An intelligent simulation generation method for sea target track

Abstract

The application provides a marine target track intelligent simulation generation method, which comprises the following steps: collecting radar track data and corresponding automatic identification system (AIS) data of a target, and preprocessing the data, converting the data into target actual track images and real track images respectively, cutting the images into fixed-size paired sub-image blocks, constructing a training data set, constructing a test data set with the remaining part of the paired sub-image blocks, training a constructed radar track generative adversarial network model using the training data set, inputting sub-image blocks of the real track images into the trained radar track generative adversarial network model, generating realistic synthetic track image sub-image blocks, and splicing the synthetic track image sub-image blocks to obtain a complete synthetic target track image. The application is based on a supervised conditional generative adversarial network, and a generative network model is constructed, so that the easily obtained target real track is converted into a radar track with a real error distribution.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent processing of radar data, and particularly relates to a method for intelligently simulating a sea target track. BACKGROUND

[0002] At present, traditional radar data intelligent processing methods such as target recognition and tracking under complex sea conditions have been difficult to meet the actual engineering needs, and the development of deep learning technology provides a new solution for intelligent processing of radar data in complex sea conditions, and has become a research hotspot at home and abroad. However, deep learning is driven by data, and the acquisition of sea target detection data is restricted by factors such as cost and non-cooperative targets, and there is a serious lack of data, which greatly limits the fitting effect of deep neural networks.

[0003] Sea target track data simulation can solve the above problem of data shortage. Existing target track data simulation techniques can be mainly divided into model-based methods, real data-based methods and deep learning-based methods. The advantages of traditional model-based methods and real data-based methods are strong interpretability and mature technology, but the disadvantages are that key details such as isolated features and high-level hidden features are lost, it is difficult to generate target track data close to real error distribution, and the actual engineering application effect is difficult to guarantee.

[0004] The deep learning-based method is a core idea of data-driven, with strong feature learning and mapping ability. Among them, the generative adversarial network is a generative model based on deep learning, which can be used to learn the latent distribution of complex data and solve the problem of generating unknown distribution data. However, the traditional generative adversarial network cannot guide the generation of samples artificially. In order to solve this problem, the conditional generative adversarial network adds an additional condition to the input end to guide the sample generation process of the generator. Pix2pix is a classic conditional generative adversarial network, which takes the input picture as the network input, can realize the conversion of image to image, and has been widely used in image denoising, style transfer, remote sensing image conversion and medical image conversion tasks, and has shown good stability and strong generation performance.

[0005] In summary, the application learns from the idea of image-to-image conversion in the field of computer vision, and based on the supervised conditional generative adversarial network, a generative network model is constructed, which can convert the real track of the sea target into the noisy track measured by the radar, so as to convert the easily obtained real track of the target into the radar track with real error distribution. At present, there is no related public research work on sea target track simulation generation based on supervised conditional generative adversarial network at home and abroad. SUMMARY

[0006] The technical problem solved by the present application is to provide an intelligent simulation generation method for sea target track in view of the deficiencies of the prior art, which draws on the idea of image-to-image conversion in the field of computer vision, generates a generative adversarial network based on supervised conditions, and constructs a generative network model capable of converting the real track of a sea target into a noisy track measured by a radar, so as to convert the easily obtained real track of the target into a radar track with real error distribution. There is no related public research work on simulation generation of sea target track based on a generative adversarial network under supervised conditions at home and abroad.

[0007] To solve the above technical problems, the technical scheme adopted by the present application is: an intelligent simulation generation method for sea target track, characterized in that the method comprises:

[0008] S1, collecting radar track data of a target and corresponding automatic identification system (AIS) data of a ship, and pre-processing the data;

[0009] S2, converting the pre-processed radar track data and AIS data into a target actual track image and a real track image respectively, and pairing the target actual track image and the real track image into sub-image blocks after cropping, randomly selecting a part of the sub-image blocks to construct a training data set, and the remaining part of the sub-image blocks to construct a test data set;

[0010] S3, constructing a radar track generative adversarial network model comprising a generator with two different network structures and a discriminator;

[0011] S4, creating a loss function of the radar track generative adversarial network model constructed in S3, training the constructed radar track generative adversarial network model using the training data set constructed in S2, and testing the trained radar track generative adversarial network model using the test data set constructed in S2 until the network converges, to obtain a trained radar track generative adversarial network model;

[0012] S5, inputting a sub-image of the real track image into the trained radar track generative adversarial network model to generate a realistic synthetic track image sub-image;

[0013] S6, splicing the generated synthetic track image sub-image to obtain a complete synthetic target track image.

