Multi-unmanned aerial vehicle cooperative path planning method for multi-target tracking

Abstract

The invention discloses a multi-unmanned aerial vehicle cooperative path planning method for multi-target tracking, and relates to the technical field of unmanned aerial vehicle path planning. The method has the advantages that a multi-unmanned-aerial-vehicle multi-target tracking problem is decomposed into a target state estimation problem, a task allocation problem and a motion planning problem so as to simplify a solution of the multi-unmanned-aerial-vehicle multi-target tracking problem, and a real-time task allocation method adopts a secondary allocation method to allocate a reasonable initial target set for an unmanned aerial vehicle group; then, selecting a current tracking target by adopting an effective tracking strategy, and meanwhile, avoiding tracking the same target through communication, so that the unmanned aerial vehicle group cooperatively executes a multi-target task; an optimization-based method is utilized to generate a smooth trajectory by considering optimization indexes such as smoothness, dynamic feasibility, interference resistance and the like, and meanwhile, the required safe distance between all interested objects is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) path planning, and in particular to a multi-UAV collaborative path planning method for multi-target tracking. Background Art

[0002] Tracking animals and understanding their habits is important for the protection of endangered animals and biodiversity conservation. In the past, scientists have used various methods to achieve tracking of animal targets, such as vision-based sensors, GPS-based tags, VHF-based tags, etc. Despite all the emerging technologies, VHF-based tags are still widely used and play an important role today. This is due to its small size, light weight, and long operating time, making it suitable for tracking most animals. However, traditional radio tracking methods require technicians to carry bulky radio receivers to obtain the signals sent by animal radio tags, which is undoubtedly a time-consuming and labor-intensive task. In addition, traditional methods have great difficulties in tracking animals that move quickly and in real time in special environments (for example, underwater or underground caves). Drones are able to carry payloads such as radio receivers and antennas, which provides a promising solution. Their deployability and high maneuverability provide the potential for achieving automatic target tracking, greatly reducing the time, effort, and expenses associated with traditional tracking methods. In addition, the advancement of software-defined radios (SDRs) is crucial. These devices facilitate rapid reconfiguration and simultaneous reception of a wide range of radio frequencies. SDRs have achieved commercial viability, providing a cost-effective, durable, and lightweight solution. This makes them ideal for integration into drones, enabling the tracking of multiple animals simultaneously.

[0003] There are some groundbreaking results in autonomous drone systems for wildlife tracking. These systems use information such as RSSI or angle of arrival (AOA) of the tag radio signal to locate VHF radio-tagged animals. In addition to improving positioning accuracy and reducing tracking errors, integrating sophisticated trajectory planning algorithms in these drones can enhance autonomy. Despite these technological advances, achieving drones that can autonomously and effectively track wildlife remains a long-term task. There are two unavoidable problems in achieving autonomous target tracking:

[0004] 1) Limited onboard resources: In practical applications, drones need to navigate challenges such as limited battery duration, limited onboard computing power, and limited payload capacity. Aerial robots should be able to fly out, track and locate animals, and return to base autonomously without the need for mid-life battery replacement or human intervention. This highlights the importance of optimizing the operational efficiency of drones in wildlife tracking applications.

[0005] 2) Avoiding disturbance to animals: The operation of drones in wildlife tracking may disturb the animals, which is counterproductive to the mission objectives. Therefore, strategies must be designed to not only ensure that drones are effective in tracking wildlife, but also minimize any negative impact on the animals.

[0006] Planning methods for UAVs can generally be divided into information gain-based planning methods and task-based planning methods. Information gain-based planning methods use the POMDP framework to calculate the optimal action to maximize the reward. The reward function can be calculated by various methods, including task-driven and information-driven strategies. When the uncertainty is high, the information-driven strategy is more suitable for reducing the position uncertainty of the target. Therefore, it is common to use information-driven strategies to calculate the reward function. There are several methods to measure information divergence, such as Shannon entropy, Cauchy-Schwarz divergence, or Renyi divergence. However, information gain-based methods are computationally intensive and are generally not suitable for real-time applications and / or require high-power ground control systems for in-loop calculations.

[0007] Task-based planning methods are lightweight, real-time methods that can be embedded on vehicle-mounted low-cost UAV platforms. These methods show that the main goal of the tracking task is to observe the target efficiently and minimize the uncertainty in the target state estimate. Therefore, a strategy that preferentially tracks the target with the smallest uncertainty in the position estimate can provide an inexpensive solution. Task-based planning strategies often require the selection of control actions. It first tracks the target with the lowest uncertainty and obtains a series of actions while maintaining a safe distance from the target, and then evaluates each action using an evaluation function to obtain the optimal control action. In the methods based on selecting control actions, they need to plan the control actions in advance and evaluate all control actions using an evaluation function in each planning. However, all control actions may not meet the requirements, in which case new control actions need to be reconsidered. Although the evaluation of control actions is computationally simple, the process of designing control actions is more complicated.

[0008] At present, there are relatively more studies on single UAV tracking multiple moving targets, but there are fewer studies on multiple UAVs collaboratively tracking multiple moving targets. Multiple UAVs can effectively improve tracking efficiency and have research value. Summary of the invention

[0009] The purpose of the present invention is to provide a multi-UAV collaborative path planning method for multi-target tracking in view of the shortcomings and defects in the prior art, and to decompose the multi-UAV multi-target tracking problem into a target state estimation problem, a task allocation problem and a motion planning problem to simplify the solution to the multi-UAV multi-target tracking problem. The real-time task allocation method adopts a secondary allocation method to first allocate a reasonable initial target set to the UAV swarm, and then adopts an effective tracking strategy to select the current tracking target, while avoiding tracking the same target through communication, so that the UAV swarm can collaboratively perform multi-target tasks; a trajectory generation method that generates a smooth trajectory by considering optimization indicators such as smoothness, dynamic feasibility and anti-interference is used based on an optimization method, while ensuring that the required safety distance exists between all objects of interest.

[0010] To achieve the above object, the present invention adopts the following technical solution: a multi-UAV collaborative path planning method for multi-target tracking, considering n r Drone The system consists of a t Targets An algorithm is designed to facilitate drones to effectively perform multi-drone target tracking tasks, and a hierarchical framework is used to simplify the multi-target tracking problem in multi-drone systems. The hierarchical framework can be divided into three layers: state estimation layer, real-time task allocation layer, and motion planning layer; the state estimation layer uses simple and fast RSSI measurement to collect and estimate the target position using a particle filter algorithm; the real-time task allocation layer uses a quadratic allocation method to solve the multi-drone task allocation problem in dynamic scenarios; the motion planning layer uses an optimization-based method to obtain a smooth trajectory that the drone can execute while ensuring that the drone can maintain the desired safe distance from all objects of interest. The process mainly involves path planning and trajectory optimization. Initially, the enhanced RRT* algorithm considers the estimated position and safe distance of the target to calculate the initial drone path point, and then the trajectory optimization considers optimization items such as controllability, dynamic feasibility, and interference avoidance to generate the final trajectory; the specific process is as follows: in the state estimation layer, the RSSI signal emitted by the radio tag is first collected using a VHF antenna, and then the drone is The machine can quickly configure the radio and receive radio signals from multiple targets at the same time. The radio signals are processed by the signal detection algorithm. Since each target is uniquely identified by the frequency of the radio transmission, there is no data correlation problem. The target state estimation problem can be solved by running multiple particle filters in parallel; the task allocation layer adopts a quadratic allocation method to solve the multi-UAV task allocation problem in dynamic scenarios. First, the task pre-allocation problem is solved by a mixed integer optimization problem. Based on the known information of the UAV and the target, the optimization problem is constructed by multiple constraints (such as distance constraints and UAV coordination constraints) to maximize the effectiveness of the task. Here, the initial position of the target can be obtained by multi-target area search. By solving this optimization problem, the initial target set is reasonably allocated to the UAV group. Subsequently, in order to more accurately locate and track the moving target, it is necessary to reallocate the appropriate tracking target for each UAV in real time. The basic requirement of task reallocation is to avoid UAVs tracking the same target, which can effectively reduce the waste of UAV onboard resources and effectively complete the problem MCTT. After determining the position of the target to be tracked, the motion planning layer generates an executable trajectory for the UAV. An optimization-based method is used to take smoothness, dynamic feasibility, and interference avoidance as optimization items. The trajectory is obtained by solving the optimization problem, thereby obtaining Algorithm 1.

