Complex scene-oriented unmanned aerial vehicle active tracking method and system
By introducing SMDP and dynamic probability transition matrix into UAV target tracking, the problems of model switching lag and tracking accuracy degradation are solved, enabling UAVs to track quickly and accurately in complex scenarios, and improving tracking stability and adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-15
- Publication Date
- 2026-04-14
AI Technical Summary
Existing UAV target tracking methods struggle to adapt to abrupt changes in target maneuvering patterns in complex scenarios. They lack online adaptive capabilities to observation degradation, occlusion, and noise abrupt changes, resulting in lag in model switching, decreased tracking accuracy, and susceptibility to divergence.
Interactive multi-model filtering (IMM) combined with semi-Markov decision process (SMDP) is employed. By constructing a dynamic probability transition matrix and utilizing environmental context feature vectors and dwell time, adaptive switching and state fusion of the model are achieved, thereby enhancing tracking robustness.
It achieves rapid and accurate target identification and smooth switching, improves tracking stability and adaptability, reduces short-term tracking errors and model oscillations, and enhances online response to complex disturbances.
Smart Images
Figure CN121857737A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of drone and moving target tracking, and in particular to a drone active tracking method and system for complex scenarios. Background Technology
[0002] Currently, various solutions exist for the problem of active target tracking by unmanned aerial vehicles (UAVs), based on interactive multi-model (IMM) and single-model filtering, such as schemes based on nonlinear motion models, flight pattern identification, and cooperative path planning. These methods have improved the UAV's ability to estimate the state of maneuvering targets in complex scenarios to some extent, but they still generally rely on preset fixed model transition probabilities and adopt a data-driven, lag-response passive tracking paradigm.
[0003] Existing methods have the following limitations when dealing with common multimodal and non-stationary motions (such as sharp turns, alternating acceleration and deceleration, and temporary hovering) in UAV tracking: First, fixed transfer probabilities are difficult to adapt to sudden changes in target maneuvering patterns, resulting in model switching lag and boundary jitter; second, they lack online adaptive capabilities to complex interferences such as observation degradation, occlusion, and sudden noise changes; and third, traditional IMM structures lack temporal memory and forward-looking decision-making mechanisms, and model switching relies on instantaneous measurement matching, which can easily lead to tracking divergence or persistent deviations when signals are missing or mismatched.
[0004] Therefore, an innovative solution is urgently needed to resolve the contradiction between model switching lag and frequent jitter in active tracking of UAVs in complex scenarios, and to improve tracking robustness and state prediction accuracy. Summary of the Invention
[0005] The purpose of this invention is to provide an active tracking method and system for UAVs in complex scenarios. It aims to solve the problems of decreased tracking accuracy, response delay and easy divergence caused by the lag in model switching and the inability of fixed transfer structures to adapt when there are complex interferences such as sudden changes in target motion patterns, degradation of observation signals or occlusion. The invention achieves rapid and accurate identification and smooth switching of targets, and enhances tracking stability and adaptability.
[0006] To achieve the above objectives, this invention provides an active tracking method for unmanned aerial vehicles (UAVs) in complex scenarios, comprising the following steps: Step S1: Data preprocessing and construction of the IMM model set; Step S2: Based on the preprocessed data and the IMM model set, perform parallel filtering to obtain the fused state estimate and posterior probability; Step S3: Generate a dynamic probability transition matrix based on the preprocessed data and posterior probabilities; Step S4: Based on the dynamic probability transition matrix, perform interaction and state fusion on each motion model, and output the target state fusion estimate; Step S5: Based on the target state fusion estimation, perform time-series updates and state estimation transfer.
[0007] Preferably, in step S1, the constructed IMM model set includes CV model, CA model and CTRV model; Define a unified state vector for each motion model in the IMM model set, and establish the state equation and observation equation for each motion model based on a preset sampling time interval.
[0008] Preferably, in step S2, parallel filtering is implemented through the IMM interactive multi-model approach, specifically including: Step S201: Based on the state estimates, covariances, and preset transition probabilities of each motion model at the previous moment, calculate the initial state and initial covariance of each motion model at the current moment. Step S202: Based on the initial state and initial covariance, perform parallel filtering on each motion model and update the state estimate and covariance; Step S203: Calculate the likelihood of each model using the likelihood function, and update the posterior probability of each model at the current time by combining the model prediction probability. Step S204: Weigh and fuse the state estimates, covariances and posterior probabilities of each model to obtain the fused state estimates and fused covariance estimates.
[0009] Preferably, in step S202, the CV model is configured with standard KF for parallel filtering, and the CA model and CTRV model are configured with EKF for parallel filtering.