[0014] Preferably, the collection of radar track data of a target and corresponding automatic identification system (AIS) data of a ship and the pre-processing of the data in S1 specifically comprise:

[0015] S1-1, obtaining measured radar measurement data of a low-speed water surface ship target and corresponding AIS data, and then pre-processing the measured radar measurement data and the AIS data by removing outliers and filling in missing values;

[0016] S1-2, selecting target data in a specified size region according to target latitude and longitude information, and representing original target position information in a rectangular coordinate system.

[0017] Preferably, in S2, the preprocessed radar track data and AIS data are respectively converted into target actual track images and real track images, and the target actual track images and real track images are cropped and paired into sub-image blocks, a part of the sub-image blocks are randomly selected to construct a training data set, and the remaining sub-image blocks construct a test data set; specifically including:

[0018] S2-1, selecting a measurement point (x0, y0) in a track from the measured radar measurement data of the preprocessed low-speed water surface ship target, selecting measurement points in the region of (x0-R W / 2, x0+R W / 2) and (y0-R H / 2, y0+R H / 2) of the track to draw a track gray image, wherein x0 is the horizontal axis coordinate of the target, y0 is the vertical axis coordinate of the target, and R W and R H are equal and fixed size values, i.e. a target actual track image is obtained;

[0019] Selecting a measurement point (x0, y0) in a track from the preprocessed AIS data of the low-speed water surface ship target, selecting measurement points in the region of (x0-R W / 2, x0+R W / 2) and (y0-R H / 2, y0+R H / 2) of the track to draw a track gray image, wherein x0 is the horizontal axis coordinate of the target, y0 is the vertical axis coordinate of the target, and R W and R H are equal and fixed size values, i.e. a target real track image of the corresponding region is obtained;

[0020] S2-2, the target actual track image and the real track image obtained in S2-1 are respectively cropped into a target actual track sub-image and a target real track sub-image of a fixed size, the target actual track sub-image and the target real track sub-image are one-to-one paired to form a sub-image block, thereby constructing a training data set and a test data set.

[0021] Preferably, a radar track generative adversarial network model is constructed, which includes two generators with different network structures and a discriminator,

[0022] The two generators with different network structures both adopt an encoder-decoder architecture, referred to as Generator 1 and Generator 2 respectively, and each consists of three parts: a downsampled network structure, a residual block part, and an upsampled network structure.

[0023] The encoder backbone network of the generator 1 adopts a proposed neural network with strong feature extraction capability and progressively smaller feature maps. Through multiple downsampling, the encoder encodes the input trajectory image into a feature map of a fixed size. The feature map output by the encoder is upsampled multiple times by the decoder to obtain a grayscale image of the same size as the input image. The decoder and encoder network structures are symmetrical to each other.

[0024] The backbone network of generator 2 is similar to that of generator 1, except that the backbone network is replaced with a neural network with strong feature extraction capabilities and feature maps that decrease layer by layer.

[0025] The discriminator uses PatchGAN. The input real image is a 2-channel image stitched together with the real trajectory and the corresponding actual track in the channel dimension. The input pseudo image is a 2-channel image stitched together with the real trajectory and the corresponding pseudo track generated by generator 2 in the channel dimension. The discriminator convolves the input image layer by layer and finally outputs a feature map with a certain scale reduction. This feature map outputs a predicted probability value for a certain region of the input image. The average value of all prediction results is used as the final judgment result.