[0011] Furthermore, the particle filter is specifically as follows: The goal of the filtering problem is to filter the historical observation data z from time t=1 to time t=k 1:k To estimate the confidence level π k (x k ∣z 1:k ), according to the Bayesian recursive idea, from the initial confidence π 0 To begin, the confidence secret is calculated sequentially following the prediction and update steps as follows: Particle filtering implements the random sampling process of the Monte Carlo method by using independent and uniformly distributed (iid) particles The confidence is approximated by a weighted set of , namely: Among them, δ(·) represents the Dirac function, and the particle weight satisfies the condition The goal of motion planning is determined by using the uncertainty of the position estimate represented by the filter confidence. Although the E-optimal uncertainty estimate can be obtained by using the maximum eigenvalue of the complete covariance matrix of the particle, a computationally less expensive method is used to calculate the maximum covariance of the estimated position along each coordinate axis to represent the position uncertainty, which has similar performance. Let is a particle set The estimated object position, σ (x) represents the standard deviation of the estimated confidence density π(·), then:

[0012]

[0013] Furthermore, Algorithm 1 is designed based on this hierarchical framework, and its specific contents are as follows:

[0014] MCTT problem algorithm;

[0015] Input: UAV initial position set Target initial position set Target radio frequency set Target initial Stdσ 0 , target end Stdσ min , the number of particles N, the target motion model f(x k ∣x k-1 ), observation model g(z k |x k );

[0016] Output: drone location set

[0017] (1) k = 0,

[0018] (2) Through get

[0019] (3) ←Task pre-allocation

[0020] (4)

[0021] (5) k = k + 1;

[0022] (6) for each drone r i do;

[0023] (7) for each target do;

[0024] (8) Obtaining measurement values

[0025] (9) ←State Estimation

[0026] (10) Value Stored in variable Ω k ;

[0027] (11) if σ tj <σ min then;

[0028] (12) Set the target t j from and Removed;

[0029] (13) ←Task Reassignment

[0030] (14)

[0031] (15)

[0032] (16) Else;

[0033] (17) Drone i has completed all tasks and returned to the starting point

[0034] Among them, line (2) is based on the initial position set and RF sets of moving targets generate Drone swarms maintain real-time communication Update, when the drone completes tracking of a specific target, it will Delete the target when When it is empty, it means that all targets have been tracked; line (3) is based on and the initial position set of the UAV The initial task set Pre-assigned to the drone group; Line (8) obtains drone r at time k i At target t j RSSI measurement value on Based on this measurement The target motion model f(x k ∣xk-1 ) and the observation model g(z k |x k ); Line (9) uses particle filtering to filter the target t j Perform state estimation and obtain the target t j Estimated location and confidence level (expressed as the standard deviation (Std) of the estimated position, see Section IV-b for details, and then Stored in variable Ω k , for later use in the reallocation algorithm, when the target's estimated position Less than a certain value σ min When considering the positioning and tracking of target tj, lines (11)-(12) are based on the variable Ω that stores the estimated position and Std of the target at time k. k , total list Initial Task Set Target initial Stdσ 0 and drone tracking target list Line (13) performs task redistribution to obtain UAV r i Current goals and updates to track In The executable trajectory of the target UAV is generated in line (15) based on the position of the tracked target.

[0035] Furthermore, the real-time task allocation layer: first, according to the known initial positions of the drones and the targets, considering the minimum distance of the drone group, the balanced number of targets to be executed and other conditions, all targets in the area are divided into drones using mixed integer programming, and then each drone obtains an initial target set. Next, the drone uses a tracking strategy, which gives priority to tracking objects with the lowest estimated belief uncertainty, where the uncertainty can be represented by the filter confidence density. When the drone is closer to the target, the signal strength received by the drone from the target is greater, and the target position estimated using particle filtering is more accurate. This strategy achieves rapid detection and real-time tracking of radio-tagged targets. At the same time, due to its movement, the target may appear within the trackable range of other drones. At this time, communication between drones is required to avoid drones from tracking the same target at the same time. The real-time task allocation layer is specifically as follows: Pre-allocation: This section establishes the multi-drone task pre-allocation problem as an optimization problem to assign reasonable initial tracking targets to drones so that drones can work more efficiently. Without losing generality, it is assumed that any task can be assigned to any drone. Since each robot can perform at most Ni tasks and a robot is required to perform each task, all tasks to be performed should have The multi-UAV task allocation problem is established as an optimization problem, so that the optimization result with the minimum total cost can be obtained under the task constraints, so that the UAV group can better collaborate as a team and improve work efficiency. Assuming there are n r drones and n t Mission objectives Among them, the status of the drone is is the three-dimensional coordinate of the drone, is the yaw angle of the UAV; the state of the mission target is It is the x, y, and z coordinates of the Cartesian coordinates. Without loss of generality, we assume that any task can be assigned to any UAV, and since each robot can perform at most N i tasks, each task requires a single robot, so for all tasks to be performed, there should be The overall goal of task allocation is to assign all tasks to robots to maximize the total benefit or minimize the total cost; is the task target t j Give UAV i The cost of the task target is determined by the certainty of the estimated position of the target. When allocating tasks, the execution order of the task targets is not considered. Here, a ij Simplified to the distance between the UAV and the mission target on the UAV flight plane, that is Therefore, the optimization goal is the cost required to execute all task goals, expressed as where f ij is the optimization variable, indicating r i Is it assigned to t j , therefore, the multi-robot task allocation problem can be expressed as an optimization problem: st, Here, constraint (1a) indicates that each mission objective must be assigned to only one UAV and each mission must be performed by one robot; constraint (1b) gives the upper limit of the mission objectives that each UAV can perform, reflecting the limited resources of UAVs; constraint (1c) indicates that f ij is a binary variable representing r i Is it assigned to t j; Reassignment: Although the task set has been assigned to the UAV, the target moves in real time. The UAV can identify other targets that the UAV needs to track during its flight. At this time, dynamic task reallocation is triggered to make task execution more efficient. The principle of the dynamic reallocation algorithm is identified as: moving to the target with the least state estimation uncertainty. When the actual estimated position is unreliable, the UAV traveling towards the target can quickly locate the target position. This is because when the UAV approaches the target, the signal strength increases and the signal is less affected by various interferences. Based on this principle, a dynamic task reallocation algorithm is proposed. The algorithm has high real-time performance and can meet the real-time task reallocation of UAVs in dynamic environments.

