A sparse trajectory recovery method based on a generative adversarial network

Abstract

A sparse trajectory recovery method based on a generative adversarial network. The method comprises: for each iteration: obtaining a first trajectory sample data set comprising real sparse trajectory data samples and real dense trajectory data samples; feeding the first trajectory sample data set to a GAN for a discriminator of the GAN to output a first discrimination value; obtaining a second trajectory sample data set generated by a generator of the GAN processing the real sparse trajectory data samples; feeding the second trajectory sample data set to the GAN for the discriminator to output a second discrimination value; optimizing an overall objective function of the GAN to obtain network parameters of the generator; and updating the generator according to the obtained generator network parameters; when the current iteration is completed, determining whether to stop iteration; and in response to determining to stop iteration, taking the updated generator as a final generator to generate a completed dense trajectory data sample based on a to-be-completed sparse trajectory data sample.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of artificial intelligence, and particularly relates to a sparse trajectory recovery method based on a generative adversarial network. BACKGROUND

[0002] With the development of information and science and technology, future combat will change towards electronic and information direction. Especially in air defense and anti-missile combat, for high-speed moving objects such as missiles and warplanes, the combat will show strong real-time change of process, strong uncertainty of combat information, strong game confrontation and other characteristics. The prediction of target trajectory is the key research direction of battlefield game confrontation, and the target trajectory information obtained in the combat often has problems such as incomplete data and insignificant time correlation.

[0003] The methods described in this section can not have been previously conceived or made. Unless otherwise indicated, it should not be assumed that any of the methods described in this section qualify as prior art merely by virtue of their inclusion in this section. Similarly, issues mentioned in this section should not be interpreted as an admission that any of the issues were already commonly known. SUMMARY

[0004] The present disclosure provides a trajectory data processing method and device, an electronic device and a computer readable storage medium.

[0005] According to an aspect of the present disclosure, a track data processing method is provided for recovering a sparse track data sample with missing data points using a generative adversarial network, the generative adversarial network comprising a discriminator and a generator, the method comprising: for each of one or more iterations, performing the following steps: obtaining a first track sample dataset, the first track sample dataset containing real track data samples, the real track data samples including both real sparse track data samples with missing data points and real dense track data samples with complete data points; feeding the first track sample dataset to the generative adversarial network for the discriminator to output a first discrimination value for the first track sample dataset, the first discrimination value being used to calculate a function value of a discriminator objective function; obtaining a second track sample dataset, the second track sample dataset being generated by the generator processing the real sparse track data samples with missing data points; feeding the second track sample dataset to the generative adversarial network for the discriminator to output a second discrimination value for the second track sample dataset, the second discrimination value being used to calculate a function value of a generator objective function; optimizing an overall objective function of the generative adversarial network to obtain network parameters of the generator when the overall objective function takes an optimal value, wherein the overall objective function is constructed based on the discriminator objective function and the generator objective function; and updating the generator according to the obtained network parameters of the generator; when all steps of the current iteration are completed, determining whether an iteration stopping condition is satisfied; and in response to determining that the iteration stopping condition is satisfied, taking the updated generator as a final generator, for the final generator to generate a completed dense track data sample with complete data points based on a sparse track data sample with missing data points to be completed; or in response to determining that the iteration stopping condition is not satisfied, performing a next iteration based on the updated generator.

[0006] According to another aspect of the present disclosure, there is provided a track data processing apparatus for recovering a sparse track data sample with missing data points using a generative adversarial network, the generative adversarial network comprising a discriminator and a generator, the apparatus comprising: a single-round execution unit configured to, for each of one or more iterations, perform the following steps: obtaining a first track sample dataset, the first track sample dataset comprising real track data samples, the real track data samples including both real sparse track data samples with missing data points and real dense track data samples with complete data points; feeding the first track sample dataset to the generative adversarial network for the discriminator to output a first discrimination value for the first track sample dataset, the first discrimination value being used to calculate a function value of a discriminator objective function; obtaining a second track sample dataset, the second track sample dataset being generated by the generator processing the real sparse track data samples with missing data points; feeding the second track sample dataset to the generative adversarial network for the discriminator to output a second discrimination value for the second track sample dataset, the second discrimination value being used to calculate a function value of a generator objective function; optimizing an overall objective function of the generative adversarial network to obtain network parameters of the generator when the overall objective function takes an optimal value, wherein the overall objective function is constructed based on the discriminator objective function and the generator objective function; and updating the generator according to the obtained network parameters of the generator; an iteration determination unit configured to determine whether an iteration stop condition is satisfied when all the steps of the current iteration are completed; a final determination unit configured to, in response to determining that the iteration stop condition is satisfied, take the updated generator as a final generator, so that the final generator generates a completed dense track data sample with complete data points based on a sparse track data sample with missing data points to be completed; and an iteration triggering unit configured to, in response to determining that the iteration stop condition is not satisfied, perform a next iteration based on the updated generator.

[0007] According to yet another aspect of the present disclosure, there is provided an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program which, when executed by the at least one processor, implements the above method.

[0008] According to yet another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above method.

[0009] According to yet another aspect of the present disclosure, there is provided a computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the above method.

[0010] According to one or more embodiments of the present disclosure, by recovering the incomplete track data into dense track data through a track recovery technique based on a generative adversarial network, on one hand, the data volume requirement is significantly reduced compared with the baseline method, and on the other hand, the accuracy and completeness of track recovery are improved.

[0011] These and other aspects of the present disclosure will become apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0012] In the following description of the exemplary embodiments in conjunction with the accompanying drawings, more details, features and advantages of the present disclosure are disclosed, in which:

[0013] Figure 1 is a schematic diagram illustrating an example system in which various methods described herein can be implemented according to an exemplary embodiment;

[0014] Figure 2 is a flowchart illustrating a track data processing method according to an exemplary embodiment;

[0015] Figure 3 is a schematic diagram illustrating an architecture of a generative adversarial network according to an exemplary embodiment;

[0016] Figure 4 is a comparison diagram of visual results of track recovery by a method according to an exemplary embodiment and a baseline method;

[0017] Figure 5 is a schematic block diagram illustrating a track data processing apparatus according to an exemplary embodiment;

[0018] Figure 6 is a block diagram illustrating an exemplary computer device that can be applied to an exemplary embodiment. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to assist in understanding, and should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the present disclosure. Also, descriptions of well-known functions and structures are omitted in the following description for clarity and conciseness.

