Unmanned aerial vehicle flight path tracking method and device based on neural network and adaptive Kalman filtering
By combining neural networks and adaptive Kalman filtering, and using long short-term memory neural networks for error compensation, the problem of low drone tracking accuracy was solved, achieving a continuous, stable, and accurate drone tracking effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-03-10
AI Technical Summary
Existing UAV tracking methods suffer from low tracking accuracy and an inability to continuously and stably track targets. In particular, when the UAV's motion is highly nonlinear, the extended Kalman filter algorithm introduces truncation error, causing the filtering results to diverge.
A method based on neural networks and adaptive Kalman filtering is adopted. By acquiring measurement data from monitoring equipment, error compensation is performed using a long short-term memory neural network, and the noise covariance is adjusted in real time by combining the adaptive Kalman filtering algorithm to improve the accuracy of state estimation.
It achieves continuous, stable, and accurate drone tracking, can quickly capture dynamic changes, improves the accuracy of state estimation vectors, and provides an effective data foundation for accurate target tracking.
Smart Images
Figure CN121639745A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of drone tracking technology, and in particular to a drone trajectory tracking method and device based on neural networks and adaptive Kalman filtering. Background Technology
[0002] The widespread application of drone technology has brought convenience to our daily lives, but it has also raised challenges in terms of safety and security. Effective detection and tracking of unauthorized drones is a crucial aspect of airspace control and preventing potential risks. Radar is one of the main means of tracking drones, but drone targets are typically small, fast, and highly maneuverable, resulting in weak radar echo signals, low signal-to-noise ratios, and their motion exhibiting typical nonlinear characteristics.
[0003] To estimate the state of a UAV's nonlinear motion, the Extended Kalman Filter (EKF) algorithm is widely used. EKF approximates the nonlinear function as a linear model by performing a first-order Taylor series expansion at the current estimation point, and then applies the standard Kalman filter framework. However, this linearization approximation introduces significant truncation errors when the system has a high degree of nonlinearity. This not only reduces tracking accuracy but can also cause the filtering results to diverge in severe cases, making it impossible to continuously and stably track the target. Summary of the Invention
[0004] To address the aforementioned problems and technical requirements, this applicant proposes a method and device for UAV trajectory tracking based on neural networks and adaptive Kalman filtering. This method solves the problems of low tracking accuracy and inability to continuously and stably track targets when using extended Kalman filtering algorithms in existing technologies, thereby achieving continuous, stable, and accurate UAV tracking.
[0005] This application provides a method for tracking UAV tracks based on neural networks and adaptive Kalman filtering, the method comprising: Acquire measurement data at time k when the monitoring equipment is monitoring the UAV; Based on the process noise covariance at time k-1, the measurement noise covariance at time k-1, and the preset adaptive Kalman filter algorithm, the measurement data is processed to obtain a preliminary state estimation vector. Obtain the historical information sequence corresponding to the preset time window, and input the historical information sequence into the preset long short-term memory neural network to obtain the prediction error output by the long short-term memory neural network. The information is obtained based on the measurement data, and the prediction error reflects the estimation error of the adaptive Kalman filter algorithm in estimating the state of the measurement data. The initial state estimation vector is corrected using the prediction error to obtain the target state estimation vector.
[0006] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the unmanned aerial vehicle trajectory tracking method based on neural networks and adaptive Kalman filtering as described above.
[0007] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the unmanned aerial vehicle trajectory tracking method based on neural networks and adaptive Kalman filtering as described above.
