Unmanned aerial vehicle autonomous positioning model construction method in GPS denial environment
By constructing a ternary relationship map dataset and differential technology with local topological consistency, the problem of degradation of drone positioning accuracy in GPS denial environment is solved, and high-precision autonomous positioning and accurate solution of velocity acceleration is achieved.
Patent Information
- Application Number
- CN202510624694.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-15
- Publication Date
- 2025-08-15
AI Technical Summary
In the GPS denial environment, the combined navigation method of inertial navigation technology combined with GPS cannot provide accurate positioning data, and the positioning accuracy of the drone in complex environments has decreased. The existing positioning method based on image matching has failed to effectively eliminate outliers, resulting in jitter in positioning results.
A ternary relationship map data set based on local topological consistency is constructed, and the triplets are screened through local affine consistency and Delaunay triangular analysis, combined with the compensation term scoring strategy and data subset refinement strategy, outliers are eliminated, real-time positioning of the drone in the GPS denial environment, and the linear velocity and line acceleration are solved through differential technology.
It improves the positioning accuracy and robustness of the drone in the GPS denial environment, ensures the autonomous positioning capability of the drone in complex environments, reduces the jitter of the positioning results, and improves the solution accuracy of linear speed and line acceleration.
Smart Images
Figure CN120491666A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of autonomous positioning of unmanned aerial vehicles (UAVs), and in particular relates to a method for constructing an autonomous positioning model of a UAV in a GPS-denied environment. Background Art
[0002] In recent years, unmanned aerial vehicles (UAVs) have played a significant role in a wide range of fields, including search and rescue, inspection, and transportation, due to their portability, flexibility, low cost, and adaptability to complex environments. Fundamental to the diverse missions that UAVs perform in these fields is accurate navigation and positioning. Only accurate positioning information enables UAVs to efficiently and accurately execute their missions. Currently, the main methods for UAV navigation and positioning include inertial navigation systems (INS), dead reckoning, and global satellite navigation. INS systems offer high autonomy, are independent of external information, and do not radiate any information. They offer high short-term accuracy, but their positioning errors accumulate rapidly over time. Furthermore, high-precision INS systems are not only heavy and bulky, but also expensive. Dead reckoning uses sensor data such as airspeed, altitude, attitude, and wind speed to estimate the UAV's position in real time. However, its errors are affected by the accuracy of onboard sensors and wind speed measurements, and they accumulate over time. Global satellite navigation systems offer high positioning accuracy, with errors that do not accumulate over time. However, they suffer from poor autonomy, are susceptible to interference, and have low information update rates, making them difficult to meet the requirements of high-dynamics and real-time control in certain scenarios. Furthermore, navigation satellites are located in outer space, far from Earth, resulting in weak signals reaching the ground and making them highly susceptible to electromagnetic interference. Furthermore, both military and civilian applications are inherently susceptible to spoofing, and navigation satellites can be attacked by foreign anti-satellite weapons, presenting security risks and potential vulnerabilities to satellite navigation technology.
[0003] Integrated navigation combines two or more navigation systems in an appropriate manner, leveraging their complementary performance to offset the shortcomings of a single navigation system, which relies too heavily on certain information and is unable to correct its errors. Currently, the most widely used and mature UAV navigation method is the combination of inertial navigation system (INS) and satellite navigation system (GPS). When GPS signals are available, GPS provides high-precision positioning and can also correct INS positioning. When GPS signals are unavailable or weak, the corrected INS system can provide short-term high-precision positioning and navigation. Combining the advantages of both systems not only achieves more accurate positioning and navigation, but also improves its applicability and robustness. However, in GPS-denied environments, the positioning accuracy of the INS system used in integrated navigation will gradually decrease over time when GPS signals are unavailable for extended periods. Furthermore, due to the lack of control over GPS, it should be used with caution in the unique context of warfare.
