Laser radar recognition adversarial sample method based on 3D point cloud reconstruction

Abstract

The invention provides a laser radar recognition adversarial sample method based on 3D point cloud reconstruction, and the method comprises the steps: monitoring the surrounding environment through a laser radar, obtaining point cloud collection data, and obtaining point cloud data feature key points through data preprocessing; encoding the feature key points of the point cloud data by using an encoder based on a multilayer perceptron, outputting global features of the point cloud acquisition data, and performing total loss analysis; 3D point cloud reconstruction and decoding are carried out according to the global features of the point cloud acquisition data, and reconstruction error analysis calculation is carried out to obtain a reconstruction error value; determining an adversarial sample detection result according to the reconstruction error value in combination with a preset threshold value; and inputting the reconstructed point cloud data into the sub-network for identification according to an adversarial sample detection result. According to the invention, 3D point cloud reconstruction is carried out by using the structural characteristics of the three-dimensional neural network, the detection capability and defense capability of the antagonistic sample are improved, the interference of the antagonistic sample on the laser radar system is reduced, and the robustness of the laser radar system is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a laser radar recognition adversarial sample method based on 3D point cloud reconstruction. Background Art

[0002] LiDAR is a commonly used sensor used in fields such as self-driving cars, robots, and drones. It is a radar system that emits laser beams to detect the position, speed, and other characteristic quantities of a target. It can generate point cloud data, which is then used to build environmental models and perform target detection and recognition. Studies have shown that LiDAR systems are vulnerable to adversarial sample attacks, which can cause LiDAR systems to produce erroneous measurement results, leading to system failures and affecting the normal use of LiDAR systems. With the advent of the "artificial intelligence revolution," the problem of adversarial sample attacks has attracted widespread attention in the field of artificial intelligence.

[0003] At present, in a laser radar system, in order to solve the problem that adversarial samples may cause errors such as interference and occlusion processing in the radar detection process, making it difficult for the laser radar to accurately perceive objects in the environment, adversarial sample defense technologies such as adversarial training, adversarial denoising and adversarial detection are usually used for adversarial sample detection. However, adversarial training is to add adversarial samples to training data when training a machine learning model, so that the model can better identify adversarial samples; adversarial denoising is to reduce the impact of adversarial samples by performing denoising on the input data; adversarial detection refers to defending against adversarial attacks by detecting whether the input data is an adversarial sample. These methods have weak defense capabilities against adversarial samples and cannot accurately detect and identify adversarial samples. Therefore, the present invention proposes a laser radar recognition adversarial sample method based on 3D point cloud reconstruction, which utilizes the structural characteristics of a three-dimensional neural network to reconstruct 3D point clouds, can more accurately detect and identify adversarial samples, improve the detection capability of adversarial samples, thereby improving the defense capability against adversarial samples, and reduce the interference of adversarial samples on the laser radar system, thereby improving the robustness of the laser radar system. Summary of the invention

[0004] The purpose of the present invention is to provide a laser radar recognition adversarial sample method based on 3D point cloud reconstruction to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a laser radar recognition adversarial sample method based on 3D point cloud reconstruction, comprising:

[0006] Use laser radar to monitor the surrounding environment and obtain point cloud data;

[0007] Perform data preprocessing on point cloud acquisition data to obtain key feature points of point cloud data;

[0008] According to the key points of the point cloud data features, an encoder based on a multi-layer perceptron is used to encode the data, output the global features of the point cloud data, and perform a total loss analysis;

[0009] Reconstruct and decode the 3D point cloud based on the global features of the point cloud acquisition data, and perform reconstruction error analysis and calculation to obtain the reconstruction error value;

[0010] Determine the adversarial sample detection result based on the reconstruction error value combined with the preset threshold;

[0011] According to the adversarial sample detection results, the reconstructed point cloud data is input into the sub-network for recognition.

[0012] Furthermore, before using the laser radar to monitor the surrounding environment, monitoring preparation work is performed for the laser radar, and the monitoring preparation work includes:

[0013] Set the laser radar parameters and confirm the laser radar with the set parameters;

[0014] Install the laser radar with set parameters, calibrate the equipment, and obtain the laser radar equipment calibration result;

[0015] Configure the LiDAR device connection based on the LiDAR device calibration results.