[0010] Preferably, in step S3, generating the dynamic probability transition matrix using SMDP specifically includes: Define system state The specific expression is: ; in, Represents the state index of the motion model; These represent three motion models. express Model, express Model, express Model; Represent the target state space; Indicates the duration of stay. ; Represents the target state vector; definition The environmental context feature vector at any given time , The specific expression is: ; in, Indicates the angular velocity of the heading; Indicates tangential angular velocity; Indicates the curvature of motion; Indicates the change in velocity; This represents the trace of the fused covariance estimate; Indicates the confidence level of measurement consistency; Indicates transpose; Define action space The specific expression is: ; in, ; This indicates that the current position is maintained after the action is executed. A motion model, and a sampling time interval. The duration of stay is accumulated. , express The duration of dwell time express Duration of stay at a given moment; Indicates that after the action is executed, from the current number... The motion model is switched to the first one. Each motion model has its dwell time reset. Switch to the current Effective immediately; Based on environmental context feature vectors and dwell time, a context-aware Weibull distribution is used to define the model dwell time distribution function. The specific expression is: ; ; in, Indicates the context scaling factor; Indicates the baseline characteristic time; Represents the dwell bias vector; Represents the natural exponential function; Introduce upper and lower bounds for dwell time [ ], This represents the minimum number of dwell steps. The time window penalty factor represents the maximum number of dwell times. The specific expression is: ; in, The non-negative truncation function is defined as follows: ; , Indicates the penalty coefficient; Based on environmental context feature vectors and time window penalty factors, the model's departure tendency quantification index is calculated. The specific expression is: ; in, Represents the context weight vector; This indicates that the context is feature-driven. Based on the model's residence time distribution function Quantitative indicators of the tendency to leave the model Calculate the probability of residence The specific expression is: ; in, The logistic function is defined as follows: ; Indicates the residency bias coefficient; A model switching strategy is introduced by combining model likelihood and environmental context feature vectors. Calculate the switching probability The specific expression is: ; ; in, This represents the minimum probability threshold. ; Indicates the first Contextual preference vectors for each motion model; This represents the weighted sum of the likelihoods of all candidate models and the context preference vectors; express Time of the first The likelihood of a motion model ( ); Indicates the model index; Indicates the first Contextual preference vectors for each motion model; Based on the probability of residence As a diagonal element, the switching probability Construct a dynamic probability transition matrix with off-diagonal elements. The specific expression is: .
[0011] The preferred action selection strategy in the action space is as follows: If the likelihood value of the current motion model is higher than the preset threshold and the dwell time is less than the preset maximum dwell time, then select... If the likelihood value of the current motion model is lower than a preset threshold, or the dwell time exceeds the maximum dwell time, then select... .
[0012] Preferably, in step S3, a reward function is set to optimize the dynamic probability transition matrix, and its specific expression is as follows: ; ; ; ; ; ; in, Indicates total return; Indicates the discount factor; express Instant rewards at any moment; , and Indicates parameter weights; This indicates a measurement adaptation feedback; This represents the IMM filter information; Indicates the new information covariance; This indicates that the length of stay is appropriate for the return. Indicates the optimal dwell time; Indicates the adaptation threshold; This indicates a context-adaptive reward; This represents the characteristic vector of measurement error.
[0013] Preferably, in step S4, the interaction and state fusion of each motion model is achieved through SMDP-IMM, specifically including: Step S401: Using the dynamic probability transition matrix and the posterior probability of the previous time step, calculate the mixed initial state and mixed covariance of each motion model at the current time step. Step S402: Based on the mixed initial state and mixed covariance, perform parallel filtering on each motion model and update the state estimate, covariance and likelihood of each motion model; Step S403: After completing parallel filtering, calculate the dynamic probability transition matrix for the next time step using SMDP based on the updated state estimates, covariance, and likelihood of each motion model. Step S404: Based on the updated likelihood and dynamic probability transition matrix, update the posterior probability of each motion model. The specific expression is as follows: ; in, express Time of the first The model probability predicted by each motion model; express Time of the first Likelihood of a motion model; express Time of the first Likelihood of a motion model; express Time of the first The posterior probability of a motion model; Step S405: Based on the updated posterior probability, perform posterior correction on the dwell time, fuse the state estimates and covariances of each motion model, and obtain the final target state fusion estimate, the specific expression of which is: ; ; ; ; in, This represents a mixture of weights based on posterior probabilities; Indicates the revised dwell time; This represents the target state fusion estimate; Indicates the first The posterior update value of each motion model; This represents the covariance of the target state fusion estimate; express Time of the first The posterior probability of each motion model.
[0014] This invention also provides an active tracking system for unmanned aerial vehicles (UAVs) in complex scenarios, employing the above-mentioned method, including: The data preprocessing and model building module is used to preprocess multi-source sensor data from UAVs and build an IMM model set. The multi-model parallel filtering and estimation module is used to perform parallel filtering based on preprocessed data and the IMM model set, calculate the likelihood of each motion model, and perform weighted fusion based on the likelihood to obtain the fused state estimate and posterior probability. The dynamic probability transition matrix generation module is used to generate a dynamic probability transition matrix based on the preprocessed data and the posterior probability. The adaptive interactive fusion module is used to perform interactive and state fusion on various motion models based on the dynamic probability transition matrix, and output the target state fusion estimate.
