A method and system for multi-target tracking of ships on the sea surface using remote sensing video.

By integrating the target association rules of Kalman filtering and kernel correlation filtering, the problem of ship target recognition in remote sensing images was solved, achieving high-accuracy multi-target tracking of ships on the sea surface and improving the robustness and occlusion re-association capability of targets in remote sensing videos.

CN117351046BActive Publication Date: 2026-11-10WUHAN UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311196667.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-15
Publication Date
2026-11-10
Estimated Expiration
2043-09-15

AI Technical Summary

Technical Problem

Existing remote sensing image multi-target tracking technologies suffer from problems such as small target pixel ratio, indistinct appearance features, difficulty in distinguishing similar targets, and video jitter in the identification of ships on the sea surface, which increases the difficulty of tracking.

Method used

The CenterNet target detection network is used in conjunction with the target association rules of Kalman filtering and kernel correlation filtering. The target is associated by fusing the Kalman filter and the kernel correlation filter. The kernel correlation filter is used to extract the target texture information and the position information before and after occlusion. The motion information of the Kalman filter is combined to perform target matching.

Benefits of technology

It improves the tracking accuracy and robustness of ship targets in remote sensing videos, effectively handles target occlusion and video jitter, and achieves high-accuracy multi-target tracking.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117351046B_ABST
    Figure CN117351046B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for multi-target tracking of ships on the sea surface using remote sensing video. First, publicly available remote sensing ship datasets are collected and input into a target detection network for training, resulting in a pre-trained model. Then, labeled remote sensing images are input into the pre-trained model for further training. Finally, unlabeled remote sensing images are input into the trained target detection network model, and the final tracking result is obtained by fusing Kalman filtering and kernel correlation filtering target association rules. This invention proposes a target association rule that fuses Kalman filtering and kernel correlation filtering. By introducing kernel correlation filtering into the multi-target tracking association method, the robustness of target tracking technology to video jitter in the field of remote sensing video tracking is improved. By comprehensively processing the target texture information extracted by the kernel correlation filter and the position information before and after occlusion, occlusion re-association of ship targets can be achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image recognition technology, specifically relating to a method and system for tracking multiple targets on the sea surface using remote sensing video. Background Technology

[0002] Real-time multi-target tracking technology for the sea surface enables automatic tracking and identification of maritime targets, which has significant practical implications. With the development of satellite technology, the resolution of satellite remote sensing imagery has greatly improved, and the sources of acquisition have become more diverse, leading to the rapid development of using remote sensing satellites for maritime vessel target identification. However, due to the large-scale nature of remote sensing images, targets occupy a small percentage of pixels. Furthermore, the appearance features of different types of targets are not obvious, easily leading to false detections; similar targets are highly similar, making them difficult to distinguish and causing target association errors. All of these factors increase the difficulty of target tracking in remote sensing images.

[0003] Multi-object tracking methods can currently be divided into two categories: two-stage and single-stage. Two-stage methods typically involve first extracting bounding boxes of the targets using an object detection network, and then post-processing the bounding boxes from two consecutive frames to correlate and match the target boxes; this method is also called detection-recordation. Two-stage methods emphasize target recordation techniques, such as SORT and deepSORT. SORT was the earliest proposed multi-object recordation technique. Its general process is as follows: first, a Kalman filter is initialized for each detected target; the Kalman filter predicts the target's bounding box in the next frame; then, the overlap between these predicted bounding boxes and the next frame's detected bounding boxes is calculated; and finally, the identity information is matched using the Hungarian algorithm. deepSORT improves upon SORT by training a neural network model to extract the target's appearance information, combining appearance and motion information to achieve ID matching, resulting in better performance than SORT. Single-stage methods, on the other hand, simultaneously implement target bounding box detection and target identity information extraction within a single network framework, achieving multi-object tracking in one step. Single-stage target tracking techniques generally improve the network framework based on target detection, simultaneously extracting target bounding boxes and target appearance information. The target identity ID is then matched by combining the bounding boxes and appearance information, such as CenterTrack and FairMOT. However, these single-stage and two-stage target tracking techniques are designed for natural scenes and struggle to handle issues such as weak texture information, target occlusion, and video jitter in remotely sensed images of ships on the sea surface. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for multi-target tracking of ships on the sea surface using remote sensing video, which can achieve high-accuracy target tracking of ships on the sea surface in complex scenarios using remote sensing images.

