Three-dimensional point cloud detection model design method and device

Abstract

The present invention discloses a method and device for designing a three-dimensional point cloud detection model. The method comprises: obtaining three-dimensional point cloud data of each submodule, designing each submodule in the three-dimensional point cloud detection model based on the three-dimensional point cloud data of each submodule; and building a three-dimensional point cloud detection model using each submodule obtained by design. The device comprises a submodule design module, which is used to obtain three-dimensional point cloud data of each submodule, and design each submodule in the three-dimensional point cloud detection model based on the three-dimensional point cloud data of each submodule; and a model construction module, which is used to build a three-dimensional point cloud detection model using each submodule obtained by design. The present invention adopts a static graph method to design all submodules, and uses a building block method to assemble all submodules into a complete three-dimensional point cloud detection model.

Description

Technical Field

[0001] The present invention relates to the technical field of model building, and in particular to a method and device for designing a point cloud 3D detection model based on a static image. Background Art

[0002] At present, the 3D point cloud detection model is composed of neural networks (sparse convolution, pointnet++ related components and their variants) and other middleware. In the prior art, the design of data preprocessing, sparse convolution and pointnet++ adopts the design method of dynamic graphs, that is, for a certain built fixed model structure, when the input point cloud data size changes, the input and output sizes of each component in the model will also change dynamically.

[0003] In the existing model design method, several frames of raw point cloud data are processed into the input format required by the model through data preprocessing, and then sent to the model for target detection. During the preprocessing process, the size of the data output will change dynamically with the actual distribution of the current frame of point cloud data.

[0004] In the existing model design method, inside the model, when sparse convolution components or pointnet++ related components and other components are used to extract features of data, the size of the component output features will change dynamically as the size of the component's current frame input features changes.

[0005] In the existing 3D point cloud detection model built based on the dynamic graph design method, the input and output sizes of each component of the model cannot be fixed during the deployment and reasoning stage. Therefore, in the actual execution of the reasoning process of each frame of point cloud data, each component needs to temporarily open up storage and computing space, resulting in a huge waste of resources and increasing the model reasoning time. At the same time, since the hardware resources occupied by the model reasoning cannot be fixed, the memory management of the detection system is greatly affected, and the memory cannot be effectively reused, which reduces the space utilization of the hardware storage. Summary of the invention

[0006] In view of this, the present invention proposes a method and device for designing a three-dimensional point cloud detection model based on a static graph, adopts a static graph method to design all sub-modules, and uses a building block method to combine all sub-modules into a complete three-dimensional point cloud detection model.

[0007] The first aspect of the present invention provides a three-dimensional point cloud detection model design method, the method comprising: obtaining three-dimensional point cloud data of each sub-module, and designing each sub-module in the three-dimensional point cloud detection model based on the three-dimensional point cloud data of each sub-module; using each sub-module obtained by design to build a three-dimensional point cloud detection model.

[0008] Furthermore, the step of designing each submodule in the three-dimensional point cloud detection model includes: based on the three-dimensional point cloud data of each submodule, evaluating the computing space required for each submodule in the three-dimensional point cloud detection model; adding additional input features to each input feature of each submodule in the three-dimensional point cloud detection model, and calculating additional output features corresponding to each output feature of each submodule; determining the valid number of input features of each submodule based on the additional input features, and determining the valid number of output features of each submodule based on the additional output features.

[0009] Furthermore, the step of evaluating the computational space required by the submodule includes: obtaining any frame of input feature three-dimensional point cloud data in the submodule; and evaluating the computational space required by the submodule according to the size of the input feature three-dimensional point cloud data.

[0010] Furthermore, the step of building a three-dimensional point cloud detection model using the designed sub-modules includes: calculating the computing space required for the three-dimensional point cloud detection model according to the computing space required for each of the sub-modules; and superimposing each input feature, each additional input feature, each output feature, and each additional output feature of each of the sub-modules in sequence to build the three-dimensional point cloud detection model.

[0011] Furthermore, the step of calculating the computing space required for the three-dimensional point cloud detection model includes: based on obtaining the computing space required by each of the sub-modules, accumulating the computing space required by each of the sub-modules to obtain the maximum memory computing space required for the three-dimensional point cloud detection model.