[0026] Preferably, the loss function comprises three parts: structural similarity index measurement (SSIM) loss, mean squared error loss, and adversarial loss;

[0027] The adversarial loss is defined as:

[0028]

[0029] Where x is the target true trajectory sub-image, y is the corresponding target track sub-image, G(x) is the synthesized track sub-image output by the generator; D(y) represents the true track sub-image entering the discriminator, while D(G(x)) represents the false track sub-image entering the discriminator.

[0030] The structural similarity loss of generator 1 is defined as:

[0031]

[0032] Where, μ x μ represents the mean of the target's true trajectory sub-image x. y y represents the mean of the target trajectory sub-image. The variance of the target's true trajectory sub-image x represents the variance of the target's true trajectory. σ represents the variance of the target trajectory sub-image y.xy is the covariance of the target real trajectory sub-image x and the target track sub-image y, and c1 and c2 are constants used to maintain stability;

[0033] The mean square error loss of the generator 2 is defined as:

[0034]

[0035] where m and n represent the width and height of the target real trajectory sub-image and the target track sub-image respectively, and x(i,j) and y(i,j) represent the pixel values of the corresponding positions of the target real trajectory sub-image and the target track sub-image respectively.

[0036] The overall loss of the network is defined as:

[0037] L total =L GAN +λ1L SSIM +λ2L MSE

[0038] where λ1 and λ2 represent the weights of L SSIM and L MSE respectively.

[0039] Compared with the prior art, the present application has the following advantages:

[0040] 1. The present application converts the radar track data of the target and the corresponding automatic identification system AIS data into the target actual track image and the real trajectory image, and cuts them into fixed-size one-to-one paired sub-image blocks, and then constructs a data set, and trains the radar track generative adversarial network model comprising two different network structures of a generator and a discriminator until the network converges and the training is completed. During the training process, the data in the data set only fixes the starting time of the target data, and does not need to pay attention to the sampling period and sampling time of the intermediate time, avoiding the strict time alignment operation of the target trajectory and the radar track.

[0041] 2. The present application finally inputs the sub-image blocks of the real trajectory image into the trained radar track generative adversarial network model, generates realistic synthetic track image sub-image blocks, and splices the generated synthetic track image sub-image blocks to obtain a complete synthetic target track image. Therefore, the target track sequence data is converted into image data, the powerful performance of the generative adversarial network in the field of visual image processing is fully utilized, the real error distribution law of the target track can be effectively learned, and a large amount of track data similar to the real error distribution can be generated.

[0042] The technical solutions of the present application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flow chart of a method for intelligent simulation generation of a marine target track is provided for Embodiment 1 of the present application.

[0044] Figure 2 A flow chart for constructing a training set for Embodiment 1 of the present application.

[0045] Figure 3 A general framework diagram of a radar track generation adversarial network model for Embodiment 1 of the present application.

[0046] Figure 4 Network architecture of two generators in Embodiment 1 of the present application.

[0047] Figure 5 Network architecture of a discriminator in Embodiment 1 of the present application.

[0048] Figure 6 Track generated for true value track; (a) and (b) are the real track and the corresponding real track respectively, (c) is the track generated by the original pix2pixGAN, and (d) is the track generated by Embodiment 1 of the present application. DETAILED DESCRIPTION

[0049] Embodiment 1

[0050] As shown in Figure 1 , a method for intelligent simulation generation of a marine target track according to an embodiment of the present application, the method comprising:

[0051] S1, collecting radar track data and corresponding automatic identification system (AIS) data of a target, and pre-processing the data;

[0052] S2, converting the pre-processed radar track data and AIS data into target actual track images and real track images respectively, and pairing the target actual track images and real track images into sub-image blocks after cropping, randomly selecting a part of the sub-image blocks to construct a training data set, and the remaining part of the sub-image blocks to construct a test data set;

[0053] S3, constructing a radar track generation adversarial network model comprising two generators with different network structures and a discriminator;

[0054] S4, creating a loss function of the radar track generation adversarial network model constructed in S3, training the constructed radar track generation adversarial network model using the training data set constructed in S2, and testing the trained radar track generation adversarial network model using the test data set constructed in S2, until the network converges, obtaining the trained radar track generation adversarial network model;

[0055] S5, input the sub-image of the real trajectory image into the trained radar trajectory generation adversarial network model to generate a sub-image of a realistic synthetic trajectory image;

[0056] S6, splice the generated sub-image of the synthetic trajectory image to obtain a complete synthetic target trajectory image.