[0020] In the present disclosure, the terms "first", "second", etc. used in the description of various described examples are not intended to limit the positional relationship, timing relationship or importance relationship of the elements, and such terms are only used to distinguish one element from another. In some examples, the first element and the second element can refer to the same instance of the element, and in some cases, based on the context of the description, they can also refer to different instances.

[0021] The terms used in the description of various described examples in the present disclosure are only for the purpose of describing the specific examples, and are not intended to be limiting. Unless the context clearly indicates otherwise, if the number of elements is not specifically limited, the element can be one or more. In addition, the term "and / or" used in the present disclosure encompasses any one of the listed items and all possible combinations thereof.

[0022] With the development of information and science and technology, future combat will change towards electronic and information direction. Especially in air defense and anti-missile combat, for high-speed moving objects such as missiles and warplanes, the combat will show strong real-time change of process, strong uncertainty of combat information, strong game confrontation and other characteristics. The prediction of target track is the key research direction of battlefield game confrontation, and the target track information obtained in the combat often has problems such as incomplete data and insignificant time correlation. How to obtain target track with expected accuracy and completeness has become a problem to be solved in the combat process.

[0023] Based on this, this disclosure proposes a trajectory data processing method. First, a first trajectory sample dataset, comprising both real sparse trajectory data samples and real dense trajectory data samples, is acquired and fed into a Generative Adversarial Network (GAN). The discriminator of the network scores the realness of the real trajectory data samples. Then, the network's generator processes the real sparse trajectory data samples to generate a second trajectory sample dataset that approximates the real dense trajectory data samples. This second dataset is also fed into the GAN, and the discriminator scores the realness of the non-real trajectory data samples in the second trajectory sample dataset. Finally, the overall objective function of the network is optimized to improve the discriminator's ability (i.e., to better distinguish between real and non-real trajectory data) and the generator's ability (i.e., to generate increasingly more accurate and accurate trajectory data). The non-real track data (approaching real track data) compete against each other during the optimization process and gradually tend towards Nash equilibrium. Finally, after all rounds of iteration are completed, the final updated generator is obtained. This allows the generator to be used to generate complete dense track data samples based on sparse track data samples with missing data points. Thus, with the help of the method disclosed in this paper, incomplete track data can be restored into dense track data through track restoration technology based on generative adversarial networks. On the one hand, it significantly reduces the amount of data required compared with the baseline method, and on the other hand, it improves the accuracy and completeness of track restoration.

[0024] Exemplary embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0025] Figure 1 This is a schematic diagram illustrating an example system 100 in which various methods described herein may be implemented according to exemplary embodiments.

[0026] refer to Figure 1 The system 100 includes a client device 110, a server 120, and a network 130 that communicatively couples the client device 110 and the server 120.

[0027] Client device 110 includes a display 114 and a client application (APP) 112 that can be displayed on the display 114. Client application 112 can be an application that needs to be downloaded and installed before running, or a lightweight application (liteapp). If client application 112 is an application that needs to be downloaded and installed before running, client application 112 can be pre-installed on client device 110 and activated. If client application 112 is a mini-app, user 102 can directly run client application 112 on client device 110 without installing it, by searching for client application 112 in the host application (e.g., by the name of client application 112) or scanning the graphic code of client application 112 (e.g., barcode, QR code, etc.). In some embodiments, client device 110 can be any type of mobile computing device, including mobile computers, mobile phones, wearable computing devices (e.g., smartwatches, head-mounted devices including smart glasses, etc.), or other types of mobile devices. In some embodiments, the client device 110 may alternatively be a fixed computer device, such as a desktop computer, server computer, or other type of fixed computer device.

[0028] Server 120 is typically a server deployed by an Internet Service Provider (ISP) or Internet Content Provider (ICP). Server 120 can represent a single server, a cluster of multiple servers, a distributed system, or a cloud server providing basic cloud services (such as cloud databases, cloud computing, cloud storage, and cloud communications). It will be understood that, although... Figure 1 The diagram shows that server 120 communicates with only one client device 110, but server 120 can provide background services to multiple client devices simultaneously.

[0029] Examples of network 130 include combinations of local area networks (LANs), wide area networks (WANs), personal area networks (PANs), and / or communication networks such as the Internet. Network 130 can be wired or wireless. In some embodiments, technologies and / or formats including Hypertext Markup Language (HTML), Extensible Markup Language (XML), etc., are used to process data exchanged through network 130. Furthermore, encryption technologies such as Secure Sockets Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), and Internet Protocol Security (IPsec) can be used to encrypt all or some of the links. In some embodiments, custom and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies.

[0030] For the purposes of this disclosure's embodiments, Figure 1In the example, client application 112 can be a track data processing program that provides various functions for recovering sparse track data samples with missing data points, such as acquiring real track data samples, initializing a generative adversarial network (GAN), and / or setting GAN iteration conditions, etc. Correspondingly, server 120 can be a server used with the track data processing program. Server 120 can provide services related to sparse track data sample recovery to client application 112 running on client device 110. Alternatively, localized services can also be provided by client application 112 running on client device 110. As a non-limiting example, the method described below can be provided by… Figure 1 The system 100 is running.

[0031] Figure 2 This is a flowchart illustrating a track data processing method 200 according to an exemplary embodiment. Method 200 can be implemented on a client device (e.g., Figure 1 The execution is performed at the client device 110 shown, that is, the execution entity of each step of method 200 can be... Figure 1 The client device 110 shown. In some embodiments, method 200 can be performed on a server (e.g., Figure 1 The method 200 is executed at server 120 (as shown in the figure). In some embodiments, the method 200 may be executed in combination by a client device (e.g., client device 110) and a server (e.g., server 120).

[0032] According to embodiments of this disclosure, method 200 can be used to recover sparse track data samples with missing data points using a generative adversarial network, which may include a discriminator and a generator.

[0033] refer to Figure 2 The track data processing method 200 includes the following steps S210 to S240, wherein step S210 includes sub-steps S211 to S216.

[0034] Step S210 may include performing sub-steps S211 to S216 for each iteration in one or more iterations.

[0035] Sub-step S211: Obtain the first track sample dataset. The first track sample dataset contains real track data samples, which include both real sparse track data samples with missing data points and real dense track data samples with complete data points.

