Unmanned aerial vehicle target tracking method and system based on flight mode identification, and storage medium

Abstract

The invention discloses an unmanned aerial vehicle target tracking method and system based on flight mode identification and a storage medium, relates to the field of target tracking, and aims to solve the problem that the tracking precision of a multi-model algorithm is reduced due to the fact that the number of models is large and competition between the models is easy to cause in an existing method. Comprises: acquiring an RF signal of an unmanned aerial vehicle; identifying and positioning the flight mode of the unmanned aerial vehicle; constructing a target state equation and a measurement equation of the unmanned aerial vehicle; constructing a total model set comprising a flight mode and a motion model, and interacting with the state estimation value of the unmanned aerial vehicle target to obtain a state input value; based on the measurement value, the target state equation and the measurement equation, a Kalman filter and an unscented Kalman filter are combined, state input values of the linear model and the non-linear model are subjected to filtering processing, and state estimation values are obtained; updating the probability corresponding to each motion model in the total model set; and updating the state estimation value of the target at the current moment.

Description

Technical Field

[0001] The present invention relates to the field of target tracking technology, and in particular to a method and system for tracking unmanned aerial vehicle (UAV) targets based on flight mode recognition, and a storage medium. Background Art

[0002] With the widespread use of drones in remote sensing, resource exploration, emergency rescue, and intelligence gathering, their technology and applications continue to attract widespread attention, prompting a keen interest in drone tracking technology. Target tracking is a key technology that uses sensor data to determine a target's position, velocity, path, and state in a noisy environment and background noise. In traditional tracking systems, the motion of the target and the measurement data provided by the sensors often exhibit nonlinear characteristics, necessitating the use of linear and nonlinear filtering methods. The mathematical model of the moving target forms the basis of the filtering algorithm, which directly impacts the accuracy of the tracking results.

[0003] Drones face significant uncertainty during flight, including hovering, rapid acceleration, and sharp turns. Drones are highly maneuverable, their flight trajectories are complex and varied, and they often remain in a single motion state for very short periods of time. Consequently, drones often need to switch between different motion states in a very short period of time. Clearly, traditional interacting multiple model (IMM) algorithms no longer meet the current requirements of drone target tracking. Therefore, the introduction of auxiliary information is necessary to correct and improve these shortcomings of the IMM algorithm. How to achieve this correction and improvement is a crucial issue.

[0004] The unique characteristics of drone motion, including strong maneuverability, small size, and complex and variable flight trajectories, require the use of multiple motion models to fully describe their motion. However, using multiple models can lead to delays in switching to the model that best matches the target's motion within the multi-model algorithm, thus reducing tracking accuracy.

[0005] When the target exhibits a wide range of motion pattern changes or frequent mode switching, the effectiveness of traditional fixed model sets (such as first-generation and second-generation multi-model algorithms) is often limited. [1] This is because covering all potential target motion patterns may require a large model set, and competition between models can lead to reduced filtering performance. To address this challenge, the VSMM algorithm was developed. It is an advanced extension of the IMM algorithm, while traditional fixed-model multi-model algorithms such as the IMM can be considered a specific instance of the VSMM.

[0006] Knowledge-assisted adjustment of model update probabilities. In interacting multi-model algorithms, model probabilities are primarily estimated using filter residuals and filter covariance matrices. However, when tracking maneuvering targets, considering only these two factors cannot effectively reflect the target's motion, especially when the target undergoes state transitions. Summary of the Invention

[0007] The technical problems to be solved by the present invention are:

[0008] The existing multi-model tracking algorithms have a large number of models, which easily leads to competition between models, thereby reducing the tracking accuracy of the multi-model algorithm.

[0009] The present invention is to solve the above technical problems using the following technical solutions:

[0010] This paper proposes a flight mode assisted multi-model tracking algorithm (FMAMM). The probability obtained from flight mode recognition is introduced into the multi-model algorithm to adjust the motion model probability, including the following steps:

[0011] Step 1: Get the RF signal of the drone;

[0012] Step 2: Identify the UAV flight mode based on the flight mode recognition network and locate the UAV based on the UAV's RF signal; use the UAV flight mode and UAV position as measurement values;

[0013] Step 3: Construct the UAV target state equation and measurement equation, and initialize the UAV target state estimation value and covariance matrix;

[0014] Step 4: Construct a model set including flight modes and motion models. Each flight mode corresponds to a motion model subset, and interact with the state estimation value of the UAV target to obtain the state input value;

[0015] Step 5: Based on the measurement value, the target state equation and the measurement equation, a combination of a Kalman filter and an unscented Kalman filter is used to filter the state input values of the linear model and the nonlinear model respectively to obtain a state estimate;

[0016] Step 6: Update the probability corresponding to each motion model in the total model set;

[0017] Step 7: Based on the probability corresponding to each motion model, the state estimation value of each motion model, and the flight mode result, update the state estimation value of the target at the current moment.