[0015] Therefore, the present invention employs the above-mentioned active tracking method and system for UAVs in complex scenarios, and the beneficial technical effects are as follows: (1) This invention solves the problems of model switching lag and boundary jitter by introducing SMDP to replace the fixed Markov switching of traditional IMM. The method of this invention combines environmental context feature vectors (such as curvature, acceleration, measurement confidence) and dwell time to realize prospective model selection based on environmental semantics and time constraints, so as to respond quickly when the target maneuvers suddenly changes and maintain sufficient dwell time in the stable phase, thereby reducing short-term tracking error and model oscillation.
[0016] (2) This invention enhances the online adaptive capability to complex interferences such as error / noise abrupt changes, outlier occlusion, and measurement degradation by using a dynamic probability transition matrix and a reward function. This invention can evaluate the model fit and observation quality in real time based on the innovation statistics and measurement confidence, and actively adjust the switching tendency and noise parameters in abnormal situations, thereby maintaining reasonable trajectory prediction during observation degradation or occlusion and avoiding filter divergence.
[0017] (3) By using the upper and lower bounds of dwell time and the early / late switching penalty mechanism, the present invention structurally unifies the constraints of “suppressing early switching” and “promoting late switching”, so that the system tends to maintain the current motion model to suppress jitter when the minimum dwell time is not reached, and forces a separation when the maximum dwell time is exceeded and the model is mismatched, thereby effectively balancing fast response and switching stability, and achieving a smoother and more reliable phase transition.
[0018] (4) This invention embeds SMDP into the classic IMM process. By simply inserting the dynamic probability transition matrix generation and update stage, it achieves closed-loop enhancement of "perception-decision-estimation". It is also compatible with KF / EKF parallel filtering and IMM probability update process, reducing computational complexity and having good scalability and integration. Attached Figure Description
[0019] Figure 1 This is a flowchart of an active tracking method for unmanned aerial vehicles (UAVs) in complex scenarios according to the present invention; Figure 2 Comparison chart of target tracking trajectories; Figure 3 A comparison chart of delay distribution for model identification; Figure 4 A comparison plot of the RMSE distribution for state estimation; where, Figure 4 (a) in the figure is the location RMSE distribution map; Figure 4 (b) in the figure is the velocity RMSE distribution diagram; Figure 4 (c) in the diagram represents the distribution of RMSE orientation. Detailed Implementation
[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0022] Example 1 like Figure 1 As shown, a method for active tracking of drones in complex scenarios includes the following steps: Step S1: Data preprocessing and construction of the IMM model set. Specifically, this includes: Based on the multiple sensors carried by the drone itself, raw data is collected: the accelerometer in the IMU (Inertial Measurement Unit) is responsible for collecting the drone's linear acceleration, the gyroscope accurately captures its angular velocity, and the GPS provides the drone's real-time position and speed information.
[0023] Preprocessing of the raw data includes: (1) Outlier detection and removal.
[0024] To address potential anomalies in sensor data caused by environmental interference (such as mechanical vibration, signal blockage, or multipath reflection), a statistical discrimination method is employed. Firstly, a reasonable data fluctuation range is established based on the 3σ criterion (Laida criterion), and data points exceeding this range are considered outliers and removed. Secondly, the Median Absolute Deviation (MAD) method is used to identify and remove isolated points deviating from the data set by calculating the dispersion of each data point from the median of the sequence. This step effectively filters out data jumps and transient distortions, improving overall data consistency.
[0025] (2) Moving average filtering.
[0026] To suppress the inherent high-frequency measurement noise in the original data, a moving average filter is used for preliminary smoothing. By setting a fixed-length sliding time window, the arithmetic mean of the continuously sampled data within the window is calculated and used as the filtered output at that moment. This method effectively smooths high-frequency fluctuations and preserves the macroscopic trend of the data, while reducing the computational complexity and numerical burden of subsequent filtering algorithms.
[0027] (3) Temporal interpolation filling.
[0028] When data sequences are missing due to sensor occlusion, signal loss, or communication interruption, a temporal interpolation strategy is employed to fill in the gaps. Based on the changing trends of valid data points before and after the missing time, local data curves are fitted using methods such as polynomial interpolation to generate continuous and reasonable filler values, thereby ensuring the integrity of the data sequence input to subsequent processing modules in the time dimension.
[0029] Construct an IMM model set, including CV model (uniform speed model), CA model (uniform acceleration model), and CTRV model (constant turning rate and speed model).
