Unmanned aerial vehicle high-precision positioning method and system based on beidou difference and visual constraint

Abstract

The application discloses a Beidou differential and visual constraint-based high-precision positioning method and system of a UAV, and belongs to the field of high-precision positioning of a UAV. The method comprises the following steps: acquiring Beidou satellite observation data corresponding to the UAV and differential observation data corresponding to the Beidou satellite observation data, constructing a double-difference carrier phase observation equation according to the Beidou satellite observation data and the differential observation data, determining a position float solution, carrier phase float ambiguity and a corresponding ambiguity covariance matrix of the UAV, determining a candidate position change amount corresponding to a current epoch according to each integer ambiguity candidate, generating a visual motion constraint corresponding to the current epoch, determining a feasible ambiguity candidate set satisfying the visual motion constraint, obtaining a high-precision positioning result of the UAV, and improving the continuity and reliability of the positioning result of the UAV in a complex low-altitude environment.

Description

Technical Field

[0001] This invention relates to the field of high-precision positioning of unmanned aerial vehicles (UAVs), and more particularly to a high-precision positioning method and system for UAVs based on BeiDou differential positioning and visual constraints. Background Technology

[0002] With the large-scale application of drones in fields such as aerial surveying, precision agriculture, power line inspection, and logistics distribution, their positioning capabilities have gradually shifted from traditional meter-level navigation to centimeter-level positioning. The basic positioning service provided by the BeiDou Navigation Satellite System has an accuracy of 5-10 meters, which is insufficient to meet the application needs of specific scenarios. However, RTK technology based on the carrier phase differential principle works in collaboration between a base station and a rover mounted on the drone. The base station continuously receives satellite signals and calculates observation errors, transmitting the correction data to the drone in real time via a communication link. The rover combines the satellite signals it receives with the correction data and uses differential algorithms to eliminate common errors such as ionospheric delay and satellite orbital errors. Ultimately, the planar positioning accuracy of the drone can be improved to the centimeter level, becoming a core technology supporting precise aerial surveying and autonomous operations.

[0003] In RTK (Real-Time Kinematic) positioning systems, carrier phase observations possess millimeter-level ranging potential, but this relies on the correct fixing of integer ambiguities in the carrier phase. Existing technologies typically employ an Integer Least Squares (ILS) framework to resolve ambiguities. This involves using LAMBDA or its extended algorithm MLAMBDA to perform an integer search on the floating-point ambiguity solution, obtaining the optimal integer vector and its suboptimal candidates. Subsequently, verification methods such as Ratio-Test are used to calculate the ratio between the suboptimal and optimal solutions in terms of the weighted sum of squared residuals. If this ratio exceeds a preset threshold, the ambiguity is locked as a fixed solution, and the rover's precise coordinates are output accordingly; otherwise, it degenerates into a floating-point solution. This pipeline of integer search—residual ratio verification—fixed solution output constitutes the mainstream solution paradigm for current BeiDou differential UAV positioning.

[0004] However, the above paradigm has inherent reliability vulnerabilities. In actual UAV operating environments, complex conditions such as urban canyons, dense buildings, high-voltage power lines, mountain obstructions, and electromagnetic interference are widespread. Satellite signals are easily blocked and generate severe multipath effects, resulting in systematic biases in carrier phase observations that are not absorbed by the model. In this case, even if integer search yields a seemingly reasonable optimal integer vector, and the residual ratio given by Ratio-Test exceeds a preset threshold, it does not mean that the spatial location corresponding to the fixed solution is truly reliable—the residual ratio only reflects the statistical difference between the optimal integer candidate and the second-best candidate, without cross-validating the physical rationality of the solution coordinates themselves. Once ambiguity is incorrectly fixed, the positioning results will exhibit substantial gross errors masked by centimeter-level claimed accuracy, while the system will continue to output a fixed solution state. This pseudo-fixation can easily lead to track deviations, flight path overruns, and even flight safety accidents in scenarios such as high-speed UAV maneuvers and low-altitude crossings. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a high-precision positioning method and system for unmanned aerial vehicles based on BeiDou differential and visual constraints.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a high-precision positioning method for unmanned aerial vehicles (UAVs) based on BeiDou differential and visual constraints, comprising the following steps:

[0007] S1: Obtain the BeiDou satellite observation data corresponding to the UAV and the differential observation data corresponding to the BeiDou satellite observation data. Construct the double-difference carrier phase observation equation based on the BeiDou satellite observation data and the differential observation data to determine the UAV's position floating-point solution, carrier phase floating-point ambiguity and the corresponding ambiguity covariance matrix.

[0008] S2: Perform an integer search based on the carrier phase floating-point ambiguity and ambiguity covariance matrix to obtain multiple integer ambiguity candidates, and determine the candidate position change corresponding to the current epoch based on each integer ambiguity candidate;

[0009] S3: Acquire continuous images that correspond to the BeiDou satellite observation data in time, determine the visual relative motion information of the UAV based on the inter-frame changes of static scene features in the continuous images, and generate the visual motion constraints corresponding to the current epoch based on the visual relative motion information.

[0010] S4: Determine the motion consistency between the change in each candidate position and the visual motion constraint, and determine the feasible ambiguity candidate set that satisfies the visual motion constraint from multiple integer ambiguity candidates based on the motion consistency.

[0011] S5: Perform carrier phase observation consistency check on the feasible ambiguity candidate set, determine the target integer ambiguity based on the check results, and perform BeiDou differential carrier phase positioning based on the target integer ambiguity to obtain high-precision positioning results for the UAV.

[0012] In a preferred embodiment of the present invention, in S3, the visual relative motion information of the UAV is determined based on the inter-frame changes of static scene features in consecutive images, specifically as follows:

[0013] Feature extraction and feature matching are performed on adjacent images or keyframe images with a preset time interval. Abnormal and dynamic features are eliminated based on the geometric consistency of the matched features. The visual motion direction, scale-free visual displacement, and visual motion credibility of the UAV are determined based on the remaining matched features.

[0014] In a preferred embodiment of the present invention, determining the visual relative motion information of the UAV based on the inter-frame changes of static scene features in consecutive images further includes:

[0015] The historical epochs in which the BeiDou differential carrier phase positioning is continuously in a reliable fixed state are obtained. The visual scale factor is determined based on the proportional relationship between the BeiDou position change corresponding to the historical epoch and the scale-free visual displacement. The scale correction parameter is determined based on the visual scale factor corresponding to multiple historical epochs.

[0016] When the BeiDou differential carrier phase positioning does not meet the reliable fixed state, the scale correction parameters are stopped from being updated, and the scale correction parameters determined before the update was stopped are used to perform scale correction on the current scaleless visual displacement.

[0017] In a preferred embodiment of the present invention, the credible fixed state is determined based on at least two of the following: the number of consecutive fixed epochs, the integer ambiguity test result, the double-difference carrier phase residual, the satellite geometry, and the continuity of motion between adjacent epochs.

[0018] The scale correction parameters are updated using historical epochs in a reliable fixed state so that integer ambiguity candidates can be constrained by a determined visual scale when the BeiDou positioning quality deteriorates.

[0019] In a preferred embodiment of the present invention, in step S3, acquiring continuous images that correspond in time to BeiDou satellite observation data includes:

[0020] The exposure time of each image and the observation epoch of the BeiDou satellite observation data are obtained. The exposure time is converted to the BeiDou time reference. The BeiDou motion information corresponding to the exposure time is determined according to the position or motion state corresponding to the adjacent BeiDou satellite observation data, so that the candidate position change and the visual relative motion information correspond to the same time interval.

[0021] In a preferred embodiment of the present invention, generating visual motion constraints corresponding to the current epoch based on visual relative motion information includes:

[0022] Visual motion confidence is determined based on at least two of the following: the number of matching features, the proportion of geometrically consistent inliers, the reprojection residual, and the spatial distribution of matching features. Allowable displacement and orientation deviations are determined based on visual motion confidence. A feasible region for visual motion is constructed based on the scale-corrected visual displacement, visual motion direction, displacement deviation, and orientation deviation.

[0023] Integer ambiguity candidates whose corresponding positional change is within the feasible region of visual motion are retained in the feasible ambiguity candidate set.

[0024] In a preferred embodiment of the present invention, the high-precision positioning method further includes:

[0025] Based on visual relative motion information, the predicted double-difference carrier phase change of each Beidou satellite between adjacent epochs is determined. The predicted double-difference carrier phase change is compared with the corresponding actual double-difference carrier phase change to determine the degree of observation anomaly of each Beidou satellite.

