Low-altitude unmanned aerial vehicle intelligent tracking and monitoring method and system

Abstract

The application discloses a low-altitude unmanned aerial vehicle intelligent tracking and monitoring method and system, and relates to the technical field of unmanned aerial vehicle tracking.The method collects target data through a multi-source sensor and unifies a space-time reference, generates fused flight tracks and identity information by adopting a hierarchical fusion architecture, drives a photoelectric turntable to track the target based on a LESO-GPC algorithm, detects a shielding state and switches three modes of normal, prediction and search to realize anti-shielding continuous tracking, and feeds back photoelectric miss distance to a fusion center to correct parameters.The application improves the continuity, precision and anti-interference capability of unmanned aerial vehicle tracking.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) tracking technology, and in particular to a method and system for intelligent tracking and monitoring of low-altitude UAVs. Background Technology

[0002] Monitoring and tracking low-altitude unmanned aerial vehicles (UAVs), especially "low, slow, and small" targets, is a crucial issue in low-altitude security. Existing technologies mainly include single radar detection, radio spectrum detection, protocol parsing and remote identification (RID), and electro-optical tracking. Radar can provide three-dimensional position and velocity of targets in all weather conditions, but it has blind spots for low-altitude, low-speed, and small radar cross-section targets, is susceptible to ground clutter and multipath interference, has a high false alarm rate, and cannot distinguish between UAVs and birds. Radio spectrum detection achieves direction finding through passive detection of communication links, offering advantages such as good concealment and low cost, but it cannot detect silently flying targets, has low positioning accuracy, and typically only provides directional information. Protocol parsing and RID technology can directly obtain UAV identity, position, and other status information, providing rich information, but it heavily relies on UAV active broadcasts and is completely ineffective against targets that do not broadcast or broadcast false information. Electro-optical tracking has advantages such as intuitive visualization, strong evidence collection capabilities, and high tracking accuracy, but its effective range is limited, it is easily affected by weather, lighting, and obstructions, and it is difficult to quickly detect and lock onto targets in complex environments.

[0003] Some existing solutions attempt to simply combine the aforementioned devices, such as the "radar-guided electro-optical" linkage scheme. However, such schemes are mostly loosely coupled, with each sensor operating independently and information not being fully shared, resulting in a high target handover failure rate; they lack anti-obstruction mechanisms, and when a target flies over a tower or tree, the electro-optical equipment is prone to losing the target, leading to poor tracking continuity; their multi-target processing capabilities are weak, and they cannot intelligently allocate tracking resources; furthermore, they cannot associate RID information with radar or electro-optical targets, making it difficult to effectively identify illegal UAVs. Therefore, there is an urgent need for a low-altitude UAV intelligent tracking and monitoring method that can deeply integrate multi-source heterogeneous sensor data, possess adaptive anti-obstruction capabilities, and intelligent task allocation capabilities. Summary of the Invention

[0004] To address the above problems, this invention provides a method and system for intelligent tracking and monitoring of low-altitude unmanned aerial vehicles (UAVs). Through deep fusion of multi-source heterogeneous data, anti-obstruction predictive tracking, and intelligent task allocation, it achieves continuous tracking and accurate identification of low-altitude UAVs.

[0005] In a first aspect, the present invention provides a method for intelligent tracking and monitoring of low-altitude unmanned aerial vehicles (UAVs), comprising: S1, collect raw data of the target based on multi-source sensors and perform preprocessing; S2, which unifies the data from different sensors in their respective time and space coordinate systems to the same spatiotemporal reference; S3 employs a layered fusion architecture to associate and fuse registered multi-source data, generating fused tracks and identity information; S4 assigns tracking tasks to the photoelectric tracking device based on fused track and identity information, and uses a generalized predictive control algorithm based on a linear extended state observer to generate servo control commands to drive the photoelectric turntable to track the target; S5 detects the occlusion status of the target in the field of view of the optoelectronic device, and uses a three-mode switching strategy to continuously track the target against occlusion based on the detection results; S6 feeds back the real-time image processing results from the photoelectric tracking device as high-precision observations to the fusion center to correct fusion parameters or calibrate the system errors of other sensors.

[0006] Furthermore, the multi-source sensors include radar detection equipment, radio spectrum detection equipment, RID resolution equipment, and photoelectric tracking equipment; the preprocessing includes filtering, noise reduction, and formatting.

[0007] By fusing four types of heterogeneous sensors—radar, radio spectrum, RID, and photoelectric—and leveraging their complementary characteristics (radar ranging and velocity measurement, passive spectrum detection, RID identity broadcasting, and photoelectric visualization tracking) to overcome the inherent limitations of single sensors, the filtering, noise reduction, and formatting processes in the preprocessing effectively improve the quality of the raw data, providing reliable data for subsequent high-precision fusion.