[0036] Further, the dynamic task reallocation algorithm is described as follows: Each target t in j , from Ω k Get its Std value It represents the uncertainty of the position estimate calculated by the previous particle filter algorithm. Then, if the Std of the target tj is lower than σ 0 , then store it in m temp In this paper, unlike the existing methods that select the estimated position with the lowest uncertainty among all targets as the target, the estimated position of the target is only considered when the target Std is below a predetermined threshold. The rationale is that the initial position estimates from particle filtering exhibit high uncertainty and significant changes over time, and using these estimates as direct destinations may result in a large number of trajectory changes and extended tracking duration; removing m temp The target is being tracked by other drones ( except) to avoid the drone tracking the same target; if m temp If there is a target, the target with the smallest Std is selected for tracking; if m temp There is no target and the drone is not already on the pre-assigned target list If the tracking target is completed in Select the nearest target to track; otherwise, is empty, which means the drone has no goals to execute and can return to its starting point. Finally, Update

[0037] Furthermore, the motion planning layer is specifically as follows: when the actual position estimate is unreliable, the drone navigating to the target can quickly locate the target position because when the drone is close, the uncertainty of the target signal measurement is low, which can be known from the distance tracking achieved by measuring the signal strength. Therefore, in the tracking task with the measurement object as the target, the planning strategy of moving to the target with the lowest position estimation uncertainty with the goal of achieving a sufficiently small position estimation uncertainty may provide a cheap planning strategy; after determining the target position to be tracked according to the position estimation uncertainty, it is necessary to generate a trajectory. The basic requirements for the trajectory required by the drone are smoothness, dynamic feasibility, mutual avoidance between drones, etc. By adopting Σ MINCO Trajectory to solve the above problem: The trajectory generation of the drone is expressed as an unconstrained optimization problem: Where λ is the weight vector that weighs each cost function; the control force J e :Control force optimization item J e It is expressed by integrating the squared norm of the third-order derivative of the trajectory over time and written as 2) Total time J t : In order to improve the efficiency of the drone swarm in executing tasks, it is hoped that the drone swarm can reach the target point as quickly as possible, thereby minimizing the total time: Obstacle Avoidance o : Use the Euclidean Signed Distance Field (ESDF) to establish the environment around the drone, and ESDF can obtain the distance between the current position of the drone and the nearest obstacle. In order to ensure safety requirements, a collision penalty item is defined. It is triggered when the distance to the obstacle is less than the safety gap. The collision penalty item is as follows: Among them, d thr is the safety threshold, The distance between the point and the nearest obstacle around it is considered, and then the obstacle avoidance optimization term J is obtained by calculating the weighted sum of the sampling constraint function o : in is the orthogonal coefficient that follows the trapezoidal rule; dynamic feasibility J d : Since the UAV needs to be constrained by its own motors and other actuators when flying, the speed and acceleration of the UAV have maximum values. The dynamic feasibility constraint term J is used d Limit the maximum values ​​of velocity and acceleration to ensure that the trajectory can be executed by the drone: J d =J v +J a #, Among them, v m ,a m is the maximum speed and acceleration allowed; anti-interference J i: To maintain a safe distance r between the UAV and the target min , an anti-interference optimization term is constructed. Let V(u, r min ) denote the anti-interference region of the target state (position) based on the UAV state u, and it is given as: Then, the interruption avoidance penalty Ji is obtained by calculating the weighted sum of the sampling constraint functions as: The gradient of Ji with respect to ci and Ti is given by the following formula where t is the relative time on the trajectory. For the case where d(x, u) < dmin, the gradient is given by the following formula: where is the gradient of the Euclidean signed distance field (ESDF) in, otherwise, the gradient becomes

[0038] A multi-UAV cooperative path planning method for multi-target tracking, characterized in that it includes a target model, a UAV model and an observation model. Target model: For wild animal targets, their dynamic behaviors are usually unpredictable. Therefore, their behaviors are modeled as a random walk model where represents the state of the target at the k-th moment; represents the Gaussian distribution with mean μ and covariance Q; is the 3×3 covariance matrix of the process noise, where In represents the n×n-dimensional matrix. UAV model: For UAVs, the concept of multi-helicopter differential flatness has been widely explored and it has been proven that it has a significant and physically interpretable flat output space. All states of the UAV, including position, attitude, velocity and angular velocity, as well as inputs such as thrust acceleration and torque, can be represented by the flat output (Z) and its corresponding derivatives. It should be noted that the flat output space overlaps with the original configuration space, and its basic structure of the flat output is consistent with where (p x , p y , p z ) T represents the translation of the UAV's center of gravity, ψ is the yaw angle. By ensuring that the trajectory obeys a set of specific constraints in the flat space, it can be effectively transformed into the original configuration space using the flat transformation. Let represent the state of the UAV at the k-th moment, where is the 3D coordinate of the UAV, is the yaw angle of the drone. Here, the drone adopts a fixed flight altitude and focuses on the horizontal control of the drone. This method avoids adjusting the altitude during trajectory calculation, thereby effectively reducing the amount of calculation of the drone's onboard computer; Observation model: The observation model is a real signal propagation model used to estimate the target state x k and the drone status u k RSSI measurement between, which plays an important role in state estimation using particle filters. The observation model is used using the logarithmic distance path loss model. The power h(x k ,u k )(dBm) contains only the line-of-sight power component, h(x k ,u k )=P 0 -10L c log(d(x k ,u k ))+G r (x k ,u k ), where P 0 (dBm) indicates reference power; L c Represents the unitless path loss constant, which characterizes the signal attenuation as the transmission distance increases, and usually ranges from 2 to 4; is the distance between the target and the UAV; G r (x k ,u k ) represents the directional antenna gain, which depends on the yaw angle of the drone and its relationship with the target x k The relative position of the RSSI value zk (dBm) is easily corrupted by noise, including thermal noise and interference from other sources. Given the assumption of white noise, the measurement likelihood model can be expressed as: Among them, Q (z) is the 1×1 covariance matrix of the measurement noise.

[0039] After adopting the above technical solution, the beneficial effects of the present invention are: the present invention has at least the following advantages:

[0040] 1) The multi-UAV multi-target tracking problem is decomposed into target state estimation problem, task allocation problem and motion planning problem to simplify the solution of the multi-UAV multi-target tracking problem.

[0041] 2) The real-time task allocation method adopts the secondary allocation method to first allocate a reasonable initial target set for the UAV swarm, and then adopts an effective tracking strategy to select the current tracking target. At the same time, communication is used to avoid tracking the same target, so that the UAV swarm can collaboratively perform multi-target tasks.

[0042] 3) Trajectory generation methods that utilize optimization-based approaches to generate smooth trajectories by considering optimization metrics such as smoothness, dynamic feasibility, and anti-interference, while ensuring the required safety distances between all objects of interest. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0044] Figure 1 It is a hierarchical framework diagram in the present invention.

[0045] Figure 2 It is a schematic diagram of the positioning results of 10 mobile targets in the present invention.

[0046] Figure 3 yes Figure 2 Enlarged structural diagram at A in the middle.

[0047] Figure 4 It is a schematic diagram of the time evolution Std value of each target tracked by the drone in scenario 1 of the present invention.

[0048] Figure 5 It is a schematic diagram of the change of the RMS value of each tracked target in the utility model.

[0049] Figure 6 This is a result diagram of four types of UAVs tracking and locating 10 moving targets in the present invention.

[0050] Figure 7 Schematic diagram of Std values ​​of targets tracked by four UAVs over time in the present invention.

[0051] Figure 8 This is a schematic diagram of the RMS value of the four UAV positioning targets changing over time in the present invention. Figure 1 .

[0052] Fig. 9 This is a schematic diagram of the RMS value of the four UAV positioning targets changing over time in the present invention. Figure 2 .

