Intelligent sensing, predicting and tracking method for dynamic target of unmanned aerial vehicle

Abstract

The invention relates to an intelligent sensing, predicting and tracking method for a dynamic target of an unmanned aerial vehicle, and belongs to the technical field of unmanned aerial vehicle autonomous control and path planning. According to the method, a target trajectory is predicted through a Kalman filter, unmanned aerial vehicle dynamics is accurately modeled by using a physical information neural network, and an optimal control sequence of a future time domain is solved through online rolling optimization by taking target trajectory prediction and an unmanned aerial vehicle dynamics model as constraints based on a model prediction control framework. And an autonomous closed loop of perception-prediction-planning-execution-updating is formed. The method effectively solves the problems of environment uncertainty, calculation real-time performance, complex scene multi-target tradeoff and the like of an existing method, provides a high-precision and low-delay dynamic target intelligent sensing prediction tracking method, and remarkably improves the precision and efficiency of dynamic target tracking of an unmanned aerial vehicle in a complex environment.

Description

TECHNICAL FIELD

[0001] The present application relates to an unmanned aerial vehicle dynamic target intelligent perception prediction tracking method, and belongs to the technical field of unmanned aerial vehicle autonomous control and path planning, in particular to an unmanned aerial vehicle dynamic target intelligent perception prediction tracking. BACKGROUND

[0002] With the rapid development of unmanned aerial vehicle technology, its application in search and rescue, security, surveying and mapping and other fields is becoming more and more widespread. Autonomous tracking of moving targets is one of the core technologies of intelligent application of unmanned aerial vehicles, which has great practical significance and wide application prospect. Autonomous tracking of dynamic targets by unmanned aerial vehicles is a core capability to realize high-level intelligence, and path planning is crucial to the success of the task. The essence of this problem is to generate a safe and efficient optimal trajectory in real time based on the prediction of target motion and physical constraints of the unmanned aerial vehicle in an uncertain environment. Existing methods cover a variety of paradigms from traditional geometric algorithms, artificial potential field methods, random algorithms, optimal control to modern machine learning algorithms.

[0003] However, these methods generally face inherent contradictions between environmental prediction uncertainty, computational real-time performance, model accuracy and simplicity, as well as multi-objective trade-off challenges in obstacle avoidance and tracking in complex scenarios. Therefore, developing a new framework that can integrate accurate prediction, high-fidelity dynamic modeling and fast online optimization has become a key direction to break through current technical bottlenecks and improve the reliability and autonomy of unmanned aerial vehicles in complex real-world scenarios. SUMMARY

[0004] The present application aims to provide an unmanned aerial vehicle dynamic target intelligent perception prediction tracking method, which predicts the target motion trajectory through a Kalman filter, accurately models the dynamics of the unmanned aerial vehicle using a physics-informed neural network (PINN), and solves the optimal control sequence online based on a model predictive control (MPC) framework, thereby achieving high-precision and high-efficiency dynamic target tracking.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solution:

[0006] An unmanned aerial vehicle dynamic target intelligent perception prediction tracking method is implemented based on an unmanned aerial vehicle system, which includes: a signal receiving module for receiving state information from a dynamic target; a positioning module for obtaining the position and speed state of the unmanned aerial vehicle itself; a data processing unit with built-in Kalman filter algorithm, physics-informed neural network model, model predictive control optimization algorithm and adaptive update logic; a flight control unit for receiving control instructions from the data processing unit and driving the unmanned aerial vehicle to perform flight tasks; and the data processing unit is connected to the signal receiving module, the positioning module and the flight control unit.

