An adaptive navigation system for unmanned aerial vehicles

Abstract

The present invention relates to the field of unmanned aerial vehicle control technology, and in particular to an adaptive navigation system for unmanned aerial vehicles, including an inertial navigation system, a visible light sensor, a data fusion module, a database and a navigation control module; the inertial navigation system provides attitude, speed and acceleration data; the visible light sensor collects images and processes them through a deep learning algorithm to obtain environmental feature points to achieve visual positioning; the data fusion module uses a dynamic weighted Kalman filter algorithm to fuse two sensor data, and dynamically adjusts the weights to optimize positioning information according to the environment. The map library and the feature library store environmental data to assist in positioning and dynamically update; the navigation control module generates instructions based on the fused positioning information, and automatically adjusts the flight strategy based on the adaptive control algorithm considering multiple factors. The system improves positioning accuracy and environmental adaptability, enhances flight safety and stability, has a high degree of intelligence and autonomy, and has flexibility and scalability, and can be widely used in the field of unmanned aerial vehicles.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle control, and in particular to an unmanned aerial vehicle adaptive navigation system. Background Art

[0002] As drone technology is booming, its applications in many fields (such as aerial photography, surveying and mapping, inspection, logistics and distribution, etc.) are becoming more and more widespread. However, the navigation and positioning accuracy and stability of drones have always been the key factors restricting their performance improvement and application expansion.

[0003] Traditional drone navigation systems mainly rely on GPS positioning, but GPS signals have problems such as signal shielding and multipath effects in some complex environments (such as urban high-rise areas, mountainous areas, forests, etc.), which leads to a significant decrease in positioning accuracy or even failure. At the same time, relying solely on GPS cannot meet the positioning requirements of drones in high-precision tasks (such as precise mapping, detailed inspections, etc.).

[0004] Although the inertial navigation system (INS) does not rely on external signals and can provide relatively stable attitude and position information in a short period of time, its errors will accumulate over time, and positioning accuracy is difficult to guarantee during long-term use.

[0005] Visible light sensors have unique advantages in acquiring environmental visual information and can achieve visual positioning through image recognition and processing. However, they are affected by lighting conditions, image noise, etc., and their positioning stability is poor when used alone.

[0006] In addition, with the diversification and complexity of UAV application scenarios, higher requirements are placed on their environmental adaptability, flight safety and intelligence, which are difficult to meet with traditional navigation systems. Therefore, an adaptive navigation system for UAV is proposed to address the above problems. Summary of the invention

[0007] The purpose of the present invention is to provide an unmanned aerial vehicle adaptive navigation system to solve the problems raised in the above-mentioned background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] An adaptive navigation system for unmanned aerial vehicles, comprising:

[0010] Inertial navigation system; the inertial navigation system provides the attitude, speed and acceleration data of the drone;

[0011] Visible light sensor; the visible light sensor includes at least one high-definition camera, which is used to collect environmental image information and obtain environmental feature points through image processing methods to achieve visual positioning;

[0012] Data fusion module: The data fusion module receives data from the inertial navigation system and the visible light sensor, and uses a dynamic weighted Kalman filter algorithm to optimize the positioning information, automatically increasing the weight of the visible light sensor in a dynamic environment and increasing the weight of the inertial navigation system in a static environment to obtain the fused positioning information;

[0013] Database: The database includes a map library and a feature library. The database stores environmental feature points and landmark data for real-time matching by visible light sensors to assist in drone positioning. The map library and feature library are dynamically updated based on environmental data obtained during the flight of the drone. The update algorithm is based on an incremental learning method.

[0014] Navigation control module: The navigation control module generates flight control instructions based on the fused positioning information to guide the UAV to fly stably. The navigation control module automatically adjusts the flight strategy based on the adaptive control algorithm.

[0015] As a preferred solution, the image processing method of the visible light sensor includes a deep learning algorithm. The deep learning algorithm performs real-time environmental feature extraction and matching through a convolutional neural network. Assume that the input image is , in, Represent the horizontal and vertical coordinates of the image pixels respectively, and the convolution kernel is , is the coordinate of the convolution kernel element, then the calculation formula of the convolution layer is: ,in, The feature map output by the convolutional layer is in coordinates The value at is a bias term. Through the combined operation of multi-layer convolution, pooling and fully connected layers, the environmental feature points are extracted and matched. The pooling operation adopts the maximum pooling formula: ,in, The output of the pooling layer is at coordinates The value at is the pooling area.

[0016] As a preferred solution, the dynamic weighted Kalman filter algorithm in the data fusion module is as follows:

[0017] Status prediction: ,in, For Always The predicted value of the state at the moment, for The best estimate of time, is the state transition matrix, which describes the system from Time has come The law of state change at every moment;

[0018] Covariance prediction: ,in, is the prediction covariance matrix, for The optimal estimated covariance matrix at time , is the process noise covariance matrix;

[0019] Kalman gain calculation: ,in, is the Kalman gain, which is used to balance the weights of prediction information and observation information in state updating. is the observation matrix, which maps the state vector to the observation vector space, is the observation noise covariance matrix, which represents the uncertainty of the observation data;

[0020] Status Update: ,in, yes The observed value at time, yes The best estimate of the time;

[0021] Covariance update: ,in, is the identity matrix, yes The optimal estimated covariance matrix at time;

[0022] Weighted processing in dynamic weighted Kalman filtering:

[0023] Assume that the estimated value of the inertial navigation system is , the covariance is , the estimated value of the visible light sensor is , the covariance is ;

[0024] Defining weights and , and satisfies , depending on the ambient light intensity , Image texture complexity And the change rate of the UAV's flight attitude Make real-time adjustments, and the weighted calculation formula is:

[0025] ,in, for , , The corresponding weight influence coefficient;

[0026] ;

[0027] The fused state estimate is: ;

[0028] The fused covariance estimate is: .

[0029] As a preferred solution, during the dynamic update process of the map library and feature library, the newly collected environmental feature point data is preliminarily screened using the clustering analysis method, and the feature library is updated and optimized after the noise data is removed; the feature point set is ,in, is the i-th environmental feature point, is the total number of feature points, and the K-means clustering algorithm is used, and its objective function is: ,in, is the number of categories into which the feature points are divided, For the A set of clusters, For the The center of the cluster, Represents the Euclidean distance. Through continuous iterative optimization, the objective function Minimum, the feature points with a set distance from the cluster center are regarded as noise points and removed, and the filtered feature points are added to the feature library for updating.

