High-precision target positioning method for UAV based on attitude adaptive perception

Through a multi-task deep learning model based on attitude adaptive perception, the system errors in drone target positioning are identified and corrected, and the problem of drone positioning accuracy is solved, and the high-precision target positioning effect is achieved.

CN118089695BActive Publication Date: 2024-11-22NANJING UNIV OF AERONAUTICS & ASTRONAUTICS +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410087921.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-22
Publication Date
2024-11-22
Estimated Expiration
2044-01-22

AI Technical Summary

Technical Problem

The system errors of various attitude angles in the target positioning of drones are difficult to eliminate, resulting in a decrease in positioning accuracy and it is difficult for the existing technology to effectively identify and correct.

Method used

A multi-task deep learning model based on pose adaptive perception, train the model through the target positioning data set to identify and correct the system errors of the yaw angle, pitch angle, roll angle, and azimuth angle of the photoelectric reconnaissance platform. , achieve high-precision target positioning.

Benefits of technology

Effectively eliminate 77% of the system error, the target positioning error is reduced from 103m to 19m, and the positioning accuracy is increased by 81%, which significantly improves the target positioning accuracy of the drone in the cruise state.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118089695B_ABST
    Figure CN118089695B_ABST
Patent Text Reader

Abstract

The present invention provides a high-precision target positioning method for unmanned aerial vehicles based on attitude adaptive perception. The attitude adaptive perception model is trained. After the unmanned aerial vehicle is launched, a cooperative target point near the route is selected, and the position of the cooperative target point is solved using the optoelectronic reconnaissance platform positioning model. The target position information sequence obtained by the solution and the navigation data and attitude data sequence of the unmanned aerial vehicle itself are substituted into the attitude adaptive perception model to obtain the attitude adaptive correction signal of this unmanned aerial vehicle. After the unmanned aerial vehicle flies to the predetermined reconnaissance area, the attitude signals of the carrier aircraft and the optoelectronic reconnaissance platform are corrected: the experimental results show that the system effectively identifies the system errors of the yaw angle, pitch angle, roll angle of the unmanned aerial vehicle and the azimuth angle and elevation angle of the optoelectronic reconnaissance platform. When the unmanned aerial vehicle is at a flight altitude of 4000m, after calibration by this method, 77% of the system errors can be effectively eliminated, the error of target positioning is reduced from 103m to 19m, and the positioning accuracy is improved by 81%.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and in particular to a high-precision target positioning method for unmanned aerial vehicles based on attitude adaptive perception. Background Art

[0002] Drones carry optoelectronic reconnaissance platforms to conduct reconnaissance of important targets, which plays a vital role in strike guidance and battlefield situation awareness in modern warfare. They are also widely used in agricultural irrigation, forest fire prevention and other fields. Drone reconnaissance can effectively avoid casualties and has the characteristics of low cost, easy deployment and recovery. Among them, the airborne optoelectronic reconnaissance platform overcomes the problem of weak real-time performance of space-based high-orbit reconnaissance, and also gets rid of the interference of land-based and ship-based platforms due to terrain and distance factors. It can provide real-time and reliable target information and cooperate with other airborne systems to complete precise strikes and targeted defense of targets. With the development of optoelectronic sensors and servo control and the continuous improvement of the integration level of optoelectronic reconnaissance platforms, higher requirements are placed on the accuracy of the three-dimensional coordinates of the target, namely latitude, longitude and altitude information. High-precision target positioning has been widely valued by various countries.

[0003] The airborne optoelectronic reconnaissance platform of the UAV is usually a turret structure, consisting of a base, azimuth and pitch frames, an imaging system, and a control system. The process of obtaining the precise geographic coordinate information of the target during the reconnaissance process is called target positioning. According to the position information of the carrier, the distance and attitude angle of the reconnaissance target, combined with the rotation angle of the airborne platform and the internal and external parameters of the camera, the target is converted from the camera coordinate system to the earth coordinate system.