[0007] The method comprises the following steps:

[0008] S1: task initialization: the UAV takes off to a safe height, initializes the communication link, positioning module and control unit, and enters the dynamic target tracking task mode;

[0009] S2: target state acquisition: the UAV receives the target state information sent by the dynamic target; the real-time position and speed information of the target are acquired by receiving the state data packet sent by the dynamic target through the airborne wireless communication module;

[0010] S3: target trajectory prediction: a Kalman filter is established, and the motion trajectory of the future N time steps is predicted based on the historical data of the target motion;

[0011] S4: UAV dynamics modeling: a physical information neural network is used to construct the dynamics model of the UAV, taking the current state and control input as the independent variable and outputting the state derivative;

[0012] S5: predictive control optimization: a predictive control optimization model is established, and the control vector of the future N time steps is calculated in a rolling time domain, minimizing the tracking error between the UAV and the predicted target and the control energy consumption;

[0013] S6: control execution and state update: the UAV flight control is executed according to the control vector of each time step;

[0014] S7: repeat steps S2-S6 to form an autonomous closed loop of "perception-prediction-planning-execution-update", until the UAV successfully tracks the target to the task area or receives the termination instruction.

[0015] Further, the target state searching target is the position and speed of the target at the current time t k , i.e. p k =[x, y, z, v x , v y , v z ], wherein x, y, z are the positions of the target in the x coordinate direction, y coordinate direction and z coordinate direction of the spatial coordinate system oxyz, and v x , v y , v z are the speeds in the x coordinate direction, y coordinate direction and z coordinate direction.

[0016] Further, the UAV state is the position and speed of the UAV at the current time t k , i.e. wherein, x, y, z are the positions of the target in the x coordinate direction, y coordinate direction and z coordinate direction of the spatial coordinate system oxyz, The velocity in the x coordinate direction, the y coordinate direction, and the z coordinate direction.

[0017] Further, the control vector u k is the speed of each motor of the unmanned aerial vehicle at the current time t k .

[0018] Further, the state equation and the measurement equation of the Kalman filter in the step S3 are as follows:

[0019]

[0020] wherein k represents the kth time step, p k is the target state vector at the time step k, z k+1 is the target measurement vector at the time step k+1, w k is the random system noise, and v k+1 is the random measurement noise; wherein the system noise is subject to a Gaussian distribution with a mean of zero and a covariance of Q k , the measurement noise is subject to a Gaussian distribution with a mean of zero and a covariance of R k , and the measurement noise and the system noise are irrelevant to each other; f(·) is a system state transformation function, and h(·) is a system state measurement function.

[0021] Preferably, the state transformation function f(·) and the state measurement function h(·) are function type determination and parameter adjustable functions, and the parameters are set through a small amount of historical data fitting.

[0022] Further, the physical information neural network NN(s, u; θ) in the step S4 has an input of the state s and the control vector u, an output of the derivative ds / dt of the state of the unmanned aerial vehicle, and a network parameter θ; and a corresponding loss function is as follows:

[0023] Loss = α·Loss_data + β·Loss_physics, wherein α and β are artificial parameters, Loss_data = MSE(NN(s, u; θ), ds / dt_measured), and ds / dt_measured is the measured derivative of the state of the unmanned aerial vehicle;

[0024] Loss_physics = MSE(NN(s, u; θ), g(s, u)), and ds / dt = g(s, u) is a differential equation obtained through analysis of the dynamics model of the unmanned aerial vehicle.

[0025] Preferably, the state of the unmanned aerial vehicle is predicted by the physical information neural network combined with the Rung-Kuta format, and specifically, the state vector of the unmanned aerial vehicle at the time step i+1 predicted for the time step k is as follows: Where, k1 = h·PINN(t) k ,s i|k ,u i ),

[0026]

[0027] k4 = h·PINN(t) k +h,s i|k +k3,u i+1 ), h = t k+1 -t k ,

[0028] Furthermore, the objective equation of the predictive control optimization model in step S6 is:

[0029] This NLP problem can be solved online using a nonlinear solver (such as IPOPT, CaSADi, ACADO) to obtain the optimal control sequence U. * ;

[0030] Where, p i+1|k s is the target state vector at time step i+1 predicted from time step k; i+1|k U is the UAV state vector predicted from time step k at time step i+1; Q is the state error weight matrix and R is the control input weight matrix, which are set manually. Adjusting Q and R essentially involves finding a balance between "tracking performance" and "control cost / smoothness"; U = [u k ,u k+1 ,...,u k+N-1 ].

[0031] Preferably, in step S6, when the error between the Kalman filter predicting the target state and the received target state at any time step exceeds a threshold, or when the execution time step reaches a set upper limit, the state transformation function and the state measurement function are updated, and the process returns to step S2.

