Unmanned aerial vehicle positioning method and simulation system based on multi-modal identification

Abstract

The invention discloses an unmanned aerial vehicle positioning method and system based on multi-modal identification. The method comprises the following specific steps: S1, obtaining measurement information of two radar stations on a target; establishing a target positioning equation set by using the measurement information; taking the measurement information as observation input, combining a dynamic model of the target, and performing filtering estimation on the position and the speed of the target by using a Kalman filter to obtain an optimized target positioning estimation result; s2, using the target positioning estimation result obtained in the step S1 to associate and weight identification results about target types from a plurality of data sources; taking the identification result of the specified data source as a weighted information source, and carrying out probability weighted adjustment on the identification results of other data sources; and an identification result with the confidence higher than a set value is obtained and fed back to a positioning and tracking link, so that positioning of the unmanned aerial vehicle is realized.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) monitoring and low-altitude traffic management technology, specifically relating to a UAV positioning method and simulation system based on multimodal recognition. This invention integrates high-precision positioning algorithms, real-time target recognition technology, and two-dimensional dynamic modeling technology. It is mainly applied to scenarios such as UAV airspace operation monitoring, UAV swarm collaborative situational awareness, and low-altitude safety training. It can realize dynamic trajectory prediction of UAVs and intelligent early warning of airspace conflicts, and is suitable for applications such as the development of urban air traffic (UAM) management systems, the construction of UAV controller training platforms, and UAV emergency response simulation exercises. Background Technology

[0002] Against the backdrop of rapid development in the low-altitude economy, unmanned aerial vehicle (UAV) target detection and identification systems have become a core infrastructure of the new airspace management system. Traditional air surveillance systems are mainly designed for high-altitude fixed-wing aircraft, and their technical characteristics present significant limitations when dealing with UAV targets: ① Spatial dimension mismatch, unable to effectively cover the complex electromagnetic environment of low altitudes from 0 to 500 meters; ② Target characteristic differences, the radar cross-section (RCS) of micro-rotor UAVs is generally less than 0.01 m² and their motion trajectory is nonlinear; ③ Data fusion defects, existing systems lack the ability to intelligently correlate and analyze multi-source heterogeneous data such as visual recognition, radio spectrum detection, and infrared data.

[0003] There are three major technical bottlenecks in the current field of UAV detection: 1) Insufficient continuous tracking capability of dynamic targets, limited by the rapid maneuverability of UAVs and brief communication interruptions; 2) Low recognition rate of micro targets, with existing optoelectronic systems having a misjudgment rate of more than 50% for UAVs with a diameter of <30cm at a distance of 200 meters; 3) Lagging early warning of airspace conflicts, as traditional trajectory prediction models do not take into account the emergent behavior characteristics of UAV swarms.

[0004] Radar signal processing is a key technology for achieving precise target localization. It involves capturing the target's electromagnetic signal with a receiving antenna, followed by low-noise amplification, mixing, and filtering by the receiver. Then, the analog signal is converted to a digital signal via an analog-to-digital converter (AD converter), and further processed through pulse accumulation, clutter suppression, and target detection to suppress unwanted signals and improve the signal-to-noise ratio. Finally, the extracted target signal undergoes parameter estimation and target tracking data processing. In existing radar systems for UAV localization, it is often necessary to modify underlying code parameters (radar frequency, pulse width, system loss, etc.) to adapt to different environments or target characteristics. However, this approach has significant drawbacks: 1. Parameter adjustment relies on specialized programming knowledge (modifying C / Python radar signal processing code), making it impossible for non-technical personnel to independently optimize the system; 2. Strong coupling between parameters, for example, modifying the noise covariance matrix of the Kalman filter requires simultaneous adjustment of related parameters, which is prone to human error through manual adjustment; 3. Poor reusability, making it difficult to quickly migrate to other radar sites via standardized interfaces for specific scenarios such as distant seas and waters; 4. Lack of visual feedback, making it difficult for operators to directly judge the impact of parameters on positioning accuracy. Summary of the Invention

[0005] To address the problems of low positioning accuracy and poor recognition accuracy of existing technologies for unmanned aerial vehicles (UAVs), this invention provides a UAV positioning method and simulation system based on multimodal recognition.

[0006] The present invention adopts the following technical solution: The UAV localization method based on multimodal recognition has the following specific steps: S1. Acquire the measurement information of the target from two radar stations; establish a target positioning equation set using the measurement information; use the measurement information as the observation input, combine it with the target's dynamic model, and use a Kalman filter to filter and estimate the target's position and velocity to obtain the optimized target positioning estimation result. S2. Using the target location estimation results obtained in step S1, correlate and weight the target type identification results from multiple data sources. Using the identification results from the specified data source as the weighting information source, perform probability weighting adjustments on the identification results from other data sources; obtain high-confidence identification results and feed them back to the positioning and tracking stage in each tracking cycle (within this cycle, feedback will be given as long as the confidence level is higher than this set value, so real-time tracking is continuously running within this cycle until the target disappears or the system shuts down and stops positioning), to achieve more accurate and stable continuous tracking and positioning of UAV targets.