[0026] Adjust the solution weights of the corresponding BeiDou satellite observation data according to the degree of observation anomalies, or remove BeiDou satellite observation data that meet the anomaly conditions, and redetermine the integer ambiguity candidates based on the adjusted BeiDou satellite observation data.

[0027] A high-precision positioning system for unmanned aerial vehicles (UAVs) based on BeiDou differential and visual constraints includes:

[0028] The solution module is used to acquire the BeiDou satellite observation data and differential observation data corresponding to the UAV, construct the double-difference carrier phase observation equation, determine the floating-point solution of the position, the floating-point ambiguity of the carrier phase and the ambiguity covariance matrix, and obtain multiple integer ambiguity candidates and the candidate position change corresponding to each integer ambiguity candidate based on the floating-point ambiguity of the carrier phase and the ambiguity covariance matrix.

[0029] The generation module is used to acquire continuous images that correspond to the BeiDou satellite observation data in time, determine visual relative motion information based on the inter-frame changes of static scene features in the continuous images, and generate visual motion constraints based on the visual relative motion information.

[0030] The filtering module is used to determine the motion consistency between the change in each candidate position and the visual motion constraints, and to determine the feasible ambiguity candidate set that satisfies the visual motion constraints based on the motion consistency.

[0031] The output module is used to perform carrier phase observation consistency checks on the feasible ambiguity candidate set, determine the target integer ambiguity, perform BeiDou differential carrier phase positioning based on the target integer ambiguity, and output the high-precision positioning results of the UAV.

[0032] In a preferred embodiment of the present invention, a management module and an alignment module are also included;

[0033] The management module is used to determine the scale correction parameters based on the BeiDou position change and scale-free visual displacement corresponding to the historical epochs when the BeiDou differential carrier phase positioning is in a credible fixed state, and to stop updating the scale correction parameters when the BeiDou differential carrier phase positioning does not meet the credible fixed state.

[0034] The alignment module is used to convert the image exposure time to the BeiDou time reference and determine the BeiDou motion information corresponding to the image exposure time, so that the visual relative motion information and the candidate position change amount correspond to the same time interval.

[0035] In a preferred embodiment of the present invention, a management module is further included. The management module is used to determine the predicted double-difference carrier phase change amount corresponding to each Beidou satellite based on visual relative motion information, determine the observation anomaly degree of the corresponding Beidou satellite based on the deviation between the predicted double-difference carrier phase change amount and the actual double-difference carrier phase change amount, and adjust the solution weight of the corresponding Beidou satellite observation data or remove Beidou satellite observation data that meet the anomaly conditions based on the observation anomaly degree.

[0036] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0037] (1) This invention provides a high-precision positioning method and system for UAVs based on BeiDou differential and visual constraints. Multiple integer ambiguity candidates are obtained by BeiDou differential carrier phase floating point solution, and the candidate position change corresponding to each integer ambiguity candidate is determined. Then, visual motion constraints are formed by using visual relative motion information obtained from continuous images to screen the candidate position change for motion consistency. Finally, the target integer ambiguity is determined by using carrier phase observation consistency. Visual information participates in candidate discrimination before the integer is fixed, so that candidates that do not match the actual movement direction or displacement amplitude of the UAV can be eliminated in advance. Compared with the method of directly selecting a fixed solution based solely on carrier phase residual or integer search results, this method can reduce the probability of incorrect integer candidates being fixed due to satellite geometric configuration variation, multipath interference or local observation anomalies, and improve the continuity and reliability of UAV positioning results in complex low-altitude environments.

[0038] (2) This invention provides a high-precision positioning method and system for UAVs based on BeiDou differential and visual constraints. By corresponding the BeiDou position change with the scale-free visual displacement according to the historical epoch of the BeiDou differential carrier phase positioning in a credible fixed state, the visual scale factor and scale correction parameters are determined. When the BeiDou positioning does not meet the credible fixed state, the scale correction parameters are stopped. The current visual displacement is corrected using the scale parameters before the stop of updating, so that the visual displacement can obtain a scale corresponding to the actual spatial movement. At the same time, it avoids the abnormal BeiDou positioning results from changing the visual scale in reverse. Compared with the method of continuously updating the visual scale using the current positioning results during the BeiDou observation quality change process, it can reduce the propagation of error fixing, observation gross error and short-term positioning fluctuations on visual motion constraints, so that the visual displacement scale can still be maintained during the period of BeiDou observation quality decline, and further improve the effectiveness of integer ambiguity candidate screening.

[0039] (3) This invention provides a high-precision positioning method and system for UAVs based on BeiDou differential and visual constraints. The visual motion credibility is determined by the number of matching features, the proportion of geometrically consistent in-points, the reprojection residual, and the spatial distribution of matching features. The allowable displacement deviation and direction deviation are adjusted accordingly. At the same time, the image exposure time is converted to the BeiDou time reference so that the candidate position change and the visual relative motion information correspond to the same time interval. This allows the visual constraint strength to change with the actual image quality and reduces the time misalignment error caused by the different sampling frequencies of the camera and BeiDou. Compared with the method of using a fixed visual constraint threshold or directly pairing adjacent sampling data, it can reduce the phenomenon of incorrect rejection of correct candidates caused by weak texture, motion blur, dynamic targets, and sensor asynchrony, and make the visual motion consistency more accurately reflect the actual motion difference corresponding to different integer ambiguity candidates.

[0040] (4) This invention provides a high-precision positioning method and system for UAVs based on BeiDou differential and visual constraints. By predicting the double-difference carrier phase change of each BeiDou satellite between adjacent epochs based on visual relative motion information, and comparing the predicted value with the actual double-difference carrier phase change, the degree of anomaly in each satellite observation is determined. Based on this, the solution weight of the corresponding observation data is adjusted or the abnormal observation is eliminated. Then, floating-point solution and integer candidate search are performed again. This process can weaken the influence of satellite observations that are inconsistent with the overall motion of the UAV on floating-point ambiguity and its covariance before integer search. Compared with the method of identifying abnormal observations mainly based on satellite elevation angle, carrier-to-noise ratio or residuals after solution, it can add independent discrimination criteria based on the actual motion state of the UAV. Visual constraints can be applied to both satellite observation quality management and integer candidate screening at the same time, further reducing the impact of abnormal satellite observations under occlusion and multipath conditions on the final positioning result. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of a preferred embodiment of the present invention;

[0043] Figure 2 This is a flowchart of a preferred embodiment of the present invention for acquiring visual relative motion information and generating visual motion constraints;

[0044] Figure 3 This is a flowchart illustrating a preferred embodiment of the present invention for filtering a feasible ambiguity candidate set based on motion consistency. Detailed Implementation

[0045] 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.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0047] Application Overview:

[0048] This application is applicable to scenarios such as low-altitude flight of UAVs, partial obstruction by tall buildings or mountains, multipath effects caused by ground or building reflections, short-term changes in the number of visible satellites, and fluctuations in carrier phase observation quality. BeiDou differential carrier phase positioning can obtain centimeter-level position results when integer ambiguity is correctly fixed. However, the process of fixing integer ambiguity mainly depends on the geometric relationship of satellite observation and the statistical characteristics of carrier phase. When the satellite geometry deteriorates, gross errors occur in local observations, or multipath deviations occur, if candidate solutions are selected only according to carrier phase residuals or traditional integer search targets, different integer candidates may show similar observation fitting degrees in a short period of time, thereby increasing the probability of incorrect fixing.

[0049] The camera on the drone can continuously observe the surrounding static scene. For adjacent images or key frame images, the inter-frame changes of static features in the image plane have a direct geometric relationship with the relative motion of the drone itself. Compared with satellite carrier phase observation, visual relative motion does not depend on the satellite geometry. Therefore, it can provide independent motion direction and relative displacement information when the BeiDou observation quality degrades. This application does not use visual results to directly replace BeiDou positioning, but converts visual relative motion into motion constraints on the position change of integer ambiguity candidates, so that visual information participates in the discrimination of integer candidates.

[0050] The BeiDou differential calculation branch of this application first obtains the floating-point solution of the position, the floating-point ambiguity of the carrier phase, and the ambiguity covariance matrix, and then performs an integer search and retains multiple integer ambiguity candidates. Each integer ambiguity candidate can be back-calculated to obtain a corresponding candidate position, and further form the candidate position change between adjacent epochs. The visual motion branch performs feature extraction, matching, and geometric consistency screening on the continuous images corresponding to the time to obtain the visual motion direction, scale-free visual displacement, and visual motion credibility. When there are credible BeiDou fixed results, the visual scale is continuously estimated, and the scale is frozen when the BeiDou quality deteriorates.

