Laser radar data resolution enhancement method based on adversarial neural network

Abstract

The invention relates to the technical field of laser radar data processing, and discloses a laser radar data resolution enhancement method based on an adversarial neural network, which converts low-resolution laser radar data into high-resolution data through adversarial training of a generator and a discriminator, and improves the accuracy of wake flow feature extraction. The generator carries out detail completion and resolution improvement on low-resolution data step by step through a multi-layer convolution module, a deconvolution module and a residual module; the discriminator dynamically optimizes the authenticity of the generated data and ensures that the geometric distribution of the generated high-resolution data is consistent with that of the real data; the generated high-resolution data supports vortex intensity analysis, vortex core position identification and accurate extraction of disturbance area boundaries, provides a three-dimensional reconstruction model of wake flow, and provides reliable support for wake flow interval optimization and flight safety management; according to the method, the accuracy and robustness of wake flow detection are remarkably improved, and a powerful guarantee is provided for flight safety and airspace utilization efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of laser radar data processing, and in particular to a laser radar data resolution enhancement method based on an adversarial neural network. Background Art

[0002] As a detection device that measures the distance, speed and characteristics of targets based on laser echoes, LiDAR is widely used in the aviation field. In order to ensure flight safety, it is necessary to conduct in-depth research on the characteristics and evolution laws of aircraft wakes, and use this as an important basis for reducing intervals. At present, LiDAR is the main tool for detecting aircraft wakes. Its high precision can effectively capture wake data, and is especially suitable for wake detection under clear sky conditions. However, due to equipment performance and environmental factors, the data resolution may be low due to factors such as weather, environment, measurement method and measurement object, resulting in the final suggestion for providing wake interval optimization is not optimal, and it is difficult to capture the detailed characteristics of the wake, resulting in insufficient accuracy of wake feature extraction, affecting detection accuracy and actual application effect. Therefore, studying how to improve the data resolution of LiDAR is the latest breakthrough in solving the bottleneck problem of capacity constraints at busy airports, and is also a key to ensuring airport safety.

[0003] Currently, existing technologies process low-resolution lidar data through traditional interpolation algorithms or convolutional neural networks (CNN), but there are the following problems: In the prior art, although traditional interpolation algorithms such as bilinear interpolation and cubic spline interpolation methods can improve low-resolution data, the generated high-resolution data is too smooth and easily loses the details of the wake characteristics; it is unable to truly restore the wake characteristics and meet the accuracy requirements of aviation wake detection.

[0004] Although convolutional neural networks have performed well in super-resolution modeling, the high-resolution data they generate is single and lacks the ability to dynamically optimize the quality of data generation. The detailed information of the output results is still relatively limited, making it difficult to fully capture complex wake characteristics, thus affecting the final data quality. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to provide a method for enhancing the resolution of lidar data based on adversarial neural networks, which applies GAN technology to lidar data resolution enhancement. Through adversarial training of the generator and the discriminator, not only the data resolution is improved, but also the detail retention of the wake characteristics is optimized, achieving higher data quality and accuracy. The technical solution is as follows: A method for enhancing the resolution of laser radar data based on an adversarial neural network, characterized by comprising the following steps: Step 1: LiDAR data collection; Collect low-resolution lidar data of aircraft wakes using lidar equipment to build an initial data set; Step 2: Data preprocessing; Perform format conversion, noise processing and normalization operations on the collected low-resolution lidar data to ensure data quality and adapt to model input; Step 3: GAN model training; The preprocessed data is input into the GAN model for adversarial neural network training. The GAN model includes a generator and a discriminator. The generator generates high-resolution data based on low-resolution lidar data. The discriminator determines the authenticity of the generated high-resolution data and performs adversarial optimization with the generator. Through multiple rounds of adversarial training, the quality of the generated data is improved. Step 4: High-resolution wake data generation; Using the trained GAN model, the low-resolution lidar data is converted into high-resolution data to refine the wake features, including vortex structure and airflow disturbances.

[0006] The beneficial effects of the present invention are: 1) High-resolution data generation: Based on the adversarial training mechanism of GAN, the high-resolution lidar data generated by the present invention retains the key features of the wake (such as vortex intensity and disturbance area), significantly improving the data quality.