[0005] This invention provides a method for multi-target tracking of ships on the sea surface using remote sensing video, comprising the following steps:

[0006] Step 1: Collect publicly available remote sensing ship datasets and input the ship data into the target detection network CenterNet for training to obtain a pre-trained model.

[0007] Step 2: Label the remote sensing images, and then input the labeled remote sensing images into the pre-trained model obtained in Step 1 for training.

[0008] Step 3: Input the unlabeled remote sensing image into the target detection network model trained in Step 2, and obtain the final tracking result by fusing the target association rules of Kalman filtering and kernel correlation filtering;

[0009] Step 3.1: Divide the test remote sensing video into image frames and input them into the target detection network CenterNet trained in Step 2 to obtain the target detection box sequence;

[0010] Step 3.2: Based on the target association rule of fused Kalman filtering and kernel correlation filtering, target identity IDs are assigned to the target detection box sequence to obtain the final target tracking result.

[0011] Furthermore, in step 1, assuming the input remote sensing image size is (H, W), it first undergoes downsampling through two convolutional layers to obtain a feature map of (H / 4, W / 4). This feature map serves as the input to the skeleton DLAseg in the object detection network CenterNet. In DLAseg, the feature map undergoes continuous downsampling to extract more effective semantic information. Simultaneously, it undergoes upsampling through deformation convolution to perform nonlinear and hybrid feature fusion, thereby enhancing the feature representation of small targets and the judgment of target location. Finally, the output feature map size is the same as the input size. Then, the feature map output by DLAseg is fed into the three branches of object detection. The three branches of object detection are: heatmap branch, target box size branch, and target center offset branch, which are used to extract the target box center point, the length and width of the target box, and the positional deviation of the target center point, respectively. The heatmap branch can obtain the position of the target center point, and the target center offset branch can obtain the positional deviation of the target center point. The sum of the two can obtain a more accurate center point position. Combined with the length and width of the target box predicted by the target box size branch, the target box can be calculated.

[0012] Calculate the loss function L det Backpropagation updates the network weights, training M batches to obtain the final network weight model, with a loss function L. det The calculation method is as follows:

[0013] L det=L k +λ size L size +λ off L off (1)

[0014] in,

[0015]

[0016]

[0017]

[0018] In the formula, L k L represents the branching loss in the heatmap. size L represents the branch loss based on the bounding box size. off λ represents the target center offset branch loss. size , λ off These are the weighting coefficients, and α and β are two hyperparameters. It is the predicted value of the heatmap at point (x,y) in channel c, where Y is the predicted value. xyc This represents the actual heat map temperature at point (x, y) in channel c. Indicates the predicted target bounding box size, s k This represents the actual target bounding box size. This represents the predicted offset of the target center point, where p represents the coordinates of the target center point in the image, and R represents the scaling scale. The approximate integer coordinates of the center point after scaling are given, and N represents the number of target center points in the image.

[0019] Furthermore, in step 2, a bounding box is drawn in each frame of video image based on the target location, and a target identity ID is assigned to the same target in consecutive frames, generating a labeling file. The labeling file includes the position of the bounding box and the target ID. The labeled remote sensing image is then input into the pre-trained model obtained in step 1 for training, and the training process is the same as in step 1.