[0053] Fig.10 This is a schematic diagram of the RMS value of the four UAV positioning targets changing over time in the present invention. Figure 3 .

[0054] Fig.11This is a schematic diagram of the RMS value of the four UAV positioning targets changing over time in the present invention. Figure 4 . DETAILED DESCRIPTION

[0055] A multi-UAV collaborative path planning method for multi-target tracking, characterized in that it includes a target model, a UAV model and an observation model:

[0056] Target model:

[0057] For wildlife targets, their dynamic behaviors are usually unpredictable, so their behaviors are modeled as random walk models:

[0058]

[0059] in, represents the state of the target at the kth moment; represents a Gaussian distribution with mean μ and covariance Q; is the 3×3 covariance matrix of process noise, where In represents an n×n dimensional matrix;

[0060] Drone Model:

[0061] For UAVs, the concept of multicopter differential flatness has been extensively explored and demonstrated to have a remarkably, physically interpretable flat output space, establishing that all states of the UAV, including position, attitude, velocity, and angular velocity, as well as inputs such as thrust acceleration and torque, can be represented by a flat output (Z) and its corresponding derivatives. It is noteworthy that the flat output space overlaps with the original configuration space, and its basic structure of flat outputs is similar to Consistent, where (p x ,p y ,p z ) T represents the translation of the center of gravity of the drone, ψ is the yaw angle, and by ensuring that the trajectory obeys a certain set of constraints in the flat space, it can be efficiently converted to the original configuration space using a planar transformation, let represents the state of the drone at the kth moment, where is the 3D coordinate of the drone, is the yaw angle of the drone. Here, the drone adopts a fixed flight altitude and focuses on the horizontal control of the drone. This method avoids adjusting the altitude during the trajectory calculation process, thereby effectively reducing the amount of calculation of the drone's onboard computer;

[0062] Observation Model:

[0063] The observation model is a real signal propagation model used to estimate the target state x kand the drone status u k RSSI measurement between, which plays an important role in state estimation using particle filters. The observation model is used using the logarithmic distance path loss model. The power h(x k ,u k )(dBm) contains only the line-of-sight power component:

[0064] h(x k ,u k )=P 0 -10L c log(d(x k ,u k ))+G r (x k ,u k ),

[0065] Among them, P 0 (dBm) indicates reference power; L c Represents the unitless path loss constant, which characterizes the signal attenuation as the transmission distance increases, and usually ranges from 2 to 4; is the distance between the target and the UAV; G r (x k ,u k ) represents the directional antenna gain, which depends on the yaw angle of the drone and its relationship with the target x k The relative position of

[0066] The RSSI value zk (dBm) is easily corrupted by noise, including thermal noise and interference from other sources. Given the assumption of white noise, the measurement likelihood model can be expressed as:

[0067]

[0068] Among them, Q (z) is the 1×1 covariance matrix of the measurement noise.

[0069] Wild animals (or mobile targets) are attached to VHF radio tags, each of which is identified by a unique frequency. These radio collar tags are available per T 0 A low power on-off keying signal is periodically transmitted and detected by the sensor system on the drone. Based on the strength of the received radio signal, the drone updates the position of the moving target using a particle filter algorithm, and an executable trajectory to the target location is generated by a motion planning algorithm.

[0070] Here, several assumptions are introduced to help define the research problem of this paper:

[0071] (1) Since this paper focuses on the tracking and positioning of the detected object rather than the signal detection problem, it is assumed that there is no missed detection or false detection.

[0072] (2) This paper assumes that the UAV can detect all targets, which is reasonable in a medium-sized search area; this paper focuses on improving the tracking performance of detected targets.

[0073] (3) It is assumed that the total number of targets is known and that no target moves out of the measurement area during the positioning and tracking process. Therefore, the tracking and localization task is terminated after all objects are tracked and located.

[0074] (4) Assuming that the starting positions of all targets are known, the starting positions can be obtained through multi-UAV target search. This paper does not consider the exploration problem.

[0075] (5) In addition, to minimize interference with wildlife, drones must maintain a safe distance when navigating to track and locate wildlife. Although drones can be flown at high altitudes (e.g., above 100 meters) to minimize interference with objects of interest, it is also recommended to avoid flying over wildlife

[16] . In addition, flying at higher altitudes results in a significant reduction in flight time because it requires a certain amount of onboard battery power to lift the drone to high altitudes, especially for small drones. Therefore, in this paper, the drone flight altitude is fixed at 30m to preserve the limited onboard battery for tracking and search tasks, and the change in the drone's altitude is not considered when calculating control actions, thereby reducing the computational requirements of the embedded computer.

[0076] Under the above assumptions, see Figure 1 As shown, the technical solution adopted in this specific implementation is: considering n r Drone The system consists of a t Targets Design an algorithm to facilitate UAVs to perform multi-UAV target tracking tasks effectively, and use a layered framework to simplify the multi-target tracking problem in a multi-UAV system. The layered framework can be divided into three layers: state estimation layer, real-time task allocation layer, and motion planning layer;

[0077] The state estimation layer uses simple and fast RSSI measurements to collect and estimate the target position using a particle filter algorithm;

[0078] The real-time task allocation layer adopts a secondary allocation method to solve the multi-UAV task allocation problem in dynamic scenarios;

[0079] The motion planning layer uses an optimization-based approach to obtain a smooth trajectory that the drone can execute while ensuring that the drone can maintain the desired safe distance from all objects of interest. The process mainly involves path planning and trajectory optimization. Initially, the enhanced RRT* algorithm calculates the initial drone path points considering the estimated position and safe distance of the target. Then the trajectory optimization considers optimization items such as control force, dynamic feasibility and interference avoidance to generate the final trajectory.

[0080] The specific process is as follows:

[0081] The state estimation layer first uses a VHF antenna to collect RSSI signals emitted by the radio tags. Then, by using SDRs, the drone can quickly configure the radio and receive radio signals from multiple targets at the same time. The radio signals are processed by the signal detection algorithm. Since each target is uniquely identified by the frequency of the radio transmission, there is no data correlation problem. The target state estimation problem can be solved by running multiple particle filters in parallel.

[0082] The task allocation layer adopts a quadratic allocation method to solve the multi-UAV task allocation problem in dynamic scenarios. First, the task pre-allocation problem is solved by a mixed integer optimization problem. Based on the known information of UAVs and targets, the optimization problem is constructed by multiple constraints (such as distance constraints and UAV coordination constraints), aiming to maximize the effectiveness of the task. Here, the initial position of the target can be obtained through multi-target area search. By solving this optimization problem, the initial target set is reasonably allocated to the UAV swarm. Subsequently, in order to more accurately locate and track moving targets, it is necessary to reallocate appropriate tracking targets to each UAV in real time. The basic requirement of task reallocation is to avoid UAVs tracking the same target, which can effectively reduce the waste of UAV on-board resources and effectively complete the problem MCTT; we achieve this requirement by broadcasting the tasks performed by UAVs.

[0083] After determining the position of the target to be tracked, the motion planning layer generates an executable trajectory for the UAV. An optimization-based approach is used, with requirements such as smoothness, dynamic feasibility, and interference avoidance as optimization terms. The trajectory is obtained by solving the optimization problem, resulting in Algorithm 1.