[0032] Preferably, in step S6, when the error between the output drone state of the drone dynamics model and the measured drone state exceeds a threshold, the network parameters of the physical information neural network need to be retrained, and the process returns to step S2.

[0033] The beneficial effects of this invention are as follows:

[0034] This invention provides an intelligent perception, prediction, and tracking method for dynamic targets on unmanned aerial vehicles (UAVs). By deeply fusing Kalman filter prediction, precise modeling using a physical information neural network (PIN) and online optimization of predictive control, a highly efficient, accurate, and robust UAV target tracking architecture is constructed. The Kalman filter is used to optimally estimate and predict the trajectory of dynamic targets, effectively overcoming the influence of environmental uncertainties and observation noise, providing reliable forward-looking information for subsequent planning. The innovative use of a PIN to perform precise modeling of UAV dynamics driven by both data and physical laws significantly improves model fidelity. This high-precision model serves as the core of predictive control, making the generated control commands more consistent with real-world physical constraints, greatly improving the smoothness and stability of the tracking process. The rolling optimization framework can solve the optimal control problem in a finite time domain online while ensuring real-time computation, effectively balancing computational efficiency and tracking accuracy. The system's built-in adaptive model and parameter update mechanism enables the overall solution to adapt to sudden changes in target motion patterns or changes in the UAV's own characteristics, demonstrating strong online learning capabilities and environmental robustness. This significantly improves the accuracy, reliability, and practicality of UAVs performing autonomous tracking tasks in complex dynamic environments. Attached Figure Description

[0035] To illustrate the objectives and technical solutions of this invention, the following figures are provided:

[0036] Figure 1 This is a flowchart of the method of the present invention;

[0037] Figure 2 This is a framework diagram of Embodiment 1 of the present invention. Detailed Implementation

[0038] Example 1: This example illustrates the method of the present invention by describing in detail the real-time tracking of a moving vehicle by a drone in an urban environment. The vehicle is equipped with a GPS module and a wireless transmitter, which can periodically broadcast its position and speed information. The method is implemented based on a drone system, which includes: a signal receiving module for receiving status data packets from the vehicle; a GPS / IMU combined positioning module for acquiring the drone's own high-precision position and speed status; an onboard data processing unit with a built-in Kalman filter algorithm, a physical information neural network model, a model predictive control solver, and model update logic; and a flight control unit for receiving control commands from the data processing unit and driving the drone to perform flight tasks. The data processing unit is connected to the signal receiving module, the positioning module, and the flight control unit.

[0039] Combination Figure 1 Specifically, it includes the following steps:

[0040] S1: Mission Initialization: The UAV takes off and ascends to a safe altitude, initializes the communication link, positioning module and control unit, and enters the dynamic target tracking mission mode.

[0041] S2: Target Status Acquisition: The UAV receives target status information from a dynamic target; it receives status data packets from the dynamic target through the onboard wireless communication module and parses them to obtain the target's real-time position and speed information.

[0042] Wherein, the target state search target is at the current time t k Position and velocity, i.e., p k =[x,y,z,v x ,v y ,v z ], where x, y, and z are the positions of the target in the x-coordinate direction, y-coordinate direction, and z-coordinate direction in the spatial coordinate system oxyz, and v x v y v z The velocities are the x-coordinate, y-coordinate, and z-coordinate.

[0043] The drone state is the current time t. k The drone's position and speed, i.e. in, Let x be the position of the target in the x, y, and z coordinates of the spatial coordinate system oxyz. The velocities are the x-coordinate, y-coordinate, and z-coordinate.

[0044] The control vector u k Let t be the current time. k The rotational speed of each motor in the drone.

[0045] S3: Target trajectory prediction: Establish a Kalman filter and predict the trajectory of the target in the next N time steps based on the target's historical motion data.

[0046] Specifically, the state equation and measurement equation of the Kalman filter are as follows:

[0047] Where k represents the k-th time step, p k Let z be the target state vector at time step k. k+1 Let w be the target measurement vector at time step k+1. k For random system noise, v k+1 The noise is a random measurement; the system noise has a mean of zero and a covariance of Q. k The signal follows a Gaussian distribution, and the measurement noise has a mean of zero and a covariance of R. kThe system noise follows a Gaussian distribution and is uncorrelated with the measurement noise; f(·) is the system state transformation function, and h(·) is the system state measurement function.

