An unmanned aerial vehicle low-altitude visual navigation method based on motion blur assistance

Abstract

The present application relates to a kind of unmanned aerial vehicle low-altitude visual navigation method based on motion blur auxiliary, comprising: acquisition monocular camera image and IMU data, quantification the image blur degree of the monocular camera image, to extract the blur parameter of dominant motion, for calculating global measurement blur data;The IMU data is pre-integrated, obtains the relative motion amount between adjacent frames, and based on system prior state estimation, the image plane motion field of current time is predicted, for simulating global theoretical blur data;The global measurement blur data and the global theoretical blur data are compared, and after being combined with adaptive weight, as motion blur consistency error, construct nonlinear least squares optimization objective function;According to the nonlinear least squares optimization objective function, Gauss-Newton method iteration is solved, generates the state vector of minimum total error, for calculating health score, to implement three-level mode switching.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) navigation and state estimation technology, and in particular to a low-altitude visual navigation method for UAVs based on motion fuzzy assistance. Background Technology

[0002] With the widespread application of drones in inspection, surveying, and logistics, their autonomous navigation capability in complex environments has become a key technological bottleneck. Accurate and reliable real-time velocity estimation is fundamental to achieving precise hovering, path tracking, and obstacle avoidance. Currently, existing solutions mainly fall into two categories: one is to preprocess blurred images using image deblurring algorithms, but this method introduces artifacts, destroys image realism, and is computationally time-consuming, failing to meet real-time requirements; the other is to simply reduce the dependence of VIO on visual features, but this leads to a rapid accumulation of estimation errors. Both methods treat motion blur as a purely interfering factor, failing to extract the valuable motion information it contains.

[0003] Current technologies are unable to provide a high-precision, high-frequency attitude estimation solution for drones operating at high speeds, while also possessing strong robustness and good generalization ability. Therefore, there is an urgent need for an innovative technology that can comprehensively utilize the advantages of multi-source information while overcoming their respective limitations, to meet the pressing need for drones to operate autonomously in extreme environments. Summary of the Invention

[0004] To address the problems existing in the prior art, the purpose of this invention is to provide a motion fuzz-assisted low-altitude visual navigation method for unmanned aerial vehicles (UAVs), which can overcome the performance degradation defects of traditional VIO in high-speed, high-ratio environments and comprehensively utilize fuzzy image information to help improve the navigation state estimation accuracy of UAVs in low-altitude, high-speed flight.

[0005] To achieve the above objectives, the present invention provides the following solution: A low-altitude visual navigation method for unmanned aerial vehicles (UAVs) based on motion fuzzing assistance includes: Acquire images from a monocular camera and IMU data, quantify the image blur level of the monocular camera image to extract blur parameters of the dominant motion, and use them to calculate global measurement blur data; The IMU data is pre-integrated to obtain the relative motion between adjacent frames, and the image planar motion field at the current moment is predicted based on the system's prior state estimation to simulate global theoretical fuzzy data. The global measurement fuzzy data and the global theoretical fuzzy data are compared and combined with adaptive weights as motion fuzzy consistency error to construct a nonlinear least squares optimization objective function. The Gauss-Newton method is used to iteratively solve the nonlinear least squares optimization objective function to generate a state vector with the minimum total error, which is used to calculate the health score and implement the three-level mode switching.

[0006] Optionally, calculating the global measurement fuzzy data includes: Perform a fast Fourier transform on each frame of monocular camera image and calculate the power spectrum. Based on the power spectrum, obtain the high-frequency attenuation rate, which is used to quantify the image blur level of the monocular camera image and obtain clear image, slightly blurred image and severely blurred image. The clear image and the slightly blurred image are subjected to standard ORB feature extraction and A-KAZE robust feature extraction respectively to obtain visual features, which are used to construct the visual reprojection error term. The severely blurred image is subjected to partitioned mixing Von Mises modeling processing, and the responsibility degree of each region is calculated to extract the blur parameters of the dominant motion from the local motion blur parameters; Radon transform is performed on the fuzzy parameters of the dominant motion to obtain the global measurement fuzzy data, namely the global fuzzy angle and fuzzy length.