[0008] The UAV trajectory tracking method and device based on neural networks and adaptive Kalman filtering provided in this application acquires measurement data at time k when the monitoring device monitors the UAV. Based on the process noise covariance at time k-1, the measurement noise covariance at time k-1, and a preset adaptive Kalman filtering algorithm, the measurement data is processed to obtain a preliminary state estimation vector. This application uses the process noise covariance and measurement noise covariance adjusted in real time to predict the preliminary state estimation vector of the measurement data, satisfying the various motion processes of the UAV's sublinear motion, enabling timely response to measurement data, rapid capture of the UAV's dynamic changes, and ensuring continuous and accurate UAV tracking. The application acquires the historical information sequence corresponding to a preset time window and inputs the historical information sequence into a preset long short-term memory neural network to obtain the prediction error output by the long short-term memory neural network. The prediction error is used to correct the preliminary state estimation vector to obtain the target state estimation vector. This application uses the prediction error of the neural network to correct the preliminary state estimation vector, improving the accuracy of the state estimation vector and providing an effective data foundation for accurate target tracking, thus achieving continuous, stable, and accurate UAV tracking. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0010] Figure 1 This is a flowchart illustrating the UAV trajectory tracking method based on neural networks and adaptive Kalman filtering provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the UAV trajectory tracking device based on neural network and adaptive Kalman filter provided in the embodiments of this application; Figure 3This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation
[0011] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0012] To clearly illustrate this application, the prior art is further described below: To overcome the limitations of EKF, Unscented Kalman Filtering (UKF) was proposed. UKF approximates the posterior probability distribution of the state through an unscented transformation and a set of Sigma sampling points. This method achieves third-order accuracy (for Gaussian distributions), far superior to the first-order linearization approximation of EKF. Nevertheless, the standard UKF algorithm still relies on a key assumption: the statistical characteristics of the system's process noise (Q) and measurement noise (R) are fixed and known a priori. This assumption often does not hold true in the real world. For example, when a UAV suddenly switches from stable cruise to high-maneuvering actions such as emergency evasion, its actual motion model deviates from the preset model. The fixed process noise Q cannot reflect this model mismatch, leading to a sluggish filter response. Similarly, sensors such as radar may be affected by time-varying external environmental interference, and the fixed measurement noise R cannot adapt to this. This lack of adaptability to dynamic models and noisy environments is the root cause of the performance bottleneck of standard UKF in complex scenarios.
[0013] To further improve tracking accuracy, researchers have begun exploring hybrid models that combine deep learning with filtering algorithms, among which Long Short-Term Memory (LSTM) networks have attracted considerable attention due to their powerful temporal data processing capabilities. These methods typically utilize LSTMs to learn and compensate for the residual errors of the underlying filters (such as EKF or UKF).
[0014] However, these existing hybrid models have two shortcomings: First, the underlying filters upon which the hybrid models rely suffer from the aforementioned problem of insufficient adaptability to dynamic changes, resulting in low-quality input benchmarks provided to the LSTM. Second, standard LSTM networks, when processing the input time series, assign relatively equal weights to all historical information points in the sequence, failing to effectively distinguish and focus on key historical nodes that have the greatest impact on the prediction of the current state, such as sharp turning points or acceleration / deceleration phases in the trajectory. This indiscriminate processing method limits the accuracy and efficiency of the error compensation model.
[0015] To address the aforementioned problems, this application provides a UAV trajectory tracking method based on neural networks and adaptive Kalman filtering. This method can be applied to smart terminals and servers. This application uses the application of this method in a server as an example for illustration, and some other descriptions in the embodiments are illustrative and not intended to limit the scope of protection of this application, and will not be described in detail thereafter. The specific implementation of the method is as follows... Figure 1 As shown: Step 101: Obtain the measurement data at time k when the monitoring equipment is monitoring the UAV.
[0016] Step 102: Based on the process noise covariance at time k-1, the measurement noise covariance at time k-1, and the preset adaptive Kalman filter algorithm, process the measurement data to obtain a preliminary state estimation vector.
[0017] Step 103: Obtain the historical information sequence corresponding to the preset time window, and input the historical information sequence into the preset long short-term memory neural network to obtain the prediction error output by the long short-term memory neural network.
[0018] Among them, the new information is obtained based on the measurement data, and the prediction error reflects the estimation error of the adaptive Kalman filter algorithm in estimating the state of the measurement data.
[0019] Step 104: Correct the preliminary state estimation vector using the prediction error to obtain the target state estimation vector.
[0020] The UAV trajectory tracking method based on neural networks and adaptive Kalman filtering provided in this application acquires measurement data at time k when the monitoring device monitors the UAV. Based on the process noise covariance at time k-1, the measurement noise covariance at time k-1, and a preset adaptive Kalman filtering algorithm, the measurement data is processed to obtain a preliminary state estimation vector. This application uses real-time adjusted process noise covariance and measurement noise covariance to predict the preliminary state estimation vector of the measurement data, satisfying the various motion processes of the UAV's sublinear motion. This allows for timely response to measurement data, rapid capture of the UAV's dynamic changes, and ensures continuous and accurate UAV tracking. The method acquires the historical information sequence corresponding to a preset time window and inputs it into a preset long short-term memory neural network to obtain the prediction error output by the long short-term memory neural network. The prediction error is used to correct the preliminary state estimation vector to obtain the target state estimation vector. This application uses the prediction error of the neural network to correct the preliminary state estimation vector, improving the accuracy of the state estimation vector and providing an effective data foundation for accurate target tracking, achieving continuous, stable, and accurate UAV tracking.