[0018] Furthermore, the flight mode recognition network described in step 2 uses a Softmax function to identify the flight mode of the drone.

[0019] Furthermore, the UAV target state equation and measurement equation are constructed as described in step 3, specifically:

[0020] x k =F k|k-1 (θ t )x k-1 +w k

[0021] z k =H k (θ t )x k +u k

[0022] where x k is the M-dimensional state vector at time k, F k|k-1 is the one-step transfer matrix from time k-1 to time k, θ t is a finite-state Markov chain, w k The mean is 0 and the covariance matrix is Q k Process noise; z k is the N-dimensional measurement vector, H k is the measurement matrix at time k, u k The mean is 0 and the covariance matrix is R k The measurement noise, and w k and u k are two uncorrelated Gaussian noises.

[0023] Furthermore, the interaction of the state estimation value of the UAV target in step 4 includes:

[0024] There are L flight modes and K motion models in the UAV during flight. The total set of motion models is M = {m1, m2, ..., m K}, the flight mode set is C = {c1, c2, ..., c L}, each flight mode c l Corresponding to a subset M of the total set of motion models M l , let the motion model j = m j ∈M,i=mi∈M,m j State Estimation and the covariance matrix for:

[0025]

[0026] Furthermore, step 6 includes:

[0027] Assume that the flight mode c at time k l The recognition probability is P k (c l ), introduce P k (c l ) to adjust the model probability To make a correction:

[0028]

[0029] Furthermore,

[0030]

[0031] If the model probability of the IMM algorithm is recorded as μ1 and the model probability is μ2, it can be obtained:

[0032] μ2 ≥ μ1

[0033] So the posterior probability corresponding to model j becomes

[0034]

[0035] Furthermore, in step 7, the result state x at time k is output k and covariance P k The estimated value is:

[0036]

[0037] The present invention also provides a UAV target tracking system based on flight mode recognition, which has a program module corresponding to the steps of the method described in any one of the above technical solutions, and executes the steps in the above UAV target tracking method based on flight mode recognition during operation.

[0038] The present invention also provides a computer-readable storage medium, which stores a computer program. The computer program is configured to implement the steps of the drone target tracking method based on flight mode recognition described in any one of the above technical solutions when called by a processor.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention provides a method, system and storage medium for tracking unmanned aerial vehicle (UAV) targets based on flight mode recognition. Since flight mode information reflects the action that the UAV is about to perform, the present invention associates the flight mode with the corresponding model set, uses the flight mode recognition probability as auxiliary information to adjust the model probability, reduces the probability of impossible models, quickly switches to the model that best matches the target motion, and increases the model probability under a specific flight mode, thereby improving the accuracy and stability of model probability estimation, reducing model switching lag, and enhancing tracking performance.

[0041] The present invention lays a foundation for further introducing other auxiliary information that may help improve tracking accuracy into the IMM algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of a method for tracking a UAV target based on flight mode recognition in an embodiment of the present invention;

[0043] Figure 2 A four-channel RTL signal receiver in an embodiment of the present invention;

[0044] Figure 3 is a pattern recognition probability graph in an embodiment of the present invention;

[0045] Figure 4 This is a UAV flight target tracking trajectory diagram in an embodiment of the present invention;

[0046] Figure 5 This is a comparison chart of RMSE in terms of position between the two methods in the embodiment of the present invention;

[0047] Figure 6 : is a model probability transfer change curve diagram of the FMAMM algorithm in an embodiment of the present invention;

[0048] Figure 7 Figures 1 and 2 show the actual, measured, and two tracking trajectory diagrams of the UAV flight in an embodiment of the present invention;

[0049] Figure 8 RMSE comparison chart of FMAMM and IMM in position and velocity in an embodiment of the present invention;

[0050] Figure 9 Graph showing the model probability transition change of the FMAMM and IMM algorithms in an embodiment of the present invention;

[0051] Figure 10 Figures 1 and 2 show the actual, measured, and two tracking trajectory diagrams of the UAV flight in an embodiment of the present invention;

[0052] Figure 11RMSE comparison chart of FMAMM and IMM in position and velocity in an embodiment of the present invention;

[0053] Figure 12 4 is a graph showing the model probability transition change of the FMAMM and IMM algorithms in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below with reference to the accompanying drawings. Obviously, the described embodiments or examples are only some of the embodiments or examples of the present invention, and not all of them. Based on the embodiments or examples of the present invention, all other embodiments or examples obtained by those skilled in the art without creative work should fall within the scope of protection of the present invention.