[0007] 2) Data processing link optimization: The present invention ensures the authenticity, smoothness and geometric consistency of the generated data through multi-stage data preprocessing, GAN model training and post-processing technology.

[0008] 3) Improved accuracy of wake feature extraction: The high-resolution data of the present invention supports accurate vortex strength calculation, vortex core position identification and three-dimensional structure reconstruction, providing reliable support for the study of wake evolution laws.

[0009] 4) Practical application value: The high-resolution data generated by the present invention can significantly improve the accuracy of wake interval optimization, support flight path planning and operation management, improve airspace utilization efficiency, and ensure flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 The present invention is a flowchart of a method for enhancing the resolution of laser radar data based on an adversarial neural network.

[0011] Figure 2 This is the data preprocessing flow chart.

[0012] Figure 3 Schematic diagram of the discriminant network.

[0013] Figure 4A schematic diagram of the generated network.

[0014] Figure 5 This is the overall adversarial network architecture diagram. DETAILED DESCRIPTION

[0015] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] The process of the technical solution of the present invention is roughly divided into the following steps: The first is lidar data collection. Low-resolution wake data is collected through lidar equipment to construct an initial data set. The data includes important features such as the velocity field and pressure field of the wake. Then comes data preprocessing. The collected low-resolution data is formatted, noise processed and normalized to ensure data quality and adapt to model input. The processed data is trained with a GAN model to train an adversarial neural network containing a generator and a discriminator. (Generator: Generates high-resolution data based on low-resolution data. Discriminator: Determines the authenticity of the generated data and performs adversarial optimization with the generator.) Through multiple rounds of adversarial training, the generator gradually improves the quality of the generated data. Use the trained GAN to convert low-resolution lidar data into high-resolution data to refine wake features, such as vortex structure and airflow disturbances. The generated high-resolution data is subjected to wake feature extraction, including key features such as vortex intensity and vortex distribution, to provide support for subsequent analysis. The wake interval is adjusted in combination with the optimization algorithm to provide more accurate data support for flight safety management, such as optimizing aircraft take-off and landing intervals or cruising paths. The specific flow chart of the present invention is as follows: Figure 1 shown.

[0017] 1) LiDAR data collection: At this stage, low-resolution wake data is collected through lidar equipment. The data content includes velocity field, pressure field and geometric features of the wake disturbance area, which is the core input of the entire process. The collected low-resolution data usually has low resolution and less information, but it can be used as input to the GAN model to generate high-resolution data. Data collection is the input source of data preprocessing, and the collected data directly enters the next step of data preprocessing.

[0018] 2) Data preprocessing: The collected low-resolution data is processed and cleaned in the following three steps: denoising, removing the environmental noise that may exist in the laser radar acquisition process; normalization, adjusting the data to a unified scale to make different data types consistent; format conversion, converting the data into a tensor format acceptable to the GAN model. This stage ensures that the data quality meets the standards and is the basis for subsequent GAN model training. Figure 2 It is the flow chart of data preprocessing.

[0019] The specific process is as follows: In the preprocessing process of LiDAR data, the first is the input layer, through which the raw Doppler LiDAR data is introduced into the system. Raw LiDAR data usually exists in the form of point clouds, each point contains three-dimensional coordinates (x, y, z), reflection intensity and timestamp. These data are unprocessed and may contain noise and redundant information, so they need to be further extracted and optimized.

[0020] Next, the data flows into the Conv Block, which is the core module for feature extraction. The first part of the Conv Block is the Conv Layer, which extracts local spatial relationships and features by using a 3×3 convolution kernel to perform a sliding window operation on the local area of ​​the data. Specifically, the convolution operation is based on the following formula: ,in, x For input data, w is the convolution kernel weight, b is the bias term. Through the convolution operation, key geometric features in the LiDAR point cloud, such as edge or plane information, can be captured.

[0021] The result of the convolution operation is transformed nonlinearly through the ReLU activation function (Rectified Linear Unit, ReLU), the formula is: ;

[0022] The role of ReLU is to introduce nonlinear characteristics, so that the model can express more complex features, and at the same time improve the sparsity of features by suppressing the interference of negative values ​​on the model. Subsequently, the feature map enters the pooling layer for dimensionality reduction. The pooling operation extracts the maximum value of the local area through a 2×2 pooling window (maximum pooling). The formula is: ; In the formula, x 1, x 2, ..., x n Represents several pixel values ​​(or activation values) of a pooling window in the feature map; The main purpose of pooling is to reduce the dimension of the feature map and reduce the computational complexity, while further enhancing the model's attention to key features by retaining the most significant values ​​of the features.