[0007] Optionally, the calculation of the responsibility level for each region includes: Calculate the weighted average vector for each region, and obtain the concentration parameter based on the weighted average vector: ; in, The square of the magnitude of the weighted average vector, ( () is the weighted average vector; when If the concentration parameter is less than the first target threshold, then it proves that the concentration parameter is: ; when If the concentration parameter is greater than the first target threshold and less than the second target threshold, then the concentration parameter is considered to be: ; when If the concentration parameter is not less than the second target threshold, then it proves that the concentration parameter is: ; in, For concentration parameters; Using the concentration parameter, calculate the degree of responsibility for each region: ; ; in, For posterior probability, For zero-order modified Bessel functions, , The weight of the first component in the mixture. This is the core exponential term in the von Mises distribution, used to describe the probability distribution of angle or direction data. For the fuzzy direction of each region, The main fuzzy direction, For concentration parameters, Pi The weight of the second component in the mixture. For integration variables, The factorial of r, For summation index.

[0008] Optionally, obtaining the global blur angle and blur length includes: Get the global blur angle: A Radon transform is performed on the blur parameters of the dominant motion. The position of the maximum value in the transform result is used as the fringe direction. The motion blur direction is then determined using the fringe direction, which is used to calculate the global blur direction. ; ; in, For fuzzy directions within each primary cluster, This is the weighting factor for the region. For global fuzzy direction, For the defined primary cluster domain, For global fuzzy direction, It is the arctangent function; Get the global fuzz length: Based on a robust fuzzy length fusion strategy using weighted median, all data points within the main cluster are sorted by fuzzy length from smallest to largest, and the total weight and cumulative weight of the sorted data are calculated. Using the total weight to set a third target threshold, data points whose first cumulative weight is not less than the third target threshold are extracted and used as the global fuzzy length.

[0009] Optionally, the method further includes: Calculate the confidence scores of the fuzzy parameters to construct the adaptive weights: ; ; ; ; in, For confidence level, For spectral stripe contrast factor, Noise level factor, The peak significance factor, + + =1, , , For preset coefficients, The maximum value of the Radon projection in the principal direction reflects the salience of the spectral fringes. This is the minimum value of the Radon projection, reflecting the background noise level. The contrast normalization threshold, The standard deviation of the Radon projection plot reflects the dispersion of the projected values. The mean of the Radon projection is... This is the second highest peak value, used to assess multi-peak interference. The difference between the main peak and the secondary peak reflects the prominence of the main peak. The term is a peak significance assessment for z-score, which measures the prominence of the main peak relative to the overall distribution.

[0010] Optionally, the simulation of the global theoretical fuzzy data includes: Pre-integrating the IMU data yields the relative motion between adjacent frames, including relative rotation integral, relative velocity integral, and relative position integral. Based on the system's prior state estimation, the planar motion field of the image at the current moment is predicted. ; ; ; in, Let be the prior velocity at time k+1. This is the optimal velocity estimate at time k. The velocity increment produced by gravitational acceleration within the time interval. To rotate the velocity increments from the body coordinate system to the world coordinate system, Let k+1 be the prior position. This is the optimal position estimate at time k. The position increment generated by uniform motion, The position increment caused by gravitational acceleration. To rotate the displacement increments in the body coordinate system to the world coordinate system; Based on the image planar motion field and the monocular camera exposure time, simulate the global theoretical blur data, namely, motion blur direction, motion blur length, and image sharpness: ; ; ; in, For motion ambiguity direction, For the motion blur length, For image clarity, For average motion, For camera exposure time, This is the resolution scale factor.

[0011] Optionally, the adaptive weights include: ; =c, , ; in, For adaptive weights, For the confidence level of the fuzzy parameter, For motion state factors, Here, c represents the system degradation factor, and c represents the confidence level. This represents the magnitude of the current velocity estimate. For velocity scaling factor, This represents the number of feature points successfully tracked in the current frame. These are the typical characteristic numbers of a system in a healthy state. Based on weights, For quality factor weights, For speed factor weights, The weights are for the degradation factor.