[0007] Preferably, step S1 specifically includes the following steps: S1.1. Suppose there is a transmitting station Tx and a spatially separated receiving station Rx. Through coordinated signal processing, the following measurement information is obtained: Distance and Rs: The total path distance of the signal from the transmitting station Tx to the target T, and then reflected back to the receiving station Rx, Rs = Rr + Rt, where Rr and Rt are the distances from the target T to the transmitting station and the receiving station, respectively; Azimuth θ : The azimuth angle of the target relative to the receiving station Rx; Pitch angle : The angle between the line connecting the target and the receiving station Rx and the horizontal plane; S1.2 Based on the geometric relationship of bistatic radar, establish and solve the target positioning equation set; The positioning equation is: The distance and equation are given by equation (1), and the spherical coordinate equation of the receiving station is given by equations (2), (3), and (4), where (x0, y0, z0) are the coordinates of the transmitter, (x, y, z) are the coordinates of the target, and the receiving station is the coordinate of the origin; Rr and Rt are the distances from the target T to the transmitter and the receiving station, respectively, and the azimuth angle is given by equation (4). θ Pitch angle ; The solution equation is as follows: Where Rs is the distance sum; Formulas (5), (6), and (7) are parameterized target coordinates. Substituting them into the three-dimensional distance and formula (1) yields the single-variable equation (8). Solving equation (8) using the Newton-Raphson method yields the numerical solution of the distance Rr from the target to the receiving station. The solution to equation (8) using the Newton-Raphson method is as follows: Equation (8) can be rewritten as the following objective function f(Rr): Here, Rs, ϕ, x0, y0, and z0 are all known quantities, and A(Rr) is an intermediate function; Calculate the derivative of the objective function f(Rr) with respect to the variable Rr. : After simplification, we get: Using trigonometric identities The derivative is further simplified to: Perform iterative calculations: Choose the initial value R for the iteration r，0 The initial value selection strategy is as follows ; Iterative Formula: Iterative calculations are performed using the iterative formula of the Newton-Raphson method. Where n is the number of iterations, and a convergence condition is set: the iteration terminates when the absolute or relative error of two consecutive iterations is less than a preset threshold, and the result R of the last iteration is given. r,final As a solution to equation (8); The solution Rr=R r,final Substituting into equations (5), (6), and (7), we obtain the absolute coordinates (x, y, z) of the target in three-dimensional space, and thus the positioning result. S1.3. After obtaining the target sequence (x, y, z) through step S1.2, input it as the observation value into the Kalman filter. The Kalman filter performs filtering estimation of the target position and velocity, and outputs a six-dimensional state vector [x, y, z, v]. x ,v y ,v z ] T .

[0008] Preferably, step S1.3 is as follows: S1.3.1 Defining the State Vector and State Equation State vector X k The target state vector at time k is defined as: Where (x,y,z) represents the position of the target in three-dimensional space, (v x ,v y ,v z ) represents the velocity components of the target along the three coordinate axes; State equations: Assuming the target is a uniform motion model, the state equations are: Where F is the state transition matrix, and for the uniform velocity model, it takes the following form: Where x is the radar sampling time interval, W k It is process noise with zero mean and covariance matrix Q, used to simulate random disturbances in the target motion; S1.3.2, Define the observation vector and observation equation Observation vector Z k The observation vector is defined as the target position obtained from the final solution in step S1.2. Observation equations establish the relationship between the state vector and the observation vector: Where H is the observation matrix; since the observation vector is the positional part of the state vector, therefore: Among them, V k It is the observation noise with zero mean and covariance matrix R, used to characterize the error of the positioning solution in step S1.2; S1.3.3, Kalman Filter Execution of Kalman Filter Recursive Algorithm The Kalman filter algorithm recursively performs the following prediction and update steps in each sampling period k: 1) Prediction steps: Based on the optimal estimate of the previous time step, predict the current state and error covariance; in, It is a state prediction value. It is the prediction error covariance matrix; 2) Update steps: Use the actual observed value Z at the current time. k Correct the predicted values: Calculate the Kalman gain K k : Updated state estimate: Update error covariance estimate: Where I is the identity matrix; Output: The state estimate after each iteration update. As output, we obtain the filtered and optimized target six-dimensional state vector [x,y,z,v]. x ,v y ,v z ] T .

[0009] Preferably, step S2 is as follows: S2.1 Weighted preprocessing of confidence-guided recognition results; S2.2, Integration of DS Evidence Theory and Decision Judgment; S2.3 Optimization of the positioning and tracking closed loop based on the recognition results.

[0010] Preferably, step S2.1 is as follows: First, we introduce a location-aware confidence-guided weighted mechanism: 1) Dynamically specify weighted information sources: The most reliable identification source is dynamically selected as the guiding source based on the target's real-time status and sensor characteristics; the decision rule is as follows: a. When the target is above the set altitude and above the set speed, the radar identification result is selected as the weighted information source first; b. When the target is below the set altitude and speed and the ambient light intensity is greater than the set value, the photoelectric result is selected as the weighted information source. c. When a specific communication protocol signal is detected, the communication reconnaissance results should be given priority as the weighted information source; 2) Extracting guidance information: Extract the target type proposition A_max with the highest confidence from the basic probability assignment function m_guide of the selected weighted information sources; Here, m_guide is a function that provides a framework for recognition. Each proposition in the equation is assigned a probability mass, representing the degree to which the information source supports the proposition being true; argmax is a mathematical operator that aims to find the parameter that maximizes the value of the subsequent function. This indicates traversing all possible propositions in the framework; 3) Probability-weighted adjustment: Guided by A_max, the basic probability allocation functions of other recognition sources are enhanced in a targeted manner. For the recognition source m_i to be adjusted, its adjusted basic probability allocation function m_i' is calculated according to the following rules: The guiding proposition is enhanced: Decrease other propositions: Where m_i: the basic probability allocation function of another information source to be adjusted, representing the original support of the information source for various propositions; A_max: the proposition with the highest confidence extracted from the most trusted guiding source m_guide; α: weighting coefficient; B is all other propositions not equal to A_max, and m_i(B) is the original support of the information source for other non-guiding propositions B.

