GNSS / VIO fusion high-precision positioning method based on graph optimization

Abstract

The invention provides a GNSS / VIO fusion high-precision positioning method based on graph optimization. A system is composed of a data acquisition module, a preprocessing module, a visual inertial odometer module and a graph optimization fusion module. The system integrates a monocular camera, an IMU (inertial measurement unit) and a GNSS-RTK (global navigation satellite system-real-time kinematic) receiver, and establishes a unified multi-factor graph optimization model by constructing a visual re-projection factor, an IMU pre-integration factor and a GNSS-RTK factor. The core innovation point of the method is that local estimation of the visual inertial odometer is corrected in real time in a sliding window by utilizing global absolute position constraint provided by a GNSS (Global Navigation Satellite System), so that the increase of accumulative errors of VIO (Visual Inertial Odometer) is effectively inhibited. The back end of the system adopts a nonlinear least square method to optimize and solve the pose, the speed and the IMU zero offset of the carrier in real time; and when the GNSS signal is shielded or lock losing occurs, high-precision positioning output is still kept depending on the continuity of the VIO. In conclusion, through collaborative modeling and joint optimization of the multi-source sensor, the limitations of insufficient precision, signal interruption and the like of a single sensor in a complex shielding environment are overcome, and high-precision and high-robustness continuous positioning is realized.

Description

Technical Field

[0001] This invention belongs to the field of satellite navigation and positioning technology, and particularly relates to a high-precision positioning method based on graph optimization GNSS / VIO fusion. Background Technology

[0002] With the rapid development of autonomous driving, real-time positioning and navigation of vehicles and robots during operation has become a critical issue. Global Navigation Satellite Systems (GNSS) offer global coverage, all-weather operation, and real-time capabilities, providing high-precision positioning in open areas. However, in complex operational scenarios such as urban areas with high-rise buildings and rural streets, satellite information suffers from factors like obstruction and multipath reflection during propagation, posing significant challenges to the positioning accuracy, stability, and precision of GNSS navigation services.

[0003] Currently, various types of sensors are used for positioning in unmanned systems, including WiFi, ultra-wideband, ground pseudosatellites, cameras, inertial measurement units (IMUs), and lidar. Monocular visual inertial odometry (VIO), by fusing visual SLAM with an IMU inertial sensor, solves the scale ambiguity problem inherent in monocular vision and significantly enhances positioning accuracy under conditions of changing lighting and motion ambiguity. Furthermore, it maintains good trajectory estimation accuracy during short-term, small-scale movements, but some problems still need to be addressed. Because monocular VIO inherently has four degrees of freedom with unobservables, and lacks an external global reference system for calibration, the system inevitably accumulates errors over long periods of operation.

[0004] Achieving high-precision navigation and positioning in complex scenarios is a key challenge. Relying solely on a single sensor often fails to simultaneously meet the dual requirements of accuracy and stability. Multi-sensor fusion methods can leverage the strengths of different sensor types, processing data uniformly and fully utilizing their respective advantages to improve the system's positioning accuracy and robustness. In summary, in complex urban scenarios, GNSS and monocular VIO can be fused for positioning to overcome the limitations of single-sensor positioning. GNSS provides a global, absolute coordinate reference for the entire system, while monocular VIO enables high-frequency relative displacement estimation within a local area. This not only effectively curbs the cumulative error of VIO but also fills the positioning gaps caused by GNSS signal obstruction. Working together, they ultimately achieve high-precision and highly robust positioning. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a high-precision positioning method based on graph optimization for GNSS / VIO fusion.

[0006] To achieve the above and other related objectives, this invention provides a graph-optimized GNSS / VIO fusion high-precision positioning method, which includes the following steps:

[0007] Step 1: Acquire monocular camera data, IMU data, and GNSS-RTK observation data;

[0008] Step 2: Perform time synchronization and coordinate system alignment of camera, IMU, and GNSS data, and perform data preprocessing;

[0009] Step 3: Extract features from the monocular image, complete inter-frame matching, obtain camera motion constraints, and pre-integrate the IMU data to obtain the preliminary pose estimate of the camera in the local coordinate system.

[0010] Step 4: Construct a graph optimization fusion framework, establish visual reprojection factor, IMU pre-integration factor and GNSS-RTK factor, and achieve multi-source data fusion through nonlinear least squares optimization;

[0011] Step 5: Graph optimization is performed using a sliding window approach. During the optimization process, VIO provides high-frequency local constraints, and GNSS-RTK provides low-frequency global constraints. The final output is the optimized state variables such as camera / carrier pose, velocity, and IMU bias.