[0030] As a preferred solution, the navigation control module takes into account the remaining power of the drone when generating flight control instructions. , Communication signal strength and traffic conditions in the surrounding airspace , let the flight control command vector be , then the control instruction generation formula based on multiple factors is: ;in, , are the weight coefficients of different factors, and , is the control function related to the remaining power, is a control function related to the communication signal strength, It is a control function related to the traffic conditions in the surrounding airspace.

[0031] It can be seen from the technical solution provided by the present invention that the adaptive navigation system for unmanned aerial vehicles provided by the present invention has the following beneficial effects:

[0032] 1. By combining the inertial navigation system (INS) with the visible light sensor and using the dynamic weighted Kalman filter algorithm for data fusion, the advantages of the two sensors are fully utilized; INS can provide stable attitude, velocity and acceleration data in the short term, especially when the GPS signal is lost or the environment is less dynamic, it can provide a reliable positioning basis for the UAV; the visible light sensor obtains high-precision position information through visual positioning in scenes with rich environmental features (such as urban environments and areas with obvious landmarks); the dynamic weighting mechanism adjusts the sensor weight in real time according to the dynamic changes of the environment, increases the weight of the visible light sensor in a dynamic environment to obtain more accurate real-time position information, and increases the weight of the INS in a static environment to utilize its long-term stability, thereby significantly improving the positioning accuracy, enabling the UAV to achieve more accurate positioning in various complex environments;

[0033] 2. The map library and feature library store a wealth of environmental feature points and landmark data. The image information collected by the visible light sensor can be matched with the data in the library in real time to further assist positioning; this not only improves the accuracy of positioning, but also enhances the adaptability of the system in different scenarios. Even after flying in an unknown environment for a period of time, the positioning accuracy can be gradually improved by continuously updating the map library and feature library, providing a solid positioning guarantee for the stable flight and mission execution of the drone;

[0034] 3. In the dynamic weighted Kalman filter algorithm, the weight is adjusted in real time according to factors such as the change in environmental light intensity, the complexity of image texture, and the rate of change of the UAV's flight attitude. In dynamic environments with large changes in light intensity (such as sunrise and sunset, entering and exiting shadow areas), complex image textures (such as densely built-up areas in cities), or frequent changes in flight attitude (such as during maneuvering flight), the system can automatically increase the weight of the visible light sensor to better utilize visual information for positioning and navigation; in static environments with stable lighting, simple textures, and stable flight attitudes, the weight of the inertial navigation system is increased to ensure the stability of positioning. This adaptive weight adjustment strategy enables the navigation system to flexibly respond to various complex and changing environmental conditions, ensuring that the UAV can fly stably and accurately in different scenarios.

[0035] 4. High-precision positioning information provides an accurate basis for the flight control of the UAV. The navigation control module generates precise flight control instructions based on the integrated positioning information, enabling the UAV to fly more accurately along the planned route and reduce the risk of deviation from the route. At the same time, during the flight, the system can monitor the position and attitude of the UAV in real time, promptly detect and correct possible flight deviations, ensure the stability of the flight trajectory, effectively avoid flight accidents caused by positioning errors, and greatly improve flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1The present invention is a schematic diagram of the overall structure of an adaptive navigation system for unmanned aerial vehicles. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0038] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0039] like Figure 1 As shown, an embodiment of the present invention provides an adaptive navigation system for a UAV, including an inertial navigation system, a visible light sensor, a data fusion module, a database and a navigation control module.

[0040] In this embodiment, the inertial navigation system provides the attitude, speed and acceleration data of the drone. The operation of the inertial navigation system includes the following steps:

[0041] 1. Inertial Measurement Unit (IMU) Measurement

[0042] Accelerometer measurements:

[0043] Based on Newton's second law, the accelerometer measures acceleration by detecting the inertial force of the internal sensitive mass block under the action of acceleration. When the drone is stationary or moving in a uniform straight line, the acceleration measured by the accelerometer is mainly the component of gravity acceleration on its measurement axis, which can be used to determine the attitude angle of the drone (including pitch angle, yaw angle and roll angle). When the drone accelerates, the accelerometer will measure additional acceleration components related to the linear acceleration of the drone.

[0044] Gyroscope measurements:

[0045] The gyroscope uses the principle of conservation of angular momentum to detect angular velocity by measuring the precession of a rotating object in inertial space. When the attitude of the drone changes, the gyroscope will measure the corresponding angular velocity signal. Its measurement data can reflect the rotational motion state of the drone in real time and quickly track the attitude change actions of the drone, such as turning and rolling.

[0046] 2. Data Preprocessing

[0047] Accelerometer data preprocessing:

[0048] The raw data measured by the accelerometer is filtered to remove high-frequency noise interference. Low-pass filtering, Kalman filtering and other methods are commonly used. Low-pass filtering can smooth the data and reduce the measurement error caused by the noise and vibration of the sensor itself. Gravity compensation is then required because the acceleration measured by the accelerometer contains the gravity acceleration component. It is necessary to combine the attitude angle information measured by the gyroscope to calculate the gravity acceleration component on each coordinate axis, and then subtract the corresponding gravity acceleration component from the accelerometer measurement data to obtain the actual linear acceleration of the drone.

[0049] Gyroscope data preprocessing:

[0050] The original angular velocity data measured by the gyroscope is filtered to reduce the impact of noise on the data, improve data accuracy and stability, and make the filtered gyroscope data more accurately reflect the rate of change of the drone's attitude; if the sensor supports it, the gyroscope data needs to be temperature compensated. Since the performance of the gyroscope may be affected by temperature changes, the working temperature is measured by a temperature sensor, and the angular velocity data measured by the gyroscope is corrected according to the pre-calibrated temperature and error relationship curve to reduce the measurement error caused by temperature changes;

[0051] 3. Posture calculation

[0052] Quaternion-based attitude update:

[0053] Using the angular velocity data measured by the gyroscope, the attitude quaternion of the drone is updated through the quaternion differential equation. Quaternions are used to represent rotations in three-dimensional space. Compared with the traditional Euler angle representation, they have the advantages of avoiding universal joint lock and simple calculation. In the attitude update process, the time interval needs to be considered, and the quaternion differential equation is solved by the integral method to obtain the attitude quaternion at different times. After the attitude is updated, the quaternion needs to be normalized to ensure that it always meets the constraint of the unit quaternion, that is, the modulus of the quaternion is 1.

