A semantic segmentation method for single object point cloud data using contour point features

Abstract

This invention discloses a method for semantic segmentation of single-object point cloud data using contour point features, belonging to the field of point cloud semantic segmentation technology. The method comprises: obtaining a single-object point cloud containing N points, and projecting the single-object point cloud in the X, Y, and Z directions to obtain three two-dimensional point sets; obtaining the contour point set of the single-object point cloud from the three two-dimensional point sets; and inputting the single-object point cloud and the contour point set into a geometric feature fusion network to obtain a classification label for each point in the single-object point cloud. The method extracts high-purity geometric features from the contour points of the single-object point cloud, and then uses these geometric features to improve the semantic segmentation accuracy of the single-object point cloud, ultimately achieving more accurate semantic segmentation accuracy for the single-object point cloud data.

Description

Technical Field

[0001] The present invention relates to the technical field of semantic segmentation of point clouds, and in particular to a semantic segmentation method for single-object point cloud data using contour point features. Background Art

[0002] With the rapid development of point cloud acquisition equipment and 3D reconstruction technology, 3D point cloud data has been widely used in fields such as autonomous driving, robotics, and immersive interaction. In point cloud data processing and applications, semantic segmentation is a key step in extracting semantic information from point cloud geometric information. Specifically, semantic segmentation involves classifying each point in a point cloud according to human-recognized categories. For example, in autonomous driving point cloud data, points representing "pedestrians" and "roads" are labeled with different categories. Previous point cloud semantic segmentation techniques tend to achieve high accuracy across all point cloud categories. However, there are significant differences between different point cloud data types. Point clouds can be simply categorized into single-object point clouds and multi-object point clouds. Single-object point clouds represent only a single object, such as a cup, a chair, or an airplane. Multi-object point clouds often represent a scene, such as an indoor scene, consisting of the floor, walls, tables, and chairs. Different point cloud data types have their own unique characteristics. Therefore, leveraging these characteristics to implement highly applicable semantic segmentation techniques often results in higher accuracy. Compared to multi-object point clouds, single-object point clouds offer simpler content and higher quality detail, but they also suffer from significant information redundancy. For example, a cup's point cloud might contain thousands of points, but only one-tenth of them might be needed to clearly distinguish the cup's shape, angle, and other information. The remaining points provide little useful information, a phenomenon known as information redundancy. This redundancy increases the burden on semantic segmentation neural networks, making the features they extract less pure or even missing useful features. Summary of the Invention

[0003] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0004] To this end, the purpose of the present invention is to propose a semantic segmentation method for single object point cloud data using contour point features, which reduces the burden of the neural network and improves the segmentation accuracy.

[0005] To achieve the above-mentioned purpose, the embodiment of the present invention proposes a semantic segmentation method for single object point cloud data using contour point features, comprising the following steps: Step S1, obtaining a single object point cloud containing N points, and projecting the single object point cloud in the X, Y, and Z directions respectively to obtain three two-dimensional point sets, denoted as V and Z respectively. x 、V y 、V z; Step S2, obtaining the contour point set of the single object point cloud from the three two-dimensional point sets; Step S3, inputting the single object point cloud and the contour point set into the geometric feature fusion network to obtain the classification label of each point in the single object point cloud.

[0006] The semantic segmentation method of single object point cloud data using contour point features of an embodiment of the present invention is highly suitable for single object point clouds. High-purity geometric features are extracted from the contour points of the single object point cloud, and then input into the neural network to achieve semantic segmentation, thereby achieving more accurate semantic segmentation accuracy of single object point cloud data and effectively reducing the burden on the neural network.

[0007] In addition, the single object point cloud data semantic segmentation method using contour point features according to the above embodiment of the present invention may also have the following additional technical features:

[0008] Furthermore, in one embodiment of the present invention, the step S2 specifically includes: step S21, obtaining the corresponding contour point sets C according to the three two-dimensional point sets respectively. x 、C y 、C z ; Step S22, the contour point set C x 、C y 、C z The contour points in are projected back to the original three-dimensional space, and the three point sets are combined to obtain the contour point set C.