[0012] Further: according to the number of additional input features and additional output features of each of the sub-modules, the number of input features and output features in the three-dimensional point cloud detection model is determined.

[0013] The second aspect of the present invention provides a three-dimensional point cloud detection model design device, which includes: a sub-module design module, used to obtain the three-dimensional point cloud data of each sub-module, and design each sub-module in the three-dimensional point cloud detection model based on the three-dimensional point cloud data of each sub-module; a model construction module, used to build a three-dimensional point cloud detection model using each sub-module obtained by design.

[0014] Furthermore, the submodule design module includes: a first space evaluation calculation module, which is used to evaluate the calculation space required for each submodule in the three-dimensional point cloud detection model based on the three-dimensional point cloud data of each submodule; a first model reasoning module, which is used to add additional input features to each input feature of each submodule in the three-dimensional point cloud detection model, and calculate the additional output features corresponding to each output feature of each submodule; a first identification module, which is used to determine the valid number of input features of each submodule based on the additional input features, and determine the valid number of output features of each submodule based on the additional output features.

[0015] Furthermore, the step of the first spatial evaluation calculation module evaluating the calculation space required for each sub-module in the three-dimensional point cloud detection model includes: obtaining any frame of input feature three-dimensional point cloud data in the sub-module; and evaluating the calculation space required for the sub-module according to the size of the input feature three-dimensional point cloud data.

[0016] Furthermore, the model construction module includes: a second space evaluation calculation module, which is used to calculate the calculation space required for the three-dimensional point cloud detection model according to the calculation space required by each sub-module; a second model inference module, which is used to superimpose each input feature, each additional input feature, each output feature and each additional output feature of each sub-module in sequence to build a three-dimensional point cloud detection model; a second identification module, which is used to determine the number of input features and output features in the three-dimensional point cloud detection model according to the number of additional input features and additional output features of each sub-module.

[0017] The above-mentioned static graph-based 3D point cloud detection model design method calculates the memory space required by the current submodule according to the size parameters of a frame of data in each submodule. During the model reasoning process, the memory space required by the component will not change; at the same time, in order to ensure the correctness of the component output results under a fixed size, additional identifiers are added to characterize the actual input features and the effective number of output features. Through this design method, the problem of dynamic changes in the space occupancy of the 3D point cloud detection model can be solved while ensuring the correctness of the results, thereby improving the model deployment performance and space utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] For purposes of illustration and not limitation, the present invention will now be described in terms of preferred embodiments thereof, with particular reference to the accompanying drawings, in which:

[0019] Figure 1 is a flow chart of the three-dimensional point cloud detection model design method provided in the first embodiment;

[0020] Figure 2 is a flow chart of the submodule design method provided in the first embodiment;

[0021] Figure 3 This is a flow chart of the method for building a three-dimensional point cloud detection model provided in the first embodiment.

[0022] Figure 4 It is a structural block diagram of the three-dimensional point cloud detection model design device provided in the first embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0024] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. The embodiments described are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0026] Glossary:

[0027] Neural network is the main component of 3D point cloud detection model, which is used for feature extraction and aggregation of data. It is mainly composed of sparse convolution, feature extractor network and its variants.

[0028] Sparse convolution includes ordinary sparse convolution (spconv) and submanifold sparse convolution (submspconv) and their variants.

[0029] The feature extractor network (pointnet++) contains modules such as spatial query, aggregation, downsampling, upsampling and their variants.

[0030] The model middleware is used to connect the various modules of the model in series to form data flow, including Dense, MeshGrid, GridSample, NMS and all other sub-modules that make up the model but do not belong to the neural network (Backbone).

[0031] Embodiment 1

[0032] Figure 1This is a flow chart of a method for designing a 3D point cloud detection model for 3D point cloud data provided in the first embodiment. The 3D point cloud detection model is composed of a neural network (Backbone) and a model middleware. The method for designing a 3D point cloud detection model uses a static graph method to design all submodules into a complete 3D point cloud detection model using a building block method.

