Intelligent space point cloud three-dimensional reconstruction method and system based on occupied grid

Abstract

The invention discloses a smart space point cloud three-dimensional reconstruction method and system based on occupied grids. Comprising the following steps: inputting an open source point cloud, obtaining a potential vector through a point cloud convolution module, and segmenting a bounding box of the point cloud p to obtain query points; for each query point, inputting the potential vector and the query point, performing relative coding to obtain a relative potential vector, inputting the relative potential vector into an attention weight module to obtain a feature code, and performing low-dimensional feature decoding on the feature code to obtain the occupancy rate of the query point; and extracting the query points of which the occupancy rate is a predefined numerical value to obtain a contour surface, and carrying out an advancing cube modeling algorithm on the contour surface to obtain a required three-dimensional model. The method can get rid of the dependence of an existing method on the normal vector of the point cloud so as to reduce the calculation and use cost, improve the modeling effect of a complex object or scene, and increase the expandability of the method, so that the method can be applied to various scenes.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision, deep learning, and three-dimensional reconstruction, and in particular to a method and system for three-dimensional reconstruction of intelligent space point clouds based on occupancy grids. Background Art

[0002] With the development of society, people pay more and more attention to the comfort and safety of the space where they live and work, and space managers also attach great importance to the efficiency of space use. The emergence of the concept of "smart space" responds to the concerns of contemporary people. Smart space builds an intelligent physical space that can perceive and connect through the comprehensive application of contemporary advanced technologies such as the Internet of Things (IoT), artificial intelligence (AI), big data, and cloud computing. It can monitor, analyze and respond to environmental changes and human activities in the space in real time, as well as the interaction between the two. The construction of smart space requires accurate spatial data as a basis, and point cloud 3D reconstruction technology provides technical support for obtaining and constructing the spatial data required for smart space. This is a method of generating a 3D model by collecting and processing point cloud data. It uses laser scanning, photogrammetry and other means to accurately obtain the 3D data of the physical space; through point cloud data processing and reconstruction algorithms, a high-precision 3D model is generated to reflect the structure and characteristics of the real physical space, thus laying the foundation for realizing the further functions of the smart space.

[0003] In recent years, deep learning has played an important role in the field of point cloud 3D reconstruction. Using deep neural networks and large-scale datasets, researchers are able to directly infer the 3D shape and structure information of point clouds from single point cloud data. These methods have broad application potential in the field of point cloud 3D reconstruction.

[0004] Existing point cloud 3D reconstruction methods based on deep learning are mostly targeted at a certain type of object or a certain scene, and lack scalability. In addition, existing methods have high requirements for point cloud data. In the modeling process, the normal vector information of the point cloud is required as an auxiliary. However, the acquisition of point clouds is usually not accompanied by normal vector information, so additional calculation of the normal vector of the point cloud is required, which increases the cost of use. In addition, in the process of deep learning of point clouds, features are extracted from the entire point cloud, lacking local structured features, resulting in poor deep learning effects. For objects with complex structures, multiple local modeling and splicing are required.

[0005] One of the current existing technologies is the patent "Constrained reconstruction method of building point cloud based on global alignment (CN117078852A)". The steps of this method include: 1. Estimating the initial normal of the input point cloud; 2. Sparsely optimizing the estimated initial normal so that points on the same surface have the same normal; 3. Combining the initial estimated normal and the optimized normal, using the region growing algorithm to segment the input point cloud to obtain a series of plane primitives; 4. Calculating the main axis direction of the building point cloud as a global reference, setting the target normal for each plane primitive to construct the parallel, vertical, coplanar and symmetric relationships in the plane primitive; 5. Globally optimize the direction and position of the plane primitive according to the constructed constraint relationship; 6. Calculate the boundary of the plane primitive by intersecting all planes to obtain the final model. The disadvantage of this method is that it requires additional calculation of the normal vector information of the point cloud, which not only increases the cost of use, but also makes it difficult to calculate the normal vector of the point cloud, and it is difficult to accurately calculate the real normal vector, resulting in poor modeling effect. In addition, this method is only applicable to buildings.