[0048] S4: UAV Dynamics Modeling: Construct a dynamics model of the UAV using a physical information neural network, with the current state and control input as independent variables, and output the state derivative.

[0049] The physical information neural network NN(s,u;θ) takes the state s and control vector u as input and outputs the derivative of the UAV state ds / dt, with network parameter θ. The corresponding loss function is: Loss = α·Loss_data + β·Loss_physics, where α and β are manually set parameters, Loss_data = MSE(NN(s,u;θ),ds / dt_measured), where ds / dt_measured is the derivative of the measured UAV state, and Loss_physics = MSE(NN(s,u;θ),g(s,u)), where ds / dt = g(s,u) is the differential equation obtained from the analysis of the UAV dynamics model.

[0050] S5: Predictive Control Optimization: Establish a predictive control optimization model, use a rolling time domain approach to calculate the control vector for the next N time steps, and minimize the tracking error between the UAV and the predicted target as well as the control energy consumption.

[0051] Specifically, the objective equation of the predictive control optimization model is: This NLP problem can be solved online using nonlinear solvers such as IPOPT, CasADi, and ACADO to obtain the optimal control sequence U. * .

[0052] Where, p i+1|k s is the target state vector at time step i+1 predicted from time step k; i+1|k U is the UAV state vector predicted from time step k at time step i+1; Q is the state error weight matrix, R is the control input weight matrix, which are set manually, and U = [u k ,u k+1 ,...,u k+N-1 ].

[0053] S6: Control Execution and State Update: Execute UAV flight control according to the control vector at each time step.

[0054] When the error between the Kalman filter prediction of the target state and the received target state at any time step exceeds the threshold, or when the execution time step reaches the set upper limit N, the state transformation function f(·) and the state measurement function h(·) are updated, and the process returns to step S2.

[0055] S7: Repeat steps S2 to S6 until the UAV successfully tracks the target to the mission area or receives a termination command.

[0056] Example 2: This example illustrates a field search and rescue mission where a drone is needed to search for and track a lost person in motion. The lost person is equipped with a BeiDou positioning terminal, which can periodically broadcast location and speed information. The system is implemented based on a drone system, which includes: a signal receiving module for receiving status data packets from the lost person; a GPS / IMU combined positioning module for obtaining the drone's own high-precision position and speed status; an onboard data processing unit with a built-in Kalman filter algorithm, a physical information neural network model, a model predictive control solver, and model update logic; and a flight control unit for receiving control commands from the data processing unit and driving the drone to execute flight tasks. The data processing unit is connected to the signal receiving module, the positioning module, and the flight control unit.

[0057] Combination Figure 1 Specifically, it includes the following steps:

[0058] S1: Mission Initialization: The UAV takes off and ascends to a safe altitude, initializes the communication link, positioning module and control unit, and enters the dynamic target tracking mission mode.

[0059] S2: Target Status Acquisition: The UAV receives target status information from a dynamic target; it receives status data packets from the dynamic target through the onboard wireless communication module and parses them to obtain the target's real-time position and speed information.

[0060] Wherein, the target state search target is at the current time t k Position and velocity, i.e., p k =[x,y,z,v x ,v y ,v z ], where x, y, and z are the positions of the target in the x-coordinate direction, y-coordinate direction, and z-coordinate direction in the spatial coordinate system oxyz, and v x v y v z The velocities are the x-coordinate, y-coordinate, and z-coordinate.

[0061] The drone state is the current time t. k The drone's position and speed, i.e. in, Let x be the position of the target in the x, y, and z coordinates of the spatial coordinate system oxyz. The velocities are the x-coordinate, y-coordinate, and z-coordinate.

[0062] The control vector u k Let t be the current time. k The rotational speed of each motor in the drone.

[0063] S3: Target trajectory prediction: Establish a Kalman filter and predict the trajectory of the target in the next N time steps based on the target's historical motion data.