[0036] In the example, real-world trajectory data samples can be obtained from target scenarios such as actual combat, exercises, and simulated combat. That is, the target environment can be a real environment (e.g., an actual combat / exercise combat environment), a simulated environment (e.g., a simulated combat environment), or a hybrid environment (e.g., including both actual combat / exercise combat environments and simulated combat environments as components), and so on. As an example and not a limitation, real-world trajectory data samples can be associated with mobile units, which are targets (e.g., blue force targets, red force targets, etc.) that can move within the target environment (e.g., move randomly, move in formation, etc.).

[0037] As used in this article, the term track data sample refers to a data sample representing the trajectory of a mobile unit such as a ship or aircraft during navigation. Since mobile units are subject to various factors and interferences during navigation, resulting in navigation errors, the track is generally not a straight line, but there are exceptions.

[0038] As used herein, the term track sample dataset refers to a dataset consisting of track data samples. In the example, the first track sample dataset may be a real track sample dataset, which may consist of or contain real track data samples.

[0039] As used herein, the term missing data points refers to the absence of one or more data points that constitute the trajectory of a movable unit represented by a given track data sample, such that the track data sample visually exhibits undesirable features such as discontinuous segments and / or breaks compared to a trajectory with complete data points. Missing data points in a track data sample can be caused by factors such as terrain occlusion, data crosstalk, and packet loss.

[0040] In this disclosure, a track data sample with missing data points in the real track data sample is referred to as a real sparse track data sample. In contrast, a track data sample with no missing data points (i.e., complete data points) in the real track data sample is referred to as a real dense track data sample.

[0041] Sub-step S212: The first track sample dataset is fed into the generative adversarial network so that the discriminator can output a first discrimination value for the first track sample dataset. The first discrimination value is used to calculate the function value of the discriminator's objective function.

[0042] In the example, the first discriminant value output by the discriminator of the generative adversarial network for the first track sample dataset can be the discriminator's score regarding the realism of the real track data samples. In the example, from the discriminator's perspective, the higher the realism of the track data samples, the higher the score; the score range could be, for example, [0,1]. In the example, it can be used... Let θ represent the discriminator of the generative adversarial network, where θ d This represents the network parameters of the discriminator. Therefore, the input-output mapping relationship of the discriminator can be established, i.e., Here, x is the input to the discriminator (e.g., a trajectory data sample fed into the generative adversarial network), while y is the discriminator's output of the judgment value on the input x, that is, the discriminator's score on the authenticity of the input x.

[0043] In the example, the first discriminant value can be used to calculate the function value of the discriminator's objective function so that the function value of the discriminator's objective function can be correlated with the discriminator's performance (i.e., its ability to distinguish between real track data samples and non-real track data samples).

[0044] In the example, the discriminator's objective function can be the optimization target targeted when optimizing the discriminator, enabling the discriminator to update its network parameters according to the network parameters corresponding to the optimal value of its objective function, thereby achieving discriminator optimization. In the example, the optimal value of the discriminator's objective function can be the extreme value (e.g., maximum, minimum) across all iterations, or it can be the extreme value (e.g., maximum, minimum) across some iterations, etc. It is understood that the selection of the optimal value of the objective function may vary depending on the actual optimization process and / or optimization requirements, and this disclosure does not impose any restrictions on it.

[0045] Sub-step S213: Obtain the second track sample dataset. The second track sample dataset is generated by the generator by processing real sparse track data samples with missing data points.

[0046] In the example, since the real sparse track data sample lacks one or more data points that form the trajectory of the movable unit represented by the track data sample, the generator's processing of the real sparse track data sample can include imputing the missing data points in the track data sample. Thus, the track data sample generated by the generator processing the track data sample can, for example, be an imputed track data sample that approximates the real dense track data sample. Of course, the generator's processing of the real sparse track data sample can also include partially imputing the missing data points in the track data sample, filling all missing data points in the track data sample with random values, filling some missing data points in the track data sample with random values, or combinations thereof, etc., and this disclosure does not impose any limitations in this regard.

[0047] In one example, the second track sample dataset may include track data samples generated by a generator of a generative adversarial network processing all real sparse track data samples in the first track sample dataset. In another example, the second track sample dataset may include track data samples generated by a generator processing a subset of real sparse track data samples in the first track sample dataset.

[0048] In the example, it can be used Let θ represent the generator of a generative adversarial network, where θ g This represents the network parameters of the generator. Therefore, the input-output mapping relationship of the generator can be established, i.e., Where u is the input of the generator (e.g., a real sparse track data sample with missing data points), and v is the generated sample obtained by processing the input data sample u.

[0049] Sub-step S214: The second track sample dataset is fed into the generative adversarial network so that the discriminator can output a second discriminant value for the second track sample dataset. The second discriminant value is used to calculate the function value of the generator's objective function.

[0050] In the example, the second discriminant value output by the discriminator of the generative adversarial network for the second track sample dataset can be a score given by the discriminator regarding the realism of the non-real track data samples (e.g., those approximating real track data samples). In the example, from the discriminator's perspective, the higher the realism of the track data samples, the higher the score, and the score range can be, for example, the range [0,1]. It is understood that in the initial (multiple) iterations, the expected value (e.g., mean) of the discriminator's second discriminant value for the second track data sample dataset on that dataset may be low, attributed to the fact that the generator's capabilities at the same time have not yet improved sufficiently to generate non-real track data samples that increasingly approximate real track data samples. In the example, as described above, it can be used... Let represent the discriminator of the generative adversarial network. As a result, the second discriminant value output by the discriminator for any track data sample in the second track sample dataset can be represented as:

[0051] In the example, the second discriminant value can be used to calculate the function value of the generator objective function so that the function value of the generator objective function can be correlated with the performance of the generator (i.e., its ability to generate non-real track data samples that are sufficiently close to real track data samples).

[0052] In the example, the generator's objective function can be the optimization goal targeted when optimizing the generator, enabling the generator to update its network parameters according to the network parameters corresponding to the optimal value of its objective function, thereby achieving generator optimization. In the example, the optimal value of the generator's objective function can be the extreme value (e.g., maximum, minimum) across all iterations, or it can be an extreme value (e.g., maximum, minimum) across some iterations, etc. It is understood that the selection of the optimal value of the objective function may vary depending on the actual optimization process and / or optimization requirements, and this disclosure does not impose any restrictions on it.

[0053] Sub-step S215 optimizes the overall objective function of the generative adversarial network to obtain the network parameters of the generator when the overall objective function takes the optimal value. The overall objective function is constructed based on the discriminator objective function and the generator objective function.

