A multi-feature LSTM-based unmanned aerial vehicle trajectory prediction method

Abstract

The application relates to a multi-feature LSTM-based unmanned aerial vehicle trajectory prediction method, and belongs to the technical field of data processing, which comprises the following steps: step 1: a six-degree-of-freedom global motion state matrix of a target unmanned aerial vehicle is constructed; step 2: a three-dimensional wind field vector acting on the target unmanned aerial vehicle is obtained through joint estimation, and a real-time speed direction of the target unmanned aerial vehicle is extracted; step 3: a historical state sequence, the three-dimensional wind field vector and a speed direction unit vector are fused to construct a multi-dimensional time sequence feature vector; step 4: an LSTM neural network learns a nonlinear maneuvering mode of the target unmanned aerial vehicle under the influence of the wind field through the gating mechanism of a forgetting gate, an input gate and an output gate; and step 5: the LSTM neural network outputs a position probability distribution of the target unmanned aerial vehicle at multiple future time points. The application has the beneficial effect that a space-time cone body representing a future trajectory is formed, and the trajectory prediction of the target unmanned aerial vehicle is completed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data processing, and particularly relates to a UAV trajectory prediction method based on multi-feature LSTM, which is suitable for probabilistic prediction of future movement trajectory of high-speed maneuvering invading UAV in an autonomous interception system. BACKGROUND

[0002] Unmanned aerial vehicle flight trajectory intelligent prediction technology, as a key link to improve the autonomy and intelligence level of unmanned aerial vehicles, plays an irreplaceable role in advanced applications such as target tracking, cluster cooperation and autonomous interception. The core goal is to accurately infer the future flight path of the target by analyzing its historical and current state, thereby providing forward-looking information input for decision and control modules. Although this technology field has undergone development for many years, in dealing with high-speed, high-maneuverability targets, especially in complex and variable real environments, existing technical solutions still have a series of essential defects and challenges to be solved.

[0003] Traditional mainstream trajectory prediction methods largely rely on linear dynamics models based on physical laws, such as constant speed models or uniform acceleration models, and usually combine Kalman filtering algorithms for state extrapolation. This kind of method has a certain effect in the scene where the target motion mode is simple and can be approximately described by a linear equation. However, when the unmanned aerial vehicle performs reconnaissance, escape or confrontation tasks, its motion mode exhibits significant nonlinear, time-varying and strong coupling characteristics, such as emergency turning, spiral climbing, irregular snake-shaped maneuvering, etc. In the face of such complex maneuvers, the assumption basis of the linear model is destroyed, leading to a rapid deviation of the predicted trajectory from the real path, resulting in unacceptable errors. More critically, such methods usually regard the unmanned aerial vehicle as an independent dynamic system, ignoring the profound influence of the environmental wind field, a key external physical factor. The wind field will have a direct and significant effect on the airspeed, ground speed and heading of the unmanned aerial vehicle. A prediction model that does not consider wind field disturbances is no different from simulating in a vacuum, and its prediction results are often quite different from the real situation in actual applications.

[0004] With the development of artificial intelligence technology, more and more research attempts to improve the prediction accuracy through data-driven methods, especially machine learning-based methods. However, the input features of many advanced models are still too single, often focusing only on the historical position coordinate sequence of the target itself. This design idea has obvious limitations: on the one hand, it fails to effectively utilize the instantaneous motion intention of the target as an explicit feature. The direction of the velocity vector is the most direct and sensitive indicator to judge the next action of the target. Indirectly obtaining speed information from the position sequence through difference not only introduces noise, but also leads to a sluggish response of the model to changes in target intention. On the other hand, such methods lack an effective mechanism for fusing multi-source heterogeneous information (such as environmental wind field, target velocity vector direction, etc.). Perception data in low-altitude complex environments often face bottlenecks such as "unclear", "inaccurate", and "incomplete". A single position sequence cannot fully describe the target and its environment, making it difficult to learn the deep rules that drive the target's motion.

