Laser radar three-dimensional image modeling method and system

Abstract

The invention provides a laser radar three-dimensional image modeling method and system, and the method comprises the steps: building a to-be-trained laser radar structure parameter estimation model based on a Res Net50 frame, a self-enhancement attention mechanism, a convolution attention mechanism and a full connection layer, and integrating a convolution attention module to the last convolution layer of the Res Net50 frame; inputting the laser radar three-dimensional image into a structure parameter estimation model, adjusting a predicted value based on the attention score and the structure parameter matrix, training the structure parameter estimation model based on the adjusted predicted value and a difference value of the number of laser radar three-dimensional structure parameters, and obtaining a trained structure parameter estimation model; and performing parametric modeling on the laser radar three-dimensional image based on a VAE network structure. According to the method, the self-enhancement attention module is introduced, so that the capturing capability of the model on key feature information is successfully enhanced, and a finer and more accurate result is realized on three-dimensional model construction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional image modeling, and in particular to a laser radar three-dimensional image modeling method and system. BACKGROUND

[0002] At present, the research on laser radar three-dimensional image modeling has made significant progress, its application range is constantly expanding, and the technical precision and efficiency have been improved to a certain extent. However, the existing technology has poor modeling quality due to the weak detail representation ability of laser radar three-dimensional image.

[0003] Therefore, it is urgent to provide a solution to improve the above problems. SUMMARY

[0004] In a first aspect, the present application provides a laser radar three-dimensional image modeling method using the following technical solution:

[0005] A laser radar structure parameter estimation model to be trained is established based on a Res Net50 framework, a self-enhanced attention mechanism, a convolution attention mechanism and a full connection layer, and the convolution attention module is integrated into the last layer convolution of the Res Net50 framework;

[0006] The laser radar three-dimensional image is input into the structure parameter estimation model, the predicted value is adjusted based on the attention score and the structure parameter matrix, the adjusted predicted value is obtained, and the structure parameter estimation model is trained based on the difference between the adjusted predicted value and the number of laser radar three-dimensional structure parameters, to obtain a trained structure parameter estimation model;

[0007] The laser radar three-dimensional image is parameterized modeled based on a VAE network structure, the VAE network structure includes an encoder and a decoder, both of which are based on a full connection structure, wherein the last layer of the decoder does not apply an activation function, and all other layers select a parameterized rectified linear unit as an activation mechanism.

[0008] The laser radar three-dimensional image modeling method provided by the present application has the beneficial effect that by introducing a self-enhanced attention module, the model's ability to capture key feature information is successfully enhanced, thereby achieving more fine and accurate results in three-dimensional model construction.

[0009] Optionally, the structure parameters include point cloud coordinates, normal vectors, curvatures, edge feature intensities, corner feature quantities, complexities, object positions, object directions and object sizes.

[0010] Optionally, the process of parameterizing modeling the laser radar three-dimensional image based on the VAE network structure includes:

[0011] A modeling loss function is established based on a divergence loss function and a reconstruction loss function, and a VAE network structure is trained based on the modeling loss function to obtain fixed network weights;

[0012] The structural parameter mean of the laser radar three-dimensional image is taken as test input data, and the test input data is encoded and decoded once by using the fixed network weights to obtain a parameterized modeling VAE model and a mean and variance associated with the encoder, and a feature point in a hidden space obtained by random sampling is obtained in combination with the mean and variance;

[0013] The feature point in the hidden space obtained by random sampling is input into a transformation layer composed of a weight parameter set of a weight matrix of the same dimension to obtain an updated feature vector, and the updated feature vector is input into a decoder to obtain a three-dimensional image modeling result image.

[0014] Optionally, a mathematical expression of the modeling loss function is:

[0015] L=-L KL +L F ;

[0016]

[0017] Wherein, L represents a modeling loss function, L KL represents a divergence loss function, L F represents a reconstruction loss function, I p is an input of a VAE network, O p is an output of the VAE network, M is a point cloud number of a laser radar three-dimensional image model, θ is a corresponding point on a generated three-dimensional model, σ 2 represents a variance, and μ represents a mean.