[0004] Scene matching originated in the terminal control of cruise missiles and later evolved into a visual navigation technology. As an auxiliary navigation method, scene matching for UAV visual navigation, combined with an inertial navigation system, can form a highly autonomous, high-precision navigation system. This gives UAVs "vision" capabilities, enabling precise positioning and navigation in GPS-denied environments, and precise strikes through image matching for search, recognition, and targeting. The rapid development of computer vision technology has led to the widespread use and research of scene matching in the UAV positioning phase, with a focus on image feature matching. Specifically, scene matching utilizes a UAV's onboard high-resolution imaging radar or electro-optical image sensor to acquire real-time images of the ground scene. Feature extraction algorithms (such as the scale-invariant feature transform (SIFT)) are used to compare the real-time image with a pre-stored digital reference image of the ground scene in the onboard computer to extract feature points. The feature matching algorithm then eliminates outliers to accurately calculate the UAV's position and correct for heading errors. However, in practical applications, such as field search and rescue and battlefield reconnaissance, the real-time image captured by the UAV is inevitably affected by numerous outliers in complex environments such as mountainous and forested areas. These outliers will provide erroneous coordinate information, causing a significant drop in the positioning accuracy of the drone. And because the positioning accuracy of the drone is reduced, the accuracy of its own real-time position solution is reduced, resulting in errors in solving the drone's linear velocity and linear acceleration. Therefore, it is crucial to design a robust model construction method to eliminate outliers to improve the autonomous positioning accuracy of the drone and accurately solve the linear velocity and linear acceleration. However, the existing positioning method based on image matching does not take into account the time-varying characteristics of the neighborhood geometry caused by the dynamic motion of the drone in a GPS-denied environment, and the outlier elimination strategy does not integrate the motion continuity constraint, resulting in jitter in the positioning results. The present invention solves the above problems in a targeted manner through dynamic neighborhood geometry constraints and motion continuity verification. Summary of the Invention
[0005] The purpose of the present invention is to overcome the problem that the combined navigation method of inertial navigation technology combined with GPS in a GPS-denied environment cannot provide accurate positioning data assistance, and drones and other aircraft are also required to have real-time calculation and anti-interference capabilities in a GPS-denied environment to cope with the problems of complex environments and sensor failures. A method for constructing an autonomous positioning model of a drone in a GPS-denied environment is provided, which can improve the accuracy of autonomous positioning of drones in a GPS-denied environment.
[0006] To achieve the above objectives, the technical solution of the present invention is: a method for constructing an autonomous positioning model for a UAV in a GPS-denied environment, comprising:
[0007] In a GPS-denied environment, a ternary relationship map dataset based on local topological consistency is constructed, where the generation of triples must satisfy neighborhood constraints.
[0008] The triplets are screened by local affine consistency verification and Delaunay triangulation, where the rank of the affine matrix verifies the validity of the triplet construction;
[0009] A compensation scoring strategy is proposed to adjust triple weights based on local geometric consistency scores.
[0010] A data subset refinement strategy is adopted to screen out inliers through iterative residual sorting stopping criteria;
[0011] Based on the correct matching points, the real-time positioning of the UAV in GPS-denied environments is achieved.
[0012] Furthermore, the method further includes:
[0013] Based on the real-time position calculation results of the UAV in a GPS-denied environment, the linear velocity and linear acceleration of the UAV are calculated using differential technology and parameterized path scalars.
[0014] Furthermore, a map dataset consisting of map data point pairs of the real-time map and the benchmark map collected by the drone is constructed as follows: the real-time map and the benchmark map collected by the drone are input, and feature extraction is performed on the real-time map and the benchmark map using a feature extraction algorithm to obtain map data point pairs corresponding to the real-time map and the benchmark map, and to construct a map dataset.
[0015] Furthermore, in a GPS-denied environment, a ternary relationship map dataset based on local topological consistency is constructed, where the generation of triples must satisfy neighborhood constraints. The specific implementation is as follows:
[0016] In a map dataset consisting of N map data point pairs For the i-th pair of map data points Z i , search Z based on spatial relationship i The K nearest neighbors of That is, K-nearest neighbors, only when Z is satisfied i 、Z j and Z k In the case of the following relationship between i 、Z j and Z k Three points form a triple Expressed as
[0017]
[0018] in, Represents Zi / k / j Belong to Z j / i / k K-nearest neighbors, Indicates when If true, the value is 1, otherwise it is 0;
[0019] After all the map data point pairs in the map dataset B0 are expressed through ternary relations, the initial data subset is obtained.
[0020] Furthermore, the triplets are screened through local affine consistency verification and Delaunay triangulation, where the rank of the affine matrix verifies the validity of the triplet construction. The specific implementation is as follows:
[0021] Extract triples based on the initial data subset obtained by expressing the ternary relationship of the map dataset (x i ,y i )(x j ,y j )、(x k ,y k ) represents a triple The corresponding coordinates, Transpose and get X i =[x i ,y i ] T 、X j =[x j ,y j ] T 、X k =[x k ,y k ] T ;
[0022] X j As the reference point, construct X i 、X k Relative to reference point X j The displacement vector is constructed into an affine matrix of three elements Calculate the rank r(A) of the affine matrix A. If r(A) < 2, it means that two or more of the three points of the triplet overlap, the triplet is invalid, and the score is recorded as 0; if r(A) = 2, the triplet is valid; after the initial data subset is effectively screened for triples, a data subset containing M data is obtained.