[0057] In this embodiment, the specific method of collecting the radar trajectory data of the target and the corresponding automatic identification system AIS data in step S1 and preprocessing the data is as follows:

[0058] Obtain the measured radar measurement sequence data of the low-speed water surface ship target and the corresponding AIS sequence data, and perform outlier removal and missing value filling data preprocessing;

[0059] According to the target latitude and longitude information, select the target data in a 50x50 km area, and represent the original position information in a rectangular coordinate system.

[0060] The specific method of step S2 is as follows:

[0061] As shown in the following formula: Figure 2 S2-1, select a measurement point (x0, y0) in a track from the preprocessed measured radar measurement data of the low-speed water surface ship target, select the measurement points in the (x0-R W / 2, x0+R W / 2) and (y0-R H / 2, y0+R H / 2) regions of the track to draw a track grayscale image, wherein x0 is the target horizontal axis coordinate, y0 is the target vertical axis coordinate, R W and R H are equal and fixed values, that is, the actual target trajectory image is obtained;

[0062] S2-2, select a measurement point (x0, y0) in a track from the preprocessed AIS data of the low-speed water surface ship target, select the measurement points in the (x0-R W / 2, x0+R W / 2) and (y0-R H / 2, y0+R H / 2) regions of the track to draw a track grayscale image, wherein x0 is the target horizontal axis coordinate, y0 is the target vertical axis coordinate, R W and R H are equal and fixed values, that is, the target real trajectory image of the corresponding region is obtained;

[0063] S2-2, the target actual track image and the real track image obtained in S2-1 are respectively cropped into target actual track sub-images and target real track sub-images of a fixed size, the target actual track sub-images and the target real track sub-images are paired one by one to form sub-image blocks, and thus a training data set and a test data set are constructed.

[0064] The actual area where the track is located has a size of 2560m x 2560m, the area is divided into 2560 x 2560 grids, the value of the grid where the target track is located is 0, and the value of other grids is 1, a track gray image with a size of 2560 x 2560 pixels is obtained, and the actual distance represented by each pixel of the image is 1m. Similarly, a track gray image with a size of 2560 x 2560 pixels is obtained. The corresponding target track image and track image are cropped into a 256 x 256 pixel sub-image pair. According to the method, 2000 sub-image pairs are obtained, 60% of which are randomly selected as a training set, and 40% of which are randomly selected as a test set.

[0065] As shown in Figure 3 The generative adversarial network model proposed in the application includes two generators with different network structures and a discriminator. Both of the two generators with different network structures adopt an encoder-decoder architecture, and are respectively denoted as generator 1 and generator 2, and are composed of a down-sampling network structure, a residual block part and an up-sampling network structure.

[0066] The encoder backbone network of the generator 1 adopts a neural network that has been proposed and has strong feature extraction capability and whose feature maps are reduced layer by layer. Through multiple down-sampling, the encoder encodes the input track image into a feature map of a fixed size. The feature map output by the encoder is up-sampled multiple times by the decoder to obtain a gray image with the same size as the input image, wherein the network structure of the decoder and the encoder is symmetrical to each other.

[0067] The backbone network of the generator 2 is similar to that of the generator 1, except that the backbone network is replaced by a neural network that has strong feature extraction capability and whose feature maps are reduced layer by layer.

[0068] The discriminator adopts PatchGAN. The input real image is a 2-channel image obtained by concatenating a real track and a corresponding actual track in the channel dimension, and the input pseudo image is a 2-channel image obtained by concatenating a real track and a corresponding pseudo track generated by the generator 2 in the channel dimension. The discriminator convolves the input image layer by layer, and finally outputs a feature map whose scale is reduced by a certain proportion. The feature map outputs a prediction probability value for a certain region of the input image, and the average value of all prediction results is used as the final determination result.

[0069] As shown in Figure 5As shown, the discriminator adopts PatchGAN, which is composed of a 5-layer fully convolutional network, and the output is a 15x15 matrix, each value in the matrix is a predicted probability value for each region of the input image, and the final output of the discriminator is obtained by averaging all the judgment results.