[0016] Furthermore, when calibrating the equipment, the LiDAR parameters are combined to predict and analyze the LiDAR monitoring area, including:

[0017] Perform lidar monitoring simulation according to the equipment parameters to obtain the lidar's predicted monitoring area;

[0018] Performing a comparative analysis on the predicted monitoring area and the target monitoring area to determine whether the predicted monitoring area is consistent with the target monitoring area, thereby obtaining a first analysis result;

[0019] According to the first analysis result, when the predicted monitoring area is inconsistent with the target monitoring area, multiple groups of predicted data are re-acquired and the correction value is confirmed; the limit distance of the predicted monitoring area is analyzed and determined in each group of predicted data to obtain multiple groups of limit distances, and the multiple limit distances are respectively combined with the target monitoring area to determine multiple monitoring error data, and it is determined whether it is caused by accidental factors based on the multiple monitoring error data to obtain the second analysis result. When the second analysis result shows that it is not caused by accidental factors, the lidar parameter correction value is determined based on the multiple monitoring error data.

[0020] Furthermore, data preprocessing is performed on the point cloud acquisition data, including:

[0021] Perform preprocessing operations on the point cloud acquisition data to obtain point cloud preprocessing data;

[0022] Perform surface reconstruction on the point cloud preprocessing data to obtain a triangular mesh model of the point cloud preprocessing data;

[0023] In the triangular mesh model, curvature analysis is performed on each point to obtain the point cloud curvature value;

[0024] Determine a first threshold value according to the point cloud curvature value, and screen and determine the feature key points of the point cloud data according to the first threshold value;

[0025] Normal analysis is performed based on the feature key points of point cloud data.

[0026] Furthermore, when encoding the key points of the feature of the point cloud data using an encoder based on a multi-layer perceptron, the number of input layers, hidden layers, and output layers is configured according to the key points of the feature of the point cloud data, and the initial weights and initial bias values ​​of the neurons are determined. Then, the data information of the key points of the feature of the point cloud data is input into the input layer, and forward propagation is performed through each layer of the network. Each neuron receives the output of the neurons in the previous layer, and the weighted sum is performed after nonlinear transformation through a modified linear unit activation function to obtain the output information. Then, the first loss function is calculated according to the output information and the true label, and based on the gradient of the first loss function, the loss gradient of each neuron is calculated layer by layer starting from the output layer, and the loss gradient is multiplied by the connection weight according to the chain rule and passed to the previous layer. Then, the weight and bias value of each neuron are updated according to the gradient descent rule. After multiple iterations, the network parameters are adjusted. When the loss function is minimized, the global characteristics of the point cloud acquisition data are obtained.

[0027] Furthermore, when performing total loss analysis, point cloud acquisition data is obtained from the encoder, and mean and variance analysis is performed on the point cloud acquisition data to obtain the mean and variance of the point cloud acquisition data, and the latent variables of the hidden layer in the variational autoencoder are determined based on the mean and variance of the point cloud acquisition data. Then, based on the latent variables of the hidden layer in the variational autoencoder, the KL divergence between the Gaussian distribution and the normal distribution is calculated using the variational autoencoder to obtain the second loss function. Finally, the total loss function is determined based on the first loss function and the second loss function.

[0028] Furthermore, when reconstructing and decoding the 3D point cloud based on the global features of the point cloud acquisition data, a decoder is used to reconstruct and decode the 3D point cloud, the global features of the point cloud acquisition data are input into the decoder, and point cloud reconstruction data samples are obtained through the decoder.

[0029] Furthermore, the reconstruction error analysis calculation includes:

[0030] For the point cloud reconstruction data samples, point cloud distance analysis is performed in turn according to the point cloud to obtain the point cloud distance before reconstruction and the point cloud distance after reconstruction;

[0031] The relative distance of the point cloud is calculated by combining the distance of the point cloud before reconstruction with the distance of the point cloud after reconstruction;

[0032] The point cloud reconstruction error value is obtained by using the EMD distance calculation based on the relative distance of the point cloud.

[0033] Furthermore, when determining the adversarial sample detection result based on the reconstruction error value combined with the preset threshold, the reconstruction error value is compared with the preset threshold. When the reconstruction error value is greater than or equal to the preset threshold, the point cloud reconstructed data sample is an adversarial sample. When the reconstruction error value is less than the preset threshold, the point cloud reconstructed data sample is a normal sample.

[0034] Furthermore, after obtaining the triangular mesh model of the point cloud preprocessing data, the triangular mesh model is optimized, including:

[0035] Perform a preliminary analysis on the triangular mesh model of the point cloud preprocessing data to determine whether the triangular mesh model of the point cloud preprocessing data needs to be optimized and obtain preliminary analysis results;

[0036] According to the preliminary analysis results, when the triangular mesh model of the point cloud preprocessing data needs to be optimized, the distribution of the point cloud preprocessing data in the triangular mesh model is analyzed to obtain the data distribution of the triangular mesh model;

[0037] The optimization scheme is matched according to the distribution of the triangular mesh model, and the triangular mesh model is adjusted based on the optimization scheme to obtain a triangular mesh model.