[0033] Please refer to the attached Figure 1 , the static graph-based 3D point cloud detection model design method includes the following steps:

[0034] S101, designing each submodule in the 3D point cloud detection model based on the 3D point cloud data of each submodule.

[0035] In some embodiments, Figure 2 As shown in the figure, the design method of each submodule in the 3D point cloud detection model is:

[0036] S1011, evaluating the computational space required for each submodule in the 3D point cloud detection model.

[0037] In some embodiments, the step of evaluating the size of the computing space required for each submodule in the three-dimensional point cloud detection model includes:

[0038] Taking a specific submodule as an example, the submodule contains N input feature data, namely Input 1, Input 2, Input 3, ... Input N. The size of a frame of input feature 3D point cloud data Input i of the submodule is used to evaluate the computational space required by the submodule in the model reasoning process. The calculation formula is as follows:

[0039]

[0040] Among them, type (*) is the input feature data input i type; sizeof(*) is the input feature data input i The type and size of the feature data; len(*) is the input feature data input i func(*) indicates the method of evaluating the occupied space based on a certain input; w is the additional weighting factor required after accumulating the computing space required for all inputs, and the value is a floating-point data greater than 1; ceil(*) is the rounding up function.

[0041] In the actual calculation process, if the space is not fully used, the remaining space will be left idle; if it exceeds the space range, the calculation of the module will be stopped immediately to avoid errors.

[0042] No matter what size of data a submodule processes, the submodule will only utilize the size of the evaluated computing space and will not occupy additional space.

[0043] S1012, adding additional input features to each input feature of each submodule in the three-dimensional point cloud detection model, and calculating additional output features corresponding to each output feature of each submodule.

[0044] Take a specific submodule as an example. The submodule contains N input features and N output features, where the N input features are Input 1, Input 2, Input 3, ... Input N; and the N output features are Output 1, Output 2, Output 3, ... Output N. Corresponding additional input features valid_i 1, valid_i 2, valid_i 3, ... valid_i N are added to the N input features Input 1, Input 2, Input 3, ... Input N of the submodule respectively, and additional output features valid_o 1, valid_o 2, valid_o 3, ... valid_o N are calculated for the N output features Output 1, Output 2, Output 3, ... Output N of the submodule respectively.

[0045] Since the number of point cloud data in different frames will change, in order to ensure that the change of point cloud data will not cause the correctness of the internal feature transmission of the model, based on the original input features and output features of the submodule, an additional input feature valid_i is added to each input feature. The additional input feature valid_i identifies the valid number of the current input data. At the same time, an additional output feature valid_o is calculated for each output feature. The additional output feature valid_o identifies the valid number of the current output data. When a submodule contains multiple input features or multiple output features, and there is a strong correlation between the features, one valid_i or valid_o can be used to identify multiple features at the same time to reduce the structural complexity of the module.

[0046] S1013, determining the valid number of input features of each submodule according to the additional input features, and determining the valid number of output features of each submodule according to the additional output features.

[0047] In order to identify the valid number of point cloud data features, the present invention adds an additional valid identifier to each input feature of the submodule to indicate the actual data size of each input feature. At the same time, while calculating the output features, the valid number of each output feature is synchronously output.

[0048] It should be noted that in a specific submodule design, any number of input features and output features are allowed. At the same time, when there is a strong correlation between multiple input features or multiple output features, a valid identifier is allowed to represent multiple input features or output features. In this case, the number of valid identifiers is not equal to the number of input features or output features.

[0049] It should be noted that this sub-module can be each sub-module that constitutes the three-dimensional point cloud detection model, including sparse convolution and pointnet++ related modules in the neural network and its variant computing modules, model middleware modules, and other modules that constitute the three-dimensional point cloud detection model that are not mentioned.

[0050] The present invention adopts a static graph method to design each sub-module, that is, the calculation space of the current sub-module is calculated according to the size parameter of a certain frame data of each sub-module; in order to ensure the correctness of the output results of each sub-module under a fixed size, additional input features and additional output features are added to characterize the effective number of actual input features and actual output features.

[0051] S102, using the designed sub-modules to build a three-dimensional point cloud detection model.

