A method for indoor point cloud segmentation of buildings based on local feature enhanced PointNet++ network

Abstract

The present invention discloses a method for segmenting indoor point cloud of a building based on a local feature enhanced PointNet++ network, comprising the following steps: inputting indoor point cloud data of a three-dimensional building, removing noise points by statistical outlier removal; encoding the filtered indoor point cloud data of a three-dimensional building; using a hierarchical feature propagation algorithm of inverse distance weighted difference, performing difference processing on k adjacent points of each point in the point cloud according to inverse distance weighting, connecting the features obtained by the difference through skip‑link, and introducing a Droput layer to prevent overfitting, thereby performing segmentation information discrimination of points in the point cloud. The method of the present invention realizes the segmentation processing of semantic information of point clouds by training a network model; by processing noise points and deeply extracting local features of a point set, the segmentation result is made more accurate and stable.

Description

Technical Field

[0001] The present invention relates to the field of building technology, and in particular to a method for indoor point cloud segmentation of a building based on a local feature enhanced PointNet++ network. Background Art

[0002] With the development of engineering surveying technology, the acquisition of spatial big data is moving towards automation, digitization and intelligence. Three-dimensional spatial scanning technology has been increasingly widely used in many engineering fields due to its improvements in measurement accuracy, scanning speed, spatial resolution and other aspects and a significant reduction in price. In recent years, it has been widely used in cultural relics restoration, 3D printing, urban construction surveying, topographic mapping, mining operations, tunnel excavation, underground pipeline design, bridge construction and reconstruction, aircraft and ship manufacturing, highway and railway construction and other fields.

[0003] At present, the processing of 3D point cloud data relies heavily on human-computer interaction, and the degree of automation is low. In particular, removing noise points, extracting point cloud features, extracting reference planes, and automatically and accurately segmenting point clouds are important issues in point cloud processing, and are also bottleneck issues. Only by solving this series of technical problems can the subsequent storage and management of point cloud data, reconstruction of 2D points, lines, and surfaces, and reconstruction of 3D models be carried out smoothly.

[0004] The PointNet++ network based on deep learning is a network structure proposed in recent years for application in 3D point cloud segmentation. The network can directly take the point cloud as input, divide the point set into overlapping local areas using the distance metric in the space, extract local features from the geometric structure, and then expand the range to extract higher-level features based on the local features until the global features of the entire point set are extracted, thereby directly outputting the segmentation label of the entire point cloud. Therefore, the PointNet++ network has good network performance for processing 3D point cloud data.

[0005] However, since the PointNet++ network directly takes the points in the point cloud as input, it does not process the noise points and does not perform in-depth extraction of the local features of the point set, resulting in inaccurate and unstable segmentation results. Summary of the invention

[0006] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0007] In view of the above-mentioned and / or existing problems in the existing building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network, the present invention is proposed.

[0008] Therefore, the problem to be solved by the present invention is how to provide a building indoor point cloud segmentation method based on a local feature enhanced PointNet++ network.

[0009] To solve the above technical problems, the present invention provides the following technical solutions: a method for indoor point cloud segmentation of a building based on a local feature enhanced PointNet++ network, comprising the following steps: inputting three-dimensional indoor point cloud data of a building, removing noise points by statistical outlier removal; encoding the filtered three-dimensional indoor point cloud data of the building; using a hierarchical feature propagation algorithm of inverse distance weighted difference, performing difference processing on k adjacent points of each point in the point cloud according to inverse distance weighting, connecting the features obtained by the difference through skip-link, and introducing a Droput layer to prevent overfitting, thereby performing segmentation information discrimination on points in the point cloud.

[0010] As a preferred solution of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network described in the present invention, the encoding includes an abstraction layer, the abstraction layer is formed after the PointNet++ network is designed, and the abstraction layer includes a sampling layer, a grouping layer and a feature layer; the abstraction layer uses the farthest point sampling to perform hierarchical extraction on the point cloud data and retains the features of each layer to ensure that the sampling points are evenly distributed on the entire point cloud set.

