A visual observation correction method for unmanned aerial vehicle countermeasure

Abstract

The application discloses a visual observation correction method for unmanned aerial vehicle countermeasure, and relates to the technical field of unmanned aerial vehicle countermeasure sensing. The technical scheme is as follows: after the target detector outputs the initial boundary box of the target unmanned aerial vehicle, the initial boundary box is used to construct the target local region and its context from the real-time image, and the initial boundary box, the detection confidence and the observation quality information are input into the observation correction network; the observation correction network comprises a shared encoder, a geometric perception branch, a domain robust branch and an uncertainty branch, and the modified boundary box, the robust observation representation and the observation uncertainty are obtained; in the reasoning stage, the real-time image, the detection box, the detection confidence and the observation quality information are relied on, and the geometric true value, the environment label or the target pose input are not needed. The application provides stable and reliable visual observation input for state estimation, trajectory prediction and tracking control in the unmanned aerial vehicle countermeasure.

Description

Technical Field

[0001] This invention relates to the field of drone countermeasure perception technology, and in particular to a visual observation correction method for drone countermeasure. Background Technology

[0002] With the widespread use of low-altitude drones in surveying, inspection, logistics, and consumer entertainment, the problems of illegal flights, incursions, and interference in low-altitude airspace are becoming increasingly prominent. Visual perception is often used as a crucial front-end sensing method for monitoring, tracking, and countering target drones. Compared to radar and radio detection, visual methods offer advantages such as lower equipment costs, richer image information, and more flexible deployment, making them suitable for use on drone platforms or in ground-based countermeasure systems.

[0003] In real-world counter-drone scenarios, target drones are often characterized by long distances, small scales, high speeds, and significant attitude changes. Furthermore, images are easily affected by complex backgrounds, lighting variations, weather conditions, motion blur, partial occlusion, and imaging noise. Therefore, even if a target detector can detect the target, its output bounding box may exhibit issues such as center shift, scale instability, boundary jitter, or inconsistency with the target's true geometry. For applications that only perform image annotation or target display, these deviations may sometimes have limited impact; however, in the countermeasures chain of detection-state estimation-prediction-control, instability in the front-end bounding box will further propagate to subsequent modules, leading to jitter in state estimation, unstable target tracking, and even frequent fluctuations in control commands.

[0004] Existing methods often focus on the detector itself, improving target detection capabilities through enhancements to the backbone network, feature fusion structure, attention mechanisms, or training sample augmentation. While these methods can improve detection rate and localization accuracy, they typically assume that the detector's output bounding boxes can be directly used as geometric observations for subsequent control or estimation. In reality, the bounding boxes output by the target detector are closer to the visible target outlines or detection annotation semantics in the image, while subsequent modules in the countermeasure system require consistent, scene-stable observations that reflect geometric changes in the target. The two may be quite similar when the target is clear, large-scale, and facing directly, but systematic differences can easily arise with small targets, oblique perspectives, partial occlusion, or motion blur. Furthermore, simply using filtering methods to temporally smooth the detection boxes, while mitigating some jitter, cannot distinguish between the true geometric changes within the detection boxes and environmental interference biases, nor can it provide a reliable assessment of the current observations.

[0005] How to solve the above-mentioned technical problems is the challenge facing this invention. Summary of the Invention

[0006] The purpose of this invention is to provide a visual observation correction method for UAV countermeasures, addressing the problems of existing visual detection results easily exhibiting bounding box jitter, scale bias, cross-scene instability, and lack of credibility description in complex environments. This method does not aim to replace the target detector, but rather, based on the initial bounding box provided by the target detector, further corrects the bounding box to a visual observation more suitable for subsequent state estimation, trajectory prediction, and countermeasure control.

