Unmanned aerial vehicle target compensation and tracking recovery method capable of resisting pose interference in complex shielding environment

Abstract

An unmanned aerial vehicle target compensation and tracking recovery method capable of resisting pose interference in a complex shielding environment relates to the technical field of unmanned aerial vehicle intelligent tracking, when a target is shielded, the pose information of the unmanned aerial vehicle is fully utilized to carry out spatial modeling on the target motion, the interference of the pose change of the unmanned aerial vehicle on a target prediction result is effectively reduced, and the target prediction efficiency is improved. The stability and the accuracy of target compensation prediction during the shielding period are improved; meanwhile, environmental prior constraints such as road topology are introduced, so that the target prediction trajectory better conforms to a real motion law, and the adaptive capacity in a complex city scene is enhanced. And after shielding is finished, stable recapture and continuous tracking of the target are realized by combining multi-level recovery strategies such as kinematics prediction, passable area constraint and appearance feature matching.

Description

Technical Field

[0001] This invention relates to the field of intelligent tracking technology for unmanned aerial vehicles (UAVs), specifically to a method for UAV target compensation and tracking recovery against pose interference in complex occlusion environments. Background Technology

[0002] With the rapid development of the urban low-altitude economy, drones are increasingly widely used in urban patrols, traffic monitoring, and other scenarios. Continuous target tracking is crucial in these scenarios. However, in real urban environments, drones are easily affected by complex environments such as buildings and vegetation, leading to target occlusion and tracking loss, resulting in mission failure. Existing algorithms largely rely on pure image prediction. Furthermore, drones inevitably experience pose changes during flight, and the motion vectors in the image coordinate system cannot accurately reflect the target's true spatial trajectory, causing tracking breaks and hindering subsequent target recovery, thus limiting the vast scope of drone patrols. On the other hand, existing algorithms lack the integration of multi-source data, such as road network topology and occlusion perception, resulting in insufficient and effective information to guide target recovery after occlusion. Therefore, there is an urgent need for a drone target compensation and tracking recovery method that can fully utilize drone pose information in complex occlusion environments, combine kinematic models and road network prior constraints, and possess stable target recovery capabilities. Summary of the Invention

[0003] In order to overcome the shortcomings of the above technologies, this invention provides a method that effectively improves the continuity and robustness of target tracking of UAVs in complex occlusion environments and reduces the impact of pose disturbances on tracking results.

[0004] The technical solution adopted by this invention to overcome its technical problems is: A method for anti-pose interference-resistant UAV target compensation and tracking recovery in complex occlusion environments includes: S1. Obtain from drone camera Video frames at any moment ; S2. For video frames The target is specified in the code, and the target bounding box is obtained. S3. Constructing comprehensive criteria Based on comprehensive criteria Determine whether to add the height and width of the tracking target bounding box to the queue. ; S4. Based on Video frames at any moment The historical smoothing point sequence is obtained by tracking the midpoint coordinates of the target bounding box. and predicted trajectory ; S5. Utilizing historical smooth point sequences The upper limit of speed was calculated. ; S6. Utilize the speed limit Obtain the predicted trajectory ; S7. Predicting the trajectory and predicted trajectory By fusion, a new predicted sequence is obtained. ; S8. Utilizing new prediction sequences Kinematic prediction box after occlusion Obtain the passable area after obstruction. ; S9. Using occlusion-based kinematic prediction boxes and areas that can be passed through after being covered Resume target tracking for the specified target.

[0005] Furthermore, in step S2, the BoT-SORT algorithm is used to process the video frames. Specify a target for target tracking.