[0011] As a preferred solution of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network described in the present invention, the method of removing noise points by using statistical outliers includes the following steps: finding all neighboring points of each point, calculating the distance from each point to its neighboring points; calculating the mean μ and standard deviation σ of the distance between all points and their neighbors according to the Gaussian distribution modeling distance parameter; calculating the mean distance μ between each point and its neighbors i ; Traverse all points, if their distance mean is within the confidence interval of the Gaussian distribution, retain them, otherwise remove them.

[0012] As a preferred solution of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network described in the present invention, the mean μ and standard deviation σ of the distance between all points and their neighbors are calculated by the following formula:

[0013]

[0014]

[0015] Where i = [1, 2, 3, ..., m] means there are m points in total, j = [1, cdots, k] means each point has k neighbors;

[0016] The mean distance μ between each point and its neighbors i Calculated by the following formula:

[0017]

[0018] The confidence interval of the Gaussian distribution is μ i ∈[μ-3σ,μ+3σ].

[0019] As a preferred solution of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network of the present invention, encoding the filtered three-dimensional building indoor point cloud data includes the following steps:

[0020] Set the sampling layer, sample the input points to obtain local features, and use the k-mediods clustering algorithm to obtain the best center point set;

[0021] Set the grouping layer, apply K nearest neighbor search and ball query algorithm to the input sampling points, create multiple sub-point clouds for the neighborhood points within a given radius, and divide the center point set into several regions;

[0022] Normalize the point coordinates in the region, encode each region with the node as the center, and transform the points in the region into relative feature coordinates around the center point, that is, C i-1 →C i , d+C i-1 →C i ;

[0023] Set the feature layer, input the set point set as N′×K×(d+C), extract the local features in the sampling area, convert the result into a structure of N′×(d+C′), and obtain the point cloud containing local spatial features after downsampling, completing the SA1 layer downsampling and feature extraction;

[0024] The above operations are repeated twice in sequence to complete the downsampling and feature extraction of the SA2 layer and SA3 layer, and three groups of point clouds containing local spatial features after sampling are obtained.

[0025] As a preferred solution of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network described in the present invention, obtaining the best center point set includes the following steps:

[0026] Obtain indoor point cloud data of three-dimensional buildings, and randomly select k objects as representative objects of the initial cluster center point set;

[0027] According to the Euclidean distance of the three-dimensional space coordinates, each remaining data point is assigned to the nearest cluster according to its distance to the representative object;

[0028] Calculate each point (x i ,y i ,z i ) to other points in the cluster (x j ,y j ,z j ), that is,

[0029] Select the point with the smallest distance and the smallest center as the new center of the cluster, and repeat this process to obtain a new set of cluster centers, that is, the optimal set of center points.

[0030] As a preferred solution of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network of the present invention, wherein: dividing the central point set into several regions includes the following steps:

[0031] Input the 3D point cloud data of the area to be divided and calculate the distance between it and each point set in the known area;

[0032] Find all points within the radius of the query point, and set the K sample data of the divided area that are closest to the point cloud samples of the area to be divided as the neighbors of the point cloud data of the area to be divided;

[0033] According to the categories to which most of the point sets in the K nearest neighbors belong, the region to which the 3D point cloud data of the area to be divided belongs is divided;

[0034] The grouping layer takes the sampling point set as the center, selects K neighbors as representatives in the local area smaller than the radius r parameter near each sampling point, aggregates the input point set N×(d+C), and the output point set is N′×K×(d+C).

[0035] As a preferred solution of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network described in the present invention, when setting the feature layer, when the density of the collected points in different areas is uneven, a multi-resolution grouping method is used for processing, which specifically includes the following steps:

[0036] The features of each layer are input in turn, which are composed of two vectors. The first vector sub-region contains sparse points, and the geometric abstraction layer is used to summarize the features of the sub-region. The second vector directly uses the feature layer to extract features from the original point cloud, and different weights are assigned to both sides according to the point cloud density.