[0051] During the candidate selection stage, the scale-corrected visual displacement, visual motion direction, and allowable displacement and direction deviations determined by the visual motion confidence level are combined into a feasible region for visual motion. Only when the candidate position change corresponding to each integer ambiguity candidate falls into this feasible region will it enter the subsequent carrier phase observation consistency test. Thus, visual information first undertakes the role of candidate space compression, and carrier phase observation then undertakes the role of final fixation confirmation. The two types of information are functionally independent but sequentially coupled, which can reduce the direct impact of a single observation source anomaly on the integer fixation result.

[0052] Furthermore, this application uses visual relative motion to predict the geometric distance change between adjacent epochs and projects it onto the double-difference observation direction formed by each Beidou satellite and the reference satellite to obtain the predicted double-difference carrier phase change. The deviation between the predicted value and the actual double-difference carrier phase change can reflect whether the corresponding satellite observation is consistent with the overall motion of the UAV. When the deviation of a certain satellite continues or increases significantly, the solution weight of that satellite is reduced or it is removed, and floating-point solution and integer candidate search are re-executed, so that visual information is not only used for candidate screening, but also for Beidou observation quality management.

[0053] Example 1:

[0054] like Figure 1 As shown, this embodiment provides a high-precision positioning method for unmanned aerial vehicles (UAVs) based on BeiDou differential positioning and visual constraints, including the following steps:

[0055] S1: Obtain the BeiDou satellite observation data corresponding to the UAV and the differential observation data corresponding to the BeiDou satellite observation data. Construct the double-difference carrier phase observation equation based on the BeiDou satellite observation data and the differential observation data to determine the UAV's position floating-point solution, carrier phase floating-point ambiguity and the corresponding ambiguity covariance matrix.

[0056] S2: Perform an integer search based on the carrier phase floating-point ambiguity and ambiguity covariance matrix to obtain multiple integer ambiguity candidates, and determine the candidate position change corresponding to the current epoch based on each integer ambiguity candidate;

[0057] S3: Acquire continuous images that correspond to the BeiDou satellite observation data in time, determine the visual relative motion information of the UAV based on the inter-frame changes of static scene features in the continuous images, and generate the visual motion constraints corresponding to the current epoch based on the visual relative motion information.

[0058] S4: Determine the motion consistency between the change in each candidate position and the visual motion constraint, and determine the feasible ambiguity candidate set that satisfies the visual motion constraint from multiple integer ambiguity candidates based on the motion consistency.

[0059] S5: Perform carrier phase observation consistency check on the feasible ambiguity candidate set, determine the target integer ambiguity based on the check result, and perform BeiDou differential carrier phase positioning based on the target integer ambiguity to obtain high-precision positioning results for the UAV.

[0060] In step S1, the UAV is equipped with a BeiDou high-precision positioning receiver and a camera. The BeiDou high-precision positioning receiver serves as the rover receiver. The differential observation data can come from a ground reference station set at a known coordinate point, or from a continuously operating reference station network that can provide equivalent differential correction information to the UAV. The rover and the reference station acquire data including at least carrier phase, pseudorange, satellite number, carrier frequency, satellite elevation angle, signal-to-noise ratio, and observation time scale at the same observation epoch. In this embodiment, the BeiDou observation frequency can be 1-20 Hz, and 10 Hz is used in this embodiment. The camera acquisition frequency can be 20-60 Hz, and 40 Hz is used in this embodiment.

[0061] To mitigate the impact of receiver clock bias, satellite clock bias, and common propagation error on carrier phase calculation, a single inter-station difference is first formed between the rover station and the base station for the same satellite at the same frequency point. Then, a double inter-satellite difference is formed using a satellite with stable visibility, a high satellite elevation angle, and continuous carrier phase as the reference satellite. Let the current epoch be k, the rover station be u, the base station be b, the satellite to be solved be i, and the reference satellite be r. The carrier phase observations are uniformly converted to m.

[0062] The formula for double-difference carrier phase observation is: .

[0063] In the above double-difference carrier phase observation formula, The double-difference carrier phase observation of satellite i relative to reference satellite r is expressed in meters. The geometric distance is the double difference between the rover, the base station, and the two satellites, expressed in meters. The corresponding frequency point carrier wavelength, in meters (m); The ambiguity is a double-difference integer; This includes the remaining observation noise, multipath residuals, and propagation errors that were not completely eliminated after double difference.

[0064] After obtaining the approximate location of the rover, the double-difference geometric distance is linearized to the rover's location in the first order. The observation equations of each non-reference satellite involved in the calculation at the same epoch are stacked row by row to obtain a linear observation model composed of the position increment and the double-difference ambiguity. For multi-frequency observations, the ambiguity of each frequency point can be established separately and the corresponding carrier wavelength can be set in the wavelength matrix.

[0065] The linearized double-difference observation formula is as follows .

[0066] In the above linearized double-difference observation formula, The double-difference observation vector is the result of subtracting the approximate geometric distance from the observations. The design matrix for the UAV's three-dimensional position based on the double-difference geometric distance; The three-dimensional position increment to be estimated; It is a coefficient matrix composed of carrier wavelengths at each frequency point; This is the double-difference carrier phase ambiguity vector; This is a joint design matrix formed by splicing the position design matrix and the wavelength matrix; It is a joint unknown consisting of the position increment and the ambiguity.

[0067] Different observations are weighted based on satellite elevation angle, signal-to-noise ratio (SNR), and historical residuals. The lower the satellite elevation angle or the lower the SNR, the smaller the corresponding observation weight. For example, the initial weights can be constructed by multiplying the elevation angle weights and the SNR weights, and the residuals from the previous epoch can be used to further reduce the weights of anomalous observations, thereby forming a weight matrix P. k We perform weighted least squares or equivalent recursive filtering on the linearized double-difference observation model to obtain the position floating-point solution, carrier phase floating-point ambiguity, and joint covariance matrix.

[0068] The formula for solving floating-point states is: ;

[0069] The formula for calculating floating-point covariance is: .

[0070] Depend on Extract the submatrix corresponding to the carrier phase floating-point ambiguity to obtain the ambiguity covariance matrix. The fuzziness covariance matrix reflects the correlation between different double-difference fuzzies and the degree of uncertainty of each fuzzyness, and serves as the basis for subsequent integer search to determine the candidate order and search range.

[0071] In step S2, several statistically superior integer candidates are retained in the integer space determined by the floating-point ambiguity and its covariance. The integer search can be performed using integer least squares search, LAMBDA decorrelation search, or equivalent tree search methods. During the search process, the ambiguity covariance matrix is ​​first decorrelated, and then the search is performed in the transformed integer space according to the objective function from small to large.

[0072] Integer candidate evaluation formula is .

[0073] In the above integer candidate evaluation formula, Candidate vectors for integer ambiguity; The carrier phase floating-point ambiguity vector obtained in step S1; The ambiguity covariance matrix; The smaller the value, the more consistent the integer candidate is with the statistical distribution of floating-point ambiguity.

[0074] In this embodiment, the top K integer ambiguity candidates with the smallest objective function are retained. K can be 2-20, and in this embodiment, it is 10. If the actual number of candidates obtained within the search radius is less than K, all candidates are retained. Retaining multiple candidates allows the subsequent visual motion constraints to play a filtering role before the integers are fixed, avoiding the risk of incorrect fixation caused by first fixing a single integer solution and then passively discovering anomalies from the positioning results in the traditional process.

[0075] For the j-th integer ambiguity candidate, substitute it back into the double-difference carrier phase observation model and fix the ambiguity, then only estimate the three-dimensional position increment.

[0076] The formula for calculating the candidate position is as follows: .

[0077] In the above candidate position calculation formula, This is the j-th integer ambiguity candidate; The position increment corresponding to the candidate is then superimposed onto the linearized approximate position to obtain the candidate position for the current epoch. .

[0078] The change in candidate position is determined by the difference between the candidate position in the current epoch and the accepted position in the previous epoch. The accepted position in the previous epoch can be a fixed position that has passed the carrier phase consistency test. If no fixed result is formed in the previous epoch, the reference position obtained by combining the most recent reliable fixed position with short-time velocity extrapolation can be used.

[0079] The formula for the change in candidate position is: .

[0080] In the above formula for the change in candidate positions, Let be the three-dimensional motion vector corresponding to the j-th integer ambiguity candidate in the current time interval, in meters; This refers to the accepted position from the previous epoch or the reference position corresponding to the starting point of the current epoch. Subsequent visual constraints do not directly compare absolute coordinates, but rather compare relative motion within the same time interval, thereby reducing the impact caused by the difference between the visual coordinate origin and the BeiDou absolute coordinate origin.

