A Trajectory Denoising Method and Apparatus Based on Generative Adversarial Networks

Abstract

This invention relates to a trajectory denoising method and apparatus based on generative adversarial networks (GANs), comprising: acquiring high-precision trajectory data and noise data respectively, and discretizing the trajectory to obtain a trajectory dataset; the noise data corresponding to the high-precision trajectory data; iteratively training a discriminator and a generator using the discretized high-precision trajectory data and noise data, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS); merging the realism discriminator (DR) and the similarity discriminator (DS) and constructing a target function for the generative adversarial network based on the traditional CGAN objective function; and performing trajectory denoising on the trajectory data to be processed using the trained generator and the newly constructed target function of the generative adversarial network. This scheme introduces a generative adversarial network (GAN), which performs better in generative models, and reconstructs a denoising generative model based on the trajectory.

Description

Technical Field

[0001] This invention relates to the field of map production technology, and specifically to a trajectory denoising method and apparatus based on generative adversarial networks. Background Technology

[0002] In the field of map production, a large amount of GPS trajectory data is used. However, GPS trajectories contain noise, which makes them difficult to use. In addition to removing abnormal trajectories, it is also necessary to denoise the trajectories to be used in order to improve the efficiency of trajectory use and reduce the impact of noise on the trajectories.

[0003] Methods for handling this problem include:

[0004] 1. Based on prior knowledge summarized by humans: including observing the changing trends of trajectory direction and position, summarizing the anomaly types, and processing them. This method is based on human knowledge and is difficult to cover all scenarios.

[0005] 2. Based on Gaussian distribution modeling, this method has strong presuppositions, predefines the distribution of noise, and is not applicable to all situations;

[0006] 3. Encoding methods based on autoencoder (AE) architecture and long short-term memory (LSTM) model learn the trajectory by itself to obtain the implicit representation of the trajectory, and then reconstruct the trajectory. However, this method often generates relatively average trajectories and is difficult to learn the true characteristics of the trajectory. Summary of the Invention

[0007] This invention addresses the technical problems existing in the prior art by providing a trajectory denoising method and apparatus based on generative adversarial networks, which solves trajectory denoising in a more generalized way.

[0008] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0009] In a first aspect, the present invention provides a trajectory denoising method based on generative adversarial networks, comprising:

[0010] High-precision trajectory data and noise data are acquired separately, and the trajectory is discretized to obtain a trajectory dataset; the noise data corresponds to the high-precision trajectory data.

[0011] The discriminator and generator are iteratively trained using discretized high-precision trajectory data and noisy data, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS).

[0012] By merging the authenticity discriminator DR and the similarity discriminator DS and constructing the objective function of the generative adversarial network based on the traditional CGAN objective function;

[0013] Trajectory denoising is performed on the trajectory data to be processed using a trained generator and the objective function of a newly constructed generative adversarial network.

[0014] Furthermore, the noise data is obtained by adding random noise to high-precision trajectory data, or the high-precision trajectory data and noise data are respectively taken from the output data of high-precision and low-precision GPS acquisition devices installed on the same measuring vehicle.

[0015] Furthermore, the high-precision trajectory data and noisy data are discretized and then regularized.

[0016] Furthermore, the high-precision trajectory data and noise data are discretized, including dividing the data into grids of 0.5m and converting latitude and longitude coordinates into integer plane coordinates.

[0017] Furthermore, after obtaining the trajectory dataset, the process also includes extracting a batch of high-precision trajectories and noisy trajectories from the trajectory dataset, and performing data augmentation on the trajectories of this batch of data.

[0018] Furthermore, the discriminator and generator are iteratively trained using discretized high-precision trajectory data and noisy data, including:

[0019] First, fix the generator parameters, then train and adjust the parameters of the realism discriminator DR and the similarity discriminator DS;

[0020] Then, fix the parameters of the authenticity discriminator DR and the similarity discriminator DS, and train and adjust the parameters of the generator;

[0021] During generator training, a new constraint is introduced: the two-dimensional distance between the newly generated trajectory and the original high-precision trajectory. The objective function for this two-dimensional distance is:

[0022]

[0023] in, Expressing expectations, It is the objective function. It is a two-dimensional distance.