[0023] Use the inverse matrix A of the affine matrix A -1 Acting on vector X k -X i , get the transformed point X′ k , point X′ k Map back to X through the inverse affine matrix jThe reference frame of X′ k At the transformed position X′ k =A -1 ·(X k -X i ), calculate the transformed point X′ k The Euclidean distance is the transformation error E; the exponential function is used to convert the transformation error E into a geometric consistency score, that is, a local affine consistency score The lengths of the three sides of the triangle are calculated using the coordinates of the triplet points. The internal angles of the triangle are then calculated using the vector cosine theorem. The angles of the triplet triangle are then compared with the angles of the regular triangle. The exponential function is used to convert the comparison results into an angle difference score, i.e., the Delaunay triangle triangulation score. Among them, ∠1, ∠2 and ∠3 are the three internal angles of the triple triangle; finally, the local affine consistency score and the Delaunay triangle triangulation score are added to obtain the final triple score S=S1+S2.
[0024] Furthermore, a compensation scoring strategy is proposed to adjust the triple weights in combination with the local geometric consistency score. The specific implementation is as follows:
[0025] By defining the compensation term The scores are adjusted by weighting the local affine consistency score and the Delaunay triangle triangulation score to minimize unreliable matches; where θ is the score threshold and β is the compensation factor; the final triple compatibility score S0 = S + C;
[0026] According to the final triple compatibility score S0, the data subset containing M data Calculate the i-th data z in the data subset B i The sampling weight Among them, z j 、z k are the jth and kth data in the data subset B, max{S0(z i , z j , z k )} represents the z-based i Associated triples The triple compatibility score S0(z i , z j , z k ) is the maximum value, if z i is not associated with any triple, then ω(z i ) is set to 0; after processing the data subset B with the triple compatibility score S0, the data subset B′ is obtained.
[0027] Furthermore, a data subset refinement strategy is adopted to screen the interior points through the iterative residual sorting stopping criterion. The specific implementation is as follows:
[0028] Based on the data subset B′, generate a model hypothesis θ l , first based on the Simpson distance calculation model, assuming θ l With map dataset The residual value between Then, the residual values are sorted in ascending order, and the data in the map dataset B0 are sorted according to the sorted residual values, and the data before and after the mth data are extracted from the sorted data to form a new data subset, and a new model hypothesis is generated; the residual values of the new model hypothesis and the map dataset B0 are calculated, and the residual values are sorted in ascending order. According to the sorted residual values, the data in the map dataset B0 are sorted, and the data before and after the mth data are extracted from the sorted data to form the next new data subset, and the next new model hypothesis is generated. This process is performed iteratively until a stable model hypothesis is obtained; the convergence of the data subset refinement process is determined based on the last three iterations, and a parameter, MaxIter, is used to represent the maximum number of iterations. At the same time, the sampling weight derived from the ternary relationship of the data subset B′ is used as the condition for the stopping criterion, and the stopping criterion is:
[0029]
[0030] Where, The jth data point is the data subset consisting of the data before and after the mth data extracted from the data sorted by the residual value in the l-1th iteration. is the mth data obtained by sorting the residual values in the lth iteration; ∧ represents the relationship between and ; ω(·) is the sampling weight; h is the size of the data subset consisting of the data before and after the mth data extracted from the data sorted by the residual values;
[0031] Finally, based on the stable model assumption, outliers and inliers are distinguished and correctly matched map data point pairs are obtained.
[0032] Furthermore, based on the real-time position calculation results of the UAV in a GPS-denied environment, the linear velocity and linear acceleration of the UAV are calculated using differential technology and parameterized path scalars. The specific implementation is as follows:
[0033] Assume that the real-time position spatial coordinate information L(p) obtained by calculating the real-time position of the drone in a GPS-denied environment is:
[0034] L(p)=(x p ,y p , z p );
[0035] The linear velocity V of the drone in the x, y, and z directions is calculated by using L(p) through a single difference technique. x (p), V y (p) and V z (p):
[0036]
[0037] Among them, x p+1 、y p+1 and z p+1 Indicates the spatial coordinate information of the drone at the next moment, x p 、y p and z p represents the spatial coordinate information of the drone at the current moment, and Δt represents the time between two moments;
[0038] Then, the linear acceleration a of the UAV in the x, y, and z directions is calculated by the quadratic difference technique. x (p), a y (p) and a z (p):
[0039]
[0040] Among them, V x (p+1), V y (p+1) and V z (p+1) represents the linear velocity of the drone at the next moment, V x (p), V y (p) and V z (p) represents the linear velocity of the UAV at the current moment;
[0041] Then, the linear velocity and acceleration obtained by L(p) solution at each moment of the UAV in the GPS-denied environment are scalarized and converted into the linear velocity and acceleration on the path.