[0008] Furthermore, the layered fusion architecture specifically includes: An improved joint probabilistic data association algorithm is used to associate radar tracks, spectrum direction finding intersections, RID broadcast locations, and established target tracks. An improved adaptive unscented Kalman filter algorithm, combined with dynamic confidence weights, is used to make optimal estimates of the target position and velocity. Based on the Dempster-Shafer evidence theory, RID parsing information and spectrum analysis information are fused at the decision level to comprehensively determine the target's identity and threat level.

[0009] An improved JPDA algorithm is used to accurately correlate radar spots, virtual spectrum points, RID points, and flight tracks, solving the problem of measurement ambiguity in multi-target environments. An improved adaptive unscented Kalman filter (IAUKF) combined with dynamic confidence weights is used to achieve optimal estimation of target position and velocity, with filtering accuracy higher than traditional methods. Based on DS evidence theory, RID and spectrum information are fused to comprehensively determine target identity and threat level, compensating for the unreliability of a single identity source.

[0010] Furthermore, the dynamic confidence weight is calculated through a dynamic confidence evaluation mechanism, which integrates three dimensions: signal-to-noise ratio factor, data residual factor, and environmental parameter factor, to evaluate the confidence of each sensor in real time and dynamically adjust the weight of each sensor data in state fusion based on this.

[0011] Sensor reliability is evaluated in real time by comprehensively considering three dimensions: signal-to-noise ratio factor, data residual factor, and environmental parameter factor. The fusion weights are dynamically adjusted to enable the system to adapt to different environmental changes such as weather, lighting, and electromagnetic interference. For example, the photoelectric weight is increased on sunny days, the radar weight is increased on foggy days, and the RID weight is increased under strong clutter, thus maintaining optimal fusion performance and improving the system's robustness and adaptability.

[0012] Furthermore, an improved Bach coefficient template matching method is used to detect occlusion status in real time, specifically: Add a background weighting factor to the initial template; The target area is divided into multiple sub-blocks, and the Bach coefficient between the color histogram of each sub-block and the corresponding sub-block of the initial template is calculated. The proportion of sub-blocks with a Bach coefficient lower than a preset threshold is used as the overall occlusion level, and the current occlusion status is determined based on the overall occlusion level.

[0013] Adding a background weighting factor to the initial template effectively suppresses the interference of the background region on the matching. The target region is divided into blocks to calculate the Bach coefficient, and the proportion of sub-blocks below the threshold is used as the overall occlusion degree. Compared with the traditional global matching method, it can detect partial occlusion and complete occlusion more accurately, reduce false detections caused by target pose changes or background clutter, and provide a reliable basis for trimodal switching.

[0014] Furthermore, the tri-modal switching strategy includes: Normal tracking mode: When the occlusion level is less than the first threshold, LESO-GPC closed-loop control is adopted, and the feedback quantity is the image miss distance. Predictive tracking mode: When the occlusion level is greater than or equal to the first threshold and less than the second threshold, the target position is extrapolated using a Kalman predictor based on the fused track, and the photoelectric turntable is driven to perform open-loop predictive tracking; Intelligent search mode: When the degree of occlusion is greater than or equal to the second threshold and the duration exceeds the time threshold, the scanning path is planned based on the last reliable position of the fused track and the target motion model to search for the target.

[0015] The three-mode switching strategy enables continuous tracking in occluded scenarios: the normal mode utilizes LESO-GPC high-precision closed-loop control; the predictive mode uses the Kalman predictor extrapolating the position based on the fused track to perform "blind tracking," ensuring uninterrupted tracking when the target passes through obstructions such as towers and trees; and the intelligent search mode quickly reacquires the target after prolonged occlusion through spiral scanning and fusion guidance.

[0016] Furthermore, a multi-stage auction mechanism is adopted to allocate tracking tasks to photoelectric tracking equipment.

[0017] A multi-stage auction mechanism is employed to allocate photoelectric tracking tasks, dynamically allocating time slices based on the target threat level. This ensures that high-threat targets receive continuous tracking resources, while low-threat targets receive periodic scanning resources. This mechanism intelligently manages limited photoelectric resources in multi-target scenarios, improving the utilization rate of measured resources while ensuring that key targets are not lost.

[0018] Furthermore, the improved joint probabilistic data association algorithm includes: At least two direction finding lines obtained by radio spectrum detection equipment are spatially intersected, the shortest distance point is calculated as a virtual positioning point, and the equivalent covariance matrix of the virtual positioning point is derived by using the error propagation law. Based on the predicted state of each target track, an ellipsoidal threshold is used to filter all measurements and construct a confirmation matrix indicating the possible correspondence between the measurements and the target track. The acknowledgment matrix is ​​decomposed into multiple independent submatrices using sparse matrix factorization. The joint correlation probability of each submatrix is ​​calculated separately, and then the results are combined.

