A target tracking method for low-altitude strong maneuvering unmanned aerial vehicle

Abstract

The embodiment of the specification discloses a target tracking method of a low-altitude strong maneuvering unmanned aerial vehicle, relates to the technical field of target tracking, and comprises the following steps: acquiring a measurement point set and a state estimation value and a state covariance matrix of a previous moment, so as to obtain a current prediction index corresponding to each strong maneuvering model at a current moment through a plurality of preset strong maneuvering models; determining an elliptical correlation gate corresponding to each current prediction index, so as to determine a candidate measurement point set falling into the elliptical correlation gate according to the measurement point set; performing intramodel correlation and intermodel correlation through a preset hierarchical correlation strategy, and screening a global optimal measurement corresponding to a current moment and a matching strong maneuvering model; performing Kalman filter updating on the current prediction index of the matching strong maneuvering model according to the global optimal measurement corresponding to the current moment, so as to obtain an updated prediction index of the target at the current moment, and target tracking is realized.

Description

Technical Field

[0001] This specification relates to the field of target tracking technology, and in particular to a target tracking method for a low-altitude, highly maneuverable unmanned aerial vehicle. Background Technology

[0002] With the rapid development and widespread application of UAV technology, reliable tracking of low-altitude flying targets has become a key focus in the field of radar surveillance. The low-altitude environment is complex, characterized not only by strong ground clutter from buildings and terrain, but also by complex electromagnetic interference, resulting in numerous false alarms and interference points in radar measurements. Simultaneously, UAVs possess strong maneuverability, including hovering, rapid turns, and U-turns, exhibiting rapid switching and strong nonlinearity in their motion patterns. This specific scenario, coupled with the high maneuverability of the target in the low-altitude environment, poses a severe challenge to traditional tracking methods.

[0003] In existing technologies, interactive multi-model (IMM) or variable structure multi-model (VSMM) and their improved algorithms are the mainstream solutions for dealing with target maneuvers. These methods improve the tracking adaptability to maneuvering targets to some extent by running multiple motion models (such as uniform speed, uniform acceleration, and coordinated turning models) in parallel and performing fusion estimation based on model probabilities. However, when dealing with low-altitude, highly maneuvering UAVs, their pre-set model sets are mostly based on idealized kinematic assumptions, making it difficult to accurately describe the highly nonlinear maneuvering modes unique to UAVs, such as sudden stops and sharp-angle turns, leading to model mismatch and increased deviation between predicted trajectories and actual motion. More importantly, existing methods mainly focus on model set optimization and state estimation algorithm improvement, failing to fully consider the reliability of measurement association, a crucial preliminary step, in low-altitude, dense clutter environments. Traditional association methods (such as the global nearest neighbor method) are prone to misassociating high-intensity clutter points as target measurements in strong clutter backgrounds, or creating association ambiguities among multiple candidate points.

[0004] Therefore, in the process of tracking low-altitude high-maneuverability UAVs, the general motion model set used by the existing multi-model method is difficult to match the actual high-maneuverability mode of the target, and its association strategy is not reliable enough in the low-altitude dense clutter environment. There are problems such as large model prediction deviation and high measurement association error rate, which leads to discontinuous target tracking trajectory, or even target loss under complex interference. Summary of the Invention

[0005] This specification provides one or more embodiments of a target tracking method for low-altitude, highly maneuverable unmanned aerial vehicles (UAVs) to address the following technical problem: In the process of tracking low-altitude, highly maneuverable UAVs, the general motion model set used in existing multi-model methods is difficult to match the actual highly maneuverable mode of the target, and its association strategy is not reliable enough in low-altitude dense clutter environments, resulting in large model prediction deviations and high measurement association error rates, leading to discontinuous target tracking trajectories, or even target loss under complex interference.

[0006] One or more embodiments of this specification employ the following technical solutions:

[0007] This specification provides one or more embodiments of a target tracking method for a low-altitude, highly maneuverable unmanned aerial vehicle (UAV). The method includes: acquiring a set of measurement points output by a radar during the current detection period, as well as the target's state estimate and state covariance matrix at the previous moment; performing one-step prediction on the previous moment using multiple preset highly maneuverable models to obtain a current prediction index corresponding to each highly maneuverable model at the current moment, wherein the current prediction index includes a predicted state and a prediction covariance matrix; determining an elliptical correlation gate corresponding to each current prediction index to determine a set of candidate measurement points falling within the elliptical correlation gate based on the set of measurement points; performing intra-model and inter-model correlation within the candidate measurement point set using a preset hierarchical correlation strategy to filter the globally optimal measurement and the corresponding matching highly maneuverable model at the current moment; updating the current prediction index of the matching highly maneuverable model using Kalman filtering based on the globally optimal measurement at the current moment to obtain an updated prediction index of the target at the current moment, thereby achieving target tracking, wherein the updated prediction index includes an updated state estimate and an updated state covariance matrix.

[0008] The above-mentioned at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: Through the technical solution of the embodiments of this specification, multiple preset strong maneuvering models are introduced in parallel during the prediction stage. By using models of maneuvering modes such as hovering, turning around, and sharp turns of low-altitude UAVs, state inference is performed simultaneously along multiple high-probability maneuvering directions at each moment. This greatly reduces the systematic risk of the predicted point deviating seriously from the real target position due to model mismatch. No matter what kind of maneuver the target actually performs, there will always be one or more models whose predicted direction roughly matches it, thus preserving the possibility for subsequent data association and avoiding the fatal weakness of traditional single model or limited model set in that the prediction completely fails when the target suddenly maneuvers, directly leading to tracking loss. Secondly, by independently establishing an elliptical correlation gate strategy for each model, personalized management of prediction uncertainty and refined spatial search are achieved. The uncertainties of different maneuvering models exhibit anisotropic characteristics. Based on the dynamic characteristics of each model, the statistically optimal search region most likely to capture the true measurement is defined. This forms a complementary and efficient distributed search network in cluttered environments, significantly improving the capture score rate for the first detection of the true target measurement in dense false alarms, while avoiding excessive clutter due to an overly large gate or missed targets due to an overly small gate. The pre-defined hierarchical correlation strategy, through a two-level progression of initial screening within the model and screening between models, transforms the complex global optimal search problem into local optimal judgment and high-level arbitration. This enables the correlation decision to possess environmental awareness and target recognition capabilities, outputting high-confidence globally optimal measurements and their corresponding matching models in environments full of uncertainty. Kalman filter updates based on a single matching model and globally optimal measurements offer a clear update path, avoiding potential mutual interference and information dilution in multi-model soft fusion. This ensures the theoretical optimality of state estimation under a given decision. Since the measurements driving the update are rigorously selected optimal measurements, their innovation terms are highly reliable, resulting in accurate state correction direction and reasonable magnitude, effectively pulling the estimated trajectory back to the true path. Multi-model active coverage addresses maneuver unpredictability, independent gate fine-grained search resolves clutter spatial interference, and hierarchical information fusion decision-making resolves correlation ambiguity. Finally, optimal estimation based on explicit decisions achieves precise state correction and tracking closure, enabling the tracking system to maintain strong robustness in maintaining track continuity and stable state estimation under the dual pressure of low-altitude dense clutter and strong target maneuvering. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0010] Figure 1 A flowchart illustrating a target tracking method for a low-altitude, highly maneuverable unmanned aerial vehicle (UAV) provided in this specification.