[0030] Define a unified state vector for each motion model in the IMM model set. And based on a preset sampling time interval State equations and observation equations for each motion model are established.
[0031] ; in, This indicates the target's position on the X-axis; This indicates the target's position on the Y-axis; Indicates the magnitude of speed; Indicates the direction of motion; This represents a variable that indicates specific dynamic information of the motion model (0 for the CV model; angular velocity for the CTRV model). For the CA model, it is acceleration. ); This indicates transpose.
[0032] The specific expression for the state equation of the CV model is: ; in, Indicates process noise; express The state vector at time t.
[0033] The state equations of the CTRV model describe the time interval Inside, position and orientation change with velocity. and angular velocity The changes. When When the x-axis is not close to zero, the state equations involve trigonometric functions and are nonlinear. When When the velocity approaches zero, the equation of state degenerates into an approximation of linear motion. When > At that time, the specific expression of the state equation is: ; in, It is a tool for processing The numerical stability threshold is close to zero.
[0034] The state equation of the CA model describes the time interval... Inside, position and velocity vary with velocity. and acceleration The change in the direction of motion. Under the CA model, the state equation remains constant and is nonlinear, specifically expressed as follows: ; in, This represents tangential acceleration.
[0035] The observation equations are consistent across the three motion models, and their specific expressions are as follows: ; ; in, Indicates measurement noise; Represents the current observation vector; This represents the measurement matrix.
[0036] Step S2: Based on the preprocessed data and the IMM model set, perform parallel filtering to obtain the fused state estimate and posterior probability.
[0037] Parallel filtering is achieved through interactive multi-model learning (IMM), specifically including: Step S201: Based on the state estimates, covariances, and preset transition probabilities of each motion model at the previous moment, calculate the initial state and initial covariance of each motion model at the current moment. The specific expression is as follows: ; ; ; in, Indicates the first A motion model in Posterior state estimation at time 1; Indicates mixed probability (interaction probability); Indicates from the first The motion model to the 1st The transition probabilities of each motion model; Indicates the first A motion model in The posterior probability at time t; This indicates the total number of motion models in the IMM model set. Indicates to the first A motion model in The initial state after mixing before filtering begins at time 1; Indicates the first A motion model in The posterior covariance matrix at time t; Indicates to the first A motion model in The initial covariance after mixing prepared before filtering begins at time 1; This indicates transpose.
[0038] Step S202: Based on the initial state and initial covariance, perform parallel filtering on each motion model to update the state estimate and covariance.
[0039] For parallel filtering of the CV model using the standard Kalman filter, the specific expression is: ; ; ; ; ; ; ; ; in, express The optimal estimate of the time step; Indicates from Time's up The state transition matrix at time t; express Time-based process noise; express Time of the first New information about a motion model; Indicates based on The optimal state estimation at time step Prior state estimation at time step; express Posterior state estimation at time 1; Indicates based on Time covariance obtained Time-prediction error covariance; express The posterior covariance matrix at time t; express Time-matrix noise covariance matrix; express Actual measurement data at any given time; express The observation matrix at each time point; express Time of the first Prior state estimation of a motion model; express Time of the first The new covariance of the motion model; express Time of the first The posterior covariance matrix of each motion model; express The noise covariance matrix is measured at any given time. express Time of the first Kalman gain of each motion model; express The inverse matrix; express Time of the first Posterior state estimation of a motion model; express Time of the first The posterior covariance matrix of each motion model; Represents the identity matrix.
[0040] Parallel filtering is performed using an EKF (Extended Kalman Filter) for both the CA and CTRV models. The specific expression is: ; ; ; ; ; ; ; ; ; ; in, express Time-based process noise; , Represents a nonlinear function; express For the state vector The Jacobian matrix; Indicates measurement noise; Indicates the estimated state; The Jacobian matrix represents the state transition function (i.e., the equivalent state transition matrix after linearization). The Jacobian matrix representing the measurement function; Represents the observation vector; express Posterior covariance estimation at time 1; This represents the prior covariance estimate; express Always Prior time estimation; This indicates the measurement residual (new information); express Time of the first The new covariance of the motion model; The process noise covariance matrix; Represents the measurement noise covariance matrix; Indicates to Post-time (posterior) state estimation; Indicates to The (posterior) covariance estimate updated at each time step.
[0041] Step S203: Calculate the likelihood of each model using the likelihood function, and update the posterior probability of each model at the current time step by combining the model prediction probability. The specific expression is as follows: ; ; ; in, express Time of the first Likelihood of a motion model; express Time of the first The model probability predicted by each motion model; The elements of the transition probability matrix (representing the first element) (elements in the column containing the motion model); express Time of the first The posterior probability of a motion model; express Time of the first The posterior probability of a motion model; Step S204: Weightedly fuse the state estimates, covariances, and posterior probabilities of each model to obtain the fused state estimate and fused covariance estimate, specifically expressed as follows: ; ; in, This represents the fusion state estimate; This represents the fusion covariance estimate; It is the first The posterior update value of each motion model; It is the first The posterior update covariance of each motion model.