[0011] Preferably, step S2.2, DS evidence theory fusion and decision judgment, is as follows: The weighted basic probability allocation functions m_guide,m1',m2',...,mn' are fused using the DS combination rule: S2.2.1. Use the DS combination formula to perform pairwise recursive fusion to obtain the fused basic allocation probability m_final; The following is a recursive fusion of DS combination rules: 1) DS combination rule formula Suppose we want to merge two basic probability assignment functions m1 and m2, and the new probability assignment function m generated after their fusion is... 12 Defined by the following formula: For all ; in, : This means finding all propositions A from m1 and propositions B from m2 whose intersection is exactly equal to proposition C; : satisfy all The sum of the products of m1(A) and m2(B) represents the total probability mass assigned to proposition C before normalization. Conflict coefficient K: K calculates the sum of the probability-mass products of all perfectly conflicting proposition pairs; K=0 indicates that there is no conflict between the two sources of evidence; the larger K is, the more serious the conflict between the sources of evidence. Normalization factor (1 / (1-K)): Its function is to redistribute the probability mass K lost due to conflicting evidence, ensuring... ; 2) Recursive fusion process When there are multiple sources of evidence, recursive fusion is performed: Step 1: Combine m_guide and m1' to obtain the fused result m_fused1; m_fused1 = m_guide ⊕ m1' Step 2: Combine m_fused1 with m2' to obtain a new m_fused2; m_fused2 = m_fused1 ⊕ m2' And so on, step N-1: Combine the result of step N-2 with the last source of evidence m. n The functions are combined to obtain the final fused basic probability allocation function m_final; S2.2.2 Calculate the trust function for each proposition based on m_final. Bel(A) represents the total confidence level in proposition A. This means summing all propositions B that are contained in A; S2.2.3 Select the proposition with the highest confidence level as the final target type identification result: Traversal recognition framework For all propositions in the given set, find the proposition A that maximizes Bel(A).

[0012] Preferably, step S2.3, the optimization of the positioning and tracking closed loop based on the recognition results, is as follows: The target type identification result T_final is fed back to the Kalman filter tracking stage in step S1 to achieve bidirectional enhancement of localization and identification: Motion model adaptation: Based on the identified target type, the key parameters in the state equation are adaptively adjusted: 1) If identified as a consumer-grade drone: adopt the standard uniform velocity model, and set the process noise covariance matrix to be less than the set value to reflect its relatively stable motion characteristics; 2) If identified as an attack drone: Increase the corresponding elements of the process noise covariance matrix based on the uniform velocity model to enhance the filter's ability to track sudden maneuvers; 3) If identified as a swarm of drones: an interactive multi-model algorithm is used to switch probabilities between multiple motion models; Observation model optimization: Optimize the observation noise covariance matrix based on the target type: 1) For small UAVs with small radar cross-sections and weak infrared characteristics, increase the noise variance of the corresponding observation channel; 2) For UAVs with typical communication signal characteristics, reduce observation uncertainty when the communication reconnaissance source provides accurate identification results.

[0013] This invention also discloses a UAV positioning simulation system based on multimodal recognition, comprising a front-end interface module and a back-end algorithm module; wherein, The front-end interface module is used to input simulated target parameters, radar system parameters, and infrared system parameters, and to call algorithms; and to display the output positioning and recognition results. The background algorithm module, based on the above method, is used to process the parameters passed from the front end and feed them back to the front end interface module.

[0014] The front-end interface designed in this invention is as follows: Figure 1 As shown. The target simulation parameter input interface is as follows: Figure 2As shown, the required drone type can be set according to the simulation scenario, including the number of drones and the flight paths drawn based on the characteristics of the drones. The radar system parameter module is as follows... Figure 3 As shown, the parameters used in the algorithm are set as input interfaces. After the front-end interface receives the input data, it performs preliminary data processing, allowing the back-end algorithm to directly call the data. The required radar site can also be directly selected on this interface, enabling rapid migration of standardized interfaces. The infrared system parameter module is as follows... Figure 4 As shown, this interface allows users to input the system parameters required by the infrared system and receive and process calls to the background algorithm. The simulated drone demonstration module is shown below. Figure 5 As shown, this interface displays the simulated drone's flight status. The positioning results display module is as follows: Figure 6 As shown, you can see the real-time positioning effect rendering here. Clicking "Start" will automatically generate a trajectory map, allowing you to view trajectory information and display positioning errors. The recognition results module is as follows: Figure 7 As shown, it can put various data sources into one architecture, automate data conversion, and use background fusion algorithms to calculate and identify results, simplifying the complex multi-source fusion process into "upload-click-view".

[0015] Finally, the link between the front-end interface module and the back-end algorithm module of this invention is explained. This invention packages the back-end algorithm into an external library and imports this library into the computer program of the front-end interface, thereby realizing communication and collaboration between the front-end and back-end. This invention relates to the field of aviation surveillance technology, specifically designing a positioning and identification system and method based on simulated unmanned aerial vehicles (UAVs). The system mainly comprises two modules: a front-end interface and a back-end algorithm. The front-end interface module, based on the algorithm's input and output interfaces, designs corresponding input data and output result modules to ensure correct data input. After data processing and invoking the back-end algorithm, the calculation results are displayed in various ways. The back-end algorithm module mainly consists of two types of algorithms: positioning and identification. Positioning primarily uses radar positioning algorithms, combining ranging and angle information with filtering and optimization techniques to achieve accurate and rapid target positioning in complex environments. Identification primarily uses a decision-level fusion algorithm, employing a guided enhancement fusion method based on Dempster-Shafer (DS) evidence theory.

[0016] Compared with the prior art, the present invention has the following significant advantages: I. It achieves the fusion of positioning and recognition, forming a closed loop of performance enhancement: This invention overcomes the limitations of traditional independent positioning and identification by combining physical positioning with type identification. The target positioning estimation result (position and velocity) obtained in step S1 is used as a reliable basis to correlate and weight the multi-source identification information in step S2, achieving confidence-guided data fusion. The high confidence level after fusion can optimize the tracking parameters and model of the Kalman filter (different UAVs have different characteristics, requiring different parameters and models), forming a closed loop of "positioning assisting identification, and identification providing feedback to positioning," thus solving the problem of insufficient accuracy in a single step.

[0017] II. Improve positioning accuracy and stability in complex environments: By combining measurements from two radar stations with a Kalman filter algorithm, this invention effectively improves the localization estimation of target trajectories and provides high-quality state association information for subsequent identification and fusion, thereby enabling continuous localization and tracking of UAV targets, especially in complex environments.