[0004] The positioning of the target by the airborne optoelectronic reconnaissance platform is the result of the coordinated work of multiple systems of the UAV. The positioning accuracy is jointly affected by multiple systems and multiple factors such as sensor accuracy, laser rangefinder accuracy, stability accuracy of the carrier and platform, navigation accuracy of the carrier, and data transmission delay. Any deviation in the positioning process will lead to serious deviation of the positioning result, and it is very difficult to improve the accuracy of target positioning. Among them, the influence of factors such as laser rangefinder accuracy and carrier navigation accuracy on the target positioning result shows the characteristics of linear translation. The various angle errors in the target positioning system are the key factors affecting the positioning accuracy of the UAV target. Angle errors can be divided into two categories: random errors and systematic errors. Optoelectronic reconnaissance platforms usually use damping shock absorbers to isolate the high-frequency vibration of the carrier, and use adaptive control to reduce the influence of mechanical vibration and low-frequency disturbances caused by target tracking. Most of the noise effects can be eliminated through filtering algorithms, but the elimination of systematic errors in attitude angles is still difficult.

[0005] In order to deal with the problem of reduced accuracy caused by the difficulty in eliminating the system errors of various attitude angles in UAV target positioning, it is necessary to propose a method for identifying and correcting system errors to achieve high-precision target positioning. To this end, the present invention proposes a method for high-precision target positioning of UAVs based on attitude adaptive perception to improve the accuracy of UAV target positioning. Summary of the invention

[0006] Purpose of the invention: In view of the above-mentioned prior art, a high-precision target positioning method for UAV based on attitude adaptive perception is proposed;

[0007] Technical solution: A high-precision target positioning method for UAV based on attitude adaptive perception, including the following steps:

[0008] Step 1): Obtain target positioning data set;

[0009] Step 2): construct a posture adaptive perception model, train the posture adaptive perception model through the target positioning data set, and obtain a trained posture adaptive perception model;

[0010] Step 3): Establish an optoelectronic reconnaissance platform positioning model;

[0011] Step 4): Use the optoelectronic reconnaissance platform positioning model to continuously locate the cooperative target point and perform position calculation. The calculated data is transmitted to the trained attitude adaptive perception model to obtain the attitude adaptive correction signal of the UAV. After the UAV flies to the predetermined reconnaissance area, the attitude signals of the carrier and the optoelectronic reconnaissance platform are corrected.

[0012] Preferably, the specific process of step 1) is: define C = {c1, c2, ..., c i ,…,c m} is the set of continuous target positioning events obtained by the UAV performing a flight along all possible trajectories, where c i represents the continuous target positioning event obtained by the UAV flying along the track i, and m represents the total number of tracks; continuous target positioning event in represents the jth positioning of the cooperative target point during a flight along trajectory i, s i Represents the positioning of the real target point, n i Represents continuous flight through n at the starting point i Second, locate the cooperation target point;

[0013] definition Locate events c for continuous targets i The characteristic matrix of is the matrix transpose, the characteristic matrix represent The characteristic vector of , the parameter j takes values ​​from 0 to n i -1, They represent the relative position offset of the result of position solution in the x, y, and z axes in the geographic coordinate system n. All are posture angles, specifically, is the attitude angle of the drone's yaw angle, is the attitude angle of pitch angle, is the attitude angle of the roll angle, The attitude angle of the camera of the optoelectronic reconnaissance platform, is the attitude angle of the high and low angles;

[0014] Construct a target positioning data set, which includes the real-time position of the drone and the relative position offset. and posture angle The target positioning dataset is divided into training set, validation set and test set with a ratio of 3:1:1.

[0015] Preferably, in step 2), the posture adaptive perception model includes a time series feature extractor and a weight regulator Target-MA connected in series, the time series feature extractor includes a classifier and a regression predictor; the classifier is used to determine the type Class of the dominant posture angle out , the regression predictor is used to determine the attitude angle offset of the drone Regress out , the attitude angle offset Regress out and the type of dominant posture angle Class out Input weight regulator adaptively adjusts weights:

[0016]

[0017]

[0018]

[0019]

[0020] Target-MA=Linear(Concat(Target-Head1,…,Target-Head i )W T )

[0021] Where Q i , K i , V i is a matrix, is the weight, represents the scaling factor, Head is the attention head, i is the sequence number of the attention head, Concat(·) is the concatenation of matrices, Linear(·) is the linear transformation; the dominant posture angles are five posture angles The attitude angle that has the greatest impact on target positioning;

[0022] The weight regulator outputs an attitude angle adaptive correction signal, and the measured relative position is adjusted according to the attitude angle adaptive correction signal. Perform reverse correction and offset the relative position after reverse correction and the measured relative position offset Substituting them into the constructed optoelectronic reconnaissance platform positioning model, two sets of positioning offsets are obtained:

[0023]