[0024] Furthermore, the aforementioned merging of the authenticity discriminator (DR) and the similarity discriminator (DS) and constructing the objective function of the generative adversarial network based on the traditional CGAN objective function includes:

[0025] Combine the authenticity discriminant DR and the similarity discriminant DS.

[0026]

[0027] The objective function of generative adversarial networks is:

[0028]

[0029] in, This represents the training objective of the discriminator. This represents the training objective of the generator. This represents the result of the generator's parameters. Used to adjust the weight ratio of 2D distance and CGAN target.

[0030] Secondly, the present invention provides a trajectory denoising device based on generative adversarial networks, comprising:

[0031] The data acquisition module acquires high-precision trajectory data and noise data respectively, and performs trajectory discretization to obtain a trajectory dataset; the noise data corresponds to the high-precision trajectory data.

[0032] The iterative training module uses discretized high-precision trajectory data and noise data to iteratively train the discriminator and the generator, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS).

[0033] The objective function construction module merges the realism discriminator (DR) and the similarity discriminator (DS) and constructs the objective function of the generative adversarial network based on the traditional CGAN objective function.

[0034] The denoising module uses the trained generator and the objective function of the newly constructed generative adversarial network to denoise the trajectory data to be processed.

[0035] Thirdly, the present invention provides an electronic device, comprising:

[0036] Memory, used to store computer software programs;

[0037] A processor is configured to read and execute the computer software program, thereby implementing the trajectory denoising method based on generative adversarial networks as described in the first aspect of the present invention.

[0038] Fourthly, the present invention provides a non-transitory computer-readable storage medium storing a computer software program for implementing the trajectory denoising method based on generative adversarial networks described in the first aspect of the present invention.

[0039] The beneficial effects of this invention are: This solution introduces Generative Adversarial Networks (GANs), which perform better in generative models, and reconstructs a denoising generative model based on trajectories. The key points are as follows:

[0040] 1. A generative adversarial network (GAN) architecture is used to build a denoising generation model. The advantage of this architecture over autoencoders (AE) is that GAN networks essentially learn the probability distribution of data, resulting in more natural generation effects.

[0041] 2. Based on the task characteristics of trajectory-to-trajectory denoising, a variant of the GAN network is sampled, namely Conditional Generative Adversarial Network (CGAN).

[0042] 3. Modify the discriminator (D) in the CGAN network by explicitly dividing it into two parts: a realism discriminator (DR) and a similarity discriminator (DS), which facilitates more precise control of the results during training;

[0043] 4. Based on the objective function of traditional CGAN, an L2 distance constraint is added to make the results of the generated model closer to the true value of the high-precision trajectory;

[0044] 5. The trajectory encoding method uses a new sequence encoding model: the self-attention model (transformer), which performs better than the original Long Short-Term Memory (LSTM) model;

[0045] In summary, the advantages of this solution compared to the previous solution are:

[0046] 1. Without making prior assumptions about the distribution of the original trajectory data, the Conditional Generative Adversarial Network (CGAN) can learn the distribution of normal trajectories from the data.

[0047] 2. The discriminator expression in the conditional generative adversarial network CGAN has been improved, which can more finely distinguish between realism and similarity, and facilitates fine control of the model;

[0048] 3. An L2 distance constraint was added to the traditional CGAN objective function, making the results of the generated model closer to the true value of the high-precision trajectory;

[0049] 4. A better self-attention model (transformer) was used for trajectory encoding.

[0050] The above improvements enhance the versatility of the solution, eliminating the need for special trajectory processing for different scenarios. Attached Figure Description

[0051] Figure 1 This is a schematic diagram illustrating how data is used when training a discriminator, as provided in an embodiment of the present invention.

[0052] Figure 2 This is a schematic diagram illustrating how data is used when training a generator, as provided in an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram of the generator structure provided in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the authenticity discriminator structure provided in an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of the similarity discriminator structure provided in an embodiment of the present invention;

[0056] Figure 6 This is a schematic diagram of the training process provided in an embodiment of the present invention;

[0057] Figure 7 A schematic diagram of an embodiment of the electronic device provided in this invention;

[0058] Figure 8 This is a schematic diagram of an embodiment of a computer-readable storage medium provided in this invention. Detailed Implementation

[0059] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0060] This invention provides a trajectory denoising method based on generative adversarial networks, comprising:

[0061] High-precision trajectory data and noise data are acquired separately, and the trajectory is discretized to obtain a trajectory dataset; the noise data corresponds to the high-precision trajectory data.