[0019] The spectrum direction finding line intersections are converted into virtual positioning points with error covariance, enabling spectrum data to participate in JPDA correlation in a unified manner with radar and RID data; an ellipsoidal threshold is used to construct the confirmation matrix, effectively filtering out impossible measurement-track combinations; and sparse matrix factorization technology is used to decompose the large matrix into independent submatrices and solve the joint correlation probability separately, reducing the computational complexity from exponential to polynomial level, meeting the millisecond-level processing requirements of real-time tracking.

[0020] Furthermore, the decision-level fusion based on Dempster-Shafer evidence theory also includes: Calculate the conflict factor K between the basic probability assignments of the two sources of evidence. rt ; When K rt If the conflict exceeds the preset conflict threshold, it is determined that there is a serious conflict between the RID information and the spectrum information. The threat level of the target is then raised by one level, the tracking priority of radar and optoelectronic equipment is increased, and an alarm is triggered to prompt manual intervention.

[0021] By calculating the conflict factor K between two sources of evidence in the DS evidence theory rt When K rt When a preset threshold is exceeded, electronic deception or information conflict is identified, immediately raising the target's threat level, increasing radar and electro-optical tracking priority, and triggering an alarm. This mechanism effectively identifies deception attacks by malicious drones using forged RIDs or suppressed spectrum signals, enhancing system security and providing timely early warnings for manual intervention.

[0022] Secondly, the present invention provides a low-altitude unmanned aerial vehicle (UAV) intelligent tracking and monitoring system, comprising: The multi-source sensor module, deployed on the communication tower, includes radar detection equipment, radio spectrum detection equipment, RID analysis equipment, and photoelectric tracking equipment, used to collect raw data of the target in real time; The spatiotemporal registration module is used to uniformly register data from various sensors to the same spatiotemporal reference. The data association and fusion module is used to perform data layer association, feature layer state estimation and decision layer identity fusion on the registered multi-source data using a hierarchical fusion architecture, and generate fused tracks and identity information. The servo tracking control module assigns tracking tasks to the photoelectric tracking device based on fused track and identity information, and generates servo control commands using a generalized predictive control algorithm based on a linear extended state observer to drive the photoelectric turntable to track the target. The anti-occlusion prediction tracking module is used to detect the occlusion status of targets in the field of view of the optoelectronic device and switch between normal tracking mode, prediction tracking mode and intelligent search mode according to the degree of occlusion. The feedback self-optimization module is used to feed back the real-time image processing results of the photoelectric tracking device as high-precision observation values ​​to the data association and fusion module in order to correct the fusion parameters or calibrate the system errors of other sensors.

[0023] Compared with existing technologies, the beneficial effects of this invention are as follows: by preprocessing multi-source sensor data and unifying spatiotemporal registration, combined with a hierarchical fusion architecture to generate fused tracks and identity information, the continuity of target tracking and the reliability of identity recognition are improved; the use of a generalized predictive control algorithm based on a linear extended state observer to drive the photoelectric turntable improves the accuracy and anti-interference capability of servo tracking; the introduction of occlusion state detection and a three-mode switching strategy enables continuous tracking of targets when they pass through occlusions; at the same time, the photoelectric tracking image processing results are fed back to the fusion center to continuously correct fusion parameters or calibrate other sensor errors, forming a closed-loop self-optimization, thereby significantly enhancing the stability, adaptability, and system accuracy of low-altitude UAV tracking. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this drawing or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this drawing. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0025] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a system overall framework diagram of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0027] This invention provides a method for intelligent tracking and monitoring of low-altitude unmanned aerial vehicles (UAVs), such as... Figure 1 , 2 Specifically, the process includes the following steps S1 to S6. The following example, using a multi-UAV surveillance scenario within the coverage area of ​​a communication tower, illustrates the implementation process of this invention in detail.

[0028] In this embodiment, three drones entered the monitoring area. The first was a registered friendly drone (actively broadcasting RID information), the second was a commercial drone (not broadcasting RID but with a clear spectrum signal), and the third was a self-made drone (flying silently, without RID broadcasting or spectrum signal). Radar detection equipment, radio spectrum detection equipment, RID parsing equipment, and photoelectric tracking equipment were all deployed on the communication tower.

[0029] S1, collects raw data of the target based on multi-source sensors and performs preprocessing.

[0030] The system first activates all sensors: the radar detection device outputs three-dimensional position and velocity data of three targets at a refresh rate of 10Hz; the radio spectrum detection device detects two signal sources (corresponding to friendly drones and commercial drones) at a refresh rate of 5Hz and outputs direction finding angles; the RID resolution device receives broadcast information (including ID, position, velocity, altitude, etc.) from the friendly drone at a refresh rate of 1Hz; and the electro-optical tracking device is initially in standby scanning mode.

[0031] The raw data were filtered, denoised, and formatted to remove obvious abnormal data, and the data from each sensor were uniformly converted into a preset data structure.

[0032] S2 registers data from different sensors in their respective time and space coordinate systems to the same spatiotemporal reference.