[0052] In some embodiments, a building block approach is adopted to build any desired 3D point cloud detection model using the designed submodules. Figure 3 ,The specific implementation method of the three-dimensional point cloud detection model construction method is as follows:

[0053] S1021, calculating the computing space required by the three-dimensional point cloud detection model according to the computing space required by each submodule.

[0054] Based on the above method, the computational space required by each submodule is obtained, and the computational space required by each submodule is accumulated together to obtain the maximum memory computational space required by the three-dimensional point cloud detection model. The corresponding memory space can be applied for in the hardware according to the maximum memory computational space required by the three-dimensional point cloud detection model for model reasoning calculation. The memory can be reused among the submodules according to the structure of the three-dimensional point cloud detection model. For the computational space of any submodule in the three-dimensional point cloud detection model, when the submodule completes all the calculation processes, the computational space applied for by the submodule can be used for calculations of other submodules, otherwise the corresponding computational space should be locked. This submodule memory reuse strategy can effectively reduce the temporary memory space required in the model reasoning process.

[0055] S1022, superimpose each input feature, each additional input feature, each output feature, and each additional output feature of each submodule in sequence to build a three-dimensional point cloud detection model.

[0056] In some embodiments, all input features, all additional input features, all output features, and all additional output features of each submodule are sequentially superimposed together in a building block manner to obtain a three-dimensional point cloud detection model.

[0057] S1023, determining the number of input features and output features in the three-dimensional point cloud detection model according to the number of additional input features and additional output features of each submodule.

[0058] The number of input features in the 3D point cloud detection model is equal to the sum of additional input features of all submodules in the 3D point cloud detection model; the number of output features in the 3D point cloud detection model is equal to the sum of additional output features of all submodules in the 3D point cloud detection model.

[0059] It should be noted that when a sub-module contains multiple input sub-features and multiple output sub-features, and there is a strong correlation between the input or output features, one additional input feature is allowed to represent multiple input features, or one additional output feature is allowed to represent multiple output features. At this time, the number of additional input features is not equal to the number of input features, or the number of additional output features is not equal to the number of output features.

[0060] The present invention uses single-frame data to evaluate the computing space required for any sub-module in the three-dimensional point cloud detection model, and then applies for the corresponding fixed memory space, so that the input feature and output feature calculation of the sub-module can be completed without changing the computing space during the model reasoning process.

[0061] In order to ensure the correctness of the model construction results, the present invention uses the parameter valid to identify the actual number of input features and output features of the submodule. At the same time, when there is a strong correlation between the input and output features, one parameter valid is allowed to identify multiple features.

[0062] In the three-dimensional point cloud detection model constructed by the present invention, the memory between submodules is reused. When a submodule has not completed calculation, the calculation space applied by the submodule is locked and cannot be reused with other modules. When the calculation of the submodule is completed, the calculation space is unlocked and can be reused for feature calculation of other modules. This reuse method can greatly reduce the memory space required for the entire model construction calculation while ensuring the correct results.

[0063] Embodiment 2

[0064] Figure 4 3D point cloud detection model design device for 3D point cloud data provided by this embodiment is a schematic diagram of the structure. Figure 4 , the three-dimensional point cloud detection model design device 200 includes:

[0065] The submodule design module 201 is used to design each submodule in the three-dimensional point cloud detection model.

[0066] The model building module 202 is used to build a three-dimensional point cloud detection model using the designed sub-modules.

[0067] Among them, the sub-module design module 201 includes a first space evaluation calculation module 2011, a first model reasoning module 2012 and a first identification module 2013; the first space evaluation calculation module 2011 is used to evaluate the calculation space required by each sub-module in the three-dimensional point cloud detection model; the first model reasoning module 2012 is used to add additional input features to each input feature of each sub-module in the three-dimensional point cloud detection model, and calculate the additional output features corresponding to each output feature of each sub-module; the first identification module 2013 is used to determine the effective number of input features of each sub-module based on the additional input features, and determine the effective number of output features of each sub-module based on the additional output features.