[0018] III. Improve the confidence and reliability of target type identification: This invention adopts a guided enhancement fusion method based on DS evidence theory. It is not a simple fusion of multi-source information. Its core lies in the front-end positioning information specifying a reliable data source as guidance, and performing probability weighting on other identification sources. This mechanism effectively amplifies the contribution of reliable information sources, suppresses the influence of unreliable or conflicting information, improves the confidence of the identification results, and reduces the probability of misjudgment and missed judgment.

[0019] IV. An efficient simulation verification platform has been built to improve R&D and evaluation efficiency: This invention presents a simulation system integrating a front-end interface and a back-end algorithm, achieving a seamless end-to-end process from parameter input and algorithm invocation to result visualization. This simulation system can intuitively demonstrate the entire localization and recognition process using animation, images, and data, facilitating rapid verification of algorithm performance, analysis of localization errors, and comparison of different recognition sources, providing a powerful tool for method optimization and iteration. Attached Figure Description

[0020] Figure 1 This is a layout diagram of the front-end interface of a preferred embodiment of the present invention; Figure 2 for Figure 1 Diagram of the target simulation module; Figure 3 for Figure 1 Diagram of radar system parameters; Figure 4 for Figure 1 Mid-infrared system parameter module diagram; Figure 5 for Figure 1 A simulated drone demonstration image; Figure 6 for Figure 1 The positioning results are shown in the image. Figure 7 for Figure 1 Image showing the recognition results; Figure 8 This is a geometrical schematic diagram of the bistatic radar in the positioning method of a preferred embodiment of the present invention; Figure 9 This is a flowchart illustrating the workflow of a drone positioning and identification simulator according to a preferred embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and specific examples. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the invention.

[0022] This embodiment discloses a UAV localization method based on multimodal recognition, including the following steps: S1: Positioning estimation step. This step aims to obtain stable and accurate target position and velocity information through radar measurement and filtering techniques. Figure 8 This describes the geometric positional relationship of a target in a bistatic radar system. The localization estimation step acquires measurement information of the target from two radar stations, including range, azimuth, and elevation angles. Based on triangulation and geometric relationships, a set of target localization equations is established using this measurement information. Using this measurement information as observation input, and combined with the target's dynamic model, a Kalman filter is used to filter and estimate the target's position and velocity, resulting in an optimized target localization estimate. Specifically, this includes the following sub-steps: S1.1 The system includes a transmitting station Tx and a spatially separated receiving station Rx. Through coordinated signal processing, the following measurement information can be obtained.

[0023] Distance and Rs: This refers to the total path distance of the signal from the transmitting station Tx to the target T, and then reflected back to the receiving station Rx. Rs = Rr + Rt, where Rr and Rt are the distances from the target T to the transmitting and receiving stations, respectively. This parameter can be calculated from observational information such as signal arrival time.

[0024] Azimuth θ : The azimuth angle of the target relative to the receiving station Rx. This parameter can be obtained through techniques such as phase difference measurement of the receiving station.

[0025] Pitch angle : The angle between the line connecting the target and the receiving station Rx and the horizontal plane. This parameter can be obtained through techniques such as phase difference measurement at the receiving station.

[0026] S1.2. Based on the geometric relationship of bistatic radar, establish and solve the positioning equations.

[0027] The positioning equation is: The distance and equation are given by equation (1), and the spherical coordinate equation of the receiving station can be obtained by equations (2), (3), and (4). Where (x0, y0, z0) are the coordinates of the transmitter, (x, y, z) are the coordinates of the target, the receiving station is the origin coordinate (0, 0, 0), Rr and Rt are the distances from the target T to the transmitter and receiver respectively, and the azimuth angle is... θ Pitch angle .

[0028] The solution equation is as follows: Rs is the distance sum, and formulas (5), (6), and (7) are the parameterized target coordinates. Substituting them into the three-dimensional distance sum equation (1) yields the single-variable equation (8). Equation (8) can be solved using the Newton-Raphson method to obtain the numerical solution of the distance Rr from the target to the receiving station.

[0029] Solving equations using the Newton-Raphson method (8): Equation (8) can be rewritten as the following objective function f(Rr): Here, Rs,ϕ,,x0,y0,z0 are all known quantities, and A(Rr) is an intermediate function.

[0030] Calculate the derivative of the objective function f(Rr) with respect to the variable Rr. : After simplification, we get: Using trigonometric identities The derivative is further simplified to: Perform iterative calculations: Choose the initial value R for the iteration r，0 A feasible initial value selection strategy is as follows: ; Iterative Formula: Iterative calculations are performed using the iterative formula of the Newton-Raphson method. Where n is the number of iterations, and a convergence condition is set: the iteration terminates when the absolute or relative error of two consecutive iterations is less than a preset threshold, and the result R of the last iteration is given. r,final As a numerical solution to equation (8).

[0031] The solution Rr=R r,final Substituting into equations (5), (6), and (7), we can obtain the absolute coordinates (x, y, z) of the target in three-dimensional space. The positioning result can be obtained from the above solution process.

[0032] S1.3. After obtaining the target sequence (x, y, z) through step S1.2, input it as the observation value into the Kalman filter. The Kalman filter uses the target's dynamic model (e.g., uniform velocity model) and a recursive algorithm to not only smooth and reduce noise at these discrete positions, but also internally estimate the target velocity component (v) based on position changes. x ,v y ,v z Therefore, the Kalman filter outputs a more complete and smooth six-dimensional state vector X. k =[x,y,z,v x ,v y ,v z ] T .

[0033] S1.3.1 Defining the State Vector and State Equation State vector X k The target state vector at time k is defined as: Where (x,y,z) represents the position of the target in three-dimensional space, (v x ,v y ,v z () represents the velocity components of the target in the three coordinate axes.

[0034] State equations: Assuming the target is in uniform motion, the state equations are as follows: Where F is the state transition matrix, and for the uniform velocity model, its specific form is: Here, x is the radar sampling time interval, W k It is a system process noise with zero mean and covariance matrix Q, used to simulate random disturbances in the target motion.

