Radar data augmentation method and apparatus, electronic device, and storage medium

Abstract

The application provides a radar data augmentation method and device, electronic equipment and storage medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: determining the first data distribution difference between the source domain data set and the new radar data set based on at least one first feature corresponding to the source domain data set and at least one second feature corresponding to the new radar data set; determining a first loss function based on the first data distribution difference, and training a feature extractor based on the first loss function; inputting each radar data in the new radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor; and inputting the at least one third feature into an augmented model respectively to obtain at least one augmented radar data output by the augmented model. The application solves the problem of model generalization performance decline caused by different data distributions in multiple fields, improves the generalization ability of the augmented model, and further improves the data quality of the augmented radar data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a radar data augmentation method and device, an electronic device and a storage medium. BACKGROUND

[0002] With the rapid development of science and technology, application scenarios based on radar data for identification are becoming more and more widespread, such as target identification, target tracking, target positioning, etc. based on radar data. For example, traditional methods identify radar emitters based on pulse arrival time, arrival angle, repetition frequency, pulse width, carrier frequency and other characteristic parameters; however, the signal pattern of radar emitters is becoming increasingly complex and variable, and traditional methods are difficult to meet the current needs of radar emitter identification. Based on this, most current radar emitter identification models are based on radar emitter identification models, and the radar data received by the radar receiver is used to identify the emitter identification result. In order to improve the identification accuracy of the emitter identification model, a large amount of radar data is needed as sample data to train, test and verify the emitter identification model.

[0003] Currently, augmented radar data is generated by perturbing or transforming the original radar data. However, data augmentation based on original radar data mostly requires the introduction of noise or distortion, resulting in a decline in the quality of augmented radar data, which in turn affects the accuracy of the required training model; and the above operation is prone to introducing repetitive information, making the augmented radar data too monotonous and lacking in diversity, thereby affecting the generalization ability of the required training model. SUMMARY

[0004] The present application provides a radar data augmentation method, device, electronic device and storage medium to solve the defects of the prior art that the quality of augmented radar data is declining and the augmented radar data is too monotonous.

[0005] The present application provides a radar data augmentation method, comprising:

[0006] Obtain a source domain data set and an additional radar data set, the source domain data set comprising at least one sample radar data, and each radar data in the additional radar data set being data of the same domain;

[0007] Respectively input each sample radar data in the source domain data set into a feature extractor to obtain at least one first feature output by the feature extractor, and respectively input each radar data in the additional radar data set into the feature extractor to obtain at least one second feature output by the feature extractor;

[0008] Based on the at least one first feature and the at least one second feature, determine the first data distribution difference between the source domain data set and the additional radar data set;

[0009] determine a first loss function based on the first data distribution difference, and train the feature extractor based on the first loss function;

[0010] input each radar data in the new radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor;

[0011] input the at least one third feature into the augmented model respectively to obtain at least one augmented radar data output by the augmented model, the augmented model being trained based on the source domain data set.

[0012] According to the radar data augmentation method provided by the application, the at least one third feature output by the trained feature extractor is obtained by inputting each radar data in the new radar data set into the trained feature extractor respectively, and the method further comprises the following steps:

[0013] input the at least one first feature into the augmented model respectively to obtain at least one first sample radar data output by the augmented model;

[0014] determine a second loss function based on a second data distribution difference between the at least one first sample radar data and the new radar data set;

[0015] input the at least one first sample radar data into the domain discriminator respectively to obtain at least one first discrimination result output by the domain discriminator, determine a third loss function based on the at least one first discrimination result, and the domain discriminator is used to discriminate whether the input data is the data in the source domain data set or the data in the new radar data set;

[0016] train the feature extractor and the augmented model based on the second loss function and the third loss function.

[0017] According to the radar data augmentation method provided by the application, the at least one third feature output by the trained feature extractor is obtained by inputting each radar data in the new radar data set into the trained feature extractor respectively, and the method further comprises the following steps:

[0018] input the at least one first feature into the augmented model respectively to obtain at least one first sample radar data output by the augmented model, and input the at least one second feature into the augmented model respectively to obtain at least one second sample radar data output by the augmented model;

[0019] determine a fourth loss function based on a third data distribution difference between the at least one first sample radar data and the at least one second sample radar data;

[0020] iteratively optimize the augmented model based on the fourth loss function.

[0021] According to the radar data augmentation method provided by the application, the augmented model is a generator, and the iteratively optimizing the augmented model based on the fourth loss function comprises:

[0022] iteratively optimizing the augmented model based on the fourth loss function and a fifth loss function;

[0023] The fifth loss function is determined based on the following steps:

[0024] inputting the at least one first sample radar data into the discriminator corresponding to the augmented model respectively to obtain at least one second discrimination result output by the discriminator;

[0025] determining the fifth loss function based on the at least one second discrimination result; and / or,

[0026] inputting the at least one second sample radar data into the discriminator corresponding to the augmented model respectively to obtain at least one third discrimination result output by the discriminator;

[0027] determining the fifth loss function based on the at least one third discrimination result.

[0028] According to the radar data augmentation method provided by the application, the obtaining the new radar data set comprises:

[0029] obtaining a plurality of first radar data received by a plurality of radar receivers, and determining the new radar data set from the plurality of first radar data; and / or,

[0030] obtaining a plurality of second radar data received by a radar receiver, and determining the new radar data set from the plurality of second radar data.

[0031] According to the radar data augmentation method provided by the application, the inputting the radar data in the new radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor further comprises:

[0032] inputting the at least one first feature into the augmented model respectively to obtain at least one first sample radar data output by the augmented model;

[0033] inputting the at least one first sample radar data into a label predictor respectively to obtain at least one label prediction result output by the label predictor;

[0034] determine a sixth loss function based on at least one label difference between the at least one label prediction result and a true label set corresponding to the source domain data set;

[0035] train the feature extractor and the augmentation model based on the sixth loss function.

[0036] According to the radar data augmentation method provided by the application, the first loss function is determined based on the first data distribution difference, the feature extractor is trained based on the first loss function, and then the method further comprises:

[0037] input each sample radar data in the source domain data set into the trained feature extractor respectively to obtain at least one fourth feature output by the trained feature extractor;

[0038] input the at least one fourth feature into the augmentation model respectively to obtain at least one augmented radar data output by the augmentation model.

[0039] The application further provides a radar data augmentation device, comprising:

[0040] a data acquisition module configured to acquire a source domain data set and acquire an added radar data set, wherein the source domain data set comprises at least one sample radar data, and each radar data in the added radar data set is data of the same domain;

[0041] a first extraction module configured to input each sample radar data in the source domain data set into a feature extractor respectively to obtain at least one first feature output by the feature extractor, and input each radar data in the added radar data set into the feature extractor respectively to obtain at least one second feature output by the feature extractor;

[0042] a difference determination module configured to determine a first data distribution difference between the source domain data set and the added radar data set based on the at least one first feature and the at least one second feature;

[0043] a model training module configured to determine a first loss function based on the first data distribution difference, and train the feature extractor based on the first loss function;

[0044] a second extraction module configured to input each radar data in the added radar data set into a trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor;

[0045] a data augmentation module configured to input the at least one third feature into an augmentation model respectively to obtain at least one augmented radar data output by the augmentation model, wherein the augmentation model is trained based on the source domain data set.