[0037] As a preferred solution of the method for indoor point cloud segmentation of buildings based on the local feature enhanced PointNet++ network described in the present invention, a hierarchical feature propagation algorithm of inverse distance weighted difference is adopted, and k adjacent points of each point in the point cloud are subjected to difference processing according to inverse distance weighting, and the features obtained by the difference are connected through skip-link, and a Droput layer is introduced to prevent overfitting, so as to discriminate the segmentation information of the points in the point cloud, which specifically includes the following steps:

[0038] Input the point set N′×(d+C′) of the SA3 layer, take k=3 and perform interpolation according to the inverse distance weighting to restore the point cloud scale;

[0039] The point cloud and features obtained by the SA3 layer are restored to the point cloud processed by the SA2 layer through upsampling;

[0040] For each point in the SA2 layer, find the three nearest points in the corresponding SA3 layer, and use the inverse distance weighted difference of the features of these three points as the difference features of the point;

[0041] The difference features are geometrically connected with the point features obtained in the SA2 layer, and the point features of the feature aggregation layer FP3 are obtained by using a multi-layer perceptron network;

[0042] The point features of the FP3 layer are used as the difference part in the next layer FP2, and the difference features are obtained by using the inverse distance weighted difference, which are connected with the point features of the SA1 layer, and then the point features of the FP2 layer are obtained through MLP;

[0043] The point features of the FP2 layer are used as the difference part in the next FP1 layer, and the difference features are obtained by inverse distance weighted difference. They are connected with the original point cloud, and then the point features of the FP1 layer are obtained through MLP.

[0044] The features obtained by the difference are stacked and connected through skip-link, and the Dropout layer is introduced to prevent the network from overfitting, and the segmentation information of the points in the point cloud is judged and finally output.

[0045] As a preferred solution of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network of the present invention, the point cloud scale is restored by the following formula:

[0046]

[0047] The beneficial effects of the present invention are as follows: by training the network model, the segmentation processing of the semantic information of the point cloud is realized; by processing the noise points and deeply extracting the local features of the point set, the segmentation result is made more accurate and stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. 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 labor. Among them:

[0049] Figure 1 The technical route of the building indoor point cloud segmentation method is based on the local feature enhanced PointNet++ network.

[0050] Figure 2 This is the framework structure of the traditional PointNet++ point cloud segmentation network.

[0051] Figure 3 This is a visualization of the original point cloud of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network.

[0052] Figure 4 Visualization of point cloud data after removing noise points preprocessing for the building indoor point cloud segmentation method based on local feature enhancement PointNet++ network.

[0053] Figure 5 The local feature enhanced PointNet++ network structure is used for the indoor point cloud segmentation method of buildings based on the local feature enhanced PointNet++ network.

[0054] Figures 6 to 8 This is a visualization of the segmentation results of the test point cloud of the building indoor point cloud segmentation method based on the local feature enhanced PointNet++ network. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0056] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0058] Example 1

[0059] Reference Figure 1 to Figure 5 , which is the first embodiment of the present invention, and provides a method for indoor point cloud segmentation of a building based on a local feature enhanced PointNet++ network. The method for indoor point cloud segmentation of a building based on a local feature enhanced PointNet++ network comprises the following steps:

[0060] S1, input 3D building indoor point cloud data, and remove noise using statistical outlier removal;

[0061] S2, encoding the filtered three-dimensional building indoor point cloud data;

[0062] S3. A hierarchical feature propagation algorithm of inverse distance weighted difference is used to perform difference on the k neighboring points of each point in the point cloud according to inverse distance weighted (IDW). The features obtained by the difference are connected through skip-link, and a Droput layer is introduced to prevent overfitting, so as to perform segmentation judgment on the points in the point cloud.

[0063] In step S2, the encoding includes an abstraction layer (setabstraction), which is formed by the PointNet++ network design. Each abstraction layer (setabstraction) consists of three operation layers: sampling layer (sampling), grouping layer (grouping) and feature layer (pointnet). It uses the farthest point sampling (FPS) to extract the point cloud data in layers and retain the features of each layer, and ensure that the sampling points are evenly distributed on the entire point cloud set.