[0042] Step S3: Based on the preprocessed data and posterior probabilities, generate a dynamic probability transition matrix using a semi-Markov decision process (SMDP). Specifically, this includes: Define system state The specific expression is: ; in, Represents the state index of the motion model; These represent three motion models. express Model, express Model, express Model; Represents the target state space, including location ( ),speed( ), angular velocity and acceleration; Indicates the duration of stay. ; This represents the target state vector.
[0043] definition The environmental context feature vector at any given time , The specific expression is: ; in, Indicates the angular velocity of the heading; Indicates tangential angular velocity; Indicates the curvature of motion; Indicates the change in velocity; This represents the trace of the fused covariance estimate; This indicates the confidence level of measurement consistency.
[0044] Define action space The specific expression is: ; in, ; This indicates that the current position is maintained after the action is executed. A motion model, and a sampling time interval. The duration of stay is accumulated. , express The duration of dwell time express Duration of stay at a given moment; Indicates that after the action is executed, from the current number... The motion model is switched to the first one. Individual motion model (target motion model), dwell time reset. Switch to the current Effective immediately.
[0045] The specific action selection strategy in the action space is as follows: If the likelihood value of the current motion model is higher than the preset threshold and the dwell time is less than the preset maximum dwell time, then select... If the likelihood value of the current motion model is lower than a preset threshold, or the dwell time exceeds the maximum dwell time, then select... .
[0046] Based on environmental context feature vectors and dwell time, a context-aware Weibull distribution is used to define the model dwell time distribution function. The specific expression is: ; ; in, Indicates the context scaling factor; Represents the natural exponential function; The reference feature time determines the basic dwell time range of the corresponding motion model. In this embodiment, it is set as follows: ; This represents the dwell bias vector; a larger value indicates a stronger bias towards dwell. As characteristics such as turning, acceleration / deceleration, and uncertainty increase, this bias becomes more pronounced. It will increase. It will decrease. The value will decrease, making it easier to leave the current motion model.
[0047] Introduce upper and lower bounds for dwell time [ ], This represents the minimum number of dwell steps. The time window penalty factor represents the maximum number of dwell times. The specific expression is: ; in, The non-negative truncation function is defined as follows: ; , This represents the penalty coefficient, which is set in this embodiment. .
[0048] when When this occurs, an early switching penalty is triggered, increasing the linear penalty. To suppress premature departure from the current motion model; when When this occurs, a late switching penalty is triggered, increasing the linear penalty. This prevents the current motion model from becoming stuck under obvious mismatch conditions.
[0049] Based on environmental context feature vectors and time window penalty factors, the model's departure tendency quantification index is calculated. The specific expression is: ; in, Represents the context weight vector; This indicates that the context is feature-driven.
[0050] Based on the model's residence time distribution function Quantitative indicators of the tendency to leave the model The residence probability is calculated using the logistic function. The specific expression is: ; in, The logistic function is defined as follows: ; This represents the residency bias coefficient.
[0051] A model switching strategy is introduced by combining model likelihood and environmental context feature vectors. Calculate the switching probability The specific expression is: ; ; in, This represents the minimum probability threshold. ; Indicates (target motion model) the first Contextual preference vectors for each motion model; This represents the weighted sum of the likelihoods of all candidate models and the context preference vectors; express Time of the first The likelihood of a motion model ( ); Indicates the model index; Indicates the first The context preference vector of a motion model.
[0052] Based on the probability of residence As a diagonal element, the switching probability Construct a dynamic probability transition matrix with off-diagonal elements. The specific expression is: .
[0053] In addition, a reward function is defined to optimize the dynamic probability transition matrix, the specific expression of which is: ; ; ; ; ; ; in, Indicates total return; This represents the discount factor, which is set in this embodiment. ; express Instant rewards at any moment; , and Indicates parameter weights; This indicates a measurement adaptation feedback; This represents the IMM filter information; Indicates the new information covariance; This indicates that the length of stay is appropriate for the return. Indicates the optimal dwell time; This represents the adaptation threshold; in this embodiment, the value is set to [value]. ; This indicates a context-adaptive reward; This represents the characteristic vector of measurement error.
[0054] Step S4: Based on the dynamic probability transition matrix, perform interaction and state fusion on each motion model to output the target state fusion estimate.