[0033] Specifically, the GPS / BeiDou high-precision timing module is used to timestamp each frame of data to achieve time synchronization with an accuracy better than 1ms. During spatial registration, the polar coordinate data output by the radar, the direction finding equations output by the spectrum detector, the latitude and longitude coordinates output by RID analysis, and the field of view azimuth parameters of the optoelectronic equipment are uniformly converted to a northeast-sky rectangular coordinate system centered on the iron tower.

[0034] S3 employs a layered fusion architecture to associate and fuse registered multi-source data, generating fused tracks and identity information.

[0035] The layered fusion architecture specifically includes three sub-steps: data layer association, feature layer state estimation, and decision layer identity fusion. S31, Data Layer Association: Using an improved joint probabilistic data association algorithm, radar tracks, spectrum direction finding cross-location points, RID broadcast location points, and established target tracks are associated.

[0036] For spectrum detection equipment, when multiple spectrum detection nodes exist (or a single node moves at different times), virtual positioning points are obtained through the intersection of direction-finding lines, enabling spectrum data to be correlated with radar and RID data. Two direction-finding lines are used. and The equations of the spatial lines are as follows: p 1 =a 1 +b 1 d 1; p 2 =a 2 +b 2 d 2; in, p 1. p 2 are the position vectors on the direction finding lines L1 and L2, respectively; a 1. a 2 represents the coordinate vector of the spatial reference point on the direction finding line; d1 and d2 are unit direction vectors; , It is the azimuth angle; , The pitch angle; b 1. b 2 is a dimensionless parameter, representing the proportion of the distance moved from the reference point along a unit direction vector.

[0037] Find the point of shortest distance between two straight lines as the virtual intersection point: ; in, The vector of intersecting virtual positioning points; This is the dimensionless parameter corresponding to the shortest distance between the two measurement lines.

[0038] The equivalent covariance matrix of the virtual location point is derived using the error propagation law. : ; Where J is the Jacobian matrix, J T Let J be the transpose of J; , Azimuth angles and pitch angle The measurement error variance; diagnosis It is a diagonal function.

[0039] Suppose there are m measurements at the current time. z j (j=1,2,…,m), there are already n target tracks. v i (i=1,2,…,n), based on the predicted state of each target trajectory, an ellipsoidal threshold is used to filter all measurements, and a confirmation matrix Ω=[ oh ji ]: ; in, oh ji Used to indicate whether there is a possible correlation between the measurement and the target; The ellipsoidal threshold is as follows: ; in, Let be the predicted measurement value of the i-th target at time k; Let be the information covariance matrix of the i-th target at time k; γ is the ellipsoidal threshold, which is usually taken as 9.21.

[0040] Based on the filtered measurements, define joint related events. The probability of joint related events for: ; in, This indicates that the j-th measurement originates from the j-th measurement. t j An event targeting a specific goal; For m Find the intersection, representing all individual events. Simultaneous occurrence; Let be the detection probability of the i-th target; A variable indicating whether the i-th target has been detected; c is the normalization constant; It is the probability density function; For the first t j The predicted measurement value of a target at time k; For the first t j The information covariance matrix of each objective at time k; The set of all measurements up to time k; This indicates that under the joint associated event θ, all events assigned to the true target (i.e. t j The probability of a measurement occurring (>0); This represents the probability that each target will be detected under the joint associated event θ.

[0041] Then measurement z j With the target track v i Association probability for: ; in, As an indicator function, in event θ, if the measurement z j Assigned to the target track v i ,but The value is 1 if it is 1, otherwise it is 0.

[0042] To address the problem of calculating combinatorial explosions in traditional JPDA under dense clutter, this invention introduces sparse matrix factorization (SMF) technology. The confirmation matrix is ​​decomposed into multiple independent submatrices, and the joint correlation probabilities are solved separately for each submatrix. The results are then combined, reducing the computational complexity to O(2^3). mn Reduced to , m i , n i Wherein are the number of measurements and the number of targets corresponding to each submatrix after decomposition, respectively.

[0043] In this embodiment, three associations were successfully established: the friendly drone was associated with radar target 1, spectrum source 1, and RID data; the commercial drone was associated with radar target 2 and spectrum source 2; and the self-made drone was associated only with radar target 3.

[0044] S32, Feature layer state estimation: An improved adaptive unscented Kalman filter algorithm is used, combined with dynamic confidence weights, to make optimal estimates of the target position and velocity.

[0045] The system first calculates the dynamic confidence weight of each sensor, which is determined by the following three factors: Signal-to-noise ratio factor α SNR ( i ): ; in, For sensors i Maximum signal-to-noise ratio; The standard deviation of the signal-to-noise ratio; For sensors i exist k The actual signal-to-noise ratio at any given moment.