[0068] In some embodiments, the specific implementation of the first space evaluation calculation module 2011 evaluating the calculation space required by each submodule in the three-dimensional point cloud detection model is as follows:

[0069] Taking a specific submodule as an example, the size of a frame of input feature data Input i of the submodule is used to evaluate the computational space required by the submodule during the model design process. The calculation formula is as follows:

[0070]

[0071] Among them, type (*) is the input feature data input i type; sizeof(*) is the input feature data input i The type and size of the feature data; len(*) is the input feature data input i func(*) indicates the method of evaluating the occupied space based on a certain input; w is the additional weighting factor required after accumulating the computing space required for all inputs, and the value is a floating-point data greater than 1; ceil(*) is the rounding up function.

[0072] In some embodiments, the first model inference module 2012 adds additional input features to each input feature of each submodule in the three-dimensional point cloud detection model, and calculates the additional output features corresponding to each output feature of each submodule. The specific implementation method is as follows:

[0073] Take a specific submodule as an example. The submodule contains N input features and N output features, where the N input features are Input 1, Input 2, Input 3, ... Input N; and the N output features are Output 1, Output 2, Output 3, ... Output N. Corresponding additional input features valid_i 1, valid_i 2, valid_i 3, ... valid_i N are added to the N input features Input 1, Input 2, Input 3, ... Input N of the submodule respectively, and additional output features valid_o 1, valid_o 2, valid_o 3, ... valid_o N are calculated for the N output features Output 1, Output 2, Output 3, ... Output N of the submodule respectively. Since the number of point cloud data in different frames will change, in order to ensure that the change of point cloud data will not cause the correctness of the internal feature transmission of the model, based on the original input features and output features of the submodule, an additional input feature valid_i is added to each input feature. The additional input feature valid_i identifies the valid number of the current input data. At the same time, an additional output feature valid_o is calculated for each output feature. The additional output feature valid_o identifies the valid number of the current output data. When a submodule contains multiple input features or multiple output features, and there is a strong correlation between the features, one valid_i or valid_o can be used to identify multiple features at the same time to reduce the structural complexity of the module.

[0074] Among them, the model building module 202 includes a second space evaluation calculation module 2021, a second model reasoning module 2022 and a second identification module 2023. The second space evaluation calculation module 2021 is used to calculate the calculation space required for the three-dimensional point cloud detection model according to the calculation space required by each submodule. The second model reasoning module 2022 is used to sequentially superimpose each input feature, each additional input feature, each output feature and each additional output feature of each submodule to build a three-dimensional point cloud detection model; the second identification module 2023 is used to determine the number of input features and output features in the three-dimensional point cloud detection model according to the number of additional input features and additional output features of each submodule.

[0075] In some embodiments, the specific implementation method of the second space evaluation calculation module 2021 calculating the calculation space required for the three-dimensional point cloud detection model is as follows:

[0076] Based on the computational space required by each submodule, the computational space required by each submodule is accumulated to obtain the maximum memory computational space required by the three-dimensional point cloud detection model. According to the maximum memory computational space required by the three-dimensional point cloud detection model, the corresponding memory space can be applied in the hardware for model inference calculation.

[0077] In some embodiments, the second model inference module 2022 builds a three-dimensional point cloud detection model in the following specific implementations:

[0078] All input features, all additional input features, all output features and all additional output features of each sub-module are superimposed together in sequence in a building block manner to obtain a three-dimensional point cloud detection model.

[0079] In some embodiments, the second identification module 2023 determines the number of input features and output features in the three-dimensional point cloud detection model in the following specific implementation manner:

[0080] The number of input features in the 3D point cloud detection model is equal to the sum of additional input features of all submodules in the 3D point cloud detection model; the number of output features in the 3D point cloud detection model is equal to the sum of additional output features of all submodules in the 3D point cloud detection model.

[0081] It should be noted that when a sub-module contains multiple input sub-features and multiple output sub-features, and there is a strong correlation between the input or output features, one additional input feature is allowed to represent multiple input features, or one additional output feature is allowed to represent multiple output features. At this time, the number of additional input features is not equal to the number of input features, or the number of additional output features is not equal to the number of output features.