[0023] The feature map output by the convolution block will pass through the normalization block, which is specifically implemented as batch normalization. The function of batch normalization is to standardize the features so that their mean is 0 and their standard deviation is 1, thereby eliminating the offset of feature distribution and stabilizing model training.

[0024] The formula is as follows: ; In the formula, is the normalized data; and are the mean and standard deviation of the features, respectively. is a minimum value used to avoid division by zero errors. Normalization can improve the convergence speed of the model and reduce the instability caused by changes in the input data distribution.

[0025] Next, the normalized features are input to the fully connected layer for feature conversion. The purpose of this step is to map the multi-dimensional features to a fixed-size feature representation to adapt to subsequent tasks (such as classification or regression). The operation of the fully connected layer can be expressed by the following formula: .

[0026] In the formula, represents the output of the fully connected layer, are input features, W is the weight matrix, b is the bias vector. Through the fully connected layer, the model can learn higher-level feature representations and model the overall structure or semantic information of the point cloud data.

[0027] Finally, the data processed by the above steps flows into the output layer to generate clean and standardized output data tensors. These data have eliminated noise and redundant information, and through feature extraction, dimensionality reduction and transformation, they form a structured and unified feature representation, which is suitable for subsequent specific tasks.

[0028] 3) GAN model training: This step aims to optimize the generative network and the discriminative network through the adversarial training mechanism of the Generative Adversarial Network (GAN), so that the generative network can generate high-quality data close to real samples based on the input data. The model training process includes network initialization, discriminative network training, generative network training, alternating optimization and model convergence. The specific workflow is as follows: a. Network initialization: Initialize the generative networkG and the discriminant network D A random initialization method (such as Xavier initialization or He initialization) is used to ensure that the initial distribution of network parameters is suitable for training.

[0029] Select an optimizer to optimize the two networks. Usually, the Adam optimizer is used, and its initial learning rate is set to , the momentum parameter is set to Define the loss functions of the generator and discriminator networks, including adversarial loss and (optional) content loss. Prepare training datasets, using input data (such as low-resolution data or noise vectors) and target data (such as high-resolution data or real samples) as inputs to the generator and discriminator networks, respectively.

[0030] b. Discriminant network training: The goal of the discriminant network is to improve its ability to distinguish between real samples and generated pseudo samples. The specific training steps are as follows: Randomly extract a batch of real samples from the training set , input to the discriminant network D , calculate the classification probability of the discriminant network for the real sample The goal is to make the output of the discriminant network close to 1, that is, to maximize .

[0031] Generate pseudo samples from input data (such as low-resolution data or random noise z) , input the generated pseudo samples into the discriminant network D , calculate the classification probability of the discriminant network for the pseudo sample ; The goal of the discriminant network is to make its output close to 0, that is, to minimize , to correctly identify the generated samples as forged.

[0032] According to the classification results of real samples and fake samples, the loss of the discriminant network is calculated : ; In the formula, It represents the ability to measure the classification of real samples by the discriminant network. It represents the ability of the discriminant network to classify the generated samples. According to the loss function , the parameters of the discriminant network are updated through back propagation, so that it can distinguish real samples from fake samples more accurately. Figure 3 It is a schematic diagram of the discriminant network.

[0033] The discriminant network consists of a convolutional layer, a pooling layer, a batch normalization (BN), and a swish activation function, and finally outputs the classification probability through a sigmoid activation function.

[0034] The first layer is a convolution layer (C(3-1-1)), which extracts local features through smaller convolution kernels and combines BN and swish activation functions to enhance nonlinear expression capabilities. The middle layer is composed of 3 convolutions, asymmetric pooling, BN and swish, which gradually reduce the resolution of feature maps and extract multi-scale features. Finally, a 1×1 convolution layer (C(1-1-1)) is used to integrate the feature maps and use global average pooling (AVGPool) to generate a single feature value. The sigmoid activation function is used to map the feature value to [0,1], indicating the probability that the input is a real high-line point cloud.