[0011] Figure 2 An example diagram illustrating the tracking results of a tracking method using a traditional single-model association, provided as an embodiment of this specification;

[0012] Figure 3 An example diagram illustrating the tracking results of a multi-model and hierarchical association strategy method provided in the embodiments of this specification;

[0013] Figure 4 This is a schematic diagram of the structure of a target tracking device for a low-altitude, highly maneuverable unmanned aerial vehicle (UAV) provided as an embodiment of this specification. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0015] This specification provides a target tracking method for a low-altitude, highly maneuverable unmanned aerial vehicle (UAV). It should be noted that the execution entity in this specification can be a server or any device with data processing capabilities. Figure 1 This is a flowchart illustrating a target tracking method for a low-altitude, highly maneuverable unmanned aerial vehicle (UAV) provided in an embodiment of this specification, as shown below. Figure 1 As shown, the main steps include the following:

[0016] Step S101: Obtain the set of measurement points output by the radar in the current detection period, as well as the target's state estimate and state covariance matrix at the previous moment. Then, through multiple preset strong maneuvering models, make a one-step prediction for the previous moment to obtain the current prediction index corresponding to each strong maneuvering model at the current moment.

[0017] The current forecasting metric includes the forecasting status and the forecasting covariance matrix.

[0018] Conventional multi-model tracking methods commonly employ general model sets such as constant velocity (CV), constant acceleration (CA), and cooperative turning (CT), which are derived from general abstractions of traditional aircraft motion. Their assumed motion patterns (such as constant acceleration and constant turn rate) are relatively mild and continuous, failing to effectively encompass the discrete, abrupt, and highly dynamic maneuvering characteristics exhibited by UAVs in complex low-altitude environments. For example, the CA model can describe acceleration but not sudden stops, and the CT model describes smooth arc turns rather than sharp right-angle turns. When the target performs such strong maneuvers, the predicted trajectory based on conventional models will deviate systematically and significantly from the target's actual path. This model mismatch directly leads to subsequent correlation gates failing to encompass the actual measurement points, causing correlation failure and tracking interruption.

[0019] In one embodiment of this specification, the high-maneuverability model includes a constant speed model, a hovering model, a U-turn model, a left-side tangential maneuvering model, and a right-side tangential maneuvering model. The state transition matrix of the constant speed model describes the motion of a target with a constant speed. The transition coefficients corresponding to the velocity components in the state transition matrix of the hovering model are less than 1, describing the motion of a target whose speed decays to near rest. The transition coefficients corresponding to the position components in the state transition matrix of the U-turn model have opposite signs to the transition coefficients corresponding to the velocity components, describing the motion of a target moving in the opposite direction. The state transition matrix of the left-side tangential maneuvering model describes the motion of a target whose velocity direction deflects counterclockwise by a preset angle in a two-dimensional plane. The state transition matrix of the right-side tangential maneuvering model describes the motion of a target whose velocity direction deflects clockwise by a preset angle in a two-dimensional plane.

[0020] In one embodiment of this specification, after the radar system completes a full scan cycle, the signal processing unit outputs a set of measurement points containing all detected points at the current moment. Each measurement point includes its position information, echo amplitude information, and timestamp in the radar coordinate system. Simultaneously, the tracking processing unit reads from its own buffer the target's final state estimate and its state covariance matrix after being updated by Kalman filtering in the previous tracking cycle (time k-1). The state estimate is a vector containing the target's position and velocity, and the state covariance matrix describes the uncertainty of the state estimate.

[0021] Subsequently, five pre-defined high-maneuverability models began working in parallel. These five models—constant velocity CV, hovering, U-turn, left tangential, and right tangential—each characterized the target's future behavior using different kinematic assumptions. Each model possessed a predefined state transition matrix, the mathematical form of which directly encoded the model's kinematic assumptions. The constant velocity model's state transition matrix assumed a constant velocity; the hovering model's matrix introduced a decay factor in the velocity dimension, causing the predicted velocity to approach zero to simulate stagnation; the U-turn model's matrix reversed the velocity direction while maintaining the previous velocity direction, thus describing a sharp turn; and the left and right tangential models' state transition matrices caused the predicted velocity direction to deflect at a fixed angle counterclockwise or clockwise, respectively, to describe rapid right-angle turns.

[0022] Based on these five different state transition matrices, and a shared previous-time state estimate and covariance matrix, five independent Kalman prediction steps are executed in parallel. Specifically, for each model, the prediction operation involves multiplying the previous-time state estimate vector by the corresponding state transition matrix to obtain the predicted state at the current time (time k) based on the model's assumptions. Simultaneously, the previous-time state covariance matrix is ​​also transformed through the state transition matrix and superimposed with the process noise covariance matrix, representing the model's own uncertainty, to finally obtain the current-time prediction covariance matrix. Thus, five different sets of current-time prediction indices are generated for the same target; that is, five sets of predicted states and their uncertainty measures derived from different motion assumptions, i.e., prediction covariance matrices, providing a complete set of prediction candidates for subsequent measurement correlation in multiple possible motion trajectory directions.

[0023] Specifically, the algorithm input is the estimated motion state of the target at time k-1. and its covariance There is also the set of measurement points detected by the radar on the k-th screen when it is operating in search mode. .in, Indicates the first There are 1 measurement, totaling 1 One measurement. First, based on the defined five types of strong maneuver models, the Kalman prediction step is used to measure... and Make a prediction.

[0024]

[0025] in , and The first The predicted state, predicted covariance, and state transition matrix of each maneuver model are given. Furthermore, the state transition matrices for the five types of maneuver models are defined as follows:

[0026]

[0027] Among them, F1, F2, F3, F4 and F5 are the state transition matrices of five strong maneuvering models: constant speed CV, hovering, U-turn, left tangential, and right tangential.

[0028] The state prediction is based on the above five maneuver models, and the specific method is as follows:

[0029] A uniform velocity CV model is defined as one where the direction and magnitude of the target's velocity remain unchanged. Based on the uniform velocity CV model, the target's motion state is predicted according to the Kalman prediction formula. and its covariance Perform a state prediction step. These represent the target's position and velocity along the x and y axes, respectively. It is a 4x4 covariance matrix describing the uncertainty. The update formula is as follows:

[0030]

[0031] in, and The predicted state and predicted covariance are obtained based on the uniform velocity CV model. This is the state transition matrix of the CV model. This indicates the data rate of the measurement.

[0032] The hovering model is defined as the target suddenly coming to a stop, i.e., the target's velocity magnitude in all dimensions becomes zero. Based on the hovering model, the target's motion state is predicted using the Kalman prediction formula. and its covariance Perform a state prediction step. These represent the target's position and velocity along the x and y axes, respectively. It is a 4x4 covariance matrix describing the uncertainty. The update formula is as follows:

[0033]

[0034] in, and The predicted state and predicted covariance are obtained based on the hovering model. This is the state transition matrix for the hovering model.

[0035] The turning-around model is defined as a target suddenly turning back within one scan cycle, specifically characterized by the target's velocity magnitude remaining unchanged, but its direction reversing. Based on the turning-around model, the target's motion state is predicted using the Kalman comprehension formula. and its covariance Perform a state prediction step. These represent the target's position and velocity along the x and y axes, respectively. It is a 4x4 covariance matrix describing the uncertainty. The update formula is as follows:

[0036]

[0037] in, and The predicted state and predicted covariance are obtained based on the U-turn model. This is the state transition matrix for the turnaround model.

