Unmanned aerial vehicle pose perception and tracking method fusing anisotropic field and view angle constraints
Patent Information
- Application Number
- CN202610690907.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-19
- Publication Date
- 2026-10-16
AI Technical Summary
现有系统缺乏对目标朝向的感知与利用，导致生成的轨迹虽然在物理上无碰撞，但在语义层面往往是不可接受的
（1）本发明通过引入各向异性语义势场，将目标朝向信息融入前端路径搜索，使规划器天然生成绕开目标正面区域的初始路径，显著降低了无人机侵入目标安全空间的概率，提升了跟踪行为的语义合规性。
Smart Images
Figure CN122921162A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of UAV path planning and target tracking technology, specifically involving a UAV pose perception and tracking method that integrates anisotropic fields and viewpoint constraints. It is applicable to dynamic target tracking scenarios that require maintaining a specific observation viewpoint, such as aerial filming, escort monitoring, and human-computer interaction. Background Technology
[0002] With the rapid development of drone technology, its application scenarios have expanded from simple aerial photography and mapping to complex dynamic environment tasks, such as film and television tracking, escort monitoring, and human-machine interaction. In these tasks, dynamic target tracking is one of the core technologies that endows drones with autonomous capabilities.
[0003] Existing mainstream tracking frameworks mainly fall into two categories: one is visual servo-based reactive control, which directly defines errors in image space and maps them to control commands. While computationally efficient, it is prone to getting trapped in local minima and lacks explicit modeling of environmental obstacles, making it difficult to guarantee flight safety in complex environments. The other is geometrically constrained trajectory planning, which generates smooth, dynamically feasible, and collision-free trajectories through a hierarchical architecture of front-end path search and back-end trajectory optimization. However, these methods typically simplify the target to a point mass or sphere, focusing optimization primarily on energy or time optimization, neglecting the crucial semantic information of target orientation.
[0004] In practical applications, simply "seeing the target" is often insufficient to meet task requirements. For example, in aerial filming, cinematographers often seek specific side angles for better expressiveness; in human-computer interaction scenarios, if a drone, in pursuit of the shortest path, passes directly in front of the target, it intrudes into the target's safe space, causing discomfort or even panic. Existing systems lack the perception and utilization of target orientation, resulting in generated trajectories that, while physically collision-free, are often semantically unacceptable.
[0005] Therefore, there is an urgent need for a drone tracking method that can sense the target orientation and incorporate semantic constraints into the planning process, in order to improve the semantic compliance of tracking behavior and the accuracy of task execution. Summary of the Invention
[0006] In view of this, the present invention provides a UAV pose perception and tracking method that integrates anisotropic fields and viewpoint constraints. By introducing target orientation information into the front-end path search and back-end trajectory optimization, it achieves detour behavior that conforms to human social habits and precise locking of the desired viewpoint.
[0007] To achieve the above objectives, the technical solution of the present invention is as follows: 1. A method for UAV pose perception and tracking that integrates anisotropic fields and viewpoint constraints, characterized by comprising the following steps:
[0008] Step 1: Perform kinematic coupling prediction filtering estimation on the target state to obtain full state prediction information including the target position and orientation.
[0009] Step 2: Construct a front-end path search algorithm based on anisotropic semantic potential field. By applying a continuously differentiable soft penalty to the front region of the target, the search is guided to generate an initial path that bypasses the front region.
[0010] Step 3: Based on the MINCO trajectory optimization framework, construct a backend optimization module with relative viewpoint constraints, derive the analytical gradient of viewpoint consistency cost with respect to the UAV position, inject viewpoint constraints into the unconstrained optimization problem, and achieve joint optimization of flight smoothness, obstacle avoidance safety and desired viewpoint.
[0011] Step 4: Implement the collaborative iteration of front-end path search and back-end trajectory optimization through a hierarchical planning framework, and output a UAV tracking trajectory that meets the requirements of dynamic feasibility, obstacle avoidance safety and viewpoint consistency.