[0081] like Figure 2 As shown, in step S3, the camera is fixedly mounted on the UAV body. The camera intrinsic parameters are calibrated before the operation, and the extrinsic rotation relationship between the camera coordinate system and the UAV coordinate system is obtained. Feature extraction and feature matching are performed on adjacent images or key frame images with a preset time interval. Feature extraction can use ORB, SIFT, AKAZE, corner point and optical flow tracking combination or other methods that can form a stable inter-frame correspondence. In this embodiment, ORB features and pyramid optical flow tracking are combined to extract 800-2000 feature points in texture-rich areas. In this embodiment, the number of target features per frame is 1200.

[0082] After feature matching, descriptor distance filtering and bidirectional consistency checks are performed first. Then, random sampling consistency estimation of the fundamental matrix or essential matrix is ​​used to eliminate abnormal matches that do not satisfy epipolar geometric relationships. For features whose motion trajectories are significantly inconsistent with the geometric model of the main scene in multiple consecutive frames, they are judged as dynamic features and excluded from the current visual motion estimation. The matching features retained are mainly from building facades, roads, ground textures, fixed facilities and other static scenes.

[0083] The geometric constraint formula for the visual epipolar line is: .

[0084] In the above formula for geometric constraints on visual epipolar lines, and These are the normalized imaging coordinates of the same static spatial feature in the two consecutive frames; The essential matrix; This refers to the relative rotation between two camera frames; This is a translation vector that determines only the direction and relative scale. By using essential matrix decomposition and combining it with the positive depth condition, the visual motion direction and scale-free visual displacement can be obtained.

[0085] The translation amount obtained from the geometric relationship between two frames in monocular vision lacks an absolute scale, so it is recorded as scale-free visual displacement. The visual displacement is first rotated to the UAV system through the camera extrinsic parameters, and then transformed to the local navigation coordinate system used by BeiDou positioning based on the UAV attitude information or the directional correspondence between the historical reliable BeiDou displacement and the visual displacement. In this embodiment, the East-North-Sky local coordinate system is used as the unified comparison coordinate system.

[0086] Step S3 simultaneously calculates the visual motion confidence level. The visual motion confidence level is determined by at least two of the following: the number of matching features, the proportion of geometrically consistent inliers, the reprojection residual, and the spatial distribution of matching features. This embodiment uses four indicators simultaneously, so that the visual constraint strength can vary with image texture, occlusion, dynamic target proportion, and motion blur, rather than using a fixed threshold under all image conditions.

[0087] like Figure 3 As shown, in step S4, for each candidate position change corresponding to an integer ambiguity candidate, the difference between the candidate position change and the visual motion direction and the visual displacement after scale correction is calculated. The displacement difference reflects whether the candidate motion distance is consistent with the visual observation, and the direction difference reflects whether the candidate motion direction is consistent with the visual observation. For cases where the UAV is nearly stationary for a short time and the displacement amplitude is lower than the motion judgment threshold, the unstable direction angle is not calculated, but the three-dimensional displacement magnitude and the absolute deviation of each coordinate axis component are used for judgment.

[0088] The formula for candidate displacement deviation is: ;

[0089] The candidate direction deviation formula is: .

[0090] In the two formulas above, The visual displacement after coordinate transformation and scale correction is expressed in meters (m). This represents the displacement amplitude deviation, in meters (m). This represents the directional angular deviation, expressed in degrees. When the visual motion reliability is high, the allowable displacement and directional deviations decrease; when the visual motion reliability decreases but still meets the usable conditions, the corresponding allowable deviations increase.

[0091] Integer ambiguity candidates that simultaneously satisfy the displacement deviation threshold and the orientation deviation threshold are retained in the feasible ambiguity candidate set. If the feasible set contains only one candidate, the carrier phase observation consistency check in step S5 is still performed instead of directly outputting the fixed result. If the feasible set contains multiple candidates, the carrier phase residuals of each candidate are further compared through step S5. If the feasible set is empty, no forced integer fixing is performed in this epoch, and the visual feasible domain can be expanded once or back to the floating-point solution to avoid incorrect fixing.

[0092] In step S5, feasible ambiguity candidates are substituted one by one into the double-difference carrier phase observation model, and the double-difference carrier phase residual, weighted residual sum of squares, and residual distribution of each satellite are calculated after fixing the candidate.

[0093] The formula for carrier phase consistency residual is: ;

[0094] The formula for the carrier phase consistency index is: .

[0095] The feasible candidates are arranged in ascending order of carrier phase consistency index, and the target integer ambiguity is determined by combining the root mean square of residuals, the ratio of the index of the best candidate to the second best candidate, and the existence of significant residuals of a single satellite. In this embodiment, the best candidate is determined as the target integer ambiguity when the root mean square of the fixed double-difference carrier phase residual is not greater than 0.03 m and the ratio of the carrier phase consistency index of the second best candidate to the best candidate is not less than 2.5. When the above conditions are not met, the floating point state is maintained or the determination is delayed to the next epoch.

[0096] Once the target integer ambiguity is determined, the UAV's 3D position is recalculated using this integer ambiguity as a fixed value. The resulting position, fixed state, set of satellites involved in the calculation, residual statistics, and visual motion confidence level are then output together. This forms a complete positioning link from BeiDou floating-point solution, integer candidate generation, visual motion screening to carrier phase confirmation.

[0097] Example 2:

[0098] The technical solution in this embodiment is further refined based on the above embodiment 1, focusing on visual scale correction, reliable fixed state management, visual motion feasible domain construction, and time alignment between image exposure time and BeiDou observation epoch.

[0099] Monocular visual relative motion can stably provide the translation direction, but its translation magnitude has arbitrary scale. If the scale-free visual displacement is directly used for integer ambiguity candidate screening, only the direction information can be used, and it is impossible to fully distinguish integer candidates with similar directions but different displacement amplitudes. This application uses the historical epochs of BeiDou differential carrier phase positioning that are continuously in a reliable and fixed state to establish a real-time scale for visual displacement, so that the high-quality stage of BeiDou becomes the calibration stage of the visual scale; when the BeiDou quality decreases, the scale is stopped from being updated to avoid erroneous or unstable BeiDou results from contaminating the visual constraints.

[0100] Specifically, at each historical epoch h that satisfies the credible fixed state, the change in BeiDou position between that epoch and the previous corresponding moment is taken, and the scale-free visual displacement within the same time interval is obtained. After both are transformed to a unified East-North-Sky coordinate system, the visual scale factor is calculated. To avoid scale anomalies caused by an excessively small denominator when the UAV is approximately stationary, the scale is updated only when the scale-free visual displacement modulus is greater than the preset minimum motion and the change in BeiDou position is greater than 0.02 m.

[0101] The formula for visual scale factor is: .

[0102] In the above formula for visual scale factor, This refers to the historical position change obtained from a fixed BeiDou location under a reliable fixed state, expressed in meters (m). Scale-free visual displacements that have completed coordinate alignment within the same time interval; This is a scaling factor for converting scaleless visual displacement into actual displacement scale.

[0103] To suppress the influence of instantaneous visual errors or slight BeiDou residuals on individual historical epochs, this embodiment retains the most recent M effective visual scale factors, where M can be 5 to 30, and is set to 20 in this embodiment. First, outlier screening is performed with the median as the center, and scale factors that deviate from the median by more than 2.5 times the median absolute deviation are removed. Then, the weights are determined according to the BeiDou fixed quality and visual motion reliability of each historical epoch, and the scale correction parameters are calculated.

[0104] The formula for the scale correction parameter is: .

[0105] In the above formula for scale correction parameters, The comprehensive weight for the h-th historical epoch can be determined by the BeiDou residual, satellite geometry, and visual motion reliability. This is the currently saved scale correction parameter. The current scale-free visual displacement is multiplied by... The scale-corrected visual displacement was then obtained.

[0106] The formula for visual displacement scale correction is: .

[0107] In the above visual displacement scale correction formula, This is the rotation transformation matrix from the visual coordinate system to the BeiDou local navigation coordinate system; This represents the scale-free visual displacement within the current time interval; The scale-corrected visual displacement used for candidate selection is in meters.

[0108] A credible fixed state is not determined by a single integer fixed criterion, but rather by a combination of at least two of the following: the number of consecutive fixed epochs, the integer ambiguity ratio test result, the double-difference carrier phase residual, the satellite geometry, and the continuity of motion between adjacent epochs. This embodiment uses five criteria: the number of consecutive fixed epochs is not less than 5; the integer ambiguity ratio test result is not less than 3.0; the root mean square of the fixed double-difference carrier phase residual is not greater than 0.03 m; the position accuracy factor (PDOP) is not greater than 3.0; and the deviation between the current BeiDou position change and the position change predicted by the estimated velocity of the most recent three fixed epochs is not greater than 0.20 m. An epoch is recorded as a credible fixed state if at least three of these criteria are met and there is no obvious cycle slip indicator.

