An eVTOL global positioning method

Abstract

The application provides an eVTOL global positioning method, in an environment with GNSS signals, an accurate environment map is constructed through image feature extraction; when GNSS signals are unavailable, an initial pose estimation is generated based on fusion of visual images and IMU data in a high-frequency visual inertial navigation thread, image and map points are matched in a low-frequency map point positioning thread, and the initial pose estimation and the matching result are fused, and finally high-precision global positioning of the eVTOL platform is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle positioning, and in particular to an eVTOL global positioning method. BACKGROUND

[0002] As a core vehicle of future urban air transportation system, eVTOL (electric vertical take-off and landing aircraft) has been widely used in logistics transportation, emergency rescue, field inspection, urban commuting and other diversified scenarios due to its vertical take-off and landing, low noise, zero emission and other characteristics. In these scenarios, the autonomous navigation and safe operation of the aircraft are highly dependent on high-precision real-time positioning information, and the current mainstream positioning scheme mainly relies on the absolute position service provided by GNSS (global navigation satellite system).

[0003] However, in actual application environments, especially in complex areas such as wild mountains, forests, valleys and urban high-rise buildings, GNSS signals are easily affected by terrain shielding, electromagnetic interference or human denial, resulting in signal loss or sharp decline in accuracy, which directly causes the failure of traditional GNSS-dependent positioning systems. At this time, the autonomous navigation capability of the aircraft is limited, which seriously threatens the safety and reliability of task execution.

[0004] To solve the problem of unstable GNSS signals, a positioning scheme that fuses vision and inertial navigation (IMU) has appeared in the prior art. This scheme can provide relatively stable pose estimation in a short time by fusing visual image information collected by a camera and motion state data output by an IMU. However, this scheme has two inherent defects: first, it cannot provide global absolute position information and can only achieve relative positioning, making it difficult to meet the global path planning requirements of long-distance and large-area tasks; second, the positioning error will accumulate over time, and in a long-term GNSS denial environment, the error will continue to expand, eventually leading to loss of control of navigation accuracy.

[0005] In addition, existing visual positioning technology also faces the problem of insufficient robustness in complex environments: traditional feature extraction methods have poor adaptability to changes in light, texture loss, dynamic obstacles and other scenarios, and are prone to feature point extraction failure or matching errors, resulting in positioning interruption; at the same time, in view of the long flight time and large coverage range of eVTOL, the existing map construction and matching scheme is difficult to balance the map accuracy and storage efficiency, and lacks a lightweight map management mechanism that adapts to the resource constraints of eVTOL platform, further limiting the practicality of the positioning system in actual scenarios.

[0006] Therefore, how to provide a technical scheme that can provide high-precision and high-robustness global positioning for eVTOL in a GNSS-limited environment has become a key problem to be solved in the field of unmanned aerial vehicle positioning. SUMMARY

[0007] The present application aims to provide an eVTOL global positioning method to solve the problems in the prior art and provide high-precision and high-robustness global positioning for eVTOL in GNSS limited environment.

[0008] The following presents a simplified summary of one or more aspects in order to provide a basic understanding of such aspects. This summary is not an extensive overview of all contemplated aspects, and is intended to neither identify key or critical elements of all aspects nor delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that is presented later.

[0009] According to an aspect of the present application, an eVTOL global positioning method is provided, comprising:

[0010] Step 100, constructing a map when GNSS signals are available, comprising:

[0011] Step 110, acquiring visual image information and GNSS positioning information collected by the eVTOL, and extracting image feature points and descriptors using a deep learning method;

[0012] Step 120, constructing a global map containing key frames and map points, wherein the key frames store poses, feature points and descriptors in the geocentric and Cartesian coordinate systems, and the map points contain three-dimensional positions, descriptors and covariance matrices;

[0013] Step 130, serializing storage through an incremental storage mechanism;

[0014] Step 200, reconstructing the map when GNSS signals are unavailable, and starting double-thread positioning, comprising:

[0015] Step 210, reading the serialized global map and reconstructing the map structure;

[0016] Step 220, visual inertial navigation thread: using a visual inertial navigation fusion method based on extended Kalman filtering, fusing real-time collected visual image data and IMU data to generate an initial pose estimate;

[0017] Step 230, map point positioning thread: extracting features from the current collected image, screening candidate key frames in combination with the initial pose estimate; matching the current image features with the map points corresponding to the candidate key frames, obtaining effective matching pairs after geometric verification, and outputting high-precision global positioning results.