[0009] Furthermore, in one embodiment of the present invention, the step S21 specifically includes: step S211, selecting the two-dimensional point set V x Any point x in i , find the two-dimensional point set V x Middle distance x i Neighborhood point set N within the preset range r i ; Step S212, select the neighborhood point set N i Any point x in j , with x i As the starting point, pass x j Draw ray v j Step S213: any two adjacent rays in the plurality of rays obtained in step S212 have an angle, and the maximum value of all angles is recorded as α i ; Step S214, determine the maximum value α i Is it greater than the preset threshold t? If so, x i Add the contour point set C x , otherwise not added.

[0010] Furthermore, in one embodiment of the present invention, in step S22, if the three contour point sets have more than n points after being combined, n points are randomly sampled; if there are less than n points, the points in the set are repeatedly sampled to make up n points.

[0011] Furthermore, in one embodiment of the present invention, the geometric feature fusion network includes four feature fusion modules and three feature upsampling modules, wherein each feature fusion module includes three feature extraction modules.

[0012] Furthermore, in one embodiment of the present invention, the step S3 specifically includes: step S31, inputting the single object point cloud and the contour point set into the four feature fusion modules of the geometric feature fusion network in sequence to obtain feature matrices at each level; step S32, starting with the highest level feature matrix, inputting them into the three feature acquisition modules together with the lower level feature matrix in sequence, and finally obtaining the classification label of each point.

[0013] Furthermore, in one embodiment of the present invention, any feature fusion module in step S301 specifically includes: step S311, determining the coordinate-feature matrix F′ of the subset point of the single object point cloud P And the coordinates of the subset points of the contour point set - the feature matrix F' C As the input of the feature fusion module; Step S312, sample M "points from the subset of the single object point cloud, denoted as P", and sample m "points from the subset of the contour point set, denoted as C"; Step S313, P "and F' P Input into the feature extraction module and get the midpoint of P″ from F′ P The extracted feature matrix is spliced to obtain the spliced feature; in step S314, the spliced feature is input into the convolution layer to obtain the feature of the point in P″, which is spliced with the coordinate matrix P″ to obtain the coordinate-feature matrix F″ of the new subset point of the single object point cloud P , as the first output of the feature fusion module; Step S315, C ″ and F ′ obtained in step 312 P Input into the feature extraction module and get the midpoint of C″ from F′ P Extracted features; Step S316, concatenate the features obtained in step 315 with C″ to obtain the coordinates of the new subset of contour points - feature matrix F″ C , as the second output of the feature fusion module.

[0014] Furthermore, in one embodiment of the present invention, the coordinate-feature matrix F′ P The size is M′*(3+c), the coordinate-feature matrix F′ C is m′*(3+c).

[0015] Furthermore, in one embodiment of the present invention, the step S313 specifically includes: step S3131, determining the coordinate matrix P′ of the subset points of the single object point cloud and the coordinate-feature matrix F′ of the subset points of the single object point cloud P is the input of the feature extraction module, where P′ is of size n*3 and F′ P The size is n*(3+c′), and the two subsets are not the same set; step S3132, find the k points closest to it in the coordinate matrix P′, and obtain the coordinate-feature matrix of the neighborhood point set of P′, with a size of n*k*(3+c′); step S3133, input the neighborhood point set coordinate-feature matrix into the first three-layer convolution to obtain a feature matrix with a dimension of n*k*c″; step S3134, input the neighborhood point set coordinate-feature matrix into the second three-layer convolution to obtain a feature rotation matrix with a dimension of n*j*k; step S3135, multiply the feature matrix obtained in step 3133 and the feature rotation matrix obtained in step 3134 to obtain a rotated feature matrix; step S3136, input the rotated feature matrix into its one-layer convolution to obtain the output of the feature extraction module.