[0109] When the BeiDou differential carrier phase positioning does not meet the credible fixed state, the current BeiDou result is immediately stopped from updating the scale correction parameters, and the scale correction parameters formed in the most recent credible state are retained. During the freeze period, the visual branch continues to update the relative motion through the image, but the scaleless displacement is converted into the actual displacement using the frozen scale. The scale update is only resumed when the BeiDou positioning continuously meets the credible fixed state again and reaches the preset recovery epoch. In this embodiment, the recovery epoch is 5. This hysteresis mechanism can avoid frequent fluctuations in the visual scale when the BeiDou state repeatedly switches near the fixed and non-fixed boundary.

[0110] The visual motion credibility is calculated as follows, assuming the number of effective matching features is N. v The geometrically consistent interior point ratio is η in The root mean square of the reprojection residual is e rep The percentage of grid cells with valid interior points after the image is divided into 4×4 grids is the spatial distribution index gsp. The four indices are normalized to 0-1 respectively.

[0111] The feature quantity scoring formula is as follows ;

[0112] The reprojection scoring formula is as follows: ;

[0113] The formula for calculating visual credibility is: .

[0114] In the above formula for calculating visual credibility, To achieve the number of reference features required for stable visual estimation, this embodiment uses 300. The reprojection error scale parameter is set to 2 px in this embodiment; , , and The weights are non-negative and sum to 1; in this embodiment, they are all taken as 0.25. For visual motion credibility.

[0115] When the visual motion confidence level is less than 0.30, the current epoch does not use visual information for hard filtering, but only retains the BeiDou carrier phase consistency check; when the visual motion confidence level is between 0.30 and 0.70, a wider visual motion feasible region is used; when the visual motion confidence level is greater than 0.70, a narrower visual motion feasible region is used. This avoids the incorrect exclusion of true integer candidates by low-quality visual results when there is weak texture, strong motion blur, or a large number of dynamic targets entering the frame.

[0116] The allowable displacement and orientation deviations vary continuously based on the visual motion confidence level. Let the lower and upper limits of displacement deviation be 0.05 m and 0.50 m, respectively, and the lower and upper limits of orientation deviation be 3° and 25°, respectively. The higher the confidence level, the closer the allowable deviations are to the lower limits.

[0117] The displacement tolerance adjustment formula is as follows: ;

[0118] The formula for adjusting the orientation tolerance is as follows: .

[0119] A feasible region for visual motion is constructed based on scale-corrected visual displacement, visual motion direction, displacement deviation, and direction deviation. For the j-th candidate, if its candidate displacement deviation is not greater than ε, the feasible region is determined. d And the directional deviation is no greater than ε θ When it is determined that it is within the feasible region of visual motion, the displacement deviation and orientation deviation can be standardized to form a continuous motion consistency index, which can be used to rank among multiple feasible candidates.

[0120] The formula for the motion consistency index is: .

[0121] In this embodiment, a hard screening is first performed using dual thresholds for displacement deviation and direction deviation. Then, the remaining candidates are arranged in ascending order of motion consistency index. The motion consistency index is only used to express the degree of proximity between the candidates and the visual motion, and does not directly replace the carrier phase observation consistency check in step S5, thus keeping the final integer fixed and still confirmed by the physical observation of the satellite carrier phase.

[0122] To ensure that the changes in candidate positions and visual relative motion information correspond to the same time interval, this embodiment performs a unified time reference conversion between image exposure time and BeiDou observation epoch. The camera records the exposure start time or exposure center time for each frame, and the BeiDou receiver records the BeiDou time corresponding to the observation epoch. If the camera supports hardware triggering, the offset relationship between the camera clock and BeiDou time is established using the 1PPS pulse output by the BeiDou receiver and the trigger count; if an independent clock is used, the clock offset and drift are periodically estimated during the startup phase and operation.

[0123] The time base conversion formula is: .

[0124] In the above time base conversion formula, Use the camera's local time; This is the converted BeiDou time. This refers to the clock scale or drift correction factor. This is the clock offset. Under hardware synchronization conditions. One value can be chosen, only estimating a fixed delay; under software synchronization conditions, multiple synchronization events are considered. and Perform linear fitting.

[0125] For exposure times located between two adjacent BeiDou observation epochs, the BeiDou motion information corresponding to the exposure time is determined based on the position or motion state of the adjacent BeiDou observations. When both BeiDou epochs are in a reliable fixed state, linear interpolation of the position can be performed; when the system simultaneously stores velocity states, constant velocity or constant acceleration short-time propagation can be used.

[0126] The BeiDou position-time interpolation formula is as follows: .

[0127] When visual relative motion is formed, the exposure time of the previous image and the exposure time of the current image are used as the visual motion interval, and the interpolation position of the same two moments is calculated in the BeiDou state sequence. The difference is used as the BeiDou position change that strictly corresponds to the visual motion.

[0128] For integer candidates, the current candidate position is propagated or interpolated to the same exposure time before the candidate position change is formed. This avoids the time misalignment error caused by directly pairing the nearest epoch under different sampling frequency conditions such as 40 Hz for the camera and 10 Hz for Beidou.

[0129] When the drone speed is 10m / s, if the visual time is out of 20ms with the BeiDou time, the time misalignment alone can produce a displacement difference of about 0.20m, which is enough to make the real integer candidate fall outside the narrow visual feasible domain. Therefore, this application converts the exposure time to the BeiDou time reference and performs time interpolation on the motion state, so that the displacement difference in the visual constraint mainly reflects the difference in real motion, rather than the asynchronous error of the sensor.

[0130] Example 3:

[0131] The technical solution of this embodiment further adds a process for identifying and adjusting the weights of BeiDou satellite observation anomalies based on visual relative motion, on the basis of the above embodiments, in order to reduce the impact of abnormal satellites on floating-point ambiguity and ambiguity covariance matrix before integer search.

[0132] The change in BeiDou carrier phase between adjacent epochs is mainly composed of the change in geometric distance between satellites and UAVs, receiver motion, and a small amount of residual error. Within a short time interval, the satellite's own position can be calculated by broadcast ephemeris or precise ephemeris, while the UAV's motion can be independently estimated by the scale-corrected visual relative displacement. Therefore, visual displacement can be used to predict the change in the double-difference geometric distance between each satellite caused by the UAV's motion, and compare it with the actual double-difference carrier phase change.

[0133] Let the line-of-sight unit vectors of satellite i and reference satellite r in the current epoch be u, respectively. i and u r Both refer to the satellite from the drone's position, and the drone's displacement Δp within a short time interval is considered. v,k Ignoring second-order minor quantities, the change in the double-difference geometric distance between satellite i and reference satellite r caused by the motion of the UAV can be determined by the inner product of the difference in the line-of-sight directions of the two satellites and the displacement of the UAV.

[0134] The formula for predicting the geometric transformation of the double difference is: ;

[0135] The formula for predicting double-difference phase changes is: .

[0136] In the above formula for predicting double-difference phase changes, Let cycle represent the predicted double-difference carrier phase change. Let λ be the carrier wavelength corresponding to the frequency point of satellite i, in meters. If the solution process uniformly adopts meters as the carrier phase unit, then the predicted double-difference geometric change is directly compared with the actual double-difference carrier phase distance change.

[0137] The actual double-difference carrier phase change is obtained by subtracting the double-difference carrier phase observation value of the current epoch from that of the previous corresponding epoch. To avoid the influence of the integer ambiguity constant term, in consecutive epochs where no cycle slip occurs and the reference satellite remains unchanged, the integer ambiguity term is canceled out after subtracting adjacent epochs. When the reference satellite is switched, the reference satellite is first converted using a common satellite so that the double-difference observations of the previous and subsequent epochs maintain the same reference standard.

[0138] The formula for satellite observation bias is: .

[0139] The standard deviation σ is calculated based on the carrier phase noise of each satellite, the satellite elevation angle, and the visual displacement uncertainty. i The observation bias is then normalized to form the degree of observation anomaly.

[0140] The formula for the degree of observed anomaly is: .

[0141] In this embodiment, when the degree of observation anomaly is less than 2, the original weight is maintained; when the degree of observation anomaly is 2-4, the observation weight of the corresponding satellite is continuously reduced according to the degree of anomaly; when the degree of observation anomaly is greater than 4, or when the degree of observation anomaly of the same satellite is greater than 3 for two consecutive epochs, the satellite is removed in the current integer fixed process. The weight adjustment can adopt exponential decay so that the higher the degree of anomaly, the smaller the effective weight entering the floating-point solution.

[0142] The formula for adjusting the weight of anomaly observations is: .

