Agricultural unmanned aerial pesticide application method based on real-time dynamic positioning technology

Abstract

The invention discloses an agricultural unmanned aerial pesticide application method based on a real-time dynamic positioning technology, and belongs to the technical field of unmanned aerial vehicles. The method comprises the following steps: 1, AI atmosphere prediction; 2, intelligent mode switching; and multi-source data fusion is fused into the two methods, and the multi-source data is preprocessed for time alignment, spatial interpolation and 3 sigma principle abnormal value elimination. Wherein in the RTK + IMU tight coupling mode, IMU mechanical arrangement data, RTK observation values, 15-dimensional state vector modeling poses and sensor errors are fused through Kalman filtering; and in a LiDAR SLAM mode, the point cloud is subjected to ground removal through RANSAC, and after feature extraction, a digital twin map is matched, so that GNSS-free environment centimeter-level positioning is realized. After the technologies are introduced into the unmanned aerial vehicle, the positioning error of the unmanned aerial vehicle in thunderstorm days can be greatly reduced, the positioning availability of the unmanned aerial vehicle in complex terrains covered by signals reaches 98% or above, and the working efficiency in complex regions can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of unmanned aerial vehicles, in particular to an agricultural unmanned aerial pesticide application method based on real-time dynamic positioning technology. BACKGROUND

[0002] At present, in the field of RTK high-precision navigation and positioning of unmanned aerial vehicles, a global navigation satellite system and real-time differential technology are combined, real-time data transmission and processing between a ground reference station and a flow station on the unmanned aerial vehicle are carried out, and centimeter-level or even millimeter-level positioning accuracy is achieved. However, the traditional RTK technology depends on a fixed reference station, atmospheric delay correction lags behind, initialization time is needed, and real-time response to weather changes cannot be achieved; and the unmanned aerial vehicle is prone to losing lock in the face of some complex farmland environments (trees / high-voltage lines). SUMMARY

[0003] The application aims to overcome the above technical deficiencies and provide a multi-source data fusion method for realizing the stability of unmanned aerial vehicle positioning signals in special weather and terrain.

[0004] The technical scheme adopted by the application is as follows: an agricultural unmanned aerial pesticide application method based on real-time dynamic positioning technology, comprising the following steps:

[0005] Step 1: Data collection and preprocessing: collect multi-source signal data, preprocess the above data, and classify the processed data by label;

[0006] Regarding data collection, the Beidou observation value is used to provide basic positioning data; the Fengyun-4 meteorological satellite data is accessed through a 5G cloud platform; the ground reference station network uploads temperature, humidity and pressure data in real time; and the unmanned aerial vehicle end multi-sensor is locally compensated.

[0007] Regarding data preprocessing, time alignment is adopted to unify the multi-source data to the GPS timestamp; spatial interpolation is adopted to interpolate the meteorological data to the unmanned aerial vehicle flight airspace; abnormal values are removed by using the 3 sigma principle to filter error data; and finally the preprocessed data is classified by label: normal weather, rainy weather, thunderstorm weather and thunderstorm weather.

[0008] Step 2: The processed data is used to generate a model through an ai atmospheric prediction algorithm, a double-channel LSTM structure algorithm is used, a loss function is combined to output a correction value for the model prediction error, and atmospheric mutation is predicted.

[0009] Regarding weather prediction, the LSTM prediction model is combined with a loss function, wherein a double-channel model is constructed: channel 1, ionospheric delay prediction; channel 2, tropospheric delay prediction; a double-channel LSTM structure algorithm is used, and TEC value + solar activity index; temperature / humidity / pressure / elevation are input respectively; and ionospheric delay and tropospheric delay are output.