[0024]

[0025] Calculate the relative cost function:

[0026]

[0027]

[0028] in y represents Class out The true value label, represents the predicted value, K represents the number of consecutive positioning times, is the cross entropy loss function;

[0029] The final model cost function is

[0030] Use FrobeniusNorm to calculate the average distance as the positioning result offset:

[0031]

[0032] Where K represents the number of consecutive positioning, Batchsize represents the number of samples in each mini-batch, They represent the relative positions of the target solution results of the jth positioning in the i-th batch of samples in the x, y, and z axes in the geographic coordinate system;

[0033] Calculate the final evaluation index Target Increase :

[0034]

[0035] Preferably, in step 3), the optoelectronic reconnaissance platform positioning model constructed is:

[0036]

[0037] Among them, T c is the coordinate of the target in the camera coordinate system of the UAV optoelectronic reconnaissance platform, including the pixel position of the target in the x and y directions and the laser ranging distance. Transpose the matrix; and It is the conversion formula between the rectangular coordinate system and the geodetic coordinate system in the theory of geodesy; Provide the UAV navigation system with the geodetic coordinate information of the carrier; B p ,L p ,H p are the latitude, longitude and altitude of the drone respectively; φ nb ,γ nb ,θ nb ,α tc ,β tc They represent the actual measured values ​​of the drone's yaw angle, drone's pitch angle, drone's roll angle, the airborne reconnaissance platform's camera's azimuth angle, and the airborne reconnaissance platform's camera's elevation angle; The three-axis attitude angle deviation between the optoelectronic reconnaissance platform and the carrier aircraft introduced by the damping vibration isolator is the vibration error matrix; is the three-axis attitude angle offset introduced between the coordinate system of the carrier aircraft and the optoelectronic reconnaissance platform, where They are the three-axis installation alignment errors between the platform base and the carrier aircraft; Because the UAV is affected by strong airflow, initial alignment of the camera's visual axis and mechanical fatigue factors, the azimuth system error rotation matrix RotΔα and the elevation system error rotation matrix RotΔβ are introduced, where Δα and Δβ are the azimuth offset and elevation offset, respectively.

[0038] Preferably, after the UAV is launched, a cooperative target point near the flight path is first selected, and the optoelectronic reconnaissance platform positioning model is used to continuously locate the cooperative target point and perform position calculation, and the real-time position and relative position offset of the UAV obtained by the position calculation are calculated. and attitude angle offset Input the trained attitude adaptive perception model to obtain the attitude angle adaptive correction signal of this UAV After the UAV flies to the predetermined reconnaissance area, the attitude angles of the carrier and the optoelectronic reconnaissance platform are corrected:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044] in represents the measured attitude angle, Represents the attitude angle corrected by the attitude angle adaptive correction signal; Re-substitute the optoelectronic reconnaissance platform positioning model to solve and obtain the high-precision target position.

[0045] Beneficial effect: The advantages of the high-precision target positioning method of the UAV based on attitude adaptive perception of the present invention are: utilizing the time series offset characteristics of the continuous positioning results of the UAV for the same target, designing an attitude adaptive perception model based on multi-task deep learning to adaptively identify and compensate for the system errors of each angle, and realizing high-precision target positioning. The experimental results show that the system effectively identifies the system errors of the yaw angle, pitch angle, roll angle of the UAV and the azimuth angle, height angle of the optoelectronic reconnaissance platform. When the UAV is at a flight altitude of 4000m, after calibration by this method, 77% of the system errors can be effectively eliminated, the error of target positioning is reduced from 103m to 19m, and the positioning accuracy is improved by 81%, which proves that this method can effectively correct the system errors of the UAV in the cruising state and significantly improve the target positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of a method for high-precision target positioning of an unmanned aerial vehicle based on attitude adaptive perception in an embodiment of the present invention;

[0047] Figure 2 is a frame diagram of a posture adaptive perception model in an embodiment of the present invention;

[0048] Figure 3 It is a flow chart of the UAV target positioning model in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further explained below in conjunction with the accompanying drawings.