[0046] The application further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the radar data augmentation method according to any one of the above when executing the program.

[0047] The application further provides a non-transitory computer-readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the radar data augmentation method according to any one of the above.

[0048] The radar data augmentation method, device, electronic device and storage medium provided by the application input each sample radar data in the source domain data set into a feature extractor respectively, obtain at least one first feature output by the feature extractor, input each radar data in the new radar data set into the feature extractor respectively, obtain at least one second feature output by the feature extractor, determine a first data distribution difference between the source domain data set and the new radar data set based on the at least one first feature and the at least one second feature, determine a first loss function based on the first data distribution difference, train the feature extractor based on the first loss function, so that the trained feature extractor can align the distribution of the source domain data set and the new radar data set, input each radar data in the new radar data set into the trained feature extractor respectively, obtain at least one third feature output by the trained feature extractor, that is, the domain-invariant representation of the source domain data set and the new radar data set can be obtained, thereby migrating the knowledge of the source domain to the target domain (the new radar data set), and input the at least one third feature into the augmentation model respectively, so that at least one augmented radar data output by the augmentation model with high quality can be obtained, thereby solving the problem of decline in model generalization performance caused by different data distributions in multiple fields, that is, improving the generalization ability of the augmented model, and further improving the data quality of the augmented radar data. Meanwhile, the at least one third feature corresponding to the new radar data set is input into the augmentation model respectively, at least one augmented radar data output by the augmentation model is obtained, and no disturbance or transformation is performed on the data of the source domain data set, so that no noise or distortion needs to be introduced, the quality of the augmented radar data is ensured, and the new radar data set can be data in each field, thereby improving the diversity of the augmented radar data, and finally improving the generalization ability of the required training model. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0050] Figure 1 One of the flowcharts of the radar data augmentation method provided by the present application;

[0051] Figure 2 One of the schematic diagrams of the augmented overall model provided by the present application;

[0052] Figure 3 One of the flowcharts of the radar data augmentation method provided by the present application;

[0053] Figure 4 One of the schematic diagrams of the augmented overall model provided by the present application;

[0054] Figure 5 One of the schematic diagrams of the augmented overall model provided by the present application;

[0055] Figure 6 One of the schematic diagrams of the augmented overall model provided by the present application;

[0056] Figure 7 One of the schematic diagrams of the augmented overall model provided by the present application;

[0057] Figure 8 One of the schematic diagrams of the radar data augmentation device provided by the present application;

[0058] Figure 9 One of the schematic diagrams of the electronic device provided by the present application. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below with reference to the drawings in the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0060] With the rapid development of science and technology, the application scenarios based on radar data recognition are more and more extensive, such as target recognition, target tracking, target positioning and the like based on radar data. For example, traditional methods identify radar emitters according to characteristic parameters such as pulse arrival time, arrival angle, repetition frequency, pulse width, carrier frequency and the like; however, with the development of multi-functional radars, the signal patterns of radar emitters are increasingly complex and variable, and the traditional methods are difficult to meet the current requirements of radar emitter identification, and generally have poor robustness and poor emitter identification effect. Based on this, at present, most of the radar emitter identification models are based on radar emitter identification models, and the radar data received by the radar receiver is used to identify the emitter identification result. In order to improve the identification accuracy of the emitter identification model, a large amount of radar data is needed as sample data to train, test and verify the emitter identification model.

[0061] Among them, the radar emitter identification is also called specific emitter identification (SEI), which refers to the process of extracting and analyzing signal characteristics by intercepting radar emitter signals to determine the individual, model and working state of the radar emitter. With the development of radar emitter identification technology in recent years, it has been applied in many scenarios, but there are still application problems. In the application of radar emitter identification technology, there is often a problem of real small sample data. Therefore, it is of great significance to study data augmentation for radar data.

[0062] At present, the original radar data is disturbed or transformed to generate augmented radar data, and the augmented radar data is used as new training data. For example, the first data augmentation method is to add Gaussian noise: Gaussian noise is added to the original radar intermediate frequency data, and the strength of the noise can be controlled by adjusting the standard deviation of the Gaussian noise; the second data augmentation method is to simulate target motion: target motion is added to the original radar intermediate frequency data, and the target motion can be simulated by moving the original radar intermediate frequency data in different directions by a certain distance, which can increase the diversity of the data and improve the recognition ability of the algorithm for moving targets; the third data augmentation method is random cropping: a part of the data in the original radar intermediate frequency data is randomly cropped; the fourth data augmentation method is to increase scaling and rotation: random scaling and rotation are added to the original radar intermediate frequency data, which can be realized by randomly scaling and rotating the original radar intermediate frequency data by a certain angle; the fifth data augmentation method is channel random arrangement: the order of the channels in the original radar intermediate frequency data is randomly arranged, which can be realized by randomly arranging the order of the channels in the original radar intermediate frequency data; the sixth data augmentation method is random translation: the original radar intermediate frequency data is randomly translated by a certain distance, which can be realized by randomly translating the original radar intermediate frequency data by a certain distance.

[0063] However, in the data augmentation based on the original radar data, noise or distortion needs to be introduced, which leads to the decline of the quality of the augmented radar data, and further affects the accuracy of the required training model. Moreover, the above operations are usually based on some simple transformation operations, such as translation, rotation, scaling, etc., which are easy to introduce repetitive information, making the augmented radar data too single and lack of diversity, thereby affecting the generalization ability of the required training model.

[0064] To solve the above problems, the present application provides the following embodiments. Figure 1 One of the flowcharts of the radar data augmentation method provided by the present application is shown in Figure 1 The radar data augmentation method comprises:

[0065] In step 110, a source domain data set is obtained, and a new radar data set is obtained.

[0066] The source domain data set comprises at least one sample radar data, and each radar data in the new radar data set is data of the same domain.

[0067] Here, the source domain data set is the original data set, i.e. the original training data set, which can be used to train the model. The data in the source domain data set can be used as source domain data. Further, the source domain data set is a labeled data set, i.e. each sample radar data has a corresponding actual label.

[0068] Here, the new radar data set is the current new radar data set, i.e. the data set newly obtained compared with the source domain data set. The data in the new radar data set can be used as target domain data. It can be understood that the new radar data set can also be used as training data to train the model. Further, the new radar data set is a data set of a different domain from the source domain data set.

[0069] In an embodiment, considering that radar intermediate frequency data is an important intermediate result of radar signal processing, it can be used for radar target recognition, target tracking, target positioning and other applications. For example, in target recognition, through feature extraction and classification of radar intermediate frequency data, automatic recognition of target type can be realized; in target tracking, through filtering, time series analysis and other processing of radar intermediate frequency data, tracking and prediction of the target can be realized. The characteristics of radar intermediate frequency data are high reliability and stability, which can adapt to the requirements of radar signal processing in different environments. Based on this, the radar data included in the source domain data set and the new radar data set can be radar intermediate frequency data, for example, LD radar intermediate frequency data.