[0016] Furthermore, in one embodiment of the present invention, the step S32 specifically includes: step S321, the coordinate matrix P' of the subset point of the single object point cloud and the coordinate-feature matrix F' of the subset point of the single object point cloud are combined. P Input to the feature upsampling module, where the sizes of the two matrices are M'*3 and m'*(3+c) respectively; Step S322, the coordinate matrix P' and the feature matrix F' P Input to the feature extraction module and get the midpoint of P′ from F′ P Extracted features; Step S323, concatenating the features obtained in Step S322 with the coordinate matrix P′ to obtain the output of the feature upsampling module.

[0017] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0019] Figure 1 This is a flow chart of a method for semantic segmentation of single object point cloud data using contour point features according to an embodiment of the present invention;

[0020] Figure 2 1 is a schematic diagram of contour point set extraction according to an embodiment of the present invention;

[0021] Figure 3 1 is a schematic diagram illustrating a specific implementation of a method for semantic segmentation of single object point cloud data using contour point features according to an embodiment of the present invention;

[0022] Figure 4 This is a structural diagram of a feature fusion module according to an embodiment of the present invention;

[0023] Figure 5 is a structural diagram of a feature extraction module according to an embodiment of the present invention;

[0024] Figure 6 4 is a structural diagram of an upsampling module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0026] The following describes a semantic segmentation method for single object point cloud data using contour point features according to an embodiment of the present invention with reference to the accompanying drawings.

[0027] Example 1:

[0028] Figure 1 This is a flowchart of a method for semantic segmentation of single object point cloud data using contour point features according to an embodiment of the present invention.

[0029] like Figure 1 As shown, the semantic segmentation method of single object point cloud data using contour point features includes the following steps:

[0030] In step S1, a single object point cloud containing N points is obtained, and the single object point cloud is projected in the X, Y, and Z directions respectively to obtain three two-dimensional point sets, which are respectively denoted as V x 、V y 、V z .

[0031] Specifically, in the embodiment of the present invention, a single object point cloud containing N points can be represented as an N*3 matrix, denoted as P, and the single object point cloud is projected in the X, Y, and Z directions respectively to obtain three two-dimensional point sets, each represented as an N*2 matrix, denoted as V respectively. x 、V y 、V z .

[0032] The projection of a single object point cloud in the X, Y, and Z directions can be easily achieved by removing the X, Y, and Z coordinates. For example, the projection of a single object point cloud in the X direction can be obtained by removing the column corresponding to the X coordinate in the N*3 matrix, resulting in an N*2 matrix.

[0033] In step S2, a contour point set of a single object point cloud is obtained from three two-dimensional point sets.

[0034] That is, the three two-dimensional point sets V obtained from step S1 x 、V y 、V z Obtain the contour point set of a single object point cloud, expressed as an n*3 matrix, denoted as C.

[0035] Furthermore, if Figure 2 As shown, in step S2, the three two-dimensional point sets V x 、V y 、V z , the process of obtaining the contour point set C of a single object point cloud is as follows:

[0036] Step S21, based on three two-dimensional point sets V x 、V y 、V z , respectively obtain the corresponding contour point set C x 、C y 、C z ;

[0037] Step S22: Contour point set C x 、C y 、C z The contour points in are projected back to the original three-dimensional space, and the three point sets are combined to obtain the contour point set C, which is represented as an n*3 matrix. If the three contour point sets have more than n points after being combined, n points are randomly sampled. If there are less than n points, the points in the union can be repeatedly sampled to make up n points.

[0038] Furthermore, in step S21, the process of obtaining the contour point set from the two-dimensional point set (by V x Get C x For example) is:

[0039] Step S211, for X-ray Figure 2 Dimensional point set V x Any point x in i , find the two-dimensional point set V x Middle distance x i Neighborhood point set N within the preset range r i ;

[0040] Step S212: Select neighborhood point set N i Any point x inj , with x i As the starting point, pass x j Draw ray v j , all rays have the same starting point, namely x i ;

[0041] Step S213: For any two adjacent rays among the multiple rays obtained in step S212, there is an angle between them. The maximum value of all angles is recorded as α. i ;

[0042] Step S214, determine the maximum value α i Is it greater than the preset threshold t? If so, x i Add contour point set C x , otherwise not added.