[0021] The following section introduces the physical model of constant acceleration for UAVs, and elaborates on this application based on it: First, define the state vector X of the drone, for example... .
[0022] in, Let this be the position of the drone in the Cartesian coordinate system. The speed of the drone.
[0023] The equation of motion for the UAV is given by formula (1): ...(1) Wherein, the state transition matrix Control Matrix .
[0024] in, It is the time step.
[0025] In one specific embodiment .
[0026] Among them, control vector , For the acceleration of the drone, This represents the control vector at time k.
[0027] The noise is assumed to be zero-mean Gaussian white noise with a covariance of . , This indicates that at time k, the process noise is a random variable with a mean of 0 and a covariance matrix of... The normal distribution (Gaussian distribution). It changes with time k.
[0028] The observation equation for the UAV is shown in formula (2): …(2) in, Indicates the distance between the radar and the drone. , , Indicates the azimuth angle in different directions detected by the radar. The measurement noise is assumed to be zero-mean Gaussian white noise with a covariance of . , , This indicates that at time k, the measurement noise is a random variable with a mean of 0 and a covariance matrix of... The normal distribution (Gaussian distribution). It changes with time k.
[0029] The following is a detailed explanation of this application based on a constant acceleration physical model: In one specific embodiment, the specific implementation of processing measurement data and obtaining a preliminary state estimation vector based on the process noise covariance at time k-1, the measurement noise covariance at time k-1, and a preset adaptive Kalman filter algorithm includes: Multiple Sigma points are generated based on the dimensions of the measurement data, and each Sigma point is propagated using a pre-defined nonlinear motion model to obtain the predicted state of the UAV, and the predicted parameters are obtained based on the predicted state and / or the process noise covariance at time k-1. The propagated Sigma points are mapped to the measurement space using a pre-defined measurement function to obtain the mapped state, and the mapped parameters are obtained based on the mapped state and / or the measurement noise covariance at time k-1. The cross-covariance matrix corresponding to the predicted state, mapped state, predicted parameters, and mapped parameters is calculated, and the Kalman gain is calculated based on the cross-covariance matrix. Based on the Kalman gain, measurement data, and mapping parameters, a preliminary state estimation vector is obtained.
[0030] In one specific embodiment, the prediction parameters include: a second mean and a first covariance corresponding to the predicted state.
[0031] Specifically, Sigma points are generated. For example, taking the dimension of the measurement data L=6 as an example, 2L+1=13 points need to be generated.
[0032] See formula (3) for details: ... (3) in, , is a scaling parameter. Let represent the posterior covariance at time k-1.
[0033] Specifically, the weights corresponding to each Sigma point are calculated, including mean weights and covariance weights.
[0034] The formula for calculating the mean weight is shown in formula (4): ………………………………(4) The formula for calculating covariance weights is given in formula (5): ………………………………(5) in, , Both b and are adjustment parameters. ,b , .
[0035] Specifically, each Sigma point is propagated using formula (6): …………………………(6) The second mean corresponding to the predicted state is calculated based on the first mean calculation formula, and the first covariance corresponding to the predicted state is calculated based on the first covariance calculation formula.
[0036] The formula for calculating the first mean is shown in formula (7): ………………………………(7) in, This represents the second mean. Indicates the dimension of the measurement data. This represents the mean weight corresponding to the Sigma point. This indicates the predicted state.
[0037] The formula for calculating covariance is given in formula (8): …(8) in, Indicates the second covariance. This represents the covariance weight corresponding to the Sigma point. This represents the process noise covariance at time k-1.
[0038] In one specific embodiment, the mapping parameters include: the first mean, the first mean, and the new information covariance corresponding to the mapping state.
[0039] Specifically, the predicted Sigma point cloud The measurement space is mapped through a measurement function, where the measurement function is given by formula (9): …………………………(9) In one specific embodiment, the first mean corresponding to the mapping state is calculated based on the second mean calculation formula, and the innovation covariance corresponding to the mapping state is calculated based on the second covariance calculation formula.