[0062] The discriminator and generator are iteratively trained using discretized high-precision trajectory data and noisy data, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS).

[0063] By merging the authenticity discriminator DR and the similarity discriminator DS and constructing the objective function of the generative adversarial network based on the traditional CGAN objective function;

[0064] Trajectory denoising is performed on the trajectory data to be processed using a trained generator and the objective function of a newly constructed generative adversarial network.

[0065] Specifically, the method includes the following steps:

[0066] 1. Data Preparation

[0067] This method requires two types of data: high-precision trajectory data and noise data. These two types of data are corresponding. Noise data can be obtained by adding random noise to high-precision trajectory data, or by installing high-precision and low-precision GPS acquisition devices on the same vehicle. The data acquired by the latter method can be used to train a denoising model with modified performance.

[0068] 2. Trajectory Discretization

[0069] 2.1 The prototype of trajectory coordinates is latitude and longitude, which is continuous spatial data. In order to reduce space and facilitate training, the trajectory coordinates need to be discretized. The grid can be divided into 0.5m grids to convert the latitude and longitude coordinates into integer planar coordinates.

[0070] 2.2 Regularization: Move all trajectory starting points to the origin, define a certain range, and place the trajectory coordinate range between (-1, 1).

[0071] The processed trajectories are then placed into the dataset.

[0072] 3. Batch augmentation of data

[0073] Extract a batch of high-precision trajectories and noisy trajectories from the trajectory dataset. Perform data augmentation on the trajectories in this batch, such as random rotation or reversal. The corresponding high-precision trajectories and noisy trajectories need to undergo the same augmentation operation. In this way, even with a small amount of data, you can still get results with good generalization.

[0074] 4. Training the discriminator

[0075] During training, the discriminator and generator are trained alternately.

[0076] First, train the discriminator.

[0077] The discriminator is divided into two parts: one is the authenticity discriminator (DR), which determines whether the trajectory is a real trajectory, and the other is the similarity discriminator (DS), which determines whether two trajectories are similar.

[0078] The training process first fixes the parameters of the generator (G), and only adjusts the parameters of the discriminators DR and DS. Data usage is as follows: Figure 1 The gray circles represent input data, the light-colored circles represent process and result data, and the squares represent the process. The simplified content is as follows:

[0079] x represents noisy data, y represents high-precision data, DR represents the authenticity discriminator, G represents the generator, Real represents the real trajectory, and Fake represents the fake trajectory.

[0080]

[0081]

[0082] DS stands for similarity discriminator, where True indicates that the similarity is true and False indicates that the similarity is false.

[0083]

[0084]

[0085] 5. Training Generator

[0086] After the discriminator is trained, the generator is trained. The generator's role is to generate noise-removed trajectories. During this process, the parameters of the discriminator DR and DS need to be fixed, while the parameters of the generator G are adjusted.

[0087] Data usage methods such as Figure 2 .

[0088] The simplified content is as follows:

[0089]

[0090]

[0091] A new constraint is introduced during generator training: the two-dimensional distance between the newly generated trajectory and the original high-precision trajectory. The objective function for this distance is:

[0092]

[0093] Expressing expectations, It is the objective function. It is a two-dimensional distance.

[0094] 6. Objective Function

[0095] The two modules DR and DS of the merge discriminator D

[0096]

[0097] The objective function based on traditional CGAN is:

[0098]

[0099] The overall objective function of combining CGAN and 2D distance is:

[0100]

[0101] This represents the training objective of the discriminator. This represents the training objective of the generator. This represents the result of the generator's parameters. Used to adjust the weight ratio of 2D distance and CGAN target.

[0102] 7. Implementation details of generator G, realism discriminator DR, and similarity discriminator DS.