[0038] The left-side tangential model is defined as a target undergoing a sharp turn to the left of its direction of motion. Specifically, the magnitude of the target's velocity remains constant, but its direction of velocity rotates 90 degrees counterclockwise within the horizontal plane. Based on the left-side tangential model, the target's motion state is predicted using the Kalman spectroscopy formula. and its covariance Perform a state prediction step. These represent the target's position and velocity along the x and y axes, respectively. It is a 4x4 covariance matrix describing the uncertainty. The update formula is as follows:

[0039]

[0040] in, and The predicted state and predicted covariance are obtained based on the U-turn model. This is the state transition matrix for the turnaround model.

[0041] The right-side tangential model is defined as a target undergoing a sharp turn to the right in its direction of motion. Specifically, the magnitude of the target's velocity remains constant, but its direction of velocity rotates 90 degrees counterclockwise within the horizontal plane. Based on the right-side tangential model, the target's motion state is predicted using the Kalman prediction formula. and its covariance Perform a state prediction step. These represent the target's position and velocity along the x and y axes, respectively. It is a 4x4 covariance matrix describing the uncertainty. The update formula is as follows:

[0042]

[0043] in, and The predicted state and predicted covariance are obtained based on the U-turn model. This is the state transition matrix for the turnaround model.

[0044] Compared to conventional interactive multiple model (IMM) methods that typically employ basic models such as uniform velocity and uniform acceleration, this step uses pre-set high-maneuver models for hovering, U-turns, and tangential turns. This approach closely matches the unique and dramatic nonlinear motion patterns of low-altitude UAVs at the prediction level. Conventional methods often suffer from significant deviations between the predicted trajectories of their basic models and the actual maneuvers of targets, such as sudden stops or sharp turns, leading to ineffective inclusion of real measurements in subsequent correlation gates. In contrast, the dedicated high-maneuver models operating in parallel in this technical solution cover various extreme maneuver scenarios that the target might encounter when extrapolating the previous state. This ensures that at least one model's predicted direction closely approximates the target's actual motion trend. Essentially, before data correlation, the search scope is selectively expanded to multiple high-probability maneuver directions, rather than relying on a single, potentially severely deviated, prediction point. This lays a solid foundation for robustly capturing real target measurements in low-altitude cluttered environments. When a target suddenly maneuvers, its actual measurement may have moved far from the predicted position based on the assumption of uniform velocity, but it may fall within the prediction gate based on the turn-around or tangential model, thus avoiding the risk of immediate tracking loss due to model mismatch. By customizing the front-end prediction model set for specific scenarios, the system's ability to pre-match the motion patterns of low-altitude, highly maneuvering targets is significantly improved. This shifts some of the pressure of maneuver response from the back-end correlation and filtering algorithms to the prediction stage, reducing a series of subsequent correlation failures and tracking divergence problems caused by inaccurate predictions at the source, and enhancing the overall tracking system's foresight and fault tolerance in complex environments.

[0045] Step S102: Determine the elliptic correlation gate corresponding to each current prediction index, so as to determine the set of candidate measurement points falling within the elliptic correlation gate based on the set of measurement points.

[0046] The process involves determining the elliptic correlation gate corresponding to each current prediction index, and then identifying a set of candidate measurement points falling within the elliptic correlation gate based on the set of measurement points. Specifically, this includes: calculating the corresponding innovation covariance matrix based on the prediction covariance matrix and a preset measurement noise covariance matrix in the current prediction index of each strong maneuvering model; determining the boundary conditions of the elliptic correlation gate for each strong maneuvering model based on the innovation covariance matrix and preset threshold parameters; calculating the normalized distance of each measurement point relative to the prediction state based on the position information of each measurement point in the measurement point set, a preset observation matrix, and the prediction state of the strong maneuvering model; and determining whether the normalized distance of each measurement point satisfies the boundary conditions of the elliptic correlation gate. If so, the measurement point is included in the candidate measurement point set of the strong maneuvering model.

[0047] Conventional methods often employ a uniform gate, or one based on optimal model predictions. This gate, with its singular shape, cannot simultaneously adapt to the varying uncertainties arising from the diverse maneuvers a target may employ. For example, when a target flies at a constant speed, prediction uncertainty primarily expands along the velocity direction; however, when the target makes a sharp turn, the uncertainty increases dramatically and its direction changes. A single gate struggles to dynamically adapt to these changes; it may be set too large, introducing excessive false alarm clutter, or set too small, missing measurements of genuine maneuvering targets.

[0048] In one embodiment of this specification, a spatial acquisition region, namely an elliptical correlation gate, is established for the predicted state of each strong maneuvering model, and candidate points that may match the predictions of each model are initially screened from all measurement points provided by the radar using the elliptical correlation gate.

[0049] Using the prediction covariance matrix from the five current prediction indices obtained in the parallel computation of the previous step, the prediction covariance matrix quantifies the uncertainty of the target's true state relative to the predicted state after a one-step prediction based on the model. This uncertainty stems from process noise (i.e., the model's incomplete description of the target's random maneuvering) and the error propagation from the previous state estimation. However, to determine whether a radar measurement originates from the target corresponding to the predicted state, it is necessary to consider not only the prediction uncertainty but also the sensor's own observation error. Therefore, for each strong maneuvering model, the corresponding innovation covariance matrix is ​​first calculated based on its prediction covariance matrix and a preset measurement noise covariance matrix. The innovation covariance matrix essentially describes the statistical characteristics of the difference between the predicted measurement and the actual measurement. Its calculation formula directly adds the uncertainty of the predicted state (projected onto the measurement space through the observation matrix) to the sensor's inherent measurement noise. The innovation covariance matrix comprehensively reflects the total uncertainty throughout the entire chain from model prediction to sensor observation.

[0050] Next, based on the new information covariance matrix and preset threshold parameters, the precise mathematical boundary conditions of the elliptical correlation gate for each strong maneuvering model are determined. The threshold parameters are typically selected based on a chi-square distribution and are used to control the gate size, i.e., the confidence level for accepting a measurement as a candidate point. The boundary conditions are specifically expressed as a quadratic inequality: for a measurement point to fall into the elliptical gate, its distance relative to the predicted position must be less than or equal to the threshold value. This distance is not ordinary Euclidean distance, but a statistical distance that considers uncertainty. Then, the original set of measurement points output by the radar in the current cycle is scanned and evaluated one by one. For each measurement point in the set, its position information (such as range, azimuth, or x and y coordinates converted to a Cartesian coordinate system) is extracted. A preset observation matrix is ​​used, which defines how to extract directly observable quantities from the target state vector (including position and velocity).

