A water-air cross-medium target positioning and tracking method

Abstract

This invention provides a method for locating and tracking targets across water and air media, comprising: obtaining a confidence score of the current frame observation image using a reference template; the confidence score being used to characterize the reliability of the target observation in the current frame; comparing the confidence score with a preset confidence threshold; determining the target bounding box of the current frame based on the comparison result; and adaptively updating the reference template for tracking processing of the next frame observation image. This invention improves the tracking robustness of targets under extreme cross-media conditions.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing technology, and in particular to a method for locating and tracking targets across water and air media. Background Technology

[0002] With the development of unmanned equipment technology, cross-medium watercraft have attracted much attention due to their combined capabilities of high-speed flight in the air and stealthy navigation underwater, and have become an important research direction in the field of autonomous unmanned systems. Target tracking is an indispensable core technology when performing autonomous recovery and docking, formation coordination, or continuous surveillance of enemy cross-medium targets.

[0003] Traditional detection and tracking systems typically employ radar (microwaves) to detect aerial targets and sonar (sound waves) to detect underwater targets. However, when a target is in a cross-medium motion (emerging from or entering the water), traditional sensors face insurmountable physical limitations. Electromagnetic waves attenuate extremely rapidly in water, failing to effectively penetrate the surface; while sound waves undergo near total reflection when crossing the water-air interface due to the significant difference in acoustic impedance, rendering sonar ineffective for aerial targets. This difference in medium propagation characteristics results in a significant detection blind zone for traditional single sensors near the water-air interface. Furthermore, GPS and other satellite positioning signals become completely ineffective underwater. Simultaneously, active detection devices such as radar and sonar are often bulky, power-hungry, and easily reveal their location, hindering covert operations or deployment on lightweight unmanned platforms. In contrast, utilizing visual sensors for cross-medium target tracking demonstrates irreplaceable convenience and necessity.

[0004] Visible light not only has excellent propagation characteristics in air but also good penetrating power in water (especially shallow water areas). This makes visual sensors one of the very few detection methods currently available capable of continuously acquiring information throughout the entire process of a target entering and exiting water. As a passive detection device, it only receives ambient light and does not actively emit signals, possessing extremely high stealth capabilities (strong anti-reconnaissance ability). Simultaneously, its small size, light weight, and low power consumption make it extremely suitable for amphibious unmanned platforms with limited space and payload. Moreover, visual images can provide rich information about the target's color, texture, shape, and local features, which is crucial for accurate target identification, anti-jamming identification, and refined attitude estimation.

[0005] Existing visual target tracking algorithms, such as KCF based on correlation filtering, SiamRPN based on Siamese networks, and depth tracking methods based on detection-tracking fusion, have achieved good results in single-medium environments. However, when targets move across water and air, existing methods still face significant challenges. Air and water have different refractive indices, and when a target crosses the water surface, the target in the image may exhibit geometric distortion, abrupt scale changes, and nonlinear deformation, making it difficult to stably match pre-trained visual features or online templates. The moment a target enters or exits the water is usually accompanied by splashing water, bubble generation, water ripple reflection, partial occlusion, abrupt changes in illumination, and contrast variations; these degradation factors significantly reduce the reliability of visual observations.

[0006] Furthermore, existing online tracking algorithms typically update the target template using a fixed learning rate or moving average. When splashes, bubbles, reflections, and severe distortions instantaneously across media are continuously written into the target template, template contamination and apparent drift can easily occur, causing the tracker to gradually deviate from the true target, ultimately resulting in target loss or re-acquisition failure. Therefore, how to avoid abnormal observations contaminating the template within the low-confidence visual observation range across media, and how to maintain target position prediction using motion continuity, are urgent problems to be solved in water-air cross-media target tracking. Summary of the Invention

[0007] This invention provides a water-air cross-medium target localization and tracking method, which solves the problems of target loss and template contamination caused by water splashes and refraction distortion when the target enters or exits the water during single-vision target tracking, and improves the tracking robustness of the model under extreme cross-medium conditions.