[0038] The present invention utilizes the structural characteristics of a three-dimensional neural network to reconstruct 3D point clouds, which can more accurately detect and identify adversarial samples, improve the detection capability of adversarial samples, thereby improving the defense capability against adversarial samples, reducing the interference of adversarial samples on the lidar system, and thereby improving the robustness of the lidar system.

[0039] Other features and advantages of the present invention will be described in the following description, and partly become obvious from the description, or be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the application documents.

[0040] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0042] Figure 1 A schematic diagram of the steps of a laser radar adversarial sample recognition method based on 3D point cloud reconstruction according to the present invention;

[0043] Figure 2 This is a schematic diagram of specific steps of step 1 in a method for laser radar recognition of adversarial samples based on 3D point cloud reconstruction according to the present invention;

[0044] Figure 3 This is a schematic diagram of the device calibration steps in step 1 of the method for laser radar recognition of adversarial samples based on 3D point cloud reconstruction described in the present invention;

[0045] Figure 4 This is a schematic diagram of step 2 in the method for laser radar recognition of adversarial samples based on 3D point cloud reconstruction described in the present invention. DETAILED DESCRIPTION

[0046] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0047] like Figure 1 As shown, an embodiment of the present invention provides a laser radar recognition adversarial sample method based on 3D point cloud reconstruction, comprising:

[0048] Step 1: Use laser radar to monitor the surrounding environment and obtain point cloud data;

[0049] Step 2: Preprocess the point cloud data to obtain the key points of the point cloud data features;

[0050] Step 3: Encode the key points of the point cloud data using an encoder based on a multi-layer perceptron, output the global features of the point cloud data, and perform a total loss analysis;

[0051] Step 4: Reconstruct and decode the 3D point cloud based on the global features of the point cloud acquisition data, and perform reconstruction error analysis and calculation to obtain the reconstruction error value;

[0052] Step 5: Determine the adversarial sample detection result based on the reconstruction error value combined with the preset threshold;

[0053] Step 6: According to the adversarial sample detection results, the reconstructed point cloud data is input into the sub-network for recognition.

[0054] In the above technical solution, when encoding the key points of the point cloud data using an encoder based on a multi-layer perceptron, the structural characteristics of the three-dimensional neural network are used for encoding.

[0055] The above technical solution utilizes the structural characteristics of the three-dimensional neural network to reconstruct the 3D point cloud, improves the detection and defense capabilities of adversarial samples, reduces the interference of adversarial samples on the laser radar system, and improves the robustness of the laser radar system. By preprocessing the point cloud acquisition data, the interference of interference factors in the point cloud acquisition data is reduced, the accuracy of the point cloud acquisition data is improved, and the error of subsequent point cloud acquisition data analysis is reduced. The structural characteristics of the three-dimensional neural network are also utilized to improve the detection and defense capabilities of adversarial samples, by encoding the key points of the point cloud data features using an encoder based on a multi-layer perceptron and reconstructing and decoding the 3D point cloud based on the global features of the point cloud acquisition data, so that the adversarial samples can be detected by making minor modifications to the original data, reducing the interference of adversarial samples on the laser radar system, thereby avoiding errors such as interference and occlusion processing caused by adversarial samples in the radar detection process, and improving the accuracy of the radar system.

[0056] like Figure 2 As shown, in one embodiment provided by the present invention, before using the laser radar to monitor the surrounding environment, monitoring preparation work is performed on the laser radar, and the monitoring preparation work includes:

[0057] A1. Set the laser radar parameters and confirm the laser radar with the set parameters;

[0058] A2. Install the laser radar with set parameters and calibrate the equipment to obtain the laser radar equipment calibration result;

[0059] A3. Configure the LiDAR device connection according to the LiDAR device calibration results.

[0060] In the above technical solution, setting the laser radar parameters includes: setting the scanning angle, scanning frequency, resolution, etc.

[0061] In the above technical solution, when calibrating the equipment, the coordinate system must be aligned.

[0062] In the above technical solution, when the laser radar device is connected and configured according to the laser radar device calibration result, the airborne laser radar device is connected to the computer, and the device is controlled by the corresponding software to collect data.

[0063] The above technical solution realizes the monitoring preparation work of the laser radar equipment, so that the laser radar can better collect point cloud data and process the collected data.