[0010] After ionospheric and tropospheric delay is derived, combine the error weight of two channels, weighted balance ionospheric / tropospheric prediction error, and constrain model complexity with L2 regularization, list the following loss function:

[0011] L = a·RMSE ion + b·RMSE trop + g·||W||2

[0012] In the above formula, L is the total loss function value, used to reflect the degree of model training optimization; a is the ionospheric error weight, the value is 0.6; b is the tropospheric error weight, the value is 0.4; g is the regularization coefficient, the coefficient is 0.001, used to prevent model prediction overfitting; RMSE ion is the ionospheric prediction root mean square error; RMSE trop is the tropospheric prediction root mean square error, and ||W||2 is the L2 regularization term.

[0013] Step three: when the GNSS signal is lost due to the complex environment of farmland, the mode switching logic of the unmanned aerial vehicle is used to judge the environment of the unmanned aerial vehicle and adjust the mode; when the signal is stable, it is divided into an open environment, the atmospheric prediction of step two is carried out, and the pure RTK mode is maintained; when the lock is lost for many times but the time of each lock loss is less than the set value, it is divided into a lightly shaded area, the RTk+IMU tightly coupled mode is used for the situation of branch and leaf shading; if the GNSS signal lock loss time exceeds 1s, it is divided into a heavily shaded area, and the LiDAR SLAM mode is used;

[0014] The unmanned aerial vehicle end continuously monitors the GNSS signal and judges the signal quality in real time, judges the environmental conditions according to the signal quality, and controls the flight of the unmanned aerial vehicle by adopting the corresponding mode; if it is adjusted to the GNSS signal recovery, the initial flight mode is maintained and the GNSS signal is continuously monitored; if the GNSS signal is not recovered, the LiDAR SLAM mode is adopted, and the positioning output is continuously maintained until the signal is recovered;

[0015] Monitoring parameters: carrier-to-noise ratio CNO, as a trigger state switching signal; number of satellites, total number of GPS+Beidou visible satellites; PDOP, position accuracy factor; the three are used to judge the shielding classification of the unmanned aerial vehicle;

[0016] Table 1: Environment classification decision table

[0017] environmental state decision condition positioning mode open environment CNO > 40 dB-Hz & number of satellites ≥ 6 pure RTK mode light occlusion 35 < CN0≤ 40 dB-Hz & 4 ≤ number of satellites < 6 RTK + IMU tight coupling heavy occlusion CN0 < 35 dB-Hz or number of satellites < 4 LiDAR SLAM

[0018] The mode switching logic algorithm is used to determine the environment state and the mode of use of the real-time UAV: when the monitored carrier-to-noise ratio of the UAV is greater than 40 dB-Hz and the number of satellites that can be connected by the UAV is not less than 6, it is determined that the UAV is in an open signal good environment, and the UAV uses the pure RTK mode; when the monitored carrier-to-noise ratio of the UAV is greater than 35 dB-Hz and not more than 40 dB-Hz and the number of connected satellites is not less than 4, it is determined that the UAV is in an environment that will be slightly blocked, and the RTK+IMU tight coupling mode is used; when the monitored carrier-to-noise ratio of the UAV is less than 35 dB-Hz and the number of connected satellites is less than 4, the environment is divided into a heavy shielding category, and the UAV uses the LiDAR SLAM mode.

[0019] Step four, pre-processing the IMU raw data, including zero offset compensation and scale factor compensation, Kalman filter prediction, fusion of RTK positioning data for RTK observation value constraint, outputting the UAV pose;

[0020] The core idea of the RTK+IMU tight coupling mode is to obtain the optimal pose output by fusing high-frequency IMU data and low-frequency high-precision RTK observation values through a Kalman filter;

[0021] The IMU data is pre-processed, and the formula is as follows:

[0022] Zero offset correction ω corr = ω raw - b g - ω bias

[0023] Scale factor compensation a corr = a raw × s a + b a

[0024] Where b g is the gyro zero offset; s a is the acceleration scale factor, ω corr , ω raw , ω bias are the correction value, initial value and error value of the flight angle of the UAV, a corr , a raw are the correction value and initial value of the flight acceleration of the UAV;