[0035] This discriminant network extracts features layer by layer and reduces dimensionality, and ultimately effectively distinguishes between real and generated high-resolution point cloud data.

[0036] c. Generate network training: The goal of the generative network is to improve its generative ability so that the generated pseudo samples can deceive the discriminative network as much as possible. The specific training steps are as follows: Similar to the training of the discriminative network, pseudo samples are generated from the input data (such as low-resolution data or random noise z) , input the generated pseudo samples into the discriminant network D The goal of the generative network is to optimize the classification probability of the discriminative network for the generated samples. Close to 1, that is, maximized , thereby deceiving the discriminative network.

[0037] Adversarial Loss for Generative Networks Defined as: ; In the image super-resolution task, the generated samples must also be similar in content to the target samples. To achieve this, content loss can be introduced : ; Content loss measures the difference between the generated pseudo samples and the target samples in pixel or feature space.

[0038] The total loss of the generator network is the weighted sum of the adversarial loss and the content loss: ; In the formula, α and β is a weight parameter used to balance the impact of content loss and adversarial loss.

[0039] According to the total loss function , the parameters of the generated network are updated through back propagation so that the pseudo samples it generates are closer to the real samples. Figure 4 It is a schematic diagram of the generated network.

[0040] The generative network consists of a convolutional layer, a residual module, and an upsampling module. It performs super-resolution processing on low-resolution point cloud data to generate high-resolution point cloud data. The single-channel distance map is input, and the initial convolution layer with a convolution kernel of 9 is used to extract shallow features such as texture and edges, enhance the learning of neighborhood similarity, and capture detailed information.

[0041] The residual module design backbone consists of multiple residual units, each of which contains two C(3-1-1) convolutional layers. Multi-scale feature reuse is achieved through cross-layer connections. Monte Carlo (MC) Dropout is deployed to simulate noise distribution, and the prediction robustness is improved through uncertainty elimination mechanism. The upsampling module reconstructs the point cloud details through convolution, asymmetric average pooling and pixel reorganization, and finally integrates two layers of C(3-1-1) convolutional layers to generate a high-line point cloud. The pooling layer is deleted to avoid resolution loss, and edge completion is introduced in the convolution process. At the same time, the residual module is used to improve the efficiency of deep feature utilization and alleviate the risks of gradient disappearance and overfitting.

[0042] d. Alternating optimization: In each training, the discriminant network is optimized first D parameters to enable it to more accurately distinguish real samples from fake samples; then optimize the generated network G The parameters of the adversarial network are adjusted so that the pseudo samples generated by it can more effectively deceive the discriminant network; the training process of the discriminant network and the generative network is carried out alternately until the generative network and the discriminant network reach an adversarial balance. The overall adversarial network architecture is as follows: Figure 5 shown.

[0043] e. Model convergence: When GAN reaches adversarial equilibrium, the distribution of samples generated by the generative network The distribution of real data Very close, at this time the discriminant network cannot effectively distinguish between real samples and fake samples, and its classification accuracy is close to the random level (about 50%).

[0044] According to the GAN theory proposed by Goodfellow et al., when the generating network and the discriminating network reach Nash equilibrium: ; at this time , the output of the discriminant network tends to 0.5 for all samples.

[0045] Use quantitative indicators to evaluate FID (Fréchet Inception Distance): measure the similarity between the distribution of pseudo samples and real samples, the smaller the value, the better. Verify the diversity of data: whether the distribution of pseudo samples is consistent with that of real samples, and whether there is mode collapse.

[0046] When the adversarial balance is achieved and the quality of the pseudo samples is verified to meet the task requirements, the parameters of the generated network are saved.

[0047] 4) High-resolution data generation: The core goal of this invention is to use the trained generative adversarial network (GAN) to convert low-resolution lidar data into high-resolution data, thereby improving the accuracy of wake feature extraction. The following is the specific implementation process of high-resolution data generation: a. Data generation: Through the trained generative network G , the low-resolution point cloud Convert to high-resolution point cloud The specific process is as follows: Input Data Includes point cloud data (or related feature data) collected by lidar, which has low resolution and limited point density.

[0048] The input form of is a preprocessed tensor with the shape ;in is the number of feature channels, H , W is the spatial distribution of the point cloud.