[0012] Optionally, constructing the nonlinear least squares optimization objective function includes: ; ; ; ; in, The objective function is a nonlinear least squares optimization function. For state vectors, This is the visual reprojection error term. For the IMU pre-integration error term, For motion fuzzy consistency error term, For adaptive weights, This is the set of keyframes available for fuzzy analysis within the sliding window. The covariance matrix of the motion fuzzy parameter measurements. This refers to the number of keyframes. For keyframe indexing, The number of feature points, For feature point index, For Huber robust kernel function, This is the visual reprojection error vector. For keyframe pair indexing, This is due to visual reprojection error. This refers to the motion fuzziness consistency error.

[0013] Optionally, implementing the three-level mode switching includes: Calculate your health score: ; ; ; ; in, As a health score, The average tracking length for all feature points. To achieve the required minimum average tracking length, For the error threshold, This represents the number of feature points successfully tracked in the current frame. These are the typical characteristic numbers of a system in a healthy state. For confidence level, As a characteristic quality indicator, As an IMU consistency indicator, For fuzzy consistency indicators, The norm of the IMU error. Let be the norm of the motion blur error; When the health score equals the fourth target threshold, it proves that the system is in an ideal working state; When the health score drops to the first threshold range, it enters the fuzzy-assisted mode to rely on motion fuzzy constraints. When the health score drops to the second threshold range, the system switches to fuzzy-dominated mode to forcibly stabilize the system using fuzzy information.

[0014] The beneficial effects of this invention are as follows: This invention fuses motion blur parameters of motion-blurred images with high-speed visual odometry data, transforming motion blur information from noise to be eliminated into usable motion constraints, and deeply couples them within the entire VIO optimization framework, thereby achieving high-precision attitude estimation for UAVs in environments such as high speed, low light, and sparse features. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a low-altitude visual navigation method for unmanned aerial vehicles (UAVs) based on motion fuzzy assistance, according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this embodiment discloses a low-altitude visual navigation method for unmanned aerial vehicles (UAVs) based on motion blur assistance, including: acquiring monocular camera images and IMU data; quantifying the image blur degree of the monocular camera images to extract the blur parameters of the dominant motion, used to calculate global measurement blur data; pre-integrating the IMU data to obtain the relative motion between adjacent frames, and predicting the image planar motion field at the current moment based on the system's prior state estimation, used to simulate global theoretical blur data; comparing the global measurement blur data and the global theoretical blur data, and combining them with adaptive weights as the motion blur consistency error, constructing a nonlinear least squares optimization objective function; performing Gauss-Newton iterative solution based on the nonlinear least squares optimization objective function to generate a state vector with the minimum total error, used to calculate the health score, to implement three-level mode switching. More specifically: This embodiment discloses a low-altitude visual navigation method for unmanned aerial vehicles (UAVs) based on motion fuzzy assistance, specifically including the following steps: Step 1. Preprocess and extract parameters from the motion-blurred image, perform frequency domain analysis and evaluate the degree of image blur, perform local blur modeling and multi-region fusion extraction on the blurred image, and estimate the dominant global blur direction. and fuzzy length And calculate the parameter confidence level c; Step 2. Simulation of motion fuzzy parameters based on state prediction: By forward simulating theoretical fuzzy parameters based on IMU pre-integration and geometric optics model of the system, a reference benchmark is provided for subsequent motion fuzzy consistency constraints. Step 3. Construct a coupled optimization framework that includes motion fuzz consistency constraints. Compare the measurement fuzzy parameters from Step 1 with the prediction fuzzy parameters from Step 2, and multiply them with adaptive weights to obtain the motion fuzzy consistency error. Introduce this error as an optimization term into the visual inertial odometry to construct a complete nonlinear least squares optimization problem. Solve for the state vector that minimizes the total error. Step 4. Calculate adaptive weights based on fuzzy quality, motion state, and system degradation degree, and implement three-level mode switching based on health score; The specific steps are as follows: Step 1. Synchronously acquire images from a monocular camera and IMU data through a hardware triggering mechanism, and accurately record the exposure time of each frame. Perform a fast Fourier transform on each frame of the image and calculate its logarithmic power spectrum. Quantify the degree of image blur by analyzing the attenuation rate of high-frequency components. Based on the quantification value of the degree of blur, a three-level decision mechanism is adopted: clear images are extracted using standard ORB features, slightly blurred images are extracted using A-KAZE robust features, and severely blurred images skip traditional feature extraction and directly enter the process of precise extraction of motion blur parameters.