[0055] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0056] Specific implementation plan 1: Figure 1 As shown, the present invention provides a UAV target tracking method based on flight mode recognition, comprising the following steps:

[0057] Step 1: Get the RF (Radio Frequency) signal of the drone;

[0058] Step 2: Identify the UAV flight mode based on the flight mode recognition network and locate the UAV based on the UAV's RF signal; use the UAV flight mode and UAV position as measurement values;

[0059] Step 3: Construct the UAV target state equation and measurement equation, and initialize the UAV target state estimation value and covariance matrix;

[0060] Step 4: Construct a model set including flight modes and motion models. Each flight mode corresponds to a motion model subset, and interact with the state estimation value of the UAV target to obtain the state input value;

[0061] Step 5: Based on the measurement value, the target state equation and the measurement equation, a combination of a Kalman filter and an unscented Kalman filter is used to filter the state input values of the linear model and the nonlinear model respectively to obtain a state estimate;

[0062] Step 6: Update the probability corresponding to each motion model in the total model set;

[0063] Step 7: Based on the probability corresponding to each motion model, the state estimation value of each motion model, and the flight mode result, update the state estimation value of the target at the current moment.

[0064] Specific implementation plan 2: The flight mode recognition network described in step 2 uses the Softmax function to make the output of the drone flight mode between (0, 1) to identify the drone flight mode.

[0065] The rest of this embodiment is the same as the specific embodiment 1.

[0066] Specific implementation plan three: Under the condition of no prior information, the discrete model of a typical UAV tracking system; the construction of the UAV target state equation and measurement equation described in step 3 is specifically:

[0067] x k =F k|k-1 (θ t )x k-1 +w k (1)

[0068] z k =H k (θ t )x k +u k (2)

[0069] where x k is the M-dimensional state vector at time k, F k-1 is the one-step transfer matrix from time k-1 to time k, θ t is a finite-state Markov chain, w k The mean is 0 and the covariance matrix is Q k The process noise represents the influence of the system noise from time k-1 to k on each state; k is the N-dimensional measurement vector, H k is the measurement matrix at time k, u k The mean is 0 and the covariance matrix is R k The measurement noise, and w k and u k are two uncorrelated Gaussian noises.

[0070] The rest of this implementation plan is the same as the second specific implementation plan.

[0071] Specific implementation plan 4: The interaction of the state estimation value of the UAV target described in step 4 includes:

[0072] There are L flight modes and K motion models in the UAV during flight. The total set of motion models is M = {m1, m2, ..., m K}, the flight mode set is C = {c1, c2, ..., c L}, each flight mode c l Corresponding to a subset M of the total set of motion models M l , from time k-1 to time k, motion model j = m j ∈M, i=m i ∈M,m j State Estimation and the covariance matrix for:

[0073]

[0074] The rest of this implementation plan is the same as the specific implementation plan three.

[0075] Specific implementation plan five: Step 6 includes:

[0076] Assume that the flight mode c at time k l The recognition probability is P k (c l ), introduce P k (c l ) to calibrate the model probability:

[0077]

[0078] Furthermore,

[0079]

[0080] From formula (7), it can be seen that the FMAMM algorithm of the present invention is closer to the probability of the real model than the model probability of the IMM. This shows that FMAMM switches to the real motion model faster, thus achieving better tracking performance than IMM. This is because when the UAV is in flight mode c l When the motion model j can only be in the model subset M l Select in M l The probability of motion models other than is 0. This makes the FMAMM algorithm more accurate in model selection.

[0081] If the model probability of the IMM algorithm is recorded as μ1 and the model probability of the FMAMM is recorded as μ2, it can be concluded that:

[0082] μ2≥μ1 (8)

[0083] So the posterior probability corresponding to model j becomes

[0084]

[0085] The rest of this implementation plan is the same as the specific implementation plan four.