[0008] A method for locating and tracking a target across water and air media includes: Confidence score of the observed image in the current frame is obtained using a benchmark template. The confidence score is used to characterize the reliability of target observations in the current frame. The confidence score is compared with a preset confidence threshold. Based on the comparison result, the target bounding box of the current frame is determined, and the baseline template is adaptively updated for tracking processing of the next frame of the observed image. Among them, the confidence score ≥Confidence upper limit At the same time, the output target bounding box is used as the tracking result of the current frame; and the benchmark template is continuously used as the basis for calculating the confidence score of the observed image in the next frame. The confidence score satisfy: ≤ First confidence level lower limit At the same time, the target bounding box predicted by the Kalman filter is used as the tracking result for the current frame; simultaneously, a cross-medium transition template is constructed. And utilize this trans-medium transition template This serves as the basis for calculating the confidence score of the next frame of the observed image, and the confidence scores of the subsequent several frames of the observed image satisfy the following: First confidence lower limit. < ≤Second confidence level lower limit At that time, with the first update rate The cross-medium transition template is updated, and the updated cross-medium transition template is used as the basis for calculating the confidence score of subsequent observed images; The confidence score satisfy: < ≤ At the same time, the corrected target bounding box is used as the tracking result of the current frame observation image; simultaneously, at the second update rate... The current cross-medium transition template is updated, and the updated cross-medium transition template is used as the basis for calculating the confidence score of subsequent observed images; the corrected target bounding box is obtained by smoothing the target bounding box output by the visual tracking model and the target bounding box predicted by the Mann filter; The confidence score After experiencing confidence level fluctuations and instability, continuous Frame, confidence score satisfy: Then, the target bounding box output by the visual tracking model is used as the tracking result for the current frame; at the same time, a third update rate is used. The current cross-medium transition template is updated, and the updated cross-medium transition template is used as the basis for calculating the confidence score of subsequent observed images. in, The higher the update rate, the faster the template updates.

[0009] Furthermore, in the water-air cross-medium target localization and tracking method described above, the acquisition of the reference template includes: Acquire stable observation images of the target in a single medium and under stable observation conditions; The target bounding box of the stable observation image is determined by manual annotation or detection module, and the target bounding box of the stable observation image is cropped, scaled, and feature extracted to finally obtain the benchmark template.

[0010] Furthermore, in the water-air cross-medium target localization and tracking method described above, the constructed cross-medium transition template is as follows:

[0011] in, As a baseline template, For critically reliable frame templates, for Kalman first frame guide template; , , To merge weights, and satisfy .

[0012] Furthermore, in the water-air cross-medium target localization and tracking method described above, the critical reliable frame template The acquisition includes: When the model calculates the confidence score of the currently observed image ≤ First confidence level lower limit At that time, extract the historical confidence scores of the three most recent historical observation images. And from the historical confidence score of these 3 The target bounding box with the highest confidence score is selected from the samples, and its features are extracted to obtain the critically reliable frame template. ; The Kalman first frame guiding template The acquisition includes: When the model calculates the confidence score of the currently observed image ≤ First confidence level lower limit At that time, the target bounding box of the next frame of the observed image predicted by the Kalman filter is obtained, and its features are extracted.

[0013] Furthermore, in the water-air cross-medium target localization and tracking method described above, the target feature template is as follows: .

[0014] Furthermore, in the water-air cross-medium target localization and tracking method described above, the first update rate... Updating the cross-media transition template includes: Every P frames, the critical reliable frame template and the Kalman first frame guiding template are recalculated, and the cross-media transition template is updated with the recalculated critical reliable frame template and the Kalman first frame guiding template. The second update rate Updating the cross-media transition template includes: Every Q frames, the critical reliable frame template and the Kalman first frame guiding template are recalculated, and the cross-media transition template is updated with the recalculated critical reliable frame template and the Kalman first frame guiding template. The third update rate Updating the cross-media transition template includes: Every R frames, the critical reliable frame template and the Kalman first frame guiding template are recalculated, and the cross-media transition template is updated with the recalculated critical reliable frame template and the Kalman first frame guiding template, where R < Q < P.

[0015] Furthermore, in the water-air cross-medium target localization and tracking method described above, the confidence score of the current frame observation image is obtained using a reference template. include: Using the benchmark template, a matching search is performed in the current frame observation image to find the search region that best matches the benchmark template; Extract the features of the search region to obtain the search region features; The search region features are matched with the baseline template to generate a response map; The confidence score of the observed image in the current frame is determined based on the response map. .

