparing (c) and (d) with the manually labeled true value (b), the effectiveness of the present invention can be verified.

[0188] A specific embodiment of the present invention also provides a computer-readable medium.

[0189] The computer readable medium is a server workstation;

[0190] The server workstation stores a computer program executed by an electronic device. When the computer program is executed on the electronic device, the electronic device executes the steps of the method for segmenting pole-shaped objects in a vehicle-mounted point cloud according to an embodiment of the present invention.

[0191] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0192] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. A method for segmenting rod-shaped objects from vehicle-mounted point clouds, characterized by: The following steps are involved: Step 1: Iteratively divide the vehicle point cloud into multiple voxels according to the spatial position, obtain the center point coordinates of each voxel, and calculate the Euclidean distance between the center point coordinates of any two voxels to represent the Euclidean distance between any two voxels; Step 2: Construct a voxel-based graph cut model by combining multiple voxels and the Euclidean distance between voxels; Step 3: Construct an energy function model, take the minimization of the energy function model as the optimization solution goal, and optimize and solve it through an exhaustive algorithm to obtain the first rod-shaped feature roughly segmented after the optimization solution and the second rod-shaped feature roughly segmented after the optimization solution; Step 4: Combine multiple voxels to build a buffer; Step 5: Use the PointNet++ network to perform fine segmentation of stick-shaped objects; Wherein, the step 2 is specifically as follows: Define the first rod-shaped feature to be segmented , the second rod-shaped feature to be segmented ; The voxel set of the rod-shaped structure of the first rod-shaped feature to be segmented is defined as ; The voxel set of the rod-shaped structure of the second rod-shaped feature to be segmented is defined as ; definition for The seed area; definition for The seed area; 、 The attribute values ​​of are marked as N; except The remaining voxel set of except The remaining voxel sets of the remaining voxel sets together constitute the remaining voxel set to be segmented, which is defined as ; The attribute value of is marked as 0; in, For all voxel sets of the rod-shaped object to be segmented, obtain multiple three-dimensional coordinates of each voxel in all voxel sets of the rod-shaped object to be segmented; Each voxel to be segmented in the remaining set of voxels to be segmented is sequentially segmented into the first rod-shaped object to be segmented Or the second rod-shaped feature to be segmented , that is, assign a category label to each voxel to be segmented in the remaining set of voxels to be segmented, and determine whether it belongs to 、 , as follows: Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in; Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in; Will The multiple three-dimensional coordinates of each voxel in are obtained by averaging The three-dimensional coordinates of the center point of each voxel in; Will The mean of the three-dimensional coordinates of the center point of each voxel is calculated The three-dimensional coordinates of the center point are defined as ; Will The mean of the three-dimensional coordinates of the center point of each voxel is calculated The three-dimensional coordinates of the center point are defined as ; The node set for building the graph model is defined as follows: in, Represents the node set of the graph model, C represents The three-dimensional coordinates of the center point of each voxel in; The set of edges that construct the graph model is defined as follows: ={t-links,n-links} in, Represents the set of edges of the graph model, and t-links represents the center point of any voxel in C and or The set of connected edges, n-links represents the set of edges between the center point of any voxel in C and the center points of its neighboring voxels; Construct the weight set of the edges of the graph model. The weight of each edge in the weight set of the edges of the graph model is The Euclidean distance between any voxels in .

2. The method for segmenting rod-shaped objects from vehicle-mounted point clouds according to claim 1, wherein: The energy function model described in step 3 is defined as follows: in, E Represents the energy function model, which contains data items and energy terms , is the Euclidean distance calculation formula; Indicates that the label Assign to voxel The penalty cost is calculated for each voxel to be segmented. Center point To the terminal node and The differences between Represents the set of all voxels of the rod-shaped features to be segmented; is the constant term empirical threshold, ln() represents logarithmic operation, exp represents exponential operation, For any voxel to be segmented Distribute separately Tags and The penalty cost when labeling; Any voxel representing the center point to be segmented The center point to The Euclidean distance, Any voxel representing the center point to be segmented The center point to The Euclidean distance, Indicates the assigned label, Used to measure the continuity between data. Represents the voxels to be segmented at the center point , the voxel to be segmented at the center point The center point of Represents the voxel at the center point to be segmented To the center point of the voxel to be segmented The Euclidean distance of the center point; The first rod-shaped feature is roughly segmented after the optimization solution is obtained through optimization solution , The second rod-shaped feature roughly segmented after optimization solution .