[0025] The data after zero offset correction and scale factor compensation is mechanically arranged and calculated, and the purpose is to convert the body coordinate system into the navigation coordinate system. The following is the data processing calculation formula:

[0026] C_bn = quat2rotm(q)

[0027] a_nav = C_bn * a_corr + g

[0028] v_new = v + a_nav * dt

[0029] p_new = p + v * dt + 0.5 * a_nav * dt**2

[0030] q_new = integrate_gyro(q, omega_corr, dt)

[0031] wherein in the formula, C_bn is a body-to-navigation rotation matrix, q: quaternion attitude, v: UAV flight speed, p: UAV position in the navigation coordinate system, a_nav is the converted acceleration, v_new is the updated speed, dt is the time interval between two measurements of the UAV, g is the gravity acceleration, p_new is the updated position, and q_new is the updated UAV attitude;

[0032] By inputting the current attitude, speed, position, and time interval of the UAV into the formula, the IMU data is processed, and the IMU motion prediction is performed in combination with the state vector.

[0033] The RTK observation equation inputs the RTK positioning result as an observation value into the Kalman filter to calculate the optimal UAV pose output; wherein the RTK observation equation is as follows:

[0034] Z RTK = H * X + v, v ~ N(0, R)

[0035] wherein Z RTK is the observation residual, the difference between the RTK positioning value and the observation value; H is the observation matrix; v is the observation noise; R is the noise covariance; and X is the state vector, and the specific formula is as follows:

[0036]

[0037] wherein δ p , δ v , and δ θ represent the position, speed, and attitude errors of the UAV, respectively, and describe the motion state error of the UAV, b g , and b a are the zero bias errors of the gyroscope and the accelerometer, which are the key error sources of the modeled IMU, i.e., the zero bias of the UAV gyroscope and accelerometer; and are three-dimensional vectors; and the 15-dimensional state vector completely describes the UAV pose error and the IMU sensor bias, and is the core of the Kalman filter;

[0038] When GNSS signals are available, the sensor fusion is realized by the above formula, which associates the position error with the state vector through the observation matrix, compensates the IMU accumulated error in feedback, and outputs the correction.

[0039] Step five, if the GNSS signal is still not restored after positioning and flight control of the unmanned aerial vehicle through pure RTK or RTK+IMU tight coupling mode, the LiDAR SLAM mode is entered. The obtained point cloud data is preprocessed, features are extracted, and matching with the digital twin map is attempted. If there is a digital twin map matching, the pose is calculated, if not, the real-time mapping is performed, and then the pose is calculated, and finally the positioning output is obtained.

[0040] The mode first preprocesses the point cloud data, including ground point removal by RANSAC algorithm, downsampling, and motion distortion compensation. Then, corner features and plane features are extracted. It is judged whether there is map matching. If there is map matching, the pose is calculated. If there is no digital twin map matching, real-time mapping is performed, and then the pose is calculated. Finally, the positioning output is obtained.

[0041] The beneficial effects of the present application are: weather mutation adaptability breakthrough, ionospheric + tropospheric delay joint correction is carried out by using a double-channel LSTM atmospheric prediction model, combined with real-time meteorological satellite data and ground observation station information, the positioning error under thunderstorm weather is greatly reduced compared with traditional RTK, and the pesticide spraying overlap rate is reduced. The active prediction mechanism shortens the system response time to sudden weather changes to within 5 seconds. Continuous operation guarantee in all terrains, unique GNSS signal quality three-level grading strategy, seamless switching technology between RTK+IMU tight coupling and LiDAR SLAM, maintains centimeter-level positioning accuracy and normal work in the fruit tree canopy shielding environment, and ensures continuous and stable work of unmanned aerial vehicles in various terrains. BRIEF DESCRIPTION OF DRAWINGS

[0042] Fig. 1 The flowchart for constructing the atmospheric picture prediction model is shown in the figure.

[0043] Fig. 2 The anti-shielding flowchart for multi-source data fusion is shown in the figure.