[0084] As a more specific description of the present invention, the particle filter is specifically:

[0085] The goal of the filtering problem is to filter the historical observations z from time t = 1 to time t = k. 1:k To estimate the confidence level π k (x k ∣z 1:k ), according to the Bayesian recursive idea, from the initial confidence π 0 To begin, the confidence secret is calculated sequentially following the prediction and update steps as follows:

[0086] π k∣k-1 (x k |z 1:k-1 )=

[0087] ∫f k∣k-1 (x k |x)π k-1 (x|z 1:k-1 )dx#

[0088]

[0089] Particle filtering implements the random sampling process of the Monte Carlo method by using independent and uniformly distributed (iid) particles The confidence is approximated by a weighted set of , namely:

[0090]

[0091] Among them, δ(·) represents the Dirac function, and the particle weight satisfies the condition

[0092] The goal of motion planning is determined by using the uncertainty of the position estimate represented by the filter confidence. Although the E-optimal uncertainty estimate can be obtained by using the maximum eigenvalue of the complete covariance matrix of the particle, a computationally less expensive method is used to calculate the maximum covariance of the estimated position along each coordinate axis to represent the position uncertainty, which has similar performance. Let is a particle set The estimated object position, σ (x) represents the standard deviation of the estimated confidence density π(·), then:

[0093]

[0094]

[0095] As a more specific description of the present invention, the algorithm 1 is designed based on this hierarchical framework, and its specific content is:

[0096] MCTT problem algorithm;

[0097] Input: UAV initial position set Target initial position set Target radio frequency set Target initial Stdσ 0 , target end Stdσ min , the number of particles N, the target motion model f(x k ∣x k-1 ), observation model g(zk |x k );

[0098] Output: drone location set

[0099] (1) k = 0,

[0100] (2) Through get

[0101] (3) ←Task pre-allocation

[0102] (4)

[0103] (5) k = k + 1;

[0104] (6) for each drone r i do;

[0105] (7) for each target do;

[0106] (8) Obtaining measurement values

[0107] (9) ←State Estimation

[0108] (10) Value Stored in variable Ω k ;

[0109] (11)

[0110] (12) Set the target t j from and Removed;

[0111] (13) ←Task Reassignment

[0112] (14)

[0113] (15)

[0114] (16) Else;

[0115] (17) Drone i has completed all tasks and returned to the starting point

[0116] Among them, line (2) is based on the initial position set and RF sets of moving targets generate Drone swarms maintain real-time communication Update, when the drone completes tracking of a specific target, it will Delete the target when When it is empty, it means that all targets have been tracked; line (3) is based on and the initial position set of the UAV The initial task set Pre-assigned to the drone group; Line (8) obtains drone r at time k i At target t j RSSI measurement value on Based on this measurement The target motion model f(x k ∣x k-1 ) and the observation model g(z k |x k ); Line (9) uses particle filtering to filter the target t j Perform state estimation and obtain the target t j Estimated location and confidence level (expressed as the standard deviation (Std) of the estimated position, see Section IV-b for details, and then Stored in variable Ω k , for later use in the reallocation algorithm, when the target's estimated position Less than a certain value σ min When considering the positioning and tracking of target tj, lines (11)-(12) are based on the variable Ω that stores the estimated position and Std of the target at time k. k , total list Initial Task Set Target initial Stdσ 0 and drone tracking target list Line (13) performs task redistribution to obtain UAV r i Current goals and updates to track In The executable trajectory of the target UAV is generated in line (15) based on the position of the tracked target.

[0117] As a more specific description of the present invention, the real-time task allocation layer: first, according to the known initial positions of the drone and the target, considering the minimum distance of the drone group, the balanced number of targets to be executed and other conditions, all targets in the area are divided into drones using mixed integer programming, and then each drone obtains an initial target set. Next, the drone uses a tracking strategy, which gives priority to tracking the object with the lowest estimated belief uncertainty, where the uncertainty can be represented by the filter confidence density. When the drone is closer to the target, the signal strength received by the drone from the target is larger, and the target position estimated using particle filtering is more accurate. This strategy realizes rapid detection and real-time tracking of radio-tagged targets. At the same time, due to its movement, the target may appear within the trackable range of other drones. At this time, it is necessary to prevent drones from tracking the same target at the same time through communication between drones. The real-time task allocation layer is specifically as follows:

[0118] 1) Pre-allocation: This section establishes the multi-UAV task pre-allocation problem as an optimization problem to assign reasonable initial tracking targets to UAVs so that they can work more efficiently. Without losing generality, it is assumed that any task can be assigned to any UAV. Since each robot can perform at most Ni tasks and one robot is required to perform each task, all tasks to be performed should have

[0119] The multi-UAV task allocation problem is established as an optimization problem, so that the optimization result with the minimum total cost can be obtained under the task constraints, so that the UAV group can better collaborate as a team and improve work efficiency;

[0120] Assume there is n r drones and n t Mission objectives Among them, the status of the drone is is the three-dimensional coordinate of the drone, is the yaw angle of the UAV; the state of the mission target is It is the x, y, and z coordinates of the Cartesian coordinates. Without loss of generality, we assume that any task can be assigned to any UAV, and since each robot can perform at most N i tasks, each task requires a single robot, so for all tasks to be performed, there should be The overall goal of task allocation is to assign all tasks to robots to maximize total benefits or minimize total costs;

[0121] make is the task target t j Give UAV iThe cost of the task target is determined by the certainty of the estimated position of the target. When allocating tasks, the execution order of the task targets is not considered. Here, a ij Simplified to the distance between the UAV and the mission target on the UAV flight plane, that is Therefore, the optimization goal is the cost required to execute all task goals, expressed as where f ij is the optimization variable, indicating r i Is it assigned to t j , therefore, the multi-robot task allocation problem can be expressed as an optimization problem:

[0122]

[0123] st

[0124]

[0125] Here, constraint (1a) indicates that each mission objective must be assigned to only one UAV and each mission must be performed by one robot; constraint (1b) gives the upper limit of the mission objectives that each UAV can perform, reflecting the limited resources of UAVs; constraint (1c) indicates that f ij is a binary variable representing r i Is it assigned to t j ;

[0126] Reassignment: Although the task set has been assigned to the UAV, the target moves in real time. The UAV can identify other targets that the UAV needs to track during its flight. At this time, dynamic task reallocation is triggered to make task execution more efficient. The principle of the dynamic reallocation algorithm is identified as: moving to the target with the least state estimation uncertainty. When the actual estimated position is unreliable, the UAV traveling towards the target can quickly locate the target position. This is because when the UAV approaches the target, the signal strength increases and the signal is less affected by various interferences. Based on this principle, a dynamic task reallocation algorithm is proposed. The algorithm has high real-time performance and can meet the real-time task reallocation of UAVs in dynamic environments.