[0054] 4. Speed ​​and position calculation

[0055] Speed ​​calculation:

[0056] The linear acceleration data measured by the accelerometer after gravity compensation is integrated to obtain the velocity change of the UAV in each coordinate axis direction. In actual calculations, numerical integration methods are usually used to approximate the integral value, such as the trapezoidal integration method.

[0057] Position calculation:

[0058] The calculated speed data is integrated again to obtain the position change of the UAV in three-dimensional space. The position change is also calculated using the numerical integration method. By repeatedly calculating the speed and position, the inertial navigation system can provide the speed and position information of the UAV in real time.

[0059] 5. Data Output

[0060] After the above series of measurement, calculation and processing steps, the inertial navigation system outputs the final attitude (quaternion representation or converted to Euler angle representation), velocity (three-dimensional velocity vector) and acceleration (linear acceleration vector) data in a certain format (such as digital signal, data format specified by the communication protocol, etc.) to other system modules of the UAV, such as data fusion module, navigation control module, etc., so that these modules can use the data for further processing and decision-making.

[0061] In this embodiment, the visible light sensor includes at least one high-definition camera for collecting environmental image information and obtaining environmental feature points through image processing methods to achieve visual positioning. During specific operation, the visible light sensor:

[0062] 1. Image Acquisition

[0063] Camera startup and initialization:

[0064] When the drone system starts, the high-definition camera in the visible light sensor is initialized, and the camera's working mode, frame rate, resolution and other parameters are set; for example, the appropriate resolution is selected according to the task requirements. If high-precision image analysis is required, it can be set to a higher resolution such as 1080p or higher; the frame rate is determined according to the drone's flight speed and the dynamics of the environment. In high-speed flight or dynamic scenes, a higher frame rate (such as 60fps or higher) is selected to ensure the continuity and real-time nature of the image;

[0065] Image Capture:

[0066] The camera starts to capture the environment image according to the set parameters, focusing the light through the lens onto the image sensor (such as CMOS or CCD), and the image sensor converts the light signal into an electrical signal, and then obtains the digital image signal through analog-to-digital conversion (ADC); during the image capture process, the camera's autofocus function adjusts the lens focal length in real time to ensure a clear image; the autofocus system drives the lens motor to move the lens by analyzing information such as image contrast or phase difference, so that the subject is clearly imaged on the image sensor; at the same time, the optical image stabilization function (if any) continuously monitors the camera's motion state, and offsets the vibration and shaking of the drone during flight by compensating for the displacement of the lens or sensor, reducing image blur caused by motion;

[0067] 2. Image processing

[0068] Image preprocessing:

[0069] The collected digital images are preprocessed to improve the accuracy and efficiency of subsequent feature extraction. First, grayscale processing is performed (if the subsequent algorithm is based on grayscale images) to convert color images into grayscale images to reduce data dimensions and calculations. The grayscale formula is as follows: in, is the coordinate The gray value at are the red, green, and blue color channel values ​​of the corresponding coordinates respectively; then, image denoising can be performed by using a filtering algorithm, such as Gaussian filtering; Gaussian filtering kernel function ,By performing weighted average operation on each pixel in the image and its neighboring pixels, the image is smoothed, and high-frequency noise components such as Gaussian noise are removed, while retaining the edge and detail information of the image;

[0070] Feature point extraction:

[0071] A feature extraction algorithm is used to extract environmental feature points from the preprocessed image. For example, the Shi-Tomasi corner detection algorithm is used. The algorithm calculates the autocorrelation matrix of each pixel in the image and determines whether the point is a corner point according to the minimum eigenvalue of the autocorrelation matrix (corner points are usually points with significant changes in the image, such as the corners of buildings, intersections of roads, etc., which can be used as environmental feature points); if the minimum eigenvalue is greater than the set threshold, the point is marked as a feature point; in addition to the corner detection algorithm, other feature extraction algorithms can also be used, such as the scale-invariant feature transform (SIFT) algorithm or the speeded up robust features (SURF) algorithm. These algorithms can detect stable feature points at different scales and rotation angles, thereby improving the robustness of feature point extraction;

[0072] Features:

[0073] For the extracted feature points, their feature description vectors are calculated for subsequent feature matching. Taking the ORB (Oriented FAST and Rotated BRIEF) feature descriptor as an example, the FAST (Features from Accelerated Segment Test) algorithm is first used to determine the position of the feature point, and then the direction of the feature point is calculated based on the intensity distribution of the pixels around the feature point, so that the feature description has rotation invariance. Then, a certain size area is selected around the feature point, and a binary feature description vector is generated by comparing the intensity of the pixel pairs in the area. The feature description vector can effectively represent the unique information of the feature point, such as the texture and shape around the feature point, so as to perform feature matching between different images.

[0074] 3. Visual positioning

[0075] Feature Matching:

[0076] Match the feature point description vector extracted from the current frame image with the feature point description vector in the map library or the previous frame image; a brute force matching method can be used to calculate the distance (such as Euclidean distance or Hamming distance) between each feature point description vector in the current frame image and all feature point description vectors in the map library or the previous frame image, and select the feature point pair with the smallest distance as the matching pair; in order to improve the matching efficiency, an algorithm based on approximate nearest neighbor search can also be used, such as the FLANN (Fast Library for Approximate Nearest Neighbors) algorithm, which can quickly find feature point description vectors close to the current feature point description vector and reduce the amount of matching calculation; in the feature matching process, wrong matching may occur, so some strategies need to be adopted to eliminate wrong matching points; for example, using the RANSAC (Random Sample Consensus) algorithm, by randomly selecting sample points multiple times to calculate the model (such as homography matrix or basic matrix), the consistency of the sample points is evaluated according to the model, and inconsistent points (i.e., wrong matching points) are eliminated, and finally an accurate feature matching result is obtained;

[0077] Attitude estimation and positioning:

[0078] Based on the matched feature point pairs, the attitude (position and orientation) of the drone is calculated using geometric relationships. For example, the relative rotation and translation relationship between different images can be calculated through epipolar geometric relationships. According to the coordinates of the matched feature points in the two images, the essential matrix or basic matrix is ​​constructed, and then the matrix is ​​decomposed to obtain the rotation and translation information of the drone. Combined with the known landmark information or initial positioning information in the map library, the position of the drone in three-dimensional space is determined by triangulation. Triangulation uses multiple matched feature point pairs to calculate the distance from the drone to the feature point based on the principle of similar triangles, thereby determining the position coordinates of the drone. In the process of attitude estimation and positioning, it is necessary to continuously optimize the calculation results to improve the positioning accuracy. Iterative optimization algorithms, such as the least squares method or the BundleAdjustment algorithm, can be used to globally optimize the feature point matching and attitude estimation results to make the calculated drone attitude and position more accurate.