[0049] Through the trained generative network G , input low-resolution lidar data , generate high-resolution data: ; The generative network gradually completes the details and improves the resolution of low-resolution data through multiple layers of convolution, deconvolution and residual modules. Output high-resolution data Contains higher density point clouds and detailed information, which can more accurately capture important details such as the vortex structure and disturbance characteristics of the wake.

[0050] b. Data processing: Use Gaussian Filter to smooth the distribution of points in the point cloud and eliminate local random noise.

[0051] For each point in the point cloud , defining its neighborhood , which is the radius rAll points within the neighborhood, calculate the weight of the points in the neighborhood , the weights follow a Gaussian distribution: ; In the formula, is the standard deviation of the Gaussian distribution and controls the smoothing range.

[0052] Update Point The position of is the weighted average of the neighborhood points: ; The effect is to smooth the point cloud and remove local noise while maintaining the geometric properties of the point cloud.

[0053] Use K-Nearest Neighbors Filtering (KNN) to remove outliers and retain the main structure of the point cloud. Calculate its k The closest point , calculation point The average distance to its nearest neighbor: ; Set a distance threshold (determined according to the distribution characteristics of the point cloud), remove > point.

[0054] The effect is to remove outliers and optimize the sparsity of the point cloud.

[0055] ICP (Iterative Closest Point) is a classic point cloud registration algorithm used to convert point clouds With reference point cloud Alignment.

[0056] For point clouds Each point in ,turn up The closest point in : ; Calculation will Align to The required rotation matrix R and translation vectors t : ; Solve the rigid transformation using the Singular Value Decomposition (SVD) method: Compute the centroid of a point set ; Compute the covariance matrix: ;

[0057] right Perform SVD decomposition to obtain the rotation matrix R and translation vectors t .

[0058] Point Cloud Updated to: ; Repeat the above steps until the alignment error converges: ; The effect is to generate a point cloud The global geometry and reference point cloud Consistent.

[0059] The ICP algorithm is simultaneously applied to a small-scale point set in the vortex core region to improve the local alignment accuracy and ensure the geometric consistency of the wake feature region.

[0060] The local geometric features of each point are extracted using neighborhood feature calculation to describe the vortex structure of the wake.

[0061] Calculate local features: Calculate the normal vector using Principal Component Analysis (PCA) : ; The curvature is based on the eigenvalues ​​of the covariance matrix of the neighborhood points , , calculate: ; Through noise removal, data alignment and feature enhancement, high-fidelity and high-resolution point cloud data can be generated, especially retaining key features (such as vortex intensity and disturbance distribution) in the wake area.

[0062] c. Data Verification: The goal of data validation is to evaluate the generated high-resolution point cloud The quality and applicability of the model are ensured to meet the requirements of wake feature extraction and application. The verification is divided into two parts: quantitative evaluation and qualitative evaluation.

[0063] The quantitative evaluation uses Chamfer Distance (CD), which is a commonly used geometric similarity indicator that can effectively measure the generated point cloud. and Similarity in geometric distribution. It comprehensively evaluates the global geometric differences of point clouds by calculating the nearest neighbor distance between two sets of point clouds: ; In the formula, d CD (i.e. Chamfer distance), is a commonly used geometric similarity index that can effectively measure the similarity between the generated point cloud and the geometric distribution; : Generated high-resolution point cloud; : Real high-resolution point cloud (Ground Truth); and : The number of points in the point cloud; u and v :Respectively represent and The point inside.

[0064] First calculation Each point in The square distance of the nearest point reflects Is it covered? The second calculation is Each point to The square distance of the nearest point reflects Whether cover. The smaller the value, the closer the geometric distribution of the generated point cloud is to the real point cloud. The Chamfer distance in the wake feature area (such as the vortex core) can be calculated separately to verify the geometric restoration effect of the key area.

[0065] Qualitative evaluation uses three-dimensional visualization, which is the most intuitive and effective qualitative evaluation method in point cloud data verification. By observing the structure, details and local features of the generated point cloud and the real point cloud, its physical authenticity and wake feature retention are evaluated.

[0066] Use open3D to render the point cloud in 3D. and real high-resolution point cloud , loaded into the view respectively.

[0067] Will and Superimposed in the same coordinate system, different point clouds are represented by different colors: is blue, is red. Then view and The structure of the point cloud is checked to see if the details of the generated point cloud conform to the laws of physics. Finally, the geometric consistency of the point cloud is checked from multiple perspectives (such as top view and side view).