[0018] In one possible embodiment, the step 120 of constructing a global map containing key frames and map points comprises:

[0019] Step 121, constructing a key frame: converting the longitude, latitude and height provided by GNSS positioning information in the Earth-Centered Earth-Fixed coordinate system to the Cartesian coordinate system with the initial pose as a reference; storing the position and pose of the key frame in the two coordinate systems, and also storing the image feature points and descriptors;

[0020] Step 122, performing feature matching between the current key frame and the previous key frame, if the matched previous feature points correspond to the constructed map points, associating the map points to the feature points of the current key frame; after data association, the feature points that are not matched are newly constructed as new map points.

[0021] Step 123, classifying the map points to construct a global map.

[0022] In one possible embodiment, step 123 further comprises constructing a local map.

[0023] Step 122 further comprises matching the current key frame with the local map.

[0024] In one possible embodiment, the classifying the map points in step 123 comprises:

[0025] The map points are divided into three types: uninitialized map points, to-be-optimized map points and optimized multiple times map points, in the processing process, for the uninitialized map points, the initial three-dimensional position is obtained by triangulation; for the to-be-optimized map points, the optimization is performed by minimizing the re-projection error through the observation constraints of multiple key frames; for the optimized multiple times map points, after the threshold determination, the optimized multiple times map points are stored in the global map, and the remaining map points that do not reach the threshold are deleted.

[0026] In one possible embodiment, the incremental storage mechanism in step 130 specifically comprises: when storing, the to-be-stored map points are constrained by the related key frames, and then a final optimization is completed; after the optimization, the Cartesian coordinates of the optimized map points are converted into the longitude, latitude and height in the Earth-Centered Earth-Fixed coordinate system, and the covariance matrix, key frame association information are combined for serialized compression storage, so as to realize the incremental expansion of the map.

[0027] In one possible embodiment, before the feature extraction of the current collected image in step 230, time synchronization processing is further included, specifically: the time stamp corresponding to the current processed image frame is kept consistent with the time stamp corresponding to the state information received in the visual inertial thread and the state to be updated.

[0028] In one possible embodiment, the filtering of the candidate key frame in step 230 adopts the nearest neighbor method, that is, based on the initial pose estimation, the K frame key frames closest to the global map are selected as the candidate key frames, wherein K is a positive integer.

[0029] In one possible embodiment, the geometric verification in step 230 includes: one-to-one matching of the current frame with the candidate key frame, screening the matching feature points by calculating the fundamental matrix, when the number of matching pairs is not less than 8 pairs, performing 3D-2D PnP verification, and retaining the matching pairs that satisfy the perspective geometric constraint.

[0030] In one possible embodiment, after the geometric verification in step 230, it further includes chi-square test, specifically: using the covariance matrix of the map point and the system covariance of the extended Kalman filter to perform statistical test on the matching pairs, and only retaining the matching pairs that pass the test as valid matching pairs.

[0031] In one possible embodiment, step 230 further includes updating the state and covariance matrix of the extended Kalman filter in the visual inertial thread based on the re-projection error constraint of the valid matching pairs.

[0032] The beneficial effects of the embodiments of the present application are:

[0033] The present application fuses GNSS information in the mapping phase to construct an environment map containing global position attributes. When GNSS signals are invalid due to obstruction, interference or restriction, the matching of the map and real-time visual image provides stable global absolute position information for eVTOL, solves the limitation of traditional visual inertial fusion scheme that can only realize relative positioning, ensures that the aircraft can still maintain global path planning and navigation capability in complex environments, breaks the dependence on GNSS, and realizes global absolute positioning.

[0034] The dual-thread architecture of "high-frequency visual inertial thread + low-frequency map point positioning thread" is adopted, the high-frequency thread guarantees real-time output, and the low-frequency thread dynamically corrects the cumulative error of visual inertial through the map matching result, forming a closed-loop correction mechanism. In the long-term limited GNSS environment, it is superior to the pure visual inertial scheme which diverges over time. It can suppress error accumulation and improve positioning accuracy.