[0018] Optionally, a mathematical expression of the feature point in the hidden space obtained by random sampling is:

[0019] z~N(μ,σ 2 );

[0020] A mathematical expression of the updated feature vector is:

[0021]

[0022] A mathematical expression of the three-dimensional image modeling result image is:

[0023]

[0024] Wherein, z represents a feature point in a hidden space obtained by random sampling, μ represents a mean, and σ 2denotes variance, z' denotes updated eigenvector, W i denotes weight parameter set, z i denotes eigenvector before updating, Q(i) denotes three-dimensional image modeling result map, Y denotes predicted value, Lz'(i) denotes latent space eigenvector.

[0025] Optionally, the mathematical expression of the adjusted predicted value is:

[0026]

[0027] wherein Y denotes the adjusted predicted value, X refers to an original certain laser radar three-dimensional image structure parameter matrix, and A refers to an attention score.

[0028] In a second aspect, the present application provides a laser radar three-dimensional image modeling system, comprising:

[0029] a Res Net50 module, configured to constitute a basic framework of the laser radar three-dimensional image;

[0030] a channel attention module, configured to perform maximum pooling and global average pooling on the input feature mapping, extract two compressed feature mappings respectively, and then perform multilayer perception and activation function processing on the two compressed feature mappings in sequence to obtain a channel attention feature map;

[0031] a spatial attention module, configured to generate two feature mappings of the same size according to average pooling and global maximum pooling respectively, and then perform merging, convolution and activation function processing on the feature mappings to obtain a final feature representation of fused spatial attention information;

[0032] a predicted value adjustment module, configured to obtain an adjusted predicted value according to the final feature representation and the attention score;

[0033] a VAE network structure module, configured to train a parameterized VAE network according to a difference between the adjusted predicted value and a true value, generate a latent space eigenvector, and perform decoding processing on the latent space eigenvector based on a decoder to obtain a laser radar three-dimensional image model. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 denotes a flow chart of a laser radar three-dimensional image modeling method provided by the present application;

[0035] Figure 2 denotes a structure parameter estimation model diagram of a laser radar three-dimensional image modeling method provided by the present application;

[0036] Figure 3 denotes a VAE network structure diagram provided by the present application;

[0037] Figure 4 A laser radar three-dimensional image modeling result map provided by the present application is shown, with a training round of 100 times;

[0038] Figure 5 A laser radar three-dimensional image modeling result map provided by the present application is shown, with a training round of 300 times;

[0039] Figure 6 A laser radar three-dimensional image modeling result map provided by the present application is shown, with a training round of 500 times;

[0040] Figure 7 A laser radar three-dimensional image modeling result map provided by the present application is shown, with a training round of 800 times;

[0041] Figure 8 A laser radar three-dimensional image modeling result map provided by the present application is shown, with a training round of 1200 times;

[0042] Figure 9 A modeling detail effect comparison map provided by the present application is shown;

[0043] Figure 10 A geometric consistency index test result provided by the present application is shown. DETAILED DESCRIPTION

[0044] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application. Unless otherwise defined, the technical terms or scientific terms used herein should be understood as the usual meanings understood by those of ordinary skill in the art to which the present application belongs. The words such as “comprise” and similar words used herein mean that the elements or objects before the words cover the elements or objects listed after the words and their equivalents, and do not exclude other elements or objects.

[0045] Referring to Figure 1 , the embodiments of the present application provide a laser radar three-dimensional image modeling method, comprising:

[0046] S1, based on a Res Net50 framework, a self-enhanced attention mechanism, a convolutional attention mechanism and a full connection layer, a laser radar structure parameter estimation model to be trained is established, and the convolutional attention module is integrated into the last layer convolution of the Res Net50 framework;

[0047] S2, inputting the laser radar three-dimensional image into the structure parameter estimation model, adjusting the predicted value based on the attention score and the structure parameter matrix, obtaining an adjusted predicted value, and training the structure parameter estimation model based on a difference between the adjusted predicted value and a number of laser radar three-dimensional structure parameters, to obtain a trained structure parameter estimation model;

[0048] S3, parameterizing modeling the laser radar three-dimensional image based on a VAE network structure, the VAE network structure including an encoder and a decoder, both based on a full connection structure, wherein the last layer of the decoder does not apply an activation function, and all other layers select a parameterized rectified linear unit as an activation mechanism.

[0049] In some embodiments, when step S1 is performed, the structure parameters include point cloud coordinates, normal vectors, curvatures, edge feature intensities, corner feature quantities, complexities, object positions, object directions, and object sizes.