[0007] To achieve the aforementioned objectives, the present invention employs the following technical solution: a visual observation correction method for countering unmanned aerial vehicles (UAVs), comprising the following steps:

[0008] Step 1: Acquire real-time images containing the target drone And use a target detector to analyze the real-time images. The initial bounding box of the target UAV was obtained through detection. and detection confidence level From the real-time image Extracting a local region of the target and its context region And construct prior quality information to characterize the current observation quality. ;

[0009] Step 2: Combine the target local region and its context region. Initial bounding box Detection confidence and quality prior information Organization for observation correction network The input is used to obtain the bounding box correction amount. Robust observational characterization and observation uncertainty The relationship is as follows:

[0010]

[0011] Step 3: Adjust the bounding box amount accordingly. For the initial bounding box Perform residual correction to obtain the corrected bounding box of the target UAV. The relationship is as follows:

[0012]

[0013] Step 4: During the training phase, construct the target loss function and use the unified geometric reference box generated by the simulation platform. For the corrected bounding box Supervision is performed, and samples from different environmental domains and under different coarse geometric conditions are used to constrain the observation correction network. ;

[0014] Step 5: During the inference phase, only real-time images are used. Initial bounding box Detection confidence and quality prior information Output corrected bounding box Robust observational characterization and observation uncertainty It is used as a visual observation input for state estimation, trajectory prediction, or tracking control in UAV countermeasure systems.

[0015] Furthermore, the initial bounding box output by the target detector in step one... Visual observations are modeled as being formed by a unified geometric reference box, semantic bias of the visual box, context-related bias, and random noise, and their relationships are as follows:

[0016]

[0017] in This indicates the geometric state of the target UAV relative to the observation equipment. This represents the unified geometric reference frame determined by the stated geometric state. This represents the semantic deviation of the visual box between the detector output box and the unified geometric reference box. This indicates environment-domain-related bias caused by changes in background, lighting, weather, motion blur, or imaging conditions. This represents the random noise generated by the detector itself. , Center of the detection frame coordinate, Center of the detection frame coordinate, For the width of the detection frame, The height of the detection frame;

[0018] The quality prior information mentioned in step one Information includes target scale, degree of boundary truncation, detection confidence, target center offset, bounding box aspect ratio, image blur index, and whether the target is close to the image edge.

[0019] Furthermore, the observation correction network mentioned in step two... Including shared encoders Geometric Perception Branch Domain robust branches and uncertainty branch The shared encoder extracts features from the input observations to obtain shared observation features. And satisfy the following relationship:

[0020]

[0021] in, Used to characterize geometric information related to target scale, center position, aspect ratio, relative distance, and viewpoint changes; Used to characterize observational information that is relatively stable in response to changes in the environmental domain; Used to estimate the degree of uncertainty in current visual observations.

[0022] Furthermore, in steps two and three, the total bounding box correction amount The following relationship must be satisfied:

[0023]

[0024] in The geometric principal correction term generated for the geometry-aware branch is primarily responsible for recovering the bounding box deviation determined by the true projected geometry of the target, satisfying... , The environmental bias compensation amount generated for the domain robust branch is mainly used to compensate for systematic observation shifts caused by environmental factors such as backlighting, low contrast, complex backgrounds, and slight blurring, satisfying the requirements of the domain robust branch. , and These are the geometric perception branch decoding network and the domain robust branch decoding network, respectively. and The geometric correction gating and environmental compensation gating values ​​are predicted for the quality control module, enabling the model to maintain a more conservative correction strategy on easy samples and release a stronger compensation capability on difficult samples. This represents element-wise multiplication; The environmental compensation scaling coefficient; the uncertainty branch prediction heteroscedasticity observation variance vector satisfies , The uncertainty branch decoding network's output describes the confidence differences of the current observation across different bounding box dimensions; the domain robust branch projection represents the low-dimensional stable observation. , This is a domain-robust branch decoding network that retains stable information that is useful for downstream state estimation and behavior prediction while being relatively insensitive to changes in the environmental domain.

