Deep learning-based generalized interactive simulation method and system for large-scale scenes

Abstract

Disclosed in the present invention are a deep learning-based generalized interactive simulation method and system for large-scale scenes, for use in the technical field of graphics. The method comprises: sampling a sampling model to obtain a multi-level coarse and fine mesh representation S; removing edge information of the finest mesh in the coarse and fine mesh representation S to obtain a multi-level coarse and fine mesh representation (I) after the removal; inputting the multi-level coarse and fine mesh representation (I) and fine mesh point cloud information V0 into a spatial discretization representation; acquiring a correspondence between coarse and fine networks in the multi-level coarse and fine mesh representation (I); and performing correction calculation on the correspondence between the coarse and fine networks to obtain a node motion information iteration process of the coarse and fine meshes. The method significantly reduces network input while maximizing the retention of mesh space information, and reduces calculation resource consumption.

Description

A large-scale scene generalization interactive simulation method and system based on deep learning Technical Field

[0001] The present invention relates to the field of graphics technology, and in particular to a large-scale scene generalization interactive simulation method and system based on deep learning. Background Art

[0002] Interactive simulation is an important field in computer graphics. Traditional interactive simulation methods are generally based on numerical calculations, and often simulate interactive behaviors by solving physical system constraints. However, the need for high-precision solutions leads to huge consumption of computing resources and low efficiency. Deep learning has been widely used in many fields due to its excellent learning ability and high computing efficiency. Compared with traditional numerical calculation methods, interactive simulation methods based on deep learning can achieve higher simulation accuracy and computing efficiency, and have become a research hotspot. With the increase in the spatial complexity of simulation scenes, the problem scale of interactive simulation scenes also expands. Especially in large-scale scenes, this will lead to a huge scale of network input data, making the network model too large, and thus consuming huge computing resources. In addition, in order to ensure the generalization of interactive simulation methods based on deep learning, different scenes need to be trained, which will lead to an increase in the scale of the data set. Especially in large-scale scenes, due to the large scale of interactive simulation problems, the sampling scale of the data set is also extremely large. Among the existing technologies, there are the following three implementation methods:

[0003] Subspace methods use dimensionality reduction to project the interaction simulation problem in the original space into a subspace with far fewer degrees of freedom for solution. However, this method can lead to the loss of topological structure and local feature information of the physical model, and the calculation of eigenvectors consumes a large amount of computing resources in large-scale scenarios.

[0004] Deformation space conversion methods utilize a contact-centric training strategy, projecting deformation information from world coordinates into the local coordinate system of the rigid tool. This significantly simplifies network training and reduces the dataset size required to ensure generalization. However, deformation space conversion methods are only applicable to rigid-soft body interactive deformations and are not suitable for other types of interactive simulations.

[0005] The substructuring method uses a hierarchical strategy to divide the mesh into layers of varying scales, replacing the original fine-scale computational structure with the comprehensive computational results from each layer. Furthermore, due to the fixed hierarchical strategy, simulations in different scenarios retain consistent hierarchical results, improving the generalization of the network model to a certain extent. However, this hierarchical strategy may lose global deformation information and overall topological structure information of the physical simulation model, thereby reducing simulation accuracy.

[0006] In order to overcome these shortcomings, this chapter proposes a large-scale scene generalized interactive simulation method and system based on deep learning, which can effectively handle large-scale generalized interactive simulation problems.

[0007] Summary of the Invention

[0008] The purpose of this application is to provide a large-scale scene generalized interactive simulation method and system based on deep learning, aiming to solve the problem of consuming a large amount of computing resources and reducing simulation accuracy in large-scale scene interactive simulation.

[0009] To achieve the above objectives, this application provides the following technical solutions:

[0010] This application provides a large-scale scenario generalized interactive simulation method based on deep learning, including:

[0011] Sampling the sampling model to obtain a multi-level coarse and fine grid expression S;

[0012] Eliminate the edge information of the finest grid in the coarse-fine grid expression S to obtain the eliminated multi-level coarse-fine grid expression

[0013] The multi-level coarse and fine grid expression And the fine grid point cloud information V0 input space discretization expression;

[0014] Get the multi-level coarse and fine grid expression The correspondence between medium and thick networks;

[0015] The corresponding relationship between the coarse and fine networks is corrected and calculated to obtain the node motion information iterative process of the coarse and fine grids.