[0127] As a more specific description of the present invention, the dynamic task reallocation algorithm is described as follows:

[0128] Step 1: For Each target t in j , from Ω k Get its Std value It represents the uncertainty of the position estimate calculated by the previous particle filter algorithm. Then, if the Std of the target tj is lower than σ 0 , then store it in m tempIn this paper, unlike existing methods that select the estimated position with the lowest uncertainty among all targets as the target, the estimated position of the target is only considered when the target Std is below a predetermined threshold. The rationale is that the initial position estimates from particle filtering exhibit high uncertainty and significant changes over time, and using these estimates as direct destinations may result in a large number of trajectory changes and extended tracking duration;

[0129] Step 2: Delete m temp The target is being tracked by other drones ( to avoid the drones tracking the same target;

[0130] Step 3: If m temp If there is a target, the target with the smallest Std is selected for tracking; if m temp There is no target and the drone is not already on the pre-assigned target list If the tracking target is completed in Select the nearest target to track; otherwise, is empty, which means the drone has no goals to execute and can return to its starting point. Finally, Update

[0131] As a more specific description of the present invention, the motion planning layer is specifically:

[0132] The reason why a UAV navigating towards a target can quickly locate the target position when the actual position estimate is unreliable is that when the UAV is close, the uncertainty of the target signal measurement is low, which can be seen from the distance tracking achieved by measuring the signal strength. Therefore, in the tracking task with the measurement object as the target, the planning strategy of moving towards the target with the lowest position estimate uncertainty with the goal of achieving a sufficiently small position estimate uncertainty may provide a cheap planning strategy;

[0133] After determining the target position to be tracked based on the position estimation uncertainty, it is necessary to generate a trajectory. The basic requirements for the trajectory required by the drone are smoothness, dynamic feasibility, mutual avoidance between drones, etc. In addition, there are additional requirements, such as minimizing execution time and minimizing interference with the target. By adopting Σ MINCO Trajectory to solve the above problems:

[0134] The trajectory generation of the UAV is formulated as an unconstrained optimization problem:

[0135]

[0136] Where λ is the weight vector weighing each cost function;

[0137] 1) Control power e :Control force optimization item J e It is expressed by integrating the squared norm of the third-order derivative of the trajectory over time and written as

[0138]

[0139] 2) Total time J t : In order to improve the efficiency of the drone swarm in executing tasks, it is hoped that the drone swarm can reach the target point as quickly as possible, thereby minimizing the total time:

[0140]

[0141] 3) Obstacle Avoidance o : Use the Euclidean Signed Distance Field (ESDF) to establish the environment around the drone, and ESDF can obtain the distance between the current position of the drone and the nearest obstacle. In order to ensure safety requirements, a collision penalty item is defined. It is triggered when the distance to the obstacle is less than the safety gap. The collision penalty item is as follows:

[0142]

[0143] Among them, d thr is the safety threshold, The distance between the point and the nearest obstacle around it is considered, and then the obstacle avoidance optimization term J is obtained by calculating the weighted sum of the sampling constraint function o :

[0144]

[0145] in are the orthogonal coefficients that follow the trapezoidal rule;

[0146] 4) Dynamic feasibility d : Since the UAV needs to be constrained by its own motors and other actuators when flying, the speed and acceleration of the UAV have maximum values. The dynamic feasibility constraint term J is used d Limit the maximum values ​​of velocity and acceleration to ensure that the trajectory can be executed by the drone:

[0147] J d =J v +J a #

[0148]

[0149]

[0150] Among them, v m ,a mis the maximum allowed speed and acceleration;

[0151] 5) Anti-interference J i : To maintain a safe distance r between the UAV and the target min , an anti-interference optimization term is constructed, letting V(u, r min ) denote the anti-interference area of the target state (position) based on the UAV state u, and given:

[0152]

[0153] Then, the interruption avoidance penalty Ji is obtained by calculating the weighted sum of the sampling constraint functions as

[0154]

[0155] The gradient of Ji with respect to ci and Ti is given by the following formula

[0156]

[0157] where t is the relative time on the trajectory. For the case of d(x, u) < dmin, the gradient is given by the following formula

[0158]

[0159] where is the gradient of the Euclidean signed distance field (ESDF) in the middle, otherwise, the gradient becomes

[0160] The following are the related embodiments of the present invention

[0161] In our simulation, 10 moving objects equipped with radio tags were set, randomly moving within a search area from 1000 meters to 1000 meters. The UAV always flew at a height of 30 meters above the ground, while the radio tags were placed at a height of 1 meter above the ground, thus restricting the planning to a two-dimensional domain. The maximum flight time was set to 3000 seconds. During the flight, a minimum safe distance of 50 meters was maintained between the UAV and the radio target to avoid interfering with the target. The particle filter used 10,000 particles. It should be noted that when the estimated uncertainty of the object was below the 20-meter threshold, the object was considered to be successfully tracked and located.

[0162] Here, we consider two scenarios. Scenario 1 describes a single UAV with position [50,50,30] locating and tracking all targets, while Scenario 2 shows four UAVs located at [50,50,30], [1000,50,30], [50,1000,30], [1000,1000,30]. The algorithm is evaluated based on the following metrics:

[0163] 1) Root Mean Square (RMS) (m): Estimates the distance between the target location and its actual location. Use to represent the estimated position of the target, to represent the actual position of the target. RMS is the square root of the mean square error and is given by:

[0164]

[0165] 2) Standard deviation of the estimated result (Std / Stdev): an indicator for evaluating the uncertainty of the estimated position, see Perform detailed calculations.

[0166] 3) Flight Time(s): The time it takes the UAV to search for and locate all assigned radio-tagged targets.

[0167] 4) Planning time (s): The time required for the UAV to track the target and calculate the optimal tracking position.

[0168] See also Figure 2-Figure 3 As shown in Figure 2, the localization results for 10 moving targets are displayed, with details such as RMS and flight time annotated along with each target location. In this case, the UAV successfully located all 10 moving targets within 542 seconds, keeping the maximum error distance below 25 meters, except for the outlier target 1, which had an RMS of 36.6907 meters. After locating the final target (target 7) within 542 seconds, the UAV received a command to return to its starting point. These results highlight the effectiveness of our algorithm in real-time search and localization of multiple targets, completing the task in approximately 10 minutes. Furthermore, it is noteworthy that the UAV's trajectory always maintains a safe distance from the radio-tagged target. The simulation results verify the effectiveness of our proposed trajectory planning algorithm in preventing the UAV from getting too close to the target while autonomously achieving the tracking goal.

[0169] See also Figure 4The figure illustrates the time evolution of the Std value for each target tracked by the UAV in Scenario 1. The data represents the uncertainty variation of different radio tags. However, over time, the UAV can effectively reduce the Std value representing the uncertainty below the pre-defined threshold of 20 meters. When the Std value of a tracked target falls below the pre-defined threshold, the target is considered successfully localized and the UAV can track other targets. It is worth noting that even if the UAV tracks other targets, the Std value of the local target remains stable, facilitating the simultaneous tracking and localization of multiple targets.

[0170] See also Figure 5 The figure shows the change in RMS value for each tracked target. The figure clearly shows that the RMS value of most targets remained below 25 meters at the time of positioning, with minimal changes observed thereafter. This verifies the high accuracy of our tracking algorithm in locating all radio-tagged targets.

[0171] See also Figure 6 The figure shows the results of four drones tracking and locating 10 moving targets. As in scenario 1, the estimation details are marked next to the target location, including two metrics: RMS and flight time. It can be seen that it takes 197 seconds for the drone swarm to locate all 10 moving targets with a maximum error distance of less than 20 meters. The results show that our algorithm can effectively collaborate with the drone swarm to solve the MCTT problem, which greatly reduces the execution time of the task, and the tracking performance is similar to that of a single drone tracking and locating all targets.

[0172] See also Figure 7 The figure depicts the Std values ​​of the targets tracked by four UAVs over time. The results show that there are differences in the estimated uncertainty of the radio tags tracked by each UAV; however, the planning strategy can effectively reduce the estimated uncertainty to complete the localization of the radio tags.

[0173] See also Figure 8-Figure 11 The figure shows how the RMS values ​​of the targets located by the four UAVs change over time.

[0174] The results show that the RMS value can be reduced to the desired value in a limited time, which indicates that our algorithm can enable the UAV to quickly locate the designated radio-tagged target.