[0064] like Figure 3As shown, in one embodiment provided by the present invention, when calibrating the equipment, the laser radar monitoring area prediction and analysis are performed in combination with the laser radar parameters, including:

[0065] S1. Perform laser radar monitoring simulation according to the parameters of the equipment to obtain the predicted monitoring area of ​​the laser radar;

[0066] S2. Compare and analyze the predicted monitoring area with the target monitoring area to determine whether the predicted monitoring area is consistent with the target monitoring area, and obtain a first analysis result;

[0067] S3. According to the first analysis result, when the predicted monitoring area is inconsistent with the target monitoring area, multiple groups of predicted data are reacquired and the correction value is confirmed; in each group of predicted data, the limit distance of the predicted monitoring area is analyzed and determined to obtain multiple groups of limit distances, and the multiple limit distances are respectively combined with the target monitoring area to determine multiple monitoring error data, and it is determined whether it is caused by accidental factors based on the multiple monitoring error data to obtain the second analysis result. When the second analysis result is that it is not caused by accidental factors, the lidar parameter correction value is determined based on the multiple monitoring error data.

[0068] In the above technical solution, if the first analysis result is that the predicted monitoring area is consistent with the target monitoring area, there is no need to calibrate the equipment.

[0069] In the above technical solution, when comparing and analyzing the predicted monitoring area in combination with the target monitoring area, the sizes of the predicted monitoring area and the target monitoring area are respectively obtained, and a preliminary comparative analysis is performed based on the sizes of the predicted monitoring area and the target monitoring area. When the sizes of the predicted monitoring area and the target monitoring area are different, the predicted monitoring area is inconsistent with the target monitoring area. When the sizes of the predicted monitoring area and the target monitoring area are the same, edge area blocks are collected for the predicted monitoring area and the target monitoring area respectively to obtain a first collection area data set and a second collection area data set. The first collection area data set and the second collection area data set are matched in combination with the relative position of the edge area block collection. According to the matching result, the area blocks in the first collection area data set and the second collection area data set are compared to obtain area block difference information. When there is area block difference information that is not zero, the predicted monitoring area is inconsistent with the target monitoring area. When all area block difference information is zero, the predicted monitoring area is consistent with the target monitoring area.

[0070] In the above technical solution, the limit distance refers to the straight-line distance of the viewing angle range in the predicted monitoring area.

[0071] In the above technical solution, when determining whether it is caused by accidental factors based on multiple monitoring error data, if at least one set of monitoring error data among the multiple sets of monitoring error data is not 0, it is not caused by accidental factors; if all the multiple sets of monitoring error data are 0, it is caused by accidental factors.

[0072] In the above technical solution, when determining the laser radar parameter correction value based on multiple monitoring error data, the device angle error value of the laser radar is determined based on multiple monitoring error data, and the mean analysis is performed based on the device angle error value of the laser radar to determine the angle correction value.

[0073] The above technical solution avoids errors in the monitoring area caused by differences in LiDAR equipment through equipment calibration, thereby ensuring the accuracy of point cloud acquisition data. Moreover, when the predicted monitoring area is inconsistent with the target monitoring area, accidental factors are avoided by re-acquiring multiple sets of predicted data, avoiding the waste of time in equipment calibration caused by the false first analysis results, thereby improving the effectiveness of equipment calibration.

[0074] like Figure 4 As shown, in one embodiment provided by the present invention, data preprocessing is performed on point cloud acquisition data, including:

[0075] B1. Perform preprocessing operations on the point cloud acquisition data to obtain point cloud preprocessing data;

[0076] B2. Perform surface reconstruction on the point cloud preprocessing data to obtain a triangular mesh model of the point cloud preprocessing data;

[0077] B3. In the triangular mesh model, perform curvature analysis on each point to obtain the point cloud curvature value;

[0078] B4. Determine a first threshold value according to the point cloud curvature value, and screen and determine the feature key points of the point cloud data according to the first threshold value;

[0079] B5. Perform normal analysis based on the key points of point cloud data features.

[0080] In the above technical solution, when preprocessing operations are performed on the point cloud acquisition data, preprocessing operations such as filtering and denoising are performed on the point cloud acquisition data.

[0081] In the above technical solution, when performing curvature analysis on each point, the curvature value of the point is calculated by the following formula:

[0082]

[0083] K h represents the curvature value of the hth point, I h represents the distance between the hth point and its adjacent point, and θ represents the angle between the line connecting the hth point and its adjacent point and the normal line.

[0084] In the above technical solution, when the first threshold is determined according to the point cloud curvature value, the first threshold is calculated by the following formula:

[0085]

[0086] In the above formula, t represents the first threshold value, represents the mean of the point cloud collection data, N represents the total number of point cloud collection data, K i Represents the curvature value of the i-th point.