[0086] Specific implementation plan six: In step 7, calculate the x at time k based on the model probability obtained after processing k and covariance P k for:

[0087]

[0088] The rest of this implementation plan is the same as the specific implementation plan five.

[0089] The UAV target tracking method (algorithm) based on flight mode recognition proposed in the present invention is the underlying technical core of the present invention, and various products can be derived based on the algorithm.

[0090] Based on the method proposed in the present invention, a UAV target tracking system based on flight mode recognition is developed using a programming language. The system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned UAV target tracking method based on flight mode recognition during operation.

[0091] The developed system (software) is stored on a computer-readable storage medium as a computer program. The computer program is configured to implement the steps of the aforementioned method for tracking unmanned aerial vehicle targets based on flight pattern recognition when invoked by a processor. This materializes the present invention on a carrier, becoming a computer program product.

[0092] Various implementations of the systems and techniques described herein can be realized in digital electronic circuitry, integrated circuitry, dedicated ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system comprising at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0093] The computer programs (also referred to as programs, software, software applications, or code) of the present invention include machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., a magnetic disk, an optical disk, a memory, a programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0094] The beneficial effects of the present invention will be described below with reference to specific embodiments.

[0095] Example 1

[0096] This embodiment builds a cost-effective multi-channel RF signal receiving system. By carefully processing the collected drone data, the compressed sensing data of the multi-channel random demodulator is extracted and the original signal is accurately reconstructed. The system is based on the software-defined radio (SDR) architecture and is centered around the cost-effective RTL-SDR device. Its open design makes it easy to add new features and perform system upgrades. Hardware devices such as Figure 2 shown.

[0097] Experimental simulations were conducted on a computer equipped with an Intel i7 3.0GHz CPU and 16GB of RAM, using Matlab R2006a software. By combining RTL devices and Matlab tools, a multi-channel system was constructed and experimentally verified. The OMP (Operational Multi-Purpose) signal reconstruction algorithm was used to reconstruct the compressed sensing signal. The experimental results validated the effectiveness of this sampling architecture for signal sampling. Finally, single-channel and multi-channel drone detection and identification were performed on drone data collected from a multi-channel random demodulator.

[0098] When obtaining the pattern recognition probability, three different types of drones (AR, Bebop, and Phantom) were included, as well as signal records of each drone in different flight modes (open connection close, hover, flight, and video recording flight); the flight mode recognition accuracy of Bebop and AR reached 98.7% and 99.4% respectively. However, these results were obtained through subsequent processing. Figure 3As shown in the figure, the actual output results are different. This is because the Softmax function is used in the drone flight mode recognition model, which makes the output of the drone flight mode between (0, 1) and the sum of all elements is 1.

[0099] Once a drone is detected, its flight mode is identified and located. The flight mode information represents the operation that the drone will perform at the current moment. In order to reduce the switching delay between models, the FMAMM algorithm of the present invention uses the probability of pattern recognition to adjust the probability of the model. During the FMAMM tracking process, the states corresponding to different flight modes of the drone will also be different. This means that a flight mode may correspond to one or more motion models, and similarly, a motion model may exist in different flight modes. Taking into account this association between flight mode and motion model, this embodiment classifies it into three scenarios:

[0100] (1) One flight mode corresponds to one motion model, which is recorded as one-to-one;

[0101] (2) One flight mode corresponds to multiple motion models, which is recorded as one-to-many;

[0102] (3) One flight mode corresponds to multiple motion models, and one motion model is included in multiple flight modes, which is recorded as many-to-many.

[0103] Next, detailed simulation experiments will be carried out for the above three scenarios.

[0104] One-on-one experiment:

[0105] After detecting the drone, the RF data is simultaneously identified and located in the same time period. While estimating the drone’s position, the probability P of the drone in three flight modes (related to tracking) is obtained. k (c l In the one-on-one experiment, the drone departs from the position (10, 10, 10) and hovers under c1 for (0-20s). It then makes a left turn with ω = -3° under c2 for (21-50s), hovers under c1 for (51-70s), and makes a right turn with ω = 3° under c3 for (71-100s), for a total of 100s, as shown in Table 1.