[0044] Fig. 3 The RTK+IMU tight coupling work flowchart is shown in the figure.

[0045] Fig. 4 The LiDAR SLAM work flowchart is shown in the figure. DETAILED DESCRIPTION

[0046] Collect multi-source signal data: Beidou observation value provides basic positioning data; access Fengyun-4 meteorological satellite data through 5G cloud platform; real-time upload temperature, humidity and pressure data from ground reference station network; local compensation of multi-sensor on unmanned aerial vehicle side.

[0047] The above data is pre-processed: time alignment, unifying multi-source data to GPS timestamp; spatial interpolation, interpolating meteorological data to the UAV flight airspace; outlier rejection, filtering error data using the 3σ principle. The pre-processed data is classified by label: normal weather, rainy weather, thunderstorm weather, and thunderstorm weather.

[0048] The processed data generates a model through an ai atmospheric prediction algorithm and its evaluation

[0049] Core algorithm: LSTM prediction model

[0050] Channel 1: Ionospheric delay prediction

[0051] Channel 2: Tropospheric delay prediction

[0052] Using a dual-channel LSTM structure algorithm, input: TEC value + solar activity index; temperature / humidity / pressure / elevation. Output: [ionospheric delay, tropospheric delay]

[0053] After deriving the ionospheric and tropospheric delays, combine the error weights of the two channels to balance the ionospheric / tropospheric prediction error, and use L2 regularization to constrain the model complexity, as follows:

[0054] L = a · RMSE ion + b · RMSE trop + g · ||W||2

[0055] In the above formula, L is the total loss function value, used to reflect the optimization degree of the model training; a is the ionospheric error weight, its value is 0.6; b is the tropospheric error weight, its value is 0.4; g is the regularization coefficient, the coefficient is 0.001, used to prevent model prediction overfitting; RMSE ion is the ionospheric prediction root mean square error; RMSE trop is the tropospheric prediction root mean square error, and ||W||2 is the L2 regularization term.

[0056] The above model can reduce the positioning error of the UAV in extreme weather, actively predict atmospheric mutations, and respond to sudden weather changes in a timely manner.

[0057] The UAV end is designed to continuously monitor GNSS signals and evaluate signal quality in real time, determine environmental conditions based on signal quality, and take responsive modes to control the flight of the UAV; if adjusted to GNSS signal recovery, maintain the original flight mode and continuously monitor GNSS signals; if the GNSS signal is not restored, use the LiDAR SLAM mode, continuously maintain the positioning output until the signal is restored.

[0058] Monitoring parameters: CNO (Carrier-to-Noise Ratio), a core indicator that serves as a trigger for state switching; Number of Satellites, the total number of visible GPS and Beidou satellites; PDOP, Position Dilution of Precision. The number of satellites and PDOP are used to determine the drone's obstruction classification.

[0059] Table 1: Environmental classification decision table

[0060] environmental state decision condition positioning mode open environment CNO > 40 dB-Hz & number of satellites ≥ 6 pure RTK mode light occlusion 35 < CN0≤ 40 dB-Hz & 4 ≤ number of satellites < 6 RTK + IMU tight coupling heavy occlusion CN0 < 35 dB-Hz or number of satellites < 4 LiDAR SLAM

[0061] A mode switching logic algorithm is used to determine the real-time drone's environmental status and operating mode: When the drone's carrier-to-noise ratio is greater than 40dB-Hz and the drone can connect to at least 6 satellites, the drone is considered to be in an environment with good open signals and uses pure RTK mode. When the drone's carrier-to-noise ratio is greater than 35dB-Hz but not more than 40dB-Hz and the number of connected satellites is at least 4, the drone is considered to be in an environment with slight obstruction and uses RTK+IMU tight coupling mode. When the drone's carrier-to-noise ratio is less than 35dB-Hz and the number of connected satellites is less than 4, the environment is classified as heavily obstructed and the drone uses LiDAR SLAM mode. The following describes the principles of RTK+IMU tight coupling mode and LiDAR SLAM mode.