[0050] The innovation of this invention is to address the problem of low positioning accuracy caused by the difficulty in identifying and eliminating the errors of various attitude systems in drone target positioning. Based on the principle of target positioning, a full-process target position solution model is established, and the locking and tracking characteristics of the optoelectronic reconnaissance platform are used to measure the cooperative target points multiple times, and the difference in the positioning results caused by the multi-dimensional attitude system error offset is analyzed. Based on the deep neural network, the attitude adaptive perception model with multi-task timing characteristics is designed to adaptively estimate and reversely compensate for the errors of each system, and the corrected parameters are used to achieve high-precision target positioning. The effectiveness of this method is verified through simulation experiments, and a set of practical solutions are formulated as follows:

[0051] The high-precision target positioning method of the UAV based on attitude adaptive perception provided by the present invention and the high-precision target positioning process of the UAV optoelectronic reconnaissance platform are as follows: Figure 1 shown.

[0052] 1. Constructing a continuous target positioning data set using the simulated flight model of the present invention according to the reconnaissance scene and the relative height of the reconnaissance target;

[0053] 2. Continuously locate the cooperative target point. Using the drone position information (B p ,L p ,H p ), and introduce the UAV’s flight attitude information (φ nb ,γ nb ,θ nb ), the rotation angle information of the optoelectronic reconnaissance platform (α tc ,β tc ), the laser ranging distance r of the cooperative target point relative to the carrier, and the target pointing vector composed of the flight attitude information of the UAV and the rotation angle information of the optoelectronic reconnaissance platform, a series of transformations are performed on the pixel coordinates of the cooperative target point in the camera coordinate system to solve the position of the cooperative target point in the geodetic coordinate system (B t ,L t ,H t ).

[0054] 3. Take the relative coordinate offset Delta between the target positioning result and the real position data coor , using relative positions in geographic coordinates Replace absolute position (B p ,L p ,H p ) and (B t ,L t ,H t ) as the original input data of the attitude adaptive perception model. The attitude adaptive perception model is used to obtain the attitude system error correction signals of the current UAV target positioning system.

[0055] 4. The drone flies to the target reconnaissance area, and the drone navigation system obtains the drone's yaw angle Drone pitch angle and the drone's roll angle The azimuth angle of the camera rotation of the optoelectronic reconnaissance platform is obtained by the optoelectronic reconnaissance platform sensor and height angle Correct the attitude signals of the carrier aircraft and optoelectronic reconnaissance platform:

[0056]

[0057]

[0058] Finally, the optoelectronic reconnaissance platform positioning model of the present invention is used to complete high-precision positioning of the target, and the UAV returns.

[0059] Specific:

[0060] First, the high-precision target positioning method of UAV based on attitude adaptive perception is described as follows: after the UAV takes off, it first continuously locates the cooperative target point multiple times, corrects the attitude parameters of the current UAV based on the attitude adaptive perception model, and uses the adaptively corrected parameters to achieve high-precision positioning in the subsequent positioning of other real targets. In the design, the core idea is to use software for modeling and simulation analysis. In order to better split the structure and function of the system and improve the performance and reliability of the system, the design scheme should include the optoelectronic reconnaissance platform positioning model, the UAV simulation flight model, and the attitude adaptive perception model.

[0061] 1. When establishing the positioning model of the optoelectronic reconnaissance platform, it is first necessary to complete the conversion of the target pointing vector between various coordinate systems to realize the solution of the target position. The basis of position solution is to establish an auxiliary coordinate system to construct a positioning model and transform the target pointing vector. This article uses a total of six coordinate systems (all right-handed coordinate systems), namely the geodetic coordinate system K, the spatial geodetic rectangular coordinate system G, the geographic coordinate system N, the body coordinate system B, the optoelectronic reconnaissance platform base coordinate system T and the camera coordinate system C. Then the software is used to build the model and perform simulation analysis, and the rotation matrix of x degrees around the x, y, and z axes is defined as The following optoelectronic reconnaissance platform positioning model is established:

[0062]

[0063] Where, T c is the coordinate of the target in the camera coordinate system of the UAV optoelectronic reconnaissance platform, including the pixel position of the target in the x and y directions and the laser ranging distance. Transpose the matrix; and It is the conversion formula between G system and K system in geodesy theory; Provide the UAV navigation system with the geodetic coordinate information of the carrier; B p ,L p ,H p are the latitude, longitude and altitude of the drone respectively; φ nb ,γ nb ,θ nb ,α tc ,β tc They represent the actual measured values ​​of the drone's yaw angle, drone's pitch angle, drone's roll angle, the airborne reconnaissance platform's camera's azimuth angle, and the airborne reconnaissance platform's camera's elevation angle; The three-axis attitude angle offset between the optoelectronic reconnaissance platform and the carrier aircraft introduced by the damping shock absorber is the vibration error matrix, which is a random error. is the three-axis attitude angle offset introduced between the coordinate system of the carrier aircraft and the optoelectronic reconnaissance platform, that is, the three-axis system error matrix, where are the three-axis installation alignment errors between the platform base and the carrier, which is the system error; Because the UAV is disturbed by strong airflow, initial alignment of the camera's visual axis, and mechanical fatigue, the azimuth system error rotation matrix RotΔα and the elevation system error rotation matrix RotΔβ are introduced, where Δα and Δβ are the azimuth offset and elevation offset, respectively. This item is the system error, and the main error parameters are shown in Table 1:

[0064] Table 1. Main error parameters

[0065]

[0066] 2. When establishing a simulated flight model for a UAV, first select the reconnaissance altitude in the scenario based on the usage scenario of the type of UAV. A complete continuous positioning is defined as the UAV taking off from the airport, climbing to the cruising altitude, and diving at high speed in the direction of the reconnaissance target. When the camera of the optoelectronic reconnaissance platform reaches the radius area where the imaging inclination angle range is allowed by the error, it maintains a uniform speed and starts continuous target tracking and positioning until it leaves the radius area where the imaging inclination angle range is located. Actual reconnaissance cannot guarantee flying directly over the target (i.e. there is a relative yaw angle). There are multiple positioning routes under the same initial orientation, and the initial orientation of the UAV is uncertain. The relevant parameters are shown in Table 2.

[0067] Table 2. Related parameters of UAV simulation flight model

[0068]

[0069] 3. The attitude adaptive perception model is used to adaptively adjust the attitude system errors of each UAV target positioning system. The same UAV performs multiple consecutive positioning of the same target in the shooting range experiment. The offset between the positioning trajectory and the actual target point shows obvious regular characteristics. The simulation experiment verifies that the offset direction is affected by the UAV's flight attitude and relative position, and is strongly correlated with the system error of the multi-dimensional attitude. The miss amount is positively correlated with the size of the angular error.

[0070] 3.1 Construct a continuous target positioning dataset based on the reconnaissance scenario and define C = {c1, c2, …, c m} is the set of continuous target positioning events of a UAV along all possible trajectories, where c i represents the continuous target positioning of the i-th UAV, and m represents the total number of events. Each group of events is defined as Where n i Represents continuous flight through n at the starting point i The positioning of the cooperation target point, represents the jth positioning of the cooperative target point in this flight, and s represents the positioning of the real target point. The dimensions of the actual navigation data in the geodetic coordinate system are quite different in three dimensions. Considering the differences in positioning of different targets, the relative coordinate offset Delta is taken between the target positioning result and the real position data. coor , using relative positions in geographic coordinates Replace absolute position (B p ,L p ,H p ) and (B t ,L t ,H t ) as the original input data. Definition Locate events c for continuous targets i The characteristic matrix of Represents the target location of this flight The eigenvector of Represents the actual measured value of angle with various error amounts.

[0071] Repeat 50,000 simulated flights and record all data to build a data set, including the real-time position of the drone and various attitude angles. The target position offset characteristics measured by each drone continuous positioning task are quite different. The division of the data set will affect the performance of the model. A five-fold cross validation is used to reduce the contingency caused by random division to reduce experimental errors. The overall distribution ratio of the training set, validation set, and test set is set to 3:1:1.

[0072] 3.2 Feature analysis of positioning offset. The present invention extracts the offset features of the continuous positioning of the UAV to the cooperative target point based on the long short-term attention network (Long Short Time Memory, LSTM), and uses the multi-head attention module Target-MA to perform adaptive adjustment of the weights.

[0073] The present invention adopts a dual parallel LSTM structure as a time series feature extractor, which can extract the features of the previous moment and infer the features of the current and subsequent moments, embedding each node into a high-dimensional vector space. The regression predictor of the dual parallel structure is used to determine the offset of each posture Regress out The classifier is used to determine the main source of posture deviation during target positioning and obtain the output of positioning deviation feature Finally, the classifier is implemented through Linear Layer and LogSoftmax Layer to obtain the main source of system error of the current UAV target positioning system. The offset features of the trajectory sequence of the UAV positioning result are extracted through the classification model to obtain Class out Then use Linear Layer to reduce the dimension, and use LogSoftmax Layer to normalize the features: Finally, the cross entropy cost function is used to measure the accuracy of the classification task prediction, where y represents the true value label. Represents the predicted value of the model, and the cost function is defined as: Where K represents the number of consecutive positioning times, is the cross entropy loss function.