[0016] Furthermore, in the step of sampling the sampling model to obtain the multi-level coarse and fine grid expression S, the following steps are specifically included:

[0017] The sampling model M0(V0,E0) is sampled at multiple levels by using the sampling method of Qslim algorithm, where V0 is a grid point, E0 is a grid edge, and M0 is a Mesh grid, and the multi-level coarse and fine grid expression S=(M0,M1,···,M l ), where M i =(V i ,E i ), represents the i-th grid in the multi-level expression S, i = 0, 1, ···, l, where V i is the point cloud location information, E i It is the topological structure edge information.

[0018] Furthermore, the edge information of the finest grid in the coarse-fine grid expression S is eliminated to obtain the multi-level coarse-fine grid expression after elimination: The steps specifically include the following steps:

[0019] The discretization strategy is used to remove the edge information M0 of the finest grid in the coarse-fine grid expression S, and the multi-level coarse-fine grid expression is obtained.

[0020] Furthermore, when expressing the multi-level coarse and fine grid The steps of inputting the fine grid point cloud information V0 into the discretized expression space specifically include the following steps:

[0021] The input size is N points of a fine grid and the number of points per layer is N i and the number of edges K j sum The sum of Much smaller than the sum of the number of points N and the number of edges K in the fine grid, N+K, that is,

[0022] Furthermore, in obtaining the multi-level coarse and fine grid expression The steps of establishing the correspondence between the coarse and fine networks specifically include the following steps:

[0023] Discretize the fine grid input into the sampling model M0 into a spatial discretization expression I;

[0024] Encode the spatial discretization expression I;

[0025] Get the correspondence between coarse and fine grids.

[0026] Furthermore, the fine grid input into the sampling model M0 is discretized into a spatial discretization expression This expression includes the fine grid point cloud information V0, point cloud position information V i And the topological structure edge information E i .

[0027] Furthermore, the step of encoding the spatial discretization expression I specifically includes the following steps:

[0028] The spatial discretization expression I is encoded into a multi-level graph in, Corresponding to the fine grid point cloud information V0, Corresponding to the i-th layer coarse grid point information V i , Corresponding to the i-th layer coarse grid edge information E i , i=1,2,···,l.

[0029] Furthermore, the step of obtaining the correspondence between the coarse and fine grids specifically includes the following steps:

[0030] Using the interpolation operator edge feature E R Represents the correspondence between coarse and fine grids, shrinking a group of vertices into one vertex, that is, The following correspondence exists:

[0031] The corresponding relationship f is an interpolation operation, which converts the vertex set (v1, v2, ..., v k ) Interpolation to obtain the contraction vertex

[0032] The multi-level coarse and fine grids express all the contraction point sets on S. Within the region center range r, and in the sampling model M0, there is a vertex set The corresponding relationship with interpolation is as follows:

[0033] Then the contraction point set Within the region center range r, the interpolation operator projection can be expressed as:

[0034] in, represents the mth point on the jth grid layer of the multi-level coarse and fine grid expression S; Represents the i-th point on the 0th layer of fine point cloud; is the corresponding interpolation operator edge feature;

[0035] There is a corresponding relationship between different coarse grid levels in the multi-level coarse and fine grid expression S, which is expressed as:

[0036] And: a=b-1

[0037] in, Respectively represent the ath grid on the ath grid and the bth grid in the multi-level coarse and fine grid expression S. i and b j points; is the corresponding edge feature between adjacent coarse grids.

[0038] Furthermore, in the step of performing correction calculation on the correspondence between the coarse and fine networks to obtain the node motion information of the coarse and fine grids through an iterative process, the following steps are specifically included:

[0039] Iterate the edge features, and the iterative process is as follows:

[0040] in, and The interpolation operator edge feature and the multi-level coarse grid edge feature are updated as and δ is the edge feature iteration network, which is a multi-layer perceptron with residual connections. After the edge feature iteration, the interpolation operator edge feature is obtained. and

[0041] The iterative process of point features is as follows:

[0042] Among them, v i , are respectively the finest grid point features and the multi-level coarse grid point features, which are updated to v′ during the point feature iteration process. i , δ is a point feature iterative network, which is a multi-layer perceptron with residual connections. The iterative process of node motion information of coarse and fine grids is obtained through point feature iteration.