[0062] The core idea of ​​the RTK+IMU tight coupling mode is to use the Kalman filter to fuse high-frequency IMU data with low-frequency and high-precision RTK observations to obtain the optimal posture output.

[0063] Specific steps:

[0064] Preprocess the IMU data using the following formula:

[0065] Bias correction ω corr =ω raw -b g -ω bias

[0066] Scale factor compensation a corr =a raw ×s a +b a

[0067] where b g is the gyro bias; s a is the acceleration scale factor, ω corr 、ω raw 、ω bias are the correction value, initial value and error value of the UAV flight angle respectively, a corr 、a raw are the correction value and initial value of the UAV flight acceleration respectively.

[0068] The data after zero offset correction and scale factor compensation is mechanically arranged and calculated, the purpose is to convert the body coordinate system into the navigation coordinate system. The following is the data processing calculation formula:

[0069] C_bn = quat2rotm(q)

[0070] a_nav = C_bn * a_corr + g

[0071] v_new = v + a_nav * dt

[0072] p_new = p + v * dt + 0.5 * a_nav * dt**2

[0073] q_new = integrate_gyro(q, omega_corr, dt)

[0074] In the formula, C_bn is the body to navigation system rotation matrix, q: quaternion attitude, v: UAV flight speed, p: UAV position in the navigation coordinate system, a_nav is the converted acceleration, v_new is the updated speed, dt is the time interval between two measurements of the UAV, g is the gravity acceleration, p_new is the updated position, q_new is the updated UAV attitude.

[0075] By inputting the current attitude, speed, position, time interval of the UAV into the formula, the IMU data processing is completed, and the state vector is combined to predict the IMU motion.

[0076] The RTK observation equation inputs the RTK positioning result as the observation value into the Kalman filter, and calculates the optimal UAV pose output. The RTK observation equation is as follows:

[0077] Z RTK = H * X + v, v ~ N(0, R)

[0078] Where, Z RTK is the observation residual, the difference between the RTK positioning value and the observation value; H is the observation matrix; v is the observation noise; R is the noise covariance; X is the state vector The specific formula is:

[0079]

[0080] Where δ p , δ v , δ θ represent the position, speed and attitude errors of the UAV respectively, which describe the motion state error of the UAV, b g , b aThe zero bias errors of the gyroscope and the accelerometer are the key error sources of the modeling of the IMU, i.e., the zero biases of the gyroscope and the accelerometer of the UAV. Both are three-dimensional vectors. The 15-dimensional state vector completely describes the UAV pose error and the IMU sensor bias, and is the core of the Kalman filter.

[0081] When the GNSS signal is available, the observation matrix is associated with the position error and the state vector by the above formula, the sensor fusion is realized, the IMU accumulated error is feedback compensated, and the output is corrected.

[0082] If the GNSS signal is still not restored after the pure RTK or RTK+IMU tight coupling mode is used for the positioning and flight control of the UAV, the LiDAR SLAM mode is entered.

[0083] This mode first preprocesses the point cloud data, which includes ground point removal by using the RANSAC algorithm, downsampling, and motion distortion compensation; then corner feature and plane feature extraction are performed; it is determined whether there is map matching, if there is map matching, the pose is settled, if there is no digital twin map matching, real-time mapping is performed, and then the pose is settled; finally, the positioning output is obtained.

[0084] Through intelligent mode switching and deep fusion algorithm, the system ensures that the positioning accuracy is still centimeter level in the environment with serious fruit tree canopy shielding, and solves the core pain point of the agricultural UAV operation in complex terrain.