[0051] For a given strong maneuvering model, the specific process for calculating the normalized distance of a measurement point relative to its predicted state is as follows: First, calculate the difference vector between the measurement point's position and the predicted measurement position obtained by projecting the model's predicted state through the observation matrix; this difference is called the innovation. Then, using the inverse of the previously calculated innovation covariance matrix as the weight matrix, calculate the weighted inner product of the innovation vectors, i.e., multiply the innovation vector on the left by its transpose and then on the right by the inverse of the innovation covariance matrix. This scalar result is the required normalized distance (or Mahalanobis distance). By whitening the difference vector using the inverse of the innovation covariance matrix, elliptical regions with different uncertainties in different directions are equivalent to an isotropic sphere when calculating the distance. Therefore, even if the prediction uncertainty is large in a certain direction (e.g., the velocity direction), causing the gate to be elongated in that direction, as long as the measurement point falls within this elongated ellipse, its calculated normalized distance will still be small. Finally, the normalized distance of each measurement point relative to each model is calculated and compared with the boundary conditions determined by the model based on threshold parameters. These boundary conditions are the threshold values. If the normalized distance of a measurement point relative to a model is less than or equal to the threshold value, the measurement point is determined to satisfy the boundary conditions of the elliptic correlation gate of the model. This means that, statistically speaking, the measurement point is highly likely to be an observation result generated by the target state predicted by the model, and therefore, the measurement point is added to the candidate measurement point set of this specific strong maneuvering model. After completing the above traversal calculation and judgment for all measurement points and all models, each strong maneuvering model obtains its own exclusive candidate measurement point set, which contains all radar traces that are statistically compatible with its predictions.

[0052] The following is a formulaic expression of the above process, based on the predicted state of each of the above strong maneuvering models. and , , used to represent the i-th strong maneuvering model, establishes an elliptical correlation gate:

[0053]

[0054] in The sensor at time k obtains the first... Individual measurement, For the observation matrix, For the preset threshold parameters, measurements that satisfy the conditions of the above formula are considered to fall into the threshold. Measurement points within the elliptic correlated gate established by the prediction model.

[0055] Compared to the fixed-size gates or gates calculated based on the covariance of a single prediction model commonly used in conventional tracking methods, the technical solution in this specification establishes an independent elliptical associated gate for each strong maneuvering model. Since the dynamic characteristics of each strong maneuvering model (such as hovering, turning, and tangential maneuvering) are different, the uncertainties introduced by one-step prediction are also drastically different. By independently calculating the innovation covariance matrix based on its own prediction covariance for each model, the shape and orientation of the elliptical gate are determined, ensuring that each gate can most accurately reflect the statistical distribution area where the measurement is most likely to occur if the target does indeed maneuver according to this model. The gate for the turning model may be elongated in the reverse velocity direction, while the gate for the tangential model may expand in the turning tangential direction. This precise matching greatly improves the probability that the gate spatially covers the actual target measurement. Secondly, a parallel and complementary search space was constructed. Five different models generated five elliptical gates with different shapes, positions, and orientations. Together, they formed a complementary search network covering multiple high-probability maneuver directions in the space at the current moment. Regardless of which of the five typical maneuvers the target actually performed, or even the transition between them, its true measurement point has a high probability of being captured by one or more correct gates. Compared with the traditional method that relies on a single gate for correlation, the parallel search mechanism significantly enhances the robustness and fault tolerance when the target performs unexpected maneuvers, and effectively reduces the risk of tracking loss caused by instantaneous mismatch of the prediction model.

[0056] Step S103: Through a preset hierarchical association strategy, perform intra-model association and inter-model association within the candidate measurement point set, and filter the global optimal measurement and the corresponding matching strong maneuver model at the current moment.

[0057] Using a pre-defined hierarchical association strategy, intra-model and inter-model associations are performed within the candidate measurement point set to filter the globally optimal measurement and the corresponding matching strong maneuver model at the current moment. Specifically, this includes: within the candidate measurement point set corresponding to each strong maneuver model, calculating the Euclidean distance between each candidate measurement point and the predicted state, and selecting the candidate measurement point with the smallest Euclidean distance as the candidate associated measurement of the strong maneuver model; if multiple candidate measurement points have the same minimum Euclidean distance, then based on the amplitude information of each candidate measurement point and the historical amplitude information of the target, the candidate associated measurement corresponding to the strong maneuver model is filtered from the multiple candidate measurement points; based on the amplitude information of each candidate associated measurement, the globally optimal measurement at the current moment is determined, and the matching strong maneuver model at the current moment is determined based on the strong maneuver model that generates the globally optimal measurement.

[0058] Based on the amplitude information of each candidate measurement point and the historical amplitude information of the target, candidate associated measurements corresponding to the strong maneuvering model are selected from multiple candidate measurement points. Specifically, this includes: obtaining the current flight altitude information of the target and determining whether the current flight altitude information of the target is lower than a preset altitude threshold; if the current flight altitude information is lower than the preset altitude threshold, obtaining the historical trajectory data of the target, which contains the amplitude value of at least one historically successfully associated measurement point; calculating the arithmetic mean of the amplitude values ​​of the at least one historically successfully associated measurement point to determine the historical amplitude reference value corresponding to the target, and calculating the difference between the amplitude information of each candidate measurement point and the historical amplitude reference value; determining the candidate measurement point with the smallest difference as the candidate associated measurement of the current strong maneuvering model; if the current flight altitude information is not lower than the preset altitude threshold, comparing the amplitude information of multiple candidate measurement points, and determining the candidate associated measurement of the current strong maneuvering model with the candidate measurement point with the largest amplitude information.

[0059] Based on the amplitude information of each candidate correlation measure, the method determines the globally optimal measure at the current moment, specifically including: comparing the amplitude information of each candidate correlation measure and determining the candidate correlation measure with the largest amplitude information as the globally optimal measure at the current moment. After comparing the amplitude information of each candidate correlation measure, the method further includes: if at least two candidate correlation measures have the same maximum amplitude information, obtaining the specified strong maneuvering model corresponding to each specified candidate correlation measure with the same maximum amplitude information; calculating the innovation distance between the maximum amplitude information and the prediction state based on the maximum amplitude information corresponding to each specified candidate correlation measure and the prediction state of the specified strong maneuvering model; and determining the globally optimal measure at the current moment using the specified candidate correlation measure with the smallest innovation distance.

[0060] Conventional tracking methods commonly employ Global Nearest Neighbor (GNN) association or Interactive Multiple Model (IMM) based soft decision-making using probability weighting. Conventional GNN methods perform a global minimum distance search across all measurements and all model predictions (or a single composite prediction), making them highly susceptible to being misled by nearest clutter points in dense clutter conditions and unable to distinguish between association quality differences under different maneuvering assumptions. While conventional IMM methods maintain multiple models, their association is essentially performed independently for each model, with the dominant model implicitly selected through model probability updates. This process is highly sensitive to association errors in single-frame measurements, and model probability convergence takes time, resulting in a lag in response to rapid, sudden maneuvers by UAVs.

[0061] In one embodiment of this specification, in order to accurately select the unique and reliable target measurement at the current moment from the numerous candidate measurements corresponding to each model gate, and determine its most matching maneuver mode, a hierarchical association strategy of intra-model association and inter-model association is used to determine the global optimal measurement at the current moment.

[0062] During the intra-model correlation phase, candidate correlation measurements are selected from the corresponding candidate measurement point set for each parallel-running strong maneuvering model. The specific implementation process is as follows:

[0063] First, for a high-maneuverability model (e.g., a U-turn model), its candidate measurement point set is read. This set contains all points statistically compatible with its prediction. Next, for each candidate measurement point in the set, the Euclidean distance between its position and the positional component included in the model's predicted state is calculated. Euclidean distance is the straight-line distance between two points; it is direct and fast, reflecting the geometrical proximity between the measurement point and the model's predicted point. After calculation, the Euclidean distances of all candidate measurement points are compared, and the measurement point with the smallest Euclidean distance is initially marked as a potential candidate associated measurement for the model. Here, it is assumed that, assuming the model's prediction is basically correct, the actual target's measurement should be closest to the predicted position.