[0068] The data verification results are intended to show that the GAN-based lidar resolution enhancement method of the present invention can significantly improve the resolution and physical reality of point cloud data, providing reliable support for wake feature extraction and practical applications.

[0069] 5) Wake feature extraction: The high-resolution lidar data generated by the present invention has important applications in wake feature extraction and can significantly improve the accuracy and efficiency of wake analysis, which is specifically reflected in the following aspects: Vortex intensity analysis: Accurately calculate the intensity of the wake vortex through high-resolution point cloud data, capture velocity gradients and rotation characteristics, and provide a scientific basis for evaluating the scope and duration of wake impact.

[0070] Identification of vortex core position and disturbance area: Utilize point cloud density distribution and local curvature characteristics to accurately identify the vortex core position and the boundary of the wake disturbance area, provide an accurate description of the wake geometry characteristics, and support wake interval optimization.

[0071] 3D reconstruction of wake: Construct a 3D geometric model of the wake based on high-resolution data, refine the vortex core structure and disturbance area characteristics, and provide reliable data support for the study of the dynamic evolution of the wake.

[0072] In summary, the present invention proposes: (a) LiDAR data resolution enhancement method based on GAN: The conversion of low-resolution lidar data to high-resolution data is achieved through the adversarial training mechanism of the generator and the discriminator. The generator network is optimized for wake characteristics, and can refine key features such as vortex structure and airflow disturbance, retaining more details. The discriminator network dynamically optimizes the authenticity of the generated data to ensure that the generated high-resolution data is consistent with the geometric distribution of the real data.

[0073] (b) Multi-stage data processing technology: Data preprocessing includes noise removal, normalization, and format conversion to ensure that the quality of input data meets the requirements of GAN model training. Post-processing techniques such as Gaussian filtering, KNN filtering, and ICP (iterative closest point) alignment algorithm further improve the smoothness, density, and geometric consistency of the generated point cloud.

[0074] (c) High-resolution point cloud generation for wake feature extraction: The generated high-resolution data supports vortex intensity analysis, vortex core location identification, and accurate extraction of disturbance region boundaries. It provides a three-dimensional reconstruction model of the wake, refines the vortex core structure and its dynamic evolution, and provides data support for wake interval optimization and flight path planning.

Claims

1. A laser radar data resolution enhancement method based on adversarial neural network, characterized in that: The following steps are involved: Step 1: LiDAR data collection; Use lidar equipment to collect low-resolution point clouds of aircraft wakes and construct an initial data set; Step 2: Data preprocessing; Perform format conversion, noise processing and normalization operations on the collected low-resolution point cloud to ensure data quality and adapt to model input; Step 3: GAN model training; The preprocessed data is input into the GAN model for adversarial neural network training. The GAN model includes a generator and a discriminator. The generator generates a high-resolution point cloud based on a low-resolution point cloud. The discriminator determines the authenticity of the generated high-resolution point cloud and performs adversarial optimization with the generator. Through multiple rounds of adversarial training, the quality of the generated data is improved. Step 4: High-resolution wake data generation; Using the trained GAN model, the low-resolution point cloud is converted into a high-resolution point cloud to refine the wake features, including vortex structure and airflow disturbances.

2. The method for enhancing the resolution of laser radar data based on an adversarial neural network according to claim 1, characterized in that: In step 2, the data preprocessing step specifically includes: Step 21: Input layer processing; The raw LiDAR data is introduced into the system through the input layer. The raw LiDAR data exists in the form of a point cloud, and each point contains three-dimensional coordinates (x, y, z), reflection intensity and timestamp; Step 22: convolution block processing; The data processed by the input layer is input into the convolution block and processed as follows: Convolutional layer processing: By using a 3×3 convolution kernel to perform sliding window operations on local areas of the data, local spatial relationships and features are extracted; ReLU activation function processing: The result of the convolution operation is transformed nonlinearly through the ReLU activation function to introduce nonlinear characteristics and improve the sparsity of features; Pooling layer processing: The feature map after ReLU activation is input into the pooling layer for dimensionality reduction. The maximum value of the local area is extracted through a 2×2 pooling window to reduce the dimension of the feature map and enhance the model's attention to key features. Step 23: Normalization module processing; The feature map processed by the pooling layer is input into the normalization module to achieve batch normalization to eliminate the offset of feature distribution and stabilize model training; Step 24: Fully connected layer processing; The normalized features are input into the fully connected layer for feature conversion to map the multi-dimensional features to a fixed-size feature representation to adapt to subsequent tasks; Step 25: Output layer processing; The data processed by the fully connected layer is input into the output layer to generate a clean and standardized output data tensor. The output data tensor has eliminated noise and redundant information, and has formed a structured and unified feature representation through feature extraction, dimensionality reduction and transformation, which is suitable for subsequent specific tasks.