[0006] The second current prior art is the patent "A method and device for three-dimensional reconstruction of sparse point clouds (CN118229893A)". The steps of this method include: after the standardized flow training is completed, based on the initial first latent code, an inverse process of the standardized flow network is performed to obtain the current second latent code, and the first signed distance function value is obtained through the target automatic decoder; then, based on the first signed distance function value, the initial first latent code is updated and optimized to obtain the best first latent code, and based on the best first latent code, a secondary inverse process of the standardized flow network is performed to obtain the best second latent code, and then the second signed distance function value is obtained through the target automatic decoder; finally, three-dimensional reconstruction is performed based on the second signed distance function value to obtain the three-dimensional model corresponding to the target sparse point cloud data. The disadvantage of this method is that it uses a signed distance function for deep learning, and also requires normal vector information as an aid. In addition, during its deep learning process, low-dimensional latent features are extracted for the entire point cloud, and local features are not extracted, so the deep learning effect is poor. In addition, this method is only applicable to vehicle models. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the existing methods and propose a method and system for 3D reconstruction of point clouds in intelligent space based on occupancy grids. The main problem solved by the present invention is how to get rid of the reliance of the existing methods on the normal vectors of point clouds to reduce the calculation and use costs, improve the modeling effect of complex objects or scenes, and increase the scalability of the method so that it can be applied in a variety of scenarios.

[0008] In order to solve the above problems, the present invention proposes a method for 3D reconstruction of smart space point cloud based on occupancy grid, the method comprising:

[0009] Input the open source point cloud p, pass through the point cloud convolution module, and get the potential vector z p , divide the bounding box of point cloud p into several small cubes according to the resolution requirements, and the vertex of each small cube is the query point q;

[0010] For each query point q, input the latent vector z p And the query point q, after relative encoding, obtain the relative potential vector z p-qq , the relative potential vector z p-q Input the attention weight module to get the feature code z q , the feature encoding z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q;

[0011] Extract query points whose occupancy rate of the query point q is a predefined value (the predefined value is 0.5) to obtain an isosurface, and use the marching cube modeling algorithm to obtain the required three-dimensional model;

[0012] Preferably, the point cloud convolution module is specifically:

[0013] The point cloud convolution module includes: a one-dimensional convolution layer with an input dimension of 3 and an output dimension of 256, a one-dimensional convolution layer with an input dimension and an output dimension of 256, a ReLU activation layer, a one-dimensional convolution layer with an input dimension of 256 and an output dimension of 512, and a ReLU activation layer.

[0014] Preferably, the input potential vector z p And the query point q, after relative encoding, obtain the relative potential vector z p-q , specifically:

[0015] Using the query point q, the nearest 64 points in the neighborhood are searched from the point cloud p according to the K nearest neighbor algorithm as support points s;

[0016] According to the support point s, from the potential vector z p Take out the potential vector z corresponding to the support point s p-k ;

[0017] The potential vector z corresponding to the support point s p-k The relative potential vector z is obtained by concatenating the distance qs between the query point q and the support point s p-q .

[0018] Preferably, the relative potential vector z p-q Input the attention weight module to get the feature code z q , specifically:

[0019] The relative potential vector z p-q After three layers of two-dimensional convolutional layers with input and output dimensions of 512, the attention feature x is obtained;

[0020] Input the attention feature x into a two-dimensional convolutional layer with an input dimension of 512 and an output dimension of 1 and a two-dimensional convolutional layer with both input and output dimensions of 512 to obtain a query matrix and a value matrix;

[0021] The query matrix is ​​normalized by the SoftMax layer and then transposed, and then multiplied by the value matrix to obtain the feature code z q .