[0025] Furthermore, the unified geometric reference frame mentioned in step four... The observation correction network is generated by the simulation platform based on the target UAV's 3D bounding box, camera intrinsic parameters, camera extrinsic parameters, and the target's relative pose; When, its total loss function Including bounding box reference alignment loss overlap loss Uncertainty loss Geometric Preservation Auxiliary Loss Total loss of counterfactual consistency Geometric sensitivity preservation loss Latent variable decoupling loss and quality-aware gating regularization It consists of, and satisfies the following relationship:

[0026]

[0027] in, Used to constrain and correct the bounding box to approximate the unified geometric reference box. Used to improve the overlap between the corrected bounding box and the unified geometric reference box. Used to match observation uncertainty with bounding box correction error. This is used to ensure that samples with similar coarse geometry but different environmental domains maintain consistent output. Used to maintain the responsiveness of geometry-aware branches under real geometric changes. Used to reduce excessive coupling between features of different branches. Used to suppress unnecessary large corrections on high-quality samples Used to help reduce the cost of true projection geometry changes;

[0028] The total loss of counterfactual consistency for:

[0029]

[0030] in To account for the loss in cross-domain consistency constraints on domain robust branches, This is the set of valid cross-domain matching samples in the current batch. If the nearest distance between two samples exceeds a threshold... If this sample does not participate in environmental intervention consistency learning in the current batch, The alignment weights are based on geometric similarity; the closer the geometry, the greater the alignment weight, and vice versa. This applies at the environmental compensation branch level. This constrains the consistency of environmental compensation terms across different environmental domains, thereby highlighting its role as "environmental deviation compensation" rather than "geometric principal correction." The corresponding loss function within the correction box is... , for A subset of, and satisfying

[0031] ; Calculate the distance between two samples. A correction factor greater than 0. and This is the balance coefficient; Distance threshold; and Indicates sample and The environment;

[0032] The geometric sensitivity preservation loss , The minimum interval for geometric branches indicates that when the coarse geometric states are significantly different and the environment is under control, the geometric branches should not be compressed too closely together by the steady-state mechanism. This is a set of control pairs where the geometric state changes significantly under controlled environmental conditions. It is used to constrain geometric branches to maintain their sensitivity to actual geometric changes, satisfying... , This is the distance threshold.

[0033] Furthermore, during inference in step five, the observation correction network no longer relies on geometric reference boxes, environment domain labels, or coarse geometric condition variables, but only uses online-available detection results and their local context to complete the observation correction. Specifically, given the current image... First, the target detector outputs a coarse detection bounding box for the target UAV. and its confidence level Then, local observation units are constructed based on the detection frame. and combined with quality prior Input the observation correction network to obtain the corrected bounding box. Domain robust representation and observation uncertainty .

[0034] Meanwhile, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method described in the present invention.

[0035] Furthermore, the present invention proposes a computer-readable storage medium having a computer program stored thereon, the computer program being configured to implement the steps of the method described in the present invention when invoked by a processor.

[0036] Finally, the present invention provides a computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described in the present invention.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] (1) The present invention sets up a visual observation correction front end after the target detector. It does not change the target detection function of the original target detector, but performs secondary correction based on the initial detection box for state estimation, trajectory prediction and tracking control. This can alleviate the problems of detection box center offset, scale jitter and geometric definition inconsistency under conditions such as complex background, long distance small target, motion blur, and local occlusion, making the output corrected bounding box more suitable as a visual observation input in the UAV countermeasure link.

[0039] (2) This invention constructs an observation correction network consisting of a shared encoder, a geometric perception branch, a domain robust branch, an uncertainty branch, and a quality perception gating module. This network can complete bounding box correction, robust representation extraction, and observation confidence estimation within the same model. This network can preserve the true geometric changes of the target, reduce observation biases caused by environmental factors such as background, illumination, weather, and blur, and can adaptively adjust the correction intensity according to the current observation quality, thereby improving the stability and reliability of visual observation in complex scenes.