[0070] The radar intermediate frequency data refers to data obtained after radar received signals (such as radio frequency signals) are subjected to intermediate frequency processing. The intermediate frequency processing is to down-convert high frequency signals received by a radar receiver to an intermediate frequency range, and perform signal processing operations such as filtering and amplification. These processed signals usually contain echo signals of targets and some noise, clutter and other interference signals.

[0071] In step 120, each sample radar data in the source domain data set is input into the feature extractor to obtain at least one first feature output by the feature extractor, and each radar data in the new radar data set is input into the feature extractor to obtain at least one second feature output by the feature extractor.

[0072] Here, the number of first features is consistent with the number of sample radar data in the source domain data set, and one sample radar data extracts one first feature. The number of second features is consistent with the number of radar data in the new radar data set, and one radar data extracts one second feature.

[0073] For example, if the data input into the feature extractor is radar intermediate frequency data, the feature extractor can extract fingerprint features to better perform radiation source identification.

[0074] In step 130, based on the at least one first feature and the at least one second feature, a first data distribution difference between the source domain data set and the new radar data set is determined.

[0075] Here, the first data distribution difference is used to represent the difference between the two, and further, to represent the difference between the probability distributions of the two.

[0076] In an embodiment, based on the at least one first feature and the at least one second feature, a maximum mean difference between the source domain data set and the new radar data set is determined; and based on the maximum mean difference, the first data distribution difference is determined. The maximum mean difference can be directly determined as the first data distribution difference, or the maximum mean difference can be further processed to obtain the first data distribution difference.

[0077] For example, the calculation formula of the maximum mean difference is as follows:

[0078]

[0079] In the formula, MMD(Xs, Xt) represents the maximum mean difference, Xs represents at least one first feature corresponding to the source domain data set, Xt represents at least one second feature corresponding to the new radar data set, n represents the number of samples of the source domain data set, k() represents a kernel function, Xi represents the i-th first feature of the at least one first feature, Let j represent the j-th first feature of at least one first feature, and m represent the number of samples in the newly added radar dataset. This represents the i-th second feature of the at least one second feature. This represents the j-th second feature of the at least one second feature.

[0080] Of course, this first data distribution difference can also be determined based on other methods, such as by measuring the Wasserstein distance (bulldozer distance).

[0081] Step 140: Determine a first loss function based on the first data distribution difference, and train the feature extractor based on the first loss function.

[0082] Considering that the newly added radar dataset may be a dataset from a different domain than the source dataset, resulting in a large data distribution difference between the newly added radar dataset and the source dataset, the training objective of the first loss function is to minimize the first data distribution difference, and thus the first loss function can be determined based on the first data distribution difference.

[0083] Understandably, since the training objective of the first loss function is to minimize the first data distribution difference, training the feature extractor based on the first loss function can align the distribution of the source domain dataset and the newly added radar dataset. This allows the trained feature extractor to extract domain-invariant representations for both the source and target domains. In other words, aligning the distribution of each pair of source domains (source dataset) and target domains (new radar dataset) in multiple specific feature spaces—that is, aligning the feature representations of the target domain with those of the source domain—reduces the distribution difference between the source and target domains. In other words, it enables the feature extractor to learn the shared features between the source and target domains and suppress their differences.

[0084] In one embodiment, to further enhance the generalization ability of the augmented model, i.e., to further reduce the distribution difference between the source and target domains, at least one first feature is input into the augmented model to obtain at least one first sample radar data output by the augmented model; a second loss function is determined based on the second data distribution difference between the at least one first sample radar data and the newly added radar dataset; and the feature extractor and the augmented model are trained based on the second loss function. That is, the feature extractor is trained jointly based on the first and second loss functions. The specific execution process of this embodiment is described in the following embodiment.

[0085] In an embodiment, in order to further enhance the generalization ability of the augmented model, i.e., to further reduce the distribution difference between the source domain and the target domain, the at least one first feature is respectively input into the augmented model to obtain at least one first sample radar data output by the augmented model; the at least one first sample radar data is respectively input into a domain discriminator to obtain at least one first discrimination result output by the domain discriminator, the domain discriminator is used to discriminate whether the input data is data in the source domain data set or data in the new radar data set; and a third loss function is determined based on the at least one first discrimination result. The feature extractor and the augmented model are trained based on the third loss function. That is, the feature extractor is trained based on the first loss function and the third loss function. The specific execution process of this embodiment is referred to the following embodiments.

[0086] In an embodiment, in order to further enhance the generalization ability of the augmented model, i.e., to further reduce the distribution difference between the source domain and the target domain, the at least one first feature and the at least one second feature are respectively input into a domain discriminator to obtain a plurality of discrimination results output by the domain discriminator, the domain discriminator is used to discriminate whether the input data is data in the source domain data set or data in the new radar data set; a discrimination loss function is determined based on the plurality of discrimination results; and the feature extractor and the domain discriminator are trained based on the discrimination loss function.

[0087] It can be understood that the training target of the feature extractor is to maximize the discrimination loss of the domain discriminator, so that the trained feature extractor can align the distributions of the source domain data set and the new radar data set based on the training of the feature extractor based on the discrimination loss function, so that the trained feature extractor can extract domain-invariant representations for the source domain data set and the new radar data set. The training target of the domain discriminator is to minimize the discrimination loss of the domain discriminator, so that the domain discriminator is trained based on the discrimination loss function, thereby realizing the adversarial training of the feature extractor and the domain discriminator. That is, the domain adversarial network is used for transfer learning, which can be regarded as a game process between two players, wherein the first player is a domain discriminator trained to distinguish the source domain and the target domain, and the second player is a feature extractor fine-tuned to confuse the ability of the domain discriminator.

[0088] Step 150, each radar data in the new radar data set is respectively input into the trained feature extractor to obtain at least one third feature output by the trained feature extractor.

[0089] Here, the number of third features is consistent with the number of radar data in the new radar data set, and one radar data extracts one third feature. The third feature is a domain-invariant representation of the source domain data set and the new radar data set.

[0090] At step 160, the at least one third feature is respectively input into the augmented model to obtain at least one augmented radar data output by the augmented model, the augmented model being trained based on the source domain data set.

[0091] Here, the number of augmented radar data is consistent with the number of third features, that is, one third feature corresponds to generate one augmented radar data. The augmented radar data can be used to train the model to further improve the robustness of the model.

[0092] It can be understood that the augmented model is trained based on the source domain data set, so it is necessary to transfer the knowledge of the source domain to the target domain (new radar data set), and the above training of the feature extractor can extract the transferable features (domain invariant representation), thereby reducing the distribution difference between the source domain and the target domain, and further enabling the augmented model to generate high-quality augmented radar data based on the third feature, solving the problem of model generalization performance decline caused by different data distributions in multiple fields, that is, improving the generalization ability of the augmented model, so that it performs better in the field of new radar data set.

[0093] In an embodiment, the augmented model is a generator, that is, the at least one third feature is respectively input into the generator to obtain at least one augmented radar data output by the generator.