[0043] This application provides a large-scale scenario generalized interactive simulation system based on deep learning, including:

[0044] Sampling module: samples the sampling model to obtain a multi-level coarse and fine grid expression S;

[0045] Elimination module: Eliminate the edge information of the finest grid in the coarse and fine grid expression S to obtain the multi-level coarse and fine grid expression after elimination

[0046] Input module: express the multi-level coarse and fine grid And the fine grid point cloud information V0 input space discretization expression;

[0047] Acquisition module: Acquisition of the multi-level coarse and fine grid expression The correspondence between medium and thick networks;

[0048] Correction module: Correction calculation is performed on the correspondence between the coarse and fine networks to obtain the node motion information iterative process of the coarse and fine grids.

[0049] The present application provides a device, which includes a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing a large-scale scene generalization interaction simulation method based on deep learning; the processor is used to execute the program instructions stored in the memory to implement large-scale scene generalization interaction simulation based on deep learning.

[0050] The present application provides a storage medium storing program instructions executable by a processor, wherein the program instructions are used to execute a large-scale scene generalized interactive simulation method based on deep learning.

[0051] This application provides a large-scale scenario generalized interactive simulation method and system based on deep learning, which has the following beneficial effects:

[0052] To address the enormous computational resource consumption in large-scale scenarios and the need for large datasets to ensure generalizability, this application designs a novel spatial discretization expression method to generate multi-level coarse and fine mesh expressions. By adopting a multi-level sampling strategy for the Mesh model, different levels of Mesh information are obtained, and finally, through a spatial discretization strategy, a complete spatial discretization expression is obtained. This method significantly reduces network input and computational resource consumption while maximally preserving the spatial information of the Mesh. In addition, it simplifies the scale of interactive simulation problems, thereby reducing the size of the dataset required to ensure generalizability. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] FIG1 is a flow chart of a large-scale scenario generalized interactive simulation method based on deep learning according to Example 1 of the present application;

[0054] FIG2 is a schematic diagram of the network structure of the network model of Example 1 of the present application;

[0055] FIG3 is a schematic diagram of a data set in Example 1 of the present application;

[0056] FIG4 is a diagram showing the qualitative evaluation results of Example 1 of the present application;

[0057] FIG5 is a diagram showing the evaluation results using the RMSE indicator in Example 1 of the present application;

[0058] FIG6 is a diagram showing the evaluation results using the R2 indicator in Example 1 of the present application;

[0059] FIG7 is a diagram showing the ablation experiment results of Example 1 of the present application.

[0060] FIG8 is a schematic structural diagram of a large-scale scenario generalized interactive simulation system based on deep learning according to Example 2 of the present application;

[0061] FIG9 is a schematic diagram of the structure of the device provided in Example 3 of the present invention;

[0062] FIG10 is a schematic diagram of the storage medium structure provided in Example 4 of the present invention. DETAILED DESCRIPTION

[0063] It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0064] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0065] Referring to FIG1 , there is a flowchart of a large-scale scenario generalized interactive simulation method based on deep learning proposed in this application; the specific steps include:

[0066] S1: Sample the sampling model to obtain a multi-level coarse and fine grid expression S.

[0067] In this embodiment, because the spatial complexity (number of vertices, edges, and faces) of the physical model is different in different scenarios, which leads to different input dimensions when using the same network model, the interactive simulation method based on deep learning needs to ensure generalization, so the consistency of the input dimension must be ensured first; in order to solve this problem, the sampling method of the Qslim algorithm is used to perform multi-level sampling on the sampling model M0 (V0, E0), where V0 is a grid point, E0 is a grid edge, and M0 is a Mesh grid, and the multi-level coarse and fine grid expression S = (M0, M1, ···, M0) of the original physical model is obtained. l ), where M i =(V i ,E i ), represents the i-th grid in the multi-level expression S, i=0,1,···,l, where V i is the point cloud location information, E i The topological structure edge information is retained as much as possible to preserve the spatial information characteristics and topological structure information of the sampling model. At the same time, the multi-level sampling strategy provides support for integrating the multi-grid method into the network framework.