[0055] Interaction and state fusion among various motion models are achieved through SMDP-IMM, specifically including: Step S401: Using the dynamic probability transition matrix and the posterior probability of the previous time step, calculate the mixed initial state and mixed covariance of each motion model at the current time step; the specific expression is as follows: ; ; ; ; ; ; in, express Time of the first The posterior probability of a motion model; express The dynamic transition probability at any given time; Indicates the first A motion model in Posterior state estimation at time 1; express Time from the first The motion model is transferred to the first... Prior mixture weights for each motion model; express Time of the first The model probability predicted by each motion model; express Time of the first The mixed initial states of several motion models; express Time of the first Mixed covariance of several motion models; express Time of the first Initial values for the expected dwell time of each motion model; Represents the time transfer function; Indicates the first A motion model in Duration of stay at a given moment; Indicates the first A motion model in The posterior covariance matrix at time t.
[0056] Step S402: Based on the mixed initial state and mixed covariance, perform parallel filtering on each motion model and update the state estimate, covariance and likelihood of each motion model.
[0057] Step S403: After completing parallel filtering, calculate the dynamic probability transition matrix for the next time step using SMDP based on the updated state estimates, covariance, and likelihood of each motion model.
[0058] Step S404: Based on the updated likelihood and dynamic probability transition matrix, update the posterior probability of each motion model. The specific expression is as follows: ; in, express Time of the first The model probability predicted by each motion model; express Time of the first Likelihood of a motion model; express Time of the first Likelihood of a motion model; express Time of the first The posterior probability of each motion model.
[0059] Step S405: Based on the updated posterior probability, perform posterior correction on the dwell time, fuse the state estimates and covariances of each motion model, and obtain the final target state fusion estimate, the specific expression of which is: ; ; ; ; in, This represents a mixture of weights based on posterior probabilities; Indicates the revised dwell time; This represents the target state fusion estimate; Indicates the first The posterior update value of each motion model; This represents the covariance of the target state fusion estimate; express Time of the first The posterior probability of each motion model.
[0060] Step S5: Based on the target state fusion estimation, perform time-series updates and state estimation transfer.
[0061] The target state output at the current moment is fused and estimated. and the target state fusion estimation covariance This serves as the prior state information for the parallel filtering of each motion model in step S2 at the next moment. Steps S2 to S5 are repeated to achieve continuous and active tracking of the moving target.
[0062] To verify the effectiveness of the method of this invention, a comparative experiment was conducted in a typical two-dimensional planar simulation scenario. The experimental scenario was a constructed two-dimensional aerial tracking scenario with varying altitude maneuvers (the UAV's flight altitude remained constant). The object being tracked was a single simulated UAV moving within a planar coordinate system (XY plane). The observation data consisted of the target's position observation vector in the planar coordinate system provided by sensors. This comparative experiment compares the performance of the classical fixed-transition-probability IMM method and the method of this invention in terms of target state estimation accuracy (RMSE) and maneuver pattern recognition delay under complex maneuvers such as rapid acceleration and continuous turns. The experimental results are as follows: Figures 2-4 As shown.
[0063] like Figure 2 As shown, traditional single models (CV, CA, CTRV) exhibit significant deviations in maneuver segments where their assumptions are inconsistent. The classic IMM method improves tracking performance through multi-model fusion, but still suffers from lag and overshoot at maneuver mode switching points. In contrast, the estimated trajectory of the method in this invention has the highest degree of fit with the true trajectory, especially in the initiation and termination phases of maneuvers, where it responds more quickly to changes in motion mode and achieves a smooth transition.
[0064] like Figure 3 As shown, the classic IMM method exhibits a wider recognition delay distribution, with higher median and upper quartile values, indicating a slower and more unstable response to maneuvers. The method of this invention reduces the median recognition delay (optimized to approximately 3.1 seconds) and compresses the delay distribution range. This demonstrates that by introducing environmental context feature vectors and dwell time distribution functions, the method of this invention can more quickly lock onto the correct motion model after a maneuver occurs, achieving proactive and forward-looking model switching.
[0065] like Figure 4 As shown, the root mean square error (RMSE) distributions of position, velocity, and heading angle estimation errors of the two methods are compared throughout the tracking process. Figure 4 (a) in the figure is the location RMSE distribution map; Figure 4 (b) in the figure is the velocity RMSE distribution diagram; Figure 4 (c) in the figure shows the orientation RMSE distribution. In the position RMSE, the median error distribution (approximately 0.31) and the overall box range of the method of this invention are significantly lower than those of the classic IMM, indicating that its overall position tracking accuracy is higher.
[0066] Similarly, the method of this invention also shows an overall decreasing trend in velocity RMSE (median approximately 0.48) and heading RMSE (median approximately 2.68 degrees). This proves that the method of this invention not only improves position tracking performance, but also brings more stable and accurate estimates of dynamic states such as velocity and heading due to more timely and accurate model switching.