[0012] Step 6: Compare the optimized pose results with the ground truth to evaluate the positioning accuracy and robustness, and verify the algorithm's continuous positioning capability in a GNSS signal obstruction environment.

[0013] Preferably, image data acquired by a monocular camera, inertial measurement unit (IMU) data, and GNSS-RTK observation data are obtained.

[0014] The GNSS-RTK observation data can provide global constraint information, and the camera and IMU data can provide high-frequency local motion constraint information.

[0015] Preferably, the camera, IMU, and GNSS-RTK data are time-synchronized and coordinate-aligned, and data preprocessing is completed to ensure that multi-source measurement information participates in fusion calculation within a unified spatiotemporal framework.

[0016] The time synchronization method includes two approaches: hardware synchronization and software timestamp alignment. The coordinate system alignment includes sensor extrinsic parameter calibration and unified coordinate system transformation.

[0017] Preferably, feature point extraction and inter-frame matching are performed on the image data from the monocular camera to construct camera motion constraints; at the same time, the IMU data is pre-integrated to obtain a preliminary pose estimate of the camera in the local coordinate system.

[0018] The feature extraction method described above uses the Tomasi corner detection algorithm to extract image feature points and combines it with KLT sparse optical flow to achieve feature point tracking.

[0019] The pre-integration process described above takes into account error models such as IMU bias and noise.

[0020] Preferably, a graph optimization-based fusion framework is constructed, in which a visual reprojection factor, an IMU pre-integration factor, and a GNSS-RTK factor are established in the graph model.

[0021] The visual reprojection factor is constructed based on feature point matching error, the IMU pre-integration factor is constructed based on kinematic constraints between consecutive frames, and the GNSS-RTK factor is constructed based on absolute position observation residuals.

[0022] Preferably, a nonlinear least squares optimization method is used to jointly optimize and solve the above factors, and output state variables such as camera / carrier pose, velocity and IMU bias.

[0023] The optimization method described is a nonlinear optimization method;

[0024] Preferably, a sliding window optimization strategy is adopted to maintain key frames and state variables with a finite time sequence within the window.

[0025] The VIO factor provides high-frequency local constraints, while the GNSS-RTK factor provides low-frequency global constraints. The two complement each other during the window optimization process, thereby ensuring accuracy locally and zero drift globally.

[0026] Preferably, comparative experiments show that, under the same test scenario, the graph optimization-based GNSS / VIO fusion positioning method proposed in this invention can effectively suppress the cumulative error generated by a single VIO, significantly reducing the average trajectory error and root mean square error. Experimental results show that the algorithm achieves an average absolute trajectory error of approximately 1.1 m and an RMSE of approximately 0.87 m. Compared with the pure VIO method, which exhibits significant drift, the fusion algorithm of this invention achieves a significant improvement in overall positioning accuracy and continuous positioning stability.

[0027] As described above, the GNSS / VIO fusion high-precision positioning method based on graph optimization proposed in this invention has the following beneficial effects:

[0028] (1) In terms of data processing, this invention first performs time alignment and coordinate system unification on the observation data of monocular camera, IMU and GNSS-RTK, so that the information of multiple sensors is under the same spatiotemporal reference. Then, through image feature extraction, adjacent frame matching and IMU pre-integration processing, a high-frequency local motion constraint relationship is constructed, which provides a reliable initial value input for the optimization algorithm, thereby enhancing the positioning accuracy of multi-source data fusion.

[0029] (2) In terms of fusion modeling, this invention proposes a multi-source constraint joint framework based on graph optimization. By introducing visual reprojection factors, IMU pre-integration factors, and GNSS-RTK factors into the graph model, local and global information are tightly coupled and fused. Compared with traditional loosely coupled methods, the tightly coupled modeling approach of this invention can better utilize the complementary advantages of various sensors, thereby significantly improving positioning accuracy and system stability.

[0030] (3) In terms of optimization algorithm, this invention adopts a nonlinear least squares optimization method combined with a sliding window strategy, so that the VIO factor provides high-frequency local constraints and the GNSS-RTK factor provides low-frequency global constraints, thus achieving an organic unity of local drift-free and global accuracy. This method can still maintain a continuous and stable positioning effect in GNSS occlusion or interference environments, significantly enhancing the robustness of the system.