[0040] The formula for calculating the second mean is shown in formula (10): ………………………………(10) in, This represents the first mean. Indicates the dimension of the measurement data. This represents the mean weight corresponding to the Sigma point. Indicates the mapping state.
[0041] The formula for calculating the second covariance is given in formula (11): …(11) in, Indicates the new information covariance. This represents the covariance weight corresponding to the Sigma point. This represents the measurement noise covariance at time k-1.
[0042] In one specific embodiment, the prediction parameters include: a second mean corresponding to the predicted state, and the mapping parameters include: a first mean corresponding to the mapped state.
[0043] The specific implementation of calculating the cross-covariance matrix corresponding to the predicted state, the mapped state, the predicted parameters, and the mapped parameters, and calculating the Kalman gain based on the cross-covariance matrix includes: Input the predicted state and the mapped state into the preset third covariance calculation formula to obtain the cross-covariance matrix output by the third covariance calculation formula.
[0044] The formula for calculating the third covariance is given in formula (12): ………(12) in, Represent the cross-covariance matrix. Indicates the dimension of the measurement data. This represents the covariance weight corresponding to the Sigma point. Indicates the predicted state. This represents the second mean corresponding to the predicted state. Indicates the mapping state. This represents the first mean value corresponding to the mapped state.
[0045] In one specific embodiment, the cross-covariance matrix is input into the Kalman gain calculation formula to obtain the Kalman gain output by the Kalman gain calculation formula.
[0046] The formula for calculating the Kalman gain is shown in formula (13): ……………………………………………(13) in, Indicates Kalman gain, This represents the new information covariance corresponding to the mapped state.
[0047] In one specific embodiment, the mapping state is corrected using measurement data to obtain a preliminary state estimation vector.
[0048] Specifically, the Kalman gain, measurement data, and mapping parameters are input into the state estimation formula to obtain the preliminary state estimation vector output by the state estimation formula.
[0049] The state estimation formula is given in formula (14): ………………………………(14) in, This represents the initial state estimation vector at time k. This represents the initial state estimation vector at time k-1. Indicates Kalman gain, Represents measurement data, This represents the first mean value corresponding to the mapped state.
[0050] In one specific embodiment, after obtaining the preliminary state estimation vector, the posterior covariance is updated based on the covariance update formula.
[0051] Among them, the posterior covariance is used to generate the Sigma point.
[0052] The covariance update formula is shown in formula (15): ………………………………………………(15) in, This represents the posterior covariance at time k. This represents the posterior covariance at time k-1. Indicates Kalman gain, This represents the new information covariance corresponding to the mapped state.
[0053] In one specific embodiment, after obtaining the preliminary state estimation vector, the innovation corresponding to time k is obtained based on the innovation calculation formula.
[0054] The formula for calculating the new information is shown in formula (16): …………………………………………(16) in, This represents the information at time k. Represents measurement data, This represents the first mean value corresponding to the mapped state.
[0055] In one specific embodiment, after obtaining the innovation corresponding to time k based on the innovation calculation formula, the noise covariance to be used at the next time is estimated using the innovation at the current time.
[0056] Specifically, the measurement noise covariance at time k is evaluated based on the measurement noise covariance evaluation formula, and the process noise covariance at time k is evaluated based on the process noise covariance evaluation formula.
[0057] The formula for evaluating the measurement noise covariance is shown in formula (17):
[0058] ….(17) Among them, represents the measurement noise covariance corresponding to the k-th moment, represents the weight factor, represents the measurement noise covariance corresponding to the (k - 1)-th moment, represents the dimension of the measurement data, represents the covariance weight corresponding to the Sigma point, represents the mapping state, represents the first mean corresponding to the mapping state.
[0059] In order to smooth the estimation and prevent the estimation value from fluctuating violently due to accidental measurement errors, a forgetting factor f (0 < f < 1) is introduced. Among them, is a time-varying weight factor, which can also be simplified to a constant .
[0060] Among them, the evaluation formula of the process noise covariance is shown in formula (18): ………………………………(18) Among them, represents the process noise covariance corresponding to the k-th moment, represents the process noise covariance corresponding to the (k - 1)-th moment, represents the Kalman gain.