[0006] Furthermore, step S3 includes the following steps: S3-1. Through formula Calculate the validity criteria of the timestamp In the formula, To track the timestamp of the last video frame in which the target appeared, For the last time added to the queue The timestamp of the video frame when the queue When it is an empty set, ; S3-2. Using the formula Calculate the stability criterion In the formula, The relative rate of change , The aspect ratio is 1. , for Video frames at any moment Track the height of the target bounding box. for Video frames at any moment Track the width of the target bounding box. The average aspect ratio of the target bounding box across all video frames. The relative rate of change threshold; S3-3. Through formula The mandatory adoption criterion is calculated. In the formula For queue Number of elements in the middle For the threshold; S3-4. Through formula The comprehensive criterion is calculated. In the formula For logical AND operation, For logical OR operation; S3-5. When considering comprehensive criteria Determine if it equals 1 Video frames at any moment The tracking target bounding box is a valid detection box, and the height of the tracking target bounding box is... and the width of the tracking target bounding box Add to queue , , For queue The number of target boxes being tracked.

[0007] Preferred, The value range is 0-1. The value is 15.

[0008] Furthermore, step S4 includes the following steps: S4-1. Kalman Filter for Constructing a Uniform Linear Motion Model ; S4-2. Video frames at any moment The coordinates of the midpoint of the tracking target bounding box The coordinates of the midpoint of the target bounding box are obtained by converting the camera pinhole imaging model to the NED coordinate system. ,in To track the x-coordinate of the midpoint of the target bounding box, To track the y-coordinate of the midpoint of the target bounding box, To track the X-axis coordinate of the midpoint of the target bounding box, To track the Y-axis coordinate of the midpoint of the target bounding box, To track the Z-axis coordinate of the midpoint of the target bounding box; S4-3. Set the X-axis coordinate of the midpoint of the target box. and Y-axis coordinate Input to Kalman filter In the process, the optimal estimated X-axis coordinate is obtained from the output. and optimal estimated Y-axis coordinate Get the first video frame that appears from the specified target. up to the last video frame The historical smoothing point sequence is obtained by taking all the optimal estimated X-axis and Y-axis coordinates. , ; S4-4. Using a Kalman filter The first video frame where the specified target disappears. The position is predicted, with each prediction time step being 100ms, iterating 100 times to form a predicted trajectory within 10 seconds. , , For the first The X-axis coordinate of the predicted target location. For the first The Y-axis coordinate of the predicted target location.

[0009] Furthermore, step S5 includes the following steps: S5-1. Kalman Filter for Constructing a Uniform Linear Motion Model ; S5-2. Transform the historical smooth point sequence The Douglas-Puk algorithm is used to thin the data, resulting in a thinned sequence of historical smooth points. ; S5-3. The thinned historical smoothed point sequence The Hidden Markov Model (HMM) algorithm is used to match trajectories and obtain the road network projection sequence. ; S5-4. Projecting the road network sequence Input to Kalman filter In the process, the output yields the optimal estimate of the velocity of the midpoint of the tracking target box in the X-axis direction. Optimal estimation of velocity in the Y-axis direction ; S5-5. Through formula The resultant velocity scalar was calculated. Through formula The upper limit of speed was calculated. In the formula, To track the upper limit of turning acceleration for a specified target, The value is 2.5 m / s 2 , The turning radius is , This is the path curvature calculated based on the first and second derivatives of the fitted spline curves after performing cubic B-spline curve fitting on the road network.

[0010] Furthermore, step S6 includes the following steps: S6-1. Through formula Calculate the first Curve arc length at the next evolution moment ,when When the value is 1, , For the first The next evolution moment to the first The difference at each evolution time; S6-2. If the first In the next evolution, if there are no forks in the road network 100m ahead of the drone, then the curve arc length is... Use the splines.UnitSpeedAdapter.evaluate() function from the splines library to get the number of... X-axis coordinate of the predicted endpoint position of the next evolution and Y-axis coordinate If the first In the next evolution, if there is a fork in the road network 100m ahead of the drone, then the fork node will be... The X-axis coordinate is used as the predicted endpoint position X-axis coordinate. The fork in the road node The Y-axis coordinate is used as the predicted endpoint position. ; S6-3. The X-axis and Y-axis coordinates of all 100 predicted endpoint positions constitute the predicted trajectory. , .