[0016] Beneficial effects: The method provided by this invention, on the one hand, determines the target position corresponding to the current frame observation image by comparing the confidence score with the confidence threshold, and adaptively configures the template update strategy to update the feature template at different update rates for tracking processing of the next frame observation image. This solves the problem in existing technologies that use fixed learning rates or moving average updates, where abnormal observations such as splashes, bubbles, and distortions at the moment of crossing the medium are continuously written into the feature template, leading to template contamination and the tracker gradually deviating from the true target, thereby improving the accuracy of target tracking. On the other hand, by constructing a cross-medium transition template, the traditional fixed update or simple freezing strategy is abandoned in the low-confidence blind zone where the target crosses the interface, thus ensuring tracking robustness under extreme conditions. Moreover, this invention improves the spatiotemporal continuity under occlusion and distortion conditions by using the target bounding box predicted by the Kalman filter as the tracking result of the current frame when the confidence is less than the lower confidence limit, and reduces tracking loss caused by short-term occlusion or severe water surface distortion.

[0017] Furthermore, the cross-medium transition template constructed in this invention comprises a baseline template that guarantees prior target identity, a critically reliable frame template that retains the latest attitude scale information before crossing the medium, and a Kalman first-frame guiding template that introduces instantaneous apparent changes based on physical inertia. By constructing the cross-medium transition template using the baseline template, the critically reliable frame template, and the Kalman first-frame guiding template, the characteristic noise of the machine during the cross-medium phase can be effectively preserved. Utilizing this characteristic noise and the initial template, relatively stable feature anchor points are provided in the low-confidence interval, ensuring tracking robustness under extreme conditions. In a set of exemplary experiments, the average tracking confidence in the cross-medium interval increased from approximately 30%-40% to approximately 90%.

[0018] In addition, the method provided by the present invention updates the template according to a set adaptive update rate under other comparison results. After confirming that the target has completely entered the new physical medium, a high update rate is adopted, which enables the system to quickly extract and adapt to the optical features of the new medium, effectively improving the stability of long-term cross-medium tracking. Attached Figure Description

[0019] 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.

[0020] Figure 1 This is one of the schematic diagrams of the water-air cross-medium target positioning and tracking method provided by the present invention; Figure 2 This is the second schematic diagram of the water-air cross-medium target positioning and tracking method provided by the present invention; Figure 3 A schematic diagram of four tracking states for a single vehicle's cross-medium navigation process provided in this application. Detailed Implementation

[0021] 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.

[0022] Considering the physical inertia and motion continuity of a vehicle during cross-medium transitions, its trajectory typically follows specific kinematic laws within a short period. Therefore, motion state estimation methods such as Kalman filtering are introduced to provide position prediction based on prior physical laws within blind zones where visual features fail. However, simply using Kalman filtering to predict target position is insufficient to address template contamination. To this end, this invention combines visual confidence assessment, Kalman motion prediction, and an adaptive template update strategy to construct a robust target tracking mechanism for cross-medium water and air scenarios. It identifies low-confidence observation intervals across media based on visual response confidence and uses Kalman prediction to constrain the search area within these intervals. Simultaneously, it constructs a cross-medium transition template using a baseline template, a critically reliable template, and a first-guided Kalman template, thereby preventing abnormal observations from contaminating the online template.

[0023] like Figure 1 , Figure 2 As shown, the water-air cross-medium target localization and tracking method provided in this application includes the following steps: Step 101: Obtain the confidence score of the observed image in the current frame using the benchmark template. The confidence score is used to characterize the reliability of the target observation in the current frame.

[0024] Specifically, the system is first initialized to acquire the sequence of observation images corresponding to the observation video, and then a stable observation image is acquired in which the target is in a single medium and in a stable observation state. Then, the target bounding box of the stable observation image is determined by manual annotation or detection module, and the target bounding box of the stable observation image is cropped, scaled, and feature extracted to obtain a benchmark template. Through this benchmark template, the position information of the target in the current observation image can be accurately detected.

[0025] Furthermore, this application uses the following method to determine the confidence score of the observed image in the current frame. : Using a benchmark template, a matching search is performed in the current frame observation image to find the search region that best matches the benchmark template; features of the search region are extracted to obtain search region features; the search region features are matched with the benchmark template to generate a response map; and the confidence score of the current frame observation image is determined based on the response map. .

[0026] Step 102: Compare the confidence score with a preset confidence threshold, determine the target bounding box of the current frame based on the comparison result, and adaptively update the baseline template for tracking processing of the next frame of observed image.