[0061] After the present application completes the estimation of the preliminary state estimation vector, it will perform a noise covariance adaptive update innovation. Using the innovation and innovation covariance at the current moment. The noise covariance matrix is iteratively updated through the innovation self-adaptive algorithm.
[0062] A simplified update strategy is that if the squared Mahalanobis distance of the innovation continuously exceeds the preset threshold value, then increase the main diagonal elements of and by a proportionality factor.
[0063] In a specific embodiment, the specific implementation of using a long short-term memory neural network to predict the prediction error (error compensation value) includes: The SA-LSTM network consists of an input layer, one or more LSTM layers, an attention layer, and an output layer, and can be implemented using frameworks such as Keras, TensorFlow, or Pytorch.
[0064] (1) Input layer: At the k-th moment, input a historical innovation sequence contained within a time window of length N into the network In this embodiment, N=20, meaning the model will review the filter performance over the past 2 seconds (20*0.1s) to predict the current error.
[0065] (2) LSTM layer: The LSTM layer acts as an encoder to process the historical information input sequence. The network will capture the time dependencies in the sequence:
[0066] Output the hidden state sequence for each time step i: .
[0067] (3) Attention layer: Since not all historical information contributes equally to the current error, an attention mechanism is introduced to dynamically focus on the most important historical time step by calculating a weight distribution.
[0068] The attention layer computes each hidden state. With the final hidden state of the sequence Similarity score (or a learnable query vector) These scores are normalized using the Softmax function to obtain the attention weights. The weights reflect the contribution of each point in the historical information sequence to the current error prediction. Then, a context vector is generated by weighted summation of all hidden states. .
[0069] (4) Output layer: Finally, The input consists of a decoder with one or more fully connected layers, and the output layer is this context vector that focuses on key information. Mapped to the predicted value of the A-UKF error at the current time. .
[0070] In one specific embodiment, the specific implementation of correcting the preliminary state estimation vector using the prediction error to obtain the target state estimation vector includes: Will As the final state estimate, it is better than the preliminary state estimate. It is closer to the real state of drones.
[0071] Training of SA-LSTM networks is typically done offline. Specifically, a training dataset containing a large number of real UAV flight paths is prepared. For each flight path in the dataset, the A-UKF filtering unit 121 of this invention is first run to obtain a series of preliminary state estimates. Then, With actual flight path By comparison, the true error at each time point is calculated. Finally, using historical information sequences as training input and the true error E as training labels, the SA-LSTM network is trained under supervised learning until the model converges.
[0072] This application employs an Adaptive Unscented Kalman Filter (A-UKF) as the base tracker. The A-UKF dynamically and adaptively updates the noise covariance matrix Q by real-time monitoring of filter information—the statistical characteristics of the difference between the actual and predicted measurements. When the UAV performs violent maneuvers, causing an increase in error, the A-UKF automatically increases the Q value to improve the response speed to the current measurement, thereby quickly capturing the dynamic changes of the target and generating a highly robust initial track reference. When the UAV returns to a stable flight phase, the A-UKF automatically decreases the Q value.
[0073] The initial state estimation sequence output by A-UKF is fed into a Long Short-Term Memory (LSTM) network with a self-attention mechanism for refined error compensation. The core innovation of this network lies in the application of the attention mechanism. Before predicting the filtering error at the current moment, this mechanism analyzes all historical points in the input trajectory sequence and calculates an "attention weight" for each point. Key information points such as maneuvering inflection points are assigned high weights, while redundant information points in the stable flight phase are assigned low weights. Subsequently, the network selectively focuses on and fuses historical information based on these weights, forming a context vector that highly condenses key dynamic features, and makes the final error prediction based on this vector.
[0074] Finally, this method integrates the preliminary state estimate output by A-UKF with the precise error compensation value output by SA-LSTM to obtain the final ultra-high precision track after dual optimization.
[0075] This application addresses the performance degradation issue of traditional filters in dynamically changing scenarios by using A-UKF to perceive and respond to the maneuvering intentions and model mismatch of UAV targets in real time, thus ensuring the robustness of the base trajectory. Simultaneously, by incorporating an attention mechanism, the error compensation model intelligently focuses on the most valuable historical information for the current prediction, avoiding interference from redundant or secondary information, making error compensation more targeted. The precise compensation of SA-LSTM also provides more accurate prior states for the next iteration of A-UKF. This mutually reinforcing deep collaborative framework improves overall tracking performance and significantly enhances tracking accuracy at key dynamic nodes (such as the start / end of maneuvers).