[0064] The premise of predictive model-based association is the assumption that the predicted maneuver model conforms to the actual maneuver pattern of the target. In other words, if the target's actual motion pattern matches a pre-defined maneuver model, then the target's measurement point should be near the predicted point. Therefore, within the association gate of each model, the nearest neighbor association criterion is used as the primary criterion to select the locally optimal measurement. Specifically, the following formula is used to calculate the measurement falling within the predicted point. The distances between all measurements within the predicted correlation gate and the predicted position of each maneuvering model.

[0065]

[0066] Among them, symbols This represents finding the L2 norm of a vector. The smallest L2 norm is chosen. Corresponding measurement As the first Measurements associated with each maneuver model.

[0067] However, the dense clutter environment at low altitudes introduces uncertainty, potentially leading to multiple clutter points being equidistant from the predicted location, causing the nearest neighbor criterion to fail. That is, after screening using the nearest neighbor criterion, multiple measurements may be equidistant from the predicted location, meaning the nearest neighbor criterion cannot determine the locally optimal measurement. In this case, the maximum measurement amplitude criterion is used as a supplement. The process involves determining if there is one and only one measurement point with the minimum Euclidean distance; if two or more measurement points have the same minimum Euclidean distance, a refined screening sub-process is performed based on amplitude information and historical information. The target's current flight altitude estimate is obtained, derived from the altitude component in the previous state estimate or obtained by filtering continuously measured altitude information. This altitude value is then compared with a preset altitude threshold. This threshold physically distinguishes two distinct interference environments. For example, it can be set to 100 meters. Below this threshold, it represents the extremely low airspace, where the radar beam has a very small downward angle and strong ground clutter (such as buildings) dominates, and its echo amplitude may far exceed that of UAV targets. Above this threshold, it represents the medium and low airspace, where the influence of ground clutter is weakened, and the main interference may become targets with smaller scattering cross-sections, such as birds.

[0068] If the current altitude is below the threshold, the system enters a low-altitude anti-clutter mode, retrieving the most recent sequence of amplitude values ​​from successfully correlated measurement points used for status updates from the target's historical track archives. This sequence reflects the statistical characteristics of the target's radar cross-section (RCS). The arithmetic mean of this amplitude sequence is calculated as a historical amplitude reference value characterizing the typical echo intensity of the target. Subsequently, the absolute difference between the amplitude information (i.e., the amplitude value detected in this instance) of each candidate measurement point with the same minimum Euclidean distance and this historical amplitude reference value is calculated. The candidate point whose amplitude value is closest to the historical reference value, i.e., the candidate measurement point with the smallest absolute difference, is selected. The principle here is that the amplitude of strong ground clutter (such as buildings) may be abnormally high and unstable, while the amplitude of a real UAV target is relatively stable over a certain period of time, fluctuating around its mean. Therefore, when the distance cannot be distinguished, selecting the measurement whose amplitude behavior is most consistent with the target's historical characteristics can effectively filter out false alarms from strong stationary clutter that differs greatly from the target's RCS characteristics. Conversely, if the current altitude is not below the threshold, the system enters a mid-to-high altitude anti-weak interference mode. At this point, the amplitude information of these equidistant candidate measurement points is directly compared, and the point with the largest amplitude value is determined as the candidate associated measurement point. At this altitude level, the RCS of the main false alarm sources (such as birds) is usually smaller than that of UAVs, so the amplitude of the true target echo has an advantage.

[0069] For multiple measurements at the same distance from the predicted location, their detection amplitude information is obtained. Taking two measurements as an example, let's call them... and When applying the maximum amplitude screening criterion, the altitude layer of the predicted target location needs to be considered, and different maximum amplitude screening criteria are applied to different altitude layers. This fully considers the limitations of low-altitude detection scenarios. When the target altitude is below 100m, the radar detection beam needs to be configured with an extremely low elevation angle to detect the target. Due to the certain beamwidth of the radar beam, stationary targets on the ground, such as buildings, can be easily detected, resulting in false alarms with large amplitudes. Therefore, when the altitude is below 100m, we use the amplitude sequence of historically associated successful points stored in the track and calculate its average. ,choose and Closest in numerical terms The measurement point is selected as the locally optimal measurement. When the target altitude is greater than 100m, the influence of ground clutter is relatively small, the detection scene is relatively clean, and UAV targets are usually larger than low-flying birds in terms of radar cross-section (RCS), so the optimal measurement point can be directly selected. and The measurement corresponding to the larger of the two values ​​is used as the local optimum measurement to avoid repeatedly calculating the mean of the amplitude sequence of the track-related points, thus reducing the computational burden. Therefore, after determining the values ​​using the nearest neighbor criterion and the maximum amplitude criterion, at most one local optimum measurement can be selected for each maneuver model.

[0070] Subsequently, inter-model correlation is performed, with the goal of determining a unique globally optimal measurement from the candidate correlation measurements of each model, and thereby identifying the most likely maneuver pattern of the target. After intra-model screening, at most one locally optimal measurement has been selected for each maneuver model, avoiding the influence of strong ground clutter. The remaining rank only needs to consider UAV targets and bird targets in low-altitude detection scenarios. Furthermore, since the radar echo intensity of UAV targets is greater than that of low-altitude birds, the globally optimal measurement is selected primarily based on the criterion of maximizing the measurement amplitude among all locally optimal measurements of the models.

[0071] First, candidate correlation measurements for five strong maneuvering models are obtained. The amplitude information of each candidate correlation measurement is read, and their magnitudes are directly compared. The measurement point with the largest amplitude is initially determined as the globally optimal measurement at the current moment. The strong maneuvering model that generated this measurement is then initially determined as the matching model. This selection assumes that, after in-model screening, especially altitude-adaptive screening, the echo intensity of the real UAV target should be the most competitive among the remaining candidate measurements. However, a special case needs to be considered: since the correlation gates set by different maneuvering models may overlap, the same measurement may be selected by multiple model correlation gates. In this case, the maximum amplitude screening criterion cannot directly determine which maneuvering model matches the true motion of the target. At this point, amplitude alone cannot distinguish which model is more accurate. In the above situation, i.e., when the maximum amplitude screening criterion cannot determine, the minimum innovation distance criterion is used as the final decision. When two or more candidate correlation measurements are detected to have the same maximum amplitude value, the minimum innovation distance criterion is used for screening. For each candidate measurement with the same amplitude and its corresponding strong maneuvering model, the innovation distance of the measurement point relative to its corresponding model's predicted state is calculated. The innovation distance here refers to the normalized distance (Madara distance) used to construct the elliptical gate, which also considers the geometric deviation between measurement and prediction, as well as the uncertainty of model prediction (innovation covariance). After calculation, the innovation distances are compared, and the candidate measurement point with the smallest innovation distance is determined as the globally optimal measurement. The corresponding model is the final matched strong maneuver model.

[0072] The innovation distance is the optimal statistical measure of the degree of agreement between a measurement and a specific model prediction. When the magnitudes are the same, the model whose prediction makes the measurement point appear with a higher probability density (i.e., a smaller statistical distance) indicates that the model's motion assumptions match the target's actual maneuvering better.