[0035] S1.3.2, Define the observation vector and observation equation Observation vector Z k The observation vector of this invention is the target position directly calculated at the end of step S1.2; therefore, the observation vector is defined as follows: Observation equations: The observation equations establish the relationship between the state vector and the observation vector. Where H is the observation matrix. Since the observation vector is the positional part of the state vector, therefore: Among them, V k It is the observation noise with zero mean and covariance matrix R, used to characterize the error in the positioning solution in step S1.2.

[0036] S1.3.3. A smoother and more accurate target estimate is provided using a Kalman filter. The Kalman filter utilizes the Kalman filtering algorithm, which recursively performs the following prediction and update steps in each sampling period k: 1. Prediction steps: Based on the optimal estimate of the previous time step, predict the current state and error covariance.

[0037] in, It is a state prediction value. It is the prediction error covariance matrix.

[0038] 2. Update steps: Use the actual observed value Z at the current time. k Correct the predicted values: Calculate the Kalman gain K k : Updated state estimate: Update error covariance estimate: Where I is the identity matrix; Output: The state estimate after each iteration update. As output, we obtain the filtered and optimized target six-dimensional state vector [x,y,z,v]. x ,v y ,v z ] T This enables high-precision and high-stability target tracking.

[0039] Step S1 first performs localization estimation. The estimation information from Step S1 is then fed to Step S2 as a weighted guide. Other identification results are then fused using the DS evidence theory based on this information to obtain the highest confidence value, which is then fed back to Step S1. Step S2 is described below.

[0040] S2: Recognition Fusion Step: After obtaining the target's location and state estimation results through step S1, the state estimation (velocity, position) is used to perform confidence fusion of target recognition results from multiple data sources. High-confidence recognition results are then fed back to the tracking and positioning stage, forming a closed-loop optimization mechanism of "location-recognition-relocation". The recognition fusion employs a guided enhancement fusion method based on Dempster-Shafer (DS) evidence theory, effectively fusing judgment results from multiple reconnaissance methods such as radar, infrared, communication reconnaissance, and visible light. Details are as follows: S2.1, Confidence-guided identification weighted preprocessing: Before fusing the S2.2D-S evidence theory, a confidence-guided weighting mechanism based on location awareness is introduced, which uses the target location estimation results obtained in step S1 for weighting.

[0041] 1. Dynamically specify weighted information sources: The most reliable identification source is dynamically selected as the guidance source based on the target's real-time status and sensor characteristics. The decision rule is as follows: (1) When the target is at high altitude and high speed, the radar identification result should be selected as the weighted information source. (2) When the target is at low altitude and low speed and the ambient light is good, the photoelectric results should be selected as the weighted information source. (3) When a specific communication protocol signal is detected, the communication reconnaissance results should be selected as the weighted information source. 2. Extract guidance information: From the basic probability assignment function m_guide of the selected weighted information sources, extract the target type proposition A_max with the highest confidence. m_guide is a function that provides a framework for recognition. Each proposition in the framework is assigned a probability mass, representing the degree to which the information source supports the proposition being true. For example, in the target type framework... ={Consumer, Reconnaissance, Attack}, where m_guide represents m_guide (Consumer) = 0.2, m_guide (Reconnaissance) = 0.3, and m_guide (Attack) = 0.5. argmax is a mathematical operator that finds the parameter whose value is maximized. This indicates that all possible propositions in the framework are traversed.

[0042] 3. Probability-weighted adjustment: Guided by A_max, the basic probability allocation functions of other recognition sources are enhanced in a targeted manner. For the recognition source m_i to be adjusted, its adjusted basic probability allocation function m_i' is calculated according to the following rules: The guiding proposition is enhanced: Decrease other propositions: m_i: The basic probability allocation function of another information source to be adjusted, representing the original support level of that information source for various propositions; A_max: The proposition with the highest confidence extracted from the most trusted guiding source (m_guide). α: A weighting coefficient, a value between 0 and 1, controlling the degree of adjustment. The larger this value, the more trusted the guiding source is, and the greater the correction force for other information sources. B represents all other propositions not equal to A_max, and m_i(B) is the original support level of this information source for other non-guiding propositions B.

[0043] S2.2, Integration of DS Evidence Theory and Decision Judgment The weighted basic probability allocation function (m_guide, m1', m2', ..., mn') is fused using the DS combination rule, which means fusing the recognition results from other information sources: S2.2.1. Use the DS combination formula to perform pairwise recursive fusion, and finally obtain the fused basic allocation probability m_final. The following is a recursive fusion of DS combination rules: 1) DS combination rule formula Suppose we want to merge two basic probability assignment functions m1 and m2, and the resulting new probability assignment function m is... 12 Defined by the following formula: For all ; This means finding all propositions A from m1 and propositions B from m2 whose intersection is exactly equal to proposition C.

[0044] : satisfy all The product of m1(A) and m2(B) is added together. This sum represents the total probability mass assigned to proposition C before normalization.

[0045] Conflict coefficient K: K calculates the sum of the probability mass products of all completely conflicting proposition pairs (i.e., the intersection of A and B is an empty set).

[0046] The value of K is between 0 and 1. K=0 indicates that there is no conflict between the two sources of evidence; the larger the value of K, the more serious the conflict between the sources of evidence.

[0047] Normalization factor : Its function is to redistribute the probability mass K lost due to conflicting evidence, ensuring... ,(1 - K) can be understood as the degree of consistency between two sources of evidence.

[0048] 2) Recursive fusion process Since the DS combination rule is a binary operation, when there are multiple sources of evidence (N>2), recursive fusion is required: Step 1: First, combine m_guide and m1' to obtain the fused result m_fused1. m_fused1 = m_guide ⊕ m1' Step 2: Combine m_fused1 with m2' to obtain a new m_fused2. m_fused2 = m_fused1 ⊕ m2' And so on; Step N-1: Combine the result of step N-2 with the last source of evidence m. n The functions are combined to obtain the final fused basic probability allocation function m_final.