[0022] Preferably, the feature encoding z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q, specifically:

[0023] Encode the feature z q After being decoded into occupancy by the linear layer and the SoftMax layer, the cross entropy loss function is used in the training process to calculate the loss between the predicted occupancy and the true occupancy marked in the point cloud p. The loss function calculation formula is as follows:

[0024] CE(y′,y)=-ylog(y′),

[0025] Among them, y′ represents the predicted occupancy rate, and y represents the actual occupancy rate marked in the dataset

[0026] Accordingly, the present invention also provides an occupancy grid-based smart space point cloud 3D reconstruction system, comprising:

[0027] The preprocessing unit is used to input the open source point cloud p, and pass it through the point cloud convolution module to obtain the potential vector z p , divide the bounding box of point cloud p into several small cubes according to the resolution requirements, and the vertex of each small cube is the query point q;

[0028] A query point occupancy calculation unit is used to input the potential vector z for each query point q. p And the query point q, after relative encoding, obtain the relative potential vector z p-q , the relative potential vector z p-q Input the attention weight module to get the feature code z q , the feature encoding z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q;

[0029] The three-dimensional modeling unit is used to extract the query points whose occupancy rate of the query point q is a predefined value (the predefined value is 0.5) to obtain an isosurface, and use the marching cube modeling algorithm to obtain the required three-dimensional model.

[0030] Correspondingly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the steps of the above-mentioned method for identifying human thermal and humidity states and predicting health based on deep learning.

[0031] Correspondingly, the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned deep learning-based human body thermal and humidity state recognition and health prediction method are implemented.

[0032] The implementation of the present invention has the following beneficial effects:

[0033] The present invention extracts isosurfaces for modeling based on an occupancy grid algorithm, and does not require the assistance of normal vector information in the point cloud modeling process. The present invention performs feature extraction and latent vector encoding for a single point in the deep learning process, and fully utilizes the point information of the neighborhood of each query point in the process of predicting the occupancy rate of the query point, which makes the present invention have a good modeling effect on objects or scenes with complex structures, and is applicable to single objects, multiple objects, or large scenes, increasing the scalability of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of a method for three-dimensional reconstruction of a smart space point cloud based on an occupancy grid according to an embodiment of the present invention;

[0035] Figure 2 is a schematic diagram of a point cloud convolution module according to an embodiment of the present invention;

[0036] Figure 3 It is a structural diagram of an occupancy grid-based intelligent space point cloud three-dimensional reconstruction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0038] Figure 1 is a flow chart of a method for 3D reconstruction of a smart space point cloud based on an occupancy grid according to an embodiment of the present invention. Figure 1As shown, the method includes:

[0039] S1, input open source point cloud p, through the point cloud convolution module, get the potential vector z p , divide the bounding box of the point cloud p into several small cubes according to the resolution requirements, and the vertex of each small cube is the query point q; open source point cloud data includes ShapeNet dataset, Synthetic Rooms dataset, Thingi10k dataset, etc., where a single point cloud contains 3000 points, each point contains three coordinate data, that is, a single point cloud is a 3000×3 matrix; the potential vector z p is a 3000×512 matrix; the query point q is a 1×3 vector;

[0040] For each query point q, the following operations S2-S4 are performed:

[0041] S2, input the potential vector z p And the query point q, after relative encoding, obtain the relative potential vector z p-q ;

[0042] S3, the relative potential vector z p-q Input the attention weight module to get the feature code z q ;

[0043] S4, encode the feature z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q;

[0044] S5, extracting query points whose occupancy rate of the query point q is a predefined value (the predefined value is 0.5) to obtain an isosurface, and using the marching cubes modeling algorithm to obtain the required three-dimensional model.

[0045] Step S1, such as Figure 2 As shown, the details are as follows:

[0046] S1-1, the point cloud convolution module includes: a one-dimensional convolution layer with an input dimension of 3 and an output dimension of 256, a one-dimensional convolution layer with an input dimension and an output dimension of 256, a ReLU activation layer, a one-dimensional convolution layer with an input dimension of 256 and an output dimension of 512, and a ReLU activation layer.