[0050] In some embodiments, when step S1 is performed, the laser radar structure parameter estimation model to be trained adopts a basic framework of Res Net50, in which a self-enhanced attention module is fused to improve the accuracy of parameter estimation, and a pre-trained model is applied in the training stage. However, the introduction of the self-enhanced attention module in the estimation model changes the network structure, resulting in the inability to use pre-trained weights. Therefore, the convolutional attention module is integrated into the last layer convolution of ResNet to more accurately capture the structure information of the laser radar three-dimensional image, thereby improving the subsequent analysis accuracy.

[0051] Further, when step S1 is performed, the self-enhanced attention mechanism is implemented based on a spatial attention module and a channel attention module. Specifically, in the channel attention module, the input feature map G first undergoes global maximum value pooling and global average value pooling, and two compressed feature maps are extracted from each, which concentrate the spatial dimension information of the original feature map. Then, these feature maps are sent to a multi-layer perceptron structure equipped with two hidden layers, aiming to generate a channel attention map. The first layer of the MLP is configured with neurons, where r is a reduction factor and c represents the number of neurons.

[0052] Further, the second layer restores to c neurons, and both layers use the ReLU function as the activation mechanism. The MLP outputs a channel attention feature map through the fusion of addition and multiplication operations and the activation processing of the Sigmoid function, where the mathematical expression of the channel attention feature map is:

[0053] p = K(MLP(Avgpool(G))) + MLP(Maxpool(G));

[0054] wherein K represents a Sigmoid function, MLP represents a multi-layer perceptron, Avgpool(G) represents an average pooling process, and Maxpool(G) represents a maximum pooling process.

[0055] Further, in the spatial attention processing unit, the first step is to generate two feature maps with a size of HxWx1 using global average pooling and global maximum pooling techniques. Then, the two feature maps are merged in the channel dimension, and a 7x7 convolution filter is applied for convolution processing. After that, the Sigmoid activation function is used to activate the feature map to generate the spatial attention weight coefficient. By performing element-level multiplication operation between these weight coefficients and the original input features, the final feature representation fused with spatial attention information is obtained, which can be expressed as:

[0056]

[0057] wherein H 7×7 is a convolution operation using a 7x7 convolution kernel.

[0058] Further, when the input is a laser radar three-dimensional image, the sample label used by the model is a continuous laser radar three-dimensional structure parameter value. The number of final outputs of the fully connected layer matches the number of parameters, and the model is trained by evaluating the difference between the predicted value and the true value. The error calculation formula is as follows:

[0059]

[0060] wherein n is the total number of laser radar three-dimensional structure parameters, y i is the true value of the i-th parameter, is the predicted value of the i-th parameter, i.e. the output value of the model.

[0061] Further, in the process of performing step S2 to train the structure parameter estimation model based on the difference between the adjusted predicted value and the number of laser radar three-dimensional structure parameters, since the predicted value output by the model is a continuous real number and the range of values of the parameter also varies in the real number domain, the model faces the problems of difficulty in convergence, strong volatility of prediction results and insufficient stability during the training process. In order to effectively solve this problem, it is necessary to output the laser radar three-dimensional image structure parameter estimation result under the condition of the lowest model error, and the specific calculation formula is as follows:

[0062]

[0063] wherein Y represents the adjusted prediction value, X refers to the original certain laser radar three-dimensional image structure parameter matrix, and A refers to the attention score.

[0064] Further, taking a certain image of an experimental data set as an example, the comparison between the laser radar three-dimensional image structure parameter estimation result and the actual result is shown in Table 1.

[0065] Table 1 Comparison between estimation result and actual result

[0066]

[0067]

[0068] It can be seen from the analysis of the results in Table 1 that the laser radar three-dimensional image structure parameter estimation result is basically consistent with the actual result, which can lay a solid foundation for subsequent laser radar three-dimensional image modeling.

[0069] In some embodiments, in the process of performing step S3 of parameterized modeling of the laser radar three-dimensional image based on the VAE network structure, the process comprises:

[0070] S3-1, establishing a modeling loss function based on the directional divergence loss function and the reconstruction loss function, and training the VAE network structure based on the modeling loss function to obtain fixed network weights;

[0071] S3-2, taking the structure parameter mean of the laser radar three-dimensional image as test input data, and using the fixed network weights to perform one-time encoding and decoding on the test input data to obtain a parameterized modeling VAE model and a mean and variance associated with the encoder, and combining the mean and variance to obtain feature points in the latent space obtained by random sampling;

[0072] S3-3, inputting the feature points in the latent space obtained by random sampling into a transformation layer composed of a weight parameter set of a weight matrix of the same dimension to obtain an updated feature vector, and inputting the updated feature vector into the decoder to obtain a three-dimensional image modeling result image.