[0143] In the above formula for adjusting the weights of abnormal observations, The original weights for satellite i are determined based on factors such as elevation angle and signal-to-noise ratio. To update the weights after incorporating visual motion consistency, satellites that meet the removal criteria are directly ordered... .

[0144] After adjusting the satellite observation weights, the double-difference carrier phase observation equations are reconstructed using updated BeiDou satellite observation data. The floating-point position solution, carrier phase floating-point ambiguity, and ambiguity covariance matrix are then redefined, and an integer search is performed again. Thus, the role of visual information in BeiDou solution calculation is divided into two levels: the first level identifies and reduces the weight of satellites whose motion is inconsistent with the overall system at the observation level; the second level filters candidates using the feasible region of visual motion at the integer candidate level. Neither level requires absolute position information from vision.

[0145] When the visual motion confidence level is below 0.30, visual information is not used to eliminate satellite observations. Only the receiver's own cycle slip detection, carrier-to-noise ratio, and residual verification are used. Once the visual motion confidence level recovers, the predicted double-difference phase change is enabled. For cases where the UAV is nearly stationary, if the scale-corrected visual displacement is less than 0.02 m, the predicted motion is considered to be near zero, and the observation anomaly threshold is appropriately relaxed to avoid misjudgment caused by amplification of minor visual jitter.

[0146] When a carrier phase cycle slip is detected, the double-difference ambiguity of the cycle slip is reinitialized to floating-point state, and the use of this satellite in the fixed candidate confirmation for the current epoch is suspended until it accumulates enough continuous observations again. If the reference satellite experiences a cycle slip or loses lock, a new reference satellite is selected, and an ambiguity reference transformation is performed on the common satellite that is still being continuously tracked.

[0147] During a specific flight, the UAV flew along the side of the building at a horizontal speed of about 6 m / s. The BeiDou observation frequency was 10 Hz, and the camera acquisition frequency was 40 Hz. When the UAV entered the edge of the building's obstruction, some low-altitude angle satellites were affected by reflection, and their actual double-difference carrier phase change was inconsistent with the visually predicted motion.

[0148] The system first reduces the weight of the corresponding satellite and recalculates the floating-point ambiguity; then it retains 10 integer candidates and eliminates candidates with significantly inconsistent displacement directions or amplitudes through the visual motion feasible domain; finally, it determines the target integer ambiguity from the remaining candidates based on the carrier phase consistency index. This process does not require continuous visual output of global absolute coordinates and is suitable for scenarios where the quality of BeiDou observation deteriorates in a short period of time but the camera can still observe the static environment.

[0149] In another scenario, when the camera image is interfered with by strong light, fog, or large-area dynamic targets, leading to a decrease in the number of effective interior points and an increase in reprojection residuals, the visual motion confidence level decreases accordingly, and the visual feasible region automatically widens. When the confidence level falls below the usable threshold, the system stops the visual hard screening and reverts to BeiDou's own floating-point and fixed checks. Thus, visual constraints are not mandatory single-point decisions, but rather auxiliary constraints controlled by confidence level, which can avoid reverse interference to BeiDou's calculations when visual conditions are poor.

[0150] In terms of BeiDou observation preprocessing, the rover and the base station first establish a common observation set according to satellite number, frequency, and observation epoch. For carrier phase observations that are lost, have abnormal half-cycle flags, have a carrier-to-noise ratio lower than the receiver's usable threshold, or fail continuity checks, they are not directly entered into the double-difference equation. The reference satellite is selected from the common observation set. In this embodiment, the satellite with the longest continuous tracking time and the largest elevation angle is preferred. When the elevation angle of the reference satellite decreases or a loss of lock occurs, the reference satellite conversion is first completed among the common satellites, and then the reference satellite is switched to reduce the number of ambiguity re-initializations caused by changes in the reference satellite.

[0151] For observation weights, in addition to elevation angle and signal-to-noise ratio, adaptive weights can also be formed by combining carrier phase historical residuals. Satellites with elevation angles below 15° are easily affected by obstruction and multipath effects, and in principle, they are not used for fixed integers in this embodiment. Satellites with elevation angles between 15° and 30° use lower weights. Satellites with elevation angles greater than 30° participate according to normal weights. The above angle range is used to reflect the physical characteristic that low elevation angle observation errors are usually large. The actual threshold can be adjusted according to the antenna installation position, the obstruction range of the UAV body, and the operating environment.

[0152] After the floating-point solution is completed, a positive definiteness and condition number check is performed on the ambiguity covariance matrix. When the matrix is ​​close to singular due to insufficient visible satellites, poor satellite geometry, or extreme imbalance of observation weights, it does not enter the integer search, but remains in the floating-point state and waits for subsequent epochs. When used for joint solution of three-dimensional position and double-difference ambiguity, the number of jointly observed satellites should at least meet the requirements for observability of unknown parameters. In actual operation, in order to improve redundancy, this embodiment does not perform visual-assisted integer fixing when there are fewer than 5 sets of available double-difference observations.

[0153] The number of integer candidates, K, is not necessarily better the larger it is. If K is too small, the true integer vector may not be retained when BeiDou observations are disturbed; if K is too large, it will increase the computational load for candidate position back-substitution and visual consistency calculations for each epoch. In this embodiment, K is set to 10, so that the search results can cover several candidates with similar statistical distances, while still allowing real-time filtering on the airborne platform. When the floating-point ambiguity covariance is small and the optimal candidate is significantly separated from other candidates, the actual number of retained candidates can be reduced; when the covariance increases, the number of candidates can be increased within the limits of computational resources.

[0154] Keyframe selection is used to control the visual motion baseline. When the displacement between adjacent images is too small, the feature disparity is insufficient, and the translation direction is easily affected by pixel noise; when the time interval is too large, the number of feature matches may decrease significantly. In this embodiment, images continue to accumulate when the average disparity of effective features between adjacent frames is less than 3px, until the average disparity reaches 3-30px or the accumulation time reaches 0.20s, at which point a keyframe is selected. This method ensures that the visual relative motion has sufficient geometric disparity while avoiding a large number of matching failures caused by excessively long time intervals.

[0155] In addition to using the geometric consistency of the essential matrix, dynamic feature removal also utilizes multi-frame trajectory consistency. For the same feature trajectory, if its motion direction is consistently inconsistent with the main motion model of the static scene in three consecutive frames, or if it forms a local cluster in the current frame and the residual of the local region is significantly higher than the global median residual, then it is removed from the static feature set. This process can reduce the impact of vehicles, pedestrians, rotating blade shadows, and locally swaying objects on the UAV's own motion estimation.

[0156] Alignment between the visual coordinate direction and the BeiDou local coordinate direction can be completed before flight through camera and airframe external parameter calibration and flight control attitude, or during flight using reliable fixed historical epoch estimation. When using historical estimation, several sets of BeiDou displacement vectors and visual displacement vectors with sufficient directional changes are collected. After normalization, the three-dimensional rotation matrix that minimizes the sum of squares of the two sets of directional errors is solved. Orthogonalization constraints are used to ensure that the rotation matrix satisfies the rigid body rotation condition. Scale estimation and directional alignment are performed separately, which can avoid erroneously absorbing directional errors into the scale factor.

[0157] When updating the visual scale window, a scale change rate limit is also set. For a newly calculated scale factor, if its change relative to the current scale correction parameter exceeds 20%, but the BeiDou fixed state and visual reliability do not change synchronously, the scale factor is marked as a value to be confirmed and is not immediately written into the scale window. The scale parameter is only allowed to be updated when two or more consecutive reliable epochs show similar changes. This mechanism is used to suppress scale abrupt changes caused by single-frame matching anomalies, short-term depth-of-field degradation, or local rotation dominating.

[0158] When the UAV turns or undergoes significant attitude changes, the rotation component in the visual relative pose is first used to compensate for changes in camera orientation, and then the translational motion is compared. The candidate position change is in the BeiDou local navigation coordinate system, and the visual translation also enters the same coordinate system after passing through the camera extrinsic parameters and attitude rotation. If the UAV mainly performs in-place yaw and the translation is less than 0.02m, then the current epoch only uses the visual result to assist in judging that it is approximately stationary, and does not use the directional angle to screen candidates, thereby avoiding numerical instability of the translation direction under extremely small displacement conditions.

[0159] The threshold of the feasible region of visual motion can also be superimposed with the displacement uncertainty caused by the time synchronization error. If the current speed of the UAV is estimated as v and the upper bound of the time alignment residual error is δt, then the upper bound of the displacement uncertainty caused by the time error is approximately v·δt. The actual displacement tolerance can be increased by adding this uncertainty on the basis of the credibility tolerance. When the UAV flies at high speed, the system automatically leaves a larger physical margin for the time synchronization error. When the flight speed is low, the candidate discrimination ability will not be weakened by increasing the displacement threshold too much.