[0103] 7.1 The generator G consists of a fully connected layer (Linear), a self-attention encoding layer (Transformer Encoder), a self-attention decoding layer (Transformer Decoder), and a fully connected layer (Linear). The fully connected layer is used for dimensionality transformation, and the self-attention layer is used for encoding and decoding. The input is a trajectory sequence, and the output is also a trajectory sequence. For details of the structure, see [link to documentation]. Figure 3 .

[0104] 7.2 The Authenticity Discriminator (DR) consists of a fully connected layer (Linear), a self-attention encoding layer (TransformerEncoder), another fully connected layer (Linear), and an activation function (sigmoid). The input is a trajectory sequence, and the output is whether the trajectory is authentic. See details of the structure. Figure 4 .

[0105] 7.3 The similarity discriminator DS consists of a fully connected layer (Linear), a concatenation layer (Concat), a self-attention encoding layer (Transformer Encoder), another fully connected layer (Linear), and an activation function (sigmoid). The input is two trajectory sequences, and the output is whether these two trajectories are similar. For details of the structure, see [link to documentation]. Figure 5 .

[0106] 8. Iterative process

[0107] like Figure 6 The data augmentation, training the discriminator, and training the generator need to be run iteratively to obtain the parameters of the generator G with sufficiently small error.

[0108] 9. Noise Reduction Process

[0109] like Figure 3 Once the generator is trained, inputting noisy data will output noisy data.

[0110] This invention also provides a trajectory denoising device based on generative adversarial networks, comprising:

[0111] The data acquisition module acquires high-precision trajectory data and noise data respectively, and performs trajectory discretization to obtain a trajectory dataset; the noise data corresponds to the high-precision trajectory data.

[0112] The iterative training module uses discretized high-precision trajectory data and noise data to iteratively train the discriminator and the generator, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS).

[0113] The objective function construction module merges the realism discriminator (DR) and the similarity discriminator (DS) and constructs the objective function of the generative adversarial network based on the traditional CGAN objective function.

[0114] The denoising module uses the trained generator and the objective function of the newly constructed generative adversarial network to denoise the trajectory data to be processed.

[0115] Please see Figure 7 , Figure 7 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 7 As shown, this embodiment of the invention provides an electronic device 500, including a memory 510, a processor 520, and a computer program 511 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 511, it performs the following steps:

[0116] High-precision trajectory data and noise data are acquired separately, and the trajectory is discretized to obtain a trajectory dataset; the noise data corresponds to the high-precision trajectory data.

[0117] The discriminator and generator are iteratively trained using discretized high-precision trajectory data and noisy data, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS).

[0118] By merging the authenticity discriminator DR and the similarity discriminator DS and constructing the objective function of the generative adversarial network based on the traditional CGAN objective function;

[0119] Trajectory denoising is performed on the trajectory data to be processed using a trained generator and the objective function of a newly constructed generative adversarial network.

[0120] Please see Figure 8 , Figure 8 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by an embodiment of the present invention. For example... Figure 8 As shown, this embodiment provides a computer-readable storage medium 600, on which a computer program 611 is stored. When the computer program 611 is executed by a processor, it performs the following steps:

[0121] High-precision trajectory data and noise data are acquired separately, and the trajectory is discretized to obtain a trajectory dataset; the noise data corresponds to the high-precision trajectory data.

[0122] The discriminator and generator are iteratively trained using discretized high-precision trajectory data and noisy data, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS).

[0123] By merging the authenticity discriminator DR and the similarity discriminator DS and constructing the objective function of the generative adversarial network based on the traditional CGAN objective function;

[0124] Trajectory denoising is performed on the trajectory data to be processed using a trained generator and the objective function of a newly constructed generative adversarial network.

