A ship target tracking method and system based on trajectory adaptive correction in low-resolution optical remote sensing images

Abstract

The invention discloses a low-resolution optical remote sensing image ship target tracking method based on trajectory adaptive correction, comprising: obtaining low-resolution optical remote sensing image sequence data to be detected, and performing preprocessing to obtain superimposed image blocks; constructing a target tracking model including an S-MOTR network and a TCNet network, and training to obtain a trained target tracking model; inputting the superimposed image block into the trained target tracking model, the S-MOTR network performs end-to-end target tracking processing on the superimposed image block to obtain an initial tracking trajectory; the TCNet network performs adaptive correction processing on the initial tracking trajectory to obtain a corrected tracking trajectory; in the training stage, the S-MOTR network and the TCNet network respectively calculate their respective loss functions and perform their respective corresponding supervised training; the TCNet network obtains a prediction frame based on the corrected tracking trajectory and feeds it back to the S-MOTR network, and the S-MOTR network calculates the loss of the tracking target frame based on the prediction frame. The accuracy of ship tracking of low-resolution optical remote sensing images is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more particularly to a method and system for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction. Background Art

[0002] Optical remote sensing image ship target tracking technology is an important part of remote sensing image analysis. Through advanced image processing and pattern recognition algorithms, it can accurately identify and track target ships from complex ocean backgrounds. It plays an irreplaceable role in maritime security monitoring, maritime search and rescue, waterway management, illegal fishing monitoring, and military operations.

[0003] At present, although the popular deep learning methods have been able to achieve good tracking of ships in optical remote sensing images, they are also facing increasingly severe challenges, including interference and occlusion caused by clouds and fog, huge changes in the visual appearance of the target caused by background clutter, lighting, shadows, etc. In particular, when performing tracking tasks under conditions of high real-time requirements, since the original remote sensing images (level 0 images) have not been precisely corrected for geo-geometry, the relative positioning errors between the frames of the sequence images make it difficult to further improve the accuracy of target tracking.

[0004] Therefore, how to improve the accuracy of ship tracking in low-resolution optical remote sensing images is an urgent problem to be solved by those skilled in the art. Summary of the invention

[0005] In view of this, the present invention provides a method and system for tracking ship targets using low-resolution optical remote sensing images based on trajectory adaptive correction, thereby improving the accuracy of ship tracking using low-resolution optical remote sensing images.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A ship target tracking method based on low-resolution optical remote sensing images with trajectory adaptive correction, comprising:

[0008] Obtaining low-resolution optical remote sensing image sequence data to be detected, and performing preprocessing to obtain superimposed image blocks;

[0009] Construct a target tracking model including an S-MOTR network and a TCNet network, and train them to obtain a trained target tracking model;

[0010] The superimposed image block is input into the trained target tracking model, and the S-MOTR network performs end-to-end target tracking processing on the superimposed image block to obtain an initial tracking trajectory;

[0011] The TCNet network performs adaptive correction processing on the initial tracking trajectory to obtain a corrected tracking trajectory;

[0012] In the training phase, the S-MOTR network and the TCNet network calculate their respective loss functions and perform their respective supervised training;

[0013] The TCNet network obtains a prediction frame based on the corrected tracking trajectory and feeds it back to the S-MOTR network. The S-MOTR network calculates the loss of the tracking target frame based on the prediction frame.

[0014] Preferably, the pretreatment specifically includes:

[0015] Quantification is performed based on the low-resolution optical remote sensing image sequence data to obtain a visualized image;

[0016] Perform cropping based on the visualized image to obtain an image block of a preset size;

[0017] The image blocks of three consecutive frames are sequentially filled into three channels of the image so that the moving target forms a pseudo-color track in the image, thereby obtaining the superimposed image block.

[0018] Preferably, the S-MOTR network comprises: a channel attention association module, a first encoder and a first decoder;

[0019] The superimposed image block is input into the channel attention association module for processing to obtain an association feature map;

[0020] The associated feature maps are input to the first encoder and the first decoder respectively;

[0021] The first encoder performs feature extraction based on the associated feature map to obtain an input feature map;

[0022] The first decoder performs processing based on the input feature map and the associated feature map, and interprets and obtains the initial tracking trajectory of each current target in the low-resolution optical remote sensing image sequence data.