[0020] Furthermore, in step 3.2, it is assumed that the remote sensing image of the k-th frame is input into the target detection model trained in step 2 to obtain the target detection box D. k First, D k With Kalman filter T km The generated k-th frame remote sensing image is within the prediction box T of the current frame. k Correlation is performed using correlation rule 1, which refers to calculating the predicted bounding box T of the Kalman filter. k and object detection box D k The intersection-union ratio (CUNR) is used as the matching weight for D in the Hungarian algorithm. k and T kPerform matching. Initialize the Kalman filter that failed to match using association rule 1 as a kernel correlation filter. Initialize the kernel correlation filter KCF using the previous frame image and detection boxes recorded by the Kalman filter, and pass the target ID recorded by the Kalman filter to the KCF. Use association rule 2 to predict the target box position F of the current frame image ID using the KCF. k The target detection box D that did not match using association rule 1 k Correlation is performed; correlation rule 2 refers to calculating the prediction frame F of the kernel correlation filter. k and object detection box D k The intersection-union ratio (CUNR) is used as the matching weight for the Hungarian algorithm on F. k and D k Perform matching. Add kernel correlation filters that fail to match target detection boxes using association rule 2 to a blocking queue. Ship targets in this queue are considered occluded targets, and the target detection box D is then... k The input is fed into a blocked kernel correlation filter, and target re-identification is performed using correlation rule 3. Correlation rule 3 refers to using the kernel correlation filter to detect the confidence of the target in the current detection box. If the confidence is higher than a set threshold, the target's motion trajectory and the position of the target detection box are combined for filtering to find the optimal match. For target detection boxes that still cannot be associated after processing by the three correlation rules, a new detection sequence number is assigned to them, and the target is considered to be the most recently appeared target. The corresponding Kalman filter is then initialized and new identity information is assigned. After the above operations, the final target tracking sequence can be output to complete the multi-target tracking task of sea surface ships.

[0021] The present invention also provides a remote sensing video multi-target tracking system for ships on the sea surface, for implementing the remote sensing video multi-target tracking method for ships on the sea surface as described above.

[0022] Furthermore, it includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute a remote sensing video multi-target tracking method for ships on the sea surface as described above.

[0023] Alternatively, it may include a readable storage medium on which a computer program is stored, which, when executed, implements a remote sensing video method for multi-target tracking of ships on the sea surface as described above.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] This invention proposes a target association rule that integrates Kalman filtering and kernel correlation filtering. The KFC filter can build a more refined model of the target's texture features, which can alleviate the problem of missed target detection. By introducing the kernel correlation filter into the multi-target tracking association method, the robustness of target tracking technology to video jitter in the field of remote sensing video tracking is improved. By comprehensively processing the target texture information extracted by the kernel correlation filter and the position information before and after occlusion, the occlusion re-association of ship targets can be realized. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the remote sensing video multi-target tracking method for ships on the sea surface according to the present invention.

[0028] Figure 2 This is a schematic diagram of the skeleton DLAseg in the CenterNet target detection network of this invention.

[0029] Figure 3 This is a schematic diagram of the three detection task heads of the CenterNet target detection network of this invention.

[0030] Figure 4 This is the execution diagram of the target correlation rule proposed in this invention, which combines Kalman filtering and kernel correlation filtering.

[0031] Figure 5 This is the result of remote sensing video target tracking in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0033] This invention provides a method and system for tracking multiple targets on the sea surface using remote sensing video. The technical solution of this invention will be further described below with reference to the accompanying drawings.

[0034] Example 1

[0035] like Figure 1As shown in the figure, this invention provides a method for multi-target tracking of ships on the sea surface using remote sensing video, which includes the following steps:

[0036] Step 1: Collect publicly available remote sensing ship datasets, input the ship data into the target detection network CenterNet for training, and obtain a pre-trained model.

[0037] This embodiment uses the HRSC2016 remote sensing ship dataset and the Airbus ship dataset. The ship data is input into the CenterNet object detection network for training, resulting in a pre-trained model. Assuming the input remote sensing image size is (H, W), it is first downsampled through two convolutional layers to obtain a (H / 4, W / 4) feature map. This feature map serves as the input to the skeleton DLAseg in the CenterNet object detection network. Figure 2 As shown, the feature map undergoes continuous downsampling in DLAseg. Each downsampling process extracts more effective semantic information. Simultaneously, it undergoes upsampling via deformation convolution for non-linear, hybrid feature fusion. This enhances the feature representation of small targets and improves the ability to determine target location. The final output feature map has the same size as the input. Then, the feature map output from the network skeleton DLAseg is fed into the three branches of object detection, as follows... Figure 3 As shown, the object detection network has three branches: a heatmap branch, a bounding box size branch, and an object center offset branch. These branches are used to extract the object center point, the length and width of the object box, and the positional deviation of the object center point, respectively. Each task branch contains a 256-channel 3×3 convolutional layer and a 1×1 convolutional layer for dimensionality reduction. The heatmap branch obtains the position of the object center point, and the object center offset branch obtains the positional deviation of the object center point. Adding the two together yields a more accurate center point position. Combined with the length and width of the object box predicted by the bounding box size branch, the object box can be calculated.