[0064] In step S1, removing noise points by using statistical outlier removal includes the following steps:

[0065] S11, find all neighboring points of each point, and calculate the distance from each point to its neighboring points;

[0066] S12. According to the Gaussian distribution model distance parameter, the mean μ and standard deviation σ of the distance between all points and their neighbors are calculated, expressed as:

[0067]

[0068]

[0069] Where i = [1, 2, 3, ..., m] means there are m points in total, j = [1, cdots, k] means each point has k neighbors;

[0070] S13, calculate the mean distance μ between each point and its neighbors i It is calculated by the following formula,

[0071]

[0072] S14. Traverse all points, and if their distance mean is within the confidence interval of the Gaussian distribution, then retain it, that is, μ i ∈[μ-3σ,μ+3σ], otherwise removed.

[0073] Through the above method, a visualization of the point cloud data after removing the noise points can be obtained, such as Figure 3 shown.

[0074] In step S2, the following steps are specifically included:

[0075] S21, set the sampling layer, sample the input points to obtain local features, use the k-mediods clustering algorithm to get the best center point set, N i-1 →N i Specifically,

[0076] (1) Obtain indoor point cloud data of three-dimensional buildings and randomly select k objects as representative objects of the initial cluster center point set;

[0077] (2) According to the Euclidean distance of the three-dimensional space coordinates, each remaining data point is assigned to the nearest cluster according to its distance to the representative object;

[0078] (3) Calculate each point (x i ,y i ,z i ) to other points in the cluster (x j ,y j ,z j ), that is,

[0079] (4) Select the point with the smallest distance and the smallest center as the new center of the cluster, and repeat this process to obtain a new set of cluster centers, that is, the optimal center point set. Assume that the input of the setabstraction module is N×(d+C), where N represents the number of points, d is the dimension, and C is the feature dimension. After the sampling layer, the coordinates of the sampling points will be N′×d, where N′ represents the number of points after sampling and d is the dimension.

[0080] S22, set the grouping layer, apply K nearest neighbor search (KNN) and ball query algorithm to the input sampling points, create multiple sub-point clouds for adjacent points within a given radius, and divide the central point set into several regions. This step includes the following steps:

[0081] (1) Calculate the distance. Input the 3D point cloud data of the area to be divided and calculate the distance between it and each point set in the known area;

[0082] (2) Find the neighborhood. Find all points within the radius of the query point and set the K sample data of the divided area that are closest to the point cloud samples of the area to be divided as the neighbors of the point cloud data of the area to be divided;

[0083] (3) Divide the region. Divide the region to which the 3D point cloud data to be divided belongs according to the category to which most of the point sets in the K nearest neighbors belong;

[0084] (4) The grouping layer takes the N′×d sampling point set as the center, selects K neighbors as representatives in the local area smaller than the radius parameter near each sampling point, aggregates the input point set N×(d+C), and the output point set is N′×K×(d+C);

[0085] S23, normalize the point coordinates in the region, encode each region with the node as the center, and convert the points in the region into relative feature coordinates around the center point, that is, C i-1 →C i , d+C i-1 →C i ;

[0086] S24, set the feature layer, input the set point set as N′×K×(d+C), extract the local features in the sampling area, convert the result into a structure of N′×(d+C′), obtain the point cloud containing local spatial features after downsampling, and complete the SA1 layer downsampling and feature extraction;

[0087] S25. Repeat the above operations twice in sequence to complete the downsampling and feature extraction of the SA2 layer and the SA3 layer, and obtain three groups of point clouds containing local spatial features after sampling.

[0088] When setting the feature layer, when the density of collected points in different areas is uneven, a multi-resolution grouping method is used for processing, which specifically includes the following steps: input the features of each layer in turn, which are composed of two vectors. The first vector sub-region contains sparse points, and the geometric abstraction layer is used to summarize the features of the sub-region. The second vector directly uses the feature layer to extract features from the original point cloud, and different weights are assigned to both sides according to the density of the point cloud.