[0023] Preferably, the channel attention association module comprises: a convolutional layer, a global pooling layer and a first attention layer;

[0024] The superimposed image blocks are sequentially input into the convolutional layer, the global pooling layer and the first attention layer to extract channel correlation features to obtain the correlation feature map.

[0025] Preferably, the TCNet network comprises: a position encoder, a second encoder, a second decoder and a protection layer;

[0026] The initial tracking trajectory is input to the position encoder for processing, and each trajectory point is position-encoded corresponding to the current frame target to obtain an encoded tracking trajectory;

[0027] The encoded tracking trajectory is input to the second encoder for processing to obtain an encoded deep feature map;

[0028] The encoded deep feature map is input to the second decoder for processing to obtain a smooth tracking trajectory;

[0029] The smoothed tracking trajectory and the encoded tracking trajectory are input to the protection layer for processing to obtain the corrected tracking trajectory.

[0030] Preferably, the second encoder comprises: a first normalization layer, a second attention layer, a second normalization layer and a fully connected layer;

[0031] The encoded tracking trajectory is sequentially input into the first normalization layer, the second attention layer, the second normalization layer and the fully connected layer for processing to obtain the encoded deep feature map.

[0032] Preferably, the protective layer treatment process is:

[0033] The smooth tracking trajectory and the encoded tracking trajectory are input to the protection layer;

[0034] The protection layer obtains a first position code P1 of the current frame target based on the code tracking trajectory;

[0035] The protection layer obtains a second position code P2 of the current frame target based on the smooth tracking trajectory;

[0036] The protection layer replaces the second position code P2 in the smooth tracking trajectory with the first position code P1, and finally outputs the corrected tracking trajectory.

[0037] Preferably, the correction trajectory loss function corresponding to the TCNet network is:

[0038] L total =λ 1 L distance +λ 2 L dircction +λ 3 L vclocity

[0039]

[0040] Among them, L total represents the correction trajectory loss function, L distance represents the distance consistency loss function, L dircctionrepresents the direction consistency loss function, L vclocity represents the speed consistency loss function, λ 1 , 2 and λ 3 They represent the weights used to balance the effects of different loss functions on trajectory correction, X represents the image trajectory sequence, T represents the trajectory length, and t represents the current time step.

[0041] Preferably, the S-MOTR network calculates the loss of tracking the target frame based on the predicted frame, specifically including:

[0042] The S-MOTR network obtains a feedback loss function based on the predicted frame, and calculates the loss of the tracking target frame based on the feedback loss function;

[0043] The feedback loss function is specifically:

[0044]

[0045] Among them, α i represents the weight coefficient of the i-th prediction box, L bbox represents the bounding box loss, b i and They represent the i-th predicted box and the corresponding i-th true value box respectively.

[0046] A low-resolution optical remote sensing image ship target tracking system based on trajectory adaptive correction, comprising: a data acquisition processing module, a model building module, a trajectory acquisition module, a trajectory correction module and a training module;

[0047] The data acquisition and processing module is used to acquire low-resolution optical remote sensing image sequence data to be detected and perform preprocessing to obtain superimposed image blocks;

[0048] The model building module is used to build a target tracking model including an S-MOTR network and a TCNet network, and perform training to obtain a trained target tracking model;

[0049] The trajectory acquisition module is used to input the superimposed image block into the trained target tracking model, and the S-MOTR network performs end-to-end target tracking processing on the superimposed image block to obtain an initial tracking trajectory;

[0050] The trajectory correction module is used to perform adaptive correction processing on the initial tracking trajectory through the TCNet network to obtain a corrected tracking trajectory;

[0051] The training module is used to calculate the respective loss functions based on the S-MOTR network and the TCNet network and perform respective supervised training during the training phase; the TCNet network obtains a prediction frame based on the corrected tracking trajectory and feeds it back to the S-MOTR network, and the S-MOTR network calculates the loss of the tracking target frame based on the prediction frame.