[0005] In addition, practical application scenarios impose strict requirements on the computational complexity and real-time performance of algorithms. Many complex models with high theoretical accuracy are difficult to meet the frame rate requirements (e.g., ≥25Hz) for real-time operation on the limited on-board computing resources of unmanned aerial vehicles. This makes many advanced algorithms only remain in the simulation or post-processing stage, and cannot be effectively deployed in tasks that require online and fast response.

[0006] In summary, current unmanned aerial vehicle flight trajectory prediction technology faces challenges such as linear models failing to adapt to non-linear maneuvers, single data sources limiting model cognition, and the contradiction between complex models and limited computing power. Therefore, there is an urgent need in the field for an intelligent trajectory prediction method that can deeply fuse multi-source information, accurately model non-linear dynamic characteristics, and meet real-time requirements. SUMMARY

[0007] The present application provides a kind of unmanned aerial vehicle trajectory prediction method based on multi-feature LSTM, to solve the technical problem of low unmanned aerial vehicle trajectory prediction accuracy, by fusing the historical trajectory of unmanned aerial vehicle target, velocity direction vector and real-time environmental wind field parameters, train LSTM neural network deep learning unmanned aerial vehicle maneuver mode, realize the high-precision probabilistic prediction of unmanned aerial vehicle target future 3-5 seconds motion trajectory, provide key and reliable input for autonomous interception decision.

[0008] To achieve the above purpose, the present application is realized by the following technical scheme:

[0009] A kind of unmanned aerial vehicle trajectory prediction method based on multi-feature LSTM, comprising the following steps:

[0010] Step 1: continuously acquire the position of the target UAV in the image through the onboard sensors of the interceptor, and combine the GPS / IMU data of the interceptor itself to perform coordinate transformation, to construct a six-degree-of-freedom global motion state matrix of the target UAV; wherein the six-degree-of-freedom global motion state matrix contains a historical state sequence of the target UAV in the past k time steps, and each historical state sequence includes the position and velocity of the target UAV in the global coordinate system;

[0011] Step 2: based on the difference between the IMU data, airspeed meter data and GPS ground speed data of the interceptor, jointly estimate a three-dimensional wind field vector acting on the target UAV, and then extract the real-time velocity direction of the target UAV to generate a velocity direction unit vector;

[0012] Step 3: fuse the historical state sequence, three-dimensional wind field vector and velocity direction unit vector to construct a multi-dimensional time series feature vector; wherein the feature vector corresponding to each time step t contains a historical state sub-vector, a three-dimensional wind field sub-vector and a velocity direction sub-vector of each time step;

[0013] Step 4: input the multi-dimensional time series feature vector into the pre-trained LSTM neural network, and the LSTM neural network learns the nonlinear maneuvering pattern of the target UAV under the influence of the wind field through the gating mechanism of the forget gate, input gate and output gate;

[0014] Step 5: output the position probability distribution of the target UAV at multiple time points in the future 3-5 seconds through the LSTM neural network, form a spatio-temporal cone representing the future trajectory, and complete the trajectory prediction of the target UAV.

[0015] Optionally, in step 1, the imaging rate of the onboard sensors of the interceptor is 25 frames / s, so 2 seconds correspond to 50 frames of data, and the historical state sequence is a time series sequence composed of position and velocity data of the past 50 frames.

[0016] The coordinate transformation converts the image pixel coordinates obtained by the onboard sensors into three-dimensional position coordinates in the global coordinate system, and calculates the real-time velocity of the target UAV in combination with the position difference between adjacent frames.

[0017] Optionally, in step 2, the estimation process of the three-dimensional wind field vector is as follows: first, obtain the acceleration and angular velocity data of the interceptor through the IMU, combine the airspeed measured by the airspeed meter, and calculate the theoretical ground speed of the interceptor; then, subtract the actual ground speed measured by the GPS from the theoretical ground speed to obtain the influence vector of the wind field on the interceptor; finally, based on the relative position relationship between the interceptor and the target UAV, the influence vector is corrected to the three-dimensional wind field vector acting on the target UAV.

[0018] Optionally, in step 3, the dimension of the multi-dimensional time sequence feature vector is (k, d), wherein k is the number of time steps, d is the feature dimension of each time step, and each feature dimension is normalized.