[0035] The feature extraction method based on deep learning improves the stability of image features in scenes such as changes in light and missing textures; the multi-level matching verification strategy combining the fundamental matrix screening, 3D-2D PnP verification and chi-square test greatly reduces the influence of false matching pair positioning results, greatly improves the feature matching accuracy in complex scenes such as forest land and valley, and ensures the reliability of the positioning system. Enhance the robustness of complex environments.

[0036] The incremental map construction and serialization storage mechanism can dynamically expand the map size according to the long-distance flight requirements of eVTOL, while significantly reducing memory occupancy; the dual-thread design balances the consumption of computing resources and the real-time positioning of the aircraft, meets the hardware constraints of dynamic navigation, and improves the practicability. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced as follows. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. Other related drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0038] The above features and advantages of the present application can be better understood after reading the detailed description of the embodiments of the present disclosure in conjunction with the following drawings. In the drawings, the components are not necessarily drawn to scale, and components having similar related properties or features can have the same or similar reference numerals.

[0039] Figure 1 is a schematic diagram of a two-stage, dual-thread architecture of an embodiment of the present application;

[0040] Figure 2 is a specific method flowchart of an embodiment of the present application;

[0041] Figure 3 is a map point classification processing flowchart of an embodiment of the present application;

[0042] Figure 4 is an incremental storage mechanism flowchart of an embodiment of the present application. DETAILED DESCRIPTION

[0043] The present application will be described in detail below in conjunction with the drawings and specific embodiments. Note that the aspects described below in conjunction with the drawings and specific embodiments are only exemplary and should not be understood as limiting the scope of protection of the present application in any way.

[0044] As shown in Figure 1 , the present embodiment provides an eVTOL map-assisted visual positioning method suitable for GNSS-restricted environments, aiming to solve the problem of high-precision global positioning of eVTOL when GNSS signals are invalid due to obstruction, interference, restriction, etc. This method, through the two-stage design of map building when GNSS signals are available and map-assisted positioning when GNSS signals are not available, combined with the dual-thread fusion architecture of "high-frequency visual inertial thread" and "low-frequency map point positioning thread", can realize stable positioning in complex environments.

[0045] Figure 2 shows a specific method flowchart. Referring to Figure 2 , each step will be described in detail as follows:

[0046] 1. Map building phase when GNSS signals are available

[0047] The map construction stage takes visual image information and GNSS positioning information as input. The visual image and GNSS positioning information are collected in real time by the camera and GNSS module carried by the eVTOL.

[0048] In one embodiment, the resolution of the real-time collected visual image can be 1920x1080, and the frame rate is 30 fps. The real-time collected GNSS positioning information includes longitude, latitude, height in the Earth-Centered Earth-Fixed coordinate system (ECEF), and a timestamp.

[0049] Feature extraction: a deep learning-based feature extraction method is used to extract feature points (e.g., about 500 per frame) and their descriptors (e.g., 256-dimensional vectors) from each frame of image by a feature extraction model. The descriptors are used for cross-frame feature matching between images, and the coordinate information of the feature points is applied to the three-dimensional position estimation of the map points. In this embodiment, the feature extraction model can be SuperPoint.

[0050] Construction of key frames: when an image frame meets the preset key frame condition (e.g., the translation amount from the previous key frame is >0.5 m or the rotation angle is >5°), it is marked as a key frame. Since the GNSS positioning information provides longitude, latitude, and height in the Earth-Centered Earth-Fixed coordinate system (ECEF), it needs to be converted to the Cartesian coordinate system with the initial pose as the reference. Through a coordinate conversion algorithm, the ECEF coordinates (X, Y, Z) provided by the GNSS are converted to the Cartesian coordinate system (x, y, z) with the initial pose as the origin, and the poses, feature points, and descriptors in the two coordinate systems are stored synchronously in the key frame. The pose includes position and attitude angle.

[0051] Data association: the feature matching between the current key frame and the previous key frame is performed, e.g., based on a brute-force search algorithm. If the matched previous feature point corresponds to an already constructed map point, the map point is associated with the feature point of the current key frame. After data association, the feature points that are not matched are newly created as a new map point, and its initial three-dimensional position is calculated based on the pose constraints of the two key frames through a triangulation algorithm.

[0052] Preferably, the current key frame can also be matched with the local map, i.e., the feature points of the current key frame are matched with the feature points of the local map (containing the map points observed by the last 10 key frames), to expand the search range and ensure that as many map points as possible are matched.