[0027] The following details how to determine the target bounding box of the current frame based on the comparison results and adaptively update the baseline template, specifically including: Specifically, Figure 3 For the four tracking states of a single vehicle's cross-medium navigation process provided in this application, please refer to [link / reference]. Figure 3 Based on the comparison results, this application corresponds to four tracking states for a single vehicle's cross-medium navigation process. These four states are as follows: State A is the high-confidence tracking state. At this time, the vehicle is located in a single medium, the target appearance is stable, and visual observation is reliable. The confidence level output by the visual tracking model at this time satisfies: ≥ At this point, the target bounding box output by the visual tracking model is used as the tracking result of the current frame; at the same time, the baseline template is continuously used as the basis for calculating the confidence score of the observed image in the next frame.

[0028] State B is a low-confidence blind zone state. At this time, the vehicle crosses the water-air interface, and the aircraft is affected by water splashes, bubbles, refractive distortion, and partial occlusion. The confidence level output by the visual tracking model at this point is... ≤ At this point, the target bounding box predicted by the Kalman filter is used as the tracking result for the current frame, while the baseline template T0 and the critically reliable template are used as the basis. and Kalman First Guiding Template Constructing a trans-medium transition template And utilize this trans-medium transition template As the basis for calculating the confidence score of the next frame of the observed image, the next frame of the observed image uses a cross-medium transition template. To calculate its confidence score.

[0029] The cross-media transition template constructed in this application is as follows:

[0030] in, As a baseline template, For critically reliable frame templates, for Kalman first frame guide template; , , To merge weights, and satisfy .

[0031] The critically reliable frame template The acquisition includes: When the model calculates the confidence score of the currently observed image ≤ First confidence level lower limit At that time, extract the confidence scores of the three most recent historical observation images. And from the historical confidence score of these 3 The target bounding box with the highest confidence score is selected from the samples; the features of the target bounding box are extracted to obtain the critically reliable frame template. .

[0032] The Kalman first frame guide template The acquisition includes: When the model calculates the confidence score of the currently observed image ≤ First confidence level lower limit At that time, the target bounding box predicted by the Kalman filter for the next frame of the observed image is obtained, and its features are extracted.

[0033] Preferably, the cross-medium transition template is as follows:

[0034] Because this cross-medium transition template is used as the basis for calculating the confidence score of the next frame of observed images, the confidence score of the next frame of observed images will be improved, and the confidence scores of subsequent frames of observed images will also be improved. satisfy < ≤ At that time, this application needs to be updated at the first update rate. The cross-medium transition template is updated and used as the basis for calculating the confidence score of subsequent observed images, so as to continuously adapt to changes in the current state and thus more accurately determine the confidence of the current observed image.

[0035] In this embodiment of the application, the first update rate Updating the cross-media transition template includes: recalculating the critical reliable frame template and the Kalman first frame guiding template every P frames, and updating the cross-media transition template with the recalculated critical reliable frame template and Kalman first frame guiding template.

[0036] State C is the recovery transition state. At this point, the vehicle has just completed its cross-medium crossing and is flying close to the water surface. The interference from water splashes and bubbles gradually weakens, and the target's appearance gradually returns to stability. At this time, the confidence level output by the visual tracking model satisfies: < ≤ .

[0037] In this state, the target bounding box output by the visual tracking model and the target bounding box predicted by the Mann filter are used for smoothing correction to obtain the corrected target bounding box, which is then used as the tracking result for the current frame's observed image. Simultaneously, a second update rate is applied. The cross-medium transition template is updated, and the updated cross-medium transition template is used as the basis for calculating the confidence score of the next observation image.

[0038] Since the confidence score increases as the target is in state C, the update rate needs to be increased. Therefore, the second update rate... Greater than the first update rate .

[0039] In this embodiment, the second update rate Updating the cross-media transition template includes: Every Q frames, the critically reliable frame template and the Kalman first-frame guiding template are recalculated, and the cross-medium transition template is updated using the recalculated critically reliable frame template and the Kalman first-frame guiding template. Where Q < P.

[0040] State D represents the cross-medium completion state. At this point, the spacecraft has entered a new physical medium and resumed stable observation. The confidence level output by the visual tracking model, after experiencing the unstable confidence levels in states B and C, now... Satisfy: N consecutive frames, < ≤ At this point, the system determines that the cross-media handover is complete, and uses the target bounding box output by the visual tracking model as the tracking result for the current frame, while employing a higher third update rate. The cross-medium transition template is updated quickly, and the updated cross-medium transition template is used as the basis for calculating the confidence score of the next observation image. This allows the updated cross-medium transition template to quickly absorb the color, brightness, texture, scale, and edge features in the new medium, thereby quickly adapting to the appearance features of the target in the new medium. .