[0054] As an example rather than a limitation, the overall objective function V of a generative adversarial network can be constructed as follows:

[0055]

[0056] For convenience, the network parameters at the subscripts of both the discriminator and the generator are omitted. Equation (1) above is the general form of the overall objective function of the generative adversarial network. Optimizing the generative adversarial network using this overall objective function can be divided into two optimization processes. The first step is to optimize the discriminator, that is, to first consider the following equation (2).

[0057]

[0058] During the optimization of the discriminator, the generator G is fixed. The first part of the right side of equation (2) above. Its function is to ensure the basic discriminant capability of the discriminator. For x represents the distribution of real data p. data The samples obtained from sampling The larger the value, the larger D(x) is, which means that the discriminator is more accurate in classifying real samples as real samples. Therefore, we have The second part on the right side of equation (2) above Its function is to ensure that the discriminator can distinguish non-real samples. For z is from a specific distribution p z The samples obtained from sampling are G(z), which are non-real samples generated by the generator that approximate the real samples. A larger value means a smaller value for D(G(z)) (equivalent to a lower score from the discriminator regarding the authenticity of non-real track data samples), meaning the discriminator is more able to correctly distinguish non-real samples. Therefore, there is also... In the example, stochastic gradient ascent can be applied to equation (2) to optimize the network parameters of the discriminator.

[0059] The second step is to optimize the generator, that is, to consider the following equation (3).

[0060]

[0061] During the optimization of the generator, the discriminator D remains constant. If G becomes more powerful, the discriminator D will make a misjudgment, and D(G(z)) will increase. It will be closer to zero, that is, the value of the whole equation (3) will be smaller, therefore we have In the example, stochastic gradient descent can be applied to equation (3) to optimize the network parameters of the generator.

[0062] Optimization in generative adversarial networks is essentially a minimax game. Theoretically, the final result of optimization will make p... G(z) =p data In other words, the distribution p of the non-real samples generated by the generator G(z) Compared with the actual data distribution p data At this point, the discriminator's score on the realism of the non-real samples generated by the generator tends to be similar to the probability of either side landing face up in a coin toss, i.e., 50%. In summary, through one or more rounds of iteration / optimization, the discriminator and the generator compete with each other during the optimization process and gradually tend towards Nash equilibrium.

[0063] It should be noted that in this disclosure, terms such as term optimization, term training, or term iteration are interchangeable, as will be apparent to those skilled in the art, and these terms have the general meanings known in the art.

[0064] In the example, in the overall objective function V of the example above, p data (x) can be the distribution of real track data samples contained in the first track sample dataset, p z (z) can be the distribution of non-real track data samples generated by the generator processing real sparse track data samples (e.g., some or all) in the first track sample dataset.

[0065] In existing technologies, the training of Generative Adversarial Networks (GANs) always begins with training a discriminator, the purpose of which is to give the discriminator an initial standard. Taking the application of GANs in the image domain as an example, this standard means that the discriminator should be able to identify a series of good images (e.g., real images), thus understanding what constitutes a good image. These images can come from various sources, such as databases, from which good images are extracted and input into the GAN discriminator. Then, after training the discriminator, which now has a standard, the generator is trained. In existing technologies, a specific distribution p... z The noise distribution can be Gaussian, uniform, or similar. Therefore, simply generating a set of noise for the generator, such as noise conforming to a Gaussian distribution, will allow the generator to produce an image. This image (and the images generated by the generator in the initial iterations) is undesirable; such images cannot be classified as real images with a high degree of realism by the discriminator. After the generator produces the image, the discriminator determines whether it is a real image from the database or a generated image from the generator. If it is classified as a real image, it is assigned a higher score, for example, a maximum of 1. If it is classified as a generated image (i.e., not a real image), it is assigned a lower score, for example, a minimum of 0. Subsequently, the overall objective function of the GAN is optimized based on the discriminator's output. The network parameters of the discriminator and generator are obtained when the overall objective function reaches its optimal value, and the discriminator and / or generator are updated accordingly.

[0066] In the example, additionally, the overall objective function of the generative adversarial network can be optimized to obtain the network parameters of the discriminator when the objective function is optimal.

[0067] In some embodiments, not p z (z) Instead of using a noise distribution such as a Gaussian or uniform distribution, p z(z) is consistent with the distribution of the track data samples generated by the generator after processing real sparse track data samples with missing data points. That is, z is taken from the group of real sparse track data samples in the real track data samples contained in the first track sample dataset (e.g., a subset of the first track sample dataset), rather than from the group of random data samples that conform to a certain noise distribution. Thus, for the multi-round adversarial training process from the first iteration to the last iteration, the generator does not start with random data samples to learn the distribution of real track data samples, but derives generated samples for the discriminator to judge the degree of authenticity based on real sparse track data samples with missing data points. This effectively reduces the number of iterations required for the objective function to converge (e.g., obtain the optimal value), and thus saves the amount of data required for the adversarial training / optimization of the generative adversarial network (e.g., the number of track data samples in the first and / or second track sample datasets).

[0068] Sub-step S216: Update the generator based on the obtained network parameters of the generator.

[0069] In the example, additionally, while updating the generator, the discriminator can be updated based on the network parameters of the discriminator obtained when the overall objective function of the generative adversarial network is optimal, so that in the next iteration (if any), the network parameters of both the discriminator and the generator are updated.

[0070] As can be understood, as described above regarding sub-step S215, the adversarial training / optimization of the generative adversarial network begins with the training of the discriminator, and each iteration is performed in an alternating manner between the discriminator and the generator. That is, in the process of optimizing the generator according to equation (3) above, the object to be minimized is V(D,G) after the discriminator has been optimized before, i.e., V(D,G) corresponding to the maximum or maximum value of equation (2) above. In the process of optimizing the generator, the network parameters of the discriminator corresponding to the maximum or maximum value of equation (2) are fixed. In other words, when optimizing / training the generator, the discriminator is considered fixed; conversely, for example, when optimizing / training the discriminator in the next iteration, the updated generator network parameters obtained in the previous iteration remain unchanged.

[0071] Step S220 may include determining whether the iteration stopping condition has been met when all steps of the current iteration have been executed.

[0072] In the example, the iteration stopping condition may include the overall objective function of the generative adversarial network reaching a preset threshold (this preset threshold may be different from the theoretical maximum or minimum value to prevent slow convergence, etc.), the mathematical expectation of the discriminator's discrimination value for the included non-real track data samples on the second track sample dataset being 0.5 (50%) (assuming the discrimination value ranges from [0,1]), or the number of iterations reaching a preset threshold, etc.