[0046] Data residual factor β res ( i ): ; in, For sensors i exist k The actual measured value at that moment; For sensors i exist k Predicted measurement value at time; It is an L2 norm; The new covariance matrix S kThe trace represents the uncertainty of prediction.

[0047] Environmental parameter factors c env ( i ): Obtained by looking up a preset environmental parameter table, for example, a photoelectric sensor: c env (Photoelectric) = f (Light intensity, visibility, rainfall).

[0048] Therefore, the overall weight for: ; Where M is the number of sensors; c env ( i ) represents the environmental parameter factor of sensor i; α SNR ( j ), β res ( j ), c env ( j ) are the signal-to-noise ratio factor, data residual factor, and environmental parameter factor of sensor j, respectively.

[0049] Then, an improved adaptive unscented Kalman filter (IAUKF) is used: Sigma points are generated through UT transformation and state prediction is propagated; the measurement noise covariance of each sensor is estimated online using a Sage-Husa adaptive filter; the measurements of each sensor are equivalently fused according to dynamic weights to obtain equivalent measurements and equivalent covariance; finally, Kalman update is performed to obtain state estimates and covariance.

[0050] Specifically, a target motion model is established, using a uniform acceleration (CA) model or the current statistical model, with the state vector defined as: X=[ x , v x , a x , y , v y , a y , z , v z , a z ] T ; The state equation is: X k =F k 1 X k 1 +w k 1 ,w k 1 N (0, Q k 1); The measurement equation is as follows (taking radar as an example): z radar,k =H radar X k +v radar,k, v radar,k N(0, R radar ); Where (x,y,z) are the coordinates of the target; v x , v y , v z These represent the target's velocities in three directions; a x , a y , a z These represent the target's acceleration in three directions; X k , X k 1 represents the complete state vector of the target at time k and time k-1, respectively; F k 1 represents the state transition matrix; w k 1 represents the process noise vector; Q k 1 represents the process noise covariance matrix; z radar,k Let be the radar measurement vector at time k; H radar For measurement matrix; v radar,k This is the radar measurement noise vector; R radar This is the radar measurement noise covariance matrix.

[0051] The Sigma points generated by the UT transformation are as follows: ; ; ; in, Let be the dimension of the state vector; For scale parameters; For the first l One Sigma point; This is the posterior state estimate at time k-1; Let be the posterior covariance matrix at time k-1. Its corresponding weights are: ; ; in, for l The weights of each Sigma point.

[0052] Propagating the Sigma point through the equation of state: ; Calculate state prediction and prediction covariance: ; ; in, The Sigma point after propagation; To control the input matrix; To control the input vector; For prior state estimation; Let be the prior covariance matrix.

[0053] The measurement noise covariance is estimated online using a Sage-Husa filter. ; in, Let be the measurement noise covariance estimated by sensor i at time k; Let be the innovation vector, representing the sensor. i exist k Actual measurement value at time Compared with predicted measurement values The differences between them; Forgetting factor, The base value for the forgetting factor is 0.95 to 0.99.

[0054] Then calculate the weighted fused equivalent measurement vector. and equivalent covariance matrix : ; .

[0055] Finally, the state estimate and covariance are obtained through the Kalman update equation: ; ; in, Here is the Kalman gain matrix. for The transpose of the matrix; The equivalent new information covariance matrix; To predict equivalent measurements; The posterior state estimate at time k; Let be the posterior covariance matrix at time k.

[0056] In this embodiment, taking a friendly UAV as an example, the current weather is good, the radar signal-to-noise ratio is high, the residual is small, the RID data is accurate, and the spectrum direction finding error is large. The calculated radar weight is 0.35, the spectrum weight is 0.15, and the RID weight is 0.50.

[0057] S33, Decision-level Identity Fusion: Based on the Dempster-Shafer evidence theory, RID parsing information and spectrum analysis information are fused at the decision level to comprehensively determine the target's identity and threat level.

[0058] Define the identification framework Θ={F,S,H,U}, where: F (friendly) represents a cooperative drone; S (suspicious) represents a drone of unknown identity but without threatening behavior; H (hostile) represents a malicious drone; and U (unknown) represents a drone that cannot be determined.

[0059] Construct a basic probability assignment based on RID information: If the RID is valid and the ID is in the whitelist, then m RID (F)=0.8, m RID (U)=0.2; if RID information is missing or verification fails, then m RID (U) = 1.0, where m RID This is the basic probability assignment function obtained from RID information.

[0060] Construct another set of probability assignments based on the spectral feature matching degree ρ∈[0,1]: If a known malicious drone signal characteristic is matched, then m Spec (H) = 0.7ρ, m Spec (U)=1 0.7ρ; If a match is found to be made with known commercial drone signal characteristics, then m Spec (F)=0.6ρ,m Spec (U)=1 0.6ρ; If no match can be found, then m Spec (U)=1.0; Where, m Spec This is the basic probability allocation function obtained from the spectral information.