[0041] In this embodiment, a higher third update rate Rapid updates to cross-media transition templates include: Every R frames, the critical reliable frame template and the Kalman first frame guiding template are recalculated, and the cross-media transition template is updated with the recalculated critical reliable frame template and the Kalman first frame guiding template, where R < Q < P.

[0042] The following example further details the proposed solution for this case: Assuming that in this embodiment, =80, =40, =60. The visual tracking model acquired a video, which includes several frames of observation images. The following describes the tracking results for each frame and the refined template strategy: Frame 1: If the target is underwater and in a stable operating state, the target bounding box of the first frame of the observed image is manually detected, and the target bounding box is cropped, scaled, and its features are extracted to obtain the baseline template.

[0043] Frame 2: The visual tracking model uses the baseline template to detect the target bounding box 2 in the second frame image. The model also provides a confidence score of 2 for the target bounding box 2, which is 95. Frame 3: The model uses the baseline template to detect in the third frame image and obtains the target bounding box 3. At the same time, the model gives the target bounding box 3 with a confidence score of 3, which is 95. Frame 4: The model uses the baseline template to detect in the 4th frame image and obtains the initial target bounding box 4. At the same time, the model gives the initial target bounding box 4 with a confidence score of 4, which is 39. Therefore, the final target bounding box of the 4th frame image is the target bounding box obtained by Kalman prediction, instead of the initial target bounding box 4.

[0044] Frame 5: The model uses a cross-media transition template to detect the target bounding box 5 in the 5th frame image. The model also gives a confidence score of 5 for the target bounding box 5, which is 42. Similarly, the target bounding box obtained by Kalman prediction is used as the target bounding box of the current observation image.

[0045] Frame 6: The model uses a cross-media transition template to detect the target bounding box 6 in the 6th frame image. The model also provides a confidence score of 6 for the target bounding box 6, which is 45. The target bounding box obtained by Kalman prediction is used as the target bounding box of the current observation image.

[0046] Frame 7: The model uses a cross-media transition template to detect the target bounding box 7 in the 7th frame image. The model also provides a confidence score of 7 for the target bounding box 7, with a score of 50. The target bounding box obtained from the Kalman prediction is used as the target bounding box for the current observed image, and the model updates at the first rate. Update the cross-media transition template to obtain the updated cross-media transition template 1.

[0047] … Frame 20: The model uses the cross-media transition template 1 to detect the target bounding box 20 in the 20th frame image, and simultaneously provides a confidence score of 20 for the target bounding box 20, with a score of 70. At this point, the target bounding box 20 is replaced with the target bounding box. ,Target bounding box =The target bounding box predicted by Kalman Observation + the target bounding box predicted by the model are obtained by smoothing both; and updated at the second rate. The updated model uses cross-media transition template 1 to obtain the updated cross-media transition template 2.

[0048] Frame 21: The model uses the cross-media transition template 2 to detect the target bounding box 21 in the 21st frame image, and at the same time, the model gives the target bounding box 21 confidence score 21, which is 75. At this time, the target bounding box 21 is obtained by smoothing the Kalman predicted target bounding box and the model predicted target bounding box.

[0049] Frame 22: The model uses the cross-media transition template 2 to detect the target bounding box 22 in the 22nd frame image, obtaining the target bounding box 22. Simultaneously, the model provides a confidence score 22 for the target bounding box 22, with a score of 78. At this point, the target bounding box 22 is obtained by smoothly adjusting the Kalman predicted target bounding box and the model predicted target bounding box. The model then continues to update at the second rate. Update the cross-media transition template.

[0050] … Frame 41: The target bounding box 41 is detected in the 41st frame image using the model's currently updated cross-media transition template. At the same time, the model provides a confidence score of 41 for the target bounding box 41, with a score of 82. At this point, the target bounding box 41 is the target bounding box predicted by the model.