[0068] The purpose of the multi-level sampling strategy is to sample the sampling model at different scales so that the number of points at different scales is N i , and the number of sampling model points is N, multi-scale sampling requires N0≈N, and N i >>N j , where i <j。

[0069] S2: Eliminate the edge information of the finest grid in the coarse-fine grid expression S to obtain the multi-level coarse-fine grid expression after elimination

[0070] In this embodiment, after multi-level sampling, coarse and fine grid expressions at different levels are obtained. In large-scale scenarios, if all edge information of Mesh models at different levels are input into the neural network, since the number of edges is much greater than the number of points, the network input will be huge, resulting in huge consumption of computing resources. However, if the edge information is removed, the topological structure information will be lost. Therefore, this application proposes a spatial discretization strategy that aims to minimize the input of physical model edge features while retaining the spatial information of the physical model; the specific contents include:

[0071] The discretization strategy is used to remove the edge information M0 of the finest grid in the coarse-fine grid expression S, and the multi-level coarse-fine grid expression is obtained. The point cloud information V0 of the fine grid and the coarse grid expression part except M0 in the multi-level expression As a joint input, a spatial discretization expression is obtained. Therefore, the spatial discretization expression includes the point information of the fine grid and the coarse grid information, and the coarse grid information provides the topological structure information of the physical model as a whole. After spatial discretization, only the fine grid point cloud information V0 and the multi-level coarse grid expression are needed. It can replace the original entire fine grid information (point-edge-surface); this method reduces the complexity of the geometric features of the simulation scene, simplifies the scale of the interactive simulation problem, and thus reduces the size of the data set required to ensure generalization.

[0072] S3: Express the multi-level coarse and fine grid And the fine grid point cloud information V0 input space discretization expression.

[0073] In this embodiment, the input size is N points of a fine grid and the number of points per layer is N. i and the number of edges K j sum The sum of Much smaller than the sum of the number of points N and the number of edges K in the fine grid, N+K, that is, This greatly reduces the problem size of interactive simulations, significantly reduces the size of the dataset required to ensure generalization, and significantly reduces network input.

[0074] S4: Obtain the multi-level coarse and fine grid expression The correspondence between medium and thick networks.

[0075] In this embodiment, when obtaining the multi-level coarse and fine grid expression The steps of establishing the correspondence between the coarse and fine networks specifically include the following steps S41 to S43, and the implementation of each step is described in detail below.

[0076] S41: Discretize the fine grid input into the sampling model M0 into a spatial discretization expression I.

[0077] After spatial discretization, the fine grid is input into the sampling model M0 and discretized into a spatial discretization expression This expression includes the fine grid point cloud information V0, point cloud position information V i And the topological structure edge information E i ; By learning the correspondence between coarse and fine grids, the projection of interpolation operators in the network is realized.

[0078] S42: Encode the spatial discretization expression I.

[0079] The step of encoding the spatial discretization expression I specifically includes the following steps:

[0080] The spatial discretization expression I is encoded into a multi-level graph in, Corresponding to the fine grid point cloud information V0, Corresponding to the i-th layer coarse grid point information V i , Corresponding to the i-th layer coarse grid edge information E i , i=1,2,···,l.

[0081] S43: Obtaining the correspondence between the coarse and fine grids.

[0082] The step of obtaining the correspondence between the coarse and fine grids specifically includes the following steps:

[0083] Using the interpolation operator edge feature E R Represents the correspondence between coarse and fine grids, shrinking a group of vertices into one vertex, that is, The following correspondence exists:

[0084] The corresponding relationship f is an interpolation operation, which converts the vertex set (v1, v2, ..., v k ) Interpolation to obtain the contraction vertex

[0085] The multi-level coarse and fine grids express all the contraction point sets on S. Within the region center range r, and in the sampling model M0, there is a vertex set (V′1, V′2, …, V′ k ), with the following interpolation correspondence:

[0086] Then the contraction point set Within the region center range r, the interpolation operator projection can be expressed as:

[0087] in, represents the mth point on the jth grid layer of the multi-level coarse and fine grid expression S; Represents the i-th point on the 0th layer of fine point cloud; is the corresponding interpolation operator edge feature.

[0088] There is a corresponding relationship between different coarse grid levels in the multi-level coarse and fine grid expression S, which is expressed as:

[0089] And: a=b-1

[0090] in, Respectively represent the ath grid on the ath grid and the bth grid in the multi-level coarse and fine grid expression S. i and b j points; It is the corresponding edge feature between the corresponding adjacent coarse grids; the coarse and fine grid iterative operation of the multi-level iterative network needs to know the corresponding relationship between the coarse and fine grids. Learning the corresponding relationship between the coarse and fine grids and realizing the projection of the interpolation operator in the network are the key to constructing the multi-level iterative network.