[0160] After the feasible ambiguity candidate set is formed, in addition to the carrier phase weighted residual, the motion continuity of the fixed position is also checked. If a candidate has a small carrier phase residual, but its fixed position produces a speed or acceleration change that exceeds the dynamic capability of the UAV relative to the previous credible position, the acceptance priority of the candidate is reduced. Motion continuity is only used as an anomaly protection condition and does not replace visual motion constraints and carrier phase consistency checks. Its threshold can be set according to the maximum design speed and maximum design acceleration of the UAV.

[0161] When the feasible ambiguity candidate set is empty, this embodiment adopts a hierarchical backoff strategy. First, it checks the visual credibility and time alignment status. If the visual credibility is at a medium level, the displacement tolerance and orientation tolerance are increased to 1.5 times the original value and the system is re-filtered. If there are still no candidates, or the visual credibility is lower than the available threshold, the visual hard constraint is no longer used, the BeiDou floating-point solution is output, and the current ambiguity status is retained to the next epoch. This strategy makes the system tend to postpone fixing when there is a constraint conflict, rather than forcibly selecting an integer solution from inconsistent candidates.

[0162] When the feasible ambiguity candidate set contains multiple candidates and the carrier phase consistency index difference is small, the system can retain candidates across epochs. For candidate paths that exist in two or three consecutive epochs, the visual motion consistency index and carrier phase consistency index are accumulated. The target integer ambiguity is determined after sufficient differences are formed between the candidate paths. Cross-epoch processing utilizes the continuity of UAV motion to increase discrimination information, which is especially suitable for scenarios with poor short-term satellite geometry but stable visual texture.

[0163] The threshold settings in this application follow the principle of matching the observation noise level with the UAV motion level. The carrier phase residual threshold should be higher than the normal double-difference carrier phase noise but lower than the obvious gross error. The visual displacement threshold should cover the combined effects of feature positioning error, scale estimation error and time synchronization error. The orientation threshold should vary with the effective parallax and inlier ratio. In actual deployment, the normal residual distribution can be statistically analyzed using flight calibration data in an unobstructed open environment, and the threshold can be determined based on the mean and standard deviation. This ensures that the thresholds are derived from specific equipment and operating conditions, rather than being arbitrarily set without regard to observation accuracy.

[0164] Through the aforementioned parameter management and boundary processing, this application rapidly eliminates erroneous integer candidates using a tighter visual feasible region under normal observation conditions. When BeiDou quality deteriorates but visual reliability remains high, the visual scale is maintained and candidate constraints continue to be provided. Visual constraints are automatically weakened or disabled when visual quality deteriorates. When both types of observations are insufficient, the application reverts to a floating-point state. The data flow between each state is always established around the candidate position changes and visual relative motion information within the same time interval, thus maintaining the physical interpretability and engineering feasibility of the algorithm chain.

[0165] Example 4:

[0166] This embodiment provides a high-precision positioning system for unmanned aerial vehicles (UAVs) based on BeiDou differential and visual constraints, used to execute the methods described in any of the above embodiments. The system includes a calculation module, a generation module, a filtering module, and an output module, and may further include a management module and an alignment module. Each module can be implemented by a software functional unit in the UAV's onboard computing platform, or it can be implemented collaboratively by the onboard end and the ground end.

[0167] The solution module communicates with the BeiDou high-precision positioning receiver to receive BeiDou satellite observation data from the UAV rover and differential observation data provided by the base station or reference station network. The solution module includes a double-difference construction unit, a floating-point solution unit, an integer search unit, and a candidate position calculation unit. The double-difference construction unit selects a reference satellite based on the common observation satellites and forms a double-difference carrier phase observation; the floating-point solution unit solves for the floating-point position solution, carrier phase floating-point ambiguity, and ambiguity covariance matrix based on the observation weights; the integer search unit obtains multiple integer ambiguity candidates based on the floating-point ambiguity and ambiguity covariance matrix; and the candidate position calculation unit substitutes each integer candidate back and calculates the corresponding candidate position change.

[0168] The generation module connects to the camera's data interface to read continuous images, image exposure times, and camera calibration parameters. The generation module includes a feature processing unit, a visual pose estimation unit, a credibility calculation unit, and a visual constraint generation unit. The feature processing unit extracts and matches features from adjacent images or keyframe images, eliminating anomalous and dynamic features based on descriptor consistency, epipolar geometric consistency, and multi-frame motion consistency. The visual pose estimation unit determines the visual motion direction and scale-free visual displacement based on the remaining static matching features. The credibility calculation unit calculates visual motion credibility based on the number of effective matches, inlier ratio, reprojection residuals, and spatial distribution. The visual constraint generation unit corrects visual displacement and credibility based on scale to form a feasible visual motion region.

[0169] The filtering module receives multiple candidate position changes output by the solving module and visual motion constraints output by the generation module. For each candidate, the filtering module calculates displacement deviation, orientation deviation, and motion consistency index, and determines whether the candidate is within the feasible region of visual motion based on the allowable deviation corresponding to the visual motion credibility. Integer ambiguity candidates that meet the conditions are added to the feasible ambiguity candidate set; candidates that do not meet the conditions are not included in the subsequent fixed confirmation.

[0170] The output module includes a carrier phase consistency verification unit and a fixed position calculation unit. The carrier phase consistency verification unit substitutes each feasible ambiguity candidate into the double-difference carrier phase observation model, calculates the fixed residual and weighted consistency index, and determines the target integer ambiguity based on the difference between the best and second-best candidates, the root mean square of the residuals, and the distribution of single-satellite residuals. The fixed position calculation unit recalculates the UAV position based on the target integer ambiguity and outputs the 3D position, positioning status, positioning quality index, and corresponding timestamp.

[0171] One function of the management module is to manage the visual scale. The management module reads the historical BeiDou fixed-state results generated by the output module and the historical scale-free visual displacements generated by the generation module, and determines whether the BeiDou differential carrier phase positioning is in a credible fixed state. It only calculates the visual scale factor and updates the scale correction parameters for historical epochs in the credible fixed state. When BeiDou positioning does not meet the credible fixed state requirement, the management module freezes the scale correction parameters and provides the unfrozen scale correction parameters to the generation module for current visual displacement scale correction.

[0172] Another function of the management module is to manage the observation quality of BeiDou satellites. The management module calculates the predicted double-difference carrier phase change of each satellite relative to a reference satellite based on the scale-corrected visual relative motion, the satellite line-of-sight unit vector, and the carrier wavelength. This predicted change is then compared with the actual double-difference carrier phase change to determine the degree of observational anomalies for each satellite. Based on the degree of anomaly, the management module sends weight adjustment information or satellite removal information to the solution module. The solution module then reconstructs the weight matrix and redetermines integer ambiguity candidates. These two management functions can be implemented by the same management module, or separate scale management and satellite observation management submodules can be set up.

[0173] The alignment module unifies the time reference between the camera and the BeiDou receiver. It reads the image exposure time, camera clock synchronization parameters, and BeiDou observation epochs, converts the image exposure time to the BeiDou time reference, and calculates the BeiDou motion information corresponding to the exposure time based on the position or motion status of adjacent BeiDou satellites. The alignment module provides unified start and end times for both the visual relative motion range generated by the generation module and the candidate position change range generated by the solution module, ensuring the comparability of the two types of motion quantities.

[0174] During system operation, the solution module and the generation module can process BeiDou observations and images within the same time window in parallel. After completing the floating-point solution and multiple integer candidates, the solution module writes the candidate position changes into the candidate cache. After completing the visual motion estimation, the generation module writes the visual motion direction, scale-corrected visual displacement, and visual motion confidence into the visual cache. The alignment module establishes a correspondence between the two based on the timestamp. The filtering module reads the corresponding data and performs candidate filtering. The output module performs carrier phase consistency checks and generates the final positioning result. The management module updates the scale parameters and satellite observation quality status according to the final positioning status and feeds the updated results back to the next epoch.

[0175] This system can be deployed on an airborne computing platform equipped with CPU, GPU, DSP, FPGA, or a combination thereof. BeiDou observation and differential data can be input via serial interface, Ethernet, wireless data transmission link, or network protocol; images can be input via MIPI, USB, Ethernet, or other camera interfaces. Modules can exchange timestamps, observation vectors, candidate ambiguities, candidate position changes, and visual constraint parameters via shared memory, message queues, or inter-process communication.