[0175] The above description is only used to illustrate the technical solution of the present invention rather than to limit it. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in this field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-UAV collaborative path planning method for multi-target tracking, characterized by: Consider n r UAVs R = {r1,…,r i ,…,r nr } system, jointly tracking n t target T={t1,…,t j ,…,t nt }, design an algorithm to facilitate UAVs to perform multi-UAV target tracking tasks effectively, and use a hierarchical framework to simplify the multi-target tracking problem in a multi-UAV system. The hierarchical framework can be divided into three layers: state estimation layer, real-time task allocation layer, and motion planning layer; The state estimation layer uses simple and fast RSSI measurements to collect and estimate the target position using a particle filter algorithm; The real-time task allocation layer adopts a secondary allocation method to solve the multi-UAV task allocation problem in dynamic scenarios; The motion planning layer uses an optimization-based approach to obtain a smooth trajectory that the drone can execute while ensuring that the drone can maintain the desired safe distance from all objects of interest. The process mainly involves path planning and trajectory optimization. Initially, the enhanced RRT* algorithm calculates the initial drone path points considering the estimated position and safe distance of the target. Then the trajectory optimization considers optimization items such as control force, dynamic feasibility and interference avoidance to generate the final trajectory. The specific process is as follows: The state estimation layer first uses a VHF antenna to collect RSSI signals emitted by the radio tags. Then, by using SDRs, the drone can quickly configure the radio and receive radio signals from multiple targets at the same time. The radio signals are processed by the signal detection algorithm. Since each target is uniquely identified by the frequency of the radio transmission, there is no data correlation problem. The target state estimation problem can be solved by running multiple particle filters in parallel. The task allocation layer adopts a secondary allocation method to solve the multi-UAV task allocation problem in dynamic scenarios. First, the task pre-allocation problem is solved by a mixed integer optimization problem. Based on the known information of UAVs and targets, the optimization problem is constructed by multiple constraints (such as distance constraints and UAV coordination constraints) to maximize the effectiveness of the task. Here, the initial position of the target can be obtained by multi-target area search. By solving this optimization problem, the initial target set is reasonably allocated to the UAV group. Subsequently, in order to more accurately locate and track moving targets, it is necessary to reallocate appropriate tracking targets to each UAV in real time. The basic requirement of task reallocation is to avoid UAVs tracking the same target, which can effectively reduce the waste of UAV onboard resources and effectively complete the MCTT problem. After determining the position of the target to be tracked, the motion planning layer generates an executable trajectory for the UAV. An optimization-based approach is used, with requirements such as smoothness, dynamic feasibility, and interference avoidance as optimization terms. The trajectory is obtained by solving the optimization problem, resulting in Algorithm 1.

2. The method for multi-UAV collaborative path planning for multi-target tracking according to claim 1, characterized in that: The particle filter is specifically: The goal of the filtering problem is to filter the historical observations z from time t = 1 to time t = k. 1:k To estimate the confidence level π k (x k ∣z 1:k ), according to the Bayesian recursive idea, starting from the initial confidence π0, the confidence secret is calculated in the order of prediction and update steps, as shown below: Particle filtering implements the random sampling process of the Monte Carlo method by using independent and uniformly distributed (iid) particles The confidence is approximated by a weighted set of , namely: Among them, δ(·) represents the Dirac function, and the particle weight satisfies the condition The goal of motion planning is determined by using the uncertainty of the position estimate represented by the filter confidence. Although the E-optimal uncertainty estimate can be obtained using the maximum eigenvalue of the particle's complete covariance matrix, a computationally less expensive method is used to calculate the maximum covariance of the estimated position along each coordinate axis to represent the position uncertainty, which has similar performance. Let is a particle set The estimated object position, σ (x) represents the standard deviation of the estimated confidence density π(·), then:

3. The method for multi-UAV collaborative path planning for multi-target tracking according to claim 1, characterized in that: The algorithm 1 is designed based on this hierarchical framework, and its specific content is: MCTT problem algorithm; Input: UAV initial position set Target initial position set Target radio frequency set Target initial Stdσ0, target end Stdσ min , the number of particles N, the target motion model f(x k ∣x k-1 ), observation model g(z k |x k ); Output: drone location set (1)k＝0, (2) Through get (3) ←Task pre-allocation (4) (5) k = k + 1; (6) for each drone r i do; (7) for each target (8) Obtaining measurement values (9) ←State Estimation (10) Value Stored in variable Ω k ; (11)ifσ tj <σ min then; (12) Set the target t j from and Removed; (13) ←Task Reassignment (14)if (15) (16) Else; (17) Drone i has completed all tasks and returned to the starting point Among them, line (2) is based on the initial position set and RF sets of moving targets generate Drone swarms maintain real-time communication Update, when the drone completes tracking of a specific target, it will Delete the target when When it is empty, it means that all targets have been tracked; line (3) is based on and the initial position set of the UAV The initial task set Pre-assigned to the drone group; Line (8) obtains drone r at time k i At target t j RSSI measurement value on Based on this measurement The target motion model f(x k ∣x k-1 ) and the observation model g(z k |x k ); Line (9) uses particle filtering to filter the target t j Perform state estimation and obtain the target t j Estimated location and the confidence level σ tj (expressed as the standard deviation (Std) of the estimated position, see Section IV-b for details, and then Stored in variable Ω k , for later use in the reallocation algorithm, when the Stdσ of the target estimated position tj Less than a certain value σ min When considering the positioning and tracking of target tj, lines (11)-(12) are based on the variable Ω that stores the estimated position and Std of the target at time k. k , total list Initial Task Set Target initial Stdσ0 and drone tracking target list Line (13) performs task redistribution to obtain UAV r i Current goals and updates to track In The executable trajectory of the target UAV is generated in line (15) based on the position of the tracked target.

4. The method for multi-UAV collaborative path planning for multi-target tracking according to claim 1, characterized in that: The real-time task allocation layer: First, based on the known initial positions of the drones and the targets, considering conditions such as the minimum distance of the drone swarm and the balance number of targets to be executed, all the targets in the area are divided into drones using mixed-integer programming. Then, each drone obtains an initial target set. Next, the drones use a tracking strategy that prioritizes tracking the object with the lowest estimated belief uncertainty, where the uncertainty can be represented using the filter confidence density. When the drone gets closer to the target, the drone receives a stronger signal intensity emitted by the target, and the target position estimated using particle filtering is more accurate. This strategy realizes the rapid detection and real-time tracking of radio-tagged targets. At the same time, due to its movement, the target may appear within the trackable range of other drones. At this time, communication between drones is required to avoid drones tracking the same target simultaneously. Its real-time task allocation layer is specifically as follows: 1) Pre-allocation: This section establishes the multi-UAV task pre-allocation problem as an optimization problem to assign reasonable initial tracking targets to UAVs so that they can work more efficiently. Without losing generality, it is assumed that any task can be assigned to any UAV. Since each robot can perform at most Ni tasks and one robot is required to perform each task, all tasks to be performed should have The multi-drone task allocation problem is formulated as an optimization problem, enabling an optimization result with the minimum total cost to be obtained under the premise of meeting task constraints, so that the drone swarm can better perform teamwork and improve work efficiency; Assume there is n r drones and n t Mission objectives Among them, the status of the drone is is the three-dimensional coordinate of the drone, θ (u ri ) is the yaw angle of the drone; the state of the mission target is It is the x, y, and z coordinates of the Cartesian coordinates. Without loss of generality, we assume that any task can be assigned to any UAV, and since each robot can perform at most N i tasks, each task requires a single robot, so for all tasks to be performed, there should be The overall goal of task allocation is to assign all tasks to robots to maximize total benefits or minimize total costs; make is the task target t j Give UAV i The cost of the task target is determined by the certainty of the estimated position of the target. When allocating tasks, the execution order of the task targets is not considered. Here, a ij Simplified to the distance between the UAV and the mission target on the UAV flight plane, that is Therefore, the optimization goal is the cost required to execute all task goals, expressed as where f ij is the optimization variable, indicating r i Is it assigned to t j , therefore, the multi-robot task allocation problem can be expressed as an optimization problem: s.t. Here, constraint (1a) indicates that each mission objective must be assigned to only one UAV and each mission must be performed by one robot; constraint (1b) gives the upper limit of the mission objectives that each UAV can perform, reflecting the limited resources of UAVs; constraint (1c) indicates that f ij is a binary variable representing r i Is it assigned to t j ; 2) Reallocation: Although the task set has been assigned to the drones, the targets move in real time. The drones can identify other targets that need to be tracked during their flight. At this time, dynamic task reallocation is triggered to make the task execution more efficient. By identifying the principle of the dynamic reallocation algorithm as: moving to the target with the least state estimation uncertainty. When the actual estimated position is unreliable, the drone moving towards the target can quickly locate the target position. This is because when the drone approaches the target, the signal intensity increases and the signal is less affected by various interferences. Based on this principle, a dynamic task reallocation algorithm is proposed. This algorithm has high real-time performance and can meet the real-time task reallocation of drones in a dynamic environment.