[0011] Furthermore, step S7 includes the following steps: S7-1. Through formula The occlusion duration factor was calculated. In the formula, For sensitivity parameters, The value ranges from 0.1 to 1; S7-2. Through formula Calculate the first Secondary evolution of road network topology disturbance term In the formula, For predicting trajectories The Middle A predicted target location, To obtain the Euclidean distance, To influence the radius parameter; S7-3. Through formula Calculate the first Fusion weights during the next evolution ; S7-4. Through formula Calculate the fused first The predicted target location X-axis coordinate Through formula Calculate the fused first The predicted target location Y-axis coordinate ; S7-5. All 100 predicted target locations after fusion constitute a new prediction sequence. , .

[0012] Furthermore, step S8 includes the following steps: S8-1. The new predicted sequence Using video frames The corresponding gimbal pose information is reprojected to pixel coordinates, and a new prediction sequence is selected. Mid-range video frames The pixel X-axis coordinate of the most recent evolution moment of the timestamp and pixel Y-axis coordinate ; S8-2. Calculation Queue The average height of all tracked target boxes is , computation queue The average width of all tracked target boxes is ; S8-3. Constructing the kinematic prediction box after occlusion , ; S8-4. Video frames The input is given to the U2-Net model, and the output is the passable area after occlusion. .

[0013] Furthermore, step S9 includes the following steps: S9-1. When the space efficiency constraint is satisfied: Kinematic prediction box after occlusion is determined To effectively predict and execute step S9-2, where The minimum area ratio, The value range is 0-1. Kinematic prediction box after occlusion area, Kinematic prediction box after occlusion Passable areas after obstruction The area of ​​the intersection; S9-2. Video frames The input is fed into a YOLOv11 network, and the output is obtained. The detection box, the first Each detection box is , ; S9-3. When the space efficiency constraint is satisfied: Time judgment Detection box To effectively detect and execute step S9-4, where For the first Detection box area, For the first Detection box Passable areas after obstruction The area of ​​the intersection; S9-4. Kinematic prediction box after occlusion With the Detection box Perform IOU matching; if a match is successful, resume tracking of the specified target.

[0014] The beneficial effects of this invention are as follows: When the target is occluded, the invention fully utilizes the UAV pose information to spatially model the target motion, effectively reducing the interference of UAV attitude changes on the target prediction results and improving the stability and accuracy of target compensation prediction during occlusion. Simultaneously, the introduction of prior environmental constraints such as road topology makes the predicted target trajectory more consistent with real motion patterns, enhancing adaptability in complex urban scenarios. After occlusion ends, this invention achieves stable re-acquisition and continuous tracking of the target by combining multi-level recovery strategies such as kinematic prediction, passable area constraints, and appearance feature matching. This method requires no additional hardware, has strong practicality and promotional value, and is suitable for low-altitude UAV applications such as urban patrols. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The following is in conjunction with the appendix Figure 1 The present invention will be further described below.

[0017] A method for anti-pose interference-resistant UAV target compensation and tracking recovery in complex occlusion environments includes: S1. Obtain from drone camera Video frames at any moment .

[0018] S2. For video frames The target is specified in the code, and the target bounding box is obtained.

[0019] S3. Constructing comprehensive criteria Based on comprehensive criteria Determine whether to add the height and width of the tracking target bounding box to the queue. .

[0020] S4. Based on Video frames at any moment The historical smoothing point sequence is obtained by tracking the midpoint coordinates of the target bounding box. and predicted trajectory .

[0021] S5. Utilizing historical smooth point sequences The upper limit of speed was calculated. .

[0022] S6. Utilize the speed limit Obtain the predicted trajectory .

[0023] S7. Predicting the trajectory and predicted trajectory By fusion, a new predicted sequence is obtained. .

[0024] S8. Utilizing new prediction sequences Kinematic prediction box after occlusion Obtain the passable area after obstruction. .

[0025] S9. Using occlusion-based kinematic prediction boxes and areas that can be passed through after being covered Resume target tracking for the specified target.

[0026] This invention aims to address the problems of existing UAV target tracking methods in complex urban scenarios, such as easy tracking interruption, large prediction errors, and difficulty in stable target recovery after occlusion, caused by target occlusion and changes in UAV pose. The technical solution of this invention fully utilizes UAV pose information to construct a unified spatial motion model when the target is partially or completely occluded, compensating for and predicting the target's actual motion state; simultaneously, it incorporates prior environmental constraints such as road topology to improve the reliability of target trajectory prediction during occlusion; and after occlusion ends, it achieves stable re-acquisition and continuous tracking of the target through a multi-level target recovery strategy.