[0012] Furthermore, in step 1, kinematic coupling prediction filtering estimation of the target state is performed in the following manner: An Extended Kalman Filter (EKF) is used to filter and estimate the target's position and velocity. Utilizing the non-holonomic constraint assumption—that the orientation of a fast-moving pedestrian is consistent with the direction of their velocity vector—the future orientation sequence is dynamically extrapolated by calculating the tangential vector of the predicted trajectory, outputting a sequence that includes the target's position. ,speed and orientation The full-state prediction window.
[0013] Furthermore, in step 2, the specific process of constructing the front-end path search algorithm based on anisotropic semantic potential field is as follows: Step 2.1: Construct an anisotropic semantic cost field based on target orientation information. Define a fan-shaped danger cone region around the target, and apply semantic penalties to spatial points within the region. As a semantic penalty item.
[0014] Step 2.2: Improve the cost function of the hybrid A* search algorithm, in the total cost Introducing semantic penalty terms This forces the search tree to actively avoid the area directly in front of the target when expanding nodes; among them, For nodes The total cost; From the starting point to the node The cumulative cost; For nodes The heuristic cost to the endpoint; This represents the semantic penalty weight.
[0015] Step 2.3: Output a collision-free polyline path with lateral detour or rear-following characteristics in the topology, as the initial solution for backend optimization.
[0016] Furthermore, the anisotropic semantic cost field The construction method is as follows: Let the target location be The orientation vector is For any node coordinate in space Define its displacement vector relative to the target. .
[0017] Define the danger cone area To meet And the included angle is less than The fan-shaped area, in which For safety, the radius is agreed to be 3m. The danger cone angle is defined as 120°.
[0018] The included angle is defined as:
[0019] The danger cone area is defined as:
[0020] Semantic value Defined as:
[0021] in, These are the coordinates of the spatial nodes; The target's current position; The target is oriented towards the unit vector; The relative displacement vector of the nodes; This represents the distance from the node to the target. for and The included angle; For safe distance radius; The angle of the dangerous cone; The semantic penalty coefficient is 0.5, and this formula makes the cost increase non-linearly as the distance decreases and the angle approaches directly in front.
[0022] Furthermore, in step 3, the backend optimization module for constructing relative viewpoint constraints based on the MINCO trajectory optimization framework specifically includes: Step 5.1: Use the MINCO trajectory class to generate the trajectory. Parameterized into a multi-segment polynomial, derived from spatial waypoints and time allocation parameters Uniquely certain; Step 5.2: Construct a composite objective function ,in As a cost of smoothness, As a result of the collision, The cost of maintaining consistency of perspective; The weighting coefficients for the smoothness cost. The weighting coefficients for collision costs. The weighting coefficient represents the cost of perspective consistency.
[0023] Step 5.3: Define the cost of perspective consistency The relative displacement vector between the UAV and the target With the expected observation vector The integral of the cosine similarity error:
[0024] Step 5.4, Derivation Regarding the location of the drone Analytical gradient:
[0025] The regularized cosine term is:
[0026] in, As a cost of smoothness; The cost of the obstacle avoidance corridor; The cost of maintaining consistency of perspective; , These are the weighting coefficients; This represents the total duration of the trajectory. The unit vector is the desired observation direction; This is a regularization term.
[0027] Step 5.5: Backpropagate the view gradient to MINCO's waypoints. and time This enables the generation of smooth trajectories that are both dynamically feasible and viewpoint-locked.
[0028] Furthermore, the cost of perspective consistency For time parameters The gradient is:
[0029] in, For the first Time period parameters; For the first End of segment; For the discrete integration time; for The drone's velocity vector at any given moment; For the integration step size, This represents the difference between coordinate points.