[0043] In step S3, the single object point cloud and the contour point set are input into the geometric feature fusion network to obtain the classification label of each point in the single object point cloud.

[0044] Furthermore, if Figure 3 As shown in FIG, the process of obtaining the classification labels of each point in the point cloud in step S3 is as follows:

[0045] Step S31, inputting the single object point cloud and the contour point set into the four feature fusion modules of the geometric feature fusion network in sequence to obtain feature matrices at each level;

[0046] Step S32: Starting with the highest-level feature matrix, it is sequentially input into the three feature acquisition modules together with the lower-level feature matrix, and finally the classification label of each point is obtained.

[0047] Furthermore, if Figure 4 As shown, the process of extracting features by the feature fusion module in the embodiment of the present invention is as follows:

[0048] Step S311: The input of the feature fusion module in step S31 is two matrices, namely the coordinates of the subset points of the single object point cloud - the feature matrix F′ P And the coordinates of the subset of contour points - feature matrix F′ C , the sizes of the two matrices are M′*(3+c) and m′*(3+c) respectively. It should be noted that the input of the first feature fusion module is the coordinate matrix P of the single object point cloud and the coordinate matrix C of the contour points. At this time, the full set is regarded as a special subset and the features are regarded as zero-dimensional, that is, M′=N, m′=n, c=0;

[0049] Step S312: Sample M″ points from the subset of the single object point cloud, expressed as an M″*3 matrix, denoted as P″; sample m″ points from the subset of the contour point set, expressed as an m″*3 matrix, denoted as C″;

[0050] Step S313: P″ and F′ P Input into the feature extraction module and get the midpoint of P″ from F′ P The extracted feature matrix, the two feature matrices are spliced to obtain the spliced features;

[0051] Step S314: Input the spliced features into the convolution layer to obtain the features of the points in P″, and splice them with the coordinate matrix P″ to obtain the coordinate-feature matrix F″ of the new subset of the single object point cloud. P , as the first output of the feature fusion module;

[0052] Step S315: C″ and F′ obtained in step 312 are P Input into the feature extraction module and get the midpoint of C″ from F′ P Extracted features;

[0053] Step S316: Concatenate the features obtained in step 315 with C″ to obtain the coordinates of the new subset of contour points - feature matrix F″ C , as the second output of the feature fusion module.

[0054] Furthermore, if Figure 5 As shown, the process of extracting features by the feature extraction module in step S313 is as follows:

[0055] Step S3131: The input of the feature extraction module is two matrices. One is the coordinate matrix P′ of the subset of the single object point cloud, with a size of n*3. The other is the coordinate matrix F′ of the subset of the single object point cloud. P , the size is n*(3+c′), the two subsets are not the same set;

[0056] Step S3132: For each point in P′, find the k points closest to it in the coordinate matrix P′, and obtain the coordinate-feature matrix of the neighborhood point set of P′, with a size of n*k*(3+c′);

[0057] Step S3133: Input the neighborhood point set coordinates-feature matrix into the first three-layer convolution to obtain a feature matrix with a dimension of n*k*c″;

[0058] Step S3134: Input the neighborhood point set coordinates-feature matrix into the second three-layer convolution to obtain a feature rotation matrix with a dimension of n*k*k;

[0059] Step S3135, multiplying the characteristic matrix obtained in step 3133 and the characteristic rotation matrix obtained in step 3134 to obtain a rotated characteristic matrix;

[0060] Step S3136: Input the rotated feature matrix into one layer of convolution to obtain the output of the feature extraction module.

[0061] Furthermore, if Figure 6 As shown, the process of extracting features by the feature upsampling module in step S32 is as follows:

[0062] Step S321: The coordinate matrix P′ of the subset of the single object point cloud and the coordinate-feature matrix F′ of the subset of the single object point cloud are combined. P Input to the feature upsampling module, where the sizes of the two matrices are M′*3 and m′*(3+c) respectively;

[0063] Step S322: coordinate matrix P′ and feature matrix F′ P Input to the feature extraction module and get the midpoint of P′ from F′ P Extracted features;

[0064] In step S323, the features obtained in step S322 are concatenated with the coordinate matrix P′ to obtain the output of the feature upsampling module.