[0091] S5: performing correction calculation on the correspondence between the coarse and fine networks to obtain the iterative process of node motion information of the coarse and fine grids.

[0092] In this embodiment, in order to further reduce the computing resource consumption of large-scale scenes, a multi-level iterative network is designed; the network uses the multi-grid method to realize the iteration of coarse and fine grid motion information, that is, the motion information calculated by the coarse grid is corrected on the fine grid; including edge feature iteration and point feature iteration.

[0093] During simulation, the correspondence between the coarse and fine meshes changes due to changes in the positions of simulation nodes, which means that the interpolation operator will change. Furthermore, for each simulation node, its relationship with its neighboring nodes will also change, meaning that the aggregated information about each node's neighbors will also change. Therefore, edge feature iteration is required to update the interpolation operator and the aggregated information about each node's neighbors. This edge feature iteration is divided into multi-level expression of adjacent mesh layer edge feature iteration and coarse and fine mesh edge feature iteration.

[0094] The iteration of edge features of adjacent mesh layers starts from the coarsest mesh in the multi-level representation, and then passes the edge feature change information to the adjacent finer meshes. The iterative process is as follows:

[0095] in, and The interpolation operator edge feature and the multi-level coarse grid edge feature are updated as and δ is the edge feature iteration network, which is a multi-layer perceptron with residual connections. After the edge feature iteration, the interpolation operator edge feature is obtained. and

[0096] In the multi-level representation, the edge features of each layer and the edge features of the fine grid are updated after edge feature iteration. The edge feature iteration is equivalent to completing the update of the interpolation operator and the aggregation information of each node's neighbor nodes during the simulation process. The following uses point feature iteration to represent the iterative process of node motion information of the coarse and fine grids during the simulation process. The iterative process is as follows:

[0097] Among them, v i , are respectively the finest grid point features and the multi-level coarse grid point features, which are updated to v′ during the point feature iteration process. i , δ is a point feature iterative network, which is a residual-connected multi-layer perceptron. The iterative process of node motion information of coarse and fine grids is obtained through point feature iteration. The multi-level iterative network uses the multi-grid method to realize the iteration of coarse and fine grid motion information. Since the computational resource consumption of coarse grid calculation and fine grid correction operations is low, the computational resource consumption is effectively reduced.

[0098] Please refer to FIG2, which is a schematic diagram of the network structure of the network model of Example 1 of the present application; wherein m0 represents the number of points in the fine grid of the deformable object, m1, ···, m l Indicates the number of points at each level of the multi-level expression of the deformable object, m Y Indicates the number of points of the rigid body tool, n1,...,n l Indicates the number of edges at each level of the multi-level expression of the deformable object, n Y Indicates the number of edges of the rigid body tool; the encoder, multi-level iterative network, and decoder of the applied network model are all multi-layer perceptrons of size 256.

[0099] Please refer to Figure 3, which is a schematic diagram of the data set of Example 1 of the present application; the first and second rows are real results, the third and fourth rows are the results obtained by the Subphysics method, the fifth and sixth rows are the results obtained by the Meshgraphnest method, and the fourth row is the simulation method of the present application.

[0100] Please refer to Figure 4, which is a qualitative evaluation result diagram of Example 1 of the present application; in each scenario, the first row is the real result, the second row is the result obtained by the Subphysics method, the third row is the result obtained by the Meshgraphnest method, and the fourth row is the simulation method of the present application.

[0101] Please refer to Figure 7 for the ablation experiment results of Example 1 of the present application; the left side is the true value, the middle is the ablation result, and the right side is the simulation method of the present application.

[0102] In summary, this application, through Example 1, employs a multi-level sampling strategy for the mesh model to obtain mesh information at different levels, providing support for the multigrid method. Finally, through a spatial discretization strategy, a complete spatial discretization expression is obtained. This method significantly reduces network input and computing resource consumption while maximally preserving the spatial information of the mesh. Furthermore, it simplifies the scale of interactive simulation problems, thereby reducing the size of the dataset required to ensure generalization.