[0042] Furthermore, the linear velocity and acceleration obtained by L(p) solution at each moment of the UAV in the GPS-denied environment are scalarized and converted into the linear velocity and acceleration on the path. The specific implementation is as follows:
[0043] Assume that the parameterized path of the UAV in the GPS-denied environment is p(s):
[0044] p(s)=[x(s),y(s),z(s)] T ;
[0045] Then, the derivative with respect to time t is obtained to obtain the component velocity and
[0046] The total linear velocity V is:
[0047]
[0048] So the speed of the parameterized path scalar for:
[0049]
[0050] in, Then, the first-order derivative of the linear velocity V with respect to time is obtained:
[0051] The acceleration of the parameterized path scalar is thus for
[0052]
[0053] Compared with the prior art, the present invention has the following beneficial effects:
[0054] (1) Based on the criterion of ternary relations of local topological consistency, the present invention proposes to combine local affine consistency and Delaunay triangulation with triple relations to construct a stable ternary relation, so as to improve the robustness of the feature matching algorithm.
[0055] (2) This paper proposes a compensation term strategy, combined with a map data subset search strategy, to further constrain triple relationships. This strategy minimizes unreliable matches, enabling the algorithm to more effectively focus on feature points with high geometric consistency and stability, ultimately improving the matching success rate. Subsequently, differential techniques and parameterized path scalars are used to further resolve the UAV's linear velocity and acceleration. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 The figure is a flow chart of UAV positioning based on the method of the present invention.
[0057] Figure 2 It is a ternary relationship process.
[0058] Figure 3 This is the UAV positioning process based on image registration; Figure 3 (a) is the input real-time map and reference map; Figure 3 (b) is the feature matching pair between map images obtained through feature extraction; Figure 3 (c) is the feature matching pair after the outliers are removed by the method of the present invention; Figure 3 (d) is the UAV positioning map after image registration.
[0059] Figure 4 Partial positioning results in the map dataset. DETAILED DESCRIPTION
[0060] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
[0061] The present invention provides a method for constructing an autonomous positioning model for a UAV in a GPS-denied environment, comprising:
[0062] In a GPS-denied environment, a ternary relationship map dataset based on local topological consistency is constructed, where the generation of triples must satisfy neighborhood constraints.
[0063] The triplets are screened by local affine consistency verification and Delaunay triangulation, where the rank of the affine matrix verifies the validity of the triplet construction;
[0064] A compensation scoring strategy is proposed to adjust triple weights based on local geometric consistency scores.
[0065] A data subset refinement strategy is adopted to screen out inliers through iterative residual sorting stopping criteria;
[0066] Based on the correct matching points, the real-time positioning of the UAV in GPS-denied environments is achieved.
[0067] Based on the real-time position calculation results of the UAV in a GPS-denied environment, the linear velocity and linear acceleration of the UAV are calculated using differential technology and parameterized path scalars.
[0068] The following is a specific implementation process of the present invention.
[0069] like Figure 1 As shown in the figure, the present invention proposes a method for constructing an autonomous positioning model of a UAV in a GPS-denied environment, which is applied to a UAV autonomous positioning platform. By improving the feature matching algorithm to enhance the outlier filtering capability and image registration accuracy, the accurate positioning of the UAV in a GPS-denied environment is achieved.
[0070] 1. Ternary Relationship Principle
[0071] This paper uses ternary relationships to guide the sampling process of valid map data subsets for robust estimation. Specifically, the triple relationships between input feature matches (i.e., input data) are analyzed. Each triple consists of three data, and the relationship between them satisfies the spatial consistency constraint, thereby obtaining a series of triple sets. The detailed explanation of the triple relationship is as follows:
[0072] like Figure 2 As shown, in the map data set consisting of N map data point pairs In the example, for the i-th pair of map data points Z i, search Z based on spatial relationship i The K nearest neighbors of That is, K-nearest neighbors, only when Z is satisfied i 、Z j and Z k In the case of the following relationship between i 、Z j and Z k Three points form a triple Expressed as
[0073]
[0074] in, Represents Z i / k / j Belong to Z j / i / k K-nearest neighbors, Indicates when If true, the value is 1, otherwise it is 0;
[0075] Thus, after all pairs of map data points in the map dataset B0 are expressed through ternary relations, an initial data subset is obtained.