[0076] Model mismatch refers to the fact that the motion equations preset by the Kalman filter cannot keep up with the complex and ever-changing actual flight trajectory of the UAV.
[0077] like Figure 2As shown in the figure, this application embodiment also provides a UAV trajectory tracking device based on neural networks and adaptive Kalman filtering. The device includes: a data acquisition module 201, a processing module 202, and an output module 203. The data acquisition module 201 is used to acquire measurement data.
[0078] For example, a radar or multi-sensor fusion system is responsible for detecting target UAVs in real time and acquiring their measurement data, such as slant range, azimuth, and pitch angle. The processing module 202 includes: an A-UKF filtering unit 2021, an SA-LSTM compensation unit 2022, and a state correction unit 2023.
[0079] For specific implementation details, please refer to the corresponding description in the methods section; repeated details will not be repeated here.
[0080] Output module 203 is used to output the target state estimation vector.
[0081] For example, a graphical user interface (GUI) is used to visualize or output the final trajectory information (obtained based on the target state estimation vector) calculated by the processing module 202 to other systems.
[0082] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 301, a communications interface 302, a memory 303, and a communication bus 304. The processor 301, communications interface 302, and memory 303 communicate with each other via the communication bus 304. The processor 301 can call logical instructions from the memory 303 to execute a UAV trajectory tracking method based on neural networks and adaptive Kalman filtering.
[0083] Furthermore, the logical instructions in the aforementioned memory 303 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0084] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, and when the program instructions are executed by a computer, the computer is able to execute the UAV trajectory tracking method based on neural networks and adaptive Kalman filtering provided by the above methods.
[0085] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the UAV trajectory tracking method based on neural networks and adaptive Kalman filtering provided in the above embodiments.
[0086] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0088] Finally, it should be noted that the above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.
Claims
1. A UAV track tracking method based on neural network and adaptive Kalman filter, characterized in that, The method comprises: obtaining measurement data of the unmanned aerial vehicle at time k when the monitoring device is monitoring the unmanned aerial vehicle; processing the measurement data based on the process noise covariance corresponding to time k-1, the measurement noise covariance corresponding to time k-1 and a preset adaptive Kalman filtering algorithm to obtain a preliminary state estimation vector; obtaining a historical innovation sequence corresponding to a preset time window, and inputting the historical innovation sequence into a preset long short-term memory neural network to obtain a prediction error output by the long short-term memory neural network, wherein the innovation is obtained based on the measurement data, and the prediction error reflects an estimation error of the adaptive Kalman filtering algorithm in state estimation of the measurement data; correcting the preliminary state estimation vector by using the prediction error to obtain a target state estimation vector. 2.The UAV track pursuit method based on neural network and adaptive Kalman filter according to claim 1, characterized in that, The method comprises: generating a plurality of Sigma points based on the dimension of the measurement data, and propagating each Sigma point by using a preset nonlinear motion model to obtain a predicted state of the unmanned aerial vehicle, and obtaining a prediction parameter based on the predicted state and / or the process noise covariance corresponding to time k-1; mapping the propagated Sigma points to a measurement space by using a preset measurement function to obtain a mapped state, and obtaining a mapping parameter based on the mapped state and / or the measurement noise covariance corresponding to time k-1; calculating a cross-covariance matrix corresponding to the predicted state, the mapped state, the prediction parameter and the mapping parameter, and calculating a Kalman gain based on the cross-covariance matrix; obtaining the preliminary state estimation vector based on the Kalman gain, the measurement data and the mapping parameter. 3.The UAV track pursuit method based on neural network and adaptive Kalman filter according to claim 2, characterized in that, The mapping parameter comprises a first mean value corresponding to the mapped state. The method comprises: inputting the Kalman gain, the measurement data and the mapping parameter into a state estimation formula to obtain the preliminary state estimation vector output by the state estimation formula; The state estimation formula comprises: ； wherein, denotes a preliminary state estimation vector corresponding to time k, denotes a preliminary state estimation vector corresponding to time k-1, denotes a Kalman gain, denotes measurement data, denotes a first mean value corresponding to a mapping state. 4.The UAV track pursuit method based on neural network and adaptive Kalman filter according to claim 2, characterized in that, The prediction parameter comprises a second mean value and a first covariance corresponding to the predicted state. The method comprises: calculating the second mean value corresponding to the predicted state based on a first mean value calculation formula, and calculating the first covariance corresponding to the predicted state based on a first covariance calculation formula; The first mean value calculation formula comprises: ； wherein, denotes the second mean value, denotes the dimension of the measurement data, denotes the mean weight corresponding to the Sigma point, denotes the predicted state; The covariance calculation formula comprises: ； wherein, represents the second covariance, represents the covariance weight corresponding to the Sigma point, represents the process noise covariance corresponding to the k-1 time.