[0094] For ease of understanding, as shown in Figure 2 First, each sample radar data in the source domain data set is respectively input into the feature extractor to obtain at least one first feature output by the feature extractor, and each radar data in the new radar data set is respectively input into the feature extractor to obtain at least one second feature output by the feature extractor; then, based on the first data distribution difference between the source domain data set and the new radar data set determined according to the at least one first feature and the at least one second feature, a first loss function for training the feature extractor is determined; finally, each radar data in the new radar data set is respectively input into the trained feature extractor to obtain at least one third feature output by the trained feature extractor, and the at least one third feature is respectively input into the augmented model to obtain at least one augmented radar data output by the augmented model.

[0095] The radar data augmentation method provided by the embodiment of the present application respectively inputs each sample radar data in the source domain data set to the feature extractor, obtains at least one first feature output by the feature extractor, respectively inputs each radar data in the new radar data set to the feature extractor, obtains at least one second feature output by the feature extractor, determines the first data distribution difference between the source domain data set and the new radar data set based on the at least one first feature and the at least one second feature, determines the first loss function based on the first data distribution difference, and trains the feature extractor based on the first loss function, so that the trained feature extractor can align the distributions of the source domain data set and the new radar data set. Each radar data in the new radar data set is respectively input to the trained feature extractor, at least one third feature output by the trained feature extractor is obtained, that is, the domain-invariant representation of the source domain data set and the new radar data set is obtained, thereby migrating the knowledge of the source domain to the target domain (the new radar data set), and the at least one third feature is respectively input to the augmentation model, and at least one augmented radar data with high quality output by the augmentation model can be obtained, thereby solving the problem of decline in model generalization performance caused by different data distributions in multiple fields, that is., improving the generalization ability of the augmented model, and further improving the data quality of the augmented radar data. At the same time, the at least one third feature corresponding to the new radar data set is respectively input to the augmentation model, at least one augmented radar data output by the augmentation model is obtained, and the data of the source domain data set is not disturbed or transformed, so that noise or distortion is not introduced, the quality of the augmented radar data is ensured, and the new radar data set can be data in each field, thereby improving the diversity of the augmented radar data, and finally improving the generalization ability of the required training model.

[0096] Based on the above embodiment, Figure 3 The flowchart of the radar data augmentation method provided by the present application is shown in Figure 2, wherein Figure 3 Before the above step 150, the method further comprises:

[0097] Step 310, respectively input the at least one first feature to the augmentation model to obtain at least one first sample radar data output by the augmentation model.

[0098] Here, the number of first sample radar data is consistent with the number of first features, that is, one first feature corresponds to generate one first sample radar data.

[0099] Step 320, based on the second data distribution difference between the at least one first sample radar data and the second data distribution of the new radar data set, determine the second loss function.

[0100] Here, the second data distribution difference is used to represent the difference between the two, and further, to represent the difference between the probability distributions of the two. ​

[0101] It is considered that the new radar data set can be a data set of a different domain from the source domain data set, so that there is a large data distribution difference between the new radar data set and the source domain data set. Therefore, the training target of the second loss function is to minimize the second data distribution difference, so that the second loss function can be determined based on the second data distribution difference.

[0102] It can be understood that, since the training target of the second loss function is to minimize the second data distribution difference, training the feature extractor and the augmentation model based on the second loss function can make the trained feature extractor align the distributions of the source domain data set and the new radar data set, so that the trained feature extractor can extract domain-invariant representations for the source domain data set and the new radar data set, and the trained augmentation model can generate more accurate augmented radar data.

[0103] In an embodiment, the second data distribution difference is determined based on a maximum mean difference between the at least one first sample radar data and the new radar data set. The maximum mean difference can be directly determined as the second data distribution difference, or the maximum mean difference can be further processed to obtain the second data distribution difference.

[0104] For example, the second loss function is as follows:

[0105] L MMD (G(Xs),Xt);

[0106] In the formula, Xs represents at least one first feature corresponding to the source domain data set, G(Xs) represents at least one first sample radar data output by the augmentation model, and Xt represents the new radar data set. The specific formula of the second loss function can refer to the calculation formula of the maximum mean difference described above.

[0107] Of course, the second data distribution difference can also be determined based on other ways, for example, by measuring the Wasserstein distance.

[0108] In step 330, the at least one first sample radar data is input into the domain discriminator respectively to obtain at least one first discrimination result output by the domain discriminator, and a third loss function is determined based on the at least one first discrimination result. The domain discriminator is used to discriminate whether the input data is data in the source domain data set or data in the new radar data set.

[0109] In consideration of the training target of the feature extractor being to maximize the discrimination loss of the domain discriminator, the third loss function is determined based on the at least one first discrimination result, so as to train the feature extractor based on the third loss function, which can align the distribution of the source domain dataset and the new radar dataset after training of the feature extractor, so that the feature extractor after training can extract domain-invariant representation for the source domain dataset and the new radar dataset. In addition, in consideration of the training target of the augmented model being to maximize the discrimination loss of the domain discriminator, so that the target domain sample generated by the augmented model is more difficult to be distinguished by the domain discriminator, so as to train the augmented model based on the third loss function, which can make the augmented radar data generated by the augmented model more accurate after training.

[0110] For example, the third loss function is as follows:

[0111]

[0112] In the formula, L adv (G(Xs),Yt) represents the third loss function, Yt represents the pseudo label corresponding to the new radar dataset, the pseudo label Yt can be input into the domain discriminator, so that the domain discriminator can obtain the first discrimination result; Xs represents at least one first feature corresponding to the source domain dataset, G(Xs) represents at least one first sample radar data output by the augmented model, n represents the number of samples of the source domain dataset, D(G(Xs i ) represents the first discrimination result, Xs i represents the i-th feature in the at least one first feature.

[0113] It should be noted that the third loss function is a loss function for training a generative adversarial network. In the field adaptation problem, the third loss function (adversarial loss function) can help to learn an augmented model that transfers the knowledge of the source domain to the target domain.

[0114] In an embodiment, at least one first sample radar data is input into the domain discriminator corresponding to the new radar dataset respectively, and at least one first discrimination result output by the domain discriminator is obtained. Specifically, a plurality of domain discriminators are trained in advance, so that the domain discriminator corresponding to the new radar dataset is determined from the plurality of domain discriminators. The plurality of domain discriminators are domain discriminators corresponding to different target domains, and the plurality of domain discriminators are domain discriminators corresponding to the same source domain. In other words, the plurality of domain discriminators can be used to capture a multi-modal structure, so that the present application can perform feature alignment on the data distribution of multiple fields.

[0115] Step 340, training the feature extractor and the augmented model based on the second loss function and the third loss function.

[0116] Specifically, a hybrid loss function is determined based on the second and third loss functions, and the feature extractor and augmentation model are trained based on the hybrid loss function.

[0117] For example, the hybrid loss function is shown below:

[0118] L G =L MMD (G(Xs),Xt)-λL adv (G(Xs),Yt);

[0119] In the formula, L G L represents the mixed loss function. MMD (G(Xs),Xt) represents the second loss function, where Xs represents at least one first feature corresponding to the source domain dataset, G(Xs) represents at least one first sample radar data output by the augmented model, Xt represents the newly added radar dataset, λ represents the weight coefficients, and L adv (G(Xs),Yt) represents the third loss function, and Yt represents the pseudo label corresponding to the newly added radar dataset.