[0064] Specifically, the state equation and measurement equation of the Kalman filter are as follows:

[0065] Where k represents the k-th time step, p k Let z be the target state vector at time step k. k+1 Let w be the target measurement vector at time step k+1. k For random system noise, v k+1 The noise is a random measurement; the system noise has a mean of zero and a covariance of Q. k The signal follows a Gaussian distribution, and the measurement noise has a mean of zero and a covariance of R. k The system noise follows a Gaussian distribution and is uncorrelated with the measurement noise; f(·) is the system state transformation function, and h(·) is the system state measurement function.

[0066] The state transformation function f(·) and the state measurement function h(·) are functions with defined function types and adjustable parameters, and the parameters are set by fitting a small amount of historical data.

[0067] S4: UAV Dynamics Modeling: Construct a dynamics model of the UAV using a physical information neural network, with the current state and control input as independent variables, and output the state derivative.

[0068] The predicted UAV state is calculated using a physical information neural network combined with the Rung-Kutta scheme. Specifically, for time step k, the predicted UAV state vector at time step i+1 is calculated. Where, k1 = h·PINN(t) k ,s i|k ,u i ),

[0069] k4 = h·PINN(t) k +h,s i|k +k3,u i+1 ), h = t k+1 -t k ,

[0070]

[0071] S5: Predictive Control Optimization: Establish a predictive control optimization model, use a rolling time domain approach to calculate the control vector for the next N time steps, and minimize the tracking error between the UAV and the predicted target as well as the control energy consumption.

[0072] In detail, the objective equation of the predictive control optimization model is: This NLP problem can be solved online using nonlinear solvers such as IPOPT, CasADi, and ACADO to obtain the optimal control sequence U. * .

[0073] Where, p i+1|k s is the target state vector at time step i+1 predicted from time step k; i+1|k U is the UAV state vector predicted from time step k at time step i+1; Q is the state error weight matrix, R is the control input weight matrix, which are set manually, and U = [u k ,u k+1 ,...,u k+N-1 ].

[0074] S6: Control Execution and State Update: Execute UAV flight control according to the control vector at each time step.

[0075] Specifically, when the error between the output UAV state of the UAV dynamics model and the measured UAV state exceeds a threshold at any time step, the network parameters θ of the physical information neural network NN(s,u;θ) need to be retrained, the weight of the prediction error of the most recent time step is increased, and the process returns to step S2.

[0076] S7: Repeat steps S2 to S6 until the UAV successfully tracks the target to the mission area or receives a termination command.

[0077] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A method for intelligent perception, prediction, and tracking of dynamic targets using unmanned aerial vehicles (UAVs), characterized in that: The method includes the following steps: S1: Mission Initialization: The UAV takes off and ascends to a safe altitude, initializes the communication link, positioning module and control unit, and enters the dynamic target tracking mission mode; S2: Target Status Acquisition: The UAV receives target status information from dynamic targets; it receives status data packets from dynamic targets through the onboard wireless communication module and parses them to obtain the target's real-time position and speed information; S3: Target trajectory prediction: Establish a Kalman filter to predict the target's trajectory in the next N time steps based on the target's historical motion data; S4: UAV Dynamics Modeling: A dynamics model of the UAV is constructed using a Physical Information Neural Network (PINN), with the current state and control input as independent variables, and the output state derivative. S5: Predictive Control Optimization: Establish a predictive control optimization model, use a rolling time domain approach to calculate the control vector for the next N time steps, and minimize the tracking error between the UAV and the predicted target as well as the control energy consumption; S6: Control Execution and State Update: Execute UAV flight control according to the control vector at each time step; S7: Repeat steps S2 to S6 to form an autonomous closed loop of "perception-prediction-planning-execution-update" until the UAV successfully tracks the target to the mission area or receives a termination command. The target state search target is at the current time t. k Position and velocity, i.e., p k =[x,y,z,v x ,v y ,v z ], where x, y, and z are the positions of the target in the x-coordinate direction, y-coordinate direction, and z-coordinate direction in the spatial coordinate system oxyz, and v x v y v z These represent the velocities in the x, y, and z coordinate directions. The drone state is the current time t. k The drone's position and speed, i.e. in, Let x be the position of the target in the x, y, and z coordinates of the spatial coordinate system oxyz. These represent the velocities in the x, y, and z coordinate directions. The control vector u k Let t be the current time. k The rotational speed of each motor in the drone.