[0019] Optionally, in step 4, the LSTM neural network adopts the forgetting gate calculation, the input gate calculation, the cell state update, and the output gate calculation to process data.

[0020] Optionally, in step 4, the training process of the LSTM neural network is as follows: supervised learning is performed by using a data set containing multiple nonlinear maneuver modes of the target UAV; wherein the data set covers trajectory data of evasion maneuvers, escape maneuvers, hovering maneuvers, and diving maneuvers, and each trajectory data contains a corresponding historical state sequence, a three-dimensional wind field vector, a velocity direction unit vector, and a real position label in the future 3-5 seconds.

[0021] Optionally, in step 5, the output form of the position probability distribution is that, for each time step in the future, the LSTM neural network outputs the Gaussian distribution parameters of the time step position, including the mean and variance of each position component, and the covariance between each position component.

[0022] Optionally, the Gaussian distribution parameters of the target position at a series of future time points are output, a space-time cone is formed by the confidence interval of each time step Gaussian distribution, and as time goes by, the confidence interval range gradually expands to represent the increase of prediction uncertainty.

[0023] Optionally, in step 5, the output layer of the LSTM neural network adopts multiple parallel fully connected layers, and each fully connected layer corresponds to one predicted trajectory.

[0024] N future trajectories are output through the parallel fully connected layers, and each trajectory contains the position coordinates of all time steps in the future 3-5 seconds; the selection of the N future trajectories is based on the position probability distribution output by the LSTM neural network, and the position sequence with the top N probability values is sampled to form a trajectory set.

[0025] The beneficial effects of the present application are as follows:

[0026] 1.The present application is to build a system with long short-term memory neural network (LSTM) as the prediction engine, through the deep fusion of the historical state of the target UAV, real-time environmental wind field parameters and its speed direction vector, to realize the high-precision probabilistic prediction of the target future 3 to 5 seconds motion trajectory. Specifically, the system first uses the airborne sensor to continuously obtain the position of the target in the image, and combines the GPS / IMU data of the interceptor itself, through coordinate transformation, to build a stable and accurate six-degree-of-freedom global motion state matrix of the target. Then, the historical state sequence containing position and speed, the three-dimensional wind field vector estimated by the interceptor state and the real-time speed direction unit vector of the target are jointly constructed into a multi-dimensional time sequence feature vector, which is input into the pre-trained LSTM network. The LSTM network can effectively learn and remember the complex nonlinear maneuvering mode of the target under the influence of the wind field by virtue of its internal complex gating mechanism, and obtain the probability distribution of the target position at a series of future time points, thereby forming a space-time cone representing the possible future trajectory.

[0027] 2.The present application has high prediction accuracy. By introducing environmental wind field parameters and speed direction vectors, the LSTM neural network is used to predict the target trajectory nonlinearly, which can effectively predict the target escape maneuver. Compared with the linear extrapolation model, it can greatly improve the interception success rate in complex confrontation scenarios. The LSTM neural network can autonomously learn various maneuvering modes from data, without pre-setting specific motion models, and has certain generalization ability for unknown new maneuvering modes. It has high practical value and can be directly integrated into the existing autonomous interception system framework as a prediction module to provide key information input and effectively improve the interception success rate. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below to the drawings needed to be used in the embodiments or prior art descriptions. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0029] Figure 1 The present application is a work flow diagram. DETAILED DESCRIPTION

[0030] The embodiments of the present application will be described in detail below with reference to the drawings.

[0031] Embodiment 1;

[0032] As Figure 1 shown, the present embodiment provides a UAV trajectory prediction method based on multi-feature LSTM, which includes the following steps:

[0033] Step 1: Continuously acquire the position of the target UAV in the image through the airborne sensors of the interceptor, and combine it with the interceptor's own GPS / IMU data to perform coordinate transformation and construct a six-degree-of-freedom global motion state matrix of the target UAV; wherein, the six-degree-of-freedom global motion state matrix contains the historical state sequence of the target UAV over the past k time steps, and each historical state sequence includes the position and velocity of the target UAV in the global coordinate system;

[0034] Step 2: Based on the difference between the interceptor's IMU data, airspeed data, and GPS ground speed data, the three-dimensional wind field vector acting on the target UAV is jointly estimated, and then the real-time velocity direction of the target UAV is extracted to generate a unit velocity direction vector.