[0030] Furthermore, in step 4, the collaborative iterative method of the hierarchical planning framework is as follows: Step 4.1: The front-end search generates a set of semantic guidance points based on the target prediction state. :
[0031] in For the desired tracking distance, The desired relative yaw angle; Let z be the rotation matrix about the z-axis. To search for the coordinates of the target point, This refers to the yaw angle of the UAV at the target point.
[0032] Step 4.2: Backend optimization. Based on the initial path provided by the frontend, perform spatiotemporal joint optimization.
[0033] Step 4.3: When environmental changes or target maneuvers cause the current trajectory to no longer meet safety or view constraints, trigger the front-end to re-search, i.e. return to step 4.1, forming a closed-loop iterative mechanism.
[0034] Furthermore, the expected observation vector Set to lateral follow mode according to task requirements, that is Orient the drone perpendicular to the target, keeping it in the area to the side of the target.
[0035] Furthermore, the method is applied to a quadcopter drone platform, which acquires environmental point cloud and target detection information through an onboard depth camera or radar, and realizes closed-loop operation of real-time perception, planning and control in the ROS robot operating system environment.
[0036] Beneficial effects: (1) By introducing an anisotropic semantic potential field, the present invention integrates target orientation information into the front-end path search, enabling the planner to naturally generate an initial path that bypasses the front area of the target, which significantly reduces the probability of UAVs intruding into the target's safe space and improves the semantic compliance of tracking behavior.
[0037] (2) Based on the MINCO framework, this invention constructs a back-end optimization module with relative viewpoint constraints, derives the analytical gradient of viewpoint consistency cost, and realizes the joint optimization of flight smoothness, obstacle avoidance safety and desired viewpoint. It can accurately lock specific observation angles (such as side tracking shots) in high-speed motion, and meet the requirements of tasks such as film and television aerial photography.
[0038] (3) The present invention adopts a hierarchical planning architecture. The front end provides a semantically reasonable initial path, and the back end performs refined spatiotemporal optimization. The two work together to iterate, which not only avoids the single optimization layer from getting stuck in local minima, but also ensures the dynamic feasibility of the trajectory, and improves the robustness and success rate of the system in complex dynamic environments. Attached Figure Description
[0039] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart of the front-end path search based on anisotropic semantic potential field in this invention; Figure 3 This is a flowchart illustrating the gradient derivation process for the backend optimization of relative perspective constraints in this invention. Detailed Implementation
[0040] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0041] 1. A UAV pose perception and tracking method integrating anisotropic fields and view constraints, characterized by comprising the following steps: Step 1: Perform kinematic coupling prediction filtering estimation on the target state to obtain full-state prediction information including the target position and orientation; In this embodiment of the invention, kinematic coupling prediction filtering estimation on the target state is performed in the following manner: An Extended Kalman Filter (EKF) is used to filter and estimate the target's position and velocity. Utilizing the non-holonomic constraint assumption—that the orientation of a fast-moving pedestrian is consistent with the direction of their velocity vector—the future orientation sequence is dynamically extrapolated by calculating the tangential vector of the predicted trajectory, outputting a sequence that includes the target's position. ,speed and orientation unit vector Full-state prediction window .
[0042] The state is defined as:
[0043] The orientation unit vector is defined as:
[0044] Discrete prediction time is defined as:
[0045] in, The target state vector; For the target location; For target speed; Target yaw angle; The target is oriented towards the unit vector; It is a numerically stable term; To predict the sampling period, it is agreed to be 0.1s; To predict the number of steps, a step length of 30 is agreed upon.
[0046] Step 2: Construct a front-end path search algorithm based on anisotropic semantic potential fields. By applying a continuously differentiable soft penalty to the front region of the target, the search is guided to generate an initial path that bypasses the front region. The specific process of constructing the front-end path search algorithm based on anisotropic semantic potential fields is as follows: Step 2.1: Construct an anisotropic semantic cost field based on target orientation information. Define a fan-shaped danger cone region around the target, and apply semantic penalties to spatial points within the region. As a semantic penalty term; anisotropic semantic cost field The construction method is as follows: Let the target location be The orientation vector is For any node coordinate in space Define its displacement vector relative to the target. ; Define the danger cone area To meet And the included angle is less than The fan-shaped area, in which For safety, the radius is agreed to be 3m. The danger cone angle is defined as 120°.