[0020] Logarithmic power spectrum calculation: Given a frame of image Its two-dimensional discrete Fourier transform is: ; Power spectrum: ; Logarithmic power spectrum: ; High-frequency attenuation rate calculation: ; in, The cutoff frequency is typically taken as 1 / 4 of the image size. For frequency domain coordinates, The width and height (in pixels) of the image. For pixel coordinates, It is an imaginary number.

[0021] in accordance with The degree of image blur is graded and judged: When the value is >0.8: the image is clear, and standard feature extraction (such as ORB) is used for VIO front-end.

[0022] 0.3< When the value is ≤0.80: the image is slightly blurry. Use features that are more robust to blur (such as A-KAZE) or perform matching at higher levels of the image pyramid.

[0023] When the value is ≤0.3: the image is severely blurred. We skip the traditional feature extraction and directly enter the precise extraction channel for motion blur parameters, using the blur information to assist in state estimation.

[0024] The process for extracting blur parameters in each region is as follows: The extraction of image blur parameters includes: performing Von Mises modeling on the blurred image in a partitioned mixing manner, calculating the responsibility degree of each region, eliminating the uncertainty and multimodality of the data in each region of the image, and extracting the blur parameters of the dominant motion from the local motion blur parameters.

[0025] The density function expression of the hybrid Von Mises model is as follows: ; In the formula, the first term, Von Mises distribution, represents the primary motion fuzziness mode, and the second term, uniform distribution, represents the noise / anomaly / secondary motion mode. , respectively, represent the mixing ratio of the two components. Represents a positive Bessel function. The average direction (obtainable from the fusion method). The fuzzy directions for each region can be determined using Radon transform. It is a concentration parameter. The weight of the first component in the mixture. The weight of the second component in the mixture.

[0026] The responsibility degree of each region is calculated to eliminate the uncertainty and multimodality of the data in each region of the image, and the blur parameters of the dominant motion are extracted from the local motion blur parameters of each region.

[0027] First, calculate the weighted average vector of each local region. ): ; in, Let i be the weight of the i-th local region. The total number of local regions, Let be the index variable, representing the i-th local region.

[0028] The dominant fuzzy direction in each local area is: ; Calculate concentration parameters : ; when hour, ; when hour, ; when hour, ; Responsibility Calculation: ; In the formula, For zero-order modified Bessel functions, If a region belongs to the principal component, the posterior probability is... If the value is greater than 0.5, it is considered to belong to the main cluster. This then participates in the subsequent calculation of global fuzzy parameters.

[0029] Radon transform is performed on the regions with a responsibility score greater than 0.5, i.e., the main cluster regions, to obtain the blur angle and blur length of each region: ; in A binarized spectral image, It is the Dirac delta function (or impulse function in discrete form).

[0030] Fuzzy parameter calculation: After Radon transform, find the location of the maximum value in the transform result and the corresponding angle. It refers to the direction of the stripes.

[0031] ; ; ; In the formula, The direction of the stripes detected by Radon. It is the direction of motion blur. It's the stripe spacing. The maximum value after Radon transform.