[0120] For ease of understanding, such as Figure 4 As shown, firstly, each sample radar data from the source domain dataset is input into the feature extractor to obtain at least one first feature output by the feature extractor. Then, each radar data from the newly added radar dataset is input into the feature extractor to obtain at least one second feature output by the feature extractor. Next, based on the first data distribution difference between the source domain dataset and the newly added radar dataset determined by at least one first feature and at least one second feature, a first loss function is determined for training the feature extractor. Simultaneously, at least one first feature is input into the augmented model to obtain at least one first sample radar data output by the augmented model. Based on the difference between the at least one first sample radar data and the first data distribution difference between the source domain dataset and the newly added radar dataset, a first loss function is determined for training the feature extractor. The second data distribution difference in the newly added radar dataset is used to determine the second loss function for training the feature extractor and the augmented model. At least one first sample radar data is then input into the domain discriminator to obtain at least one first discrimination result output by the domain discriminator. Based on at least one first discrimination result, a third loss function is determined for training the feature extractor and the augmented model. Finally, each radar data in the newly added radar dataset is input into the trained feature extractor to obtain at least one third feature output by the trained feature extractor. At least one third feature is then input into the trained augmented model to obtain at least one augmented radar data output by the trained augmented model.

[0121] The radar data augmentation method provided by the embodiment of the present application trains the feature extractor and the augmentation model based on the second loss function and the third loss function to combine the two functions to form a hybrid loss function, thereby minimizing the second data distribution difference while maximizing the discrimination loss of the domain discriminator to further enhance the generalization ability of the augmentation model, that is, to further reduce the distribution difference between the source domain and the target domain, and finally enable the augmentation model to generate more realistic augmented radar data, that is, to improve the data quality of the augmented radar data.

[0122] Based on any of the above embodiments, considering that the above embodiments mainly focus on learning domain-invariant representations without considering domain-specific decision boundaries between classes, based on this, before the step 160, the method further comprises:

[0123] inputting the at least one first feature into the augmentation model respectively to obtain at least one first sample radar data output by the augmentation model, and inputting the at least one second feature into the augmentation model respectively to obtain at least one second sample radar data output by the augmentation model;

[0124] determining a fourth loss function based on a third data distribution difference between the at least one first sample radar data and the at least one second sample radar data;

[0125] iteratively optimizing the augmentation model based on the fourth loss function.

[0126] Here, the third data distribution difference is used to represent the difference between the two, and further, to represent the difference between the probability distributions of the two. The third data distribution difference can be measured based on the Wasserstein distance, the maximum mean difference, etc.

[0127] Considering that the new radar data set may be a data set of a different domain from the source domain data set, thereby causing a large data distribution difference between the new radar data set and the source domain data set, the training target of the fourth loss function is to minimize the third data distribution difference, thereby the fourth loss function can be determined based on the third data distribution difference. And considering that the target samples near the specific domain decision boundary generated by different augmentation models may obtain different labels, therefore, the output of the augmentation model is aligned using the specific domain decision boundary.

[0128] It can be understood that since the training target of the fourth loss function is to minimize the third data distribution difference, the augmentation model is trained based on the fourth loss function, so that the samples generated by the augmentation model for the source domain data set and the new radar data set are similar to a certain extent, thereby aligning the distributions of the two data sets, and further enabling the trained augmentation model to generate more accurate augmented radar data.

[0129] Further, the feature extractor is iteratively optimized based on the fourth loss function. Since the training target of the fourth loss function is to minimize the third data distribution difference, training the augmented model based on the fourth loss function can align the distribution of the source domain dataset and the new radar dataset extracted by the trained feature extractor, so that the trained feature extractor can extract domain-invariant representations for the source domain dataset and the new radar dataset.

[0130] For ease of understanding, as shown in Figure 5 As shown in the first place, each sample radar data in the source domain dataset is input into the feature extractor to obtain at least one first feature output by the feature extractor, and each radar data in the new radar dataset is input into the feature extractor to obtain at least one second feature output by the feature extractor; then, based on the first data distribution difference between the source domain dataset and the new radar dataset determined according to the at least one first feature and the at least one second feature, a first loss function for training the feature extractor is determined; at the same time, the at least one first feature is input into the augmented model to obtain at least one first sample radar data output by the augmented model, and the at least one second feature is input into the augmented model to obtain at least one second sample radar data output by the augmented model, and based on the third data distribution difference between the at least one first sample radar data and the at least one second sample radar data, a fourth loss function for training the augmented model is determined; finally, each radar data in the new radar dataset is input into the trained feature extractor to obtain at least one third feature output by the trained feature extractor, and the at least one third feature is input into the trained augmented model to obtain at least one augmented radar data output by the trained augmented model.

[0131] The radar data augmentation method provided by the embodiment of the present application inputs at least one first feature into the augmented model to obtain at least one first sample radar data output by the augmented model, and inputs at least one second feature into the augmented model to obtain at least one second sample radar data output by the augmented model, so as to determine the fourth loss function based on the third data distribution difference between the at least one first sample radar data and the at least one second sample radar data, and then iteratively optimize the augmented model based on the fourth loss function, so that the augmented model generated by the augmented model for the source domain dataset and the new radar dataset is similar to a certain extent, thereby aligning the distribution of the two datasets, and further enabling the trained augmented model to generate more accurate augmented radar data, i.e., improving the data quality of the augmented radar data.

[0132] Based on any of the above embodiments, in the method, the augmented model is a generator, and the iteratively optimizing the augmented model based on the fourth loss function comprises:

[0133] The augmented model is iteratively optimized based on the fourth and fifth loss functions.

[0134] The fifth loss function is determined based on the following steps:

[0135] The at least one first sample radar data is input into the discriminator corresponding to the augmented model to obtain at least one second discrimination result output by the discriminator; the fifth loss function is determined based on the at least one second discrimination result.

[0136] Alternatively, the fifth loss function can be determined based on the following steps:

[0137] The at least one second sample radar data is input into the discriminator corresponding to the augmented model to obtain at least one third discrimination result output by the discriminator; the fifth loss function is determined based on the at least one third discrimination result.

[0138] Alternatively, the fifth loss function can be determined based on the following steps:

[0139] The at least one first sample radar data is input into the discriminator corresponding to the augmented model to obtain at least one second discrimination result output by the discriminator. The at least one second sample radar data is input into the discriminator corresponding to the augmented model to obtain at least one third discrimination result output by the discriminator. The fifth loss function is determined based on the at least one second discrimination result and the at least one third discrimination result.

[0140] Understandably, the augmented model (generator) is responsible for generating synthetic data that is similar to real data, while the discriminator is responsible for distinguishing between synthetic and real data. Based on this, the augmented model and the discriminator undergo adversarial training. Therefore, the augmented model is trained based on the fifth loss function, that is, the generator and the discriminator play against each other and continuously optimize until the trained augmented model can generate synthetic data that is indistinguishable from real data. This allows the trained augmented model to generate more accurate augmented radar data, that is, improve the data quality of augmented radar data.