[0053] Map point processing: see Figure 3 The map points can be divided into three states:

[0054] Uninitialized: points that are just generated and have not calculated three-dimensional positions, which are upgraded to the to-be-optimized state after triangulation;

[0055] Optimization: After the initial 3D position is obtained, the reprojection error is minimized (target error < 10 pixels) by using bundle adjustment (BA) to observe the data of 5-10 key frames in the sliding window, and the 3D position is iteratively optimized.

[0056] Optimized multiple times: points with stable error (change < 0.1 m) after 10 consecutive optimizations are added to the global map candidate set; points that have not been observed for more than 20 frames and have been optimized less than 5 times are directly deleted from the map; points that have not been observed for more than 10 frames and have not been successfully triangulated are also deleted.

[0057] By classifying the map points, the storage efficiency of the map can be improved, and the accuracy of the map points can be guaranteed.

[0058] Incremental storage mechanism: Since eVTOL generally flies a long distance and involves a large scene, waiting for the complete map to be built will consume a lot of memory. This method uses an incremental storage method to save memory during runtime.

[0059] As shown in Figure 4 When storing, the related key frames of the map points to be stored are used to constrain them, and a final optimization is performed. Based on the optimization results, the covariance matrix of the position uncertainty is calculated (based on the residual estimation during the optimization process), and the observation key frame ID is associated. At the same time, the Cartesian coordinates are converted back to the ECEF coordinate system. The ProtocolBuffers serialization format is used to determine whether the proportion of shared map points between the two saved key frames meets the set threshold. If it is less than the set threshold, the new key frame is saved. For example, if the proportion of shared map points between the current key frame and the last saved key frame is less than 0.6, the new key frame is saved. The stored data includes feature points, descriptors, map point numbers, latitude, longitude, height, covariance, and key frame association information. After compression, a single frame of data is about 500KB, which can better adapt to the storage resource limitations of eVTOL.

[0060] 2. Positioning phase when GNSS signal is not available

[0061] The positioning phase is a method of real-time online global positioning using visual inertial information and maps without GNSS global positioning signals.

[0062] Since the online image feature extraction uses a deep learning model, its extraction frequency will be lower than the fusion frequency of visual inertial navigation, so this method divides the positioning phase into a high-frequency visual inertial thread and a low-frequency map point positioning thread.

[0063] High-frequency visual-inertial thread is based on MSCKF method of extended Kalman filter: fusion of image and IMU data, provide high frequency and real-time initial pose estimation, ensure short-time positioning stability.

[0064] Low-frequency map point positioning thread: based on map matching results, low-frequency correction of visual-inertial system cumulative error, the low-frequency matches the frequency of deep learning feature extraction, and outputs the optimized global pose.

[0065] High-frequency visual-inertial thread provides initial pose for low-frequency map point positioning thread, reducing the search range of map matching; map positioning thread updates the state and covariance matrix of extended Kalman filter through information fusion (reprojection error constraint), which in turn optimizes the visual-inertial system and forms a dynamic correction closed loop.

[0066] The specific process is as follows:

[0067] 2.1 Map reconstruction: the map is stored as a serialized data stream, so it is first read as corresponding key frame and map point information, and the map structure that can be used for positioning is reconstructed through the association relationship between key frames and map points;

[0068] 2.2 High-frequency visual-inertial thread: MSCKF (Multi-State Constraint Kalman Filter) based on extended Kalman filter data fusion method is used to fuse camera image (10Hz) and IMU data (sampling rate 200Hz, output acceleration and angular velocity), to provide initial position for map positioning thread.

[0069] MSCKF is a data fusion algorithm specially optimized for visual-inertial system (VI-SLAM), which is a variant of extended Kalman filter (EKF). It forms a sliding window by retaining multiple camera states (key frames), and uses the time redundancy of visual measurement to improve estimation accuracy.

[0070] MSCKF estimates the IMU state and multiple camera poses jointly, which allows constraints to be established between the current frame and historical frames, and uses visual reprojection error to optimize the IMU state. In addition, MSCKF uses IMU pre-integration technology to process high-frequency IMU data. Through multi-frame constraints and pre-integration technology, the limitations of traditional EKF in visual-inertial systems are effectively solved, providing key support for high-precision positioning of eVTOL in complex environments.

[0071] In particular, the state and covariance matrix information updated by the map point positioning thread will also be continuously applied to the system, further improving the overall positioning accuracy.