[0176] In one specific configuration, the BeiDou receiver outputs dual-frequency carrier phase observations at 10Hz, while the camera outputs grayscale or color images with a resolution of 1280×720px at 40Hz. The onboard computing platform performs feature processing on the images at a scale of 640×360px or equivalent to reduce real-time computation. Each BeiDou epoch retains 10 integer ambiguity candidates, and the visual scale window contains the 20 most recent reliable fixed epochs. Visual hard filtering is disabled when the visual motion confidence level falls below 0.30. This configuration enables epoch-level coordination between BeiDou differential and visual constraints while ensuring real-time performance.

[0177] Through the above system architecture, BeiDou differential observation provides absolute coordinate reference and precise carrier phase ranging information, while visual observation provides independent short-term motion constraints. Trusted fixed-state management ensures that high-quality BeiDou results are used to calibrate the visual scale, while low-quality BeiDou results do not contaminate the visual scale. Time alignment ensures that both types of motion quantities correspond to the same physical time interval. The visual feasible region excludes candidates inconsistent with the actual motion at the integer candidate layer. Satellite observation anomaly management further reduces the impact of anomalous satellites at the observation layer, and finally, the integer ambiguity of the target is confirmed by carrier phase observation consistency verification. This enables high-precision, continuous positioning of UAVs in complex BeiDou observation environments with anomaly suppression capabilities.

[0178] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A high-precision positioning method for unmanned aerial vehicles (UAVs) based on BeiDou differential and visual constraints, characterized in that, Includes the following steps: S1: Obtain the BeiDou satellite observation data corresponding to the UAV and the differential observation data corresponding to the BeiDou satellite observation data. Construct the double-difference carrier phase observation equation based on the BeiDou satellite observation data and the differential observation data to determine the UAV's position floating-point solution, carrier phase floating-point ambiguity and the corresponding ambiguity covariance matrix. S2: Perform an integer search based on the carrier phase floating-point ambiguity and ambiguity covariance matrix to obtain multiple integer ambiguity candidates, and determine the candidate position change corresponding to the current epoch based on each integer ambiguity candidate; S3: Acquire continuous images that correspond to the BeiDou satellite observation data in time, determine the visual relative motion information of the UAV based on the inter-frame changes of static scene features in the continuous images, and generate the visual motion constraints corresponding to the current epoch based on the visual relative motion information. S4: Determine the motion consistency between the change in each candidate position and the visual motion constraint, and determine the feasible ambiguity candidate set that satisfies the visual motion constraint from multiple integer ambiguity candidates based on the motion consistency. S5: Perform carrier phase observation consistency check on the feasible ambiguity candidate set, determine the target integer ambiguity based on the check results, and perform BeiDou differential carrier phase positioning based on the target integer ambiguity to obtain high-precision positioning results for the UAV.

2. The high-precision positioning method for UAVs based on BeiDou differential and visual constraints according to claim 1, characterized in that: In step S3, the visual relative motion information of the UAV is determined based on the inter-frame changes of static scene features in consecutive images, specifically as follows: Feature extraction and feature matching are performed on adjacent images or keyframe images with a preset time interval. Abnormal and dynamic features are eliminated based on the geometric consistency of the matched features. The visual motion direction, scale-free visual displacement, and visual motion credibility of the UAV are determined based on the remaining matched features.

3. The high-precision positioning method for UAVs based on BeiDou differential and visual constraints according to claim 2, characterized in that: Determining the visual relative motion information of the UAV based on the inter-frame changes of static scene features in continuous images also includes: The historical epochs in which the BeiDou differential carrier phase positioning is continuously in a reliable fixed state are obtained. The visual scale factor is determined based on the proportional relationship between the BeiDou position change corresponding to the historical epoch and the scale-free visual displacement. The scale correction parameter is determined based on the visual scale factor corresponding to multiple historical epochs. When the BeiDou differential carrier phase positioning does not meet the trusted fixed state, the update of the scale correction parameters is stopped, and the scale correction parameters determined before the update is stopped are used to perform scale correction on the current scaleless visual displacement.

4. The high-precision positioning method for UAVs based on BeiDou differential and visual constraints according to claim 3, characterized in that: The credible fixed state is determined based on at least two of the following: the number of consecutive fixed epochs, the integer ambiguity test result, the double-difference carrier phase residual, the satellite geometry, and the motion continuity of adjacent epochs. The scale correction parameters are updated using historical epochs in the trusted fixed state so that integer ambiguity candidates are constrained by the determined visual scale when the BeiDou positioning quality deteriorates.

5. The high-precision positioning method for UAVs based on BeiDou differential and visual constraints according to claim 1, characterized in that: In step S3, acquiring continuous images that correspond in time to the BeiDou satellite observation data includes: The exposure time of each image and the observation epoch of the BeiDou satellite observation data are obtained. The exposure time is converted to the BeiDou time reference. The BeiDou motion information corresponding to the exposure time is determined according to the position or motion state corresponding to the adjacent BeiDou satellite observation data, so that the candidate position change and the visual relative motion information correspond to the same time interval.

6. The high-precision positioning method for UAVs based on BeiDou differential and visual constraints according to claim 2, characterized in that: The step of generating visual motion constraints corresponding to the current epoch based on visual relative motion information includes: The visual motion confidence level is determined based on at least two of the following: the number of matching features, the proportion of geometrically consistent inliers, the reprojection residual, and the spatial distribution of matching features. The allowable displacement deviation and orientation deviation are determined based on the visual motion confidence level. The feasible region of visual motion is constructed based on the scale-corrected visual displacement, visual motion direction, displacement deviation, and orientation deviation. Integer ambiguity candidates whose corresponding positional change is within the feasible region of visual motion are retained in the feasible ambiguity candidate set.

7. The high-precision positioning method for UAVs based on BeiDou differential and visual constraints according to claim 1, characterized in that: High-precision positioning methods also include: Based on visual relative motion information, the predicted double-difference carrier phase change of each Beidou satellite between adjacent epochs is determined. The predicted double-difference carrier phase change is compared with the corresponding actual double-difference carrier phase change to determine the degree of observation anomaly of each Beidou satellite. Adjust the solution weights of the corresponding BeiDou satellite observation data according to the degree of observation anomalies, or remove BeiDou satellite observation data that meet the anomaly conditions, and redetermine the integer ambiguity candidates based on the adjusted BeiDou satellite observation data.

8. A high-precision positioning system for unmanned aerial vehicles based on BeiDou differential and visual constraints, wherein the method described in any one of claims 1-7 is characterized in that, include: The solution module is used to acquire the BeiDou satellite observation data and differential observation data corresponding to the UAV, construct the double-difference carrier phase observation equation, determine the floating-point solution of the position, the floating-point ambiguity of the carrier phase and the ambiguity covariance matrix, and obtain multiple integer ambiguity candidates and the candidate position change corresponding to each integer ambiguity candidate based on the floating-point ambiguity of the carrier phase and the ambiguity covariance matrix. The generation module is used to acquire continuous images that correspond to the BeiDou satellite observation data in time, determine visual relative motion information based on the inter-frame changes of static scene features in the continuous images, and generate visual motion constraints based on the visual relative motion information. The filtering module is used to determine the motion consistency between the change in each candidate position and the visual motion constraint, and to determine the feasible ambiguity candidate set that satisfies the visual motion constraint based on the motion consistency. The output module is used to perform carrier phase observation consistency checks on the feasible ambiguity candidate set, determine the target integer ambiguity, perform BeiDou differential carrier phase positioning based on the target integer ambiguity, and output the high-precision positioning results of the UAV.

9. The high-precision positioning system for unmanned aerial vehicles based on BeiDou differential and visual constraints according to claim 8, characterized in that: It also includes a management module and an alignment module; The management module is used to determine the scale correction parameters based on the BeiDou position change and scale-free visual displacement corresponding to the historical epochs when the BeiDou differential carrier phase positioning is in a credible fixed state, and to stop updating the scale correction parameters when the BeiDou differential carrier phase positioning does not meet the credible fixed state. The alignment module is used to convert the image exposure time to the BeiDou time reference and determine the BeiDou motion information corresponding to the image exposure time, so that the visual relative motion information and the candidate position change amount correspond to the same time interval.

10. The high-precision positioning system for unmanned aerial vehicles based on BeiDou differential and visual constraints according to claim 8, characterized in that: It also includes a management module, which is used to determine the predicted double-difference carrier phase change amount corresponding to each Beidou satellite based on the visual relative motion information, determine the observation anomaly degree of the corresponding Beidou satellite based on the deviation between the predicted double-difference carrier phase change amount and the actual double-difference carrier phase change amount, and adjust the solution weight of the corresponding Beidou satellite observation data or remove Beidou satellite observation data that meet the anomaly conditions based on the observation anomaly degree.

Images