[0061] Calculate the probability of the fusion of the two probability assignments using the Dempster synthesis rule: ; Among them, the conflict factor K rt for: ; A, B, and C are focal elements; The basic probability assignment (confidence level) for focal element A.

[0062] If the conflict factor K rt Exceeding threshold K th If the value is 0.3 to 0.5, it is determined that there is a serious conflict between the RID information and the spectrum information, which may indicate electronic deception. In this case, the target threat level is raised by one level, the tracking priority of radar and electro-optical devices is increased, an alarm is triggered, and manual intervention is requested.

[0063] In this embodiment, the friendly drone: RID provides clear identity information, spectrum feature matching, and after DS fusion, m(F)=0.95, and it is determined to be friendly and low threat; Commercial drones: No RID, spectrum feature matching commercial database, after DS fusion m(S)=0.8, judged as suspicious, medium threat; Homemade drone: No RID, no spectrum signal, m(U)=0.9 after DS fusion, judged as unknown and high threat.

[0064] S4 assigns tracking tasks to the photoelectric tracking device based on fused track and identity information, and uses a generalized predictive control algorithm based on a linear extended state observer to generate servo control commands to drive the photoelectric turntable to track the target.

[0065] Since the optoelectronic device can only continuously track one target at a time, time slices need to be allocated to multiple targets. The system conducts a multi-stage auction based on threat level (homemade drones are the highest, commercial drones are medium, and friendly drones are the lowest): high-threat targets receive higher bids, and the optoelectronic device prioritizes tracking the homemade drone (continuous tracking); commercial drones are quickly scanned every 10 seconds to update their positions; and friendly drones are periodically scanned when idle. The allocation result determines which target the optoelectronic turntable should be pointing at at the current moment.

[0066] Once a tracking target is assigned, the system extracts the target's position from the fused track and calculates the azimuth and pitch commands required by the photoelectric turntable. Traditional PID control struggles to overcome turntable friction, unbalanced torque, and image processing lag. This invention employs Generalized Predictive Control (LESO-GPC) based on a Linear Extended State Observer: the extended state observer estimates and compensates for the total disturbance in real time, while the generalized predictive controller optimizes the control input in the prediction time domain, adaptively adjusting the prediction time domain based on the tracking residual.

[0067] Specifically, the azimuth / pitch axis dynamics model of the photoelectric turntable is simplified into a second-order system: ; Where H is the moment of inertia; G is the damping coefficient; τ d For disturbance torque; D t This is the torque coefficient; u is the control input; This refers to the rotation angle (azimuth / pitch) of the photoelectric turntable. Angular velocity, ω is angular acceleration.

[0068] Expand the total internal and external disturbances of the system into new state variables. x 3=f Establish the third-order extended state equation: ; in, For the corner of the photoelectric turntable, Angular velocity; x 3. f For the total disturbance, for f The derivative; , , They are respectively x 1. x 2. x The derivative of 3; b 0 represents the estimated parameter for the control input. .

[0069] Design a linearly extended state observer: ; in, , , They are respectively x 1. x 2. x The observed value of 3; y represents the actual system output; , , They are respectively , , The derivative; To achieve gain with the observer band, , This represents the observer bandwidth.

[0070] Design a generalized predictive control law: Based on a continuous-time model, use Taylor expansion to predict future outputs. Define the prediction time domain as T. p The control time domain is T c The objective function is: ; in, For predicting output, i.e., the future The reference trajectory of the step; , These are the control weighting coefficients in the GPC prediction model; For the nominal value of control gain N1 and N2 are the start and end points of the prediction time domain, respectively; This is the control input at time k; For the future Step control increment; The total perturbation estimated by the extended state observer at time k; To control the length of the time domain.

[0071] The reference trajectory is generated by the tracking differentiator: ; in, Let be the expected turning angle at time k; Let be the desired angular velocity at time k.

[0072] By minimizing the objective function, the analytical form of the control law is obtained: ; in, For the prediction matrix; I is the identity matrix; Y r The reference trajectory vector; To control the weighting coefficients; F is the free response coefficient matrix; The state vector estimated for the extended state observer.

[0073] The calculated control quantity Convert to an analog quantity that the driver can recognize: ; in, This is the analog voltage output to the servo driver; This represents the PWM modulation depth coefficient. This refers to the bus voltage. This represents the number of bits in the DA converter.

[0074] Prediction time-domain adaptive adjustment: based on tracking residuals Gradient descent method is used to adjust the prediction time domain T. p : ; in, This is the adjusted prediction time domain; This represents the prediction time domain for the current moment. For reference trajectory in The value at time; For the output of the prediction model The predicted value at any given time; The learning rate; This is the observed angle of rotation of the photoelectric turntable at the current moment.

[0075] Control commands drive the photoelectric turntable to move, ensuring the target remains centered in the image. The miss distance (the deviation between the target center and the image center) obtained from image processing is recorded in real time and used for closed-loop control as well as as a high-precision observation feedback to the fusion center.