[0032] By employing a weighted vector fusion formula tailored to the periodicity of motion fuzzy directions, and a robust fuzzy length fusion strategy based on the weighted median, the fuzzy parameters of the main cluster are fused and extracted, thereby estimating the fuzzy directions that dominate the global motion. and fuzzy length .

[0033] Global fuzzy direction calculation: ; ; in, For each fuzzy direction within the main cluster, This is the weighting factor for the region. For global fuzzy direction, For the defined primary cluster domain, For global fuzzy direction, It is the arctangent function.

[0034] Global fuzzy length calculation: All data points within the main cluster are sorted by fuzzy length. Sort the data from smallest to largest and calculate the total weight of the sorted data. Calculate the cumulative weights and find the first one that results in a cumulative weight. Data The fuzzy length of this data point is the weighted median. , which serves as the global fuzzy length.

[0035] Calculate the confidence of parameters: By constructing confidence factors, the calculation ratio and credibility of each fuzzy parameter are measured, which improves the reliability of fuzzy parameter extraction in complex scenes. The calculation of confidence c for fuzzy images includes: spectral stripe contrast factor, noise level factor, and peak significance factor.

[0036] The expression for the spectral stripe contrast factor is as follows: ; In the formula, The maximum value of the Radon projection in the principal direction reflects the salience of the spectral fringes. This is the minimum value of the Radon projection, reflecting the background noise level. The contrast normalization threshold is typically set to 50, and the exponential term is used to suppress falsely high scores in cases of extremely low contrast.

[0037] The expression for the noise level factor is as follows: ; In the formula, The standard deviation of the Radon projection plot reflects the dispersion of the projected values. The mean of the Radon projection is... The second highest peak value is used to assess multi-peak interference.

[0038] The expression for the peak significance factor is as follows: ; In the formula, The difference between the main peak and the secondary peak reflects the prominence of the main peak. The term is a peak significance assessment for z-score, which measures the prominence of the main peak relative to the overall distribution.

[0039] Comprehensive confidence model: ; In the formula, + + =1.

[0040] Step 2. Pre-integrate the IMU data to obtain the relative motion between adjacent frames; Relative rotation integral: ; Relative velocity integral: ; Relative position integral: ; in, These are the angular velocities and accelerations measured by the IMU. The gyroscope and accelerometer have zero bias. For image frame indexing, For time intervals.

[0041] The prior state is calculated based on the pre-integral quantity, and combined with the optimal estimate at time k, the planar motion field of the image at the current time is predicted. System prior state estimation calculation: Attitude priors: ; Speed ​​Prior: ; in, Let be the prior velocity at time k+1. This is the optimal velocity estimate at time k. The velocity increment produced by gravitational acceleration within the time interval. To rotate the velocity increments from the body coordinate system to the world coordinate system.

[0042] Location Priors: ; in, Let k+1 be the prior position. This is the optimal position estimate at time k. The position increment generated by uniform motion, The position increment caused by gravitational acceleration. To rotate the displacement increment in the body coordinate system to the world coordinate system.

[0043] Based on the motion field and camera exposure time, the theoretical motion blur direction is simulated and predicted. Fuzzy length and image clarity ; Back projection calculation: ; in, A 3D point in the camera coordinate system. Image pixel coordinates, Let these be the coordinates of the camera's principal point. Z is the camera focal length (in pixels) and Z is the assumed depth value.

[0044] Three-dimensional motion velocity: ; in, The velocity of a 3D point in the camera coordinate system. For the a priori linear velocity, For the prior angular velocity, It is the cross product of vectors.

[0045] Image planar motion speed: ; The Jacobian matrix for: ; Average motion calculation: ; Theoretical fuzzy parameters: ; ; ; in, The camera exposure time (in seconds). This is the resolution scale factor (typically 100 pixels per second).

[0046] Step 3. Construct an optimization framework for motion fuzzy coupling by combining motion fuzzy consistency error; Specifically, the state vector is defined as follows: ; in, For position vectors, For velocity vector, For attitude quaternions, To achieve zero bias in the accelerometer, To achieve zero bias in the gyroscope, Let be the inverse depth of the i-th feature point.