[0106] Table 1 Changes in the trajectory of UAV targets 1

[0107]

[0108] Set the flight mode set C = {c1, c2, c3} and its corresponding motion model set M = {M1, M2, M3}, where M1 = {m HM}, M2={m CT1}, M3={m CT2}, the total set of models is:

[0109] M={M1,M2,M3}={m HM , m CT1 , m CT2} (14)

[0110] The transition probability matrix [P ij ] Initialization is set according to the fixed value method, the main diagonal element p ii =q i , the off-diagonal elements p ij =(1-q i ) / (r-1), so:

[0111]

[0112] Assuming an observation error of 10 meters and a sampling interval of T = 1 second, the HM model uses KF filtering, while the CT1 and CT2 models use UKF filtering. Monte Carlo simulations are performed 100 times and evaluated using the root mean square error (RMSE). Figure 4 The real trajectory and tracking trajectory of the drone flight for the two methods are: Figure 5 is the RMSE curve of the two tracking algorithms, Figure 6 is the change in model probability transfer.

[0113] One-to-many experiment:

[0114] In the one-to-many experiment, the drone starts from the position (0,0,0) and performs uniform linear motion (0-20s) in C2 flight mode with an initial velocity of (6,6,3)m / s. After quickly flying to the designated destination, it performs uniform deceleration linear motion (21-26s) with an acceleration of (-1,-1,-0.5)m / s. 2 After the speed drops to 0, the drone enters hovering mode (27-50s); after hovering, it performs uniformly accelerated linear motion in C2 flight mode (51-57s), with an acceleration of (1, 1, 0) m / s 2 After the acceleration is completed, a left turn of ω = -7° is performed (58-75 seconds) and the video recording flight mode is entered. Finally, a right turn is performed in the video recording flight mode (76-100 seconds), for a total of 100 seconds, as shown in Table 2.

[0115] Table 2 Changes in the trajectory of UAV targets

[0116]

[0117] Assume that the flight mode set is C = {c1, c2, c3}, and its corresponding motion model set is M = {M1, M2, M3}, where: hovering mode c1: the velocity and acceleration are both 0, so M1 is composed only of HM models; flight mode c2 and flight mode with video recording c3 may contain multiple motion models, and a motion model may exist in both flight modes at the same time. Assume M2 = {m CV , m CA}, M3={m CT1 , m CT2}, the total set of models is:

[0118] M={M1,M2,M3}={m HM ,m CV , m CT1 , m CT2 , m CA} (16)

[0119] The transition probability matrix [P ij ]for

[0120]

[0121] Assuming an observation error of 10 meters and a sampling interval of T = 1 second, the CV and HM models use Kalman filtering, while the CT and CA models use UKF filtering. Monte Carlo simulations are performed 100 times and evaluated using the root mean square error (RMSE). Figure 7 The real trajectory and tracking trajectory of the drone flight under the two methods are shown. Figure 8 The RMSE curves of the two tracking algorithms are shown. Figure 9 This shows the change in the model probability transfer.

[0122] Many-to-many experiments:

[0123] In the many-to-many experiment, the drone starts at the position (10, 10, 10) and performs uniform linear motion (0-20s) in flight mode C2 with an initial velocity of (6, 6, 3) m / s. When approaching the designated destination, it performs uniform deceleration linear motion (21-26s) with an acceleration of (-1, -1, -0.5) m / s. 2 When the speed drops to 0, the drone enters hovering mode (27-50s); after hovering, the drone performs uniformly accelerated linear motion (51-57s) with an acceleration of (1, 1, 0) m / s 2 After the acceleration is completed, a turn of ω = -7° is performed (58-75 seconds), and the video recording flight mode is entered. Finally, a uniform linear motion is performed in the video recording flight mode (76-100 seconds), for a total of 100 seconds, as shown in Table 3.

[0124] Table 3 Changes in the trajectory of UAV targets

[0125]

[0126] Assume that the flight mode set C = {c1, c2, c3}, and the corresponding total set of motion models is M = {M1, M2, M3}, where: hovering mode c1: the speed and acceleration are both 0, so M1 is composed only of HM models; flight c2 and flight with video recording c3 may contain multiple motion models, and a motion model may exist in both flight modes at the same time. Therefore, suppose that c2 and c3 both contain CV models, CA is unique to c2 (generally, the drone does not record video when accelerating), and CT is unique to c3. That is, M2 = {m CV , m CA}, M3={m CV ,m CT}, the total set of models is:

[0127] M={M1,M2,M3}={m HM , m CV ,m CT ,,m CA} (18)

[0128] The transition probability matrix [P ij ] Initialization is set according to the fixed value method, the main diagonal element p ii =q i , the off-diagonal elements p ij =(1-q i ) / (r-1), therefore,

[0129]

[0130] Assuming an observation error of 10 meters and a sampling interval of T = 1 second, the CV and HM models use Kalman filtering, while the CT and CA models use UKF filtering. Monte Carlo simulations are performed 100 times, and the root mean square error (RMSE) is used for evaluation. Figure 10 The real trajectory and tracking trajectory of the drone flight for the two methods are: Figure 11 is the RMSE curve of the two tracking algorithms, Figure 12 is the change in model probability transfer.