Claims

1. A method for agricultural unmanned aerial pesticide application based on real-time dynamic positioning technology, characterized by: Step 1: Data collection and preprocessing: Collect signal data from multiple sources, preprocess the above data, and label and classify the processed data; Step 2: The processed data is used to generate a model through the AI ​​atmospheric prediction algorithm. The dual-channel LSTM structure algorithm is used in combination with the loss function to output a correction value for the model prediction error and predict atmospheric mutations. Step 3: When the GNSS signal is lost due to the complex farmland environment, the drone's mode switching logic is used to determine the drone's environment and adjust the mode. When the signal is stable, it is classified as an open environment, and the atmospheric prediction in step 2 is performed while maintaining the pure RTK mode. If the lock is lost multiple times but the lock loss time is less than the set value, it is classified as light obscuration. If the GNSS signal is blocked by branches and leaves, the RTK+IMU tight coupling mode is used. If the GNSS signal is lost for more than 1 second, it is classified as a heavy obscuration area and the LiDAR SLAM mode is used. Step 4: Preprocess the IMU raw data, including zero bias compensation and scale factor compensation, perform Kalman filter prediction, integrate RTK positioning data to constrain RTK observation values, and output the UAV pose; Step 5: If the GNSS signal has not recovered after positioning and controlling the drone in pure RTK or RTK+IMU tightly coupled mode, the system will enter LiDAR SLAM mode. The acquired point cloud data will be preprocessed, features will be extracted, and then matched with the digital twin map. If a match is found on the digital twin map, pose settlement will be performed. If not, real-time mapping will be used, pose settlement will be performed, and finally positioning output will be generated.

2. The method for agricultural unmanned aerial pesticide application based on real-time dynamic positioning technology according to claim 1, characterized in that: In step 1, regarding data collection, Beidou observation values ​​are used to provide basic positioning data; Fengyun-4 meteorological satellite data is accessed through the 5G cloud platform; the ground base station network uploads temperature, humidity, and air pressure data in real time; and local compensation of multiple sensors on the drone end is performed.

3. The method for agricultural unmanned aerial pesticide application based on real-time dynamic positioning technology according to claim 1, characterized in that: In the step 1, regarding the data preprocessing, time alignment is used to unify the multi-source data into the GPS timestamp; spatial interpolation is used to interpolate the meteorological data into the UAV flight airspace; Outliers are eliminated and erroneous data are filtered using the 3σ principle; finally, the preprocessed data is labeled and classified as normal weather, rainy weather, thunderstorm weather, and thunderstorm weather.

4. The method for agricultural unmanned aerial pesticide application based on real-time dynamic positioning technology according to claim 1, characterized in that: In step 2, the weather forecast is based on an LSTM prediction model combined with a loss function. The model is constructed with two channels: channel 1, ionospheric delay prediction; channel 2, tropospheric delay prediction. The dual-channel LSTM structure algorithm is used to input the following: TEC value + solar activity index; temperature / humidity / pressure / elevation; and output: ionospheric delay and tropospheric delay. After obtaining the ionospheric and tropospheric delays, the error weights of the two channels are combined to weightedly balance the ionospheric / tropospheric prediction errors. At the same time, L2 regularization is used to constrain the model complexity. The following loss function is listed: L=α·RMSE ion +β·RMSE trop +γ·||W||2 In the above formula, L is the total loss function value, which is used to reflect the degree of optimization of model training; α is the ionosphere error weight, whose value is 0.6; β is the troposphere error weight, whose value is 0.4; γ is the regularization coefficient, which is 0.001, used to prevent overfitting of model prediction; RMSE ion RMSE is the root mean square error of ionospheric prediction trop is the root mean square error of tropospheric prediction, and ||W||2 is the L2 regularization term.