5. The neural network and adaptive Kalman filter based UAV track-while-scan method of claim 2, wherein, The mapping parameter comprises a first mean value and an innovation covariance corresponding to the mapped state. The method comprises: calculating the first mean value corresponding to the mapped state based on a second mean value calculation formula, and calculating the innovation covariance corresponding to the mapped state based on a second covariance calculation formula; The second mean value calculation formula comprises: ； wherein, denotes the first mean value, denotes the dimension of the measurement data, denotes the mean weight corresponding to the Sigma point, denotes the mapped state; The second covariance calculation formula comprises: ； wherein, denotes the innovation covariance, denotes the covariance weight corresponding to the Sigma point, denotes the measurement noise covariance corresponding to the k-1 time.
6. The neural network and adaptive Kalman filter based UAV track-while-scan method of claim 2, wherein, The prediction parameter comprises a second mean value corresponding to the prediction state, and the mapping parameter comprises a first mean value corresponding to the mapping state; The method further comprises: calculating a cross-covariance matrix corresponding to the prediction state, the mapping state, the prediction parameter and the mapping parameter, and calculating a Kalman gain based on the cross-covariance matrix, comprising: inputting the prediction state and the mapping state into a third covariance calculation formula to obtain a cross-covariance matrix output by the third covariance calculation formula; The third covariance calculation formula comprises: ； wherein, represents a cross-covariance matrix, represents a dimension of the measurement data, represents a covariance weight corresponding to the Sigma point, represents a predicted state, represents a second mean value corresponding to the predicted state, represents a mapped state, represents a first mean value corresponding to the mapped state; inputting the cross-covariance matrix into a Kalman gain calculation formula to obtain a Kalman gain output by the Kalman gain calculation formula; The Kalman gain calculation formula comprises: ； wherein, denotes the Kalman gain, denotes the innovation covariance corresponding to the mapped state.
7. The neural network and adaptive Kalman filter based UAV track-while-scan method of claim 2, wherein, After obtaining the preliminary state estimation vector, the method further comprises: updating a posterior covariance based on a covariance update formula, wherein the posterior covariance is used to generate a Sigma point; The covariance update formula comprises: ； wherein, denotes the posterior covariance corresponding to time k, denotes the posterior covariance corresponding to time k-1, denotes the Kalman gain, denotes the innovation covariance corresponding to the mapped state.
8. The neural network and adaptive Kalman filter based UAV track-while-scan method of claim 2, wherein, After obtaining the preliminary state estimation vector, the method further comprises: obtaining a new input corresponding to the k time based on a new input calculation formula; The new input calculation formula comprises: ； wherein, denotes the innovation at time k, denotes the measurement data, denotes the first mean value corresponding to the mapping state.
9. The neural network and adaptive Kalman filter based UAV track-while-scan method of claim 8, wherein, After obtaining the new input corresponding to the k time based on the new input calculation formula, the method further comprises: evaluating a measurement noise covariance corresponding to the k time based on a measurement noise covariance evaluation formula, and evaluating a process noise covariance corresponding to the k time based on a process noise covariance evaluation formula; The measurement noise covariance evaluation formula comprises: ； wherein, denotes the measurement noise covariance at time k, denotes a weight factor, denotes the measurement noise covariance at time k-1, denotes the dimension of the measurement data, denotes the covariance weight of the Sigma point, denotes the mapped state, denotes the first mean value of the mapped state; The process noise covariance evaluation formula comprises: ； wherein, denotes the process noise covariance corresponding to time k, denotes the process noise covariance corresponding to time k-1, denotes the Kalman gain.
10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the method for unmanned aerial vehicle track tracking based on a neural network and an adaptive Kalman filter according to any one of claims 1 to 9.