[0073] Through initial selection within the model, local optima are determined independently and in parallel within five different motion hypothesis spaces. This ensures that even if a model's prediction deviates significantly from actual maneuvers, its association error remains confined to that model and does not pollute the global decision, improving the system's survivability when some models fail. Secondly, it organically integrates various heterogeneous information such as geometric distance (Euclidean distance), target feature information (radar amplitude and its historical statistics), environmental context (flight altitude), and prediction consistency (innovation distance). The adaptive amplitude selection based on flight altitude is not a fixed logic but a dynamically switching intelligent strategy according to the type of environmental threat. This allows the association criteria to closely align with the actual constraints of the physical world, effectively suppressing highly deceptive strong stationary clutter at extremely low altitudes, while leveraging the RCS advantage of UAVs against birds at medium and high altitudes for rapid screening. Finally, inter-model association utilizes the most discriminative amplitude information for rapid initial judgment. In the event of competition, it reverts to statistical distance for final arbitration, ensuring both processing efficiency under normal conditions and optimal decision-making under extreme conflict situations.

[0074] Step S104: Based on the global optimal measurement at the current moment, perform Kalman filtering to update the current prediction index of the matching strong maneuver model to obtain the updated prediction index of the target at the current moment, so as to achieve target tracking.

[0075] The updated prediction index includes the updated state estimate and the updated state covariance matrix.

[0076] Based on the global optimal measurement at the current moment, the current prediction index of the matched strong maneuvering model is updated using Kalman filtering. Specifically, this includes: calculating the Kalman gain based on the predicted state and prediction covariance matrix of the matched strong maneuvering model; correcting the predicted state of the matched strong maneuvering model based on the Kalman gain, the difference between the global optimal measurement and the predicted state of the strong maneuvering model, and obtaining the updated state estimate; and updating the updated state covariance matrix based on the Kalman gain and the prediction covariance matrix of the matched strong maneuvering model.

[0077] In one embodiment of this specification, a globally optimal measurement and a matched strong maneuvering model are received. Simultaneously, the current prediction index calculated by this matched strong maneuvering model in the previous prediction stage is read, namely, the model's predicted state vector and prediction covariance matrix. The Kalman filter update is based on this prediction value as a priori, and uses the globally optimal measurement as new information (i.e., new information brought by observations) for data fusion.

[0078] First, the predicted covariance matrix is ​​projected onto the measurement space through the observation matrix and its transpose. Then, the measurement noise covariance matrix is ​​added to obtain the inverse or correlation form of the innovation covariance matrix. Finally, the predicted covariance matrix and the transpose of the observation matrix are multiplied by the inverse of this intermediate result to solve for the Kalman gain matrix. The calculation principle of this gain matrix is ​​based on the minimum mean square error estimation criterion, and its mathematical result ensures that the final state correction can reduce the estimation error most effectively in a statistical sense. If the model prediction is very uncertain in a certain state component (corresponding to a large predicted covariance), while the sensor's observation in that direction is relatively accurate (low measurement noise), then the gain value corresponding to that component is large, meaning that more measurement information will be adopted during the update; conversely, the predicted value will be trusted more.

[0079] After obtaining the Kalman gain, the innovation is calculated, which is the difference vector between the actual position of the globally optimal measurement and the predicted measurement position obtained by projecting the predicted state of the matching model through the observation matrix. This difference directly reflects the inconsistency between sensor observations and model predictions. The calculated Kalman gain matrix is ​​then multiplied by this innovation vector. Using the Kalman gain as the optimal weight, the positional difference between observation and prediction is reasonably allocated and fed back to the correction of the complete state vector (including position, velocity, etc.). The result of this multiplication is the state correction. Finally, this correction is added to the original predicted state vector of the matching model to obtain the updated state estimate for the current moment. This ensures that the final state estimate incorporates both the dynamic recursion of the model based on historical information and the evidence provided by the latest observations at the current moment, theoretically achieving the optimal estimate given the information.

[0080] After correcting the state estimate, the state covariance matrix is ​​updated synchronously. Using the previously obtained Kalman gain and the original predicted covariance matrix of the matched model, the standard formula is used to subtract a term related to the Kalman gain and innovation covariance from the predicted covariance through specific matrix operations. Due to the injection of observational information, some uncertainties in the original state are eliminated. The calculated updated state covariance matrix will be numerically smaller than the predicted covariance matrix. The updated covariance matrix serves as one of the inputs for the five strong maneuvering models to make a new round of one-step predictions at the start of the next tracking cycle, thus initiating the next recursive loop of prediction, association, and update. In this way, not only is the most accurate state estimate of the target at the current moment output, but an optimized uncertainty measure is also provided for the continuous tracking of its future motion, achieving a closed loop and continuation of the tracking process.

[0081] Specifically, using the associated globally optimal measurement, the Kalman update formula is applied to the predicted state. , Make corrections to obtain the current state update of the maneuvering target. and By repeating the above steps, the motion state of the maneuvering target at each time step can be iteratively solved, thus achieving robust tracking of the maneuvering target. Figure 2 This is an example diagram illustrating the tracking results of a tracking method using a traditional single-model association, provided in an embodiment of this specification. Figure 3 An example diagram illustrating the tracking results using a multi-model and hierarchical association strategy method provided in this specification is shown below. Figure 2 and Figure 3 As shown, the method implemented in this specification can effectively improve the trajectory continuity of maneuvering UAV targets in complex low-altitude scenarios.

[0082] After determining the set of candidate measurement points that fall within the elliptical associated gate based on the set of measurement points, the method further includes: if the set of candidate measurement points for all strong maneuvering models is empty, then a specified model is selected from the multiple strong maneuvering models based on a preset rule, so that the predicted state and predicted covariance matrix of the specified model are directly used as the updated state estimate and updated state covariance matrix of the target at the current time.

[0083] Selecting a designated model from multiple strong maneuvering models based on preset rules specifically includes: obtaining a sequence of association success flags for each strong maneuvering model within multiple preset historical tracking periods, wherein each element in the association success flag sequence indicates whether the strong maneuvering model has successfully associated with a candidate association measurement within the corresponding historical tracking period; based on the association success flag sequence of each strong maneuvering model, calculating the cumulative number of association successes for each strong maneuvering model within the multiple historical tracking periods; calculating the association success weight corresponding to each strong maneuvering model based on the cumulative number of association successes, wherein the higher the cumulative number of association successes, the greater the corresponding association success weight; and selecting the designated model from the multiple strong maneuvering models according to the association success weight corresponding to each strong maneuvering model.

[0084] In one embodiment of this specification, fault tolerance and state maintenance under extreme conditions are provided to address the technical problem of how to continue tracking when the target fails to capture any candidate measurements by all preset elliptical correlation gates of the strong maneuvering model due to strong maneuvering, brief obstruction, or extremely dense clutter.