[0103] Example 2

[0104] Please refer to FIG8 , which is a schematic diagram of the structure of a large-scale scenario generalized interactive simulation system based on deep learning according to Example 2 of the present application; specifically, it includes:

[0105] Sampling module: samples the sampling model to obtain a multi-level coarse and fine grid expression S;

[0106] Elimination module: Eliminate the edge information of the finest grid in the coarse and fine grid expression S to obtain the multi-level coarse and fine grid expression after elimination

[0107] Input module: express the multi-level coarse and fine grid And the fine grid point cloud information V0 input space discretization expression;

[0108] Acquisition module: Acquisition of the multi-level coarse and fine grid expression The correspondence between medium and thick networks;

[0109] Correction module: Correction calculation is performed on the correspondence between the coarse and fine networks to obtain the node motion information iterative process of the coarse and fine grids.

[0110] In summary, the present application obtains a multi-level coarse-fine grid expression S through the sampling module in Example 2, and then removes the edge information of the finest grid in the coarse-fine grid expression S, and converts the multi-level coarse-fine grid expression S into a coarse-fine grid expression S through the input module. The fine grid point cloud information V0 is discretized into input space to obtain the corresponding relationship between the two. Finally, the corresponding relationship between the coarse and fine networks is corrected and calculated to realize the iterative process of node motion information of the coarse and fine grids. The multi-level iterative network uses the idea of ​​multiple grid method. Since the computational resource consumption of coarse grid calculation and fine grid correction operation is low, the computational resource consumption is effectively reduced.

[0111] Example 3

[0112] Please refer to FIG9 , which is a schematic diagram of the device structure of an embodiment of the present application. The device 50 includes a processor 51 and a memory 52 coupled to the processor 51 .

[0113] The memory 52 stores program instructions for implementing the above-mentioned large-scale scene generalized interactive simulation method based on deep learning.

[0114] The processor 51 is used to execute program instructions stored in the memory 52 to implement large-scale scene generalization interactive simulation based on deep learning.

[0115] The processor 51 may also be referred to as a CPU (Central Processing Unit).

[0116] Processor 51 may be an integrated circuit chip with signal processing capabilities. Processor 51 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. A general-purpose processor may be a microprocessor or any conventional processor.

[0117] Example 4

[0118] Please refer to Figure 10, which is a schematic diagram of the structure of the storage medium of the embodiment of the present application. The storage medium of the embodiment of the present application stores a program file 61 that can implement all the above methods, wherein the program file 61 can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a computer, server, mobile phone, tablet or other device.

[0119] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0120] The above description is only a preferred embodiment of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

[0121] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

[0122] Of course, the present invention may have many other implementations. Based on this implementation, other implementations obtained by ordinary technicians in this field without any creative work are all within the scope of protection of the present invention.

Claims

1. A large-scale scenario generalized interactive simulation method based on deep learning, characterized in that: include: Sampling the sampling model to obtain a multi-level coarse and fine grid expression S; Eliminate the edge information of the finest grid in the coarse-fine grid expression S to obtain the multi-level coarse-fine grid expression after elimination The multi-level coarse and fine grid expression and the fine grid point cloud information V0 input space discretization expression; Get the multi-level coarse and fine grid expression The correspondence between medium and thick networks; The corresponding relationship between the coarse and fine networks is corrected and calculated to obtain the node motion information iteration process of the coarse and fine grids.

2. According to the method of large-scale scene generalization interactive simulation based on deep learning in claim 1, it is characterized in that: The step of sampling the sampling model to obtain a multi-level coarse and fine grid expression S specifically includes the following steps: The sampling model M0(V0,E0) is sampled at multiple levels by using the sampling method of the Qslim algorithm, where V0 is a grid point, E0 is a grid edge, and M0 is a Mesh grid, and a multi-level coarse and fine grid expression S=(M0,M1,···,M l ), where M i =(V i ,E i ), represents the i-th grid in the multi-level expression S, i = 0, 1, ···, l, where V i is the point cloud location information, E i It is the topological structure edge information.

3. The method for large-scale scene generalization interactive simulation based on deep learning according to claim 1, characterized in that: After removing the edge information of the finest grid in the coarse-fine grid expression S, the multi-level coarse-fine grid expression after removal is obtained. The steps specifically include the following steps: The discretization strategy is used to remove the edge information M0 of the finest grid in the coarse-fine grid expression S, and the multi-level coarse-fine grid expression is obtained.