[0079] 4. Data output and update

[0080] Data output:

[0081] The drone posture (including position coordinates and direction angles) obtained by visual positioning and related environmental feature point information are output in a certain data format; the output data can be transmitted to other system modules of the drone, such as the data fusion module, and fused with the data of other sensors such as the inertial navigation system to improve the accuracy and reliability of drone positioning and navigation; at the same time, some data can also be stored in the local database for subsequent map updates or as historical data for analysis;

[0082] Map library and feature library updates:

[0083] The map library and feature library are updated according to the newly collected image data and visual positioning results; if new environmental feature points are found in the new image, these feature points and their related information (such as position, feature description vector, etc.) are added to the feature library; at the same time, the UAV's posture information and feature point matching results are used to update the map information in the map library, such as adding new landmark positions or correcting the coordinates of existing landmarks; the update of the map library and feature library adopts the incremental learning method to continuously accumulate and optimize environmental information and improve the visual positioning performance of the UAV during subsequent flights; during the update process, the newly added feature points and map information need to be verified and screened to ensure the accuracy and consistency of the data; cluster analysis and other methods can be used to classify and screen new feature points, remove noise points and duplicate points, and improve the quality of the feature library and map library.

[0084] In this embodiment, the data fusion module receives data from the inertial navigation system and the visible light sensor, and optimizes the positioning information using a dynamic weighted Kalman filter algorithm, automatically increasing the weight of the visible light sensor in a dynamic environment and increasing the weight of the inertial navigation system in a static environment to obtain fused positioning information. The specific operation steps of the data fusion module are as follows:

[0085] 1. Data reception and preprocessing

[0086] Data reception:

[0087] The data fusion module receives data from the inertial navigation system (INS) and visible light sensor (VIS) in real time; the data sent by the INS includes the attitude (such as pitch angle, yaw angle, roll angle), velocity (three-dimensional velocity vector) and acceleration (linear acceleration vector) information of the drone, and the data sent by the VIS is the environmental feature point information obtained through image processing and the visual positioning results calculated based on these feature points (such as the estimated position of the drone relative to the environment); the data fusion module ensures the stability and accuracy of data transmission, and adopts appropriate communication protocols (such as serial communication, Ethernet communication, etc.) and data verification mechanisms (such as CRC verification) to receive data;

[0088] Time Synchronization:

[0089] Since the data acquisition frequencies of INS and VIS may be different, in order to ensure the accuracy of fusion, the two sets of data need to be time synchronized; hardware synchronization methods can be used, such as using a synchronous clock signal to trigger the data acquisition of INS and VIS; or software synchronization methods can be used to add timestamp information to the data packet header and align the data according to the timestamp in the data fusion module; the accuracy of time synchronization directly affects the effect of data fusion. For high-precision navigation systems, the time synchronization error should be controlled within a very small range (such as microseconds);

[0090] Data format conversion and preprocessing:

[0091] Convert the received INS and VIS data into a format suitable for processing by the dynamic weighted Kalman filter algorithm; for INS data, it may be necessary to convert the angle data into radians, unify the units of the velocity and acceleration data, etc.; for VIS data, it may be necessary to normalize the feature point coordinates, or convert the visual positioning results into the same coordinate system as the INS data (such as the geographic coordinate system or the body coordinate system); at the same time, perform a preliminary validity check on the data and remove obviously abnormal data points (such as speed values ​​beyond a reasonable range, incorrect feature point coordinates, etc.) to avoid the impact of abnormal data on subsequent fusion calculations;

[0092] 2. Dynamic weighted Kalman filter algorithm initialization

[0093] State vector definition:

[0094] Determine the state vector of the Kalman filter. The state vector should contain the key state information of the drone, such as position (3D coordinates), velocity (3D velocity vector), attitude (such as quaternion representation or Euler angle representation), etc. For example, the state vector in is the position coordinate of the drone in space, is the velocity component, is the attitude quaternion; the selection of the state vector should be reasonably determined according to the UAV's motion model and navigation requirements to ensure that the state changes of the UAV can be accurately described;

[0095] Initial state estimate:

[0096] At the beginning of the filtering algorithm, an initial estimate of the state vector needs to be set; the initial state estimate can be based on the initial data of the INS or other prior information; for example, before the drone takes off, the INS can provide initial attitude and position information as the initial state estimate; the accuracy of the initial state estimate has a certain impact on the convergence speed and final accuracy of the filtering algorithm. If the initial estimate deviation is large, the filtering algorithm may require more iterations to converge to an accurate state estimate;

[0097] Covariance matrix initialization:

[0098] Initialize the covariance matrix P of the state estimate. The covariance matrix describes the uncertainty of the initial state estimate. For the initial state estimates of INS and VIS, determine the initial value of the covariance matrix based on the accuracy of the sensor and prior knowledge. For example, the position and velocity estimation accuracy of INS is high, and the initial value of the corresponding covariance matrix element can be set to a smaller value. The initial positioning accuracy of VIS is relatively low, and the initial value of the covariance matrix element can be set to a larger value. The correct initialization of the covariance matrix helps the Kalman filter algorithm to reasonably weigh the weights of the prediction information and the observation information in subsequent iterations.