[0052] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction. The present invention can improve the trajectory fluctuation problem caused by directly tracking sequence images that have not been geometrically corrected under high real-time tasks, and can also improve the target tracking accuracy to a certain extent, and obtain a trajectory that can reflect the true value movement of the ship target and a highly accurate target tracking result. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0054] Figure 1 A flow chart of a method for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction is provided by the present invention.

[0055] Figure 2 This is a schematic diagram of the S-MOTR network structure provided by the present invention.

[0056] Figure 3 This is a schematic diagram of the TCNet network structure provided by the present invention.

[0057] Figure 4 This is a flow chart of the protective layer processing method provided by the present invention.

[0058] Figure 5 This is a schematic diagram of the preprocessed superimposed image blocks provided by the present invention.

[0059] Figure 6 This is a schematic diagram of the initial tracking trajectory provided by the present invention.

[0060] Figure 7 This is a schematic diagram of the correction tracking trajectory provided by the present invention.

[0061] Figure 8 A schematic structural diagram of a low-resolution optical remote sensing image ship target tracking system based on trajectory adaptive correction provided by the present invention.

[0062] Fig. 9 A schematic structural diagram of a low-resolution optical remote sensing image ship target tracking device based on trajectory adaptive correction provided by the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] Example 1

[0065] like Figure 1 As shown, the embodiment of the present invention discloses a low-resolution optical remote sensing image ship target tracking method based on trajectory adaptive correction, comprising:

[0066] Obtaining low-resolution optical remote sensing image sequence data to be detected, and performing preprocessing to obtain superimposed image blocks;

[0067] Construct a target tracking model including an S-MOTR network and a TCNet network, and train them to obtain a trained target tracking model;

[0068] The superimposed image blocks are input into the trained target tracking model, and the S-MOTR network performs end-to-end target tracking processing on the superimposed image blocks to obtain the initial tracking trajectory;

[0069] The TCNet network performs adaptive correction processing on the initial tracking trajectory to obtain the corrected tracking trajectory;

[0070] In the training phase, the S-MOTR network and the TCNet network calculate their respective loss functions and perform their respective supervised training;

[0071] The TCNet network obtains the predicted box based on the corrected tracking trajectory and feeds it back to the S-MOTR network. The S-MOTR network calculates the loss of the tracked target box based on the predicted box.

[0072] Example 2

[0073] The embodiment of the present invention discloses a low-resolution optical remote sensing image ship target tracking method based on trajectory adaptive correction, comprising:

[0074] Obtain the low-resolution optical remote sensing image sequence data to be detected and perform preprocessing to obtain the superimposed image blocks:

[0075] Preferably, the acquired data to be detected are: original 16-bit optical remote sensing image sequence data of level 0 taken by a satellite, that is, original 16-bit wide-width low-resolution single-channel sequence remote sensing image data.

[0076] Preferably, the pretreatment specifically includes:

[0077] Quantification is performed based on low-resolution optical remote sensing image sequence data to obtain a visual image;

[0078] Cropping is performed based on the visualized image to obtain an image block of a preset size;

[0079] The image blocks of three consecutive frames are filled into the three channels of the image in order, so that the moving target forms a pseudo-color track in the image and obtains the superimposed image block.

[0080] Preferably, in this embodiment, the original 16-bit low-resolution optical remote sensing image sequence data is linearly quantized to 0.3% to obtain a clear 8-bit visualized image; the visualized image is cut into image blocks of 512×512 size for making model training samples; the image blocks of three consecutive frames are sequentially filled into the three channels of the image so that the moving target forms a pseudo-color trajectory in the image, thereby enhancing the characteristics of the small target and obtaining a superimposed image block; the target truth box is annotated according to the obtained superimposed image block for model training.

[0081] A target tracking model including an S-MOTR network and a TCNet network is constructed and trained to obtain a trained target tracking model.

[0082] The superimposed image blocks are input into the trained target tracking model, and the S-MOTR network performs end-to-end target tracking processing on the superimposed image blocks to obtain the initial tracking trajectory:

[0083] Preferably, Figure 2 As shown, the S-MOTR network includes: a channel attention association module, a first encoder and a first decoder;

[0084] The superimposed image blocks are input into the channel attention association module for processing to obtain the association feature map;

[0085] The associated feature maps are input to the first encoder and the first decoder respectively;

[0086] The first encoder performs feature extraction based on the associated feature map to obtain an input feature map;

[0087] The first decoder processes the input feature map and the associated feature map to interpret the initial tracking trajectory of each current target in the low-resolution optical remote sensing image sequence data.