[0073] Step S230 may include, in response to the determination that the iteration stopping condition is met, using the updated generator as the final generator, so that the final generator can generate complete dense track data samples with missing data points based on the sparse track data samples to be completed.

[0074] Step S240 may include proceeding to the next iteration based on the updated generator in response to the determination that the iteration stopping condition has not been met.

[0075] In the example, before proceeding to the next iteration, the network parameters of the discriminator can be obtained when the overall objective function of the generative adversarial network is optimal, so that the next iteration can also be based on the updated discriminator.

[0076] Therefore, it should be understood that in each iteration, the generator is specifically designed to generate a flight trajectory with missing data points based on sparse track data samples. Compared to the generators in conventional generative adversarial networks, which are trained to generate samples based on random data samples to approximate the distribution of real data samples, this effectively avoids the undesirable situation where the generator generates a track that approximates the real track in appearance but may actually be fabricated (for example, for some sections where track data points are impossible to exist, but the data points supplemented in the generated samples fall into such sections). This can effectively ensure the accuracy and completeness of track recovery.

[0077] In the example, the first track sample dataset in each iteration can remain unchanged; that is, the network parameters of the discriminator are not updated with each iteration, which is especially true if the representativeness and coverage of the real track data samples contained in the first track data sample are good. Of course, the first track data sample in each iteration can also change or contain different real track data samples, and this disclosure does not impose any restrictions on this.

[0078] In the example, the real sparse track data samples that constitute the second track sample dataset in each iteration can be a distinct subset of all real sparse track data samples in the first track sample dataset. Of course, these distinct subsets can have overlap with each other, especially when the number of real sparse track data samples is not too redundant, and this disclosure does not impose any restrictions on this.

[0079] As an example and not a limitation, the discriminator objective function may include only the first part of the right-hand side of equation (2) above, or both the first and second parts of the right-hand side of equation (2) above. It is understood that the example overall objective function described with respect to substep S215 is essentially minimizing the JS divergence between the distributions of the first track sample dataset and the second track sample dataset. There are many other ways to measure the difference (e.g., distance) between the distributions of two datasets, and JS divergence is just one of them. Therefore, different objective functions can be obtained by using different metrics. Metrics may include, for example, Inception Score, Mode Score, Kernel MMD (Maximum Mean Discrepancy), Wasserstein distance, Fréchet Inception Distance (FID), etc., and this disclosure does not impose any restrictions on these or the discriminator objective function constructed based on them. Similarly, this disclosure does not impose any restrictions on the generator objective function.

[0080] According to embodiments of this disclosure, firstly, a first track sample dataset, comprising both real sparse track data samples and real dense track data samples, is acquired and fed into a generative adversarial network (GAN). This yields a score from the network's discriminator regarding the realism of the real track data samples. Then, the network's generator processes the real sparse track data samples to generate a second track sample dataset that approximates the real dense track data samples. This dataset is also fed into the GAN to obtain a score from the discriminator regarding the realism of the non-real track data samples in the second track sample dataset. Finally, the overall objective function of the network is optimized to improve the discriminator's ability (i.e., to better distinguish between real and dense track data). The generator's ability (i.e., generating non-real track data that increasingly approximates real track data) and the non-real track data compete against each other during the optimization process and gradually tend towards Nash equilibrium. Finally, after all rounds of iteration are completed, the final updated generator is obtained, which enables the generator to generate complete dense track data samples based on sparse track data samples with missing data points. Thus, with the help of the method disclosed in this paper, incomplete track data can be restored into dense track data through track restoration technology based on generative adversarial networks. On the one hand, it significantly reduces the amount of data required compared with the baseline method, and on the other hand, it improves the accuracy and completeness of track restoration.

[0081] In some embodiments, the generator may include an encoder and a decoder. The encoder may be used to receive and encode real sparse track data samples with missing data points to obtain an encoded vector corresponding to the real sparse track data samples with missing data points. The decoder may be used to receive and decode a concatenated vector obtained by concatenating the encoded vector and a noise vector to obtain a complete dense track data sample with complete data points corresponding to the real sparse track data samples with missing data points.

[0082] In the example, the noise distribution can be Gaussian, uniform, etc., and this disclosure does not impose any restrictions on it.

[0083] In the example, concatenating the encoded vector and the noise vector may include adding corresponding elements of the encoded vector and the noise vector (i.e., equivalent to adding additive noise to the encoded vector) or multiplying them (i.e., equivalent to adding multiplicative noise to the encoded vector), or concatenating the beginning and end of the encoded vector and the noise vector, etc., and this disclosure does not impose any restrictions on this.

[0084] In some embodiments, the encoder may be a time-aware long short-term memory network (T-LSTM).

[0085] Given that the intervals between adjacent time points within a single real track data point (e.g., a sample of real sparse track data with missing data points) are not fixed, an LSTM suitable for processing data samples with irregular time intervals is needed to encode such track data. T-LSTM has the forget gate, input gate, and output gate of a standard LSTM, but the memory cells are adjusted. Specifically, T-LSTM uses a time decay function to convert elapsed time into weights. That is, T-LSTM converts the elapsed time Δt between two consecutive records into weights to reduce the influence of short-term memory. T-LSTM adjusts short-term memory based on the time span between records from time step t to t-1. If the difference between time t and t-1 is large, it indicates that the mobile unit (e.g., an enemy target) has not recorded new information for a long period; therefore, reliance on short-term memory should not play a significant role in the prediction of the current output. This processing of short-term memory can be expressed mathematically as follows: in, Representing short-term memory, g(Δ t ) represents about Δ t discount function, This represents the processed short-term memory. Therefore, as a supplement to the inherent adversarial training / optimization of generative adversarial networks (GANs), configuring the generator with a structure consisting of an encoder and a decoder further ensures the accuracy and completeness of the generated samples, which is beneficial for the rapid convergence of GANs in multiple iterations, thus saving data requirements and computational resources.

[0086] In some embodiments, the discriminator objective function may be E y～p(y) log(D(y)), where y is the input of the discriminator, p(y) is the distribution of the real track data samples, y~p(y) means that y is taken from the distribution p(y), D(·) represents the discriminant value output by the discriminator after receiving the input, which represents the degree of authenticity of the input, and E y～p(y) This represents the mathematical expectation over all track data samples in the first track sample dataset. Considering that the generator does not learn the distribution of real track data samples from random data samples, but rather derives generated samples for the discriminator to judge the degree of authenticity based on real sparse track data samples with missing data points, focusing the discriminator's objective function on ensuring the discriminator's basic discrimination ability can improve the efficiency of each iteration and shorten the time for the overall objective function to converge (e.g., to obtain the optimal or suboptimal value), thereby saving computational resources and meeting the time requirement for rapid data point completion of sparse tracks in combat scenarios.