[0049] S2.2.2 Calculate the trust function for each proposition based on m_final. , Bel(A) represents the total confidence level in proposition A. This means summing all propositions B that are contained within A. For example, if A = {Aggressive}, then B can only be {Aggressive}. If A = {Aggressive, Scouting}, then B can be {Aggressive}, {Scouting}, or {Aggressive, Scouting}.

[0050] S2.2.3 Select the proposition with the highest confidence level as the final target type identification result: Traversal recognition framework For all propositions in the given set, find the proposition A that maximizes Bel(A).

[0051] S2.3 Optimization of Localization and Tracking Closed-Loop Based on Recognition Results The high-confidence target type identification result T_final is fed back to the Kalman filter tracking stage of S1 to achieve bidirectional enhancement of localization and identification: Motion model adaptation: In a Kalman filter, the state equation is used to describe the target's dynamic model (such as a uniform velocity model). Its key parameters include the state transition matrix F and the process noise covariance matrix Q. These parameters are adaptively adjusted for optimized tracking training based on the identified target type.

[0052] 1. If identified as a consumer-grade drone: adopt the standard uniform velocity model, that is, the state equation assumes that the target moves at a constant velocity, and the state transition matrix F is as shown in formula (17) in S1.3.1, a 6×6 matrix. The process noise covariance matrix Q should be set to a small value. Q is usually a diagonal matrix. For example, the diagonal elements of Q can be set to [0.1,0.1,0.1,0.01,0.01,0.01]. The first three correspond to position noise, and the last three correspond to velocity noise. This reflects the characteristics of consumer-grade drones, which have relatively smooth motion and low maneuverability, making the filter track more smoothly.

[0053] 2. If identified as an attack drone: Increase the velocity-related elements in the process noise covariance matrix Q based on the uniform velocity model to enhance the filter's ability to track sudden maneuvers (such as acceleration and turning). For example, since the Q matrix is ​​a diagonal matrix, the noise variance of its velocity components can be increased from 0.01 to 0.1 or higher, which allows the filter to respond to motion changes faster and reduce hysteresis errors.

[0054] 3. If identified as a swarm of drones: The Interactive Multiple Model (IMM) algorithm is employed. IMM is an advanced tracking technique that maintains multiple motion models simultaneously (e.g., uniform motion model, acceleration model) and switches between models based on probability weights. Specifically, a set of models is defined (e.g., Model 1 for uniform motion, Model 2 for acceleration), each with its own state equation and Q-matrix. IMM calculates the likelihood probability and state estimate of each model and performs weighted fusion. This approach is suitable for the diverse motion patterns that swarm of drones may exhibit, improving tracking robustness.

[0055] Observation model optimization: The observation equation describes the relationship between measured values ​​and state values, and its key parameter is the observation noise covariance matrix R. Optimizing R according to the target type can adjust the filter's confidence in the observation data. (1) For small UAVs with small radar cross-sections and weak infrared characteristics: Since the target observation signals of such UAVs are weak and the errors are large, the corresponding elements of the observation noise covariance matrix R should be appropriately increased. For example, R is usually a diagonal matrix, and its diagonal values ​​can be increased from 1 to 5 (indicating increased observation uncertainty), so that the filter depends more on the predicted values ​​than the observed values, thereby reducing noise interference.

[0056] (2) For UAVs with typical communication signal characteristics: When the communication reconnaissance source provides reliable identification results (such as accurate signal parameter estimation), the observation uncertainty should be reduced accordingly, that is, the elements of the observation noise covariance matrix R should be reduced. For example, if the communication reconnaissance determines the target type, the diagonal line in the R matrix can be changed from 1 to 0.1 (indicating that the observation is more reliable), so that the filter trusts the observation input more and improves the positioning accuracy.

[0057] The front-end interface of this invention will be described below.

[0058] A preferred embodiment of the present invention designs a UAV positioning simulation system based on multimodal recognition, which consists of a front-end interface module and a back-end algorithm module.

[0059] The front-end interface module is used to input simulated target parameters, radar system parameters, and infrared system parameters, and to call algorithms; it is also used to display the output positioning and recognition results. Input parameter processing includes receiving and encapsulating the data; output result display includes animations, images, and calculation results. Front-end interface ( Figure 1 It mainly includes five modules, namely: the target simulation parameter module ( Figure 2 ), Radar system parameter module ( Figure 3 Infrared system parameter module ( Figure 4 ), simulated drone demonstration module ( Figure 5 ), Location recognition result display module ( Figure 6 ).

[0060] In the processing of input parameters, receiving and encapsulating refer to saving the input parameters and storing them in a specified location, and converting the data format to meet the format required by the algorithm call, so that the background algorithm can directly read them when it starts to call.

[0061] The output results are displayed using a 2D animation to show the drone's localization process, directly demonstrating the calculation process, and showing the calculation results and localization errors in the side border. The individual drone identification data sources are integrated into a switchable frame, and their respective identification results are combined into a table to display the final fused identification result.

[0062] The background algorithm module, based on the above method embodiment, is used to process and calculate parameters in a specific format input from the front end. Positioning algorithm: First, it acquires measurement information such as the target's distance and / or distance difference, azimuth, etc., from two radar stations, and establishes a set of equations for target positioning using triangulation and geometric relationships. Then, it uses these measurement data as observation input, combined with a dynamic model of the target's motion state, and uses Kalman filtering to filter and estimate the target's position and velocity to reduce measurement noise and errors, and improve the accuracy and stability of positioning. Recognition algorithm: Primarily a data source decision-level fusion algorithm, it employs a guided enhancement fusion method based on Dempster-Shafer (DS) evidence theory. Using a certain type of reconnaissance result as a weighted information source, it guides other recognition results to undergo probability weighting adjustments before fusion, achieving weight enhancement under confidence guidance.

[0063] Real-time positioning effect rendering: input parameters and click start to automatically generate a trajectory map, view trajectory information, and display positioning error.