[0141] It should be noted that the above scheme can solve the decision boundary problem in a specific domain. Specifically, a generator (augmented model) generates new samples, and a discriminator identifies which samples are real. In the two-stage alignment scheme, the augmented model in the first stage is used to learn the distribution of the source domain dataset and generate samples similar to that dataset (at least one first sample radar data). The augmented model in the second stage is used to learn the distribution of the newly added radar dataset and generate samples similar to that dataset (at least one second sample radar data).

[0142] For ease of understanding, such as Figure 6 As shown, firstly, each sample radar data from the source domain dataset is input into the feature extractor to obtain at least one first feature output by the feature extractor. Then, each radar data from the newly added radar dataset is input into the feature extractor to obtain at least one second feature output by the feature extractor. Next, based on the first data distribution difference between the source domain dataset and the newly added radar dataset determined by at least one first feature and at least one second feature, a first loss function is determined for training the feature extractor. Simultaneously, at least one first sample radar data is input into the discriminator corresponding to the augmented model to obtain at least one second discrimination result output by the discriminator. Based on at least one second discrimination result, a fifth loss function is determined for training the augmented model, and / or, at least one second sample radar data is input into the discriminator corresponding to the augmented model to obtain at least one third discrimination result output by the discriminator, and a fifth loss function is determined for training the augmented model based on at least one third discrimination result; finally, each radar data in the newly added radar dataset is input into the trained feature extractor to obtain at least one third feature output by the trained feature extractor, and at least one third feature is input into the trained augmented model to obtain at least one augmented radar data output by the trained augmented model.

[0143] The radar data augmentation method provided in this embodiment of the invention trains the augmentation model based on the fifth loss function. That is, the generator (augmentation model) and the discriminator play each other and continuously optimize until the trained augmentation model can generate synthetic data that is indistinguishable from real data. In this way, the trained augmentation model can generate more accurate augmented radar data, that is, improve the data quality of augmented radar data.

[0144] Based on any of the above embodiments, step 110 above, obtaining the newly added radar dataset, includes:

[0145] Acquire multiple first radar data received by multiple radar receivers, and determine the newly added radar dataset from the multiple first radar data; and / or,

[0146] Acquire multiple types of second radar data received by the radar receiver, and determine the newly added radar dataset from the multiple types of second radar data.

[0147] Considering the influence of different radar receivers, the characteristic distribution of the first radar data received by different radar receivers is different. Therefore, the first radar data received by different radar receivers are divided into different domains. In this embodiment of the invention, the first radar data emitted by the same radar source and received by different radar receivers are considered to be in different domains.

[0148] Specifically, the new radar data set belonging to the same radar receiver is determined from the plurality of first radar data. It can be understood that the first radar data received by different radar receivers can be regarded as different target domains, and thus the above steps 110-160 are performed with the source domain data set (source domain) respectively.

[0149] Considering that the plurality of second radar data received by the same radar receiver can also be different from the source domain data set (source domain), and the plurality of second radar data can also be different from each other, different second radar data can be divided into different domains. For example, the second radar data received by the same radar receiver and emitted by different radar emitters are regarded as different domains.

[0150] The radar data augmentation method provided by the embodiment of the present application can obtain the new radar data set in the above manner, so that data augmentation based on the source domain data set is not necessarily required, thereby improving the quality of the augmented radar data and the diversity of the augmented radar data, and finally improving the generalization ability of the required training model.

[0151] Based on any of the above embodiments, before the above step 150, the method further comprises:

[0152] inputting the at least one first feature into the augmentation model respectively to obtain at least one first sample radar data output by the augmentation model;

[0153] inputting the at least one first sample radar data into the label predictor respectively to obtain at least one label prediction result output by the label predictor;

[0154] determining a sixth loss function based on at least one label difference between the at least one label prediction result and a true label set corresponding to the source domain data set;

[0155] training the feature extractor and the augmentation model based on the sixth loss function.

[0156] Here, the label predictor is used to predict whether the input data is data in the source domain data set or data in the new radar data set.

[0157] Here, the true label set includes the true labels corresponding to each data in the source domain data set.

[0158] Here, the training target of the sixth loss function is to minimize the label difference, i.e., to minimize the prediction loss of the label predictor, so as to determine the sixth loss function based on at least one label difference, so that training the feature extractor based on the sixth loss function can make the trained feature extractor align the distribution of the source domain data set and the newly added radar data set while ensuring the accuracy of label prediction as much as possible, i.e., aligning the data distribution while ensuring that there is a difference between them; training the augmented model based on the sixth loss function can make the trained augmented model generate augmented data as much as possible to ensure the accuracy of label prediction, i.e., aligning the data distribution while ensuring that there is a difference between them.

[0159] For ease of understanding, as shown in Figure 7 First, each sample radar data in the source domain data set is input into the feature extractor to obtain at least one first feature output by the feature extractor, and each radar data in the newly added radar data set is input into the feature extractor to obtain at least one second feature output by the feature extractor; then, based on the first data distribution difference between the source domain data set and the newly added radar data set determined according to the at least one first feature and the at least one second feature, a first loss function for training the feature extractor is determined; at the same time, the at least one first feature is input into the augmented model to obtain at least one first sample radar data output by the augmented model, the at least one first sample radar data is input into the label predictor to obtain at least one label prediction result output by the label predictor, and based on at least one label difference between the at least one label prediction result and the real label set corresponding to the source domain data set, a sixth loss function for training the feature extractor and the augmented model is determined; finally, each radar data in the newly added radar data set is input into the trained feature extractor to obtain at least one third feature output by the trained feature extractor, and the at least one third feature is input into the trained augmented model to obtain at least one augmented radar data output by the trained augmented model.

[0160] The radar data augmentation method provided by the embodiments of the present application inputs at least one first feature into an augmented model respectively, obtains at least one first sample radar data output by the augmented model, inputs the at least one first sample radar data into a label predictor respectively, obtains at least one label prediction result output by the label predictor, determines a sixth loss function based on at least one label difference between the at least one label prediction result and a real label set corresponding to the source domain data set, and trains the feature extractor based on the sixth loss function, so that the trained feature extractor can align the distribution of the source domain data set and the new radar data set while ensuring the accuracy of label prediction as much as possible, that is, aligning the data distribution while ensuring that there is a difference between them, thereby improving the effectiveness of the augmented radar data; the augmented model is trained based on the sixth loss function, so that the trained augmented model can generate augmented data as much as possible to ensure the accuracy of label prediction, that is, aligning the data distribution while ensuring that there is a difference between them, thereby improving the effectiveness of the augmented radar data.

[0161] Based on any of the above embodiments, after the above step 140, the method further comprises:

[0162] Inputting each sample radar data in the source domain data set into the trained feature extractor respectively to obtain at least one fourth feature output by the trained feature extractor;

[0163] Inputting the at least one fourth feature into the augmented model respectively to obtain at least one augmented radar data output by the augmented model.

[0164] Here, the number of fourth features is consistent with the number of sample radar data in the source domain data set, and one sample radar data extracts one fourth feature. The fourth feature is a domain-invariant representation of the source domain data set and the new radar data set.

[0165] Here, the number of augmented radar data is consistent with the number of fourth features, that is, one fourth feature corresponds to generate one augmented radar data. The augmented radar data can be used to train the model to further improve the robustness of the model.