[0027] This invention effectively improves the continuity and robustness of target tracking in complex obstructed environments for UAVs without requiring additional hardware modifications, and reduces the impact of pose disturbances on tracking results. It is suitable for low-altitude UAV applications such as urban patrol and traffic monitoring, and has good practical value and promotion prospects.

[0028] In one embodiment of the present invention, the BoT-SORT algorithm is used to process video frames in step S2. Specify a target for target tracking.

[0029] In one embodiment of the present invention, step S3 includes the following steps: S3-1. Through formula Calculate the validity criteria of the timestamp In the formula, To track the timestamp of the last video frame in which the target appeared, For the last time added to the queue The timestamp of the video frame when the queue When it is an empty set, .

[0030] S3-2. Using the formula Calculate the stability criterion In the formula, The relative rate of change represents the aspect ratio. The degree of deviation from the historical average , The aspect ratio is 1. , for Video frames at any moment Track the height of the target bounding box. for Video frames at any moment Track the width of the target bounding box. The average aspect ratio of the target bounding boxes across all video frames is used. Excessive deviation from this ratio indicates an abnormal bounding box, and the box is rejected from the queue. , This is the threshold for the relative rate of change.

[0031] S3-3. Through formula The mandatory adoption criterion is calculated. In the formula For queue Number of elements in the middle The threshold value is used.

[0032] S3-4. Through formula The comprehensive criterion is calculated. In the formula For logical AND operation, For logical OR operation.

[0033] S3-5. When considering comprehensive criteria Determine if it equals 1 Video frames at any moment The tracking target bounding box is a valid detection box, and the height of the tracking target bounding box is... and the width of the tracking target bounding box Add to queue , , For queue The number of target bounding boxes tracked. Comprehensive criteria. When the value is 0, it is judged as abnormal jitter or duplicate data and is rejected.

[0034] In this embodiment, preferably, The value range is 0-1. The value is 15.

[0035] In one embodiment of the present invention, step S4 includes the following steps: S4-1. Kalman Filter for Constructing a Uniform Linear Motion Model .

[0036] S4-2. Video frames at any moment The coordinates of the midpoint of the tracking target bounding box By converting the camera pinhole imaging model to the North-East-Ground (NED) coordinate system, the coordinates of the midpoint of the target bounding box can be obtained. ,in To track the x-coordinate of the midpoint of the target bounding box, To track the y-coordinate of the midpoint of the target bounding box, To track the X-axis coordinate of the midpoint of the target bounding box, To track the Y-axis coordinate of the midpoint of the target bounding box, The Z-axis coordinate of the midpoint of the target bounding box.

[0037] S4-3. Set the X-axis coordinate of the midpoint of the target box. and Y-axis coordinate Input to Kalman filter In the process, the optimal estimated X-axis coordinate is obtained from the output. and optimal estimated Y-axis coordinate Get the first video frame that appears from the specified target. up to the last video frame The historical smoothing point sequence is obtained by taking all the optimal estimated X-axis and Y-axis coordinates. , .

[0038] S4-4. When the target is completely occluded, in the first video frame where the target disappears. The target detector cannot obtain valid observation information. Due to the lack of observations, the BOTSORT association fails, and the target is determined to be in an occluded state. At this point, a Kalman filter is used. The first video frame where the specified target disappears. The position is predicted, with each prediction time step being 100ms, iterating 100 times to form a predicted trajectory within 10 seconds. , , For the first The X-axis coordinate of the predicted target location. For the first The Y-axis coordinate of the predicted target location.

[0039] In one embodiment of the present invention, step S5 includes the following steps: S5-1. Kalman Filter for Constructing a Uniform Linear Motion Model .

[0040] S5-2. Transform the historical smooth point sequence The Douglas-Puk algorithm is used to thin the data, resulting in a thinned sequence of historical smooth points. .