[0072] 2.3 Low-frequency map point positioning thread:

[0073] 2.3.1 Time synchronization: The timestamp of the image frame processed by the map localization thread should be consistent with the timestamp of the state information received from the visual-inertial system and the state to be updated. Therefore, the timestamps of the current image frame and the state output by the visual-inertial system are calibrated (error < 1 ms), and if there is asynchronization, the next synchronized data is waited for.

[0074] 2.3.2 Feature extraction: The same as 1.1, a deep learning-based feature extraction method can be used to extract image feature points and their descriptors.

[0075] 2.3.3 Alternative key frame screening: The visual-inertial system provides rough initial position information, but the large map matching range can cause low efficiency. Therefore, in this embodiment, after receiving the initial position information, the map localization thread calculates the Euclidean distance between the initial position and all key frames in the map, and uses the nearest neighbor method to screen the K (for example, 3) frames closest to the initial position in the map, thereby narrowing the matching range and improving the real-time positioning.

[0076] 2.3.4 Feature matching and geometric verification: One-to-one matching is performed between the current frame and the alternative key frame, and the matching feature points are screened by calculating the fundamental matrix, for example, using the RANSAC algorithm with 100 iterations. If the matching feature points of the alternative key frame correspond to the map points, a matching pair of feature points and map points under the current frame is generated, at least 8 matching pairs are required, and 3D-2D PnP verification is performed to verify the spatial geometric consistency, thereby reducing the influence of false matching pairs on positioning accuracy.

[0077] 2.3.5 Chi-square test: In order to further eliminate false map points, a chi-square test is performed using the map points, their covariance matrix, and the system covariance in the extended Kalman filter. Based on the map point covariance and the Kalman filter system covariance, the statistic is calculated for the matching pairs that pass the geometric verification, and only the matching pairs that satisfy the χ² distribution are retained as valid matching pairs. Only the map point matching that passes the chi-square test can enter the information fusion and state update module, thereby improving the matching robustness and reducing the influence of outliers.

[0078] 2.3.6 Information fusion and state update: The re-projection error of the valid matching pair is used as a constraint to update the state vector such as position, velocity, and attitude of the extended Kalman filter and the covariance matrix, so that the positioning error is corrected to within 10 m.

[0079] In the above process, the high-frequency thread provides the initial pose to the low-frequency thread to narrow the matching range, and the optimization result of the low-frequency thread is fed back to the high-frequency thread to reset the inertial navigation error accumulation, forming a dynamic correction closed loop, ensuring that the horizontal error is within 5 meters and the vertical error is within 10 meters within 30 kilometers of GNSS restricted positioning error.

[0080] In summary, the embodiment of the present application provides an eVTOL map-assisted visual positioning method suitable for GNSS restricted environment, which is used to realize high-precision and real-time global positioning based on image information collected by the aircraft and the pre-constructed map in the case of GNSS unavailability or unreliability. In the environment with GNSS signal, an accurate environment map is constructed through image feature extraction; when GNSS signal is unavailable, the system generates an initial pose estimation based on the fusion of visual image and IMU data, and fuses the initial estimation with the matching result of the image and the three-dimensional points in the map, finally realizes high-precision global positioning of the eVTOL platform.

[0081] Through actual scene test, the technical effects of the method are as follows:

[0082] After the GNSS signal is completely lost, the global positioning can be restored through map matching within 3 seconds, solving the problem of absolute position loss; compared with the pure visual inertial navigation scheme, the 10-kilometer positioning error is reduced from horizontal error 30m and vertical error 5m to horizontal error 3m and vertical error 3m. The 30-kilometer positioning error is reduced from horizontal error 150m and vertical error 10m to horizontal error 5m and vertical error 3m.

[0083] Robustness is improved. In the scene of serious feature blocking in the wild such as forest land, the matching accuracy is improved from 65% of the traditional method to 92% due to deep learning features and multi-layer verification.

[0084] Dynamic scalability. Incremental map construction and serialized storage mechanism can dynamically expand the map size according to the long-distance flight requirements of eVTOL, while significantly reducing the memory occupation.

[0085] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between each embodiment can be referred to each other.