[0087] In the above technical solution, when normal analysis is performed based on the key points of the point cloud data, the following formula is used for calculation:

[0088]

[0089] In the above formula, n represents the normal of the key point of the point cloud data feature, e j Represents the location data information of the key point j of the point cloud data feature, e i Represents the location data information of the key point i of the point cloud data feature.

[0090] The above technical solution improves the accuracy and efficiency of feature extraction by preprocessing the point cloud data, and visualizes the point cloud data by surface reconstruction of the preprocessed point cloud data, so that the features of the point cloud data can be presented more intuitively, providing convenience for key point analysis. The triangular mesh model can not only improve the efficiency of curvature analysis through parallel processing, but also spatialize the point cloud data, making it more consistent with actual application scenarios in the fields of self-driving cars, robots and drones, and improving the accuracy of the lidar system. In addition, normal analysis based on the key points of the point cloud data features is convenient for subsequent analysis.

[0091] In one embodiment provided by the present invention, when encoding is performed using an encoder based on a multi-layer perceptron according to the feature key points of the point cloud data, the number of input layers, hidden layers, and output layers is configured according to the feature key points of the point cloud data, and the initialization weights and initial bias values ​​of the neurons are determined. Then, the data information of the feature key points of the point cloud data is input into the input layer, and forward propagation is performed through each layer of the network. Each neuron receives the output of the neurons in the previous layer, and the weighted sum is performed after nonlinear transformation through a modified linear unit activation function to obtain output information. Then, the first loss function is calculated according to the output information and the true label, and based on the gradient of the first loss function, the loss gradient of each neuron is calculated layer by layer starting from the output layer, and the loss gradient is multiplied by the connection weight according to the chain rule and passed to the previous layer. Then, the weight and bias value of each neuron are updated according to the gradient descent rule, and the network parameters are adjusted for multiple iterations. When the loss function is minimized, the global features of the point cloud acquisition data are obtained.

[0092] In the above technical solution, when configuring the number of input layers, hidden layers and output layers according to the key points of the point cloud data features, the number of input layer features, the number of neurons in the hidden layer and the number of layers in the output layer are set according to the input sample data. For example, the number of input layer features is set to 64, the number of neurons in the hidden layer is set to 128 and the number of layers in the output layer is set to 1024.

[0093] In the above technical solution, the rectified linear unit is Rectified Linear Unit, ReLU. When the output of the layer is set to 0, ReLU only retains positive elements and discards negative elements. If an element x is given, ReLU is defined as the maximum value of the element and 0 ReLU(x)=max(x, 0).

[0094] In the above technical solution, when calculating the first loss function according to the output information and the true label, the first loss function is calculated by the following formula:

[0095]

[0096] In the above formula, d CD represents the first loss function, S represents the point cloud before encoding, S′ represents the point cloud after encoding, N represents the total number of point clouds, x represents the point cloud data before encoding, and y represents the point cloud data after encoding;

[0097] In the above technical solution, when updating the weight and bias value of each neuron, the first update is to update the initial weight and initial bias value of the determined neuron, and each subsequent iterative update is to update the weight and bias value of the last update.

[0098] The above technical solution realizes loss optimization analysis in the encoding process through multiple iterations, so as to optimize the weight and bias value data, thereby reducing the encoding error and ensuring the accuracy of the global features of the point cloud acquisition data when encoding according to the key points of the point cloud data features using an encoder based on a multi-layer perceptron.

[0099] In one embodiment provided by the present invention, when performing the total loss analysis, point cloud acquisition data is obtained from the encoder, and mean and variance analysis is performed on the point cloud acquisition data to obtain the mean and variance of the point cloud acquisition data, and the latent variables of the hidden layer in the variational autoencoder are determined based on the mean and variance of the point cloud acquisition data. Then, based on the latent variables of the hidden layer in the variational autoencoder, the KL divergence between the Gaussian distribution and the normal distribution is calculated using the variational autoencoder to obtain the second loss function. Finally, the total loss function is determined based on the first loss function and the second loss function.

[0100] In the above technical solution, when determining the hidden variables of the hidden layer in the variational autoencoder according to the mean and variance of the point cloud acquisition data, in the sampling module, the corresponding Gaussian distribution random numbers are generated according to the mean and variance, and the Gaussian distribution random numbers are the hidden variables of the hidden layer in the variational autoencoder.

[0101] In the above technical solution, based on the latent variables of the hidden layer in the variational autoencoder, when the variational autoencoder is used to calculate the KL divergence between the Gaussian distribution and the normal distribution, the KL divergence between the Gaussian distribution N(μ, σ) and the normal distribution N(0, 1) is calculated by the following formula:

[0102]

[0103] In the above formula, d KL (N(μ, σ)||N(0, 1)) represents the KL divergence between the Gaussian distribution N(μ, σ) and the normal distribution N(0, 1), m represents the total number of points in the point cloud collection data, μ represents the mean of the point cloud collection data, and σ represents the variance of the point cloud collection data.