[0038] Calculate the loss function L det Backpropagation is used to update the network weights, and this process is repeated for 30 batches to obtain the final network weight model. The loss function L... det The calculation method is as follows:

[0039] L det =L k +λ size L size +λ off L off (1)

[0040] in,

[0041]

[0042]

[0043]

[0044] In the formula, L k L represents the branching loss in the heatmap. size L represents the branch loss based on the bounding box size. off λ represents the target center offset branch loss. size , λ off λ is the weighting coefficient, and in this embodiment it is... size Set to 0.1, λ off Set to 1, α and β are two hyperparameters. It is the predicted value of the heatmap at point (x,y) in channel c, where Y is the predicted value. xyc This represents the actual heat map temperature at point (x, y) in channel c. Indicates the predicted target bounding box size, s k This represents the actual target bounding box size. This represents the predicted offset of the target center point, where p represents the coordinates of the target center point in the image, and R represents the scaling scale. The approximate integer coordinates of the center point after scaling are given, and N represents the number of target center points in the image.

[0045] Step 2: Label the remote sensing images, and then input the labeled remote sensing images into the pre-trained model obtained in Step 1 for training.

[0046] In each frame of video imagery, a bounding box is drawn based on the target location. A target identity ID is assigned to the same target in consecutive frames, generating a labeling file that includes the location of the bounding box and the target ID. The labeled remote sensing imagery is then input into the pre-trained model obtained in step 1 for training, following the same training process as in step 1.

[0047] Step 3: Input the unlabeled remote sensing image into the target detection network model trained in Step 2, and obtain the final tracking result by fusing the target association rules of Kalman filtering and kernel correlation filtering.

[0048] Step 3.1: Divide the test remote sensing video into image frames and input them into the target detection network CenterNet trained in Step 2 to obtain a sequence of target detection boxes.

[0049] Step 3.2: Based on the target association rule of fusion Kalman filtering and kernel correlation filtering, target identity IDs are assigned to the target detection box sequence to obtain the final target tracking result.

[0050] The target association rule, which combines Kalman filtering and kernel correlation filtering, is used to perform identity matching on the input detection box sequence. The specific workflow is as follows: Figure 4 As shown, the k-th frame of remote sensing image is first input into the target detection network to obtain the target detection box D. k Then, a Kalman filter is used to generate the predicted bounding box T of the current remote sensing image. k Using association rule 1 on D k and T k A matching process is performed. If a match is successful, target association is achieved; otherwise, the Kalman filter is transformed into a kernel correlation filter, and the kernel correlation filter is used to predict the target bounding box position F of the current frame image ID. k Apply association rule 2 to D k and F k Perform a match; if the match is successful, the target is associated; otherwise, the target detection box D is removed. k The input is fed into a blocked kernel correlation filter, and target re-identification is performed using association rule 3. If the target detection box still fails to be associated after processing by the three association rules, a new detection sequence number is assigned to it, assuming that the target is the most recently appeared target, and the corresponding Kalman filter is initialized and new identity information is assigned.

[0051] Assuming the remote sensing image of frame k is input into the target detection model, the target detection bounding box D is obtained. k First, D k With Kalman filter T km The generated k-th frame remote sensing image is within the prediction box T of the current frame. k Using correlation rule 1 (to calculate the prediction box T of the Kalman filter) k and object detection box D k The intersection-union ratio (CUNR) is used as the matching weight for D in the Hungarian algorithm. k and T k (Matching) and associating.