[0087] In some embodiments, the generator objective function may be E z～γ，x～q(x) log(1-D(G(z,x))), where z and x are the inputs to the generator, z~γ means z is taken from the noise distribution γ, x~q(x) means x is taken from the distribution q(x) of real sparse track data samples with missing data points, G(·) means the output generated by the generator after receiving and processing the input, D(·) means the discriminator outputs the discrimination value representing the authenticity of the input after receiving the input, and E z～γ，x～q(x) It represents the mathematical expectation over all track data samples in the second track sample dataset.

[0088] In some embodiments, the discriminator is pre-trained, and the track training dataset used for pre-training includes: real track training data samples, distorted track training data samples, and labels corresponding to the real track training data samples and distorted track training data samples, respectively. This ensures the discriminator's ability to distinguish real track data samples from the outset, preventing adverse effects on generator training / optimization caused by fluctuations in the discriminator's discriminative ability (e.g., the generator continuously derives generated samples that do not closely approximate real dense tracks because the discriminator's scores indicate that such generated samples are sufficiently realistic).

[0089] In the example, different real track training data samples used to pre-train the discriminator can be associated with the same real label (e.g., 1), and different false track training data samples used to pre-train the discriminator can be associated with the same false label (e.g., 0).

[0090] In the example, the distorted track training data sample may be obtained by partially modifying a real track data sample. In another example, the distorted track training data sample may be a data sample with randomized data point values, and so on. This disclosure does not impose any limitations in this regard.

[0091] In some embodiments, before feeding the first track sample dataset to the generative adversarial network, the track data samples in the first track sample dataset may be scaled according to the following formula:

[0092]

[0093] Where p represents a track data sample in the first track sample dataset, p new This represents the track data sample obtained after scaling the track data sample p. max Let p represent the largest track data sample in the first track sample dataset, and p min This represents the smallest track data sample in the first track sample dataset. This eliminates the differences between different tracks and the impact of the original track feature dimensions on model performance.

[0094] Figure 3 This is a schematic diagram illustrating the architecture of a generative adversarial network 300 according to an exemplary embodiment.

[0095] As shown in the figure, the Generative Adversarial Network (GAN) 300 includes a generator 310 and a discriminator 320, wherein the generator 310 includes an encoder 311 and a decoder 312. Optionally, the encoder 311 includes a T-LSTM 314. The input 301 of the GAN includes both real sparse track data samples and real dense track data samples. As described above, the real sparse track data samples with missing data points (e.g., partially or completely) are encoded into an encoded vector 302 after being encoded by the encoder 311. The encoded vector 302 is concatenated with a noise vector 303 taken from a noise distribution 313 to obtain a concatenated vector 304. The concatenated vector 304 is then decoded by the decoder 312 into a decoded vector 305, which is a complete dense track data sample (i.e., a generated sample) corresponding to the real sparse track data samples with missing data points. After the decoded vector 305 is discriminated by the discriminator 320, an adversarial loss 306 is obtained (the adversarial loss is based on the second discriminant value as described above). The adversarial loss 306 is then backpropagated to the generator 310, thereby updating the network parameters of the generator 310.

[0096] It is important to note that Figure 3 The components shown are illustrative and are illustrated for the purpose of facilitating understanding of this disclosure only, and are not intended to limit this disclosure in any way.

[0097] Figure 4 This is a visualization comparing the results of track recovery using the method according to the exemplary embodiment and the baseline method.

[0098] As shown in the figure, the baseline methods include linear prediction, Long Short-Term Memory (LSTM) networks, bidirectional LSTM, and Graph Aggregation Inference Network (GAIN). It can be seen that, compared to ground truth (i.e., real track data samples), the dense track data obtained by iteratively training / optimizing the GAN generator using the method disclosed herein to complete the track data is visually more consistent with real track data. Furthermore, it can be expected that the discriminator's judgment on whether such dense track data is real track data will tend to be random (50%).

[0099] Figure 5 This is a schematic block diagram illustrating a track data processing apparatus 500 according to an exemplary embodiment. According to embodiments of this disclosure, apparatus 500 can be used to recover sparse track data samples with missing data points using a generative adversarial network (GAN), which includes a discriminator and a generator. Figure 5As shown, the apparatus 500 includes: a single-round execution unit 510, configured to perform the following steps for each iteration in one or more rounds of iteration: acquiring a first track sample dataset, the first track sample dataset containing real track data samples, the real track data samples including both real sparse track data samples with missing data points and real dense track data samples with complete data points; feeding the first track sample dataset to a generative adversarial network (GAN) for a discriminator to output a first discriminant value for the first track sample dataset, the first discriminant value being used to calculate the function value of the discriminator's objective function; acquiring a second track sample dataset, the second track sample dataset being generated by a generator processing real sparse track data samples with missing data points; feeding the second track sample dataset to the GAN for a discriminator to output a second discriminant value for the second track sample dataset, the second discriminant value being used to calculate the function value of the discriminator's objective function; acquiring a second track sample dataset, the second track sample dataset being generated by a generator processing real sparse track data samples with missing data points; feeding the second track sample dataset to the GAN for a discriminator to output a second discriminant value for the second track sample dataset, the second discriminant value being used to calculate the function value of the discriminator's objective function; and acquiring a second track sample dataset, the second track sample dataset being generated by a generator processing real sparse track data samples with missing data points. The algorithm calculates the function value of the generator's objective function; optimizes the overall objective function of the generative adversarial network to obtain the generator's network parameters when the overall objective function is optimal, wherein the overall objective function is constructed based on the discriminator's objective function and the generator's objective function; and updates the generator based on the obtained generator's network parameters; an iteration determination unit 520 is used to determine whether the iteration stopping condition is met when all steps of the current iteration are completed; a final determination unit 530 is used to use the updated generator as the final generator in response to the determination that the iteration stopping condition is met, so that the final generator can generate complete dense track data samples based on the sparse track data samples with missing data points; and an iteration triggering unit 540 is used to perform the next iteration based on the updated generator in response to the determination that the iteration stopping condition is not met.