[0104] In the above technical solution, the total loss function is equal to the first loss function plus the second loss function, that is:

[0105] Loss = d CD +d KL

[0106] In the above formula, Loss represents the total loss function, d CD represents the first loss function, d KL represents the second loss function.

[0107] The above technical solution calculates the KL divergence between the Gaussian distribution and the normal distribution through the variational compiler, so that the second loss function can be obtained efficiently and accurately. At the same time, the total loss function is determined according to the first loss function and the second loss function to avoid negligence in loss calculation, thereby ensuring the comprehensiveness and accuracy of the total loss function.

[0108] In one embodiment provided by the present invention, when 3D point cloud is reconstructed and decoded based on the global features of point cloud acquisition data, a decoder is used to reconstruct and decode the 3D point cloud, the global features of the point cloud acquisition data are input into the decoder, and point cloud reconstruction data samples are obtained through the decoder.

[0109] In the above technical solution, the steps of the decoder are the same as the steps of encoding according to the feature key points of the point cloud data using an encoder based on a multi-layer perceptron.

[0110] The above technical solution can not only improve the efficiency of 3D point cloud reconstruction and decoding by using a decoder to reconstruct and decode 3D point cloud, but also reduce the error of 3D point cloud reconstruction and decoding, and ensure the accuracy of point cloud reconstruction data samples.

[0111] In one embodiment provided by the present invention, when performing reconstruction error analysis calculation, it includes:

[0112] For the point cloud reconstruction data samples, point cloud distance analysis is performed in turn according to the point cloud to obtain the point cloud distance before reconstruction and the point cloud distance after reconstruction;

[0113] The relative distance of the point cloud is calculated by combining the distance of the point cloud before reconstruction with the distance of the point cloud after reconstruction;

[0114] The point cloud reconstruction error value is obtained by using the EMD distance calculation based on the relative distance of the point cloud.

[0115] In the above technical solution, when performing point cloud distance analysis on the point cloud reconstruction data samples in sequence according to the point clouds, the distance between each reconstructed point cloud S′ and the point cloud before reconstruction S is calculated by the following formula:

[0116]

[0117] In the above formula, D total (S, S′) represents the distance between the reconstructed point cloud S′ and the reconstructed point cloud S, Δh i represents the height difference of the i-th point before and after reconstruction, Δr i represents the distance difference between the i-th point and the mass point before and after reconstruction, d i It represents the Euclidean distance between the i-th point and the nearest corresponding point in another point cloud before and after reconstruction.

[0118] In the above technical solution, in order to better reflect the actual differences between point clouds, we can change from directly accumulating the point pair differences to considering the complexity of the optimal distribution mapping. When the EMD distance is used to calculate the point cloud reconstruction error value, the point cloud reconstruction error value is analyzed by the following formula:

[0119]

[0120] In the above formula, d EMD (S, S′) represents the EMD distance between the point clouds before and after reconstruction, φ represents the mapping from the point cloud S before reconstruction to the point cloud S′ after reconstruction, which attempts to “move” the points in S to the points in S′ with the minimum total distance, and ||x-φ(x)||2 represents the Euclidean distance between the point x in the point cloud S before reconstruction and its mapping point φ(x) in the point cloud S′ after reconstruction.

[0121] The above technical solution realizes distance analysis of all points by performing point cloud distance analysis on point cloud reconstruction data samples in sequence according to the point cloud, thereby making it possible to analyze the reconstruction error based on the distance between points in the point cloud before and after reconstruction.

[0122] In one embodiment provided by the present invention, when determining the adversarial sample detection result based on the reconstruction error value combined with a preset threshold, the reconstruction error value is compared with the preset threshold. When the reconstruction error value is greater than or equal to the preset threshold, the point cloud reconstructed data sample is an adversarial sample. When the reconstruction error value is less than the preset threshold, the point cloud reconstructed data sample is a normal sample.

[0123] In the above technical solution, the preset threshold value is equal to the first threshold value in terms of numerical value.

[0124] The above technical solution realizes the screening of adversarial samples by presetting a threshold, so that adversarial samples can be detected when the pet error value is large, thereby improving the detection capability of adversarial samples.