2. The intelligent perception, prediction, and tracking method for dynamic targets of unmanned aerial vehicles according to claim 1, characterized in that, The state equation and measurement equation of the Kalman filter mentioned in step S3 are as follows: Where k represents the k-th time step, p k Let z be the target state vector at time step k. k+1 Let w be the target measurement vector at time step k+1. k For random system noise, v k+1 The noise is a random measurement; the system noise has a mean of zero and a covariance of Q. k The signal follows a Gaussian distribution, and the measurement noise has a mean of zero and a covariance of R. k The system noise follows a Gaussian distribution and is uncorrelated with the measurement noise; f(·) is the system state transformation function, and h(·) is the system state measurement function.

3. The intelligent perception, prediction, and tracking method for dynamic targets of unmanned aerial vehicles according to claim 2, characterized in that, The state transformation function f(·) and the state measurement function h(·) are functions with defined function types and adjustable parameters, and the parameters are set by fitting a small amount of historical data.

4. The intelligent perception, prediction, and tracking method for dynamic targets of unmanned aerial vehicles according to claim 1, characterized in that, The physical information neural network NN(s,u;θ) described in step S4 takes the state s and control vector u as input and outputs the derivative ds / dt of the UAV state. The network parameter is θ. The corresponding loss function is: Loss=α·Loss_data+β·Loss_physics, where α and β are manually set parameters, Loss_data=MSE(NN(s,u;θ),ds / dt_measured), where ds / dt_measured is the derivative of the measured UAV state, and Loss_physics=MSE(NN(s,u;θ),g(s,u)), where ds / dt=g(s,u) is the differential equation obtained from the analysis of the UAV dynamics model.

5. The intelligent perception, prediction, and tracking method for dynamic targets of unmanned aerial vehicles according to claim 1, characterized in that, The predicted state of the UAV is calculated using a physical information neural network combined with the Rung-Kuta format. Specifically, For the UAV state vector predicted at time step k at time step i+1 Where h = t k+1 -t k k1 = h·PINN(t) k ,s i|k ,u i ), k4 = h·PINN(t) k +h,s i|k +k3,u i+1 ), 6. The intelligent perception, prediction, and tracking method for dynamic targets of unmanned aerial vehicles according to claim 1, characterized in that, The objective equation of the predictive control optimization model described in step S6 is: This NLP problem is solved online using a nonlinear solver to obtain the optimal control sequence U. * ; where p i+1|k s is the target state vector at time step i+1 predicted from time step k; i+1|k U is the UAV state vector predicted from time step k at time step i+1; Q is the state error weight matrix, R is the control input weight matrix, which are set manually, and U = [u k ,u k+1 ,...,u k+N-1 ].

7. The intelligent perception, prediction, and tracking method for dynamic targets of unmanned aerial vehicles according to claim 1, characterized in that, In step S6, when the error between the Kalman filter prediction of the target state and the received target state at any time step exceeds a threshold or when the execution time step reaches the set upper limit N, the state transformation function f(·) and the state measurement function h(·) are updated, and the process returns to step S2.

8. The intelligent perception, prediction, and tracking method for dynamic targets of unmanned aerial vehicles according to claim 1, characterized in that, In step S6, when the error between the output drone state of the drone dynamics model and the measured drone state exceeds a threshold, the network parameter θ of the physical information neural network NN(s,u;θ) needs to be retrained, and the process returns to step S2.

9. An unmanned aerial vehicle system for performing the method as described in any one of claims 1-8, characterized in that, include: The signal receiving module is used to receive the status information of dynamic targets; The positioning module is used to obtain the drone's own position and speed; The data processing unit incorporates a Kalman filter, a physical information neural network model, and a predictive control optimization model. Flight control unit, used to execute control vector commands; The data processing unit is connected to the signal receiving module, the positioning module, and the flight control unit, respectively.

Images