[0125] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0126] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0127] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0128] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0129] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0130] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0131] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A trajectory denoising method based on generative adversarial networks, characterized in that, include: High-precision trajectory data and noisy data are acquired separately, and the trajectory is discretized to obtain a trajectory dataset; The noise data corresponds to the high-precision trajectory data; The discriminator and generator are iteratively trained using discretized high-precision trajectory data and noisy data, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS). By merging the authenticity discriminator DR and the similarity discriminator DS and constructing the objective function of the generative adversarial network based on the traditional CGAN objective function; Trajectory denoising is performed on the trajectory data to be processed using a trained generator and the objective function of a newly constructed generative adversarial network. The discriminator and generator are iteratively trained using discretized high-precision trajectory data and noisy data, including: First, fix the generator parameters, then train and adjust the parameters of the realism discriminator DR and the similarity discriminator DS; Then, fix the parameters of the authenticity discriminator DR and the similarity discriminator DS, and train and adjust the parameters of the generator; During generator training, a new constraint is introduced: the two-dimensional distance between the newly generated trajectory and the original high-precision trajectory. The objective function for this two-dimensional distance is: in, Expressing expectations, It is the objective function. It is a two-dimensional distance; The aforementioned method of merging the authenticity discriminator (DR) and the similarity discriminator (DS) and constructing the objective function of the generative adversarial network based on the traditional CGAN objective function includes: Combine the authenticity discriminant DR and the similarity discriminant DS. The objective function of generative adversarial networks is: in, This represents the training objective of the discriminator. This represents the training objective of the generator. This represents the result of the generator's parameters. Used to adjust the weight ratio of 2D distance and CGAN target; x represents noisy data, y represents high-precision data, D represents discriminator, DR represents realism discriminator, DS represents similarity discriminator, G represents generator, and G(x) represents the generator output given noisy data x as input. This represents the objective function based on the traditional CGAN.

2. The method according to claim 1, characterized in that, The noise data is obtained by adding random noise to high-precision trajectory data, or by taking high-precision trajectory data and noise data from the output data of high-precision and low-precision GPS acquisition devices installed on the same measuring vehicle, respectively.

3. The method according to claim 1, characterized in that, High-precision trajectory data and noisy data are discretized and then regularized.

4. The method according to claim 1, characterized in that, Trajectory discretization is performed on high-precision trajectory data and noisy data, including dividing the data into grids of 0.5m and converting latitude and longitude coordinates into integer plane coordinates.

5. The method according to claim 1, characterized in that, After obtaining the trajectory dataset, the process also includes extracting a batch of high-precision trajectories and noisy trajectories from the trajectory dataset, and performing data augmentation on the trajectories of this batch of data.

6. A trajectory denoising device based on generative adversarial networks, characterized in that, include: The data acquisition module acquires high-precision trajectory data and noise data respectively, and performs trajectory discretization to obtain a trajectory dataset; The noise data corresponds to the high-precision trajectory data; The iterative training module uses discretized high-precision trajectory data and noise data to iteratively train the discriminator and the generator, wherein the discriminator includes a realism discriminator (DR) and a similarity discriminator (DS). The objective function construction module merges the realism discriminator (DR) and the similarity discriminator (DS) and constructs the objective function of the generative adversarial network based on the traditional CGAN objective function. The denoising module utilizes the trained generator and the objective function of the newly constructed generative adversarial network to denoise the trajectory data to be processed. The iterative training module is specifically used for: First, fix the generator parameters, then train and adjust the parameters of the realism discriminator DR and the similarity discriminator DS; Then, fix the parameters of the authenticity discriminator DR and the similarity discriminator DS, and train and adjust the parameters of the generator; During generator training, a new constraint is introduced: the two-dimensional distance between the newly generated trajectory and the original high-precision trajectory. The objective function for this two-dimensional distance is: in, Expressing expectations, It is the objective function. It is a two-dimensional distance; The objective function construction module is specifically used for: Combine the authenticity discriminant DR and the similarity discriminant DS. The objective function of generative adversarial networks is: in, This represents the training objective of the discriminator. This represents the training objective of the generator. This represents the result of the generator's parameters. Used to adjust the weight ratio of 2D distance and CGAN target; x represents noisy data, y represents high-precision data, D represents discriminator, DR represents realism discriminator, DS represents similarity discriminator, G represents generator, and G(x) represents the generator output given noisy data x as input. This represents the objective function based on the traditional CGAN.

7. An electronic device, characterized in that, include: Memory, used to store computer software programs; A processor is configured to read and execute the computer software program, thereby implementing the trajectory denoising method based on generative adversarial networks as described in any one of claims 1-5.

8. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores a computer software program for implementing the trajectory denoising method based on generative adversarial networks as described in any one of claims 1-5.

Images