[0051] ...the confidence level did not decrease during this period. Under these circumstances, the default target has now returned to a stable state, during which time... To update the cross-media transition template.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for locating and tracking a water-air cross-medium target, characterized in that, include: Obtaining a confidence score for a current frame observation image using a reference template ; the confidence score is used to represent the reliability of the current frame target observation; The confidence score is compared with a preset confidence threshold. Based on the comparison result, the target bounding box of the current frame is determined, and the baseline template is adaptively updated for tracking processing of the next frame of the observed image. Among them, the confidence score ≥Confidence upper limit At the same time, the output target bounding box is used as the tracking result of the current frame; and the benchmark template is continuously used as the basis for calculating the confidence score of the observed image in the next frame. The confidence score satisfy: ≤ First confidence level lower limit At the same time, the target bounding box predicted by the Kalman filter is used as the tracking result for the current frame; simultaneously, a cross-medium transition template is constructed. And utilize this cross-medium transition template This serves as the basis for calculating the confidence score of the next frame of the observed image, and the confidence scores of the subsequent several frames of the observed image satisfy the following: First confidence lower limit. < ≤Second confidence level lower limit At that time, with the first update rate The cross-medium transition template is updated, and the updated cross-medium transition template is used as the basis for calculating the confidence score of subsequent observed images; The confidence score satisfy: < ≤ At the same time, the corrected target bounding box is used as the tracking result of the current frame observation image; simultaneously, at the second update rate... The current cross-medium transition template is updated, and the updated cross-medium transition template is used as the basis for calculating the confidence score of subsequent observed images; the corrected target bounding box is obtained by smoothing the target bounding box output by the visual tracking model and the target bounding box predicted by the Mann filter; The confidence score After experiencing confidence level fluctuations and instability, continuous Frame, confidence score satisfy: Then, the target bounding box output by the visual tracking model is used as the tracking result for the current frame; at the same time, a third update rate is used. The current cross-medium transition template is updated, and the updated cross-medium transition template is used as the basis for calculating the confidence score of subsequent observed images. in, The higher the update rate, the faster the template updates.

2. The water-air cross-medium target positioning and tracking method according to claim 1, characterized in that, The acquisition of the benchmark template includes: Acquire stable observation images of the target in a single medium and under stable observation conditions; The target bounding box of the stable observation image is determined by manual annotation or detection module, and the target bounding box of the stable observation image is cropped, scaled, and feature extracted to finally obtain the benchmark template.

3. The water-air cross-medium target positioning and tracking method according to claim 1, characterized in that, The constructed cross-media transition template is as follows: in, As a baseline template, For critically reliable frame templates, for Kalman first frame guide template; , , To merge weights, and satisfy .

4. The water-air cross-medium target positioning and tracking method according to claim 3, characterized in that, The critical reliable frame template The acquisition includes: When the model calculates the confidence score of the currently observed image ≤ First confidence level lower limit At that time, extract the historical confidence scores of the three most recent historical observation images. And from the historical confidence score of these 3 The target bounding box with the highest confidence score is selected from the samples, and its features are extracted to obtain the critically reliable frame template. ; The Kalman first frame guiding template The acquisition includes: When the model calculates the confidence score of the currently observed image ≤ First confidence level lower limit At that time, the target bounding box of the next frame of the observed image predicted by the Kalman filter is obtained, and its features are extracted.

5. The water-air cross-medium target positioning and tracking method according to claim 3, characterized in that, The target feature template is as follows: 。 6. The water-air cross-medium target positioning and tracking method according to claim 1, characterized in that, The first update rate Updating the cross-media transition template includes: Every P frames, the critical reliable frame template and the Kalman first frame guiding template are recalculated, and the cross-media transition template is updated with the recalculated critical reliable frame template and the Kalman first frame guiding template. The second update rate Updating the cross-media transition template includes: Every Q frames, the critical reliable frame template and the Kalman first frame guiding template are recalculated, and the cross-media transition template is updated with the recalculated critical reliable frame template and the Kalman first frame guiding template. The third update rate Updating the cross-media transition template includes: Every R frames, the critical reliable frame template and the Kalman first frame guiding template are recalculated, and the cross-media transition template is updated with the recalculated critical reliable frame template and the Kalman first frame guiding template, where R < Q < P.

7. The water-air cross-medium target positioning and tracking method according to claim 1, characterized in that, The confidence score of the observed image in the current frame is obtained using a benchmark template. include: Using the benchmark template, a matching search is performed in the current frame observation image to find the search region that best matches the benchmark template; Extract the features of the search region to obtain the search region features; The search region features are matched with the baseline template to generate a response map; The confidence score of the observed image in the current frame is determined based on the response map. .

Images