[0041] S5-3. The thinned historical smoothed point sequence The Hidden Markov Model (HMM) algorithm is used to match trajectories and obtain the road network projection sequence. .

[0042] S5-4. Projecting the road network sequence Input to Kalman filter In the process, the output yields the optimal estimate of the velocity of the midpoint of the tracking target box in the X-axis direction. Optimal estimation of velocity in the Y-axis direction .

[0043] S5-5. Due to the curvature of the road network, there is an upper limit to the target speed when turning, using the formula... The resultant velocity scalar was calculated. Through formula The upper limit of speed was calculated. In the formula, To track the upper limit of turning acceleration for a specified target, The value is 2.5 m / s 2 , The turning radius is , This is the path curvature calculated based on the first and second derivatives of the fitted spline curves after performing cubic B-spline curve fitting on the road network.

[0044] In one embodiment of the present invention, step S6 includes the following steps: S6-1. Through formula Calculate the first Curve arc length at the next evolution moment ,when When the value is 1, , For the first The next evolution moment to the first The difference at each evolution time.

[0045] S6-2. If the first In the next evolution, if there are no forks in the road network 100m ahead of the drone, then the curve arc length is... Use the splines.UnitSpeedAdapter.evaluate() function from the splines library to get the number of... X-axis coordinate of the predicted endpoint position of the next evolution and Y-axis coordinate If the first In the next evolution, if there is a fork in the road network 100m ahead of the drone, then the fork node will be... The X-axis coordinate is used as the predicted endpoint position X-axis coordinate. The fork in the road node The Y-axis coordinate is used as the predicted endpoint position. .

[0046] S6-3. The X-axis and Y-axis coordinates of all 100 predicted endpoint positions constitute the predicted trajectory. , .

[0047] In one embodiment of the present invention, step S7 includes the following steps: S7-1. Through formula The occlusion duration factor was calculated. In the formula, This is a sensitivity parameter used to adjust the effect of occlusion duration on the predicted trajectory. Predicting the rate of credibility decay, when When it is large, the occlusion duration factor It will quickly approach 1. The value ranges from 0.1 to 1.

[0048] S7-2. Through formula Calculate the first Secondary evolution of road network topology disturbance term In the formula, For predicting trajectories The Middle A predicted target location, To obtain the Euclidean distance, The influence radius parameter is used to adjust the influence distance at the intersection. Secondary evolution of road network topology disturbance term The closer the predicted point is to the intersection, the larger it becomes, indicating a decreasing level of distrust in the predicted trajectory. .

[0049] S7-3. Through formula Calculate the first Fusion weights during the next evolution .

[0050] S7-4. Through formula Calculate the fused first The predicted target location X-axis coordinate Through formula Calculate the fused first The predicted target location Y-axis coordinate .

[0051] S7-5. All 100 predicted target locations after fusion constitute a new prediction sequence. , .

[0052] In one embodiment of the present invention, step S8 includes the following steps: S8-1. The new predicted sequence Using video frames The corresponding gimbal pose information is reprojected to pixel coordinates, and a new prediction sequence is selected. Mid-range video frames The pixel X-axis coordinate of the most recent evolution moment of the timestamp and pixel Y-axis coordinate .

[0053] S8-2. Calculation Queue The average height of all tracked target boxes is , computation queue The average width of all tracked target boxes is .

[0054] S8-3. Constructing the kinematic prediction box after occlusion , .

[0055] S8-4. Video frames The input is given to the U2-Net model, and the output is the passable area after occlusion. .

[0056] In one embodiment of the present invention, step S9 includes the following steps: S9-1. When the space efficiency constraint is satisfied: Kinematic prediction box after occlusion is determined To effectively predict and execute step S9-2, where The minimum area ratio, The value range is 0-1. Kinematic prediction box after occlusion area, Kinematic prediction box after occlusion Passable areas after obstruction The area of ​​the intersection of the two.

[0057] S9-2. Video frames The input is fed into a YOLOv11 network, and the output is obtained. The detection box, the first Each detection box is , .