[0040] (3) During the training phase of this invention, the unified geometric reference frame, environment domain labels and coarse geometric conditions generated by the simulation platform can be used for supervised learning and sample organization. However, during the inference phase, it does not rely on geometric truth, environment labels, coarse geometric labels or the target's true pose, which makes it easy to deploy in actual UAV countermeasure systems. Attached Figure Description

[0041] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0042] Figure 1 The present invention provides an overall flowchart of a visual observation correction method for countering unmanned aerial vehicles.

[0043] Figure 2 This is a schematic diagram of the conditional domain robust observation correction network structure in this invention.

[0044] Figure 3 This is a schematic diagram showing the results of the robust observation correction method and comparison algorithm of this invention on the MAE index.

[0045] Figure 4 This is a schematic diagram showing the results of the algorithm of this invention and the comparison algorithm on the IoU index. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] Example 1

[0048] See Figure 1 and Figure 4 This embodiment 1 provides a visual observation correction method for UAV countermeasures. This method is applicable to scenarios where a first UAV, a ground-based countermeasure platform, or other visual observation equipment detects, tracks, estimates the state of, predicts the trajectory of, or controls the approach of a second UAV (target UAV). For example... Figure 1 As shown, it includes the following steps:

[0049] 1): Acquire real-time images containing the target drone. And use a target detector to analyze the real-time images. The initial bounding box of the target UAV was obtained through detection. and detection confidence level From the real-time image Extracting a local region of the target and its context region And construct prior quality information to characterize the current observation quality. ;

[0050] 2): The target local region and its context region Initial bounding box Detection confidence and quality prior information Organization for observation correction network The input is used to obtain the bounding box correction amount. Robust observational characterization and observation uncertainty The relationship is as follows:

[0051]

[0052] 3): Based on the bounding box correction amount For the initial bounding box Perform residual correction to obtain the corrected bounding box of the target UAV. The relationship is as follows:

[0053]

[0054] 4) During the training phase, construct the target loss function and utilize the unified geometric reference box generated by the simulation platform. For the corrected bounding box Supervision is performed, and samples from different environmental domains and under different coarse geometric conditions are used to constrain the observation correction network. ;

[0055] 5): During the inference phase, only real-time images are used. Initial bounding box Detection confidence and quality prior information Output corrected bounding box Robust observational characterization and observation uncertainty It is used as a visual observation input for state estimation, trajectory prediction, or tracking control in UAV countermeasure systems.

[0056] Further, the initial bounding box output by the target detector in step 1) Visual observations are modeled as being formed by a unified geometric reference box, semantic bias of the visual box, context-related bias, and random noise, and their relationships are as follows:

[0057]

[0058] in This indicates the geometric state of the target UAV relative to the observation equipment. This represents the unified geometric reference frame determined by the stated geometric state. This represents the semantic deviation of the visual box between the detector output box and the unified geometric reference box. This indicates environment-domain-related bias caused by changes in background, lighting, weather, motion blur, or imaging conditions. This represents the random noise generated by the detector itself. , Center of the detection frame coordinate, Center of the detection frame coordinate, For the width of the detection frame, The height is the detection box height; the target detector can be a YOLO series detector, or other target detection models that can output target bounding boxes and detection confidence scores.

[0059] The quality prior information mentioned in step 1) Information includes target scale, degree of boundary truncation, detection confidence, target center offset, bounding box aspect ratio, image blur index, and whether the target is close to the image edge.

[0060] Furthermore, such as Figure 2 As shown, the observation correction network in step 2) Including shared encoders Geometric Perception Branch Domain robust branches and uncertainty branch The shared encoder extracts features from the input observations to obtain shared observation features. And satisfy the following relationship:

[0061] ;

[0062] in, Used to characterize geometric information related to target scale, center position, aspect ratio, relative distance, and viewpoint changes; Used to characterize observational information that is relatively stable in response to changes in the environmental domain; Used to estimate the degree of uncertainty in current visual observations.