[0047] The included angle is defined as:
[0048] The danger cone area is defined as:
[0049] Semantic value Defined as:
[0050] in, These are the coordinates of the spatial nodes; The target's current position; The target is oriented towards the unit vector; The relative displacement vector of the nodes; This represents the distance from the node to the target. for and The included angle; For safe distance radius; The angle of the dangerous cone; The semantic penalty coefficient is 0.5, and this formula makes the cost increase non-linearly as the distance decreases and the angle approaches directly in front.
[0051] Step 2.2: Improve the cost function of the hybrid A* search algorithm, in the total cost Introducing semantic penalty terms This forces the search tree to actively avoid the area in front of the target when expanding nodes; in, For nodes The total cost; From the starting point to the node The cumulative cost; For nodes The heuristic cost to the endpoint; This represents the semantic penalty weight.
[0052] Step 2.3: Output a collision-free polyline path with lateral detour or rear-following characteristics in the topology, as the initial solution for backend optimization.
[0053] Step 3: Construct a backend optimization module with relative viewpoint constraints based on the MINCO trajectory optimization framework, derive the analytical gradient of the viewpoint consistency cost with respect to the UAV's position, and inject the viewpoint constraints into the unconstrained optimization problem to achieve joint optimization of flight smoothness, obstacle avoidance safety, and desired viewpoint; In this embodiment of the invention, the backend optimization module with relative viewpoint constraints based on the MINCO trajectory optimization framework specifically includes: Step 5.1: Use the MINCO trajectory class to generate the trajectory. Parameterized into a multi-segment polynomial, derived from spatial waypoints and time allocation parameters Uniquely certain; Step 5.2: Construct a composite objective function ,in As a cost of smoothness, As a result of the collision, The cost of maintaining consistency of perspective; Step 5.3: Define the cost of perspective consistency The relative displacement vector between the UAV and the target With the expected observation vector The integral of the cosine similarity error:
[0054] Step 5.4, Derivation Regarding the location of the drone Analytical gradient:
[0055] The regularized cosine term is:
[0056] in, As a cost of smoothness; The cost of the obstacle avoidance corridor; The cost of maintaining consistency of perspective; , These are the weighting coefficients; d_{des}(t) represents the total trajectory duration; d_{des}(t) represents the unit vector of the desired observation direction. This is a regularization term.
[0057] Cost of Perspective Consistency For time parameters The gradient is:
[0058] in, For the first Time period parameters; For the first End of segment; For the discrete integration time; The velocity vector of the UAV; For the integration step size, This represents the difference between coordinate points.
[0059] Step 5.5: Backpropagate the view gradient to MINCO's waypoints. and time This enables the generation of smooth trajectories that are both dynamically feasible and viewpoint-locked.
[0060] Step 4: Implement the collaborative iteration of front-end path search and back-end trajectory optimization through a hierarchical planning framework, and output a UAV tracking trajectory that meets the requirements of dynamic feasibility, obstacle avoidance safety and viewpoint consistency.
[0061] The collaborative iterative method of the hierarchical planning framework is as follows: Step 4.1: The front-end search generates a set of semantic guidance points based on the target prediction state. :
[0062] in For the desired tracking distance, The desired relative yaw angle; Let z be the rotation matrix about the z-axis. To search for the coordinates of the target point, This refers to the yaw angle of the UAV at the target point.
[0063] Step 4.2: Backend optimization involves performing spatiotemporal joint optimization based on the initial path provided by the frontend. Step 4.3: When environmental changes or target maneuvers cause the current trajectory to no longer meet safety or view constraints, trigger the front-end to re-search, i.e. return to step 4.1, forming a closed-loop iterative mechanism.