[0047] The complete least squares optimization problem: ; Among them, the visual reprojection error term is: ; in, ; IMU pre-integration error term: in, ; Motion fuzzy consistency error term: ; in, Let j be the motion blur consistency error of the j-th frame. This is the set of keyframes available for fuzzy analysis within the sliding window. The covariance matrix of the motion fuzzy parameter measurements. The residual is between frame i and frame i+1. For state vectors, Let be the pre-integral rotation increment from frame i to frame i+1. Let be the rotation matrix for the i-th frame. Let be the rotation matrix for the (i+1)th frame. Let i be the velocity vector of the i-th frame. Let i be the velocity vector of the (i+1)th frame. It is the acceleration due to gravity. Let be the time interval from frame i to frame i+1. Let be the position vector of the i-th frame. Let i be the position vector of the (i+1)th frame. The acceleration bias is for the (i+1)th frame. Let be the acceleration bias for the i-th frame. The gyroscope bias for the (i+1)th frame. The gyroscope bias is for the i-th frame. Let be the direction and angle of motion blur in the j-th frame image. Let be the length of the motion blur in the j-th frame. For the motion blur features of the j-th frame, The motion blur direction of the predicted j-th frame is... Let j be the predicted motion blur length of the j-th frame. Let be the motion blur feature of the predicted j-th frame.

[0048] Using adaptive weighting factors Multiplying by the fuzzy error term can be adjusted It can control the trade-off between vision, inertia, and blur, and the calculation of adaptive weighting factors includes: ; 1. Fuzzy quality factor: Confidence level of fuzzy parameters: =c; 2. Motion state factors: ; in, This represents the magnitude of the current velocity estimate. The velocity scaling factor is usually taken as 5 m / s. 3. System degradation factors: ; in, This represents the number of feature points successfully tracked in the current frame. The typical number of features in a healthy system state is usually set to 100.

[0049] Solve using the Gauss-Newton method iteratively: Linearization: Estimating the current state At this point, perform a first-order Taylor expansion on all error terms: ; in, This is an estimate of the current working status. This represents the update amount of the state increment. Is the residual function in The Jacobian matrix at that location.

[0050] Construct the normal equation: ; Status Update: ; Where W is the weight matrix, This is the updated status. For the residual function, For update operations on manifolds, especially for pose quaternions: ; in, For the updated pose quaternion, The current attitude quaternion, This is quaternion multiplication. This is an exponential mapping.

[0051] Iteration termination condition: (The increment is small enough); (The residual change is small enough); the maximum number of iterations is reached. .

[0052] Step 4. Health Score Calculation: ; ; ; ; in, The average tracking length for all feature points. For the required minimum average tracking length (typically) =3), This is the error threshold.

[0053] when =1 means that the visual features are rich and stable, the IMU data is highly consistent with the estimation, the motion fuzzy constraints are also satisfied, and the system is in an ideal working state.

[0054] when A drop to 0.4-0.7: This is usually due to Decrease (fewer features) and / or The effects begin to show, the system enters fuzzy-assisted mode, and begins to rely more on motion fuzzy constraints.

[0055] when ≤0.4: usually means The value is already very low (features are extremely sparse), and the system mainly relies on IMU and fuzzy constraints. At this point, traditional VIO is on the verge of failure and must be switched to fuzzy-dominated mode to forcibly stabilize the system using fuzzy information.

[0056] In summary, this invention synchronously acquires images and IMU data via hardware, performs fuzzy classification and adaptive processing of images based on frequency domain analysis, accurately extracts motion blur parameters through Radon transform, simulates theoretical fuzzy parameters based on IMU pre-integration and geometric optics models, constructs a tightly coupled optimization framework including motion blur consistency constraints, and calculates adaptive weights based on multiple factors to achieve three-level mode switching. This invention transforms motion blur from noise to be eliminated into usable motion constraints, deeply coupling it into the VIO optimization framework, effectively solving the state estimation degradation problem of UAVs in high-speed flight and visual feature sparse environments, and significantly improving the accuracy and robustness of the navigation system.