5. The method for multi-UAV collaborative path planning for multi-target tracking according to claim 1, characterized in that: The described dynamic task reallocation algorithm is described as follows: S1: For Each target t in j , from Ω k Get its Std value σ tj , which represents the uncertainty of the position estimate calculated by the previous particle filter algorithm. Then, if the Std of the target tj is lower than σ0, it is stored in m temp In this paper, unlike existing methods that select the estimated position with the lowest uncertainty among all targets as the target, the estimated position of the target is only considered when the target Std is below a predetermined threshold. The rationale is that the initial position estimates from particle filtering exhibit high uncertainty and significant changes over time, and using these estimates as direct destinations may result in a large number of trajectory changes and extended tracking duration; S2: Delete m temp The target is being tracked by other drones ( to avoid the drones tracking the same target; S3: If m temp If there is a target, the target with the smallest Std is selected for tracking; if m temp There is no target and the drone is not already on the pre-assigned target list If the tracking target is completed in Select the nearest target to track; otherwise, is empty, which means the drone has no goals to execute and can return to its starting point. Finally, Update 6. The method for multi-UAV collaborative path planning for multi-target tracking according to claim 1, characterized in that: The specific motion planning layer is as follows: In the case where the actual position estimation is unreliable, the drone navigating towards the target can quickly locate the target position because when the drone is closer, the uncertainty of the target signal measurement is lower, which can be known from distance tracking by measuring the signal intensity. Therefore, in the tracking task with the measured object as the target, a planning strategy of moving towards the target with the lowest position estimation uncertainty with the goal of achieving a sufficiently small position estimation uncertainty may provide a cheap planning strategy; After determining the target position to be tracked based on the position estimation uncertainty, it is necessary to generate a trajectory. The basic requirements for the trajectory required by the drone are smoothness, dynamic feasibility, mutual avoidance between drones, etc. By adopting Σ MINCO Trajectory to solve the above problems: The trajectory generation of the drone is formulated as an unconstrained optimization problem: where λ is a weight vector that weighs each cost function; 1) Control power e :Control force optimization item J e It is expressed by integrating the squared norm of the third-order derivative of the trajectory over time and written as 2) Total time J t : In order to improve the efficiency of the drone swarm in executing tasks, it is hoped that the drone swarm can reach the target point as quickly as possible, thereby minimizing the total time: 3) Obstacle Avoidance o : Use the Euclidean Signed Distance Field (ESDF) to establish the environment around the drone, and ESDF can obtain the distance between the current position of the drone and the nearest obstacle. In order to ensure safety requirements, a collision penalty item is defined. It is triggered when the distance to the obstacle is less than the safety gap. The collision penalty item is as follows: Among them, d thr is the safety threshold, is the distance between the considered point and the nearest obstacle around it; Then, the obstacle avoidance optimization term J is obtained by calculating the weighted sum of the sampling constraint functions o : in are the orthogonal coefficients that follow the trapezoidal rule; 4) Dynamic feasibility d : Since the UAV needs to be constrained by its own motors and other actuators when flying, the speed and acceleration of the UAV have maximum values. The dynamic feasibility constraint term J is used d Limit the maximum values ​​of velocity and acceleration to ensure that the trajectory can be executed by the drone: I d =J v +J a # Among them, v m ,a m are the maximum speed and acceleration allowed; 5) Anti-interference i : In order to maintain a safe distance between the drone and the target min , construct an anti-interference optimization term, let V(u,r min ) represents the anti-interference area based on the target state (position) of the drone state u, given by: Then, the interruption avoidance penalty Ji is obtained by calculating the weighted sum of the sampled constraint functions as The gradient of Ji with respect to ci and Ti is given by the following formula where t is the relative time on the trajectory. For the case of d(x,u) < dmin, the gradient is given by the following formula in, yes The gradient of the Euclidean signed distance field (ESDF) in , otherwise, the gradient becomes 7. A multi-UAV collaborative path planning method for multi-target tracking, characterized in that: It includes a target model, a drone model, and an observation model Target model: For wild animal targets, their dynamic behaviors are usually unpredictable. Therefore, their behaviors are modeled as a random walk model in, represents the state of the target at the kth moment; represents a Gaussian distribution with mean μ and covariance Q; is the 3×3 covariance matrix of process noise, where In represents an n×n dimensional matrix; Drone model: For UAVs, the concept of multicopter differential flatness has been extensively explored and demonstrated to have a remarkably, physically interpretable flat output space, establishing that all states of the UAV, including position, attitude, velocity, and angular velocity, as well as inputs such as thrust acceleration and torque, can be represented by a flat output (Z) and its corresponding derivatives. It is noteworthy that the flat output space overlaps with the original configuration space, and its basic structure of flat outputs is similar to Consistent, where (p x ,p y ,p z ) T represents the translation of the center of gravity of the drone, ψ is the yaw angle, and by ensuring that the trajectory obeys a certain set of constraints in the flat space, it can be efficiently converted to the original configuration space using a planar transformation, let represents the state of the drone at the kth moment, where is the 3D coordinate of the drone, is the yaw angle of the drone. Here, the drone adopts a fixed flight altitude and focuses on the horizontal control of the drone. This method avoids adjusting the altitude during the trajectory calculation process, thereby effectively reducing the amount of calculation of the drone's onboard computer; Observation model: The observation model is a real signal propagation model used to estimate the target state x k and the drone status u k RSSI measurement between, which plays an important role in state estimation using particle filters. The observation model is used using the logarithmic distance path loss model. The power h(x k ,u k )(dBm) only contains the line-of-sight power component h(x k ,u k )=P0-10L c log(d(x k ,u k ))+G r (x k ,u k ), Where, P0 (dBm) represents the reference power; L c Represents the unitless path loss constant, which characterizes the signal attenuation as the transmission distance increases, and usually ranges from 2 to 4; is the distance between the target and the UAV; G r (x k ,u k ) represents the directional antenna gain, which depends on the yaw angle of the drone and its relationship with the target x k The relative position of The RSSI value zk (dBm) is easily corrupted by noise, including thermal noise and interference from other sources. Given the assumption of white noise, the measurement likelihood model can be expressed as: Among them, Q (z) is the 1×1 covariance matrix of the measurement noise.

Images