3. The method for laser radar data resolution enhancement based on adversarial neural network according to claim 1, characterized in that: In step 3, the steps of GAN model training specifically include: Step 31: Network initialization; Initialize the parameters of the generator network and the discriminant network, select the optimizer to optimize the two networks, and define the loss function of the generator network and the discriminant network; use the input low-resolution wake data as the input of the generator network, and use the high-resolution wake data as the target data as the input of the discriminant network; Step 32: discriminant network training; The goal of the discriminant network is to improve its own resolution ability so that it can distinguish between real samples and generated pseudo samples; the specific training steps are as follows: Randomly extract a batch of real samples from the training set , input to the discriminant network, calculate the classification probability of the discriminant network for the real sample ; The goal is to make the output of the discriminant network close to 1, that is, to maximize ; Generate pseudo samples from input low-resolution wake data , input the generated pseudo samples into the discriminant network D Calculate the classification probability of the discriminant network for the pseudo sample ; The goal is to make the output of the discriminant network close to 0, that is, to minimize ; According to the classification results of real samples and fake samples, the loss of the discriminant network is calculated : ； In the formula, Represents the real data distribution Real sample Expected calculation of ; Represents the pseudo samples generated by the generator Expected calculation of ; It represents the ability to measure the classification of real samples by the discriminant network. It represents the ability to measure the classification of the generated samples by the discriminant network; Step 33: Generate network training; The goal of the generative network is to improve its own generative ability so that the generated pseudo samples can deceive the discriminative network as much as possible. The specific training steps are as follows: Generate pseudo samples from input low-resolution wake data , input the generated pseudo samples into the discriminant network D , calculate the classification probability of the discriminant network for the pseudo sample ; The goal is to make the output of the discriminant network close to 1, that is, to maximize ; Adversarial Loss for Generative Networks Defined as: ； In the formula, Represents the generation network G Input Noise z Obey the noise distribution Expected value at time Represents the logarithmic value of the probability of the discriminator outputting the sample generated by the generator, is the output of the generator; Introducing content loss : ； Content loss measures the difference between the generated pseudo sample and the target sample in pixel or feature space; The total loss of the generator network is the weighted sum of the adversarial loss and the content loss: ； In the formula, α and β is a weight parameter used to balance the impact of content loss and adversarial loss; According to the total loss function , update the parameters of the generated network through back propagation, so that the pseudo samples it generates are closer to the real samples; Step 34: Alternate optimization; The training process of the discriminant network and the generative network is performed alternately according to step 32 and step 33 until the generative network and the discriminant network reach a confrontation equilibrium; Step 35: Model convergence; When the GAN model reaches adversarial equilibrium, the parameters of the generated network are saved.

4. The method for enhancing the resolution of laser radar data based on adversarial neural network according to claim 3, characterized in that: The discriminant network includes a convolution layer, a pooling layer, a batch normalization and a swish activation function, and finally outputs the classification probability through a sigmoid activation function; The first convolution layer C (3-1-1) extracts local features through convolution kernels, and combines batch normalization and swish activation function to enhance nonlinear expression capabilities; the middle layer includes three convolution layers, asymmetric pooling layer Pool, batch normalization and swish activation function, which gradually reduce the resolution of feature maps and extract multi-scale features at the same time; Finally, the feature maps are integrated through a 1×1 convolution layer C(1-1-1), and a single feature value is generated using global average pooling AVGPool; Use the sigmoid activation function to map the feature value to [0,1], indicating the probability that the input is a real high-beam point cloud; The generation network includes a convolution layer, a residual module and an upsampling module, which performs super-resolution processing on low-resolution point clouds to generate high-resolution point clouds; a single-channel distance map is input, and the initial convolution layer extracts shallow features of textures and edges, enhances the learning of neighborhood similarity, and captures detailed information; The residual module includes multiple residual units, each of which contains two C(3-1-1) convolutional layers. Multi-scale feature reuse is achieved through cross-layer connections. Monte Carlo Dropout is deployed to simulate noise distribution, and the prediction robustness is improved through the uncertainty elimination mechanism. The upsampling module reconstructs the point cloud details through convolution, asymmetric average pooling and pixel reorganization, and finally integrates two layers of C (3-1-1) convolutional layers to generate a high-line point cloud.