[0052] In remote sensing videos, the motion of the remote sensing satellites inevitably introduces video jitter, posing a challenge to motion-based matching. Therefore, target detection boxes that fail to match using Association Rule 1 may be due to video jitter. The Kalman filter that fails to match using Association Rule 1 is transformed into a kernel correlation filter. This involves initializing the kernel correlation filter (KCF) using the previous frame image and detection boxes recorded by the Kalman filter, and passing the target IDs recorded by the Kalman filter to the KCF. Then, Association Rule 2 is used to calculate the predicted box F of the kernel correlation filter. k and object detection box D k The intersection-union ratio (CUNR) is used as the matching weight for the Hungarian algorithm on F. k and D k (Perform matching) Predict the bounding box position F of the current frame image ID using KCF. kThe target detection box D that did not match using association rule 1 k Establish a connection.

[0053] Remote sensing videos may contain obstructions such as clouds, fog, and bridges, which pose a challenge to target tracking. Kernel correlation filters that fail to match target detection boxes using association rule 2 are added to a blocking queue; ship targets in this queue are considered occluded. As the target becomes increasingly obstructed, the confidence level of the kernel correlation filter gradually decreases below a threshold, and the kernel correlation filter is sensitive to the size of the target box. The target detection box D... k The input is fed into a blocked kernel correlation filter, and target re-identification is performed using correlation rule 3 (which uses the kernel correlation filter to detect the confidence of the target in the current detection box; if the confidence is higher than a set threshold, it combines the target's motion trajectory and the position of the target detection box for filtering to find the optimal match). Finally, for target detection boxes that still cannot be associated after processing by the three correlation rules, a new detection sequence number is assigned to them, assuming that the target is the most recently appeared target, and the corresponding Kalman filter is initialized and assigned new identity information. After the above operations, the final target tracking sequence can be output, completing the multi-target tracking task for sea surface ships.

[0054] Figure 5 The target tracking results of the proposed method are shown. At frame t, a total of 8 targets were detected. At frame t+n, a new target number 9 appeared. It can be seen that target number 4 was able to be re-matched after crossing a bridge, which verifies that the algorithm proposed in this invention successfully achieves target matching under occlusion.

[0055] Example 2

[0056] Based on the same inventive concept, the present invention also provides a remote sensing video multi-target tracking system for ships on the sea surface, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the remote sensing video multi-target tracking method for ships on the sea surface as described above.

[0057] Example 3

[0058] Based on the same inventive concept, the present invention also provides a remote sensing video multi-target tracking system for ships on the sea surface, including a readable storage medium on which a computer program is stored. When the computer program is executed, it implements the remote sensing video multi-target tracking method for ships on the sea surface as described above.

[0059] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0060] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for multi-target tracking of ships on the sea surface using remote sensing video, characterized in that, Includes the following steps: Step 1: Collect publicly available remote sensing ship datasets and input the ship data into the target detection network CenterNet for training to obtain a pre-trained model. Step 2: Label the remote sensing images, and then input the labeled remote sensing images into the pre-trained model obtained in Step 1 for training. Step 3: Input the unlabeled remote sensing image into the target detection network CenterNet trained in Step 2, and obtain the final tracking result by fusing the target association rules of Kalman filtering and kernel correlation filtering; Step 3.1: Divide the test remote sensing video into image frames and input them into the target detection network CenterNet trained in Step 2 to obtain the target detection box sequence; Step 3.2: Based on the target association rule of fusion Kalman filtering and kernel correlation filtering, the target detection box sequence is assigned target identity IDs to obtain the final target tracking result; The target association rule combining Kalman filtering and kernel correlation filtering is as follows: First, the remote sensing image is input into the target detection network CenterNet to obtain target detection boxes. Then, the Kalman filter is used to generate a predicted box for the current remote sensing image. Association rule 1 is used to match the target detection box and the predicted box generated by the Kalman filter. If the match is successful, target association is achieved. Otherwise, the Kalman filter is initialized as a kernel correlation filter, that is, the kernel correlation filter KCF is initialized using the previous frame image and detection boxes recorded by the Kalman filter, and the target identity ID recorded by the Kalman filter is passed to the KCF. The kernel correlation filter is used to predict the position of the target box with the ID in the current frame image. Association rule 2 is used to match the target detection box and the predicted box generated by the kernel correlation filter. If the match is successful, target association is achieved. Otherwise, the target detection box is input into the blocked kernel correlation filter, that is, kernel correlation filters that failed to match the target detection box using association rule 2 are added to the blocking queue. Ship targets in this queue are considered occluded targets. The target detection boxes are input into a blocked kernel correlation filter, and target re-identification is performed using association rule 3. If the target detection box still fails to be associated after processing by the three association rules, a new detection sequence number is assigned to it, assuming that the target is the most recently appeared target, and the corresponding Kalman filter is initialized and a new identity ID is assigned. Among them, association rule 1 refers to calculating the intersection-union ratio (IU) of the predicted box of the Kalman filter and the target detection box, and using the IU as the matching weight of the Hungarian algorithm to match the target detection box and the predicted box; association rule 2 refers to calculating the IU of the predicted box of the kernel correlation filter and the target detection box, and using the IU as the matching weight of the Hungarian algorithm to match the predicted box and the target detection box; association rule 3 refers to using the kernel correlation filter to detect the confidence of the target in the current detection box. If the confidence is higher than a set threshold, the target's motion trajectory and the position of the target detection box are combined for filtering to find the optimal match.