[0166] The radar data augmentation method provided in the embodiments of the present application inputs at least one fourth feature corresponding to the source domain data set into the augmentation model respectively, obtains at least one augmented radar data output by the augmentation model, which is not a certain perturbation or transformation on the data of the source domain data set, thereby without introducing noise or distortion, ensuring the quality of the augmented radar data, improving the diversity of the augmented radar data, and finally improving the generalization ability of the required training model. Meanwhile, the at least one fourth feature corresponding to the source domain data set is extracted based on the trained feature extractor, compared with the data augmentation based on the features obtained based on the feature extractor before training, the embodiments of the present application can improve the number of augmented radar data, and finally further improve the accuracy of the required training model.

[0167] In order to facilitate the understanding of the above embodiments, a specific embodiment is described as follows. Figure 2 、 Figure 4 、 Figure 5 、 Figure 6 、 Figure 7As shown, first, each sample radar data in the source domain data set is input into the feature extractor to obtain at least one first feature output by the feature extractor, and each radar data in the new radar data set is input into the feature extractor to obtain at least one second feature output by the feature extractor; then, based on the first data distribution difference between the source domain data set and the new radar data set determined according to the at least one first feature and the at least one second feature, a first loss function for training the feature extractor is determined; at the same time, the at least one first feature is input into the augmentation model to obtain at least one first sample radar data output by the augmentation model, based on the second data distribution difference between the at least one first sample radar data and the new radar data set, a second loss function for training the feature extractor and the augmentation model is determined, and the at least one first sample radar data is input into the domain discriminator to obtain at least one first discrimination result output by the domain discriminator, based on the at least one first discrimination result, a third loss function for training the feature extractor and the augmentation model is determined; at the same time, the at least one second feature is input into the augmentation model to obtain at least one second sample radar data output by the augmentation model; based on the third data distribution difference between the at least one first sample radar data and the at least one second sample radar data, a fourth loss function for training the augmentation model is determined; at the same time, the at least one first sample radar data is input into the discriminator corresponding to the augmentation model to obtain at least one second discrimination result output by the discriminator, based on the at least one second discrimination result, a fifth loss function for training the augmentation model is determined, and / or the at least one second sample radar data is input into the discriminator corresponding to the augmentation model to obtain at least one third discrimination result output by the discriminator, based on the at least one third discrimination result, a fifth loss function for training the augmentation model is determined; at the same time, the at least one first sample radar data is input into the label predictor to obtain at least one label prediction result output by the label predictor, based on at least one label difference between the at least one label prediction result and the real label set corresponding to the source domain data set, a sixth loss function for training the feature extractor and the augmentation model is determined; finally, each radar data in the new radar data set is input into the trained feature extractor to obtain at least one third feature output by the trained feature extractor, and the at least one third feature is input into the trained augmentation model to obtain at least one augmented radar data output by the trained augmentation model; at the same time, each sample radar data in the source domain data set is input into the trained feature extractor to obtain at least one fourth feature output by the trained feature extractor, and the at least one fourth feature is input into the trained augmentation model to obtain at least one augmented radar data output by the trained augmentation model.

[0168] Based on the above embodiments, the application provides a scheme based on multi-source domain adaptive radar data augmentation, which is used to solve the distribution difference problem between different domain data and generate new augmented radar data. The scheme is realized based on a domain adversarial network, which includes two main components: a feature extractor and a domain discriminator. During training, the scheme uses multiple domain discriminators to capture the multi-modal structure and realizes feature alignment of different data distributions through adversarial-based sub-domain feature alignment. In addition, the scheme also solves the domain-specific decision boundary problem between classes.

[0169] The radar data augmentation device provided by the application is described below. The radar data augmentation device described below can be referred to with the radar data augmentation method described above.

[0170] Figure 8 The structure diagram of the radar data augmentation device provided by the application is shown as Figure 8 The radar data augmentation device comprises:

[0171] The data acquisition module 810 is configured to acquire a source domain data set and acquire a new radar data set, wherein the source domain data set comprises at least one sample radar data, and each radar data in the new radar data set is data of the same domain;

[0172] The first extraction module 820 is configured to input each sample radar data in the source domain data set into a feature extractor respectively to obtain at least one first feature output by the feature extractor, and input each radar data in the new radar data set into the feature extractor respectively to obtain at least one second feature output by the feature extractor;

[0173] The difference determination module 830 is configured to determine a first data distribution difference between the source domain data set and the new radar data set based on the at least one first feature and the at least one second feature;

[0174] The model training module 840 is configured to determine a first loss function based on the first data distribution difference, and train the feature extractor based on the first loss function;

[0175] The second extraction module 850 is configured to input each radar data in the new radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor;

[0176] The data augmentation module 860 is configured to input the at least one third feature into an augmentation model respectively to obtain at least one augmented radar data output by the augmentation model, wherein the augmentation model is trained based on the source domain data set.

[0177] The radar data augmentation device provided by the embodiment of the present application respectively inputs each sample radar data in the source domain data set to the feature extractor, obtains at least one first feature output by the feature extractor, respectively inputs each radar data in the new radar data set to the feature extractor, obtains at least one second feature output by the feature extractor, determines the first data distribution difference between the source domain data set and the new radar data set based on the at least one first feature and the at least one second feature, determines the first loss function based on the first data distribution difference, trains the feature extractor based on the first loss function, so that the trained feature extractor can align the distribution of the source domain data set and the new radar data set, respectively inputs each radar data in the new radar data set to the trained feature extractor, obtains at least one third feature output by the trained feature extractor, that is, the domain invariant representation of the source domain data set and the new radar data set can be obtained, thereby migrating the knowledge of the source domain to the target domain (the new radar data set), and by respectively inputting the at least one third feature to the augmentation model, at least one augmented radar data with high quality output by the augmentation model can be obtained, thereby solving the problem of decline of model generalization performance caused by different data distributions in multiple fields, that is, improving the generalization ability of the augmented model, and further improving the data quality of the augmented radar data. At the same time, the present application respectively inputs at least one third feature corresponding to the new radar data set to the augmentation model, obtains at least one augmented radar data output by the augmentation model, and is not a certain disturbance or transformation of the data of the source domain data set, so that noise or distortion does not need to be introduced, the quality of the augmented radar data is ensured, and the new radar data set can be data of each field, thereby improving the diversity of the augmented radar data, and finally improving the generalization ability of the required training model.

[0178] Figure 9 An example of an entity structure diagram of an electronic device is shown in FIG. 1. Figure 9As shown, the electronic device can include a processor 910, a communications interface 920, a memory 930, and a communications bus 940, wherein the processor 910, the communications interface 920, and the memory 930 complete mutual communication through the communications bus 940. The processor 910 can invoke a logic instruction in the memory 930 to execute a radar data augmentation method, which includes: obtaining a source domain data set, and obtaining an added radar data set, the source domain data set including at least one sample radar data, and each radar data in the added radar data set being data of the same domain; inputting each sample radar data in the source domain data set into a feature extractor respectively to obtain at least one first feature output by the feature extractor, and inputting each radar data in the added radar data set into the feature extractor respectively to obtain at least one second feature output by the feature extractor; determining a first data distribution difference between the source domain data set and the added radar data set based on the at least one first feature and the at least one second feature; determining a first loss function based on the first data distribution difference, training the feature extractor based on the first loss function; inputting each radar data in the added radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor; and inputting the at least one third feature into an augmentation model respectively to obtain at least one augmented radar data output by the augmentation model, the augmentation model being trained based on the source domain data set.