[0088] Preferably, since the low-resolution optical remote sensing image sequence data to be detected uses a continuous frame superposition method to enhance the target features during preprocessing, three consecutive frames are superimposed on the three channels of the image, so the three channels themselves contain the motion information of the target. The S-MOTR network adds a channel attention association module before feature extraction to extract feature maps with stronger channel correlation from the input image for network learning.

[0089] Preferably, the channel attention association module includes: a convolution layer, a global pooling layer and a first attention layer;

[0090] The superimposed image blocks are sequentially input into the convolutional layer, the global pooling layer, and the first attention layer for channel correlation feature extraction to obtain the correlation feature map.

[0091] Preferably, after processing through the attention association module, a correlation feature map is obtained, which is not only used as the input of the first encoder, but also used as the input of the first decoder through a jump connection, so as to supplement the motion information contained in the three channels into the first decoder, so as to further enhance the learning performance of the S-MOTR network for the target motion features.

[0092] Preferably, the decoding process of the first decoder is:

[0093] The first decoder performs processing based on the input feature map and the associated feature map, adds the detected targets to the generated tracking sequence, assigns new IDs to the previously undetected targets, and generates the initial tracking trajectory of the target corresponding to each current ID.

[0094] Preferably, the S-MOTR network is an end-to-end target tracking network based on Transformer. The first encoder and the first decoder both use Deformable DETR to extract image features. Deformable DETR combines the sparse sampling capability of DCN and the global relationship modeling capability of Transformer, and can aggregate multi-scale features, which is conducive to feature extraction of small-size targets.

[0095] Preferably, the S-MOTR network is trained with the preprocessed image as input, the target position box and the target ID as true values, so that the S-MOTR network can autonomously locate and track the target from the input image.

[0096] The TCNet network performs adaptive correction processing on the initial tracking trajectory to obtain the corrected tracking trajectory:

[0097] Preferably, Figure 3 As shown, the TCNet network includes: a position encoder, a second encoder, a second decoder and a protection layer;

[0098] The initial tracking trajectory is input to the position encoder for processing, and the position of each trajectory point corresponding to the current frame target is encoded to obtain the encoded tracking trajectory;

[0099] The encoded tracking trajectory is input to the second encoder for processing to obtain an encoded deep feature map;

[0100] The encoded deep feature map is input into the second decoder for processing to obtain a smooth tracking trajectory;

[0101] The smoothed tracking trajectory and the encoded tracking trajectory are input to the protection layer for processing to obtain the corrected tracking trajectory.

[0102] Preferably, the second encoder comprises: a first normalization layer, a second attention layer, a second normalization layer and a fully connected layer;

[0103] The encoded tracking trajectory is sequentially input into the first normalization layer, the second attention layer, the second normalization layer and the fully connected layer for processing to obtain the encoded deep feature map.

[0104] Preferably, since trajectory smoothing will change the position of the trajectory points, in order to ensure that the trajectory smoothing process does not affect the positioning information of the current frame target, the TCNet network adds a protection layer after the output of the second encoder. The output after processing by the protection layer is the final smooth trajectory, that is, the corrected tracking trajectory.

[0105] Preferably, Figure 4 As shown in the figure, the protection layer processing process is:

[0106] The smooth tracking trajectory and the encoded tracking trajectory are input to the protection layer;

[0107] The protection layer obtains the first position code P1 of the target in the current frame based on the coding tracking trajectory;

[0108] The protection layer obtains the second position code P2 of the current frame target based on the smooth tracking trajectory;

[0109] The protection layer replaces the second position code P2 in the smooth tracking trajectory with the first position code P1, and finally outputs the corrected tracking trajectory.

[0110] Preferably, the protection layer can protect the target position point of the current frame, ensuring that the trajectory correction will not affect the target positioning of the current frame, and ensuring that only the historical trajectory curve will be smoothed during trajectory correction without changing the target position of the current frame.