[0064] Desired observation vector Set to lateral follow mode according to task requirements, that is Orient the drone perpendicular to the target, keeping it in the area to the side of the target.
[0065] The method provided by this invention is applied to a quadcopter unmanned aerial vehicle (UAV) platform. It acquires environmental point cloud and target detection information through an onboard depth camera or radar, and achieves closed-loop operation of real-time perception, planning, and control within the ROS robot operating system environment. Specific Example 1: Construction of a Simulation Verification Platform To verify the effectiveness of the method of this invention, a high-fidelity simulation platform based on ROS Noetic was built under the Ubuntu 20.04 operating system. The simulation environment integrates the Gazebo dynamics engine and the Rviz visualization interface, which can simulate the nonlinear dynamic characteristics of a quadcopter UAV, the field of view limitations of the depth camera (90° horizontal, 60° vertical), and ranging noise (±0.05m). Several static obstacles (cylinders, walls) are randomly placed in the environment, and the dynamic movement of pedestrians is simulated using the Gazebo Actor plugin. The pedestrian movement speed ranges from 1.0 to 2.0 m / s, and typical maneuvers such as straight walking and sharp turns can be performed. The simulation parameters of the UAV platform are: maximum speed 3.0 m / s, maximum acceleration 2.0 m / s², and rotational inertia configured according to the standard quadcopter configuration. The airborne sensor simulates the point cloud output of the depth camera for real-time mapping and obstacle detection. The target detection module provides observation data of pedestrian position and orientation, with observation noise set to 0.1m for position and 5° for orientation.
[0066] Target state kinematic coupling prediction (corresponding to step 1) In the simulation, the target state kinematic coupling prediction, after receiving observation data from the target detection module, uses an extended Kalman filter to estimate the target state. The state vector includes position... ,speed A constant velocity motion model is adopted. To obtain the target orientation, a non-holonomic constraint assumption is used: when the target velocity amplitude is greater than a threshold... At that time, it is assumed that the target's orientation is consistent with the velocity direction, that is... When the speed is low, maintain the orientation estimate from the previous moment. Extrapolate the future trajectory using the difference vector of the predicted trajectory. The orientation sequence of steps (0.2s) forms a full-state prediction window, providing forward-looking information for subsequent planning modules.
[0067] Anisotropic semantic front-end path search (corresponding to step 2) Anisotropic semantic front-end path search constructs anisotropic semantic cost field based on the target prediction state. A safe distance radius is set. Danger cone angle Semantic penalty coefficient In a raster map, for each node Calculate semantic value This is then superimposed onto the cost function of the hybrid A* search. The heuristic function of the hybrid A* search employs a path length estimate that considers dynamic constraints and semantic penalty weights. When the target moves in a straight line, the initial path generated by the front-end search naturally detours to the side of the target, avoiding direct passage through the frontal area; when the target suddenly turns, the front-end replans based on the predicted state, generating a new detour path, providing a reasonable initial solution for subsequent optimization.
[0068] Relative perspective constraint back-end trajectory optimization (corresponding to step 3) Use the polyline path generated by the front end as the initial waypoint of the MINCO trajectory. Time allocation Initialize uniformly in the predicted time domain. Construct a composite objective function. The smoothness cost Using the Minimum Jerk form, the collision cost Based on Euclidean distance field construction, viewpoint consistency cost Defined using formula (3). The weight parameters are set as follows: Time-varying weights of perspective cost Set to a constant of 1.0. Desired observation vector. Set to a lateral direction perpendicular to the target's orientation, i.e. The specific choice between left and right can be determined based on the initial position. This is achieved using the derived analytical gradient. Waypoints were determined using the L-BFGS algorithm. and time Optimization is performed. During the optimization process, calculations are also performed simultaneously. Adjust the time allocation to match the target motion.