[0099] 3. Dynamic Weighted Kalman Filter Algorithm Iteration Process

[0100] Status prediction: ,in, For Always The predicted value of the state at the moment, for The best estimate of time, is the state transition matrix, which describes the system from Time has come The law of state change at every moment;

[0101] Covariance prediction: ,in, is the prediction covariance matrix, for The optimal estimated covariance matrix at time , is the process noise covariance matrix;

[0102] Kalman gain calculation: ,in, is the Kalman gain, which is used to balance the weights of prediction information and observation information in state updating. is the observation matrix, which maps the state vector to the observation vector space, is the observation noise covariance matrix, which represents the uncertainty of the observation data;

[0103] Status Update: ,in, yes The observed value at time, yes The best estimate of the time;

[0104] Covariance update: ,in, is the identity matrix, yes The optimal estimated covariance matrix at time;

[0105] Weighted processing in dynamic weighted Kalman filtering:

[0106] Assume that the estimated value of the inertial navigation system is , the covariance is , the estimated value of the visible light sensor is , the covariance is ;

[0107] Defining weights and , and satisfies , depending on the ambient light intensity , Image texture complexity And the change rate of the UAV's flight attitude Make real-time adjustments, and the weighted calculation formula is:

[0108] ,in, for , , The corresponding weight influence coefficient is used to adjust the influence of light intensity, image texture complexity and flight attitude change rate on the weight;

[0109] ; In dynamic environments (such as large lighting changes, rich image textures, or frequent changes in flight posture), the weight of VIS will increase, because VIS may provide more accurate positioning information in this environment; in a static environment (such as stable lighting, simple image texture and stable flight attitude), the weight of INS will increase, because INS has advantages in long-term stability;

[0110] The fused state estimate is: ; By weighted fusion of prediction information and observation information, a more accurate state estimation is obtained;

[0111] The fused covariance estimate is: Here, the weighted combination of the inverse of the covariance is used. This method can better consider the accuracy of different sensors in the fusion estimation, because the covariance reflects the uncertainty of the sensor estimation. Through this weighted combination, the information of different sensors can be more reasonably integrated.

[0112] 4. Iteration and convergence judgment

[0113] Iteration process:

[0114] Repeat the calculation process of the above prediction phase and update phase. As time goes by, new INS and VIS data are continuously received and Kalman filter iterative calculations are performed. In each iteration, the state estimation and covariance matrix are updated according to the new data and dynamic weighting strategy to gradually improve the accuracy of positioning information. The iterative process continues until the UAV mission ends or a specific stop condition is met.

[0115] Convergence judgment:

[0116] During the iteration process, it is necessary to determine whether the Kalman filter algorithm has converged. Convergence can be determined by monitoring the changes in state estimation, such as calculating the difference in state estimation between two adjacent iterations. If the difference is less than the set threshold (such as the position error is less than a certain distance, the speed error is less than a certain value, etc.), the algorithm is considered to have converged. In addition, the changes in the covariance matrix can also be monitored. When the element values ​​of the covariance matrix are stable within a small range, it also indicates that the algorithm has converged. Once the algorithm is determined to have converged, it means that the fused positioning information has reached a relatively stable and accurate state, and can be used as the final positioning result output or used for subsequent operations such as navigation control of the drone.

[0117] 5. Output of positioning information after fusion

[0118] Data format conversion and output:

[0119] Convert the fused positioning information (including the position, speed, attitude, etc. of the drone) into a format suitable for use by other system modules of the drone; for example, convert the position information into longitude, latitude and altitude information in the geographic coordinate system, convert the speed information into the speed magnitude and direction relative to the ground, and convert the attitude information into Euler angles or other commonly used representations of other navigation systems; then output the converted positioning information to the navigation control module, flight management system, etc. according to the agreed communication protocol and data interface, so as to provide accurate position and attitude information for the flight control and mission execution of the drone;

[0120] Data recording and monitoring:

[0121] At the same time, the fused positioning information is recorded for subsequent analysis and evaluation of the performance of the navigation system; the recorded data may include the curve of the change of positioning information over time, the statistics of positioning error under different environmental conditions, etc.; through the analysis of the recorded data, possible problems with the navigation system can be discovered (such as large positioning error under specific conditions), providing a basis for further optimization of algorithms and system parameters; in addition, during the flight of the UAV, the quality of the fused positioning information is monitored in real time, such as checking the rationality of the position and attitude data (whether it is within a reasonable range) and the continuity of the data (whether there are abnormal conditions such as data jumps). If the positioning information is found to be abnormal, take corresponding measures in time (such as alarming, switching to backup navigation mode, etc.) to ensure the safe flight of the UAV.

[0122] In this embodiment, the database includes a map library and a feature library. The database stores environmental feature points and landmark data for real-time matching by the visible light sensor to assist in the positioning of the drone. The map library and the feature library are dynamically updated according to the environmental data obtained during the flight of the drone. The update algorithm is based on the incremental learning method. The specific operations are as follows:

[0123] 1. Data storage and initialization

[0124] Map library and feature library construction:

[0125] During the initialization phase of the UAV system, the data structures of the map library and feature library are created; the map library is used to store map information of the environment, such as geographic coordinates, topography, landmark locations, etc.; the feature library is used to store information of environmental feature points extracted from images, including the locations of feature points, feature description vectors, etc.; appropriate data storage formats, such as database management systems (such as MySQL, SQLite, etc.) or custom file formats, can be used to organize and store data to facilitate data management, query, and update;

[0126] Initial data entry:

[0127] If there is prior map data (such as satellite maps, geographic information system data, etc.), import it into the map library to provide the UAV with initial environmental reference information; at the same time, for some known landmarks or key feature points, enter their feature information into the feature library manually or through preliminary image analysis as the basis for subsequent visual positioning and map updates; the entry of initial data can improve the positioning accuracy and navigation performance of the UAV in the initial stage and reduce the initialization time of the system;

[0128] 2. Real-time matching process

[0129] Feature point extraction and description (visible light sensor side):

[0130] After the visible light sensor collects the environmental image, it extracts the environmental feature points in the image and calculates the description vector of the feature points according to the image processing method described above. These feature points and description vectors will be used to match the data in the feature library. The accuracy of feature point extraction and description directly affects the effect of subsequent matching. Therefore, it is necessary to select appropriate feature extraction algorithms (such as Shi-Tomasi corner detection, SIFT, SURF, ORB, etc.) and feature description algorithms to ensure that stable and representative feature points can be extracted and effective feature description vectors can be generated.

[0131] Matching algorithm execution (data fusion module or related processing unit):

[0132] Match the feature point description vector extracted from the current frame image with the feature point description vector in the feature library. A variety of matching algorithms can be used, such as brute force matching (calculating the distance between all feature points and selecting the matching pair with the smallest distance), algorithms based on approximate nearest neighbor search (such as FLANN), etc. In the matching process, in order to improve the accuracy and robustness of the matching, some strategies are usually combined to remove erroneous matching points, such as using the RANSAC algorithm to remove abnormal matching points through geometric consistency checks to ensure the reliability of the matching results. The matching results will obtain the correspondence between the feature points in the current image and the feature points in the feature library. These correspondences provide key information for subsequent positioning calculations.