[0074] The present invention adopts a multi-head attention module as a weight regulator for local feature learning of target positioning offset, so as to better learn the offset effect of various characteristics of multi-dimensional posture on target positioning and realize adaptive adjustment of weights. The module splits and maps features into multiple subspaces, uses multiple independent attention heads to understand the positioning offset features generated under the influence of different relative yaw angles and system errors from different dimensions, and uses the output of regression prediction and classification tasks to generate the local feature of target positioning offset. out and Class out Multi-head attention is introduced to adaptively adjust the weights, thereby achieving adaptive correction of multi-dimensional posture angles. Each attention head and the overall definition of the module are

[0075]

[0076]

[0077]

[0078]

[0079] Target-MA=Linear(Concat(Target-Head1,…,Target-Head i )W T )

[0080] in K i The length of is the dimension of the hidden layer output, is the weight, num heads is the total number of attention heads, and i is the sequence number of the attention head.

[0081] The correction data of the Target-MA module reversely corrects the multi-dimensional attitude angle measured in continuous target positioning, and substitutes the attitude data before correction into the target positioning model to obtain two sets of positioning offsets:

[0082]

[0083]

[0084] The mean square error is used to calculate the positioning loss during continuous positioning measurement and the positioning loss after model prediction optimization, and the relative cost function of the regression task is obtained:

[0085]

[0086] The final model cost function is

[0087] The Frobenius Norm is used to calculate the average distance as the positioning result offset:

[0088]

[0089] Where K represents the number of consecutive positioning, and Batchsize represents the number of samples in each mini-batch.

[0090] The original offset of the drone's positioning of the real target point Target measure , the corrected high-precision positioning offset is Target predict , and finally obtain the improvement of positioning accuracy as the final evaluation index:

[0091]

[0092] 4. High-precision positioning of real target points

[0093] The attitude adaptive perception model is trained according to the current reconnaissance scene and the relative height of the reconnaissance target. After the UAV is launched, a cooperative target point near the route is first selected, and the optoelectronic reconnaissance platform positioning model is used to continuously track the cooperative target point and calculate the position. The target position information sequence obtained by the calculation and the UAV's own navigation data and attitude data sequence are substituted into the attitude adaptive perception model to obtain the attitude adaptive correction signal of this UAV. After the UAV flies to the predetermined reconnaissance area, the attitude signals of the carrier and the optoelectronic reconnaissance platform are corrected:

[0094]

[0095]

[0096] in Represents the actual measured attitude signal, Represents the adaptively corrected signal. After re-substituting into the optoelectronic reconnaissance platform positioning model, the target position can be calculated with high precision.