[0063] Furthermore, in steps 1) and 3), the total bounding box correction amount The following relationship must be satisfied:

[0064]

[0065] in The geometric principal correction term generated for the geometry-aware branch is primarily responsible for recovering the bounding box deviation determined by the true projected geometry of the target, satisfying... , The environmental bias compensation amount generated for the domain robust branch is mainly used to compensate for systematic observation shifts caused by environmental factors such as backlighting, low contrast, complex backgrounds, and slight blurring, satisfying the requirements of the domain robust branch. , and These are the geometric perception branch decoding network and the domain robust branch decoding network, respectively. and The geometric correction gating and environmental compensation gating values ​​are predicted for the quality control module, enabling the model to maintain a more conservative correction strategy on easy samples and release a stronger compensation capability on difficult samples. This represents element-wise multiplication; The environmental compensation scaling coefficient; the uncertainty branch prediction heteroscedasticity observation variance vector satisfies , The uncertainty branch decoding network's output describes the confidence differences of the current observation across different bounding box dimensions; the domain robust branch projection represents the low-dimensional stable observation. , This is a domain-robust branch decoding network that retains stable information that is useful for downstream state estimation and behavior prediction while being relatively insensitive to changes in the environmental domain.

[0066] Furthermore, the unified geometric reference frame mentioned in step 4) The observation correction network is generated by the simulation platform based on the target UAV's 3D bounding box, camera intrinsic parameters, camera extrinsic parameters, and the target's relative pose; When, its total loss function Including bounding box reference alignment loss overlap loss Uncertainty loss Geometric Preservation Auxiliary Loss Total loss of counterfactual consistency Geometric sensitivity preservation loss Latent variable decoupling loss and quality-aware gating regularization It consists of, and satisfies the following relationship:

[0067]

[0068] in, Used to constrain and correct the bounding box to approximate the unified geometric reference box. Used to improve the overlap between the corrected bounding box and the unified geometric reference box. Used to match observation uncertainty with bounding box correction error. This is used to ensure that samples with similar coarse geometry but different environmental domains maintain consistent output. Used to maintain the responsiveness of geometry-aware branches under real geometric changes. Used to reduce excessive coupling between features of different branches. Used to suppress unnecessary large corrections on high-quality samples Used to help reduce the cost of true projection geometry changes;

[0069] The total loss of counterfactual consistency for:

[0070]

[0071] in To account for the loss in cross-domain consistency constraints on domain robust branches, This is the set of valid cross-domain matching samples in the current batch. If the nearest distance between two samples exceeds a threshold... If this sample does not participate in environmental intervention consistency learning in the current batch, The alignment weights are based on geometric similarity; the closer the geometry, the greater the alignment weight, and vice versa. This applies at the environmental compensation branch level. This constrains the consistency of environmental compensation terms across different environmental domains, thereby highlighting its role as "environmental deviation compensation" rather than "geometric principal correction." The corresponding loss function within the correction box is... , for A subset of , and satisfying:

[0072] ; Calculate the distance between two samples. A correction factor greater than 0. and This is the balance coefficient; Distance threshold; and Indicates sample and The environment;

[0073] The geometric sensitivity preservation loss , The minimum interval for geometric branches indicates that when the coarse geometric states are significantly different and the environment is under control, the geometric branches should not be compressed too closely together by the steady-state mechanism. This is a set of control pairs where the geometric state changes significantly under controlled environmental conditions. It is used to constrain geometric branches to maintain their sensitivity to actual geometric changes, satisfying... , This is the distance threshold.

[0074] Furthermore, during inference in step 5), the observation correction network no longer relies on geometric reference boxes, environment domain labels, or coarse geometric condition variables, but only uses online-available detection results and their local context to complete the observation correction. Specifically, given the current image... First, the target detector outputs a coarse detection bounding box for the target UAV. and its confidence level Then, local observation units are constructed based on the detection frame. and combined with quality prior Input the observation correction network to obtain the corrected bounding box. Domain robust representation and observation uncertainty .