[0133] 3. Positioning assistance and navigation decision

[0134] UAV positioning calculation:

[0135] According to the feature point matching results, combined with the landmark positions and map information in the map library, the current position and attitude of the drone are calculated; the geometric relationship (such as epipolar geometry, perspective transformation, etc.) and triangulation principle can be used to determine the position of the drone in the map coordinate system through the matching relationship between the map coordinates of the known feature points and the image; at the same time, according to the matching of multiple feature points, the attitude estimation of the drone (such as pitch angle, yaw angle, roll angle) can be further optimized to improve the accuracy and stability of positioning; the results of the positioning calculation will be used for the navigation control of the drone to ensure that the drone can fly according to the predetermined route and maintain an accurate position and attitude during the flight;

[0136] Navigation Decision Support:

[0137] The map library and feature library in the database are not only used for positioning calculations, but also provide support for the navigation decisions of drones. For example, based on the terrain information in the map library, the navigation system can plan the flight path in advance to avoid obstacles and dangerous areas. Based on the landmark information in the feature library, drones can identify specific targets or areas during flight and perform corresponding mission operations (such as monitoring and photographing the target area). The information in the map library and feature library is combined with the navigation algorithm of the drone to realize intelligent navigation decisions and improve the flight safety and mission execution efficiency of the drone in complex environments.

[0138] 4. Dynamic update trigger and data preparation

[0139] Update trigger condition judgment:

[0140] During the flight of the drone, the flight status and environmental data are continuously monitored to determine whether the update conditions of the map library and feature library are met; the update conditions can be based on a variety of factors, such as the flight distance of the drone (after flying a certain distance, it is considered that the environment may have changed, and the map and feature library need to be updated), the degree of change of image features (if the number, distribution or feature description of feature points in multiple consecutive frames of images has changed significantly, it may indicate that a new environmental area has been entered), task requirements (such as when performing specific map surveying tasks, it is necessary to continuously update the map library to obtain more detailed and accurate environmental information), etc.; when the update conditions are met, the update process of the map library and feature library is started;

[0141] New data collection and arrangement:

[0142] When the update is triggered, the visible light sensor continues to collect environmental images and extract new feature points and related information in the images; at the same time, the current position, posture, flight trajectory and other information of the drone are recorded, which will be associated with the newly collected feature point data for subsequent map library and feature library updates; the newly collected data needs to be sorted and preprocessed, such as removing noise points and screening out representative feature points, to ensure the quality of the updated data; in addition, other sensor data (such as lidar data, depth camera data, etc., if any) can be combined to obtain richer environmental information and improve the accuracy and completeness of map library and feature library updates;

[0143] 5. Incremental learning and update algorithm execution

[0144] Incremental learning model training (optional):

[0145] If an incremental learning method based on deep learning is used to update the map library and feature library, the newly collected data can be used to train or fine-tune the existing deep learning model; for example, a neural network model for environmental feature classification or map generation can be trained, new data can be input into the model, and the model parameters can be adjusted so that the model can adapt to new environmental features and changes; during the training process, some incremental learning techniques can be used, such as knowledge distillation and regularization methods, to ensure that the model does not forget the previously learned knowledge while learning new data, and maintains the ability to recognize existing environmental features;

[0146] Map Library Updates:

[0147] Update the map information in the map library based on the newly collected feature point data and the location information of the UAV. A variety of methods can be used to update the map, such as a map stitching algorithm based on feature point matching, which matches and merges new image features with existing maps in the map library to expand the scope and accuracy of the map; or use a simultaneous localization and mapping (SLAM) algorithm to locate the UAV in real time while building the map, and continuously optimize the map and positioning results; the updated map library should be able to accurately reflect the environmental changes that the UAV has passed through during flight, providing a more reliable reference for subsequent navigation and positioning;

[0148] Feature library update:

[0149] Add the newly extracted feature points and their description vectors to the feature library, and optimize and manage the feature library. When adding new feature points, check the repeatability of the feature points to avoid storing too many similar feature points, which will affect the retrieval efficiency and matching accuracy of the feature library. Cluster analysis and other methods can be used to classify and screen the feature points and remove redundant feature points. At the same time, sort or weight the feature points in the feature library according to their frequency of use and importance, give priority to matching with high-quality and high-reliability feature points, and improve the overall performance of the feature library.

[0150] 6. Update result verification and feedback

[0151] Verification method selection:

[0152] Verify the update results of the map library and feature library to ensure that the updated database can accurately reflect environmental changes and improve positioning and navigation performance; a variety of verification methods can be used, such as field measurement verification (if possible, use high-precision measurement equipment to conduct field measurements in the drone flight area, and compare the updated database information with the field measurement results), cross-validation (use the updated database for drone positioning and navigation in different time periods or different flight missions to check its accuracy and stability in different scenarios), and comparison verification with other reliable data sources (such as comparison with more accurate map data, satellite images, etc.); the selection of verification methods should be determined based on actual conditions and operability to ensure that the quality of the update results can be effectively evaluated;

[0153] Feedback and Adjustment:

[0154] According to the verification results, if problems are found in the updated map library or feature library (such as inaccurate map information, increased feature point matching error rate, etc.), timely feedback and adjustments will be made; feedback information can be used to improve the update algorithm, adjust parameter settings or re-collect data for updating; for example, if a large error is found in the map update of a certain area, it may be due to inaccurate feature point extraction in that area or the map stitching algorithm is not applicable. At this time, the feature point extraction algorithm can be improved or a more appropriate map update method can be adopted; through continuous verification, feedback and adjustment, the update process of the map library and feature library can be gradually optimized, the quality and reliability of the database can be improved, and strong support can be provided for the continuous and stable operation of the UAV.