[0076] S5 detects the occlusion status of the target in the field of view of the optoelectronic device, and uses a three-mode switching strategy to continuously track the target against occlusion based on the detection results.

[0077] During photoelectric tracking, the target may fly over obstructions such as towers or trees, causing it to disappear from the image. This step detects the occlusion and switches the control mode to ensure uninterrupted tracking.

[0078] A background weighting factor is added to the initial template to suppress background interference. The histogram of the initial template after adding the background weighting factor is shown. ,in This is a color histogram of the background area. To prevent small constants from being divided by zero, The initial template histogram before adding the background weighting factor.

[0079] The target area is divided into 4×4 sub-blocks, and the Baumkal coefficient between the color histogram of each sub-block and the corresponding sub-block of the initial template is calculated. Where E is the total number of intervals in the histogram. The Bach distance of the sub-block in the i-th row and j-th column in the current frame. This is the grayscale / color histogram of the target sub-block in the current frame. This is the initial template histogram for the corresponding sub-block.

[0080] Define the sub-block occlusion state occ ij (t): If ρ ij (t)<ρ block , ρ block If the preset sub-block occlusion threshold is used, then occupancy... ij (t)=1; otherwise, occ ij(t)=0; Overall degree of obstruction for: ; Determine the current occlusion status based on the overall degree of occlusion: Normal state: occ total (t)<0.3; Partial occlusion: 0.3≤occ total (t)<0.7; Completely occluded: occ total (t)≥0.7.

[0081] The three-mode switching strategy specifically includes: Normal tracking mode: Triggered under normal conditions, using LESO-GPC closed-loop control, the feedback quantity is the image miss distance, and this feedback quantity is fed back to the fusion center as a high-precision observation value to correct the filter parameters.

[0082] Predictive tracking mode: The system automatically switches when partial occlusion is detected. In this mode, it no longer relies on image feedback, but instead extrapolates the target position using a Kalman predictor based on the fused trajectory output from S3. ( The target state vector is extrapolated. The current state estimation vector, F k Here is the state transition matrix. B k To control the input matrix, To control the input vector, drive the turntable to perform open-loop predictive tracking.

[0083] Intelligent search modality: When complete occlusion is detected and the duration exceeds the time threshold T lost The system plans a spiral scanning path based on the last credible position of the fused track and the target motion model. At the same time, the fusion center pushes the suspected target points detected by radar / spectrum to the optoelectronic equipment in real time to guide the field of view to be quickly reacquired.

[0084] In this embodiment, the self-made drone was obstructed when flying over the tower, occ total As the t gradually rises to 0.45 (partial occlusion), the system switches to predictive tracking mode. The system extrapolates the target position based on the Kalman predictor of the fused track. The turntable moves according to the predicted trajectory for 2.5 seconds. After the target leaves the occlusion area, the system switches back to normal tracking mode. During the occlusion period, the prediction error is ≤0.3° (better than the turntable beamwidth), achieving seamless crossing.

[0085] S6 feeds back the real-time image miss distance from the photoelectric tracking device as a high-precision observation value to the fusion center. This is used to correct fusion parameters or calibrate system errors of other sensors, including adjusting the process noise covariance of the target motion model and adjusting the time threshold T. lost and the threshold ρ of the Barthel coefficient block If the prediction data provided by a sensor during the occlusion period is too biased, its dynamic confidence weight λ should be reduced. k(i) .

[0086] It should be noted that the present invention is not limited to the above-described embodiments. The above embodiments are merely examples, and any embodiments that have the same structure and perform the same effects as the technical concept within the scope of the present invention are included within the scope of the present invention. Furthermore, various modifications that can be conceived by those skilled in the art to the embodiments, and other ways of constructing by combining some of the constituent elements of the embodiments, without departing from the spirit of the present invention, are also included within the scope of the present invention.

Claims

1. A method for intelligent tracking and monitoring of low-altitude unmanned aerial vehicles (UAVs), characterized in that, include: S1, collect raw data of the target based on multi-source sensors and perform preprocessing; S2, which unifies the data from different sensors in their respective time and space coordinate systems to the same spatiotemporal reference; S3 employs a layered fusion architecture to associate and fuse registered multi-source data, generating fused tracks and identity information; S4 assigns tracking tasks to the photoelectric tracking device based on fused track and identity information, and uses a generalized predictive control algorithm based on a linear extended state observer to generate servo control commands to drive the photoelectric turntable to track the target; S5 detects the occlusion status of the target in the field of view of the optoelectronic device, and uses a three-mode switching strategy to continuously track the target against occlusion based on the detection results; S6 feeds back the real-time image processing results from the photoelectric tracking device as high-precision observations to the fusion center to correct fusion parameters or calibrate the system errors of other sensors.