[0100] It should be understood that Figure 5 The various modules / units of the device 500 shown can be used with reference to Figure 2 The steps in method 200 described correspond to each other. Therefore, the operations, features, and advantages described above for method 200 also apply to apparatus 500 and its included modules. For the sake of brevity, some operations, features, and advantages will not be repeated here.

[0101] While specific functions have been discussed above with reference to specific modules, it should be noted that the functions of the modules discussed herein can be divided into multiple modules, and / or at least some functions of multiple modules can be combined into a single module. The specific actions performed by the modules discussed herein include the specific module itself performing the action, or alternatively, the specific module calling or otherwise accessing another component or module that performs the action (or performs the action in conjunction with the specific module). Therefore, a specific module performing an action can include the specific module performing the action itself and / or another module that performs the action, called or otherwise accessed by the specific module.

[0102] It should also be understood that this article can describe various technologies in the general context of software and hardware components or program modules. The above regarding... Figure 5 The various modules and / or units described may be implemented in hardware or in hardware in combination with software and / or firmware. For example, these modules and / or units may be implemented as computer program code / instructions configured to execute in one or more processors and stored in a computer-readable storage medium. Alternatively, these modules and / or units may be implemented as hardware logic / circuit. One or more of the modules and / or units may be implemented together in a System on Chip (SoC). The SoC may include an integrated circuit chip (which includes a processor (e.g., a Central Processing Unit (CPU), microcontroller, microprocessor, digital signal processor (DSP), etc.), memory, one or more communication interfaces, and / or one or more components of other circuitry) and may optionally execute received program code and / or include embedded firmware to perform functions.

[0103] According to one aspect of this disclosure, a computer device is provided, including a memory, a processor, and a computer program stored in the memory. The processor is configured to execute the computer program to implement the steps of any of the method embodiments described above.

[0104] According to one aspect of this disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of any of the method embodiments described above.

[0105] According to one aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps of any of the method embodiments described above.

[0106] In the following text, combined with Figure 6 Illustrative examples describing such computer devices, non-transitory computer-readable storage media, and computer program products.

[0107] Figure 6 An example configuration of a computer device 600 that can be used to implement the methods described herein is shown. For example, Figure 1 The server 120 and / or client device 110 shown may include an architecture similar to computer device 600. The aforementioned apparatus 500 may also be implemented wholly or at least partially by computer device 600 or similar devices or systems.

[0108] Computer device 600 can be a variety of different types of devices. Examples of computer device 600 include, but are not limited to: desktop computers, server computers, laptop or netbook computers, mobile devices (e.g., tablet computers, cellular or other wireless phones (e.g., smartphones), notebook computers, mobile stations), wearable devices (e.g., glasses, watches), entertainment devices (e.g., entertainment appliances, set-top boxes communicatively coupled to a display device, game consoles), televisions or other display devices, automotive computers, and so on.

[0109] Computer device 600 may include at least one processor 602, memory 604, multiple communication interfaces 606, display device 608, other input / output (I / O) devices 610, and one or more mass storage devices 612 capable of communicating with each other, such as via system bus 614 or other suitable connections.

[0110] Processor 602 may be a single processing unit or multiple processing units, and all processing units may include single or multiple computing units or multiple cores. Processor 602 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, and / or any device that manipulates signals based on operating instructions. Among other capabilities, processor 602 may be configured to acquire and execute computer-readable instructions stored in memory 604, mass storage device 612, or other computer-readable media, such as program code of operating system 616, program code of application program 618, program code of other program 620, etc.

[0111] Memory 604 and mass storage device 612 are examples of computer-readable storage media for storing instructions that are executed by processor 602 to perform the various functions described above. For example, memory 604 may generally include both volatile and non-volatile memory (e.g., RAM, ROM, etc.). Furthermore, mass storage device 612 may generally include hard disk drives, solid-state drives, removable media, including external and removable drives, memory cards, flash memory, floppy disks, optical disks (e.g., CDs, DVDs), storage arrays, network-attached storage, storage area networks, etc. Both memory 604 and mass storage device 612 may be collectively referred to herein as memory or computer-readable storage media, and may be non-transitory media capable of storing computer-readable, processor-executable program instructions as computer program code, which may be executed by processor 602 as a specific machine configured to perform the operations and functions described in the examples herein. Multiple programs may be stored on mass storage device 612. These programs include an operating system 616, one or more application programs 618, other programs 620, and program data 622, and they can be loaded into memory 604 for execution.

[0112] Although Figure 6 The blocks 616, 618, 620, and 622 are illustrated as being stored in memory 604 of computer device 600, but blocks 616, 618, 620, and 622, or portions thereof, may be implemented using any form of computer-readable medium accessible by computer device 600. As used herein, “computer-readable medium” includes at least two types of computer-readable media: computer-readable storage media and communication media.

[0113] Computer-readable storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, DVD, or other optical storage devices, magnetic cassettes, magnetic tapes, disk storage devices or other magnetic storage devices, or any other non-transmission medium that can be used to store information for access by computer devices. In contrast, communication media can embody computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms. Computer-readable storage media as defined herein do not include communication media.

[0114] One or more communication interfaces 606 are used for exchanging data with other devices, such as via a network, direct connection, etc. Such communication interfaces can be one or more of the following: any type of network interface (e.g., a network interface card (NIC)), wired or wireless (such as IEEE 802.11 Wireless LAN (WLAN)) wireless interface, Wi-MAX interface, Ethernet interface, Universal Serial Bus (USB) interface, cellular network interface, Bluetooth. TM Interfaces, near field communication (NFC) interfaces, etc. Communication interface 606 can facilitate communication across various network and protocol types, including wired networks (e.g., LAN, cable, etc.) and wireless networks (e.g., WLAN, cellular, satellite, etc.), the Internet, etc. Communication interface 606 can also provide communication with external storage devices (not shown) such as storage arrays, network-attached storage, storage area networks, etc.

[0115] In some examples, a display device 608, such as a monitor, may be included for displaying information and images to the user. Other I / O devices 610 may be devices that receive various inputs from the user and provide various outputs to the user, and may include touch input devices, gesture input devices, cameras, keyboards, remote controls, mice, printers, audio input / output devices, and so on.