[0086] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein, but should be granted the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. An eVTOL global positioning method, characterized in that: include: Step 100: When GNSS signals are available, construct a map, including: Step 110: Obtain visual image information and GNSS positioning information collected by the eVTOL, and use a deep learning method to extract image feature points and descriptors; Step 120: construct a global map including keyframes and map points, wherein the keyframes store poses, feature points, and descriptors in an Earth-centered Earth-fixed coordinate system and a Cartesian coordinate system, and the map points include three-dimensional positions, descriptors, and covariance matrices; Step 130: Serialize and store data using an incremental storage mechanism. Step 200: When the GNSS signal is unavailable, rebuild the map and start dual-thread positioning, including: Step 210: read the serialized stored global map and rebuild the map structure; Step 220, visual inertial navigation thread: adopt a visual inertial navigation fusion method based on extended Kalman filter to fuse the real-time collected visual image data and IMU data to generate an initial pose estimate; Step 230, map point positioning thread: extract features from the currently captured image, and filter candidate key frames based on the initial pose estimation; match the current image features with the map points corresponding to the candidate key frames, obtain a valid match after geometric verification, and output a high-precision global positioning result.

2. The eVTOL global positioning method according to claim 1, characterized in that: In step 120, constructing a global map including keyframes and map points includes: Step 121: Construct a keyframe: convert the latitude, longitude, and altitude in the Earth-centered Earth-fixed coordinate system provided by the GNSS positioning information into a Cartesian coordinate system with the initial pose as a reference; store the position and pose of the keyframe in both coordinate systems, and also store the image feature points and descriptors; Step 122: Perform feature matching between the current keyframe and the previous keyframe. If the matched feature point of the previous keyframe corresponds to a constructed map point, the map point is associated with the feature point of the current keyframe. After data association, any feature point that is not matched is created as a new map point. Step 123: Classify the map points and construct a global map.

3. The eVTOL global positioning method according to claim 2, characterized in that: Step 123 also includes constructing a local map; Step 122 also includes: matching the current key frame with the local map.

4. The eVTOL global positioning method according to claim 2, characterized in that: The classification process of the map points in step 123 includes: Map points are divided into three types: uninitialized map points, map points to be optimized, and map points that have been optimized multiple times. During the processing, for uninitialized map points, the initial three-dimensional position is obtained through triangulation; for map points to be optimized, optimization is performed by minimizing the reprojection error through observation constraints of multiple key frames; for map points that have been optimized multiple times, they are stored in the global map after the optimization number threshold is determined, and the remaining map points that do not reach the threshold are deleted.

5. The eVTOL global positioning method according to claim 1, characterized in that: The incremental storage mechanism in step 130 is as follows: during storage, the map points to be stored are constrained by their associated keyframes, and a final optimization is performed. The Cartesian coordinates of the optimized map points are converted into longitude, latitude, and altitude in an Earth-centered, Earth-fixed coordinate system, and serialized and compressed storage is performed in combination with the covariance matrix and keyframe association information to achieve incremental expansion of the map.

6. The eVTOL global positioning method according to claim 1, characterized in that: In step 230, before feature extraction is performed on the currently captured image, time synchronization processing is also included, specifically: the timestamp corresponding to the currently processed image frame is made consistent with the state information received in the visual inertial navigation thread and the timestamp corresponding to the state to be updated.

7. The eVTOL global positioning method according to claim 2, characterized in that: In step 230 , the candidate key frames are screened using a nearest neighbor method, that is, based on the initial pose estimation, the nearest K key frames are screened from the global map as candidate key frames, where K is a positive integer.

8. The eVTOL global positioning method according to claim 7, characterized in that: The geometric verification in step 230 includes: matching the current frame with the candidate key frames one by one, screening the matching feature points by calculating the basic matrix, and when the number of matching pairs is not less than 8, performing 3D-2DPnP verification to retain the matching pairs that meet the perspective geometry constraints.

9. The eVTOL global positioning method according to claim 8, characterized in that: In step 230 , after the geometric verification, a chi-square test is also included, specifically: using the covariance matrix of the map points and the system covariance of the extended Kalman filter, a statistical test is performed on the matching pairs, and only the matching pairs that pass the test are retained as valid matching pairs.

10. The eVTOL global positioning method according to claim 9, characterized in that: Step 230 further includes updating the state and covariance matrix of the extended Kalman filter in the visual inertial navigation thread based on the reprojection error constraint of the valid matching pair.

Images