[0070] The loss function of the model includes three parts, structural similarity index measurement (SSIM) loss, mean square error (MSE) loss and adversarial loss, and the adversarial loss is defined as:

[0071]

[0072] Where x is the target real trajectory sub-image, y is the target track sub-image corresponding thereto, G(x) is the synthesized track sub-image output by the generator, D(y) represents the real track sub-image entering the discriminator, and D(G(x)) represents the false track sub-image entering the discriminator.

[0073] The structural similarity loss of the generator 1 is defined as:

[0074]

[0075] Where μ x represents the mean of the target real trajectory sub-image x, μ y represents the mean of the target track sub-image y, represents the variance of the target real trajectory sub-image x, represents the variance of the target track sub-image y, σ xy is the covariance of the target real trajectory sub-image x and the target track sub-image y, and c1 and c2 are constants used to maintain stability.

[0076] The mean square error loss of the generator 2 is defined as:

[0077]

[0078] Where m and n represent the width and height of the target real trajectory sub-image and the target track sub-image respectively, and x(i,j) and y(i,j) represent the pixel values of the corresponding positions of the target real trajectory sub-image and the target track sub-image respectively.

[0079] The overall loss of the network is defined as:

[0080] L total = L GAN + λ1L SSIM + λ2L MSE

[0081] Where λ1 and λ2 represent the weights of L SSIM and L MSE respectively.

[0082] The experiment adopts an Adam optimizer, a batch size of 16, 200 epochs, and learning rates of 1e -4 , 2e -4 , and 2e -4 for the generator 1, the generator 2, and the discriminator, respectively. The weights of L SSIM and L MSE are set to 1 and 10, i.e., λ1=1 and λ2=10. The weights of the generator or the discriminator are fixed during training, and the other is updated. The network is optimized in alternating training.

[0083] A comparative experiment was performed, and some of the true value trajectory generated tracks are shown in FIG. 3. Figure 6 (a) and (b) are the real trajectory and the corresponding real track, respectively, (c) is the track generated by the original pix2pixGAN, and (d) is the track generated by the method of the embodiment 1 of the present application. From the experimental results, both models can generate clear track samples. The generated tracks have increased errors compared with the true value trajectory, and the error distribution characteristics are similar to the real track. In the turning motion and irregular motion, the track generated by the original pix2pixGAN has an underfitting condition. The track generated by the method of the present application can better match the real track in the overall and local structure, indicating that the method of the present application generates the sea target track more in line with the real track error distribution, and has good effect.

[0084] The above description is only a preferred embodiment of the present application, and does not limit the present application in any way. Any simple modification, change, and equivalent change made according to the technical essence of the present application to the above embodiment are still within the protection scope of the technical solution of the present application.

Claims

1. A method for intelligent simulation and generation of maritime target tracks, characterized in that, The method includes: S1. Collect the radar track data of the target and the corresponding Automatic Identification System (AIS) data, and preprocess the data; S2. Convert the preprocessed radar track data and AIS data into actual target track images and true trajectory images, respectively. Then, crop the actual target track images and true trajectory images and pair them into sub-image blocks. Randomly select a portion of the sub-image blocks to construct the training dataset, and the remaining portion of the sub-image blocks to construct the test dataset. S3. Construct a radar trajectory generative adversarial network model comprising two generators with different network structures and one discriminator. The two generators with different network structures both adopt an encoder-decoder architecture, denoted as Generator 1 and Generator 2 respectively, and each consists of a downsampling network structure, a residual block, and an upsampling network structure. The encoder backbone of Generator 1 uses a proposed neural network with strong feature extraction capabilities and progressively smaller feature maps. Through multiple downsampling operations, the encoder encodes the input trajectory image into a fixed-size feature map. The feature map output by the encoder is then upsampled multiple times by the decoder to obtain a grayscale image of the same size as the input image. The decoder and... The encoder network structures are symmetrical. The backbone network of generator 2 is similar to that of generator 1, except that the backbone network is replaced by a neural network with strong feature extraction capabilities and a feature map that is progressively smaller. The discriminator uses PatchGAN. The input real image is a 2-channel image of the real trajectory and the corresponding actual track in the channel dimension. The input pseudo image is a 2-channel image of the real trajectory and the corresponding pseudo track generated by generator 2 in the channel dimension. The discriminator convolves the input image layer by layer and finally outputs a feature map that is scaled down by a certain ratio. This feature map outputs a prediction probability value for a certain region of the input image. The average value of all prediction results is used as the final judgment result. S4. Create the loss function for the radar trajectory generative adversarial network model constructed in S3. Use the training dataset constructed in S2 to train the radar trajectory generative adversarial network model and use the test dataset constructed in S2 to test the trained radar trajectory generative adversarial network model until the network converges and the trained radar trajectory generative adversarial network model is obtained. S5. Input the sub-images of the real trajectory image into the trained radar trajectory generative adversarial network model to generate realistic sub-images of the synthetic trajectory image. S6. Stitch together the sub-images of the generated synthetic track image to obtain a complete synthetic target track image.