[0031] (4) In terms of application value, the method of this invention provides a solution to the limitations of a single navigation system in confined environments. This method integrates multi-source sensor data and uses a graph optimization framework to achieve joint estimation, improving positioning accuracy and system stability while ensuring real-time computation. The method is applicable to scenarios such as urban canyons, tree-covered areas, and indoor-outdoor transitions, and can be applied to autonomous driving, mobile robots, and precision measurement fields. Attached Figure Description

[0032] To further illustrate the contents described in this invention, the specific embodiments of the invention will be described in more detail below with reference to the accompanying drawings. It should be understood that these drawings are merely typical examples and should not be construed as limiting the scope of the invention.

[0033] Figure 1 This is a flowchart of a graph-optimized GNSS / VIO fusion high-precision positioning method provided in this invention.

[0034] Figure 2 This is a flowchart of the image information preprocessing process;

[0035] Figure 3 This is a flowchart of the vision-inertial initialization process;

[0036] Figure 4This is a framework diagram of a graph-optimized GNSS / visual / inertial tightly coupled positioning algorithm;

[0037] Figure 5 This is a heatmap of the absolute error of the localization trajectory sequence in the KITTI dataset. Detailed Implementation

[0038] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following specific embodiments are only used to explain this application.

[0039] This invention provides a graph optimization-based GNSS / VIO fusion high-precision positioning method. Specifically, by fusing observation data from a monocular camera, IMU, and GNSS-RTK within a graph optimization framework, it overcomes the cumulative error of VIO and the GNSS signal obstruction problem, achieving high-precision positioning. The method includes the following steps:

[0040] Step 1: Acquire monocular camera data, IMU data, and GNSS-RTK observation data;

[0041] Step 2: Perform time synchronization and coordinate system alignment of camera, IMU, and GNSS data, and perform data preprocessing. The specific process is as follows:

[0042] 1) GNSS uses the global ECEF coordinate system, while VIO operates in a local coordinate system with the initial camera optical center as the origin. When not aligned with the global system, the local system has four unobservables relative to ECEF: three-dimensional translation and yaw angle (θ); pitch and roll can be determined by gravity constraints. Rotation is first represented by the offset in the θ direction. This is done by rotating the local coordinate system to the northeast-sky coordinate system corresponding to its origin. Then, based on the position of the origin of the local coordinate system in the ECEF coordinate system... The specific formula for transforming the northeast-sky coordinate system to the ECEF coordinate system is as follows:

[0043]

[0044] in, This indicates the position of a point in the local coordinate system. This indicates the position of the point in the ECEF coordinate system. The pose and velocity obtained from visual inertia are transformed to the ECEF coordinate system through the above coordinate transformation.

[0045] 2) Due to the different time bases used by GNSS and visual-inertial systems (VIS), it is necessary to unify their time descriptions. Specifically, the IMU frequency is 200Hz, the camera frequency is 20Hz, the user-end GNSS receiver frequency is 10Hz, and the base station GNSS receiver frequency is 1Hz. Time fusion between the GNSS and VIS systems is achieved by comparing the time reported by the user-end GNSS receiver with the time reported by the VIS. Based on the frequency relationship between the user-end GNSS and the camera, within the time frame of receiving one set of GNSS information, image information from two sets of cameras will be received. The specific formula for this principle is as follows:

[0046]

[0047] in, Indicates the time of the current image information. The time of the previous image information. For GNSS

[0048] The difference between time and time inertia time, This indicates the GNSS time at which the user terminal receives GNSS information.

[0049] The monocular camera data includes image frame sequences and their timestamps, and the IMU data includes the raw measurement values ​​of the three-axis accelerometer and the three-axis gyroscope and their corresponding timestamps.

[0050] Step 3: Extract features from the monocular image, complete inter-frame matching, obtain camera motion constraints, and pre-integrate the IMU data to obtain the preliminary pose estimate of the camera in the local coordinate system.

[0051] The feature extraction described herein uses the Tomasi corner detection algorithm to extract image feature points and combines it with KLT sparse optical flow to achieve feature point tracking. To address the difficulty in extraction caused by excessively dark or bright lighting, adaptive histogram equalization is introduced to enhance contrast.