[0069] Simulation Experiments and Result Analysis In the same simulation environment, the method of this invention was compared with the existing technology Elastic Tracker. The target motion trajectory included straight lines, sharp turns, and random obstacles. The following metrics were used for quantitative evaluation: - Forward Intrusion Rate (FIR): The total time a drone spends in front of a target (distance <2.0m, angle ±30°).
[0070] - Mean relative heading error (MBE): The average absolute deviation between the actual observed angle and the expected lateral angle.
[0071] - Expected View Hold Rate (SVR): The percentage of time the drone spends in the expected lateral area (angle ±45°).
[0072] - Average distance error (ADE): The average deviation between the actual tracking distance and the expected distance (2.0m). The experimental results are shown in Table 1.
[0073] Table 1 Performance comparison between the method of this invention and Elastic Tracker
[0074] As shown in Table 1, the method of this invention is slightly better than the comparative method in terms of average distance error (ADE) (0.165m vs 0.182m), while reducing the average relative heading error (MBE) by 72%, improving the expected view hold rate (SVR) by nearly 5 times, and reducing the forward intrusion time (SVD) from 3.20 seconds to 0.90 seconds, which verifies the advantages of this method in terms of semantic compliance and view hold.
[0075] System integration and real-time analysis The above modules are encapsulated as ROS nodes, and their states are managed and tracked using a finite state machine. In closed-loop operation within a simulation environment, the average planning time per frame (including front-end search and back-end optimization) is 45ms, meeting real-time requirements (control cycle 50ms). The system can quickly replan when the target suddenly turns or the environment changes, maintaining safety and viewpoint lock.
[0076] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A UAV pose perception and tracking method integrating anisotropic fields and viewpoint constraints, characterized in that, Includes the following steps: Step 1: Perform kinematic coupling prediction filtering estimation on the target state to obtain full-state prediction information including the target position and orientation; Step 2: Construct a front-end path search algorithm based on anisotropic semantic potential field. By applying a continuously differentiable soft penalty to the front region of the target, the search is guided to generate an initial path that bypasses the front region. Step 3: Based on the MINCO trajectory optimization framework, construct a backend optimization module with relative viewpoint constraints, derive the analytical gradient of viewpoint consistency cost with respect to the UAV position, inject viewpoint constraints into the unconstrained optimization problem, and achieve joint optimization of flight smoothness, obstacle avoidance safety and desired viewpoint. Step 4: Implement the collaborative iteration of front-end path search and back-end trajectory optimization through a hierarchical planning framework, and output a UAV tracking trajectory that meets the requirements of dynamic feasibility, obstacle avoidance safety and viewpoint consistency.
2. The UAV pose perception and tracking method integrating anisotropic fields and view constraints as described in claim 1, characterized in that, In step 1, the kinematic coupling prediction filter estimation for the target state is performed in the following manner: An Extended Kalman Filter (EKF) is used to filter and estimate the target's position and velocity. Utilizing the non-holonomic constraint assumption—that the orientation of a fast-moving pedestrian is consistent with the direction of their velocity vector—the future orientation sequence is dynamically extrapolated by calculating the tangential vector of the predicted trajectory, outputting a sequence that includes the target's position. ,speed and orientation The full-state prediction window.
3. The UAV pose perception and tracking method integrating anisotropic fields and view constraints as described in claim 1, characterized in that, In step 2, the specific process of constructing the front-end path search algorithm based on anisotropic semantic potential field is as follows: Step 2.1: Construct an anisotropic semantic cost field based on target orientation information. Define a fan-shaped danger cone region around the target, and apply semantic penalties to spatial points within the region. As a semantic penalty item; Step 2.2: Improve the cost function of the hybrid A* search algorithm, in the total cost Introducing semantic penalty terms This forces the search tree to actively avoid the area directly in front of the target when expanding nodes; among them, For nodes The total cost; From the starting point to the node The cumulative cost; For nodes The heuristic cost to the endpoint; This represents the semantic penalty weight. Step 2.3: Output a collision-free polyline path with lateral detour or rear-following characteristics in the topology, as the initial solution for backend optimization.