2. The intelligent simulation generation method for maritime target tracks according to claim 1, characterized in that, S1 describes collecting radar track data of the target and corresponding Automatic Identification System (AIS) data, and preprocessing the data, specifically including: S1-1. Obtain the measured radar measurement data and corresponding AIS data of low-speed surface vessel targets, and then perform preprocessing on the measured radar measurement data and AIS data, including outlier removal and missing value filling. S1-2. Select target data within a specified area based on the target's latitude and longitude information, and represent the original target location information in a rectangular coordinate system.

3. The intelligent simulation generation method for maritime target tracks according to claim 2, characterized in that, In S2, the preprocessed radar track data and AIS data are converted into actual target track images and true trajectory images, respectively. These images are then cropped and paired into sub-image patches. A subset of these sub-image patches is randomly selected to construct the training dataset, while the remaining patches are used to construct the test dataset. Specifically, this includes: S2-1. Select a measurement point from a trajectory in the preprocessed radar measurement data of low-speed surface vessel targets. Select this flight path and The measurement points within the area are plotted as a grayscale image of the flight path, where, The x-axis coordinate of the target The target's vertical axis coordinate, and The values ​​are equal and of a fixed size, which gives us the actual trajectory image of the target; Select a measurement point from a trajectory in the preprocessed AIS data of low-speed surface vessel targets. Select this trajectory and The measurement points within the area are plotted as a trajectory grayscale image, where, The x-axis coordinate of the target The target's vertical axis coordinate, and The values ​​are equal and of a fixed size, which gives the true trajectory image of the target in the corresponding region; S2-2. The actual target trajectory image and the true target trajectory image obtained in S2-1 are cropped into fixed-size sub-images of the actual target trajectory and sub-images of the true target trajectory. The sub-images of the actual target trajectory and sub-images of the true target trajectory are paired one-to-one to form sub-image blocks, thereby constructing the training dataset and the test dataset.

4. The intelligent simulation generation method for maritime target tracks according to claim 3, characterized in that, The loss function consists of three parts: structural similarity index measurement loss, mean squared error loss, and adversarial loss. The adversarial loss is defined as: ； in, It is a sub-image of the target's true trajectory. It is the corresponding target track sub-image. It is the synthesized track sub-image output by the generator; This represents the actual track sub-image entering the discriminator, while This represents a false track sub-image entering the discriminator; The structural similarity loss of generator 1 is defined as: ； in, Sub-image representing the true trajectory of the target The mean, Represents the target track sub-image The mean, Sub-image representing the true trajectory of the target variance Represents the target track sub-image variance It is the target's true trajectory sub-image and target track sub-image covariance, and It is a constant used to maintain stability; The mean squared error loss of generator 2 is defined as: ； in, and Let represent the width and height of the target's true trajectory sub-image and the target's flight path sub-image, respectively. and These represent the pixel values ​​at corresponding positions in the target's true trajectory sub-image and the target's flight path sub-image, respectively. The overall loss of the network is defined as: ； in and They represent and The weight.

Images