[0073] Specifically, when performing step S3-1, the mathematical expressions of the directional divergence loss function and the modeling loss function are as follows:

[0074]

[0075] wherein L KL represents the directional divergence loss function, L F represents the reconstruction loss function, and σ 2denotes variance, μ denotes mean, M denotes the number of point clouds of the laser radar three-dimensional image model, θ denotes the corresponding point on the generated three-dimensional model, O p denotes the output of the VAE network, I p denotes the input of the VAE network.

[0076] Further, the modeling loss function is composed of the divergence loss function and the modeling loss function, and the mathematical expression is:

[0077] L = -L KL + L F ;

[0078] Wherein, L denotes the modeling loss function.

[0079] In some embodiments, after the process of performing step S3-1 to train the VAE network structure based on the modeling loss function to obtain fixed network weights, the training is completed, the network weights are fixed, and then the subsequent stage of parameterized modeling is entered, at this time, the mean value of the structural parameters of the laser radar three-dimensional image is taken as the test input data, and the fixed network weights are used to pass the average value through an encoding-decoding process to construct the VAE model of parameterized modeling.

[0080] Specifically, in the process of performing step S3-2 to obtain the VAE model of parameterized modeling, the output parameters μ and σ 2 associated with the mean model encoder are also obtained, both of which are generated by sampling the hidden space features through normal distribution, and the sampling process follows the normal distribution, and the feature points in the hidden space are obtained by random sampling, wherein the mathematical expression of the feature points in the hidden space obtained by random sampling is:

[0081] z ~ N(μ,σ 2 );

[0082] Wherein, z denotes the feature points in the hidden space obtained by random sampling, N(μ,σ 2 ) denotes the normal distribution with mean μ and variance σ 2 .

[0083] Further, these feature points in the hidden space obtained by random sampling are then sent to a transformation layer composed of a set of weight parameters W of the same dimension, wherein the weight parameter set W contains 256 weight matrices from W1 to W 256 , and the transformation layer generates a new feature vector after implementing linear transformation on the feature points in the hidden space obtained by random sampling, wherein the mathematical expression of the updated feature vector is:

[0084]

[0085] Wherein, z' represents the updated feature vector, W i represents a set of weight parameters, z i represents the feature vector before updating.

[0086] Further, the updated feature vector z' is converted into a new three-dimensional model by the decoder, thereby realizing the laser radar three-dimensional image modeling, wherein the mathematical expression of the three-dimensional image modeling result map is:

[0087]

[0088] Wherein, Q(i) represents the three-dimensional image modeling result map, Y represents the predicted value, Lz'(i) represents the latent space feature vector, W i represents a set of weight parameters.

[0089] Specifically, the mathematical expression of the adjusted predicted value is:

[0090]

[0091] Wherein, Y represents the adjusted predicted value, X refers to the original certain laser radar three-dimensional image structure parameter matrix, and A refers to the attention score.

[0092] The application also provides a laser radar three-dimensional image modeling system, comprising:

[0093] The Res Net50 module is used to constitute the basic framework of the laser radar three-dimensional image.

[0094] The channel attention module is used to perform maximum pooling and global average pooling on the input feature mapping, and then extract two compressed feature mappings, and then sequentially perform multilayer perception and activation function processing to obtain a channel attention feature map.

[0095] The spatial attention module is used to generate two feature mappings of the same size according to average pooling and global maximum pooling, and then perform merging, convolution and activation function processing on the feature mappings to obtain a final feature representation of fused spatial attention information.

[0096] The predicted value adjustment module is used to obtain an adjusted predicted value according to the final feature representation and the attention score.

[0097] The VAE network structure module is used to train the parameterized VAE network according to the difference between the adjusted predicted value and the true value, generate a latent space feature vector, and decode the latent space feature vector by the decoder to obtain a laser radar three-dimensional image model.

[0098] In order to more clearly describe the implementability of the technical scheme of the present application, the present application takes a certain building image as an example, and the laser radar three-dimensional image modeling results of the designed method when the training rounds of the parameterized VAE neural network are 100, 300, 500, 800 and 1200 are as shown in Figures 4-8 .