[0089] In step S3, it specifically includes the following steps:

[0090] S31, input the point set N′×(d+C′) of the SA3 layer, take k=3 and perform difference according to the inverse distance weighting (IDW) to restore the point cloud scale, that is,

[0091] S32, restore the point cloud and features obtained by the SA3 layer to the point cloud processed by the SA2 layer through upsampling;

[0092] S33, for each point in the SA2 layer, find the three closest points in the corresponding SA3 layer, and use the inverse distance weighted difference of the features of these three points as the difference features of the point;

[0093] S34, geometrically connect the difference feature with the point feature obtained in the SA2 layer, and obtain the point feature of the feature aggregation layer FP3 layer by using a multi-layer perceptron network;

[0094] S35, taking the point features of the FP3 layer as the difference part in the next layer FP2, using the inverse distance weighted difference to obtain the difference features, connecting them with the point features of the SA1 layer, and then obtaining the point features of the FP2 layer through MLP;

[0095] S36, taking the point features of the FP2 layer as the difference part in the next FP1 layer, obtaining the difference features by using the inverse distance weighted difference, connecting with the original point cloud, and then obtaining the point features of the FP1 layer by MLP;

[0096] S37. The features obtained by the difference are stacked and connected through skip-link, and the Dropout layer is introduced to prevent the network from overfitting, and the segmentation information of the points in the point cloud is judged and finally output.

[0097] It should be noted that the framework structure of the traditional PointNet++ network is as follows Figure 2 As shown in the figure, since the point cloud is only downsampled twice, that is, there are only two layers of SA in the encoding layer and FP in the decoding layer, the local spatial information contained in each point is insufficient, and the local features of the point set cannot be deeply extracted, resulting in the segmentation result being inaccurate and unstable.

[0098] like Figure 5As shown, the local feature enhanced PointNet++ network of the present invention changes the original network encoding and decoding structure from 2 layers to 3 layers, namely, encoding layers SA1, SA2, SA3, and decoding layers FP1, FP2, FP3. This will facilitate more effective feature extraction and learning after the network depth is enhanced, so as to obtain better point cloud segmentation effects. The input point cloud itself has the feature d+c. Through the SA module of the first layer, the local spatial features of the points are preliminarily aggregated, and then through the remaining two SA modules, the spatial information aggregation encoding module of the point cloud is completed. After that, the point cloud feature decoding process is carried out, through the three FP feature transfer layers, and the corresponding skip link connection is performed in the encoding layer to obtain the feature information of each point. At this time, the features include its own features and the features of the surrounding neighborhood. Finally, the features of each point are passed through the fully connected layer and the DP layer to complete the segmentation prediction of each point.

[0099] As the network depth increases, the feature dimension of the point cloud increases, resulting in a large number of MLP structures in the network. The network parameters also increase accordingly. In order to avoid overfitting of the network, the Dropout layer is introduced in the network training to enhance the stability of the entire network.

[0100] In summary, the present invention uses the local features of the sampling layer, grouping layer and feature layer (pointnet), adopts the hierarchical feature propagation algorithm of inverse distance weighted difference, and performs difference on the k adjacent points of each point in the point cloud according to the inverse distance weighted (IDW). The feature point set obtained by the difference is connected through skip-link, and the Droput layer is introduced to prevent overfitting, so as to distinguish the segmentation information of the points in the point cloud.

[0101] Example 2

[0102] Reference Figure 6 to Figure 8 , which is the second embodiment of the present invention. This embodiment uses the real 3D building indoor point cloud dataset as the main data. In order to facilitate the comparison with the network prediction result, the test point cloud is colored according to the test visualization color standard. 8 categories of visualization RGB color standards are set. The point cloud after the test point cloud is visualized after denoising is shown as follows Figure 5 After training, the local feature enhancement PointNet++ network is used to perform semantic segmentation on the point cloud. The results are shown in Figure 6 to Figure 8 shown.

[0103] The accuracy evaluation indicators of the final network's three-dimensional building indoor point cloud segmentation results are shown in Tables 1 and 2 below.