[0035] Step 3: Fuse the historical state sequence, the three-dimensional wind field vector, and the velocity direction unit vector to construct a multi-dimensional temporal feature vector; wherein, the feature vector corresponding to each time step t contains the historical state sub-vector, the three-dimensional wind field sub-vector, and the velocity direction sub-vector for each time step;

[0036] Step 4: Input the multidimensional temporal feature vector into the pre-trained LSTM neural network. The LSTM neural network learns the nonlinear maneuvering mode of the target UAV under the influence of wind field through the gating mechanism of forget gate, input gate and output gate.

[0037] Step 5: Output the probability distribution of the target UAV's position at multiple time points within the next 3-5 seconds through the LSTM neural network, forming a spatiotemporal cone representing the future trajectory, thus completing the trajectory prediction of the target UAV.

[0038] Specifically, in step 1, the airborne sensor imaging rate of the interceptor is 25 frames / s, so 2 seconds corresponds to 50 frames of data, and the historical state sequence is a time sequence composed of the position and velocity data of the past 50 frames; that is, the airborne sensor imaging rate of the interceptor is 25Hz, the past k time steps correspond to a duration of 2 seconds, k=50 frames, and the historical state sequence is a time sequence composed of the position and velocity data of the past 50 frames.

[0039] The coordinate transformation uses the interceptor's GPS / IMU data as a reference to convert the image pixel coordinates obtained by the airborne sensors into three-dimensional position coordinates in the global coordinate system. At the same time, it combines the position difference between adjacent frames to calculate the real-time speed of the target UAV.

[0040] In step 2, the estimation process of the three-dimensional wind field vector is as follows: first, the acceleration and angular velocity data of the interceptor are obtained through the IMU of the interceptor, and the theoretical ground speed of the interceptor is calculated in combination with the airspeed measured by the airspeed meter; then, the actual ground speed measured by the GPS is subtracted from the theoretical ground speed to obtain the influence vector of the wind field on the interceptor; finally, based on the relative position relationship between the interceptor and the target UAV, the influence vector is corrected to the three-dimensional wind field vector acting on the target UAV.

[0041] In step 3, the dimension of the multi-dimensional time sequence feature vector is (k, d); wherein k is the number of time steps, and d is the feature dimension of each time step, and each feature dimension is normalized, and d is 9, which specifically includes: 3 position components (longitude, latitude and height) of the historical state sub-vector, 3 velocity components (x-axis velocity, y-axis velocity and z-axis velocity), 3 components (x-axis wind speed, y-axis wind speed and z-axis wind speed) of the three-dimensional wind field sub-vector, and 3 components (x-axis direction, y-axis direction and z-axis direction) of the velocity direction unit vector, and each feature dimension component is normalized to the [0, 1] interval.

[0042] In step 4, the LSTM neural network adopts forgetting gate calculation, input gate calculation, cell state update and output gate calculation for data processing.

[0043] In addition, in step 4, the training process of the LSTM neural network is as follows: supervised learning is performed using a data set containing multiple nonlinear maneuver modes of the target UAV; wherein the data set covers trajectory data of evasion maneuvers, escape maneuvers, hovering maneuvers and diving maneuvers, and each trajectory data includes a corresponding historical state sequence, a three-dimensional wind field vector, a velocity direction unit vector and a real position label for the next 3-5 seconds.

[0044] In step 5, the output form of the position probability distribution is that for each future time step, the LSTM neural network outputs the Gaussian distribution parameters of the time step position, including the mean and variance of each position component, and the covariance between each position component. The Gaussian distribution parameters of the target position at a series of future time points are output, and the space-time cone is composed of the confidence intervals of each Gaussian distribution, and as time goes on, the confidence interval range gradually expands to represent the increase of prediction uncertainty.