[0099] When the training rounds are 1200, the modeling of the designed method is close to the end, and can reflect all the detailed information of the research object, indicating that the laser radar three-dimensional image modeling effect of the method is good.

[0100] The laser radar three-dimensional image modeling method based on the self-enhanced attention mechanism is used to implement laser radar three-dimensional image modeling, and its modeling performance is tested. The experimental data set includes contents in multiple fields such as buildings, power transmission lines, indoor complex scenes, outdoor complex scenes and engineering measurement, to verify the effectiveness of different methods.

[0101] The configuration of the experimental environment is as follows: development environment: Microsoft Visual Studio Community 2023; third-party software: Sketchup; third-party libraries and frameworks: Qt, Open GL, Caffe, Py Torch. The actual application effects of the modeling method based on the self-enhanced attention mechanism, the modeling method based on the fuzzy C-means clustering algorithm, the modeling method based on the elevation model and the modeling method based on the tight coupling of laser radar and IMU are compared.

[0102] Further, taking an engineering measurement project as an example, the modeling effects of the four methods are tested, and the specific detail comparison results are as shown in Figure 9 .

[0103] As can be seen from Figure 9 , the modeling method based on the self-enhanced attention mechanism can accurately restore the fine structures and lines in complex scenes, and exhibits high clarity. In contrast, although the three experimental comparison methods perform well in overall layout and general form, they are poor in detail processing, so the images presented are slightly rough, and some subtle features cannot be accurately presented, so the modeling effects of the three methods are not good. Through analysis and comparison, it can be seen that the designed method has the best modeling effect and clarity in laser radar three-dimensional image modeling due to its excellent detail presentation ability.

[0104] Taking the contents in multiple fields such as buildings, power transmission lines, indoor complex scenes, outdoor complex scenes and engineering measurement as research objects, the four methods are used for cross-scene testing of these scenes, the geometric consistency index of the modeling results is tested, and the applicability of different methods in other scenes is verified. The test results are as shown in Figure 5 .

[0105] Depend on Figure 10 It is evident that among all methods, the modeling method based on the self-enhancing attention mechanism exhibits the best geometric consistency index. This directly reflects the significant advantage of the design method in preserving and accurately representing geometric elements (such as lines, planes, and surfaces) when constructing 3D models. This is because the design method employs a self-enhancing attention mechanism in the structural parameter estimation stage. This mechanism can more accurately capture key structural information in the LiDAR 3D images, thus providing high-quality input for subsequent parametric modeling. Furthermore, the VAE network, through the encoding-decoding process, can learn the latent distribution in the data and generate a 3D model similar to the input data but with new features. This helps the network learn more accurate and stable feature representations in the latent space, thereby generating a model with better geometric consistency in the decoding stage.

[0106] Based on this, the cross-union ratio (CUI) of the models of the four methods with actual buildings, power transmission lines, complex indoor scenes, complex outdoor scenes, and engineering surveying scenes was calculated. The CUI can comprehensively evaluate the degree of matching between the model and the actual scene. The higher the value, the closer the model is to the real situation of the actual scene. The specific test results are shown in Table 2.

[0107] Table 2 Test Results

[0108]

[0109] Analysis of the results in Table 2 shows that the average cross-union ratio (CUI) of the modeling method based on the self-reinforcing attention mechanism is 0.95, the average CUI of the modeling method based on fuzzy C-means clustering is 0.66, the average CUI of the modeling method based on the elevation model is 0.80, and the average CUI of the modeling method based on the tight coupling of LiDAR and IMU is 0.81. The comparison shows that the modeling method based on the self-reinforcing attention mechanism has a higher CUI, indicating that this method can accurately reproduce the real scene and achieves good modeling results.

[0110] See Figure 3 A variational autoencoder architecture for parametric modeling of LiDAR 3D images is constructed. The input of this architecture is the extracted structural parameters of the LiDAR 3D image. Both the encoder and decoder are based on a fully connected structure. The last layer of the decoder does not apply an activation function, while all other layers use parametric modified linear units as the activation mechanism.

[0111] Specifically, the parametric modeling of VAE networks involves two core stages: the first stage is training the parametric VAE network, which uses the part connected by the black line from left to right to train and generate latent space feature vectors.

[0112] While the embodiments of the application have been illustrated and described in detail, it will be readily apparent to those skilled in the art that various modifications and changes can be made to the embodiments without departing from the scope and spirit of the application, as described in the claims. Moreover, the application described is not limited in its application to the details set forth in the description or illustrated in the drawings. The application is capable of other embodiments and of being practiced or carried out in various ways.