[0076] 2. Local Affine Consistency
[0077] When a drone is performing positioning, the target images it collects are often complex and do not have a simple global correspondence in terms of distance and rotation. However, even under complex non-rigid transformations, the topological structure of small areas in the map image pair often remains unchanged to a certain extent. Therefore, the present invention introduces local affine transformations into the triple relationship to approximate the image correspondence. Specifically, the coordinates of the triple points are first extracted. These points are located on the 2D plane, and the coordinates are:
[0078]
[0079] Then, Transpose and get X i =[x i ,y i ] T 、X j =[x j ,y j ]、X k =[x k ,y k ] T , from the reference point (e.g. select X j Starting from the reference point, construct the displacement vectors of the other two points relative to the reference point, and construct these two vectors into a triplet affine matrix A:
[0080]
[0081] Calculate the rank of the affine matrix A. If r(A) < 2, it means that there are more than two points in the three points that overlap, which is invalid and the score is recorded as 0. If r(A) = 2, the triple is valid. Thus, after the initial data subset is effectively screened for triples, a data subset containing M data is obtained.
[0082] Use the inverse matrix A -1 Acting on vector X k -X i , get the transformed point X′ k , so that point X′ k Map back to X through the inverse of the affine matrix j , check the position of the point after the transformation:
[0083] X′ k =A -1 ·(X k -X i )
[0084] Calculate the transformed point X′ k The Euclidean distance is the transformation error E. The error is then converted into a geometric consistency score S1 using an exponential function. The smaller the error, the higher the geometric consistency score; the larger the error, the faster the score approaches 0:
[0085]
[0086] 3. Delaunay triangulation
[0087] Delaunay triangulation provides an ideal neighboring structure for matching points during feature matching. The classic Delaunay triangulation forms a neighborhood structure of randomly distributed feature points in the image, and at the same time divides the image plane into several blocks of approximately regular triangles, ensuring the reliability of the calculation and the high precision of the transformation parameters between two corresponding triangular facets. Since each triple will form a corresponding triangle when constructing a triple, based on the idea of Delaunay triangulation, it is only necessary to calculate the angle difference between the triplets to retain a stable triple structure. Specifically, first extract the coordinates of the triple point, calculate the lengths of the three sides of the triangle through the coordinates of the triple point, and then use the vector cosine theorem to calculate the internal angle of the triangle:
[0088]
[0089] Among them, ∠1, ∠2, and ∠3 are the three internal angles of the triangle. Then, the angles of the triple triangle are compared with the angles of the equilateral triangle. The comparison result is converted into an angle difference score S2 using an exponential function. The closer the triple triangle is to an equilateral triangle, the higher the score; otherwise, the score will quickly approach 0:
[0090]
[0091] Finally, the local affine consistency score and the simplified Delaunay triangulation score are added to get the final triple score S:
[0092] S=S1+S2
[0093] 4. Compensation scoring strategy
[0094] After introducing local affine consistency and simplified Delaunay triangulation into the calculation of triplet scores, the present invention introduces a compensation term mechanism to adjust the score, minimize unreliable matches, and thus improve the robustness and accuracy of feature matching. By weighting local affine consistency and angular differences, it is possible to more effectively focus on feature points with high geometric consistency and stability, ultimately improving the success rate of matching. Specifically, the calculation of the compensation term C can be defined as:
[0095]
[0096] Where θ is the score threshold and β is the compensation factor (usually a positive number). The final triple compatibility score S0 can be expressed as:
[0097] S0=S+C
[0098] According to the final triple compatibility score S0, the data subset containing M data Calculate the i-th data z in the data subset B i The sampling weights are:
[0099]
[0100] Among them, z j 、z k are the jth and kth data in the data subset B, max{S0(z i , z j , z k )} represents the z-based i Associated triples The triple compatibility score S0(z i , z j , z k ) is the maximum value, if Z iis not associated with any triple, then ω(z i ) is set to 0; after processing the data subset B with the triple compatibility score S0, the data subset B′ is obtained.
[0101] 5. Map Data Subset Refinement Strategy
[0102] Given a model hypothesis θ l (Generated by data subset B′), first based on the Sampson distance calculation model assumption θ l With map dataset The residual value between Then, the residual values are sorted in ascending order, and the data in the map dataset B0 are sorted according to the sorted residual values, and the data before and after the mth data are extracted from the sorted data to form a new data subset to generate a new model hypothesis; the residual values of the new model hypothesis and the map dataset B0 are calculated, and the residual values are sorted in ascending order. According to the sorted residual values, the data in the map dataset B0 are sorted, and the data before and after the mth data are extracted from the sorted data to form the next new data subset to generate the next new model hypothesis. This process is performed iteratively until a stable solution (i.e., a stable model hypothesis) is obtained. Whether the data subset refinement process converges is determined based on the last three iterations, and a parameter (i.e., MaxIter) is used to represent the maximum number of iterations. At the same time, the weight value derived from the ternary relationship of the sampled data is used as the condition of the stopping criterion to reduce sensitivity to outliers. Specifically, the stopping criterion is expressed as:
[0103]
[0104] Where, The jth data point is the data subset consisting of the data before and after the mth data extracted from the data sorted by the residual value in the l-1th iteration. is the mth data obtained by sorting the residual values in the lth iteration; ∧ represents the relationship between and ; ω(·) is the sampling weight; h is the size of the data subset consisting of the data before and after the mth data extracted from the data sorted by the residual values;
[0105] The stopping criterion indicates that the iterations are stopped when the data subsets sampled from the last three iterations are more likely to belong to the same model. This further improves the performance of data subset sampling, resulting in a valid data subset and a significant model hypothesis for the estimated model. Finally, based on the estimated model, outliers and inliers can be distinguished, resulting in correctly matched pairs of map data points, allowing the drone to autonomously calculate its real-time position in GPS-denied environments. Furthermore, differential techniques and parameterized path scalars can be used to further calculate the drone's linear velocity and acceleration.