[0047] Step S2 is as follows:

[0048] S2-1, using the query point q, searching the 64 nearest points in the neighborhood from the point cloud p as support points s (64×3 matrix) according to the K nearest neighbor algorithm;

[0049] S2-2, according to the support point s, from the potential vector zp Take out the potential vector z corresponding to the support point s p-k (64×512 matrix);

[0050] S2-3, the potential vector z corresponding to the support point s p-k The relative potential vector z is obtained by concatenating the distance qs between the query point q and the support point s p-q (64×515 matrix).

[0051] Step S3 is as follows:

[0052] S3-1, the relative potential vector z p-q After three layers of two-dimensional convolutional layers with input and output dimensions of 512, the attention feature x (64×512 matrix) is obtained;

[0053] S3-2, input the attention feature x into a two-dimensional convolution layer with an input dimension of 512 and an output dimension of 1 and a two-dimensional convolution layer with both input and output dimensions of 512, respectively, to obtain a query matrix (a 64×1 matrix) and a value matrix (a 64×512 matrix);

[0054] S3-3, normalize the query matrix through the SoftMax layer and then transpose it, and then multiply it with the value matrix to obtain the feature code z q (a 1×512 vector).

[0055] Step S4 is as follows:

[0056] S4-1, encode the feature z q After being decoded into occupancy rate (occupancy rate value is 0-1) by the linear layer and the SoftMax layer, the cross entropy loss function is used to calculate the loss between the predicted occupancy rate and the actual occupancy rate marked in the point cloud p during the training process. The loss function calculation formula is as follows:

[0057] CE(y′,y)=-ylog(y′),

[0058] Among them, y′ represents the predicted occupancy rate, and y represents the true occupancy rate marked in the dataset.

[0059] Step S5 is as follows:

[0060] S5-1, extracting query points with an occupancy rate of 0.5 of the query point q to obtain an isosurface with an occupancy rate of 0.5;

[0061] S5-2, the isosurface with occupancy rate of 0.5 is modeled by the marching cubes algorithm to obtain the required 3D model. The marching cubes algorithm is a classic and effective 3D surface reconstruction algorithm that can process various types of volume data and has good robustness and performance at different resolutions. Due to its simple and intuitive implementation and good results, the marching cubes algorithm has been widely used in the field of 3D modeling and visualization. The basic steps of the algorithm include the following four steps:

[0062] Cube segmentation: divide the three-dimensional space into a series of cubic units, each of which contains a volume data point or volume data value;

[0063] Isosurface extraction: traverse each cube unit and determine the isosurface (i.e. the surface where the volume data value is equal to or exceeds the threshold) by comparing the data value inside the cube unit with the preset threshold. This step usually uses some interpolation method to estimate the position of the isosurface;

[0064] Triangulation: For each cubic unit, the intersection point between the cubic unit and the isosurface is determined according to the position of the isosurface, and the intersection points are connected using the triangulation method to form a triangular mesh, which ultimately constitutes the mesh representation of the entire surface;

[0065] Surface mesh merging: Merge the triangular meshes of all cube units into an overall surface mesh to obtain the final three-dimensional surface model.

[0066] Accordingly, the present invention also provides a smart space point cloud 3D reconstruction system based on occupancy grid, such as Figure 3 As shown, including:

[0067] Preprocessing unit 1 is used to input the open source point cloud p, and pass through the point cloud convolution module to obtain the potential vector z p , divide the bounding box of point cloud p into several small cubes according to the resolution requirements, and the vertex of each small cube is the query point q;

[0068] Specifically, the point cloud convolution module includes: a one-dimensional convolution layer with an input dimension of 3 and an output dimension of 256, a one-dimensional convolution layer with an input dimension and an output dimension of 256, a ReLU activation layer, a one-dimensional convolution layer with an input dimension of 256 and an output dimension of 512, and a ReLU activation layer.