5. The method for agricultural unmanned aerial pesticide application based on real-time dynamic positioning technology according to claim 1, characterized in that: In step 3, the drone is designed to continuously monitor GNSS signals and assess signal quality in real time. Based on the signal quality, it determines environmental conditions and adopts a responsive mode to control the flight of the drone. If the GNSS signal is restored, the drone maintains the initial flight mode and continues to monitor the GNSS signal. If the GNSS signal does not return, the drone adopts LiDAR SLAM mode and continues to output positioning until the signal is restored. Monitoring parameters: Carrier-to-noise ratio (CNO), used as a trigger for state switching; number of satellites, the total number of GPS+Beidou visible satellites; PDOP, position dilution of precision; these three parameters are used to determine the drone's occlusion classification; Table 1: Environmental classification decision table A mode switching logic algorithm is used to determine the real-time drone's environmental status and the mode it is using: when the drone's carrier-to-noise ratio is greater than 40dB-Hz and the number of satellites it can connect to is no less than 6, the drone is considered to be in an environment with good open signals and uses pure RTK mode; when the drone's carrier-to-noise ratio is greater than 35dB-Hz but not more than 40dB-Hz and the number of satellites connected is no less than 4, the drone is considered to be in an environment with slight obstruction and uses RTK+IMU tight coupling mode; when the drone's carrier-to-noise ratio is less than 35dB-Hz and the number of satellites connected is less than 4, the environment is classified as heavily obstructed and the drone uses LiDAR SLAM mode.

6. The method for agricultural unmanned aerial pesticide application based on real-time dynamic positioning technology according to claim 1, characterized in that: In step 4, the core idea of ​​the RTK+IMU tight coupling mode is to use a Kalman filter to fuse high-frequency IMU data with low-frequency, high-precision RTK observations to obtain the optimal pose output; Preprocess the IMU data using the following formula: Bias correction ω corr =ω raw -b g -ω bias Scale factor compensation a corr =a raw ×s a +b a where b g is the gyro bias; s a is the acceleration scale factor, ω corr 、ω raw 、ω bias are the correction value, initial value and error value of the UAV flight angle respectively, a corr 、a raw are the correction value and initial value of the UAV flight acceleration respectively; The data after zero bias correction and scale factor compensation is mechanically arranged and calculated to convert the body coordinate system into the navigation coordinate system. The following is the data processing calculation formula: C_bn=quat2rotm(q) a_nav=C_bn*a_corr+g v_new=v+a_nav*dt p_new=p+v*dt+0.5*a_nav*dt**2 q_new=integrate_gyro(q,omega_corr,dt) In the formula, C_bn is the rotation matrix from the body to the navigation system, q: quaternion attitude, v: drone flight speed, p: drone position in the navigation coordinate system, a_nav is the converted acceleration, v_new is the updated speed, dt is the time interval between two drone measurements, g is the gravitational acceleration, p_new is the updated position, and q_new is the updated drone attitude; By inputting the current attitude, speed, position, and time interval of the drone into the formula, the IMU data is processed and the IMU motion prediction is performed in combination with the state vector. The RTK observation equation uses the RTK positioning results as observation values ​​and inputs them into the Kalman filter to calculate the optimal posture output of the drone. The RTK observation equation is as follows: Z RTK =H·X+v,v~N(0,R) Among them, Z RTK is the observation residual, the difference between the RTK positioning value and the observation value; H is the observation matrix; v is the observation noise; R is the noise covariance; X is the state vector. The specific formula is: where δ p ,δ v ,δ θ Represent the errors of the position, speed and attitude of the UAV, describing the error of the UAV's motion state. g ,b a The gyroscope and accelerometer bias errors are the key error sources for modeling IMUs, i.e., the gyroscope and accelerometer biases of drones. Both are three-dimensional vectors. The 15-dimensional state vector fully describes the drone's posture error and IMU sensor bias, and is the core of the Kalman filter. When the GNSS signal is available, the above formula is used to associate the position error with the state vector through the observation matrix to achieve sensor fusion, feedback compensation IMU accumulated error, and output correction.

7. The method for agricultural unmanned aerial pesticide application based on real-time dynamic positioning technology according to claim 1, characterized in that: In step 5, the model first pre-processes the point cloud data, including ground point one using the RANSAC algorithm, downsampling, and motion distortion compensation; then performs corner feature and plane feature extraction; Determine whether there is a map match. If so, perform pose settlement. If there is no digital twin map match, perform real-time mapping and then perform pose settlement. Finally, position the output.

Images