[0106] 6. Calculation of UAV linear velocity and linear acceleration in GPS-denied environments
[0107] Assume that the real-time position spatial coordinate information L(p) obtained by calculating the real-time position of the drone in a GPS-denied environment is:
[0108] L(p)=(x p ,y p , z p );
[0109] The linear velocity V of the drone in the x, y, and z directions is calculated by using L(p) through a single difference technique. x (p), V y (p) and V z (p):
[0110]
[0111] Among them, x p+1 、y p+1 and z p+1 Indicates the spatial coordinate information of the drone at the next moment, x p 、y p and z p represents the spatial coordinate information of the drone at the current moment, and Δt represents the time between two moments;
[0112] Then, the linear acceleration a of the UAV in the x, y, and z directions is calculated by the quadratic difference technique. x (p), a y (p) and a z (p):
[0113]
[0114]
[0115] Among them, V x (p+1), V y (p+1) and V z (p+1) represents the linear velocity of the drone at the next moment, V x (p), V y (p) and V z (p) represents the linear velocity of the UAV at the current moment;
[0116] Then, the linear velocity and acceleration obtained by L(p) solution at each moment of the UAV in the GPS-denied environment are scalarized and converted into the linear velocity and acceleration on the path.
[0117] Assume that the parameterized path of the UAV in the GPS-denied environment is p(s):
[0118] p(s)=[x(s),y(s),z(s)] T ;
[0119] Then, the derivative with respect to time t is obtained to obtain the component velocity and
[0120]
[0121] The total linear velocity V is:
[0122]
[0123] So the speed of the parameterized path scalar for:
[0124]
[0125] in,
[0126] Then, the first-order derivative of the linear velocity V with respect to time is obtained:
[0127]
[0128] in,
[0129] The acceleration of the parameterized path scalar is thus for
[0130]
[0131] Therefore, through differential technology and parameterized path scalars, the drone can solve the speed and acceleration in three directions of space based on real-time position coordinate information.
[0132] 7. Principle of UAV positioning based on image registration in GPS-denied environment
[0133] The drone positioning of the present invention is based on image registration. By extracting and matching feature points of the real-time map collected by the input drone and the reference map, the feature matching pairs between the two map images are screened out, and then positioning is performed based on the coordinate information provided by the feature matching pairs, thereby achieving drone positioning. The basic process of image registration is as follows: Figure 2 shown.
[0134] The basic process of UAV positioning based on image registration is as follows: Figure 3 As shown. Figure 3 (a) is the input real-time map and reference map; Figure 3 (b) is the feature matching pair between map images obtained through feature extraction; Figure 3 (c) in the middle is the feature matching pair after removing outliers by our method; Figure 3 (d) is the UAV positioning map after image registration.
[0135] 8. Simulation analysis
[0136] To demonstrate the effectiveness of our proposed method, we compared it with four state-of-the-art feature matching methods on four map datasets: RANSAC, RANSAC++, LPM, and TRESAC. We evaluated each algorithm based on average correct match rate, average runtime, and effective localization rate. The results are shown in Table 1.
[0137] Table 1 Positioning performance of different algorithms
[0138]
[0139] As can be seen in Table 1, the proposed method achieves superior localization performance, with an effective localization rate of 93.75% and an average runtime of 0.1190 seconds. This represents a 79.34% improvement in computational efficiency over RANSAC, and an average correct match rate of 98.73%. Due to the need for iterative data optimization to obtain the best-fit model, both RANSAC and RANSAC++ exhibit higher computational costs. Compared to TRESAC++ and LATSAC, LPM demonstrates superior computational efficiency. However, despite LPM's average runtime of 0.0308 seconds, its average correct match rate is the lowest compared to other advanced algorithms. This results in residual outliers providing erroneous coordinate information, reducing localization accuracy. Compared to TRESAC, LATSAC effectively retains inliers by using a compensation term strategy and the proposed improved initial subset selection strategy. Consequently, it achieves a higher correct match rate, and more accurate inliers provide more precise map image coordinate information.