[0064] Target recognition technologies include optical recognition, electronic reconnaissance recognition, infrared recognition, and multi-source fusion. Optical recognition relies on optical sensors to capture images of drones and then uses image processing and computer vision techniques for identification. Electronic reconnaissance recognition primarily acquires electronic signal information through electronic means to achieve identification. Infrared recognition mainly utilizes infrared sensors to detect and identify the thermal radiation of drones. Multi-source fusion technology integrates data from different recognition methods (such as optical, infrared, and electronic reconnaissance) and fuses the multi-source information to obtain more comprehensive and reliable target features, thereby improving recognition performance. Traditional multi-source recognition systems suffer from a "black box" operation problem. They exhibit the following drawbacks: Drawback 4: Data silos – optical, electronic reconnaissance, and infrared data are scattered across different terminals, requiring comparison on different screens; Drawback 5: The entire process must be re-executed every time the data source or algorithm parameters are modified.

[0065] This invention designs and identifies interactive software that centralizes various data sources under one framework, automates data conversion, and simplifies the complex multi-source fusion process into a three-step operation of "upload-click-view".

[0066] The above description is merely a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. It should be emphasized that any minor adjustments and improvements made by those skilled in the art without departing from the basic principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A UAV localization method based on multimodal recognition, characterized by: The specific steps are as follows: S1. Acquire the measurement information of the target from two radar stations; establish a target positioning equation set using the measurement information; use the measurement information as the observation input, combine it with the target's dynamic model, and use a Kalman filter to filter and estimate the target's position and velocity to obtain the optimized target positioning estimation result. S2. Using the target positioning estimation results obtained in step S1, correlate and weight the target type identification results from multiple data sources; use the identification results from the specified data source as the weighting information source to perform probability weighting adjustment on the identification results from other data sources; obtain identification results with a confidence level higher than the set value and feed them back to the positioning and tracking stage in each tracking cycle to achieve the positioning of the UAV.

2. The UAV localization method based on multimodal recognition as described in claim 1, characterized in that, Step S1 specifically includes the following steps: S1.

1. Suppose there is a transmitting station Tx and a spatially separated receiving station Rx. Through coordinated signal processing, the following measurement information is obtained: Distance and Rs: The total path distance of the signal from the transmitting station Tx to the target T, and then reflected back to the receiving station Rx, Rs = Rr + Rt, where Rr and Rt are the distances from the target T to the transmitting station and the receiving station, respectively; Azimuth θ : The azimuth angle of the target relative to the receiving station Rx; Pitch angle : The angle between the line connecting the target and the receiving station Rx and the horizontal plane; S1.2 Based on the geometric relationship of bistatic radar, establish and solve the target positioning equation set; The positioning equation is: The distance and equation are given by equation (1), and the spherical coordinate equation of the receiving station is given by equations (2), (3), and (4), where (x0, y0, z0) are the coordinates of the transmitter, (x, y, z) are the coordinates of the target, and the receiving station is the coordinate of the origin; Rr and Rt are the distances from the target T to the transmitter and the receiving station, respectively, and the azimuth angle is given by equation (4). θ Pitch angle ; The solution equation is as follows: Where Rs is the distance sum; Formulas (5), (6), and (7) are parameterized target coordinates. Substituting them into the three-dimensional distance and formula (1) yields the single-variable equation (8). Solving equation (8) using the Newton-Raphson method yields the numerical solution of the distance Rr from the target to the receiving station. The solution to equation (8) using the Newton-Raphson method is as follows: Equation (8) can be rewritten as the following objective function f(Rr): Here, Rs, ϕ, x0, y0, and z0 are all known quantities, and A(Rr) is an intermediate function; Calculate the derivative of the objective function f(Rr) with respect to the variable Rr. : After simplification, we get: Using trigonometric identities The derivative is further simplified to: Perform iterative calculations: Selecting the initial value R for iteration r，0 The initial value selection strategy is as follows ; Iterative Formula: Iterative calculations are performed using the iterative formula of the Newton-Raphson method. Where n is the number of iterations, and a convergence condition is set: the iteration terminates when the absolute or relative error of two consecutive iterations is less than a preset threshold, and the result R of the last iteration is given. r,final As a solution to equation (8); The solution Rr=R r,final Substituting into equations (5), (6), and (7), we obtain the absolute coordinates (x, y, z) of the target in three-dimensional space, and thus the positioning result. S1.

3. After obtaining the target sequence (x, y, z) through step S1.2, input it as the observation value into the Kalman filter. The Kalman filter performs filtering estimation of the target position and velocity, and outputs a six-dimensional state vector [x, y, z, v]. x ,v y ,v z ] T .

3. The UAV positioning method based on multimodal recognition as described in claim 2, characterized in that the steps are as follows: S1.3 is as follows: S1.3.1 Defining the State Vector and State Equation State vector X k The target state vector at time k is defined as: Where (x,y,z) represents the position of the target in three-dimensional space, (v x ,v y ,v z ) represents the velocity components of the target along the three coordinate axes; State equations: Assuming the target is a uniform motion model, the state equations are: Where F is the state transition matrix, and for the uniform velocity model, it takes the following form: Where x is the radar sampling time interval, W k It is process noise with zero mean and covariance matrix Q, used to simulate random disturbances in the target motion; S1.3.2, Define the observation vector and observation equation Observation vector Z k The observation vector is defined as the target position obtained from the final solution in step S1.

2. Observation equations establish the relationship between the state vector and the observation vector: Where H is the observation matrix; since the observation vector is the positional part of the state vector, therefore: Among them, V k It is the observation noise with zero mean and covariance matrix R, used to characterize the error of the positioning solution in step S1.2; S1.3.3, Kalman Filter Execution of Kalman Filter Recursive Algorithm The Kalman filter algorithm recursively performs the following prediction and update steps in each sampling period k: 1) Prediction steps: Based on the optimal estimate of the previous time step, predict the current state and error covariance; in, It is a state prediction value. It is the prediction error covariance matrix; 2) Update steps: Use the actual observed value Z at the current time. k Correct the predicted values: Calculate the Kalman gain K k : Updated state estimate: Update error covariance estimate: Where I is the identity matrix; Output: The state estimate after each iteration update. As output, we obtain the filtered and optimized target six-dimensional state vector [x,y,z,v]. x ,v y ,v z ] T .