5. The method for laser radar data resolution enhancement based on adversarial neural network according to claim 1, characterized in that: In step 4, the low-resolution point cloud is converted into a high-resolution point cloud as follows: Input low-resolution point cloud , including point cloud data collected by LiDAR; the input form is a preprocessed tensor with a shape of ,in, is the number of feature channels; H , W is the spatial distribution of the point cloud; through the trained generative network , input low-resolution point cloud , generate a high-resolution point cloud: ； The generative network gradually completes the details and improves the resolution of low-resolution data through multi-layer convolution, deconvolution and residual modules; the output high-resolution point cloud , containing higher density point clouds and detailed information.

6. The method for enhancing the resolution of laser radar data based on adversarial neural network according to claim 5, characterized in that: In step 4, the conversion of the low-resolution point cloud into a high-resolution point cloud also includes: (a) Use Gaussian filtering to smooth the distribution of points in the point cloud and eliminate local random noise; For high-resolution point clouds Each point , defining its neighborhood , which is the radius r All points within the neighborhood, calculate the weight of the points in the neighborhood , the weights follow a Gaussian distribution: ； In the formula, is the standard deviation of the Gaussian distribution, controlling the smoothing range; and Respectively represent the three-dimensional coordinates of two points in the point cloud; Update Point The position of is the weighted average of the neighborhood points: ； (b) Use K-neighbor point filtering to remove outliers and retain the important structure of the point cloud; For each point Calculate its k Nearest Neighbor , calculation point Its nearest neighbor The average distance : ； Set a distance threshold , remove the satisfying > point; For high-resolution point clouds Each point in , find the low-resolution point cloud The closest point in : ； In the formula, q For low-resolution point clouds The point in Compute high-resolution point clouds Align to low resolution point cloud The required rotation matrix R and translation vectors t : ； Solve the rigid transformation using the singular value decomposition method: Compute the centroid of a point set ; Compute the covariance matrix: ； Covariance matrix Perform singular value decomposition to obtain the rotation matrix R and translation vectors t ; T is the transpose symbol; represents the centroid of the high-resolution point cloud, Represents the centroid of the low-resolution point cloud; High-resolution point cloud Updated to: ； In the formula, Points in high-resolution point cloud data after Gaussian filtering and smoothing ; Repeat the above steps until the alignment error E Convergence; the alignment error E for: ； In the formula, N is the total number of points in the point cloud that participate in the alignment calculation; Calculate local features and calculate normal vectors and curvature using principal component analysis; Computing normal vectors using principal component analysis : ； In the formula, For point The normal vector represents the direction of the local features of the point cloud; To describe the normal vector direction of the local plane of the point cloud; The curvature is based on the eigenvalues ​​of the covariance matrix of the neighborhood points , and calculate: ； In the formula, For point The curvature.

7. The method for enhancing the resolution of laser radar data based on adversarial neural network according to claim 6, characterized in that: Step 4 also includes wake feature extraction of the generated high-resolution point cloud to support: Vortex intensity analysis: Accurately calculate the intensity of the wake vortex through high-resolution point cloud data, capture velocity gradients and rotation characteristics, and provide a scientific basis for the assessment of the scope and duration of the wake impact; Identification of vortex core position and disturbance area: Using point cloud density distribution and local curvature characteristics, the vortex core position and the boundary of the wake disturbance area are accurately identified, providing an accurate description of the wake geometry characteristics and supporting wake interval optimization; 3D reconstruction of wake: Construct a 3D geometric model of the wake based on high-resolution point clouds, refine the vortex core structure and disturbance area characteristics, and provide reliable data support for the study of the dynamic evolution of the wake.

Images