[0140] In contrast, the method of the present invention utilizes the triplet relationship guided by local affine consistency to achieve high accuracy while maximally retaining the correct matches (inliers), thereby improving the positioning accuracy and real-time performance of the UAV. Figure 4 This is a partial positioning result in the map data set using the method of the present invention.
[0141] At the same time, the autonomous positioning of the UAV in the GPS-denied environment of the present invention can be verified by semi-physical simulation based on the four-rotor test bench system - Links-UAV-MiniBench developed by Lingsi Chuangqi. Specifically, the pan-tilt camera configured on the UAV is responsible for capturing the actual environment screen and sending these real-time map images to the main control computer through the image transmission device. Subsequently, the received map image is compared and analyzed with the reference map image pre-stored in the UAV system using Matlab software, and then the linear velocity and linear acceleration of the UAV are further solved by differential technology and parameterized path scalars. After the test is completed, all image registration results will be exported in the Matlab environment and saved in the folder specified by the main control computer. It is convenient to analyze, evaluate and use the proposed UAV positioning results based on image registration in the GPS-denied environment.
[0142] The above is a preferred embodiment of the present invention. Any modifications made according to the technical solution of the present invention that produce functional effects within the scope of the technical solution of the present invention shall fall within the scope of protection of the present invention. Furthermore, this embodiment can also be used on aircraft, that is, by combining the patented autonomous positioning model based on image registration with the aircraft's flight control system to achieve autonomous real-time positioning of the aircraft and real-time calculation of linear velocity and linear acceleration.
Claims
1. A method for constructing an autonomous positioning model for a UAV in a GPS-denied environment, characterized in that: include: In a GPS-denied environment, a ternary relationship map dataset based on local topological consistency is constructed, where the generation of triples must satisfy neighborhood constraints. Triples are screened by local affine consistency verification and Delaunay triangulation, where the rank of the affine matrix verifies the validity of triple construction; A compensation scoring strategy is proposed to adjust triple weights based on local geometric consistency scores. A data subset refinement strategy is adopted to screen out inliers through iterative residual sorting stopping criteria; Based on the correct matching points, the real-time positioning of the UAV in GPS-denied environments is achieved.
2. The method for constructing an autonomous positioning model for a UAV in a GPS-denied environment according to claim 1, wherein: Also includes: Based on the real-time position calculation results of the UAV in a GPS-denied environment, the linear velocity and linear acceleration of the UAV are calculated using differential technology and parameterized path scalars.
3. The method for constructing an autonomous positioning model for a UAV in a GPS-denied environment according to claim 1, wherein: The map dataset composed of the map data point pairs of the real-time map and the benchmark map collected by the UAV is constructed as follows: the real-time map and the benchmark map collected by the UAV are input, and the feature extraction algorithm is used to extract features from the real-time map and the benchmark map, and the map data point pairs corresponding to the real-time map and the benchmark map are obtained to construct the map dataset.
4. The method for constructing an autonomous positioning model for a UAV in a GPS-denied environment according to claim 2, wherein: In a GPS-denied environment, a ternary relationship map dataset based on local topological consistency is constructed, where the generation of triples must satisfy neighborhood constraints. The specific implementation is as follows: In a map dataset consisting of N map data point pairs For the i-th pair of map data points , based on spatial relationship search The K nearest neighbors of That is, K-nearest neighbors, only if 、 and When the relationship between the following formula is 、 and Three points form a triple , expressed as ; ； in, express Z i / k / j belong Z j / i / k K-nearest neighbors, Indicates when If true, the value is 1, otherwise it is 0; Map dataset After all pairs of map data points in are expressed through ternary relations, the initial data subset is obtained.
5. The method for constructing an autonomous positioning model for a UAV in a GPS-denied environment according to claim 4, wherein: The triplets are screened by local affine consistency verification and Delaunay triangulation, where the rank of the affine matrix verifies the validity of the triplet construction. The specific implementation is as follows: Extract triples based on the initial data subset obtained by expressing the ternary relationship of the map dataset , Represents a triple The corresponding coordinates, Transpose and get ; by As a reference point, construct Relative to a reference point The displacement vector is constructed into an affine matrix of three elements , calculate the rank r(A) of the affine matrix A. If r(A)<2, it means that there are more than two points in the three-tuple that overlap, the triple is invalid, and the score is recorded as 0; if r(A=2, the triple is valid; after the initial data subset is effectively screened for triples, a data subset containing M data is obtained ; Use the inverse matrix A of the affine matrix A -1 Acting on a vector , get the transformed point , point Map back through the inverse affine matrix The reference frame, check At the transformed position , calculate the transformed point The Euclidean distance is the transformation error E; the exponential function is used to convert the transformation error E into a geometric consistency score, that is, a local affine consistency score Calculate the lengths of the three sides of the triangle using the coordinates of the triplet points, then use the vector cosine theorem to calculate the internal angle of the triangle, then compare the angles of the triplet triangle with the angles of the regular triangle, and use an exponential function to convert the comparison results into an angle difference score, i.e., the Delaunay triangle triangulation score. ,in, 、 and are the three internal angles of the triplet triangle; finally, the local affine consistency score and the Delaunay triangle triangulation score are added to get the final triplet score .