[0057] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A low-altitude visual navigation method for unmanned aerial vehicles (UAVs) based on motion fuzzy assistance, characterized in that, include: Acquire images from a monocular camera and IMU data, quantify the image blur level of the monocular camera image to extract blur parameters of the dominant motion, and use them to calculate global measurement blur data; The IMU data is pre-integrated to obtain the relative motion between adjacent frames, and the image planar motion field at the current moment is predicted based on the system's prior state estimation to simulate global theoretical fuzzy data. The global measurement fuzzy data and the global theoretical fuzzy data are compared and combined with adaptive weights as motion fuzzy consistency error to construct a nonlinear least squares optimization objective function. The Gauss-Newton method is used to iteratively solve the nonlinear least squares optimization objective function to generate a state vector with the minimum total error, which is used to calculate the health score and implement the three-level mode switching.

2. The UAV low-altitude visual navigation method based on motion fuzzy assistance according to claim 1, characterized in that, Calculating the global measurement fuzzy data includes: Perform a fast Fourier transform on each frame of monocular camera image and calculate the power spectrum. Based on the power spectrum, obtain the high-frequency attenuation rate, which is used to quantify the image blur level of the monocular camera image and obtain clear image, slightly blurred image and severely blurred image. The clear image and the slightly blurred image are subjected to standard ORB feature extraction and A-KAZE robust feature extraction respectively to obtain visual features, which are used to construct the visual reprojection error term. The severely blurred image is subjected to partitioned mixing Von Mises modeling processing, and the responsibility degree of each region is calculated to extract the blur parameters of the dominant motion from the local motion blur parameters; Radon transform is performed on the fuzzy parameters of the dominant motion to obtain the global measurement fuzzy data, namely the global fuzzy angle and fuzzy length.

3. The UAV low-altitude visual navigation method based on motion fuzzy assistance according to claim 2, characterized in that, The calculation of the responsibility level for each region includes: Calculate the weighted average vector for each region, and obtain the concentration parameter based on the weighted average vector: ； in, The square of the magnitude of the weighted average vector, ( () is the weighted average vector; when If the concentration parameter is less than the first target threshold, then it proves that the concentration parameter is: ; when If the concentration parameter is greater than the first target threshold and less than the second target threshold, then the concentration parameter is considered to be: ; when If the concentration parameter is not less than the second target threshold, then it proves that the concentration parameter is: ; in, For concentration parameters; Using the concentration parameter, calculate the degree of responsibility for each region: ； ； in, For posterior probability, For zero-order modified Bessel functions, , The weight of the first component in the mixture. This is the core exponential term in the von Mises distribution, used to describe the probability distribution of angle or direction data. For the fuzzy direction of each region, The main fuzzy direction, For concentration parameters, Pi The weight of the second component in the mixture. For integration variables, The factorial of r, For summation index.

4. The UAV low-altitude visual navigation method based on motion fuzzy assistance according to claim 2, characterized in that, Obtaining the global blur angle and blur length includes: Get the global blur angle: A Radon transform is performed on the blur parameters of the dominant motion. The position of the maximum value in the transform result is used as the fringe direction. The motion blur direction is then determined using the fringe direction, which is used to calculate the global blur direction. ； ； in, For fuzzy directions within each primary cluster, This is the weighting factor for the region. For global fuzzy direction, For the defined primary cluster domain, For global fuzzy direction, It is the arctangent function; Get the global fuzz length: Based on a robust fuzzy length fusion strategy using weighted median, all data points within the main cluster are sorted by fuzzy length from smallest to largest, and the total weight and cumulative weight of the sorted data are calculated. Using the total weight to set a third target threshold, data points whose first cumulative weight is not less than the third target threshold are extracted and used as the global fuzzy length.