[0085] After evaluating the candidate measurement point sets for all high-maneuverability models, a global detection logic is initiated to check the candidate sets for the five models: constant speed, hovering, U-turn, left tangential, and right tangential. If the detection finds that all five sets are empty (meaning no radar measurement point falls within the statistically reliable region of any model in the current cycle), it is determined that an unrelated measurement event has been triggered, and model selection rules are then applied to avoid interruption of the tracking filter update. First, a sequence of association success flags is recorded for each high-maneuverability model (e.g., the U-turn model). This sequence is a fixed-length first-in-first-out queue, with its length corresponding to multiple preset historical tracking cycles, such as the most recent 30 radar scan cycles. Each element in the sequence corresponds to a specific historical cycle, such as time k-1, time k-2, etc., and its value is a binary flag. A value of 1 indicates that within that historical cycle, the model successfully output a candidate associated measurement during the hierarchical association process at that time, meaning its candidate set is not empty and it is preferred during the model's association phase. A value of 0 indicates that the model failed to associate successfully at that time, its candidate set is empty, or it was not selected in its internal filtering. When a new tracking period is completed, regardless of whether the above process is triggered, the sequence is updated, removing the oldest historical flag and adding the success or failure result of the current period's association as a new flag to the sequence. Therefore, this sequence reflects the activity and effectiveness of each model in recent history in real time.

[0086] For each high-maneuverability model, the complete sequence of successful association flags is read, and all flag values ​​(1 and 0) in the sequence are summed. The sum represents the cumulative number of successful associations for that model across multiple historical tracking periods, intuitively and quantitatively reflecting the model's effective matching degree to target motion patterns in the recent past. For example, a model that frequently describes target maneuvers (such as a target performing a series of evasive turns) in the past (e.g., a left-side tangential model) will have a significantly higher cumulative number of associations than a model that has not been matched for a long time (e.g., a hovering model). A higher cumulative number indicates stronger reliability and relevance of the model in the recent tracking context.

[0087] Based on the cumulative number of successful associations, the association success weight for each model is calculated. A pre-defined normalization function is used to process the data. The cumulative number of successful associations for all five models is summed to obtain the total number. Then, the number of each model's associations is divided by this total number, resulting in a normalized weight value between 0 and 1. The sum of the weights for all models is 1. Alternatively, a non-linear mapping function can be introduced to square or exponentially calculate the number of associations before normalization, amplifying the weight advantage of high-number models. The higher the cumulative number of successful associations, the greater the corresponding association success weight, ensuring that the model selection decision is based on objective, data-driven historical performance, rather than fixed or random selection.

[0088] Finally, based on the calculated association success weights of each model, a unique designated model is selected from the five models using a preset strategy. The strong maneuvering model with the highest association success weight is directly selected as the designated model. When the weight calculation uses the aforementioned normalization method, it is equivalent to selecting the model with the highest cumulative number of recent association successes. After selecting the designated model, Kalman filtering updates are no longer needed (because no new information measurements are available). Instead, the predicted state and prediction covariance matrix obtained by the designated model in one step at the current time are directly assigned to the updated state estimate and updated state covariance matrix of the target at the current time, respectively. In the absence of observational information, the prediction of the future using the most reliable motion model in the recent time allows the target's state to continue to extrapolate along the most likely motion pattern in the recent time. At the same time, the state covariance will increase reasonably due to the lack of new observations, which is already included in the prediction covariance, reflecting the increase in estimation uncertainty. This state will serve as the starting point for the next tracking iteration, thus maintaining a reasonable transient estimate consistent with the recent motion trend for tracking when the target temporarily disappears into clutter or undergoes extreme unmodeled maneuvers. This greatly improves the survivability and continuity of the track in harsh environments.

[0089] Based on historical performance-adaptive model selection, this approach addresses the frequent short-term data loss issues in low-altitude, highly maneuverable UAV tracking. It dynamically determines which model to trust when observations are unavailable by considering the recent successful association history of each model. Essentially, it leverages target maneuver pattern preferences learned from the recent tracking history. If the target was frequently performing tangential maneuvers before signal loss, the tangential model will have a high weight, and it will naturally be selected for state extrapolation after signal loss. The predicted turning trend is far more likely to be consistent with the target's actual behavior than blindly choosing a uniform speed model. This ensures that state estimation does not deviate significantly during signal interruption, but rather follows a trajectory consistent with recent behavioral logic, preserving greater flexibility for re-association after signal recovery. By introducing historical performance-based learning and decision-making in extreme situations, the system significantly improves its fault tolerance, adaptability, and track maintenance capabilities in highly uncertain scenarios such as complex low-altitude electromagnetic environments.

[0090] The technical solution of the embodiments in this specification introduces multiple preset strong maneuvering models in parallel during the prediction stage. By using models of maneuvering modes such as hovering, turning, and sharp turns of low-altitude UAVs, state inference is performed simultaneously along multiple high-probability maneuvering directions at each moment. This greatly reduces the systemic risk of the predicted point deviating significantly from the actual target position due to model mismatch. Regardless of what kind of maneuver the target actually performs, there will always be one or more models whose predicted direction roughly matches it, thus preserving the possibility for subsequent data association. This avoids the fatal weakness of traditional single models or limited model sets where predictions completely fail when the target suddenly maneuvers, directly leading to tracking loss. Secondly, by independently establishing an elliptical correlation gate strategy for each model, personalized management of prediction uncertainty and refined spatial search are achieved. The uncertainties of different maneuvering models exhibit anisotropic characteristics. Based on the dynamic characteristics of each model, the statistically optimal search region most likely to capture the true measurement is defined. This forms a complementary and efficient distributed search network in cluttered environments, significantly improving the capture score rate for the first detection of the true target measurement in dense false alarms, while avoiding excessive clutter due to an overly large gate or missed targets due to an overly small gate. The pre-defined hierarchical correlation strategy, through a two-level progression of initial screening within the model and screening between models, transforms the complex global optimal search problem into local optimal judgment and high-level arbitration. This enables the correlation decision to possess environmental awareness and target recognition capabilities, outputting high-confidence globally optimal measurements and their corresponding matching models in environments full of uncertainty. Kalman filter updates based on a single matching model and globally optimal measurements offer a clear update path, avoiding potential mutual interference and information dilution in multi-model soft fusion. This ensures the theoretical optimality of state estimation under a given decision. Since the measurements driving the update are rigorously selected optimal measurements, their innovation terms are highly reliable, resulting in accurate state correction direction and reasonable magnitude, effectively pulling the estimated trajectory back to the true path. Multi-model active coverage addresses maneuver unpredictability, independent gate fine-grained search resolves clutter spatial interference, and hierarchical information fusion decision-making resolves correlation ambiguity. Finally, optimal estimation based on explicit decisions achieves precise state correction and tracking closure, enabling the tracking system to maintain strong robustness in maintaining track continuity and stable state estimation under the dual pressure of low-altitude dense clutter and strong target maneuvering.

[0091] This specification also provides a target tracking device for a low-altitude, highly maneuverable unmanned aerial vehicle (UAV), such as... Figure 4 As shown, the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method.

[0092] This specification also provides a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0093] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0094] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0095] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A target tracking method of a low-altitude strong maneuvering unmanned aerial vehicle, characterized in that, The method comprises: acquiring a set of measurement points output by a radar in a current detection period, and state estimation value and state covariance matrix of a target at a previous time, to perform one-step prediction on the previous time through a plurality of preset strong maneuvering models, to obtain current prediction indexes corresponding to the current time for each strong maneuvering model, wherein the current prediction indexes comprise predicted state and predicted covariance matrix; determining an elliptical association gate corresponding to each current prediction index, to determine a set of candidate measurement points falling within the elliptical association gate according to the set of measurement points; performing intra-model association and inter-model association in the set of candidate measurement points through a preset hierarchical association strategy, to screen a globally optimal measurement corresponding to the current time and a matching strong maneuvering model; performing Kalman filter updating on the current prediction index of the matching strong maneuvering model according to the globally optimal measurement corresponding to the current time, to obtain updated prediction indexes of the target at the current time, to realize target tracking, wherein the updated prediction indexes comprise updated state estimation value and updated state covariance matrix.