6. The method for constructing an autonomous positioning model for a UAV in a GPS-denied environment according to claim 5, wherein: A compensation scoring strategy is proposed to adjust the triple weights based on the local geometric consistency score. The specific implementation is as follows: By defining the compensation term The scores are adjusted by weighting the local affine consistency score and the Delaunay triangle triangulation score to minimize unreliable matches; is the score threshold, is the compensation factor; the final triple compatibility score ; According to the final triple compatibility score , for a data subset containing M data Calculate the i-th data in the data subset B The sampling weight ;in, are the jth and kth data in the data subset B respectively, Representation and based on Associated triples The triple compatibility score of The maximum value of is not associated with any triple, then Set to 0; score by triple compatibility After processing data subset B, we get data subset .
7. The method for constructing an autonomous positioning model for a UAV in a GPS-denied environment according to claim 6, wherein: The data subset refinement strategy is adopted to screen the interior points through the iterative residual sorting stopping criterion. The specific implementation is as follows: Based on a subset of data , generating a model hypothesis , first based on the Simpson distance calculation model assumption With map dataset The residual value between ; Then, sort the residual values in ascending order, and map the dataset according to the sorted residual values. Sort the data in the data set, and extract the data before and after the mth data from the sorted data to form a new data subset, generate a new model hypothesis; calculate the new model hypothesis and the map data set The residual values are sorted in ascending order, and the map dataset is sorted according to the sorted residual values. The data in the data set are sorted, and the data before and after the mth data are extracted from the sorted data to form the next new data subset, and the next new model hypothesis is generated. This process is iteratively performed until a stable model hypothesis is obtained; the convergence of the data subset refinement process is determined based on the last three iterations, and a parameter, MaxIter, is used to represent the maximum number of iterations. The sampling weights obtained from the ternary relationship of are used as the conditions for the stopping criterion, and the stopping criterion is: ； Where, For the In the iteration, the jth data point of the data subset consisting of the data before and after the mth data is extracted from the data sorted by the residual value. For the The mth data obtained by sorting by residual value in the iteration; Express the relationship between and; is the sampling weight; h is the size of the data subset consisting of the data before and after the mth data in the data sorted by the residual value; Finally, based on the stable model assumption, outliers and inliers are distinguished and correctly matched map data point pairs are obtained.
8. The method for constructing an autonomous positioning model for a UAV in a GPS-denied environment according to claim 7, wherein: Based on the real-time position calculation results of the UAV in a GPS-denied environment, the linear velocity and acceleration of the UAV are calculated using differential technology and parameterized path scalars. The specific implementation is as follows: Assume that the real-time position spatial coordinate information of the drone is obtained by calculating its own real-time position in a GPS-denied environment for: ； Depend on The linear velocity of the drone in the x, y, and z directions is calculated using a single difference technique. 、 and : ； ； ； in, 、 and Indicates the spatial coordinate information of the drone at the next moment, 、 and Indicates the spatial coordinate information of the drone at the current moment, Indicates the time between two moments; Then, the linear acceleration of the drone in the x, y, and z directions is calculated by the quadratic difference technique. 、 and : ； ； ； in, 、 and Indicates the linear velocity of the drone at the next moment, 、 and Indicates the linear velocity of the drone at the current moment; Then the drone will pass through the GPS denial environment at every moment. The solved linear velocity and acceleration are scalarized on the path and converted into the linear velocity and acceleration on the path.
9. The method for constructing an autonomous positioning model for a UAV in a GPS-denied environment according to claim 8, wherein: The drone passes every moment in a GPS-denied environment The linear velocity and acceleration obtained by the solution are scalarized on the path and converted into the linear velocity and linear acceleration on the path. The specific implementation is as follows: Assume that the parameterized path of the UAV in the GPS denial environment is : ； Then, the derivative with respect to time t is obtained to obtain the component velocity 、 and : ； ； ； The total linear velocity V is: ； So the speed of the parameterized path scalar for: ； in, ; Then, the first-order derivative of the linear velocity V with respect to time is obtained: ； in, ; The acceleration of the parameterized path scalar is thus for 。