[0131] like Figure 11 As shown in Figure 2, the RMSE of the FMAMM algorithm in both position and velocity is lower than that of the IMM algorithm, which effectively proves that the tracking accuracy of the FMAMM algorithm is effectively improved compared to the IMM algorithm. In particular, during the 0-20s, when the drone is in flight mode c2, the corresponding motion model subset does not include the CT and HM models, as shown in Figure 2. Figure 12 It can also be found that in this time period, the probability of the CT and HM models in FMAMM is close to 0, and the model probability of the matching model CV increases accordingly. Figure 12 As can be seen from the figure, when the drone switches from flight mode c2 to c1, the drone quickly switches to the matching model HM. Because the model subset M1 has only one model HM, the probability of other mismatching models is close to 0. This is also the case in Figure 11 One of the key reasons why the RMSE of FMAMM is much lower than that of the IMM method is that at 51 seconds, when the drone switches from c1 to c2, the IMM algorithm is still influenced by the previous model HM and takes about 3 seconds to switch to the matching model CA. However, the FMAMM algorithm, by adjusting the pattern recognition probability, effectively reduces the influence of the mismatching model and quickly switches to the matching model CA. The same situation occurs at 75 seconds.

[0132] In summary, detailed simulations of the three scenarios above demonstrate that the FMAMM algorithm effectively excludes motion models that do not correspond to the current flight mode, ensuring that the filtering process is not affected by mismatched models. Furthermore, the algorithm enhances the probability of motion model matching, bringing the probability of the motion model closer to that of the true model during iteration. These optimizations collectively improve the system's tracking accuracy.

[0133] The motion model used in the algorithm of this embodiment is:

[0134] Drones (UAVs) are characterized by their high maneuverability, compact size, and ever-changing flight paths. Furthermore, the environments they fly in and the objects they encounter are even more complex. Accurately describing UAV motion requires the integration of multiple motion models. However, due to the unique characteristics of UAV flight, a complete description of the flight state requires a large number of motion models, which leads to a bloated set of motion models, causing competition between them and switching lag. Therefore, tracking algorithms must not only design a model set that covers the actual trajectory variations of UAVs, but also minimize competition between them and increase the probability of matching models.

[0135] Models include: Constant Velocity (CV), Constant Acceleration (CA), Coordinated Turn (CT), Singer, and Current Statistical (CS). This example uses four motion models for analysis: Hover (HM), 3-Dimensional Constant Velocity (3DCV), 3-Dimensional Constant Acceleration (3DCA), and 3-Dimensional Coordinated Turn (3DCT).

[0136] The specific process of pattern recognition probability assisted multi-model filtering in this embodiment is as follows:

[0137] In this embodiment, the current flight mode of the UAV is obtained through the UAV recognition network. The Softmax function is used in the UAV recognition network to make the range of each flight mode between (0, 1) and the sum of all elements is 1. The mode recognition probability is P{c k This embodiment uses the pattern recognition probability as measurement information and, based on the correspondence between the unmanned flight mode and the motion model, performs auxiliary adjustment on the motion model probability obtained by the likelihood function and the filter residual, as shown below.

[0138] In one filtering cycle, from one measurement update to the next measurement update, the mixed state Markov process (x k ,θ k ) is recursively looped, where x k is the state estimate, θ k is a motion model. Assume that there are L flight modes and K motion models in the UAV during flight. The total set of motion models is marked as M = {m1, m2, ..., m K}, the flight mode set is C = {c1, c2, ..., c L}, each flight mode c l Corresponding to a subset M of the total set of motion models M l :

[0139]

[0140] Let z k ={Y k-1 , c k ,y k}, according to P{ck} under θ k The probability P{θ k |c k} for θ k Make adjustments:

[0141]

[0142]

[0143] Through the total probability formula, we get:

[0144]

[0145] In the above loop derivation, since w k and u k are independent of each other. Through the Bayes formula, equations (35) and (36) can be easily derived from equation (2), and equations (32) to (34) can be derived from equation (1).