[0045] In addition, in step 5, the output layer of the LSTM neural network adopts multiple parallel fully connected layers, and each fully connected layer corresponds to one predicted trajectory.

[0046] According to the historical trajectory of the UAV target, the velocity vector, the environmental wind field parameter perceived by the FPV interceptor is introduced, and then based on the LSTM neural network, a dynamic trajectory prediction model of the UAV target is established.

[0047] The N future trajectories are output by a parallel fully connected layer, each trajectory containing position coordinates of all time steps within the future 3-5 seconds; the selection of the N future trajectories is based on the position probability distribution output by the LSTM neural network, and the position sequence with the top N probability values is sampled to form a trajectory set.

[0048] Embodiment 2;

[0049] Based on Embodiment 1, the present application trains the LSTM neural network to learn the maneuvering mode of the UAV target by fusing the historical trajectory of the UAV target, the velocity direction vector, and the real-time environmental wind field parameters, realizes high-precision probabilistic prediction of the future 3-5 second motion trajectory of the UAV target, and provides key and reliable input for autonomous interception decision.

[0050] Traditional interception systems usually use a simple linear dynamic model (such as: uniform motion or uniform acceleration motion model) for extrapolation prediction. This method is still applicable for targets performing regular flight, but when facing targets that take high-intensity, nonlinear escape maneuvers after realizing being tracked (such as sudden turning, spiral climbing, or diving), the prediction error will grow exponentially, leading to failure of the interception path.

[0051] Therefore, the present application discards the assumption of the traditional linear model and introduces a dynamic trajectory prediction model based on a long short-term memory (LSTM) neural network. This model can autonomously learn the complex maneuvering mode of the target from its historical motion data and generate a probabilistic trajectory distribution at the future time.

[0052] The input of the LSTM dynamic prediction model is a multi-dimensional, time-dependent feature sequence, which is designed to comprehensively capture internal and external factors affecting the target motion.

[0053] The core of the input sequence is the historical state information of the target at the past k time steps (for example: the past 2 seconds, if the imaging rate of the camera on the interceptor is 25 frames / s, then 2 seconds correspond to 50 frames of data), that is, Each historical state vector or historical state information contains the position and velocity of the target in the global coordinate system, wherein, represents the position coordinates of the target in the x, y, and z directions of the global coordinate system (such as the earth rectangular coordinate system, the inertial coordinate system), which describes the spatial position of the target; represents the velocity components of the target in the x, y, and z directions, which describes the motion rate of the target in the corresponding coordinate axis direction; represents the matrix transpose, which converts the original row vector to a column vector, facilitating matrix operations in filtering and state estimation algorithms, and comprehensively represents the "position + velocity" motion state of the target.

[0054] In addition, the present application also takes environmental wind field parameters As an important feature input model, the wind field is an external factor that has a significant impact on the movement of the UAV, the value of which can be estimated by the IMU (Inertial Measurement Unit) data of the interceptor itself, the difference between the airspeed meter and the GPS ground speed, and is represented as wherein, is the wind speed component in the x direction and at time step t; is the wind speed component in the y direction and at time step t; is the wind speed component in the z direction and at time step t. The target current speed vector direction is also added as a key indicator embodying its instantaneous movement intention to the input features. Therefore, at each time step t, the input feature vector of the model is a fusion of state, environment and intention information: .

[0055] Embodiment 3;

[0056] Based on Embodiments 1-2, the long short-term memory (LSTM) neural network: is a special recurrent neural network (RNN), the disadvantage of the RNN algorithm is that with the passage of time and the increase of network layers, problems such as gradient disappearance or gradient explosion will occur, and the LSTM neural network is used to solve this fatal problem of traditional RNN. The LSTM neural network includes a forgetting gate, an input gate and an output gate, which are composed of a sigmoid neural network layer and a dot multiplication operation and can selectively allow information to pass. Its working process is: first, the forgetting gate decides what old information to discard from the long-term state; then, the input gate decides what new information to store in the long-term state; then, the long-term state is updated; finally, the output gate decides the final output (i.e. the next short-term state) according to the current long-term state, the current input and the previous short-term state.