[0125] In one embodiment provided by the present invention, after obtaining the triangular mesh model of the point cloud preprocessing data, optimization processing is further performed on the triangular mesh model, including:

[0126] Perform a preliminary analysis on the triangular mesh model of the point cloud preprocessing data to determine whether the triangular mesh model of the point cloud preprocessing data needs to be optimized and obtain preliminary analysis results;

[0127] According to the preliminary analysis results, when the triangular mesh model of the point cloud preprocessing data needs to be optimized, the distribution of the point cloud preprocessing data in the triangular mesh model is analyzed to obtain the data distribution of the triangular mesh model;

[0128] The optimization scheme is matched according to the distribution of the triangular mesh model, and the triangular mesh model is adjusted based on the optimization scheme to obtain a triangular mesh model.

[0129] In the above technical solution, when performing preliminary analysis on the triangular mesh model of the point cloud preprocessing data, the shape of the triangular mesh model of the exhaled point cloud preprocessing data is determined, and an overall characteristic analysis is performed based on the shape of the triangular mesh model to preliminarily determine whether the triangular mesh model of the point cloud preprocessing data needs to be optimized.

[0130] In the above technical solution, when analyzing the distribution of point cloud preprocessing data in the triangular mesh model, the vertices of the triangular mesh model are obtained, and the structure of the triangular mesh model is determined based on the vertices, thereby determining the data distribution of the triangular mesh model.

[0131] In the above technical solution, the optimization solution is matched according to the distribution of the triangular mesh model, including:

[0132] Perform edge analysis based on the data distribution of the triangular mesh model to determine the edges to be optimized;

[0133] Combine the network adjustment criteria to screen the edges to be optimized and determine the target optimization edges;

[0134] Analyze the characteristics of the target optimization edge to obtain characteristic information of the target optimization edge; wherein the characteristics include: connection characteristics of vertices, attribute characteristics of edges, relative positions, etc.;

[0135] According to the feature information, the optimization scheme is matched in the data storage unit to obtain the optimization scheme of the target optimization edge;

[0136] The optimization adjustment plan is determined according to the number of target optimization edges. When the number of target optimization edges is one, the optimization plan of the target optimization edge is the final optimization plan. When the number of target optimization edges is not unique, the edge optimization adjustment sequence is determined in combination with the structure of the triangular mesh model, and the final optimization plan is determined based on the optimization plan of the target optimization edge according to the edge optimization adjustment sequence.

[0137] In the above technical solution, a plurality of optimization solutions are stored in the data storage unit, and each optimization solution corresponds to a different edge feature.

[0138] The above technical solution performs preliminary analysis on the triangular mesh model of the point cloud preprocessing data so that when the triangular mesh model of the point cloud preprocessing data needs to be optimized, the triangular mesh model is optimized, thereby optimizing the triangular mesh model, making the triangular mesh model more uniform and smooth, thereby facilitating the analysis of the midpoints of the triangular mesh model, and also improving the accuracy of the point cloud curvature value.

[0139] Those skilled in the art should understand that the first and second in the present invention merely refer to different application stages.

[0140] Other embodiments of the present disclosure will be readily apparent to those skilled in the art after considering the specification and practicing the disclosure disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0141] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A laser radar recognition adversarial sample method based on 3D point cloud reconstruction, characterized in that: include: Use laser radar to monitor the surrounding environment and obtain point cloud data; Perform data preprocessing on point cloud acquisition data to obtain key feature points of point cloud data; According to the key points of the point cloud data features, an encoder based on a multi-layer perceptron is used to encode the data, output the global features of the point cloud data, and perform a total loss analysis; Reconstruct and decode the 3D point cloud based on the global features of the point cloud acquisition data, and perform reconstruction error analysis and calculation to obtain the reconstruction error value; Determine the adversarial sample detection result based on the reconstruction error value combined with the preset threshold; According to the adversarial sample detection results, the reconstructed point cloud data is input into the sub-network for recognition.

2. The method according to claim 1, characterized in that Before using LiDAR to monitor the surrounding environment, the monitoring preparation work for LiDAR is carried out, and the monitoring preparation work includes: Set the laser radar parameters and confirm the laser radar with the set parameters; Install the laser radar with set parameters, calibrate the equipment, and obtain the laser radar equipment calibration result; Configure the LiDAR device connection based on the LiDAR device calibration results.