4. The UAV pose perception and tracking method integrating anisotropic fields and view constraints as described in claim 3, characterized in that, The anisotropic semantic cost field The construction method is as follows: Let the target location be The orientation vector is For any node coordinate in space Define its displacement vector relative to the target. ; Define the danger cone area To meet And the included angle is less than The fan-shaped area, in which For safety, the radius is agreed to be 3m. The angle of the danger cone is defined as 120°. The included angle is defined as: The danger cone area is defined as: Semantic value Defined as: in, These are the coordinates of the spatial nodes; The target's current position; The target is oriented towards the unit vector; The relative displacement vector of the nodes; This represents the distance from the node to the target. for and The included angle; For safe distance radius; The angle of the dangerous cone; The semantic penalty coefficient is 0.5, and this formula makes the cost increase non-linearly as the distance decreases and the angle approaches directly in front.
5. The UAV pose perception and tracking method integrating anisotropic fields and view constraints as described in claim 1, characterized in that, In step 3, the backend optimization module for constructing relative viewpoint constraints based on the MINCO trajectory optimization framework specifically includes: Step 5.1: Use the MINCO trajectory class to generate the trajectory. Parameterized into a multi-segment polynomial, from spatial waypoints and time allocation parameters Uniquely certain; Step 5.2: Construct a composite objective function ,in As a cost of smoothness, As a result of the collision, The cost of maintaining consistency of perspective; The weighting coefficients for the smoothness cost. The weighting coefficients for collision costs. The weighting coefficient represents the cost of perspective consistency. Step 5.3: Define the cost of view consistency The relative displacement vector between the UAV and the target With the expected observation vector The integral of the cosine similarity error: Step 5.4, Derivation Regarding the location of the drone Analytical gradient: The regularized cosine term is: in, As a cost to smoothness; The cost of the obstacle avoidance corridor; The cost of maintaining consistency of perspective; , These are the weighting coefficients; This represents the total duration of the trajectory. The unit vector is the desired observation direction; It is a regularization term; Step 5.5: Backpropagate the view gradient to MINCO's waypoints. and time This enables the generation of smooth trajectories that are both dynamically feasible and viewpoint-locked.
6. The UAV pose perception and tracking method integrating anisotropic fields and view constraints as described in claim 5, characterized in that, The cost of perspective consistency For time parameters The gradient is: in, For the first Time period parameters; For the first End of segment; For the discrete integration time; for The drone's velocity vector at any given moment; The integration step size is... This represents the difference between coordinate points.
7. The UAV pose perception and tracking method integrating anisotropic fields and view constraints as described in claim 1, characterized in that, In step 4, the collaborative iterative method of the hierarchical planning framework is as follows: Step 4.1: The front-end search generates a set of semantic guidance points based on the target prediction state. : in For the desired tracking distance, The desired relative yaw angle; Let z be the rotation matrix about the z-axis. To search for the coordinates of the target point, This refers to the yaw angle of the UAV at the target point. Step 4.2: Backend optimization involves performing spatiotemporal joint optimization based on the initial path provided by the frontend. Step 4.3: When environmental changes or target maneuvers cause the current trajectory to no longer meet safety or view constraints, trigger the front-end to re-search, i.e. return to step 4.1, forming a closed-loop iterative mechanism.
8. The UAV pose perception and tracking method integrating anisotropic fields and view constraints as described in claim 1, characterized in that, The desired observation vector Set to lateral follow mode according to task requirements, that is Orient the drone perpendicular to the target, keeping it in the area to the side of the target.
9. The UAV pose perception and tracking method integrating anisotropic fields and view constraints as described in claim 1, characterized in that, The method is applied to a quadcopter drone platform, which acquires environmental point cloud and target detection information through an onboard depth camera or radar, and realizes closed-loop operation of real-time perception, planning and control in the ROS robot operating system environment.