[0104] Table 1 Overall accuracy (OA) and category intersection over union (IoU) evaluation results of each network

[0105] network OA IoU RANSAC 60.1％ 29.9％ MVCNNN 69.8％ 37.7％ PointNet 77.6％ 41.8％ PointNet++ 82.3％ 48.6％ Local feature enhancement PointNet++ 84.5％ 55.3％

[0106] Table 2 Intersection over Union (IoU) evaluation results of each segmentation region category of each network

[0107]

[0108] The overall accuracy (OA) of the improved local feature enhanced PointNet++ network is 84.5%, which is 1.2% higher than that of PointNet++. For the intersection-over-union (IoU) of each segmented region category, we can see that the IoU value of the local feature enhanced PointNet++ network is 51.3%, which is 6.7% higher than that of PointNet++. It also exceeds the PointNe++t network in some categories that originally performed poorly, such as ceilings, tables, partitions, and walls. Among them, the IoU value of the ceiling is 89.5.0%, the IoU value of the table is 77.3%, the IoU value of the column is 3.1%, the IoU value of the partition is 15.0%, the IoU value of wall 1 is 73.4%, the IoU value of wall 2 is 74.5%, the IoU value of wall 3 is 74.3%, and the IoU value of wall 4 is 75.2%.

[0109] From the analysis of the above segmentation results, we can see that by training the network model, the segmentation processing of the semantic information of the point cloud is realized. This network has certain advantages over other algorithms in terms of the overall segmentation accuracy and the intersection-over-union ratio of each segmented area category. It is an effective point cloud segmentation method.