[0116] The technologies described herein can be supported by these various configurations of computer device 600, and are not limited to specific examples of the technologies described herein. For example, the functionality can also be implemented wholly or partially on a “cloud” using a distributed system. A cloud includes and / or represents a platform for resources. The platform abstracts the underlying functionality of the cloud’s hardware (e.g., servers) and software resources. Resources may include applications and / or data that can be used when performing computational processing on a server remote from computer device 600. Resources may also include services provided via the Internet and / or via subscriber networks such as cellular or Wi-Fi networks. The platform can abstract resources and functionality to connect computer device 600 to other computer devices. Therefore, the implementation of the functionality described herein can be distributed throughout the cloud. For example, the functionality may be implemented partly on computer device 600 and partly through a platform that abstracts the functionality of the cloud.

[0117] Although this disclosure has been described and illustrated in detail in the accompanying drawings and the foregoing description, such description and illustration should be considered illustrative and suggestive, not restrictive; this disclosure is not limited to the disclosed embodiments. By studying the drawings, the disclosure, and the appended claims, those skilled in the art will be able to understand and implement variations of the disclosed embodiments in practice with respect to the claimed subject matter. In the claims, the word "comprising" does not exclude other elements or steps not listed, the indefinite article "a" or "an" does not exclude a plurality, the term "a plurality" means two or more, and the term "based on" should be interpreted as "at least partially based on". The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be beneficial.

Claims

1. A track data processing method for recovering sparse track data samples with missing data points using a generative adversarial network (GAN), wherein the GAN includes a discriminator and a generator, and the method includes: For each iteration in one or more iterations, perform the following steps: Obtain a first track sample dataset, which contains real track data samples, including both real sparse track data samples with missing data points and real dense track data samples with complete data points. The first track sample dataset is fed into the generative adversarial network so that the discriminator can output a first discrimination value for the first track sample dataset. The first discrimination value is used to calculate the function value of the discriminator's objective function. Obtain a second track sample dataset, which is generated by the generator by processing real sparse track data samples with missing data points; The second track sample dataset is fed into the generative adversarial network so that the discriminator can output a second discriminant value for the second track sample dataset. The second discriminant value is used to calculate the function value of the generator's objective function. The overall objective function of the generative adversarial network is optimized to obtain the network parameters of the generator when the overall objective function reaches its optimal value, wherein the overall objective function is constructed based on the discriminator objective function and the generator objective function; and The generator is updated based on the obtained network parameters of the generator; Once all steps of this iteration have been executed, determine whether the iteration stopping condition has been met; and In response to the determination that the iteration stopping condition has been met, the updated generator is used as the final generator, which generates complete dense track data samples based on the sparse track data samples with missing data points; or In response to the determination that the iteration stopping condition has not been met, the next iteration is performed based on the updated generator.

2. The method according to claim 1, wherein, The generator includes an encoder and a decoder. The encoder is used to receive and encode real sparse track data samples with missing data points to obtain an encoded vector corresponding to the real sparse track data samples with missing data points. The decoder is used to receive and decode a concatenated vector obtained by concatenating the encoded vector and a noise vector to obtain a complete dense track data sample with complete data points corresponding to the real sparse track data samples with missing data points.

3. The method according to claim 2, wherein, The encoder is a time-aware long short-term memory network (T-LSTM).

4. The method according to any one of claims 1-3, wherein, The objective function of the discriminator is: E y～p(y) log(D(y)) Where y is the input of the discriminator, p(y) is the distribution of the real track data samples, y~p(y) means that y is taken from the distribution p(y), D(·) represents the discrimination value output by the discriminator after receiving the input, which characterizes the authenticity of the input, and E y～p(y) It represents the mathematical expectation over all track data samples in the first track sample dataset.

5. The method according to any one of claims 1-3, wherein, The objective function of the generator is: E z～γ，x～q(x) log(1-D(G(z,x))), Where z and x are both inputs to the generator, z~γ represents z taken from the noise distribution γ, x~q(x) represents x taken from the distribution q(x) of the real sparse track data samples with missing data points, G(·) represents the output generated by the generator after receiving and processing the input, D(·) represents the discrimination value output by the discriminator after receiving the input, which characterizes the authenticity of the input, and E z～γ，x～q(x) It represents the mathematical expectation over all track data samples in the second track sample dataset.

6. The method according to any one of claims 1-3, wherein, The discriminator is pre-trained, and the trajectory training dataset used for pre-training includes: real trajectory training data samples, distorted trajectory training data samples, and labels corresponding to the real trajectory training data samples and the distorted trajectory training data samples, respectively.

7. The method according to any one of claims 1-3, further comprising, prior to the step of feeding the first track sample dataset to the generative adversarial network: The trajectory data samples in the first trajectory sample dataset are scaled using the following formula: in, p represents a certain track data sample in the first track sample dataset, p new This represents the track data sample obtained after scaling the track data sample p. max Let p represent the largest track data sample in the first track sample dataset, and p min This represents the smallest track data sample in the first track sample dataset.

8. A track data processing apparatus for recovering sparse track data samples with missing data points using a generative adversarial network, wherein the generative adversarial network includes a discriminator and a generator, and the apparatus comprises: A single-round execution unit is used to execute the following steps for each iteration in one or more rounds of iteration: Obtain a first track sample dataset, which contains real track data samples, including both real sparse track data samples with missing data points and real dense track data samples with complete data points. The first track sample dataset is fed into the generative adversarial network so that the discriminator can output a first discrimination value for the first track sample dataset. The first discrimination value is used to calculate the function value of the discriminator's objective function. Obtain a second track sample dataset, which is generated by the generator by processing real sparse track data samples with missing data points; The second track sample dataset is fed into the generative adversarial network so that the discriminator can output a second discriminant value for the second track sample dataset. The second discriminant value is used to calculate the function value of the generator's objective function. The overall objective function of the generative adversarial network is optimized to obtain the network parameters of the generator when the overall objective function reaches its optimal value, wherein the overall objective function is constructed based on the discriminator objective function and the generator objective function; and The generator is updated based on the obtained network parameters of the generator; The iterative determination unit is used to determine whether the iteration stopping condition has been met when all steps of the current iteration have been executed. The final determination unit is configured to, in response to determining that the iteration stopping condition is met, use the updated generator as the final generator, so that the final generator can generate complete dense track data samples based on the sparse track data samples with missing data points; and An iteration triggering unit is configured to initiate the next iteration based on the updated generator in response to determining that the iteration stop condition has not been met.

9. An electronic device, comprising: At least one processor; as well as A memory that is communicatively connected to the at least one processor; in The memory stores a computer program that, when executed by the at least one processor, implements the method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing a computer program, wherein, The computer program, when executed by a processor, implements the method according to any one of claims 1-7.

Images