[0057] The LSTM network is very suitable for processing such time series prediction problems due to its gating mechanism, and its core lies in three gates: the forgetting gate , the input gate , the output gate , the cell state and the hidden state , and the calculation of these gates is as follows:

[0058] ;

[0059] wherein, is the output of the forgetting gate, which is a vector with the same dimension as the cell state , and the element value is between (0, 1) (determined by the activation function determined by the sigmoid activation function) for controlling how much information is forgotten in the cell state;

[0060] sigmoid activation function, which maps the input to the interval (0, 1) to realize gating (i.e., on-off control of information, 0 means complete forgetting, and 1 means complete retention);

[0061] is the weight matrix of the forget gate, which is used for linear transformation of the input and needs to be learned during the training process; is the hidden state of the previous time step, which contains the information processed by the LSTM neural network at the previous time step and is the short-term memory carrier of the network; the input at the current time step; is the original data at the current time step; is the bias term of the forget gate, which is used to adjust the result of linear transformation and is also learned during training.

[0062] The forget gate is a key component of the LSTM neural network that realizes selective memory by dynamically determining the retention and forgetting of information in the cell state, thereby effectively capturing long-distance dependencies.

[0063] ;

[0064] wherein, is the output of the input gate, which is a vector with the same dimension as the hidden layer, and the element value is between 0 and 1 (determined by the sigmoid activation function ), indicating how much new information is allowed to enter the cell state; is the weight matrix of the input gate, which is used for linear transformation of the input; is the concatenation operation, which concatenates the hidden state of the previous time step (containing historical information) and the input at the current time step (containing current information) into a vector as the joint input of the input gate; is the bias term of the input gate, which is used to adjust the result of linear transformation.

[0065] The output of the input gate is multiplied element-wise with the candidate cell state (obtained by tanh activation of the current input and historical hidden state), and finally determines how much new information is added to the cell state.

[0066] ;

[0067] wherein, is the candidate cell state, which is an intermediate variable used to update the cell state in the LSTM neural network; is the hyperbolic tangent activation function, which maps the input to the interval (−1, 1) and introduces nonlinearity for candidate cell state, and controls the range of candidate state and the amplitude of cell state update; is the weight matrix, which is used to linearly transform the concatenated vector of input, and contains the weight parameters for the hidden state at the previous time step and the input at the current time step; is the bias vector, which is used to adjust the result of linear transformation and introduce more expressive power for the model.

[0068] After concatenating the historical information and the current input , the linear transformation is performed by the weight matrix and the bias , and then the tanh activation is performed to obtain the candidate cell state . The subsequent combination with the output of the forget gate and the input gate completes the update of the LSTM cell state .

[0069] ;

[0070] wherein is the LSTM cell state; represents element-wise multiplication, which embodies the dynamic management of information, i.e., forgetting irrelevant information and remembering important new information.

[0071] ;

[0072] wherein is the output of the output gate, which takes a value between 0 and 1 and controls how much information in the cell state is transmitted to the hidden state at the current time step; is the weight matrix of the output gate, which is used to linearly transform the concatenated vector of input, and contains the weight parameters for the hidden state at the previous time step and the input at the current time step; is the bias vector of the output gate, which is used to adjust the result of linear transformation and increase the expressive power of the model.

[0073] The function of the output gate is to generate a gating signal by combining the historical hidden state and the current input through the Sigmoid activation function, so as to control how much information in the cell state is output to the hidden state at the current time step.

[0074] ;

[0075] where, is the hidden state at time step t, is the information summary outputted by LSTM at current time, which will also be passed to next time as historical information, i.e. will become part of the input of at next time; denotes element-wise multiplication.

[0076] is the cell state at time t or time step t is passed through a tanh function to map the value of cell state to the interval (−1, 1), introducing nonlinearity for the generation of hidden state and normalizing the output range of cell state.

[0077] The cell state is nonlinearly transformed and range-constrained by the tanh function, and then filtered by the output gate to let the information allowed by the output gate pass to the hidden state . This ensures that the hidden state can carry the core information of the cell state, and also enables fine-grained control of information output by the output gate, allowing LSTM to flexibly decide how much memory information to expose externally at the current time.