[0110] 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 preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for indoor point cloud segmentation of buildings based on local feature enhanced PointNet++ network, characterized by: The following steps are included: Input 3D building indoor point cloud data and use statistical outlier removal to remove noise; Encoding the filtered three-dimensional building indoor point cloud data; The hierarchical feature propagation algorithm of inverse distance weighted difference is adopted to perform difference processing on the k adjacent points of each point in the point cloud according to the inverse distance weighted difference. The features obtained by the difference are connected through skip-link, and the Dropout layer is introduced to prevent overfitting, so as to discriminate the segmentation information of the points in the point cloud. The encoding of the filtered 3D building indoor point cloud data includes the following steps: Set the sampling layer, sample the input points to obtain local features, and use the k-mediods clustering algorithm to obtain the best center point set; Set the grouping layer, apply K nearest neighbor search and ball query algorithm to the input sampling points, create multiple sub-point clouds for the neighborhood points within a given radius, and divide the center point set into several regions; Normalize the point coordinates in the region, encode each region with the node as the center, and transform the points in the region into relative feature coordinates around the center point, that is, C i-1 →C i , d+C i-1 →C i ; Set the feature layer, input the set point set as N′×K×(d+C), extract the local features in the sampling area, convert the result into a structure of N′×(d+C′), and obtain the point cloud containing local spatial features after downsampling, completing the SA1 layer downsampling and feature extraction; Repeat the above steps of setting the sampling layer to setting the feature layer twice, complete the downsampling and feature extraction of the SA2 layer and SA3 layer, and obtain 3 groups of point clouds containing local spatial features after sampling; Obtaining the best center point set includes the following steps: Obtain indoor point cloud data of three-dimensional buildings, and randomly select k objects as representative objects of the initial cluster center point set; According to the Euclidean distance of the three-dimensional space coordinates, each remaining data point is assigned to the nearest cluster according to its distance to the representative object; Calculate each point (x i ,y i ,z i ) to other points in the cluster (x j ,y j ,z j ), that is, Select the point with the smallest distance and the smallest center as the new center of the cluster, and repeat this process to obtain a new set of cluster centers, that is, the optimal center point set; Dividing the center point set into several regions includes the following steps: Input the 3D point cloud data of the area to be divided and calculate the distance between it and each point set in the known area; Find all points within the radius of the query point, and set the K sample data of the divided area that are closest to the point cloud samples of the area to be divided as the neighbors of the point cloud data of the area to be divided; According to the categories to which most of the point sets in the K nearest neighbors belong, the region to which the 3D point cloud data of the area to be divided belongs is divided; The grouping layer takes the sampling point set as the center, selects K neighbors as representatives in the local area around each sampling point that is smaller than the radius r parameter, aggregates the input point set N×(d+C), and the output point set is N′×K×(d+C); When setting the feature layer, when the density of collected points is uneven in different areas, a multi-resolution grouping method is used for processing, which specifically includes the following steps: Input the features of each layer in turn, which are composed of two vectors. The first vector sub-region contains sparse points, and the geometric abstraction layer is used to summarize the features of the sub-region. The second vector directly uses the feature layer to extract features from the original point cloud, and different weights are assigned to both sides according to the point cloud density. The hierarchical feature propagation algorithm of inverse distance weighted difference is adopted to perform difference processing on the k adjacent points of each point in the point cloud according to the inverse distance weighted difference. The features obtained by the difference are connected through skip-link, and the Droput layer is introduced to prevent overfitting, so as to discriminate the segmentation information of the points in the point cloud. The specific steps include the following: Input the point set N′×(d+C′) of the SA3 layer, take k=3 and perform interpolation according to the inverse distance weighting to restore the point cloud scale; The point cloud and features obtained by the SA3 layer are restored to the point cloud processed by the SA2 layer through upsampling; For each point in the SA2 layer, find the three nearest points in the corresponding SA3 layer, and use the inverse distance weighted difference of the features of these three points as the difference features of the point; The difference features are geometrically connected with the point features obtained in the SA2 layer, and the point features of the feature aggregation layer FP3 are obtained by using a multi-layer perceptron network; The point features of the FP3 layer are used as the difference part in the next layer FP2, and the difference features are obtained by using the inverse distance weighted difference, which are connected with the point features of the SA1 layer, and then the point features of the FP2 layer are obtained through MLP; The point features of the FP2 layer are used as the difference part in the next FP1 layer, and the difference features are obtained by inverse distance weighted difference. They are connected with the original point cloud, and then the point features of the FP1 layer are obtained through MLP. The features obtained by the difference are stacked and connected through skip-link, and the Dropout layer is introduced to prevent the network from overfitting, and the segmentation information of the points in the point cloud is judged and finally output; The point cloud scale is restored using the following formula: In the formula, i = [1, 2, 3, ..., m] means there are m points in total, and j = [1, cdots, k] means each point has k neighbors.

2. The method for indoor point cloud segmentation of a building based on a local feature enhanced PointNet++ network as claimed in claim 1, characterized in that: The encoding includes an abstraction layer, which is formed by the PointNet++ network design, and the abstraction layer includes a sampling layer, a grouping layer and a feature layer; The abstract layer uses farthest point sampling to perform hierarchical extraction of point cloud data and retains the features of each layer to ensure that the sampling points are evenly distributed on the entire point cloud set.

3. The method for indoor point cloud segmentation of a building based on a local feature enhanced PointNet++ network as claimed in claim 1 or 2, characterized in that: Using statistical outlier removal to remove noise includes the following steps: Find all neighboring points of each point and calculate the distance from each point to its neighboring points; According to the Gaussian distribution model distance parameter, the mean μ and standard deviation σ of the distance between all points and their neighbors are calculated; Calculate the mean distance μ between each point and its neighbors i ; Traverse all points, if their distance mean is within the confidence interval of the Gaussian distribution, then retain it, otherwise remove it.

4. The method for indoor point cloud segmentation of a building based on a local feature enhanced PointNet++ network as claimed in claim 3, characterized in that: The mean μ and standard deviation σ of the distance between all points and their neighbors are calculated by the following formula: Where i = [1, 2, 3, ..., m] means there are m points in total, j = [1, cdots, k] means each point has k neighbors; The mean distance μ between each point and its neighbors i Calculated by the following formula: The confidence interval of the Gaussian distribution is μ i ∈[μ-3σ,μ+3σ].

Images