Claims

1. A method for three-dimensional image modeling of a lidar system, characterized in that, Includes the following steps: A LiDAR structural parameter estimation model to be trained is established based on the ResNet50 framework, self-enhancing attention mechanism, convolutional attention mechanism and fully connected layer. The convolutional attention module is integrated into the last convolutional layer of the ResNet50 framework. The three-dimensional image of the LiDAR is input into the structural parameter estimation model. The predicted value is adjusted based on the attention score and the structural parameter matrix to obtain the adjusted predicted value. The structural parameter estimation model is trained based on the difference between the adjusted predicted value and the number of three-dimensional structural parameters of the LiDAR to obtain the trained structural parameter estimation model. Parametric modeling of LiDAR 3D images is performed based on a VAE network structure. The VAE network structure includes an encoder and a decoder, both of which are based on a fully connected structure. In the decoder, no activation function is applied to the last layer, while all other layers use parametrically modified linear units as the activation mechanism.

2. The method for three-dimensional image modeling of a lidar according to claim 1, characterized in that, The structural parameters include: point cloud coordinates, normal vector, curvature, edge feature intensity, number of corner features, complexity, object position, object orientation, and object size.

3. The method for three-dimensional image modeling of a lidar according to claim 1, characterized in that, The process of parametric modeling of LiDAR 3D images based on VAE network structure includes: A modeling loss function is established based on the divergence loss function and the reconstruction loss function, and the VAE network structure is trained based on the modeling loss function to obtain fixed network weights; The mean of the structural parameters of the 3D image of the LiDAR is used as the test input data. The test input data is encoded and decoded once using the fixed network weights to obtain the parameterized VAE model and the mean and variance associated with the encoder. The mean and variance are combined to obtain the feature points in the latent space obtained by random sampling. The feature points in the latent space obtained by random sampling are input into a transformation layer composed of a set of weight parameters consisting of weight matrices of the same dimension to obtain an updated feature vector. The updated feature vector is then input into a decoder for decoding to obtain a 3D image modeling result.

4. The method for three-dimensional image modeling of a lidar according to claim 3, characterized in that, The mathematical expression for the modeling loss function is: l=-L KL +L F ; Where L represents the modeling loss function, L KL L represents the divergence loss function. F I represents the reconstruction loss function. p This refers to the input of the VAE network, O p This refers to the output of the VAE network, where M is the number of point clouds in the LiDAR 3D image model, θ is the corresponding point on the generated 3D model, and σ is the output of the VAE network. 2 denoted by , and μ represents the mean.

5. A method for three-dimensional image modeling of a lidar according to claim 3, characterized in that, The mathematical expression for the feature points in the latent space obtained through random sampling is: z~N(μ,σ 2 ); The mathematical expression for the updated feature vector is: The mathematical expression for the three-dimensional image modeling result is: Where z represents a feature point in the latent space obtained through random sampling, μ represents the mean, and σ 2 Let z' represent the variance, z′ represent the updated feature vector, and W represent the variance. i Represents the set of weight parameters, z i Let represent the feature vector before the update, Q(i) represent the 3D image modeling result, Y represent the predicted value, and Lz′(i) represent the latent space feature vector.

6. The lidar 3D image modeling system according to claim 1, wherein the mathematical expression for the adjusted predicted value is: in, Y represents the adjusted prediction value, X refers to the original structural parameter matrix of a certain LiDAR 3D image, and A refers to the attention score.

7. A lidar 3D image modeling system, comprising: The Res Net50 module is used to construct the basic framework of a LiDAR 3D image. The channel attention module is used to perform max pooling and global average pooling on the input feature map, and then extract the two compressed feature maps respectively. After processing by multilayer perception and activation function, the channel attention feature map is obtained. The spatial attention module is used to generate two feature maps of the same size based on average pooling and global max pooling respectively, and then process the feature maps by merging, convolution and activation functions in sequence to obtain the final feature representation fused with spatial attention information; The prediction value adjustment module is used to obtain the adjusted prediction value based on the final feature representation and attention score; The VAE network structure module is used to train a parameterized VAE network based on the difference between the adjusted predicted value and the true value, generate a latent space feature vector, and decode the latent space feature vector based on the decoder to obtain a 3D image model of the LiDAR.

Images