[0069] The query point occupancy calculation unit 2 is used to input the potential vector z for each query point q. p And the query point q, after relative encoding, obtain the relative potential vector z p-q , the relative potential vector z p-q Input the attention weight module to get the feature code zq , the feature encoding z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q;

[0070] Specifically, using the query point q, the nearest 64 points in the neighborhood are searched from the point cloud p according to the K nearest neighbor algorithm as support points s;

[0071] According to the support point s, from the potential vector z p Take out the potential vector z corresponding to the support point s p-k ;

[0072] The potential vector z corresponding to the support point s p-k The relative potential vector z is obtained by concatenating the distance qs between the query point q and the support point s p-q ;

[0073] The relative potential vector z p-q After three layers of two-dimensional convolutional layers with input and output dimensions of 512, the attention feature x is obtained;

[0074] Input the attention feature x into a two-dimensional convolutional layer with an input dimension of 512 and an output dimension of 1 and a two-dimensional convolutional layer with both input and output dimensions of 512 to obtain a query matrix and a value matrix;

[0075] The query matrix is ​​normalized by the SoftMax layer and then transposed, and then multiplied by the value matrix to obtain the feature code z q ;

[0076] Encode the feature z q After being decoded into occupancy by the linear layer and the SoftMax layer, the cross entropy loss function is used in the training process to calculate the loss between the predicted occupancy and the true occupancy marked in the point cloud p. The loss function calculation formula is as follows:

[0077] CE(y′,y)=-ylog(y′),

[0078] Among them, y′ represents the predicted occupancy rate, and y represents the true occupancy rate marked in the dataset.

[0079] The three-dimensional modeling unit 3 is used to extract the query points whose occupancy rate of the query point q is a predefined value (the predefined value is 0.5) to obtain an isosurface, and use the marching cube modeling algorithm to obtain the required three-dimensional model.

[0080] Therefore, the present invention extracts isosurfaces for modeling based on the occupancy grid algorithm, and does not require the assistance of normal vector information in the point cloud modeling process. The present invention performs feature extraction and latent vector encoding for a single point in the deep learning process, and fully utilizes the point information of the neighborhood of each query point in the process of predicting the occupancy rate of the query point, which makes the present invention have a good modeling effect on objects or scenes with complex structures, and is applicable to single objects, multiple objects, or large scenes, increasing the scalability of the method.

[0081] The above is a detailed introduction to the occupancy grid-based intelligent spatial point cloud three-dimensional reconstruction method and system provided by an embodiment of the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for 3D reconstruction of smart space point cloud based on occupancy grid, characterized in that: The method comprises: Input the open source point cloud p, pass through the point cloud convolution module, and get the potential vector z p , divide the bounding box of point cloud p into several small cubes according to the resolution requirements, and the vertex of each small cube is the query point q; For each query point q, input the latent vector z p And the query point q, after relative encoding, obtain the relative potential vector z p-q , the relative potential vector z p-q Input the attention weight module to get the feature code z q , the feature encoding z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q; The query points whose occupancy rates of the query points q are predefined values ​​are extracted to obtain isosurfaces, and the isosurfaces are modeled by marching cubes algorithm to obtain the required three-dimensional model.

2. The method for 3D reconstruction of smart space point cloud based on occupancy grid according to claim 1, characterized in that: The point cloud convolution module is specifically: The point cloud convolution module includes: a one-dimensional convolution layer with an input dimension of 3 and an output dimension of 256, a one-dimensional convolution layer with an input dimension and an output dimension of 256, a ReLU activation layer, a one-dimensional convolution layer with an input dimension of 256 and an output dimension of 512, and a ReLU activation layer.

3. The method for 3D reconstruction of smart space point cloud based on occupancy grid as claimed in claim 1, characterized in that: The input latent vector z p And the query point q, after relative encoding, obtain the relative potential vector z p-q , specifically: Using the query point q, the nearest 64 points in the neighborhood are searched from the point cloud p according to the K nearest neighbor algorithm as support points s; According to the support point s, from the potential vector z p Take out the potential vector z corresponding to the support point s p-k ; The potential vector z corresponding to the support point s p-k The relative potential vector z is obtained by concatenating the distance qs between the query point q and the support point s p-q .