2. The intelligent tracking and monitoring method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The multi-source sensors include radar detection equipment, radio spectrum detection equipment, RID resolution equipment, and photoelectric tracking equipment; the preprocessing includes filtering, noise reduction, and formatting.

3. The intelligent tracking and monitoring method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The layered fusion architecture specifically includes: An improved joint probabilistic data association algorithm is used to associate radar tracks, spectrum direction finding intersections, RID broadcast locations, and established target tracks. An improved adaptive unscented Kalman filter algorithm, combined with dynamic confidence weights, is used to make optimal estimates of the target position and velocity. Based on the Dempster-Shafer evidence theory, RID parsing information and spectrum analysis information are fused at the decision level to comprehensively determine the target's identity and threat level.

4. The intelligent tracking and monitoring method for low-altitude unmanned aerial vehicles according to claim 3, characterized in that, The dynamic confidence weights are calculated through a dynamic confidence assessment mechanism, which integrates three dimensions: signal-to-noise ratio factor, data residual factor, and environmental parameter factor. This mechanism evaluates the confidence of each sensor in real time and dynamically adjusts the weights of each sensor's data in state fusion based on these dimensions.

5. The intelligent tracking and monitoring method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, An improved Bach coefficient template matching method is used to detect occlusion status in real time, specifically: Add a background weighting factor to the initial template; The target area is divided into multiple sub-blocks, and the Bach coefficient between the color histogram of each sub-block and the corresponding sub-block of the initial template is calculated. The proportion of sub-blocks with a Bach coefficient lower than a preset threshold is used as the overall occlusion level, and the current occlusion status is determined based on the overall occlusion level.

6. The intelligent tracking and monitoring method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The tri-mode switching strategy includes: Normal tracking mode: When the occlusion level is less than the first threshold, LESO-GPC closed-loop control is adopted, and the feedback quantity is the image miss distance. Predictive tracking mode: When the occlusion level is greater than or equal to the first threshold and less than the second threshold, the target position is extrapolated using a Kalman predictor based on the fused track, and the photoelectric turntable is driven to perform open-loop predictive tracking; Intelligent search mode: When the degree of occlusion is greater than or equal to the second threshold and the duration exceeds the time threshold, the scanning path is planned based on the last reliable position of the fused track and the target motion model to search for the target.

7. The intelligent tracking and monitoring method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, A multi-stage auction mechanism is used to allocate tracking tasks to photoelectric tracking equipment.

8. The intelligent tracking and monitoring method for low-altitude unmanned aerial vehicles according to claim 3, characterized in that, The improved joint probabilistic data association algorithm includes: At least two direction finding lines obtained by radio spectrum detection equipment are spatially intersected, the shortest distance point is calculated as a virtual positioning point, and the equivalent covariance matrix of the virtual positioning point is derived by using the error propagation law. Based on the predicted state of each target track, an ellipsoidal threshold is used to filter all measurements and construct a confirmation matrix indicating the possible correspondence between the measurements and the target track. The acknowledgment matrix is ​​decomposed into multiple independent submatrices using sparse matrix factorization. The joint correlation probability of each submatrix is ​​calculated separately, and then the results are combined.

9. The intelligent tracking and monitoring method for low-altitude unmanned aerial vehicles according to claim 3, characterized in that, The decision-level fusion based on Dempster-Shafer evidence theory also includes: Calculate the conflict factor K between the basic probability assignments of the two sources of evidence. rt ; When K rt If the conflict exceeds the preset conflict threshold, it is determined that there is a serious conflict between the RID information and the spectrum information. The threat level of the target is then raised by one level, the tracking priority of radar and optoelectronic equipment is increased, and an alarm is triggered to prompt manual intervention.

10. A low-altitude unmanned aerial vehicle (UAV) intelligent tracking and monitoring system, characterized in that: include: The multi-source sensor module, deployed on the communication tower, includes radar detection equipment, radio spectrum detection equipment, RID analysis equipment, and photoelectric tracking equipment, used to collect raw data of the target in real time; The spatiotemporal registration module is used to uniformly register data from various sensors to the same spatiotemporal reference. The data association and fusion module is used to perform data layer association, feature layer state estimation and decision layer identity fusion on the registered multi-source data using a hierarchical fusion architecture, and generate fused tracks and identity information. The servo tracking control module assigns tracking tasks to the photoelectric tracking device based on fused track and identity information, and generates servo control commands using a generalized predictive control algorithm based on a linear extended state observer to drive the photoelectric turntable to track the target. The anti-occlusion prediction tracking module is used to detect the occlusion status of targets in the field of view of the optoelectronic device and switch between normal tracking mode, prediction tracking mode and intelligent search mode according to the degree of occlusion. The feedback self-optimization module is used to feed back the real-time image processing results of the photoelectric tracking device as high-precision observation values ​​to the data association and fusion module in order to correct the fusion parameters or calibrate the system errors of other sensors.

Images