[0078] and are the parameter matrices and vectors to be trained, the forget gate decides which information from the previous cell state to discard, the input gate decides which new information to store in the cell state, and the output gate decides the current hidden state output based on the cell state. Through this mechanism, LSTM neural network can selectively remember and forget long-term and short-term dependencies, effectively capturing the complex transition patterns of target UAV from level flight to evasion.

[0079] LSTM neural network is trained to predict the target state at the next T time steps (corresponding to 3 to 5 seconds in the future) based on the input historical sequence by supervised learning on a large amount of trajectory data containing various UAV maneuvers (including evasion, escape, hovering, and diving, etc.). The output layer of the network is usually a fully connected layer, and its output is not a single, deterministic trajectory, but a probability distribution of future trajectories.

[0080] There are two ways to implement this embodiment:

[0081] 1) Define the output as the Gaussian distribution parameters of the target position at a series of future time points, i.e., for each future time step , the network outputs a mean and a covariance matrix The covariance matrix represents the uncertainty of the prediction. This means that the output of the model is a probabilistic "spacetime cone" that diverges over time, where, is the mean, which is the central tendency indicator of the probability distribution, in time series prediction, it represents the model's prediction of the most likely value of the system state at a certain future time; is the future time step, represents the current time to the future time interval; is the component mean in three-dimensional space, which corresponds to the mean prediction in x, y and z directions of the three-dimensional coordinate system (such as Cartesian coordinate system) respectively.

[0082] 2) Use multiple parallel fully connected heads to predict N most likely future trajectories, forming a trajectory set Each trajectory is a sequence of future state points. Among them, represents the trajectory of the future, which is a vector or matrix variable, which contains the position, velocity and other motion state sequence of the target in the future, and the superscript represents the first, second, …, Nth trajectory, which embodies the characteristics of multi-trajectory prediction. represents that N trajectories form a set, that is, the whole of all prediction results.

[0083] This probabilistic prediction method based on LSTM neural network has a fundamental advantage over traditional linear models. It does not rely on any prior kinematic model assumption, but learns the motion pattern directly from the data, so it can accurately predict the "unpredictable" dramatic maneuver of the target UAV. The output probability distribution provides a crucial uncertainty measure for the subsequent interception decision module, and the decision algorithm can plan a path that successfully intercepts under the maximum probability, or develop a robust strategy for multiple possible paths, thereby greatly improving the intelligence level and success probability of the entire interception system in a real combat environment.

[0084] The primary purpose of the present application is to overcome the defects and limitations of the prior art, and to provide a high-precision and reliable UAV trajectory prediction method. Specifically, the present application discards the traditional extrapolation model based on unrealistic linear assumptions, and adopts an artificial intelligence model that can learn complex nonlinear maneuver patterns from data autonomously, namely long short-term memory (LSTM) neural network, which fundamentally improves the prediction ability of sudden and dramatic maneuver behavior of the target.

[0085] Another object of the present application is to break through the limitation of single position sequence input and realize deep fusion of multi-source heterogeneous features. The present application designs a new input feature construction method, which synchronously takes the environmental wind field parameters, target speed vector direction and historical trajectory data as model inputs. In this way, the prediction model can simultaneously perceive the historical motion law of the target, the current physical environment constraint and the instantaneous motion intention, so as to realize more comprehensive and more physically realistic modeling and inference on the future motion of the target.