[0075] Example 2: Based on Example 1, the following experiment was conducted. The experiment used a unified validation set containing three different environment domains, with a total of 997 test samples. In the experiment, the original detection boxes output by the YOLO target detector were used as the baseline results, and the method of the present invention was compared with the YOLO method and the YOLO combined with the multilayer perceptron residual regression model (YOLO+MLP). The experimental evaluation indicators included the mean bounding box error (MAE) and the intersection-over-union ratio (IoU). The lower the MAE, the closer the corrected box is to the unified geometric reference box, and the higher the IoU, the greater the overlap between the corrected box and the unified geometric reference box.

[0076] like Figure 3As shown, the MAE of the YOLO baseline method is 0.002924, indicating that the original detection boxes still have certain center bias and scale bias. After introducing ordinary residual regression on the basis of the YOLO detection boxes, the MAE of the YOLO+MLP method is reduced to 0.002771, indicating that subsequent correction of the detection boxes can improve the observation accuracy to a certain extent. Using the method proposed in this invention, the MAE is further reduced to 0.002725, which is better than the YOLO baseline and YOLO+MLP methods. As can be seen from the MAE histogram, the method of this invention can more effectively correct the coarse detection boxes output by the target detector to observation results close to the unified geometric reference box, thereby improving the accuracy of visual observation and providing a more stable input for subsequent state estimation, trajectory prediction and tracking control.

[0077] like Figure 4 As shown, the IoU of the YOLO baseline method is 0.852167. After applying ordinary residual regression to the YOLO detection boxes, the IoU of the YOLO+MLP method increases to 0.858115, indicating that post-detection correction can improve the overlap between the bounding boxes and the reference boxes. The method proposed in this invention further improves the IoU to 0.860146, which is significantly better than the YOLO baseline and YOLO+MLP methods. This demonstrates that the method of this invention can better maintain its responsiveness to real geometric changes while suppressing the influence of environmental disturbances, thereby obtaining higher quality corrected bounding boxes.

[0078] Example 3: This example proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method steps of the present invention.

[0079] Example 4: This example proposes a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the method described in this invention, which will not be repeated here.

[0080] Example 5: This example proposes a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the steps of the method described in this invention, which will not be repeated here.

[0081] It should be noted that the processing flow of embodiments 2-5 corresponds to the specific steps of the method provided in embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in embodiment 1 of the present invention.

[0082] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A visual observation correction method for countering unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: S1. Acquire real-time images containing the target drone. And use a target detector to analyze the real-time images. The initial bounding box of the target UAV was obtained through detection. and detection confidence level From the real-time image Extracting a local region of the target and its context region And construct prior quality information to characterize the current observation quality. ; S2, The target local region and its context region are... Initial bounding box Detection confidence and quality prior information Organization for observation correction network The input is used to obtain the bounding box correction amount. Robust observational characterization and observation uncertainty The relationship is as follows: ； In S2, the observation correction network Including shared encoders Geometric Perception Branch Domain robust branches and uncertainty branch The shared encoder extracts features from the input observations to obtain shared observation features. And satisfy the following relationship: ； in, Used to characterize geometric information related to target scale, center position, aspect ratio, relative distance, and viewpoint changes; Used to characterize observational information that is relatively stable in response to changes in the environmental domain; Used to estimate the degree of uncertainty in current visual observations; The bounding box correction amount The following relationship must be satisfied: ； in, The principal geometric corrections generated for the geometric perception branch. Responsible for recovering the bounding box deviation determined by the true projected geometry of the target, satisfying , The environmental deviation compensation amount generated for the domain robust branch satisfies , and These are the geometric perception branch decoding network and the domain robust branch decoding network, respectively. and Predict the geometric correction gating and environmental compensation gating values ​​for the quality control module; This represents element-wise multiplication; The environmental compensation scaling coefficient; the uncertainty branch prediction heteroscedasticity observation variance vector satisfies , For uncertainty branch decoding network, The output describes the confidence differences of the current observations across different bounding box dimensions; the domain robust branch projection represents the low-dimensional stable observations. , Decode the domain robust branch network; S3, Based on the bounding box correction amount For the initial bounding box Perform residual correction to obtain the corrected bounding box of the target UAV. The relationship is as follows: ； S4. During the training phase, construct the target loss function and use the unified geometric reference box generated by the Rflysim simulation platform. For the corrected bounding box The observation correction network is supervised and constrained using samples from different environmental domains and under different coarse geometric conditions. ; S5. During the inference phase, utilize real-time images. Initial bounding box Detection confidence and quality prior information Output corrected bounding box Robust observational characterization and observation uncertainty It is used as a visual observation input for state estimation, trajectory prediction, or tracking control in UAV countermeasure systems.