4. The method of large-scale scene generalization interactive simulation based on deep learning according to claim 1, characterized in that: In expressing the multi-level coarse and fine grid The step of inputting the fine grid point cloud information V0 into the spatial discretization expression specifically includes the following steps: The input size is N points of a fine grid and the number of points per layer is N. i With the number of edges K j sum Nowa, that is Much smaller than the sum of the number of points N and the number of edges K of the fine grid, N+K, that is, 5. The method for large-scale scene generalization interactive simulation based on deep learning according to claim 1, characterized in that: In obtaining the multi-level coarse and fine grid expression The step of establishing the correspondence between the coarse and fine networks specifically includes the following steps: Discretize the fine grid input into the sampling model M0 into a spatial discretization expression I; Encode the spatial discretization expression I; Get the correspondence between coarse and fine grids.

6. A large-scale scene generalization interactive simulation method based on deep learning according to claim 5, characterized in that: The fine grid input into the sampling model M0 is discretized into a spatial discretization expression This expression includes the fine grid point cloud information V0, point cloud position information V i And the topological structure edge information E i .

7. The method for large-scale scene generalization interactive simulation based on deep learning according to claim 5, characterized in that: The step of encoding the spatial discretization expression I specifically includes the following steps: The spatial discretization expression I is encoded into a multi-level graph in, Corresponding to the fine grid point cloud information V0, Corresponding to the i-th layer coarse grid point information V i , Corresponding to the i-th layer coarse grid edge information E i , i=1,2,···,l.

8. The large-scale scene generalization interactive simulation method based on deep learning according to claim 5, characterized in that: The step of obtaining the correspondence between the coarse and fine grids specifically includes the following steps: Using the interpolation operator edge feature E R Represents the correspondence between coarse and fine grids, shrinking a group of vertices into one vertex, that is, The following corresponding relationships exist: The corresponding relationship f is an interpolation operation, which converts the vertex set (v1, v2, ..., v k ) Interpolate to get the contraction vertex The multi-level coarse and fine grids express all the contraction point sets on S. Within the region center range r, and in the sampling model M0, there is a vertex set (V′1, V′2, …, V′ k ), with the following interpolation correspondence: The contraction point set Within the region center range r, the interpolation operator projection can be expressed as: in, represents the mth point on the jth grid layer of the multi-level coarse and fine grid expression S; Represents the i-th point on the 0th layer of fine point cloud; is the corresponding interpolation operator edge feature; There is a corresponding relationship between different coarse grid levels in the multi-level coarse and fine grid expression S, which is expressed as: and: a=b-1 in, a, b = 1, 2, ..., l, respectively represent the ath grid on the ath and bth grids in the multi-level coarse and fine grid expression S i and b j points; is the corresponding edge feature between the corresponding adjacent coarse grids.

9. The method for large-scale scene generalization interactive simulation based on deep learning according to claim 1, characterized in that: The step of performing correction calculation on the correspondence between the coarse and fine networks to obtain the node motion information iteration process of the coarse and fine grids specifically includes the following steps: Iterate the edge features, and the iterative process is as follows: in, and is the interpolation operator edge feature and the multi-level coarse grid edge feature, which are updated as and δ is the edge feature iteration network, which is a multi-layer perceptron with residual connection. After edge feature iteration, the interpolation operator edge feature is obtained. and The iterative process of point features is as follows: in, are the finest grid point features and the multi-level coarse grid point features, which are updated to v′ during the point feature iteration process. i , δ is a point feature iteration network, which is a multi-layer perceptron with residual connections. The iterative process of node motion information of coarse and fine grids is obtained through point feature iteration.

10. A system for a large-scale scenario generalized interactive simulation method based on deep learning according to claims 1-9, characterized in that: include: Sampling module: samples the sampling model to obtain a multi-level coarse and fine grid expression S; Elimination module: Eliminate the edge information of the finest grid in the coarse and fine grid expression S to obtain the eliminated multi-level coarse and fine grid expression Input module: express the multi-level coarse and fine grid and the fine grid point cloud information V0 input space discretization expression; Acquisition module: Acquisition of the multi-level coarse and fine grid expression The correspondence between medium and thick networks; Correction module: Correction calculation is performed on the correspondence between the coarse and fine networks to obtain the iterative process of node motion information of the coarse and fine grids.

11. A device, characterized in that: The device includes a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing a large-scale scene generalized interactive simulation method based on deep learning as described in any one of claims 1 to 9; the processor is used to execute the program instructions stored in the memory to implement a large-scale scene generalized interactive simulation based on deep learning.

12. A storage medium, characterized in that: Processor-executable program instructions are stored, and the program instructions are used to execute a large-scale scene generalized interactive simulation method based on deep learning as described in any one of claims 1-9.

Images