4. The UAV localization method based on multimodal recognition as described in claim 3, characterized in that, Step S2 is as follows: S2.1 Weighted preprocessing of confidence-guided recognition results; S2.2, Integration of DS Evidence Theory and Decision Judgment; S2.3 Optimization of the positioning and tracking closed loop based on the recognition results.

5. The UAV localization method based on multimodal recognition as described in claim 4, characterized in that, Step S2.1 is as follows: First, we introduce a location-aware confidence-guided weighted mechanism: 1) Dynamically specify weighted information sources: The most reliable identification source is dynamically selected as the guiding source based on the target's real-time status and sensor characteristics; the decision rule is as follows: a. When the target is above the set altitude and above the set speed, the radar identification result is selected as the weighted information source first; b. When the target is below the set altitude and speed and the ambient light intensity is greater than the set value, the photoelectric result is selected as the weighted information source. c. When a specific communication protocol signal is detected, the communication reconnaissance results should be given priority as the weighted information source; 2) Extracting guidance information: Extract the target type proposition A_max with the highest confidence from the basic probability assignment function m_guide of the selected weighted information sources; Here, m_guide is a function that provides a framework for recognition. Each proposition in the equation is assigned a probability mass, representing the degree to which the information source supports the proposition being true; argmax is a mathematical operator that aims to find the parameter that maximizes the value of the subsequent function. This indicates traversing all possible propositions in the framework; 3) Probability-weighted adjustment: Guided by A_max, the basic probability allocation functions of other recognition sources are enhanced in a targeted manner. For the recognition source m_i to be adjusted, its adjusted basic probability allocation function m_i' is calculated according to the following rules: The guiding proposition is enhanced: Decrease other propositions: Where m_i: the basic probability allocation function of another information source to be adjusted, representing the original support of the information source for various propositions; A_max: the proposition with the highest confidence extracted from the most trusted guiding source m_guide; α: weighting coefficient; B is all other propositions not equal to A_max, and m_i(B) is the original support of the information source for other non-guiding propositions B.

6. The UAV localization method based on multimodal recognition as described in claim 5, characterized in that, Step S2.2, DS Evidence Theory Fusion and Decision Judgment, is as follows: The weighted basic probability allocation functions m_guide,m1',m2',...,mn' are fused using the DS combination rule: S2.2.

1. Use the DS combination formula to perform pairwise recursive fusion to obtain the fused basic allocation probability m_final; The following is a recursive fusion of DS combination rules: 1) DS combination rule formula Suppose we want to merge two basic probability assignment functions m1 and m2, and the new probability assignment function m generated after their fusion is... 12 Defined by the following formula: For all ; in, : This means finding all propositions A from m1 and propositions B from m2 whose intersection is exactly equal to proposition C; : satisfy all The sum of the products of m1(A) and m2(B) represents the total probability mass assigned to proposition C before normalization. Conflict coefficient K: K calculates the sum of the probability-mass products of all perfectly conflicting proposition pairs; K=0 indicates that there is no conflict between the two sources of evidence; the larger K is, the more serious the conflict between the sources of evidence. Normalization factor (1 / (1-K)): Its function is to redistribute the probability mass K lost due to conflicting evidence, ensuring... ; 2) Recursive fusion process When there are multiple sources of evidence, recursive fusion is performed: Step 1: Combine m_guide and m1' to obtain the fused result m_fused1; m_fused1 = m_guide ⊕ m1' Step 2: Combine m_fused1 with m2' to obtain a new m_fused2; m_fused2 = m_fused1 ⊕ m2' And so on, step N-1: Combine the result of step N-2 with the last source of evidence m. n The functions are combined to obtain the final fused basic probability allocation function m_final; S2.2.2 Calculate the trust function for each proposition based on m_final. Bel(A) represents the total confidence level in proposition A. This means summing all propositions B that are contained in A; S2.2.3 Select the proposition with the highest confidence level as the final target type identification result: Traversal recognition framework For all propositions in the given set, find the proposition A that maximizes Bel(A).

7. The UAV localization method based on multimodal recognition as described in claim 6, characterized in that, Step S2.3, optimization of the localization and tracking closed loop based on the recognition results, is as follows: The target type identification result T_final is fed back to the Kalman filter tracking stage in step S1 to achieve bidirectional enhancement of localization and identification: Motion model adaptation: Based on the identified target type, the key parameters in the state equation are adaptively adjusted: 1) If identified as a consumer-grade drone: adopt the standard uniform velocity model, and set the process noise covariance matrix to be less than the set value to reflect its relatively stable motion characteristics; 2) If identified as an attack drone: Increase the corresponding elements of the process noise covariance matrix based on the uniform velocity model to enhance the filter's ability to track sudden maneuvers; 3) If identified as a swarm of drones: an interactive multi-model algorithm is used to switch probabilities between multiple motion models; Observation model optimization: Optimize the observation noise covariance matrix based on the target type: 1) For small UAVs with small radar cross-sections and weak infrared characteristics, increase the noise variance of the corresponding observation channel; 2) For UAVs with typical communication signal characteristics, reduce observation uncertainty when the communication reconnaissance source provides accurate identification results.

8. A UAV positioning simulation system based on multimodal recognition, characterized in that: It includes a front-end interface module and a back-end algorithm module; among which, The front-end interface module is used to input simulated target parameters, radar system parameters, and infrared system parameters, and to call algorithms; and to display the output positioning and recognition results. The background algorithm module, based on the method described in any one of claims 1-7, is used to process the parameters passed from the front end and feed them back to the front end interface module.

Images