[0065] Example 2

[0066] The difference between Example 2 and Example 1 is that the contour points of this embodiment are replaced by any subset of the single object point cloud. In essence, the contour points are also a subset of the single object point cloud, so any method of replacing contour points with a subset can be regarded as a special case of Example 1.

[0067] Example 3

[0068] The difference between Example 3 and Example 1 is that the geometric feature fusion network in this embodiment can be integrated into a single deep neural network through simple transformations. In principle, the distinction between different network modules in Steps S31 and S32 is for convenience of description and is based on function. During training and deployment, the entire network is an end-to-end form, so the conceptual distinction between network modules is a special case of Example 1.

[0069] Example 4

[0070] The difference between Example 4 and Example 1 is that in the single object point cloud semantic segmentation method of this embodiment, the parameters of the contour extraction technology in step S2 can be arbitrary parameters, and the method and process after adjusting the parameters of the contour extraction algorithm are the same as those in Example 1.

[0071] Example 5

[0072] The difference between Example 5 and Example 1 is that in the single object point cloud semantic segmentation method of this embodiment, the number of layers and parameters of each neural network module in step S3 can be the parameters and number of layers of any neural network, and the method and process of adjusting the parameters and number of layers of the neural network structure are the same as those in Example 1.

[0073] Example 6

[0074] The difference between Example 6 and Example 1 is that in the single-object point cloud semantic segmentation method of this embodiment, the point cloud data not only contains coordinate information, but also contains attribute information such as reflection intensity, color, normal vector, etc. of each point. The method and process after adding input information are the same as those in Example 1.

[0075] Example 7

[0076] The difference between Example 7 and Example 1 is that this example is applied to semantic segmentation of multi-object point cloud data, and the method and process of applying the present invention to other types of data are the same as those in Example 1.

[0077] In addition, the above embodiment of the present invention is further illustrated below through a comparative experiment.

[0078] Taking Example 1 as an example, we tested it on the ShapeNet Parts dataset and compared it with existing point cloud semantic segmentation techniques: PointNet, PointNet++, and PointCNN. The results are shown below. It can be seen that the method of the present invention outperforms all three existing methods.

[0079] method PointNet PointNet++ PointCNN The present invention Partial Average Intersection-Union Ratio 83.7％ 85.1％ 86.1％ 87.4％ Classification Average Intersection-Union Ratio 80.4％ 81.9％ 84.6％ 85.8％

[0080] In summary, the semantic segmentation method for single object point cloud data using contour point features proposed in the embodiment of the present invention has the following beneficial effects:

[0081] 1. By extracting features from the contour points of a single object point cloud, the burden on the neural network can be reduced, making it easier to extract high-purity semantic features. On the ShapeNet Parts dataset, compared to existing neural network-based point cloud semantic segmentation methods, it can improve the part mean intersection-over-union ratio by 1.3% and the class mean intersection-over-union ratio by 1.2%.

[0082] 2. Based on the maximum central angle contour point extraction technology, it can effectively extract clear and complete contour points from the two-dimensional point set;

[0083] 3. The feature fusion module used can extract features with higher purity by fusing the features of single object point cloud and contour points, which is conducive to improving the final semantic segmentation accuracy.

[0084] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0085] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0086] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A semantic segmentation method for single object point cloud data using contour point features, characterized in that: The following steps are involved: Step S1: obtain a single object point cloud containing N points, and project the single object point cloud in the X, Y, and Z directions respectively to obtain three two-dimensional point sets, which are respectively recorded as 、 、 ; Step S2, obtaining a contour point set of a single object point cloud from the three two-dimensional point sets; Step S3, inputting the single object point cloud and the contour point set into a geometric feature fusion network to obtain a classification label for each point in the single object point cloud; The step S2 specifically includes: Step S21: Obtain the corresponding contour point sets according to the three two-dimensional point sets. 、 、 ; Step S22: The contour point set 、 、 Project the contour points in back to the original three-dimensional space, and merge the three point sets to obtain the contour point set ; The geometric feature fusion network includes four feature fusion modules and three feature upsampling modules, wherein each feature fusion module includes three feature extraction modules; The step S3 specifically includes: Step S31, inputting the single object point cloud and the contour point set into the four feature fusion modules of the geometric feature fusion network in sequence to obtain feature matrices at each level; Step S32, starting with the highest-level feature matrix, it is sequentially input into three feature acquisition modules together with the lower-level feature matrix, and finally the classification label of each point is obtained.