2. The visual observation correction method for UAV countermeasures according to claim 1, characterized in that, In S1, the initial bounding box output by the target detector Visual observations are modeled as being formed by a unified geometric reference box, semantic bias of the visual box, context-related bias, and random noise, and their relationships are as follows: ； in This indicates the geometric state of the target UAV relative to the observation equipment. This represents the unified geometric reference frame determined by the stated geometric state. This represents the semantic deviation of the visual box between the detector output box and the unified geometric reference box. This indicates environment-domain related bias caused by changes in background, lighting, weather, motion blur, or imaging conditions. This represents the random noise generated by the detector itself. , Center of the detection frame coordinate, Center of the detection frame coordinate, For the detection frame width, This is the height of the detection frame.

3. The visual observation correction method for UAV countermeasures according to claim 1, characterized in that, In S1, the quality prior information This includes target scale, boundary truncation degree, detection confidence, target center offset, bounding box aspect ratio, image blur index, and information on whether the target is close to the image edge.

4. The visual observation correction method for UAV countermeasures according to claim 1, characterized in that, In S4, the unified geometric reference frame The observation correction network is generated by the Rflysim simulation platform based on the target UAV's 3D bounding box, camera intrinsic parameters, camera extrinsic parameters, and the target's relative pose; Total loss function Including bounding box reference alignment loss overlap loss Uncertainty loss Geometric Preservation Auxiliary Loss Total loss of counterfactual consistency Geometric sensitivity preservation loss Latent variable decoupling loss and quality-aware gating regularization The composition satisfies the following relationship: ； in, Used to constrain and correct the error of the unified geometric reference box of the bounding box. Used to improve the overlap between the corrected bounding box and the unified geometric reference box. Used to match observation uncertainty with bounding box correction error. Used to keep the outputs of samples with coarse geometric similarities but different environmental domains consistent. Used to maintain the responsiveness of geometry-aware branches under real geometric changes. Used to reduce excessive coupling between features of different branches. Used to suppress unnecessary large corrections on high-quality samples Used to help reduce the cost of true projection geometry changes; The total loss of counterfactual consistency for: ； in, To account for the loss in cross-domain consistency constraints on domain robust branches, This is the set of valid cross-domain matching samples in the current batch. If the geometric distance between two samples exceeds a threshold... If the cross-domain matching samples are valid, they will not participate in the environmental intervention consistency learning in the current batch. Geometric similarity weights; at the environmental compensation branch level The loss function corresponding to the correction box is , for A subset of , and satisfying: ; Calculate the geometric distance between two samples. A correction factor greater than 0. and This is the balance coefficient; Distance threshold; and Representing samples respectively and The environment; The geometric sensitivity preservation loss , Minimum interval for geometric branches This is a set of controls for changes in geometric state under controlled environmental conditions, used to constrain geometric branches to maintain sensitivity to actual geometric changes, satisfying... , This is the distance threshold.

5. The visual observation correction method for UAV countermeasures according to claim 1, characterized in that, During inference in S5, the observation correction network uses the online-obtained detection results and their local context to complete the observation correction; given the current image First, the target detector outputs a coarse detection bounding box for the target UAV. and its confidence level Then, local observation units are constructed based on the detection frame. and combined with quality prior Input the observation correction network to obtain the corrected bounding box. Domain robust representation and observation uncertainty .

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is configured to implement the steps of the method according to any one of claims 1 to 5 when invoked by a processor.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 5.

Images