[0111] Preferably, the TCNet network is an end-to-end encoder-decoder structure, which takes the initial tracking trajectory generated by the S-MOTR network as input and the smooth trajectory accurately corrected by geographic geometry as the true value, and trains the TCNet network to autonomously recover the smooth trajectory from the jittered trajectory of the initial tracking trajectory.

[0112] During the training phase, the S-MOTR network and the TCNet network calculate their respective loss functions and perform their respective supervised training:

[0113] Preferably, in order to achieve consistent target tracking results between frames, in the design of the loss function of the TCNet network, according to the ship dynamics model, the direction and speed of the ship in different frames remain roughly unchanged. For the distance mutation, speed mutation and direction mutation caused by the inter-frame positioning error, distance consistency loss, direction consistency loss and speed consistency loss are designed respectively to restore the real motion trend of the ship from the fluctuating trajectory.

[0114] Preferably, the correction trajectory loss function corresponding to the TCNet network is:

[0115] L total =λ 1 L distance +λ 2 L dircction +λ 3 L vclocity

[0116]

[0117] Among them, L total represents the correction trajectory loss function, L distance represents the distance consistency loss function, L dircction represents the direction consistency loss function, L vclocity represents the speed consistency loss function, λ 1 , 2 and λ 3 They represent the weights used to balance the effects of different loss functions on trajectory correction, X represents the image trajectory sequence, T represents the trajectory length, and t represents the current time step.

[0118] The TCNet network obtains the predicted box based on the corrected tracking trajectory and feeds it back to the S-MOTR network. The S-MOTR network calculates the loss of the tracked target box based on the predicted box:

[0119] Preferably, the TCNet network obtains the target's moving direction and speed according to the corrected tracking trajectory, predicts the target's position information for the next frame, obtains the prediction box information based on the position information, and the TCNet network integrates the prediction box information into the trackquery and feeds it back to the S-MOTR network. The S-MOTR network calculates the loss of the tracking target box based on the prediction box.

[0120] Preferably, the S-MOTR network calculates the loss of the tracking target box based on the predicted box, specifically including:

[0121] The S-MOTR network obtains a feedback loss function based on the predicted box, and calculates the loss of the tracked target box based on the feedback loss function;

[0122] The feedback loss function is specifically:

[0123]

[0124] Among them, α i represents the weight coefficient of the i-th prediction box, L bbox represents the bounding box loss, b i and They represent the i-th predicted box and the corresponding i-th true value box respectively.

[0125] Preferably, providing an S-MOTR network to calculate the loss of the tracked target box can increase the weight of the predicted box, so that the tracker pays more attention to the target area, thereby improving the tracking performance of the tracker.

[0126] Preferably, the S-MOTR network and the TCNet network are combined into an end-to-end network, and each is supervised but integrated for network training.

[0127] Preferably, the S-MOTR network and the TCNet network calculate their own loss functions and perform gradient backpropagation respectively; wherein the loss function of the S-MOTR network is weighted by the current frame target prediction box output by TCNet according to the historical trajectory information, so that S-MOTR pays more attention to the area around the prediction box.

[0128] Preferably, Figure 5-Figure 7 As shown, a method for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction according to the present invention is a schematic diagram of the results obtained by performing trajectory tracking and trajectory correction on the low-resolution optical remote sensing image sequence data to be detected. Based on the results, it can be clearly seen that the method of the present invention improves the target tracking accuracy to a certain extent, and obtains a trajectory that can reflect the true movement of the ship target and a target tracking result with high accuracy.

[0129] Example 3

[0130] like Figure 8 As shown, a low-resolution optical remote sensing image ship target tracking system based on trajectory adaptive correction includes: a data acquisition processing module, a model building module, a trajectory acquisition module, a trajectory correction module and a training module;

[0131] A data acquisition and processing module is used to acquire low-resolution optical remote sensing image sequence data to be detected and perform preprocessing to obtain superimposed image blocks;

[0132] The model building module is used to build a target tracking model including an S-MOTR network and a TCNet network, and to train the model to obtain a trained target tracking model.