2. The target tracking method of a low-altitude strong maneuvering unmanned aerial vehicle according to claim 1, characterized in that, The strong maneuvering model comprises a uniform speed model, a hovering model, a U-turn model, a left tangential maneuvering model and a right tangential maneuvering model; wherein the state transition matrix of the uniform speed model is used to describe motion of the target with constant speed; the transition coefficient corresponding to the speed component in the state transition matrix of the hovering model is less than 1, and is used to describe motion of the target with speed decaying to near static; the transition coefficient corresponding to the position component in the state transition matrix of the U-turn model is opposite in sign to the transition coefficient corresponding to the speed component, and is used to describe motion of the target with motion direction reversing; the state transition matrix of the left tangential maneuvering model is used to describe motion of the target with speed direction deflecting by a preset angle counterclockwise in a two-dimensional plane; the state transition matrix of the right tangential maneuvering model is used to describe motion of the target with speed direction deflecting by a preset angle clockwise in a two-dimensional plane.

3. The target tracking method of a low-altitude high-maneuverable unmanned aerial vehicle according to claim 1, wherein, The method comprises: calculating an innovation covariance matrix corresponding to each strong maneuvering model according to the predicted covariance matrix in the current prediction index of each strong maneuvering model and a preset measurement noise covariance matrix; determining boundary conditions of the elliptical association gate of each strong maneuvering model based on the innovation covariance matrix and a preset threshold parameter; calculating a normalized distance of each measurement point relative to the predicted state according to position information of each measurement point in the set of measurement points, a preset observation matrix and the predicted state of the strong maneuvering model; judging whether the normalized distance of each measurement point satisfies the boundary conditions of the elliptical association gate, and if yes, the measurement point is classified into the set of candidate measurement points of the strong maneuvering model.

4. The target tracking method of a low-altitude high-maneuverable unmanned aerial vehicle according to claim 1, wherein, The method comprises: performing intra-model association and inter-model association in the set of candidate measurement points through a preset hierarchical association strategy, to screen a globally optimal measurement corresponding to the current time and a matching strong maneuvering model, specifically comprising: calculating the Euclidean distance between each candidate measurement point and the predicted state in each candidate measurement point set corresponding to the strong maneuvering model, and selecting a candidate measurement point with the minimum Euclidean distance as the candidate associated measurement of the strong maneuvering model; if there are multiple candidate measurement points with the same minimum Euclidean distance, screening the candidate associated measurement corresponding to the strong maneuvering model from the multiple candidate measurement points based on the amplitude information of each candidate measurement point and the historical amplitude information of the target; determining the globally optimal measurement at the current time according to the amplitude information of each candidate associated measurement, and determining the matching strong maneuvering model at the current time with the strong maneuvering model that produces the globally optimal measurement.

5. The target tracking method of a low-altitude high-maneuverable unmanned aerial vehicle according to claim 4, wherein, screening the candidate associated measurement corresponding to the strong maneuvering model from the multiple candidate measurement points based on the amplitude information of each candidate measurement point and the historical amplitude information of the target, specifically including: obtaining the current flight height information of the target, and judging whether the current flight height information of the target is lower than a preset height threshold; if the current flight height information is lower than the preset height threshold, obtaining the historical track data of the target, the historical track data containing the amplitude values of at least one historically associated successful measurement point; calculating the arithmetic mean of the amplitude values of the at least one historically associated successful measurement point to determine the historical amplitude reference value corresponding to the target, and calculating the difference between the amplitude information of each candidate measurement point and the historical amplitude reference value; determining the candidate associated measurement of the current strong maneuvering model as the candidate measurement point with the minimum difference; if the current flight height information is not lower than the preset height threshold, comparing the amplitude information of the multiple candidate measurement points, and determining the candidate associated measurement of the current strong maneuvering model as the candidate measurement point with the maximum amplitude information.

6. The target tracking method of a low-altitude high-maneuverable unmanned aerial vehicle according to claim 4, wherein, determining the globally optimal measurement at the current time according to the amplitude information of each candidate associated measurement, specifically including: comparing the amplitude information of each candidate associated measurement, and determining the candidate associated measurement with the maximum amplitude information as the globally optimal measurement at the current time.

7. The target tracking method of a low-altitude high-maneuverable unmanned aerial vehicle according to claim 6, wherein, After comparing the amplitude information of each candidate associated measurement, the method further includes: if there are at least two candidate associated measurements with the same maximum amplitude information, obtaining the specified strong maneuvering model corresponding to each specified candidate associated measurement with the same maximum amplitude information; calculating the innovation distance between the maximum amplitude information of each specified candidate associated measurement and the predicted state of the specified strong maneuvering model according to the maximum amplitude information of each specified candidate associated measurement and the predicted state of the specified strong maneuvering model; determining the globally optimal measurement at the current time as the specified candidate associated measurement with the minimum innovation distance.

8. The target tracking method of a low-altitude high-maneuverable unmanned aerial vehicle according to claim 1, wherein, performing Kalman filter update on the current prediction index of the matching strong maneuvering model according to the globally optimal measurement corresponding to the current time, specifically including: calculating the Kalman gain according to the predicted state and the predicted covariance matrix of the matching strong maneuvering model; According to the Kalman gain, a difference between the global optimal measurement and a predicted state of the matched strong maneuver model, the predicted state of the matched strong maneuver model is corrected to obtain the updated state estimation value; According to the Kalman gain and a predicted covariance matrix of the matched strong maneuver model, the updated state covariance matrix is obtained.

9. The target tracking method of a low-altitude high-maneuverable unmanned aerial vehicle according to claim 1, wherein, After determining the candidate measurement point set falling into the elliptical association gate according to the measurement point set, the method further comprises: If all the candidate measurement point sets of the strong maneuver models are empty, a specified model is selected from the plurality of strong maneuver models based on a preset rule, and a predicted state and a predicted covariance matrix of the specified model are directly taken as the updated state estimation value and the updated state covariance matrix of the target at the current time, respectively.

10. The target tracking method of a low-altitude high-maneuverable unmanned aerial vehicle according to claim 9, wherein, The method for selecting the specified model from the plurality of strong maneuver models based on the preset rule comprises: An association success flag sequence of each strong maneuver model in a plurality of preset historical tracking periods is obtained, wherein each element in the association success flag sequence is used to indicate whether the strong maneuver model is successfully associated to a candidate association measurement in a corresponding historical tracking period; Based on the association success flag sequence of each strong maneuver model, a cumulative association success number of each strong maneuver model in the plurality of historical tracking periods is obtained; Based on the cumulative association success number of each strong maneuver model, a corresponding association success weight of the strong maneuver model is calculated, wherein the higher the cumulative association success number is, the greater the corresponding association success weight is; According to the corresponding association success weight of each strong maneuver model, the specified model is selected from the plurality of strong maneuver models.

Images