3. The method for segmenting rod-shaped objects from vehicle-mounted point clouds according to claim 2, wherein: Step 4 combines multiple voxels to construct a buffer as follows: If any voxel in step 1 and multiple 26 neighborhoods belong to the first rod-shaped object to be segmented defined in step 2 Or the second rod-shaped feature to be segmented , then the arbitrary voxel exists in the first rod-shaped object to be segmented Or the second rod-shaped feature to be segmented the interior; If the multiple 26 neighborhoods of any voxel in step 1 contain both the first rod-shaped object to be segmented The voxel in also contains the second rod-shaped feature to be segmented If the voxel in the first rod-shaped feature is Or the second rod-shaped feature to be segmented The adhesion area between them is defined as the voxel set: in, The set of voxels representing the adhesion area, The voxel set representing the adhesion area i voxels in the adhesion region, The number of voxels in the adhesion region in the voxel set representing the adhesion region; buffer The voxel set defined as the adhesion region and the twenty-six neighboring sets of the voxel set of the adhesion region; The twenty-six neighborhood sets of the voxel set in the adhesion area are defined as follows: M Where M represents the twenty-six neighborhood sets of the voxel set in the adhesion area, The voxel in the twenty-six neighborhood sets representing the adhesion area is the voxel in the twenty-six neighborhood sets representing the adhesion area. i Twenty-six neighborhoods; In the region where the voxels are located, repeated voxels are not counted repeatedly, i.e. .

4. The method for segmenting rod-shaped objects from vehicle-mounted point clouds according to claim 3, wherein: The step 5 is as follows: Step 5.1: High-dimensional deep learning feature calculation; Introduce the PointNet++ network and use multiple groups of sticky rod-shaped objects with coordinates and category labels as training data. Input each group of sticky rod-shaped objects in the training data into the PointNet++ network for two abstraction collection layers, feature transfer layers, and loss function calculations to obtain the buffer described in step 4. Multi-dimensional 32-dimensional high-dimensional deep learning features for each point in; Among them, the loss function of PointNet++ is defined as: Among them, min means minimization, Indicates the i Points in the feature space Features in j Points in the feature space The Euclidean distance of the features in ; Where: in, Respectively represent i Points in the feature space Features in j Points in the feature space Features in Construct a single-point graph cut model in the buffer zone through steps 2 and 3; Defining a buffer zone All points in are: in, represents a set of buffer points, Indicates the first i points; The first rod-shaped feature after rough segmentation Except in the buffer zone The three-dimensional coordinates of all points in the mean are calculated to get the new center point as ; The second rod-shaped feature after rough segmentation Except in the buffer zone The three-dimensional coordinates of all points in the mean are calculated to get the new center point as ; Construct a node set based on the point graph model, which is defined as follows: in, Represents the node set of the graph model, C represents The three-dimensional coordinates of the center point of each voxel in; Construct a set of edges for a vertex-based graph model, defined as follows: ={t-links,n-links} in, Represents the set of edges of the graph model, t-links represents Any point in and or The set of connected edges, represented by n-links Any point in Points within its k nearest neighbors The set of edges between the edges; Construct the weight set of the edges of the graph model. The weight of each edge in the weight set of the edges of the graph model is The Euclidean distance of high-dimensional deep learning features between any two points in ; Energy function model calculation; Taking the minimization of the energy function model as the optimization solution goal, the exhaustive algorithm is used to optimize the solution and obtain the first rod-shaped object with fine segmentation after optimization solution. , the second rod-shaped feature after fine segmentation after optimization solution ; . in, E represents the energy function model, Represents a data item, represents the energy term; Indicates that the label Give points The penalty cost is calculated as follows: Where, is the constant term empirical threshold, for Distribute separately Tags and The penalty cost when labeling; in, Indicates the first i Point to The high-dimensional deep learning feature distance between Indicates the first i Point to High-dimensional deep learning feature distance between them; It is used to measure the continuity between data and is calculated as follows: in, Indicates the first i point to the buffer j The Euclidean distance of points. exp represents exponential calculation.

5. A computer-readable medium, characterized in that It stores a computer program executed by an electronic device, and when the computer program runs on the electronic device, the electronic device executes the steps of the method according to any one of claims 1 to 4.

Images