[0133] The trajectory acquisition module is used to input the superimposed image blocks into the trained target tracking model, and the S-MOTR network performs end-to-end target tracking processing on the superimposed image blocks to obtain the initial tracking trajectory;

[0134] The trajectory correction module is used to perform adaptive correction processing on the initial tracking trajectory through the TCNet network to obtain the corrected tracking trajectory;

[0135] The training module is used to calculate the respective loss functions based on the S-MOTR network and the TCNet network and perform corresponding supervised training during the training phase; the TCNet network obtains the prediction frame based on the corrected tracking trajectory and feeds it back to the S-MOTR network, and the S-MOTR network calculates the loss of the tracking target frame based on the prediction frame.

[0136] Example 4

[0137] A low-resolution optical remote sensing image ship target tracking device based on trajectory adaptive correction is provided to achieve accurate detection of small ship targets. The target detection device can be implemented by software and / or hardware.

[0138] like Fig. 9 As shown, the apparatus 300 includes an image acquisition device 301 , a memory 302 , and a processor 303 .

[0139] The image acquisition device 301, the memory 302 and the processor 303 may be connected via a bus or other means.

[0140] The image acquisition device 301 is used to acquire original remote sensing image data and send it to the processor 303 .

[0141] The processor 303 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components and other chips, or a combination of the above chips.

[0142] The memory 302, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as programs or instructions corresponding to the ship target tracking method in the first embodiment of the present application.

[0143] The processor 303 executes various functional applications and data processing of the processor by running the non-transitory software program or instructions stored in the memory 302, that is, implements the ship target tracking method in the above method embodiment.

[0144] The memory 302 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created by the processor 302 and the like.

[0145] In addition, the memory 302 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage devices.

[0146] In some aspects, the memory 302 optionally includes a memory remotely located relative to the processor 303 , and these remote memories may be connected to the processor 303 via a network.

[0147] Optionally, the above network includes but is not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0148] Example 5

[0149] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the ship target tracking method as described in any one of the above technical solutions is implemented.

[0150] It can be seen from the above technical solutions that the present invention discloses a method and system for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction. The present invention can improve the trajectory fluctuation problem caused by directly tracking sequential images that have not been geometrically corrected under high real-time tasks, and can also improve the target tracking accuracy to a certain extent, and obtain a trajectory that can reflect the true value movement of the ship target and a highly accurate target tracking result.

[0151] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0152] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A ship target tracking method based on low-resolution optical remote sensing images with trajectory adaptive correction, characterized in that: include: Obtaining low-resolution optical remote sensing image sequence data to be detected, and performing preprocessing to obtain superimposed image blocks; Construct a target tracking model including an S-MOTR network and a TCNet network, and train them to obtain a trained target tracking model; The S-MOTR network includes: a channel attention association module, a first encoder and a first decoder; The TCNet network includes: a position encoder, a second encoder, a second decoder and a protection layer; The superimposed image block is input into the trained target tracking model, and the S-MOTR network performs end-to-end target tracking processing on the superimposed image block to obtain an initial tracking trajectory; The TCNet network performs adaptive correction processing on the initial tracking trajectory to obtain a corrected tracking trajectory; In the training phase, the S-MOTR network and the TCNet network calculate their respective loss functions and perform their respective supervised training; The TCNet network obtains a prediction frame based on the corrected tracking trajectory and feeds it back to the S-MOTR network. The S-MOTR network calculates the loss of the tracking target frame based on the prediction frame.

2. The ship target tracking method based on low-resolution optical remote sensing images and trajectory adaptive correction according to claim 1 is characterized in that: The preprocessing specifically includes: Quantification is performed based on the low-resolution optical remote sensing image sequence data to obtain a visualized image; Perform cropping based on the visualized image to obtain an image block of a preset size; The image blocks of three consecutive frames are sequentially filled into three channels of the image so that the moving target forms a pseudo-color track in the image, thereby obtaining the superimposed image block.

3. The ship target tracking method based on low-resolution optical remote sensing images and trajectory adaptive correction according to claim 1 is characterized in that: The superimposed image block is input into the channel attention association module for processing to obtain an association feature map; The associated feature maps are input to the first encoder and the first decoder respectively; The first encoder performs feature extraction based on the associated feature map to obtain an input feature map; The first decoder performs processing based on the input feature map and the associated feature map, and interprets and obtains the initial tracking trajectory of each current target in the low-resolution optical remote sensing image sequence data.