[0086] The above merely describes the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1.A method for UAV trajectory prediction based on multi-feature LSTM, characterized in that, Comprise the following steps: Step 1: continuously acquire the position of the target UAV in the image through the onboard sensor of the interceptor, and combine the GPS / IMU data of the interceptor itself to perform coordinate transformation to construct a six-degree-of-freedom global motion state matrix of the target UAV; wherein the six-degree-of-freedom global motion state matrix contains a historical state sequence of the target UAV in the past k time steps, and each historical state sequence includes the position and velocity of the target UAV in the global coordinate system; Step 2: based on the difference between the IMU data, airspeed meter data and GPS ground speed data of the interceptor, a three-dimensional wind field vector acting on the target UAV is estimated, and a real-time velocity direction of the target UAV is extracted to generate a velocity direction unit vector; Step 3: fuse the historical state sequence, the three-dimensional wind field vector and the velocity direction unit vector to construct a multi-dimensional time sequence feature vector; wherein the feature vector corresponding to each time step t contains a historical state subvector, a three-dimensional wind field subvector and a velocity direction subvector of each time step; Step 4: input the multi-dimensional time sequence feature vector into the pre-trained LSTM neural network, and the LSTM neural network learns the nonlinear maneuvering mode of the target UAV under the influence of the wind field through the gating mechanism of the forget gate, input gate and output gate; Step 5: output the position probability distribution of the target UAV at multiple future time points through the LSTM neural network to form a spatio-temporal cone representing the future trajectory, and complete the trajectory prediction of the target UAV. 2.The multi-feature LSTM based trajectory prediction method for UAVs according to claim 1, wherein, In the step 1, the imaging rate of the onboard sensor of the interceptor is 25 frames / s, so 2 seconds correspond to 50 frames of data, and the historical state sequence is a time sequence composed of position and velocity data of the past 50 frames; The coordinate transformation converts the image pixel coordinates obtained by the onboard sensor into three-dimensional position coordinates in the global coordinate system, and calculates the real-time velocity of the target UAV in combination with the position difference of adjacent frames. 3.The multi-feature LSTM based trajectory prediction method for UAVs according to claim 1, wherein, In the step 2, the estimation process of the three-dimensional wind field vector is as follows: first, obtain the acceleration and angular velocity data of the interceptor through the IMU, calculate the theoretical ground speed of the interceptor in combination with the airspeed measured by the airspeed meter; then, the actual ground speed measured by the GPS is subtracted from the theoretical ground speed to obtain the influence vector of the wind field on the interceptor; finally, based on the relative position relationship between the interceptor and the target UAV, the influence vector is corrected to the three-dimensional wind field vector acting on the target UAV. 4.The multi-feature LSTM based trajectory prediction method for UAVs according to claim 1, wherein, In the step 3, the dimension of the multi-dimensional time sequence feature vector is (k, d); wherein k is the number of time steps, and d is the feature dimension of each time step, and each feature dimension is normalized. 5.The multi-feature LSTM based trajectory prediction method for UAVs according to claim 1, wherein, In the step 4, the LSTM neural network adopts forget gate calculation, input gate calculation, cell state update and output gate calculation for data processing. 6.The multi-feature LSTM based trajectory prediction method for UAVs according to claim 1, wherein, In the step 4, the training process of the LSTM neural network is to perform supervised learning using a dataset containing multiple nonlinear maneuver modes of the target UAV; wherein the dataset covers trajectory data of evasion maneuvers, escape maneuvers, hovering maneuvers, and diving maneuvers, and each trajectory data contains a corresponding historical state sequence, a three-dimensional wind field vector, a velocity direction unit vector, and a real position label for the next 3-5 seconds. 7.The multi-feature LSTM based trajectory prediction method for UAVs according to claim 1, wherein, In the step 5, the output form of the position probability distribution is that for each time step in the future, the LSTM neural network outputs Gaussian distribution parameters of the time step position, including the mean and variance of each position component, and the covariance between each position component. 8.The multi-feature LSTM based trajectory prediction method for UAVs according to claim 7, wherein, The Gaussian distribution parameters of the target position at a series of future time points are output, and the spatiotemporal cone is composed of the confidence intervals of each time step Gaussian distribution, and the confidence interval range gradually expands over time to represent the increase of prediction uncertainty. 9.The multi-feature LSTM based trajectory prediction method for UAVs according to claim 1, wherein, In the step 5, the output layer of the LSTM neural network uses multiple parallel fully connected layers, and each fully connected layer corresponds to a predicted trajectory. N future trajectories are output through the parallel fully connected layers, and each trajectory contains position coordinates at all time steps within the next 3-5 seconds; the selection of the N future trajectories is based on the position probability distribution output by the LSTM neural network, and the top N position sequences are sampled to form a trajectory set.

Images