2. The semantic segmentation method of single object point cloud data using contour point features according to claim 1, characterized in that: The step S21 specifically includes: Step S211, select the two-dimensional point set Any point in , find the two-dimensional point set Middle distance In the preset range Neighborhood point set within ; Step S212: Select the neighborhood point set Any point in ,by As a starting point, Draw rays ; Step S213: For any two adjacent rays among the multiple rays obtained in step S212, each of them has an angle, and the maximum value of all angles is recorded as ; Step S214, determining the maximum value Is it greater than the preset threshold? If it is greater than Add the contour point set , otherwise not added.

3. The semantic segmentation method of single object point cloud data using contour point features according to claim 1, characterized in that: In step S22, if the three contour point sets are combined and exceed points, then randomly sample Points, if not enough points, then repeatedly sample the points in the set to make up points.

4. The method for semantic segmentation of single object point cloud data using contour point features according to claim 1, characterized in that: Any feature fusion module in step S31 specifically includes: Step S311, determine the coordinate-feature matrix of the subset points of the single object point cloud and the coordinate-feature matrix of the subset points of the contour point set As the input of the feature fusion module; Step S312: sampling from a subset of the single object point cloud points, denoted as , sampled from a subset of the contour point set points, denoted as ; Step S313: and Input into the feature extraction module and get Midpoint from The extracted feature matrix is concatenated to obtain the concatenated features; Step S314: input the splicing feature into a convolutional layer to obtain The characteristics of the points in, and the coordinate matrix Splicing to obtain the coordinate-feature matrix of the new subset of single object point cloud , as the first output of the feature fusion module; Step S315: the and Input into the feature extraction module and get Midpoint from Extracted features; Step S316: Combine the features obtained in step S315 with Splicing to obtain the coordinates of the new subset of contour points - feature matrix , as the second output of the feature fusion module.

5. The method for semantic segmentation of single object point cloud data using contour point features according to claim 4, characterized in that: The coordinate-feature matrix The size is , the coordinate-feature matrix for .

6. The method for semantic segmentation of single object point cloud data using contour point features according to claim 4, characterized in that: The step S313 specifically includes: Step S3131, determine the coordinate matrix of the subset points of the single object point cloud And the coordinate-feature matrix of the subset points of the single object point cloud is the input of the feature extraction module, where Size , Size , the two subsets are not the same set; Step S3132, find the coordinate matrix The closest Points, get The coordinate-feature matrix of the neighborhood point set of ; Step S3133: Input the neighborhood point set coordinate-feature matrix into the first three-layer convolution to obtain a dimension of The characteristic matrix of Step S3134: Input the neighborhood point set coordinate-feature matrix into the second three-layer convolution to obtain a dimension of The characteristic rotation matrix of Step S3135, multiplying the characteristic matrix obtained in step S3133 by the characteristic rotation matrix obtained in step S3134 to obtain a rotated characteristic matrix; Step S3136: input the rotated feature matrix into a convolution layer to obtain the output of the feature extraction module.

7. The method for semantic segmentation of single object point cloud data using contour point features according to claim 1, characterized in that: The step S32 specifically includes: Step S321: The coordinate matrix of the subset of the single object point cloud is And the coordinate-feature matrix of the subset points of the single object point cloud Input to the feature upsampling module, where the sizes of the two matrices are and ; Step S322: the coordinate matrix and the characteristic matrix Input into the feature extraction module and get Midpoint from Extracted features; Step S323: Combine the features obtained in step S322 with the coordinate matrix Splicing to get the output of the feature upsampling module.

Images