4. The method for 3D reconstruction of smart space point cloud based on occupancy grid as claimed in claim 1, characterized in that: The relative potential vector z p-q Input the attention weight module to get the feature code z q , specifically: The relative potential vector z q-q After three layers of two-dimensional convolutional layers with input and output dimensions of 512, the attention feature x is obtained; Input the attention feature x into a two-dimensional convolutional layer with an input dimension of 512 and an output dimension of 1 and a two-dimensional convolutional layer with both input and output dimensions of 512 to obtain a query matrix and a value matrix; The query matrix is ​​normalized by the SoftMax layer and then transposed, and then multiplied by the value matrix to obtain the feature code z q .

5. The method for 3D reconstruction of smart space point cloud based on occupancy grid as claimed in claim 1, characterized in that: The feature encoding z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q, specifically: Encode the feature z q After being decoded into occupancy by the linear layer and the SoftMax layer, the cross entropy loss function is used in the training process to calculate the loss between the predicted occupancy and the true occupancy marked in the point cloud p. The loss function calculation formula is as follows: CE(y′,y)=-ylog(y′), Among them, y′ represents the predicted occupancy rate, and y represents the true occupancy rate marked in the dataset.

6. A smart space point cloud 3D reconstruction system based on occupancy grid, characterized in that: The system comprises: The preprocessing unit is used to input the open source point cloud p, and pass it through the point cloud convolution module to obtain the potential vector z p , divide the bounding box of point cloud p into several small cubes according to the resolution requirements, and the vertex of each small cube is the query point q; A query point occupancy calculation unit is used to input the potential vector z for each query point q. p And the query point q, after relative encoding, obtain the relative potential vector z p-q , the relative potential vector z p-q Input the attention weight module to get the feature code z q , the feature encoding z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q; The three-dimensional modeling unit is used to extract the query points whose occupancy rate of the query point q is a predefined value to obtain an isosurface, and use the marching cube modeling algorithm to obtain the required three-dimensional model.

7. The occupancy grid-based intelligent space point cloud 3D reconstruction system according to claim 6, characterized in that: The preprocessing unit is used to input the open source point cloud p, and obtain the potential vector z through the point cloud convolution module. p , divide the bounding box of point cloud p into several small cubes according to the resolution requirements, and the vertex of each small cube is the query point q, specifically: The point cloud convolution module includes: a one-dimensional convolution layer with an input dimension of 3 and an output dimension of 256, a one-dimensional convolution layer with an input dimension and an output dimension of 256, a ReLU activation layer, a one-dimensional convolution layer with an input dimension of 256 and an output dimension of 512, and a ReLU activation layer.

8. The occupancy grid-based intelligent space point cloud 3D reconstruction system according to claim 6, characterized in that: The query point occupancy calculation unit is used to input the potential vector z for each query point q. p And the query point q, after relative encoding, obtain the relative potential vector z p-q , the relative potential vector z p-q Input the attention weight module to get the feature code z p , the feature encoding z q Perform low-dimensional feature decoding to obtain the occupancy rate of the query point q, specifically: Using the query point q, the nearest 64 points in the neighborhood are searched from the point cloud p according to the K nearest neighbor algorithm as support points s; According to the support point s, from the potential vector z p Take out the potential vector z corresponding to the support point s p-k ; The potential vector z corresponding to the support point s p-k The relative potential vector z is obtained by concatenating the distance qs between the query point q and the support point s p-q ; The relative potential vector z p-q After three layers of two-dimensional convolutional layers with input and output dimensions of 512, the attention feature x is obtained; Input the attention feature x into a two-dimensional convolutional layer with an input dimension of 512 and an output dimension of 1 and a two-dimensional convolutional layer with both input and output dimensions of 512 to obtain a query matrix and a value matrix; The query matrix is ​​normalized by the SoftMax layer and then transposed, and then multiplied by the value matrix to obtain the feature code z q ; Encode the feature z q After being decoded into occupancy by the linear layer and the SoftMax layer, the cross entropy loss function is used in the training process to calculate the loss between the predicted occupancy and the true occupancy marked in the point cloud p. The loss function calculation formula is as follows: CE(y′,y)=-ylohg(y′), Among them, y′ represents the predicted occupancy rate, and y represents the true occupancy rate marked in the dataset.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images