[0067] In summary, the experiments demonstrated the superiority of the method described in this invention. Compared to the classical IMM method, which relies on fixed transition probabilities, the method of this invention offers faster recognition speed, reduces the latency of maneuver pattern recognition, and enables rapid response to target intent. Furthermore, the method achieves lower steady-state errors in the estimation of key state variables such as position, velocity, and heading, thereby improving overall tracking accuracy.
[0068] This embodiment verifies the effectiveness of the method of the present invention in a two-dimensional plane. Since the method framework of the present invention relies on a modular kinematic model of the platform, it is also applicable to other platforms that require tracking complex maneuvering targets, such as unmanned ground vehicles and surface vessels. Further improvements can be made by extending the state space to three dimensions and enriching the model library (e.g., by introducing a height change model) to further enhance the tracking capability and physical consistency of the present invention for maneuvering targets in all space, thus meeting the active perception needs of a wider range of unmanned systems.
[0069] An active tracking system for unmanned aerial vehicles (UAVs) in complex scenarios, employing the aforementioned active tracking method for UAVs in complex scenarios, includes: The data preprocessing and model building module is used to preprocess multi-source sensor data from UAVs and build IMM model sets.
[0070] The multi-model parallel filtering and estimation module is connected to the data preprocessing and model building module. It is used to perform parallel filtering based on the preprocessed data and the IMM model set, calculate the likelihood of each motion model, and perform weighted fusion based on the likelihood to obtain the fused state estimate and posterior probability.
[0071] The dynamic probability transition matrix generation module, connected to the multi-model parallel filtering and estimation module, is used to generate a dynamic probability transition matrix based on preprocessed data and posterior probabilities.
[0072] The adaptive interactive fusion module is connected to the dynamic probability transition matrix generation module and the multi-model parallel filtering and estimation module, respectively. It is used to perform interactive and state fusion on each motion model based on the dynamic probability transition matrix and output the target state fusion estimate.
[0073] Therefore, the present invention adopts the above-mentioned active tracking method and system for UAVs in complex scenarios, which solves the problems of decreased tracking accuracy, response delay and easy divergence caused by the lag in model switching and the inability of fixed transfer structure to adapt when there are complex interferences such as sudden changes in target motion mode, degradation of observation signal or occlusion. It achieves rapid and accurate identification and smooth switching of target, and enhances tracking stability and adaptability.
[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for active tracking of unmanned aerial vehicles (UAVs) in complex scenarios, characterized in that, Includes the following steps: Step S1: Data preprocessing and construction of the IMM model set; Step S2: Based on the preprocessed data and the IMM model set, perform parallel filtering to obtain the fused state estimate and posterior probability; Step S3: Generate a dynamic probability transition matrix based on the preprocessed data and posterior probabilities; Step S4: Based on the dynamic probability transition matrix, perform interaction and state fusion on each motion model, and output the target state fusion estimate; Step S5: Based on the target state fusion estimation, perform time-series updates and state estimation transfer.
2. The active tracking method for UAVs in complex scenarios according to claim 1, characterized in that, In step S1, the constructed IMM model set includes CV model, CA model and CTRV model; Define a unified state vector for each motion model in the IMM model set, and establish the state equation and observation equation for each motion model based on a preset sampling time interval.
3. The active tracking method for UAVs in complex scenarios according to claim 1, characterized in that, In step S2, parallel filtering is implemented through the IMM interactive multi-model approach, specifically including: Step S201: Based on the state estimates, covariances, and preset transition probabilities of each motion model at the previous moment, calculate the initial state and initial covariance of each motion model at the current moment. Step S202: Based on the initial state and initial covariance, perform parallel filtering on each motion model and update the state estimate and covariance; Step S203: Calculate the likelihood of each model using the likelihood function, and update the posterior probability of each model at the current time by combining the model prediction probability. Step S204: Weigh and fuse the state estimates, covariances and posterior probabilities of each model to obtain the fused state estimates and fused covariance estimates.
4. The active tracking method for UAVs in complex scenarios according to claim 3, characterized in that, In step S202, the CV model is configured with standard KF for parallel filtering, and the CA and CTRV models are configured with EKF for parallel filtering.