[0155] In this embodiment, the navigation control module generates flight control instructions based on the fused positioning information to guide the UAV to fly stably. The navigation control module automatically adjusts the flight strategy based on the adaptive control algorithm. When the navigation control module is running,

[0156] 1. Positioning information reception and preprocessing

[0157] Information Receiving:

[0158] The navigation control module receives the fused positioning information from the data fusion module, including the position (such as three-dimensional geographic coordinates or relative reference point), speed (three-dimensional speed vector) and attitude (such as pitch angle, yaw angle, roll angle expressed by Euler angle or quaternion) of the UAV; ensure the accuracy and timeliness of data transmission, and use reliable communication interface and data verification mechanism to prevent data loss or error;

[0159] Coordinate conversion and unit unification (optional):

[0160] According to the requirements of the UAV flight control system, the received positioning information is converted into coordinates (such as converting the coordinates in the geographic coordinate system into the coordinates in the body coordinate system for subsequent calculations) and the units are unified; for example, the speed unit is converted from meters per second to the specific speed unit used by the UAV flight control system to ensure the consistency and compatibility of the data in subsequent calculations;

[0161] Data filtering and smoothing (optional):

[0162] In order to reduce the impact of noise and fluctuations in positioning information on flight control, the data is filtered and smoothed; low-pass filtering, Kalman filtering and other methods can be used to remove high-frequency noise to make the positioning data more stable and reliable; the filtered positioning information will serve as the basis for the subsequent generation of flight control instructions to improve the accuracy and stability of flight control;

[0163] 2. Target trajectory planning

[0164] Task analysis and goal determination:

[0165] According to the flight mission of the UAV (such as autonomous inspection, aerial photography, material transportation, etc.), analyze the mission requirements and determine the target location and flight trajectory; for example, in an autonomous inspection mission, the target location may be the coordinate points of a series of facilities or areas that need to be inspected, and the flight trajectory can be a pre-planned path or a path dynamically generated based on real-time environmental information to ensure that the UAV can cover all target areas and complete the mission;

[0166] Trajectory generation algorithm execution:

[0167] Use a suitable trajectory generation algorithm to generate the expected flight trajectory of the drone based on the target position and the drone's current position, speed, and attitude information; common trajectory generation algorithms include polynomial interpolation and spline curve interpolation. For example, using cubic polynomial interpolation, the expected position, speed, and acceleration of the drone at different times are calculated based on the drone's current position and target position, as well as the speed and acceleration constraints during flight, to form a smooth flight trajectory. The drone's dynamic constraints (such as maximum speed, maximum acceleration, minimum turning radius, etc.) need to be considered during trajectory generation to ensure that the generated trajectory is a feasible trajectory that the drone can actually fly.

[0168] 3. Initialization of adaptive control algorithm

[0169] Control parameter setting:

[0170] According to the model, performance characteristics and flight mission requirements of the UAV, the initial parameters of the adaptive control algorithm are set; these parameters include control gain (such as proportional gain, integral gain, differential gain, etc.), adaptive law parameters (used to adjust the control gain to adapt to different flight conditions), constraints (such as maximum control output limit, attitude angle limit, etc.), etc. The selection of initial parameters directly affects the performance and stability of the adaptive control algorithm, and needs to be reasonably set based on experience and theoretical analysis;

[0171] State variable initialization:

[0172] Initialize the state variables in the adaptive control algorithm, such as the error integral term (used to eliminate steady-state errors) and the adaptive parameter estimation value (used to track changes in system parameters). The initial values ​​of the state variables are usually set according to the initial state of the UAV and the mission requirements. For example, when the UAV takes off, the error integral term is initialized to zero, and the adaptive parameter estimation value can be set according to the pre-identification results of the system or the empirical value.

[0173] 4. Flight control command generation

[0174] Error calculation:

[0175] According to the fused positioning information and the expected flight trajectory, the error between the current state of the drone and the expected state is calculated; the error includes position error (position deviation in three-dimensional direction), speed error (speed magnitude and direction deviation), attitude error (attitude deviation expressed by Euler angle or quaternion), etc. For example, position error (in, is the desired position, is the current position), speed error (in, is the expected speed, is the current speed), attitude error (in, It is a posture of expectation. is the current posture);

[0176] Control law calculation:

[0177] Based on the calculated error, the flight control command is calculated using an adaptive control algorithm; the adaptive control algorithm adjusts the control gain in real time according to the size and change trend of the error to adapt to different flight conditions and environmental changes; for example, using an adaptive proportional-integral-derivative (PID) control algorithm, the control command (control command (in They are proportional gain, integral gain, and differential gain, which are adjusted in real time according to the adaptive law); the calculation of the control law also needs to consider the UAV's dynamic model and control constraints to ensure that the control instructions are within the executable range of the UAV and can stably control the UAV to track the desired trajectory;

[0178] Instruction optimization and constraint processing:

[0179] The generated flight control instructions are optimized and constrained to ensure the rationality and safety of the instructions. For example, check whether the control instructions exceed the maximum control output limit of the UAV (such as maximum throttle, maximum rudder angle, etc.). If the limit is exceeded, the instructions are limited to prevent the UAV from losing control due to excessive control. At the same time, considering the flight stability and comfort of the UAV, the control instructions are smoothed to avoid sudden changes in instructions that cause the UAV to shake violently. In addition, according to the flight status and environmental conditions of the UAV (such as wind speed, wind direction, etc.), the control instructions are appropriately compensated and adjusted to improve the flight performance of the UAV in complex environments.

[0180] 5. Flight strategy adjustment

[0181] Environmental Monitoring and Assessment:

[0182] The navigation control module monitors the environmental information around the drone in real time, such as meteorological information such as wind speed, wind direction, air pressure, temperature, and geographical information such as topography, obstacle distribution, etc. (which can be obtained through communication with other sensors or external data sources); based on this environmental information, it evaluates the impact of the current flight environment on the drone's flight, such as strong winds may cause the drone to deviate from its route, and complex terrain may require adjustment of flight altitude to avoid collision, etc.

[0183] Strategy Adjustment Algorithm Execution:

[0184] According to the environmental assessment results, the flight strategy adjustment algorithm is executed to dynamically change the flight strategy of the UAV; for example, if a strong headwind is encountered, the throttle output of the UAV is increased to maintain the flight speed; if an obstacle is detected ahead, a new flight path is planned based on the location and shape of the obstacle to avoid the obstacle; the flight strategy adjustment algorithm can be based on a rule library (such as flight strategy rules pre-set according to different environmental conditions) or a real-time optimization algorithm (such as a model predictive control algorithm, which optimizes the flight strategy by predicting the flight status and environmental changes of the UAV in the future) to ensure that the UAV can fly safely and efficiently in a complex and changing environment;

[0185] 6. Control instruction output and execution

[0186] Instruction format conversion and output:

[0187] Convert the generated flight control instructions into a format that can be recognized and executed by the UAV flight control system, such as PWM (pulse width modulation) signals, digital instructions, etc.; send the control instructions to the UAV flight control actuators, such as motors, servos, etc., through appropriate communication interfaces (such as serial ports, CAN buses, etc.);

[0188] Instruction execution and feedback:

[0189] The flight control actuator receives control instructions and executes corresponding actions to control the flight attitude and trajectory of the UAV; at the same time, the flight control system feeds back the execution results (such as actual attitude, speed, position change, etc.) to the navigation control module. The navigation control module evaluates the execution effect of the control instructions based on the feedback information and determines whether the UAV flies according to the expected trajectory; if deviations are found, the flight control instructions are adjusted in time to form a closed-loop control to ensure that the UAV flies stably and accurately completes the flight mission; during the entire flight process, the above steps are repeated continuously to realize the real-time monitoring, control and strategy adjustment of the UAV flight by the navigation control module.