[0146]

[0147] In the formula

[0148]

[0149] in

[0150]

[0151] P{c k =l} is the UAV’s pattern recognition probability, P{θ k =i|c k =l} represents the proportion of each motion model in flight mode l at time k-1. Each flight mode contains one or more motion models.

[0152]

[0153] p[x k-1 |θ k , z k-1 ]=p[x k-1 |θ k , Y k-1 , c k ,y k ] (43)

[0154] represents a kind of mixing in formula (33). And formula (34) is because c k With the state estimate x k are independent of each other, and knowing θ k In the case ofk Also independent of x k-1 , so Equation (34) is the same as that of the traditional IMM algorithm.

[0155]

[0156] Although the present invention is disclosed as above, the scope of protection disclosed by the present invention is not limited thereto. Those skilled in the art of the present invention may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the scope of protection of the present invention.

[0157] The documents cited in this invention include:

[0158] [1] J.Ru and

Claims

1. A UAV target tracking method based on flight mode recognition, characterized in that: The steps include: Step 1: Get the RF signal of the drone; Step 2: Identify the UAV’s flight mode based on the flight mode recognition network, locate the UAV based on its RF signal, and use the UAV’s flight mode and position as measurement values. Step 3: Construct the UAV target state equation and measurement equation, and initialize the UAV target state estimation value and covariance matrix; Step 4: Construct a model set including flight modes and motion models. Each flight mode corresponds to a motion model subset, and interact with the state estimation value of the UAV target to obtain the state input value; Step 5: Based on the measurement value, the target state equation and the measurement equation, a combination of a Kalman filter and an unscented Kalman filter is used to filter the state input values of the linear model and the nonlinear model respectively to obtain a state estimate; Step 6: Update the probability corresponding to each motion model in the total model set; Step 7: Based on the probability corresponding to each motion model, the state estimation value of each motion model, and the flight mode result, update the state estimation value of the target at the current moment.

2. The method for tracking a UAV target based on flight mode recognition according to claim 1, characterized in that: The flight mode recognition network described in step 2 uses the Softmax function to identify the flight mode of the drone.

3. The method for tracking unmanned aerial vehicle targets based on flight mode recognition according to claim 2, characterized in that: The target state equation and measurement equation of the UAV are constructed as described in step 3, specifically: x k =F k|k-1 (θ t )x k-1 +w k z k =H k (θ t )x k +u k where x k is the M-dimensional state vector at time k, F k|k-1 is the one-step transfer matrix from time k-1 to time k, θ t is a finite-state Markov chain, w k The mean is 0 and the covariance matrix is Q k Process noise; z k is the N-dimensional measurement vector, H k is the measurement matrix at time k, u k The mean is 0 and the covariance matrix is R k The measurement noise, and w k and u k are two uncorrelated Gaussian noises.

4. The method for tracking unmanned aerial vehicle targets based on flight mode recognition according to claim 3, characterized in that: The interaction of the state estimate of the UAV target described in step 4 includes: There are L flight modes and K motion models in the UAV during flight. The total set of motion models is M = {m1, m2, ..., m K }, the flight mode set is C = {c1, c2, ..., c L }, each flight mode c l Corresponding to a subset M of the total set of motion models M l , let the motion model j = m j ∈M, i=m i ∈M,m j State Estimation and the covariance matrix for:

5. The method for tracking a UAV target based on flight mode recognition according to claim 4, characterized in that: Step 6 includes: Assume that the flight mode c at time k l The recognition probability is P k (c l ), introduce P k (c l ) to adjust the model probability To make a correction: Furthermore, If the model probability of the IMM algorithm is recorded as μ1 and the model probability is μ2, it can be obtained: μ2 ≥ μ1 So the posterior probability corresponding to model j becomes 6. The method for tracking unmanned aerial vehicle targets based on flight mode recognition according to claim 5, characterized in that: In step 7, the result state x at time k is output k and covariance P k The estimated value is:

7. A UAV target tracking system based on flight mode recognition, characterized in that: The system has a program module corresponding to the steps of the method described in any one of claims 1 to 6, and executes the steps of the above-mentioned UAV target tracking method based on flight mode recognition when running.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the drone target tracking method based on flight mode recognition according to any one of claims 1 to 6 when called by a processor.

Images