5. The active tracking method for UAVs in complex scenarios according to claim 1, characterized in that, In step S3, the dynamic probability transition matrix is generated using SMDP, specifically including: Define system state The specific expression is: ； in, Represents the state index of the motion model; These represent three motion models. express Model, express Model, express Model; Represent the target state space; Indicates the duration of stay. ; Represents the target state vector; definition The environmental context feature vector at any given time , The specific expression is: ； in, Indicates the angular velocity of the heading; Indicates tangential angular velocity; Indicates the curvature of motion; Indicates the change in velocity; This represents the trace of the fused covariance estimate; Indicates the confidence level of measurement consistency; Indicates transpose; Define action space The specific expression is: ； in, ; This indicates that the current position is maintained after the action is executed. A motion model, and a sampling time interval. The duration of stay is accumulated. , express The duration of dwell time express Duration of stay at a given moment; Indicates that after the action is executed, from the current number... The motion model is switched to the first one. Each motion model has its dwell time reset. Switch to the current Effective immediately; Based on environmental context feature vectors and dwell time, a context-aware Weibull distribution is used to define the model dwell time distribution function. The specific expression is: ； ； in, Indicates the context scaling factor; Indicates the baseline characteristic time; Represents the dwell bias vector; Represents the natural exponential function; Introduce upper and lower bounds for dwell time [ ], This represents the minimum number of dwell steps. The time window penalty factor represents the maximum number of dwell times. The specific expression is: ； in, The non-negative truncation function is defined as follows: ; , Indicates the penalty coefficient; Based on environmental context feature vectors and time window penalty factors, the model's departure tendency quantification index is calculated. The specific expression is: ； in, Represents the context weight vector; This indicates that the context is feature-driven. Based on the model's residence time distribution function Quantitative indicators of the tendency to leave the model Calculate the probability of residence The specific expression is: ； in, The logistic function is defined as follows: ; Indicates the residency bias coefficient; A model switching strategy is introduced by combining model likelihood and environmental context feature vectors. Calculate the switching probability The specific expression is: ； ； in, This represents the minimum probability threshold. ; Indicates the first Contextual preference vectors for each motion model; This represents the weighted sum of the likelihoods of all candidate models and the context preference vectors; express Time of the first The likelihood of a motion model ( ); Indicates the model index; Indicates the first Contextual preference vectors for each motion model; Based on the probability of residence As a diagonal element, the switching probability Construct a dynamic probability transition matrix with off-diagonal elements. The specific expression is: 。 6. The active tracking method for UAVs in complex scenarios according to claim 5, characterized in that, The specific action selection strategy in the action space is as follows: If the likelihood value of the current motion model is higher than the preset threshold and the dwell time is less than the preset maximum dwell time, then select... If the likelihood value of the current motion model is lower than a preset threshold, or the dwell time exceeds the maximum dwell time, then select... .
7. The active tracking method for UAVs in complex scenarios according to claim 5, characterized in that, In step S3, a reward function is defined to optimize the dynamic probability transition matrix. The specific expression is as follows: ； ； ； ； ； ； in, Indicates total return; Indicates the discount factor; express Instant rewards at any moment; , and Indicates parameter weights; This indicates a measurement adaptation feedback; This represents the IMM filter information; Indicates the new information covariance; This indicates that the length of stay is appropriate for the return. Indicates the optimal dwell time; Indicates the adaptation threshold; This indicates a context-adaptive reward; This represents the characteristic vector of measurement error.
8. The active tracking method for UAVs in complex scenarios according to claim 7, characterized in that, In step S4, the interaction and state fusion of the various motion models are achieved through SMDP-IMM, specifically including: Step S401: Using the dynamic probability transition matrix and the posterior probability of the previous time step, calculate the mixed initial state and mixed covariance of each motion model at the current time step. Step S402: Based on the mixed initial state and mixed covariance, perform parallel filtering on each motion model and update the state estimate, covariance and likelihood of each motion model; Step S403: After completing parallel filtering, calculate the dynamic probability transition matrix for the next time step using SMDP based on the updated state estimates, covariance, and likelihood of each motion model. Step S404: Based on the updated likelihood and dynamic probability transition matrix, update the posterior probability of each motion model. The specific expression is as follows: ； in, express Time of the first The model probability predicted by each motion model; express Time of the first Likelihood of a motion model; express Time of the first Likelihood of a motion model; express Time of the first The posterior probability of a motion model; Step S405: Based on the updated posterior probability, perform posterior correction on the dwell time, fuse the state estimates and covariances of each motion model, and obtain the final target state fusion estimate, the specific expression of which is: ； ； ； ； in, This represents a mixture of weights based on posterior probabilities; Indicates the revised dwell time; This represents the target state fusion estimate; Indicates the first The posterior update value of each motion model; This represents the covariance of the target state fusion estimate; express Time of the first The posterior probability of each motion model.
9. An active tracking system for unmanned aerial vehicles (UAVs) in complex scenarios, comprising the method described in any one of claims 1 to 8, characterized in that, include: The data preprocessing and model building module is used to preprocess multi-source sensor data from UAVs and build an IMM model set. The multi-model parallel filtering and estimation module is used to perform parallel filtering based on preprocessed data and the IMM model set, calculate the likelihood of each motion model, and perform weighted fusion based on the likelihood to obtain the fused state estimate and posterior probability. The dynamic probability transition matrix generation module is used to generate a dynamic probability transition matrix based on the preprocessed data and the posterior probability. The adaptive interactive fusion module is used to perform interactive and state fusion on various motion models based on the dynamic probability transition matrix, and output the target state fusion estimate.