[0190] In this embodiment, the image processing method of the visible light sensor includes a deep learning algorithm. The deep learning algorithm performs real-time environmental feature extraction and matching through a convolutional neural network. Assume that the input image is , in, Represent the horizontal and vertical coordinates of the image pixels respectively, and the convolution kernel is , is the coordinate of the convolution kernel element, then the calculation formula of the convolution layer is: ,in, The feature map output by the convolutional layer is in coordinates The value at is a bias term. Through the combined operation of multi-layer convolution, pooling and fully connected layers, the environmental feature points are extracted and matched. The pooling operation adopts the maximum pooling formula: ,in, The output of the pooling layer is at coordinates The value at is the pooling area.

[0191] In this embodiment, during the dynamic update process of the map library and the feature library, the newly collected environmental feature point data is initially screened using a clustering analysis method, and the feature library is updated and optimized after the noise data is removed; the feature point set is ,in, is the i-th environmental feature point, is the total number of feature points, and the K-means clustering algorithm is used, and its objective function is: ,in, is the number of categories into which the feature points are divided, For the A set of clusters, For the The center of the cluster, Represents the Euclidean distance. Through continuous iterative optimization, the objective function Minimum, the feature points with a set distance from the cluster center are regarded as noise points and removed, and the filtered feature points are added to the feature library for updating.

[0192] In this embodiment, the navigation control module takes into account the remaining power of the drone when generating flight control instructions. , Communication signal strength and traffic conditions in the surrounding airspace , let the flight control command vector be , then the control instruction generation formula based on multiple factors is: ;in, , are the weight coefficients of different factors, and , is the control function related to the remaining power, is a control function related to the communication signal strength, It is a control function related to the traffic conditions in the surrounding airspace.

[0193] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An adaptive navigation system for unmanned aerial vehicles, characterized in that: include: Inertial navigation system; the inertial navigation system provides attitude, speed and acceleration data of the drone; Visible light sensor; the visible light sensor includes at least one high-definition camera, which is used to collect environmental image information and obtain environmental feature points through image processing methods to achieve visual positioning; Data fusion module: The data fusion module receives data from the inertial navigation system and the visible light sensor, optimizes the positioning information using a dynamic weighted Kalman filter algorithm, automatically increases the weight of the visible light sensor in a dynamic environment, and increases the weight of the inertial navigation system in a static environment, to obtain fused positioning information; Database; the database includes a map library and a feature library. The database stores environmental feature points and landmark data for real-time matching by visible light sensors to assist in drone positioning. The map library and the feature library are dynamically updated based on environmental data acquired during the flight of the drone. The update algorithm is based on an incremental learning method. Navigation control module; the navigation control module generates flight control instructions based on the fused positioning information to guide the UAV to fly stably, and the navigation control module automatically adjusts the flight strategy based on the adaptive control algorithm; The dynamic weighted Kalman filter algorithm in the data fusion module is as follows: Status prediction: ,in, For Always The predicted value of the state at the moment, for The best estimate of time, is the state transition matrix, which describes the system from Time has come The law of state change at every moment; Covariance prediction: ,in, is the prediction covariance matrix, for The optimal estimated covariance matrix at time , is the process noise covariance matrix; Kalman gain calculation: ,in, is the Kalman gain, which is used to balance the weights of prediction information and observation information in state updating. is the observation matrix, which maps the state vector to the observation vector space, is the observation noise covariance matrix, which represents the uncertainty of the observation data; Status Update: ,in, yes The observed value at time, yes The best estimate of the time; Covariance update: ,in, is the identity matrix, yes The optimal estimated covariance matrix at time; Weighted processing in dynamic weighted Kalman filtering: Assume that the estimated value of the inertial navigation system is , the covariance is , the estimated value of the visible light sensor is , the covariance is ; Defining weights and , and satisfies , depending on the ambient light intensity , Image texture complexity And the change rate of the UAV's flight attitude Make real-time adjustments, and the weighted calculation formula is: ,in, , , for , , The corresponding weight influence coefficient; ； The fused state estimate is: ; The fused covariance estimate is: .

2. The UAV adaptive navigation system according to claim 1, characterized in that: The image processing method of the visible light sensor includes a deep learning algorithm, which performs real-time environmental feature extraction and matching through a convolutional neural network. Assume that the input image is , in, Represent the horizontal and vertical coordinates of the image pixels respectively, and the convolution kernel is , is the coordinate of the convolution kernel element, then the calculation formula of the convolution layer is: ,in, The feature map output by the convolutional layer is in coordinates The value at is a bias term. Through the combined operation of multi-layer convolution, pooling and fully connected layers, the environmental feature points are extracted and matched. The pooling operation adopts the maximum pooling formula: ,in, The output of the pooling layer is at coordinates The value at is the pooling area.

3. The UAV adaptive navigation system according to claim 1, characterized in that: During the dynamic update process of the map library and the feature library, the newly collected environmental feature point data is initially screened by using a cluster analysis method, and the feature library is updated and optimized after noise data is removed; set up The feature point set is ,in, is the i-th environmental feature point, is the total number of feature points, and the K-means clustering algorithm is used, and its objective function is: ,in, is the number of categories into which the feature points are divided, For the A set of clusters, For the The centers of the clusters, Represents the Euclidean distance. Through continuous iterative optimization, the objective function Minimum, the feature points with a set distance from the cluster center are regarded as noise points and removed, and the filtered feature points are added to the feature library for updating.

4. The UAV adaptive navigation system according to claim 1, characterized in that: The navigation control module takes into account the remaining power of the drone when generating flight control instructions. , Communication signal strength and traffic conditions in the surrounding airspace , let the flight control command vector be , then the control instruction generation formula based on multiple factors is: ;in, , , , are the weight coefficients of different factors, and , is the control function related to the remaining power, is a control function related to the communication signal strength, It is a control function related to the traffic conditions in the surrounding airspace.

Images