2. The method for multi-target tracking of ships on the sea surface using remote sensing video as described in claim 1, characterized in that: The target detection network CenterNet takes a remote sensing image of size (H, W) as input. First, it undergoes downsampling through two convolutional layers to obtain a feature map of size (H / 4, W / 4). This feature map serves as the input to the skeleton DLAseg in CenterNet. Within DLAseg, the feature map undergoes continuous downsampling to extract more effective semantic information. Simultaneously, it undergoes upsampling through deformation convolutions for non-linear, hybrid feature fusion, enhancing the feature representation of small targets and improving the ability to determine target location. The final output feature map has the same size as the input. The feature map output by DLAseg is then fed into three branches of the target detection network CenterNet: a heatmap branch, a target bounding box size branch, and a target center offset branch. These branches are used to extract the target bounding box center point, the length and width of the target bounding box, and the positional deviation of the target center point, respectively. The heatmap branch obtains the target center point's position, and the target center offset branch obtains the target center point's positional deviation. Adding these two together yields a more accurate center point position. Combined with the target bounding box size branch's predicted length and width, the target bounding box can be calculated.

3. The method for multi-target tracking of ships on the sea surface using remote sensing video as described in claim 2, characterized in that: Calculate the loss function in step 1. Backpropagation updates the network weights, training M batches to obtain the final network weight model, and the loss function is... The calculation method is as follows: (1) in, (2) (3) (4) In the formula, This indicates the branching loss in the heatmap. Indicates the branch loss based on the target bounding box size. This represents the target center offset branch loss. , These are the weighting coefficients. and There are two hyperparameters. The predicted heatmap is in channel c. Predicted value at point This indicates the true heatmap in channel c. The heat value at point Indicates the predicted target bounding box size. This represents the actual target bounding box size. This represents the offset value of the predicted target center point. p Indicates the coordinates of the center point of the target in the image. R Indicates the scaling factor. The approximate integer coordinates of the center point after scaling are given, and N represents the number of target center points in the image.

4. The method for multi-target tracking of ships on the sea surface using remote sensing video as described in claim 3, characterized in that: In step 2, a bounding box is drawn in each frame of video image based on the target location. A target identity ID is assigned to the same target in consecutive frames, and a label file is generated. The label file includes the position of the bounding box and the target identity ID. The remote sensing image with the label is input into the pre-trained model obtained in step 1 for training. The training process is the same as in step 1.

5. A remote sensing video multi-target tracking system for ships on the sea surface, characterized in that, It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the remote sensing video multi-target tracking method for ships on the sea surface as described in any one of claims 1-4.

6. A remote sensing video multi-target tracking system for ships on the sea surface, characterized in that, It includes a readable storage medium on which a computer program is stored, and when the computer program is executed, it implements a remote sensing video method for multi-target tracking of ships on the sea surface as described in any one of claims 1-4.

Citation Information

Patent Citations

  • Multi-target tracking method and device and electronic equipment

    CN113344975A

  • Large-scale remote sensing image multi-target tracking method

    CN114066937A