[0179] In addition, the logic instruction in the memory 930 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0180] In another aspect, the present application also provides a computer program product comprising a computer program, which can be stored on a non-transitory computer-readable storage medium, and the computer program can be executed by a processor to enable a computer to perform the radar data augmentation method provided by the above-mentioned method, which comprises: obtaining a source domain data set and obtaining an added radar data set, the source domain data set comprising at least one sample radar data, and each radar data in the added radar data set being data of the same domain; inputting each sample radar data in the source domain data set into a feature extractor respectively to obtain at least one first feature output by the feature extractor, and inputting each radar data in the added radar data set into the feature extractor respectively to obtain at least one second feature output by the feature extractor; determining a first data distribution difference between the source domain data set and the added radar data set based on the at least one first feature and the at least one second feature; determining a first loss function based on the first data distribution difference, and training the feature extractor based on the first loss function; inputting each radar data in the added radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor; and inputting the at least one third feature into an augmentation model respectively to obtain at least one augmented radar data output by the augmentation model, the augmentation model being trained based on the source domain data set.

[0181] In yet another aspect, the present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and the computer program can be executed by a processor to implement the radar data augmentation method provided by the above-mentioned method, which comprises: obtaining a source domain data set and obtaining an added radar data set, the source domain data set comprising at least one sample radar data, and each radar data in the added radar data set being data of the same domain; inputting each sample radar data in the source domain data set into a feature extractor respectively to obtain at least one first feature output by the feature extractor, and inputting each radar data in the added radar data set into the feature extractor respectively to obtain at least one second feature output by the feature extractor; determining a first data distribution difference between the source domain data set and the added radar data set based on the at least one first feature and the at least one second feature; determining a first loss function based on the first data distribution difference, and training the feature extractor based on the first loss function; inputting each radar data in the added radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor; and inputting the at least one third feature into an augmentation model respectively to obtain at least one augmented radar data output by the augmentation model, the augmentation model being trained based on the source domain data set.

[0182] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0183] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0184] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A radar data augmentation method, characterized by, The method comprises the following steps: obtaining a source domain data set and a new radar data set, wherein the source domain data set comprises at least one sample radar data, and each radar data in the new radar data set is data of the same domain; inputting each sample radar data in the source domain data set into a feature extractor respectively to obtain at least one first feature output by the feature extractor, and inputting each radar data in the new radar data set into the feature extractor respectively to obtain at least one second feature output by the feature extractor; determining a first data distribution difference between the source domain data set and the new radar data set based on the at least one first feature and the at least one second feature; training the feature extractor based on a first loss function determined based on the first data distribution difference; inputting each radar data in the new radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor; inputting the at least one third feature into an augmentation model respectively to obtain at least one augmented radar data output by the augmentation model, wherein the augmentation model is trained based on the source domain data set.

2. The radar data augmentation method of claim 1, wherein, Before the step of inputting each radar data in the new radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor, the method further comprises the following steps: inputting the at least one first feature into the augmentation model respectively to obtain at least one first sample radar data output by the augmentation model; determining a second loss function based on a second data distribution difference between the at least one first sample radar data and the new radar data set; inputting the at least one first sample radar data into a domain discriminator respectively to obtain at least one first discrimination result output by the domain discriminator, determining a third loss function based on the at least one first discrimination result, wherein the domain discriminator is used to discriminate whether the input data is data in the source domain data set or data in the new radar data set; training the feature extractor and the augmentation model based on the second loss function and the third loss function.

3. The radar data augmentation method of claim 1, wherein, Before the step of inputting the at least one third feature into the augmentation model respectively to obtain at least one augmented radar data output by the augmentation model, the method further comprises the following steps: inputting the at least one first feature into the augmentation model respectively to obtain at least one first sample radar data output by the augmentation model, and inputting the at least one second feature into the augmentation model respectively to obtain at least one second sample radar data output by the augmentation model; determining a fourth loss function based on a third data distribution difference between the at least one first sample radar data and the at least one second sample radar data; iteratively optimizing the augmentation model based on the fourth loss function.

4. The radar data augmentation method of claim 3, wherein, The augmentation model is a generator, and the step of iteratively optimizing the augmentation model based on the fourth loss function comprises the following steps: iteratively optimizing the augmentation model based on the fourth loss function and a fifth loss function. The fifth loss function is determined based on the following steps: The at least one second sample radar data is respectively input into the discriminator corresponding to the augmented model to obtain at least one third discrimination result output by the discriminator; The fifth loss function is determined based on the at least one third discrimination result; and / or The at least one second sample radar data is respectively input into the discriminator corresponding to the augmented model to obtain at least one third discrimination result output by the discriminator; The fifth loss function is determined based on the at least one third discrimination result.

5. The radar data augmentation method of claim 1, wherein, The new radar data set is obtained, including: The first radar data received by the radar receiver is obtained, and the new radar data set is determined from the first radar data; and / or The second radar data received by the radar receiver is obtained, and the new radar data set is determined from the second radar data.

6. The radar data augmentation method of claim 1, wherein, The at least one third feature output by the trained feature extractor is obtained by respectively inputting each radar data in the new radar data set into the trained feature extractor. The at least one first sample radar data is respectively input into the label predictor to obtain at least one label prediction result output by the label predictor; The sixth loss function is determined based on at least one label difference between the at least one label prediction result and the real label set corresponding to the source domain data set; The feature extractor and the augmented model are trained based on the sixth loss function. The at least one fourth feature output by the trained feature extractor is obtained by respectively inputting each sample radar data in the source domain data set into the trained feature extractor.

7. The radar data augmentation method of claim 1, wherein, The at least one augmented radar data output by the augmented model is obtained by respectively inputting the at least one fourth feature into the augmented model. Including: The data acquisition module is used for acquiring the source domain data set and acquiring the new radar data set, the source domain data set includes at least one sample radar data, and each radar data in the new radar data set is data of the same domain; 8. A radar data augmentation apparatus, characterized by, The first extraction module is used for respectively inputting each sample radar data in the source domain data set into the feature extractor to obtain at least one first feature output by the feature extractor, and respectively inputting each radar data in the new radar data set into the feature extractor to obtain at least one second feature output by the feature extractor; The difference determination module is used for determining the first data distribution difference between the source domain data set and the new radar data set based on the at least one first feature and the at least one second feature; The model training module is used for determining the first loss function based on the first data distribution difference, and training the feature extractor based on the first loss function; ​ ​ a second extraction module configured to input each radar data in the new radar data set into the trained feature extractor respectively to obtain at least one third feature output by the trained feature extractor; a data augmentation module configured to input the at least one third feature into an augmentation model respectively to obtain at least one augmented radar data output by the augmentation model, the augmentation model being trained based on the source domain data set.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the radar data augmentation method according to any one of claims 1 to 7 when executing the program.

10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program enables the processor to implement the radar data augmentation method according to any one of claims 1 to 7 when executed.

Images