3. The method according to claim 2, characterized in that When calibrating the equipment, the LiDAR parameters are combined to predict and analyze the LiDAR monitoring area, including: Perform lidar monitoring simulation according to the equipment parameters to obtain the lidar's predicted monitoring area; Performing a comparative analysis on the predicted monitoring area and the target monitoring area to determine whether the predicted monitoring area is consistent with the target monitoring area, thereby obtaining a first analysis result; According to the first analysis result, when the predicted monitoring area is inconsistent with the target monitoring area, multiple groups of predicted data are re-acquired and the correction value is confirmed; the limit distance of the predicted monitoring area is analyzed and determined in each group of predicted data to obtain multiple groups of limit distances, and the multiple limit distances are respectively combined with the target monitoring area to determine multiple monitoring error data, and it is determined whether it is caused by accidental factors based on the multiple monitoring error data to obtain the second analysis result. When the second analysis result shows that it is not caused by accidental factors, the lidar parameter correction value is determined based on the multiple monitoring error data.

4. The method according to claim 1, characterized in that Data preprocessing is performed on point cloud acquisition data, including: Perform preprocessing operations on the point cloud acquisition data to obtain point cloud preprocessing data; Perform surface reconstruction on the point cloud preprocessing data to obtain a triangular mesh model of the point cloud preprocessing data; In the triangular mesh model, curvature analysis is performed on each point to obtain the point cloud curvature value; Determine a first threshold value according to the point cloud curvature value, and screen and determine the feature key points of the point cloud data according to the first threshold value; Normal analysis is performed based on the feature key points of point cloud data.

5. The method according to claim 1, characterized in that: When encoding the key points of the feature of the point cloud data using an encoder based on a multi-layer perceptron, the number of input layers, hidden layers, and output layers is configured according to the key points of the feature of the point cloud data, and the initial weights and initial bias values ​​of the neurons are determined. Then, the data information of the key points of the feature of the point cloud data is input into the input layer, and forward propagation is performed through each layer of the network. At each neuron, the output of the neurons in the previous layer is received, and the weighted sum is performed after nonlinear transformation through the modified linear unit activation function to obtain the output information. Then, the first loss function is calculated according to the output information and the true label, and based on the gradient of the first loss function, the loss gradient of each neuron is calculated layer by layer starting from the output layer, and the loss gradient is multiplied by the connection weight according to the chain rule and passed to the previous layer. Then, the weight and bias value of each neuron are updated according to the gradient descent rule. After multiple iterations, the network parameters are adjusted. When the loss function is minimized, the global characteristics of the point cloud acquisition data are obtained.

6. The method according to claim 5, characterized in that When performing total loss analysis, point cloud acquisition data is obtained from the encoder, and mean and variance analysis is performed on the point cloud acquisition data to obtain the mean and variance of the point cloud acquisition data, and the latent variables of the hidden layer in the variational autoencoder are determined based on the mean and variance of the point cloud acquisition data. Then, based on the latent variables of the hidden layer in the variational autoencoder, the KL divergence between the Gaussian distribution and the normal distribution is calculated using the variational autoencoder to obtain the second loss function. Finally, the total loss function is determined based on the first loss function and the second loss function.

7. The method according to claim 1, characterized in that When reconstructing and decoding 3D point clouds based on the global features of point cloud acquisition data, a decoder is used to reconstruct and decode 3D point clouds, the global features of the point cloud acquisition data are input into the decoder, and point cloud reconstruction data samples are obtained through the decoder.

8. The method according to claim 7, characterized in that When performing reconstruction error analysis calculation, it includes: For the point cloud reconstruction data samples, point cloud distance analysis is performed in turn according to the point cloud to obtain the point cloud distance before reconstruction and the point cloud distance after reconstruction; The relative distance of the point cloud is calculated by combining the distance of the point cloud before reconstruction with the distance of the point cloud after reconstruction; The point cloud reconstruction error value is obtained by using the EMD distance calculation based on the relative distance of the point cloud.

9. The method according to claim 8, characterized in that When determining the adversarial sample detection result based on the reconstruction error value combined with the preset threshold, the reconstruction error value is compared with the preset threshold. When the reconstruction error value is greater than or equal to the preset threshold, the point cloud reconstructed data sample is an adversarial sample. When the reconstruction error value is less than the preset threshold, the point cloud reconstructed data sample is a normal sample.

10. The method according to claim 4, characterized in that After obtaining the triangular mesh model of the point cloud preprocessing data, the triangular mesh model is optimized, including: Perform a preliminary analysis on the triangular mesh model of the point cloud preprocessing data to determine whether the triangular mesh model of the point cloud preprocessing data needs to be optimized and obtain preliminary analysis results; According to the preliminary analysis results, when the triangular mesh model of the point cloud preprocessing data needs to be optimized, the distribution of the point cloud preprocessing data in the triangular mesh model is analyzed to obtain the data distribution of the triangular mesh model; The optimization scheme is matched according to the distribution of the triangular mesh model, and the triangular mesh model is adjusted based on the optimization scheme to obtain a triangular mesh model.

Images