[0082] The present invention uses single-frame data to evaluate the computing space required for any sub-module in the three-dimensional point cloud detection model, and then applies for the corresponding fixed memory space, so that the input feature and output feature calculation of the sub-module can be completed without changing the computing space during the model reasoning process.

[0083] In order to ensure the correctness of the model reasoning results, the present invention uses the parameter valid to identify the actual number of input features and output features of the submodule. At the same time, when there is a strong correlation between the input and output features, one parameter valid is allowed to identify multiple features.

[0084] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A three-dimensional point cloud detection model design method, characterized in that: include: Acquire the three-dimensional point cloud data of each submodule, and design each submodule in the three-dimensional point cloud detection model based on the three-dimensional point cloud data of each submodule; Using the designed sub-modules, a three-dimensional point cloud detection model is constructed; The steps of designing each submodule in the three-dimensional point cloud detection model include: Based on the three-dimensional point cloud data of each sub-module, evaluating the computing space required by each sub-module in the three-dimensional point cloud detection model; Adding additional input features to each input feature of each of the submodules in the three-dimensional point cloud detection model, and calculating additional output features corresponding to each output feature of each of the submodules; The valid number of input features of each of the submodules is determined according to the additional input features, and the valid number of output features of each of the submodules is determined according to the additional output features.

2. The three-dimensional point cloud detection model design method according to claim 1, characterized in that: The step of evaluating the computational space required by each of the submodules comprises: Get any frame of input feature 3D point cloud data in the submodule; The computational space required by the submodule is evaluated according to the size of the input feature three-dimensional point cloud data.

3. The three-dimensional point cloud detection model design method according to claim 1, characterized in that: The step of building a three-dimensional point cloud detection model using the designed sub-modules includes: Calculate the computing space required by the three-dimensional point cloud detection model according to the computing space required by each of the submodules; Each input feature, each additional input feature, each output feature and each additional output feature of each submodule are sequentially superimposed together to build the three-dimensional point cloud detection model.

4. The three-dimensional point cloud detection model design method according to claim 3, characterized in that: The step of calculating the computational space required for the three-dimensional point cloud detection model comprises: Based on obtaining the computing space required by each of the sub-modules, the computing space required by each of the sub-modules is accumulated to obtain the maximum memory computing space required by the three-dimensional point cloud detection model.

5. The three-dimensional point cloud detection model design method according to claim 1, further comprising: The number of input features and output features in the three-dimensional point cloud detection model is determined according to the number of additional input features and additional output features of each submodule.

6. A three-dimensional point cloud detection model design device, characterized in that: include: A submodule design module, used to obtain the three-dimensional point cloud data of each submodule, and design each submodule in the three-dimensional point cloud detection model based on the three-dimensional point cloud data of each submodule; A model building module, used to build a three-dimensional point cloud detection model using the designed sub-modules; The submodule design module includes: A first space evaluation calculation module, used for evaluating the calculation space required by each submodule in the three-dimensional point cloud detection model based on the three-dimensional point cloud data of each submodule; A first model inference module, used for adding additional input features to each input feature of each submodule in the three-dimensional point cloud detection model, and calculating additional output features corresponding to each output feature of each submodule; The first identification module is used to determine the valid number of input features of each of the submodules according to the additional input features, and to determine the valid number of output features of each of the submodules according to the additional output features.

7. The three-dimensional point cloud detection model design device according to claim 6, characterized in that: The step of the first space evaluation calculation module evaluating the calculation space required by each submodule in the three-dimensional point cloud detection model includes: Get any frame of input feature 3D point cloud data in the submodule; The computational space required by the submodule is evaluated according to the size of the input feature three-dimensional point cloud data.

8. The three-dimensional point cloud detection model design device according to claim 6, characterized in that: The model building module includes: A second space evaluation calculation module is used to calculate the calculation space required by the three-dimensional point cloud detection model according to the calculation space required by each submodule; The second model inference module is used to sequentially stack each input feature, each additional input feature, each output feature and each additional output feature of each submodule to build a three-dimensional point cloud detection model; The second identification module is used to determine the number of input features and output features in the three-dimensional point cloud detection model according to the number of additional input features and additional output features of each sub-module.

Images