5. The UAV low-altitude visual navigation method based on motion fuzzy assistance according to claim 1, characterized in that, The method also includes: Calculate the confidence scores of the fuzzy parameters to construct the adaptive weights: ； ； ； ； in, For confidence level, For spectral stripe contrast factor, Noise level factor, The peak significance factor, + + =1, , , For preset coefficients, The maximum value of the Radon projection in the principal direction reflects the salience of the spectral fringes. This is the minimum value of the Radon projection, reflecting the background noise level. The contrast normalization threshold, The standard deviation of the Radon projection plot reflects the dispersion of the projected values. The mean of the Radon projection is... This is the second highest peak value, used to assess multi-peak interference. The difference between the main peak and the secondary peak reflects the prominence of the main peak. The term is a peak significance assessment for z-score, which measures the prominence of the main peak relative to the overall distribution.

6. The UAV low-altitude visual navigation method based on motion fuzzy assistance according to claim 1, characterized in that, The simulated global theoretical fuzzy data includes: Pre-integrating the IMU data yields the relative motion between adjacent frames, including relative rotation integral, relative velocity integral, and relative position integral. Based on the system's prior state estimation, the planar motion field of the image at the current moment is predicted. ； ； ； in, Let be the prior velocity at time k+1. This is the optimal velocity estimate at time k. The velocity increment produced by gravitational acceleration within the time interval. To rotate the velocity increments from the body coordinate system to the world coordinate system, Let k+1 be the prior position. This is the optimal position estimate at time k. The position increment generated by uniform motion, The position increment caused by gravitational acceleration. To rotate the displacement increments in the body coordinate system to the world coordinate system; Based on the image planar motion field and the monocular camera exposure time, simulate the global theoretical blur data, namely, motion blur direction, motion blur length, and image sharpness: ； ； ； in, For motion ambiguity direction, For the motion blur length, For image clarity, For average motion, For camera exposure time, This is the resolution scale factor.

7. The UAV low-altitude visual navigation method based on motion fuzzy assistance according to claim 1, characterized in that, The adaptive weights include: ； =c， ， ； in, For adaptive weights, For the confidence level of the fuzzy parameter, For motion state factors, Here, c represents the system degradation factor, and c represents the confidence level. This represents the magnitude of the current velocity estimate. For velocity scaling factor, This represents the number of feature points successfully tracked in the current frame. These are the typical characteristic numbers of a system in a healthy state. Based on weights, For quality factor weights, For speed factor weights, The weights are for the degradation factor.

8. The UAV low-altitude visual navigation method based on motion fuzzy assistance according to claim 1, characterized in that, Constructing the nonlinear least squares optimization objective function includes: ； ； ； ； in, The objective function is a nonlinear least squares optimization function. For state vectors, This is the visual reprojection error term. For the IMU pre-integration error term, For motion fuzzy consistency error term, For adaptive weights, This is the set of keyframes available for fuzzy analysis within the sliding window. The covariance matrix of the motion fuzzy parameter measurements. This refers to the number of keyframes. For keyframe indexing, The number of feature points, For feature point index, For Huber robust kernel function, This is the visual reprojection error vector. For keyframe pair indexing, This is due to visual reprojection error. This refers to the motion fuzziness consistency error.

9. The low-altitude visual navigation method for UAVs based on motion fuzzy assistance according to claim 1, characterized in that, Implementing the three-level mode switching includes: Calculate your health score: ； ； ； ； in, As a health score, The average tracking length for all feature points. To achieve the required minimum average tracking length, For the error threshold, This represents the number of feature points successfully tracked in the current frame. These are the typical characteristic numbers of a system in a healthy state. For confidence level, As a characteristic quality indicator, As an IMU consistency indicator, For fuzzy consistency indicators, The norm of the IMU error. Let be the norm of the motion blur error; When the health score equals the fourth target threshold, it proves that the system is in an ideal working state; When the health score drops to the first threshold range, it enters the fuzzy-assisted mode to rely on motion fuzzy constraints. When the health score drops to the second threshold range, the system switches to fuzzy-dominated mode to forcibly stabilize the system using fuzzy information.

Images