4. The ship target tracking method based on low-resolution optical remote sensing images and trajectory adaptive correction according to claim 3 is characterized in that: The channel attention association module includes: a convolution layer, a global pooling layer and a first attention layer; The superimposed image blocks are sequentially input into the convolutional layer, the global pooling layer and the first attention layer to extract channel correlation features to obtain the correlation feature map.

5. The method for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction according to claim 1, characterized in that: The initial tracking trajectory is input to the position encoder for processing, and each trajectory point is position-encoded corresponding to the current frame target to obtain an encoded tracking trajectory; The encoded tracking trajectory is input to the second encoder for processing to obtain an encoded deep feature map; The encoded deep feature map is input to the second decoder for processing to obtain a smooth tracking trajectory; The smoothed tracking trajectory and the encoded tracking trajectory are input to the protection layer for processing to obtain the corrected tracking trajectory.

6. The method for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction according to claim 5, characterized in that: The second encoder includes: a first normalization layer, a second attention layer, a second normalization layer and a fully connected layer; The encoded tracking trajectory is sequentially input into the first normalization layer, the second attention layer, the second normalization layer and the fully connected layer for processing to obtain the encoded deep feature map.

7. The method for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction according to claim 5, characterized in that: The protective layer processing process is as follows: The smooth tracking trajectory and the encoded tracking trajectory are input to the protection layer; The protection layer obtains a first position code P1 of the current frame target based on the code tracking trajectory; The protection layer obtains a second position code P2 of the current frame target based on the smooth tracking trajectory; The protection layer replaces the second position code P2 in the smooth tracking trajectory with the first position code P1, and finally outputs the corrected tracking trajectory.

8. The method for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction according to claim 1, characterized in that: The correction trajectory loss function corresponding to the TCNet network is: L total =λ1L distance +λ2L dircction +λ3L vclocity Among them, L total represents the correction trajectory loss function, L distance represents the distance consistency loss function, L dircction represents the direction consistency loss function, L vclocity represents the speed consistency loss function, λ1, λ2 and λ3 represent the weights used to balance the influence of different loss functions on trajectory correction, X represents the image trajectory sequence, T represents the trajectory length, and t represents the current time step.

9. The method for tracking ship targets in low-resolution optical remote sensing images based on trajectory adaptive correction according to claim 1, characterized in that: The S-MOTR network calculates the loss of tracking the target frame based on the predicted frame, specifically including: The S-MOTR network obtains a feedback loss function based on the predicted frame, and calculates the loss of the tracking target frame based on the feedback loss function; The feedback loss function is specifically: Among them, α i represents the weight coefficient of the i-th prediction box, L bbox represents the bounding box loss, b i and They represent the i-th predicted box and the corresponding i-th true value box respectively.

10. A low-resolution optical remote sensing image ship target tracking system based on trajectory adaptive correction, characterized in that: include: Data acquisition processing module, model building module, trajectory acquisition module, trajectory correction module and training module; The data acquisition and processing module is used to acquire low-resolution optical remote sensing image sequence data to be detected and perform preprocessing to obtain superimposed image blocks; The model building module is used to build a target tracking model including an S-MOTR network and a TCNet network, and perform training to obtain a trained target tracking model; The S-MOTR network includes: a channel attention association module, a first encoder and a first decoder; the TCNet network includes: a position encoder, a second encoder, a second decoder and a protection layer; The trajectory acquisition module is used to input the superimposed image block into the trained target tracking model, and the S-MOTR network performs end-to-end target tracking processing on the superimposed image block to obtain an initial tracking trajectory; The trajectory correction module is used to perform adaptive correction processing on the initial tracking trajectory through the TCNet network to obtain a corrected tracking trajectory; The training module is used to calculate the respective loss functions based on the S-MOTR network and the TCNet network and perform respective supervised training during the training phase; the TCNet network obtains a prediction frame based on the corrected tracking trajectory and feeds it back to the S-MOTR network, and the S-MOTR network calculates the loss of the tracking target frame based on the prediction frame.

Images