[0058] S9-3. When the space efficiency constraint is satisfied: Time judgment Detection box To effectively detect and execute step S9-4, where For the first Detection box area, For the first Detection box Passable areas after obstruction The area of ​​the intersection of the two.

[0059] S9-4. Kinematic prediction box after occlusion With the Detection box Perform IOU matching; if a match is successful, resume tracking of the specified target.

[0060] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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 method for anti-pose interference-resistant UAV target compensation and tracking recovery in complex occlusion environments, characterized in that, include: S1. Obtain from drone camera Video frames at any moment ; S2. For video frames The target is specified in the code, and the target bounding box is obtained. S3. Constructing comprehensive criteria Based on comprehensive criteria Determine whether to add the height and width of the tracking target bounding box to the queue. ; S4. Based on Video frames at any moment The historical smoothing point sequence is obtained by tracking the midpoint coordinates of the target bounding box. and predicted trajectory ; S5. Utilizing historical smooth point sequences The upper limit of speed was calculated. ; S6. Utilize the speed limit Obtain the predicted trajectory ; S7. Predicting the trajectory and predicted trajectory By fusion, a new predicted sequence is obtained. ; S8. Utilizing new prediction sequences Kinematic prediction box after occlusion Obtain the passable area after obstruction. ; S9. Using occlusion-based kinematic prediction boxes and areas that can be passed through after being covered Resume target tracking for the specified target.

2. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 1, characterized in that: In step S2, the BoT-SORT algorithm is used to process the video frames. Specify a target for target tracking.

3. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 1, characterized in that, Step S3 includes the following steps: S3-1. Through formula Calculate the validity criteria of the timestamp In the formula, To track the timestamp of the last video frame in which the target appeared, For the last time added to the queue The timestamp of the video frame when the queue When it is an empty set, ; S3-2. Using the formula Calculate the stability criterion In the formula, The relative rate of change , The aspect ratio is 1. , for Video frames at any moment Track the height of the target bounding box. for Video frames at any moment Track the width of the target bounding box. The average aspect ratio of the target bounding box across all video frames. The relative rate of change threshold; S3-3. Through formula The mandatory adoption criterion is calculated. In the formula For queue Number of elements in the middle For threshold; S3-4. Through formula The comprehensive criterion is calculated. In the formula For logical AND operation, For logical OR operation; S3-5. When considering comprehensive criteria Determine if it equals 1 Video frames at any moment The tracking target bounding box is a valid detection box, and the height of the tracking target bounding box is... and the width of the tracking target bounding box Add to queue , , For queue The number of target boxes tracked in the middle.

4. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 3, characterized in that: The value range is 0-1. The value is 15.

5. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 1, characterized in that, Step S4 includes the following steps: S4-1. Kalman Filter for Constructing a Uniform Linear Motion Model ; S4-2. Video frames at any moment The coordinates of the midpoint of the tracking target bounding box The coordinates of the midpoint of the target bounding box are obtained by converting the camera pinhole imaging model to the NED coordinate system. ,in To track the x-coordinate of the midpoint of the target bounding box, To track the y-coordinate of the midpoint of the target bounding box, To track the X-axis coordinate of the midpoint of the target bounding box, To track the Y-axis coordinate of the midpoint of the target bounding box, To track the Z-axis coordinate of the midpoint of the target bounding box; S4-3. Set the X-axis coordinate of the midpoint of the target box. and Y-axis coordinate Input to Kalman filter In the process, the optimal estimated X-axis coordinate is obtained from the output. and optimal estimated Y-axis coordinate Get the first video frame that appears from the specified target. Until the last video frame The historical smoothing point sequence is obtained by taking all the optimal estimated X-axis and Y-axis coordinates. , ; S4-4. Using a Kalman filter The first video frame where the specified target disappears. The position is predicted, with each prediction time step being 100ms, iterating 100 times to form a predicted trajectory within 10 seconds. , , For the first The X-axis coordinate of the predicted target location. For the first The Y-axis coordinate of the predicted target location.

6. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 1, characterized in that, Step S5 includes the following steps: S5-1. Kalman Filter for Constructing a Uniform Linear Motion Model ; S5-2. Transform the historical smooth point sequence The Douglas-Puk algorithm is used to thin the data, resulting in a thinned sequence of historical smooth points. ; S5-3. The thinned historical smoothed point sequence The Hidden Markov Model (HMM) algorithm is used to match trajectories and obtain the road network projection sequence. ; S5-4. Projecting the road network sequence Input to Kalman filter In the process, the output yields the optimal estimate of the velocity of the midpoint of the tracking target box in the X-axis direction. Optimal estimation of velocity in the Y-axis direction ; S5-5. Through formula The resultant velocity scalar was calculated. Through formula The upper limit of speed was calculated. In the formula, To track the upper limit of turning acceleration for a specified target, The value is 2.5 m / s 2 , The turning radius, , This is the path curvature calculated based on the first and second derivatives of the fitted spline curves after performing cubic B-spline curve fitting on the road network.

7. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 5, characterized in that, Step S6 includes the following steps: S6-1. Through formula Calculation yields the first Curve arc length at the next evolution moment ,when When the value is 1, , For the first The next evolutionary moment to the first The difference at each evolution time; S6-2. If the first In the next evolution, if there are no forks in the road network 100m ahead of the drone, then the curve arc length is... Use the splines.UnitSpeedAdapter.evaluate() function from the splines library to get the number of... X-axis coordinate of the predicted endpoint position of the next evolution and Y-axis coordinate If the first In the next evolution, if there is a fork in the road network 100m ahead of the drone, then the fork node will be... The X-axis coordinate is used as the predicted endpoint position X-axis coordinate. The fork in the road node The Y-axis coordinate is used as the predicted endpoint position. ; S6-3. The X-axis and Y-axis coordinates of all 100 predicted endpoint positions constitute the predicted trajectory. , .

8. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 7, characterized in that: Step S7 includes the following steps: S7-1. Through formula The occlusion duration factor was calculated. In the formula, For sensitivity parameters, The value ranges from 0.1 to 1; S7-2. Through formula Calculation yields the first Secondary evolution of road network topology disturbance term In the formula, For predicting trajectories The Middle A predicted target location, To obtain the Euclidean distance, To influence the radius parameter; S7-3. Through formula Calculation yields the first Fusion weights during the next evolution ; S7-4. Through formula Calculate the fused first The predicted target location X-axis coordinate Through formula Calculate the fused first The predicted target location Y-axis coordinate ; S7-5. All 100 predicted target locations after fusion constitute a new prediction sequence. , .

9. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 5, characterized in that, Step S8 includes the following steps: S8-1. The new predicted sequence Using video frames The corresponding gimbal pose information is reprojected to pixel coordinates, and a new prediction sequence is selected. Mid-range video frames The pixel X-axis coordinate of the most recent evolution moment of the timestamp and pixel Y-axis coordinate ; S8-2. Calculation Queue The average height of all tracked target boxes is , computation queue The average width of all tracked target boxes is ; S8-3. Constructing the kinematic prediction box after occlusion , ; S8-4. Video frames The input is given to the U2-Net model, and the output is the passable area after occlusion. .

10. The method for anti-pose interference recovery of UAV targets under complex occlusion environment as described in claim 9, characterized in that, Step S9 includes the following steps: S9-1. When the space efficiency constraint is satisfied: Kinematic prediction box after occlusion is determined To effectively predict and execute step S9-2, where The minimum area ratio, The value range is 0-1. Kinematic prediction box after occlusion area, Kinematic prediction box after occlusion Passable areas after obstruction The area of ​​the intersection; S9-2. Video frames The input is fed into a YOLOv11 network, and the output is obtained. The detection box, the first Each detection box is , ; S9-3. When the space efficiency constraint is satisfied: Time judgment Detection box To effectively detect and execute step S9-4, where For the first Detection box area, For the first Detection box Passable areas after obstruction The area of ​​the intersection; S9-4. Kinematic prediction box after occlusion With the Detection box Perform IOU matching; if a match is successful, resume tracking of the specified target.

Images