[0097] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A high-precision target positioning method for UAV based on attitude adaptive perception, characterized in that: The following steps are involved: Step 1): Obtain target positioning data set; Step 2): construct a posture adaptive perception model, train the posture adaptive perception model through the target positioning data set, and obtain a trained posture adaptive perception model; Step 3): Establish an optoelectronic reconnaissance platform positioning model; Step 4): Use the optoelectronic reconnaissance platform positioning model to continuously locate the cooperative target point and perform position calculation, and transmit the calculated data to the trained attitude adaptive perception model to obtain the attitude adaptive correction signal of the UAV. After the UAV flies to the predetermined reconnaissance area, the attitude signals of the carrier and the optoelectronic reconnaissance platform are corrected; The specific process of step 1) is: define C = {c1, c2, ..., c i ,…,c m } is the set of continuous target positioning events obtained by the UAV performing a flight along all possible trajectories, where c i represents the continuous target positioning events obtained by the UAV performing a flight along track i, and m represents the total number of tracks; Continuous Targeting Events where r i j represents the jth positioning of the cooperative target point during a flight along trajectory i, s i Represents the positioning of the real target point, n i Represents continuous flight through n at the starting point i Second, locate the cooperation target point; definition Locate events c for continuous targets i The characteristic matrix of is the matrix transpose, the characteristic matrix Represents r i j The characteristic vector of , the parameter j takes values ​​from 0 to n i -1, They represent the relative position offset of the result of position solution in the x, y, and z axes in the geographic coordinate system n. All are posture angles, specifically, is the attitude angle of the drone's yaw angle, is the attitude angle of pitch angle, is the attitude angle of the roll angle, The attitude angle of the camera of the optoelectronic reconnaissance platform, is the attitude angle of the high and low angles; Construct a target positioning data set, which includes the real-time position of the drone and the relative position offset. and posture angle The target positioning dataset is divided into training set, validation set and test set with a ratio of 3:1:1; In step 2), the posture adaptive perception model includes a time series feature extractor and a weight regulator Target-MA connected in series, and the time series feature extractor includes a classifier and a regression predictor; the classifier is used to determine the type Class of the dominant posture angle out , the regression predictor is used to determine the attitude angle offset of the drone Regress out , the attitude angle offset Regress out and the type of dominant posture angle Class out Input weight regulator adaptively adjusts weights: Q i =Class out W i Q K i =Regress out W i K V i =Regress out W i V Target-MA=Linear(Concat(Target-Head1,…,Target-Head i )W T ) Where Q i , K i , V i is a matrix, W i Q ,W i K ,W i V ,W T is the weight, represents the scaling factor, Head is the attention head, i is the sequence number of the attention head, Concat(·) is the concatenation of matrices, Linear(·) is the linear transformation; the dominant posture angles are five posture angles The attitude angle that has the greatest impact on target positioning; The weight regulator outputs an attitude angle adaptive correction signal, and the measured relative position is adjusted according to the attitude angle adaptive correction signal. Perform reverse correction and offset the relative position after reverse correction and the measured relative position offset Substituting them into the constructed optoelectronic reconnaissance platform positioning model, two sets of positioning offsets are obtained: Calculate the relative cost function: in y represents Class out The true value label of represents the predicted value, K represents the number of consecutive positioning times, is the cross entropy loss function; The final model cost function is The Frobenius Norm is used to calculate the average distance as the positioning result offset: Where K represents the number of consecutive positioning, Batchsize represents the number of samples in each mini-batch, They represent the relative positions of the target solution results of the jth positioning in the i-th batch of samples in the x, y, and z axes in the geographic coordinate system; Calculate the final evaluation index Target Increase : In step 3), the optoelectronic reconnaissance platform positioning model constructed is: Among them, T c is the coordinate of the target in the camera coordinate system of the UAV optoelectronic reconnaissance platform, including the pixel position of the target in the x and y directions and the laser ranging distance. Transpose the matrix; and It is the conversion formula between the rectangular coordinate system and the geodetic coordinate system in the theory of geodesy; Provide the UAV navigation system with the geodetic coordinate information of the carrier; B p ,L p ,H p are the latitude, longitude and altitude of the drone respectively; φ nb ,γ nb ,θ nb ,α ct ,β ct They represent the actual measured values ​​of the drone's yaw angle, drone's pitch angle, drone's roll angle, the airborne reconnaissance platform's camera's azimuth angle, and the airborne reconnaissance platform's camera's elevation angle; The three-axis attitude angle deviation between the optoelectronic reconnaissance platform and the carrier aircraft introduced by the damping vibration isolator is the vibration error matrix; is the three-axis attitude angle offset introduced between the coordinate system of the carrier aircraft and the optoelectronic reconnaissance platform, where They are the three-axis installation alignment errors between the platform base and the carrier aircraft; Because the UAV is affected by strong airflow, initial alignment of the camera's visual axis and mechanical fatigue factors, the system error rotation matrix RotΔα for azimuth and the system error rotation matrix RotΔβ for elevation are introduced, where Δα and Δβ are the azimuth offset and elevation offset, respectively.

2. The method for high-precision target positioning of an unmanned aerial vehicle based on attitude adaptive perception as claimed in claim 1, characterized in that: After the UAV is launched, a cooperative target point near the flight path is first selected, and the optoelectronic reconnaissance platform positioning model is used to continuously locate the cooperative target point and perform position calculation. The real-time position and relative position offset of the UAV obtained by the position calculation are calculated. and attitude angle offset Input the trained attitude adaptive perception model to obtain the attitude angle adaptive correction signal of this UAV After the UAV flies to the predetermined reconnaissance area, the attitude angles of the carrier and the optoelectronic reconnaissance platform are corrected: in represents the measured attitude angle, Represents the attitude angle corrected by the attitude angle adaptive correction signal; Re-substitute the optoelectronic reconnaissance platform positioning model to solve and obtain the high-precision target position.