[0052] During the tracking process, the fundamental matrix and RANSAC method are used to remove false matches and outliers. When the number of effective feature points is insufficient to meet a set threshold, corner point extraction is performed again to supplement them. A feature point spacing threshold is set to ensure uniform distribution of features in the image, and a priority is set based on the frequency of feature point occurrences, prioritizing the retention of feature points with longer durations.

[0053] When selecting keyframes, if the number of co-view feature points between the current frame and the latest keyframe exceeds 20, it is determined whether the average disparity of these co-view points exceeds a set threshold. The specific formula is as follows:

[0054]

[0055] in and Indicates the two shaft and Axis pixel difference, This indicates the number of feature points shared by both.

[0056] The IMU data preprocessing mainly involves pre-integrating the raw acceleration and angular velocity measurements from the IMU to obtain the rotation and translation relationship between two image frames. This is done at the corresponding moments of the keyframes in the two images. During this period, the corresponding IMU device state to state The position, velocity, and rotation relationships in the local coordinate system can be obtained by integrating the IMU measurements, as shown in the following formula:

[0057]

[0058]

[0059]

[0060] in Represents the time interval between corresponding moments of two image keyframes. .

[0061] The transformation from the local coordinate system to the IMU coordinate system allows for the separation of integral terms that are only related to the IMU's acceleration and angular velocity measurements. Therefore, pre-integration can be performed simultaneously with the IMU measurements, allowing for subsequent calculations; this is called the pre-integration process. The specific formula is as follows:

[0062]

[0063]

[0064]

[0065] in

[0066]

[0067]

[0068]

[0069] At the initial moment , Set to 0, Let it be a unit quaternion. Due to the noise term... , Since these are unknown terms, the pre-integration is used to estimate the values ​​of , , and , as shown in the following formula:

[0070]

[0071] The IMU pre-integration process first performs time alignment and noise filtering on the original measurements and corrects the accelerometer and gyroscope biases; then, between two keyframes, the continuous acceleration and angular velocity observations are discretely integrated to accumulate the pre-integrated quantities, including displacement increment, velocity increment and attitude increment.

[0072] Step 4: Construct a graph optimization fusion framework, establish visual reprojection factor, IMU pre-integration factor, and GNSS-RTK factor, and achieve multi-source data fusion through nonlinear least squares optimization. The specific steps are as follows:

[0073] 1) Establish a visual-inertial probabilistic model, and express the state variables to be determined in the form of factors through the relationship between the observed quantities and the state variables. The specific formula is as follows:

[0074]

[0075] Where 𝑓(𝑥) represents the relationship between the observed quantity and the state quantity. It is the residual of the pose calculated from the IMU observed quantity between two adjacent carrier pose state quantities. The factor is proportional to the probability, i.e.

[0076]

[0077]

[0078] The above formula can be transformed into the following formula:

[0079]

[0080] The optimized state variables within the system's sliding window This includes the pose of the IMU in the local coordinate system at each corresponding moment of each image frame. and speed Rotation And the IMU's own acceleration bias term and angular velocity offset term The inverse depth of all feature points within the sliding window in the camera coordinate system of the image first observed. And external parameters of the camera and IMU, The formula for representing the size of the sliding window is as follows:

[0081]

[0082] The factor graph method described above characterizes the constraint relationship between state variables and observations through a residual model, thereby constructing an optimization problem. This method achieves joint optimization of the system state variables by minimizing prior information and the residual terms corresponding to all observations within the sliding window. The factors used to express the nonlinear relationship between observations and state variables mainly include: marginalization factors, IMU pre-integration factors, and visual reprojection factors.

[0083] 2) Establish the IMU pre-integration factor, where the state variable to be optimized in the IMU is the pose. and speed Rotation Acceleration bias term Angular velocity bias term The corresponding observations are the pre-integral terms obtained by integrating acceleration and angular velocity. , , The specific formula is as follows:

[0084]

[0085] in This indicates that the vector portion of the quaternion has been extracted, and the three-dimensional variable is represented. Rotation. The residuals from solving the equations were used to optimize the rotation between two adjacent image keyframes within the sliding window. , , , , State variables for factor optimization The specific formula is as follows:

[0086]

[0087] 3) Establish visual reprojection factor

[0088] The core of visual measurement error lies in the reprojection error of feature points. Taking feature point A as an example, it is observed in both frame i and frame j. The coordinates of point A in the camera coordinate system of frame i are transformed to the coordinate system of frame j using relevant state variables. The difference between this transformed coordinates and the actual observed coordinates in frame j yields the visual measurement residual. (The coordinates of the feature point in the camera coordinate system are then used.) Its corresponding camera normalized coordinates The relationship is proportional, and the specific formula is as follows:

[0089]

[0090] in, Let be the depth value of the feature point. Starting from the camera coordinate system of frame i, the feature points are sequentially transformed to the IMU coordinate system of frame i, then to the world coordinate system, then to the IMU coordinate system of frame j, and finally projected onto the camera coordinate system of frame j. The specific formula is as follows:

[0091]

[0092] The above relationship can be expressed using state variables, as shown in the following formula:

[0093]

[0094] Among them, the optimized state variables are: , , and feature point depth value .

[0095] Step 5: Graph optimization is performed using a sliding window approach. During the optimization process, VIO provides high-frequency local constraints, and GNSS-RTK provides low-frequency global constraints. The final output is the optimized state variables such as camera / carrier pose, velocity, and IMU bias.

[0096] The sliding window method reduces computational complexity and improves real-time performance by setting a fixed window size and optimizing calculations only for the current moment and a few historical frames. The VIO high-frequency local constraints provide relative motion constraints between the camera and the IMU within a short time frame through IMU pre-integration factors and visual reprojection factors. The GNSS-RTK low-frequency global constraints provide positional information constraints on a longer time scale within the window through GNSS-RTK factors, ensuring global consistency.

[0097] The method described above minimizes global error by jointly optimizing multiple constraint factors, thereby obtaining accurate optimization results for state variables such as camera / carrier pose, velocity, and IMU bias.

[0098] Step six involves comparing the optimized pose results with the ground truth to evaluate positioning accuracy and robustness, and verifying the algorithm's continuous positioning capability under GNSS signal obstruction conditions. The comparative experiments show that, under the same test scenario, the graph optimization-based GNSS / VIO fusion positioning method proposed in this invention can effectively suppress the cumulative error generated by a single VIO, significantly reducing the average trajectory error and root mean square error. Experimental results show that the algorithm's mean absolute trajectory error is approximately 1.1 m, and the RMSE is approximately 0.87 m. Compared with the pure VIO method, which exhibits significant drift, the fusion algorithm of this invention achieves a significant improvement in overall positioning accuracy and continuous positioning stability.

[0099] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. By referring to the detailed description of the embodiments, any person skilled in the art can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A graph optimization-based GNSS / VIO fusion high-precision positioning method is proposed. By fusing observation data from a monocular camera, IMU, and GNSS-RTK within a graph optimization framework, the cumulative error of VIO and the GNSS signal obstruction problem are overcome, achieving high-precision positioning. Its key features are: The method includes the following steps: Step 1: Acquire monocular camera data, IMU data, and GNSS-RTK observation data; Step 2: Perform time synchronization and coordinate system alignment of camera, IMU, and GNSS data, and perform data preprocessing; Step 3: Extract features from the monocular image, complete inter-frame matching, obtain camera motion constraints, and pre-integrate the IMU data to obtain the preliminary pose estimate of the camera in the local coordinate system. Step 4: Construct a graph optimization fusion framework, establish visual reprojection factor, IMU pre-integration factor and GNSS-RTK factor, and achieve multi-source data fusion through nonlinear least squares optimization; Step 5: Graph optimization is performed using a sliding window approach. During the optimization process, VIO provides high-frequency local constraints, and GNSS-RTK provides low-frequency global constraints. The final output is the optimized state variables such as camera / carrier pose, velocity, and IMU bias. Step 6: Compare the optimized pose results with the ground truth to evaluate the positioning accuracy and robustness, and verify the algorithm's continuous positioning capability in a GNSS signal obstruction environment.

2. The GNSS / VIO fusion high-precision positioning method based on graph optimization according to claim 1, characterized in that, Step two specifically involves: synchronizing the time and aligning the coordinate systems of the camera, IMU, and GNSS data, and performing data preprocessing.

3. The monocular camera data includes image frame sequences and their timestamps; the IMU data includes the raw measurement values ​​of the three-axis accelerometer and the three-axis gyroscope and their corresponding timestamps; the GNSS-RTK observation data includes carrier phase, pseudorange, ephemeris information, differential corrections and observation timestamps, used to provide centimeter-level positioning accuracy; The time synchronization includes using a unified timestamp reference to interpolate or match camera image frames, IMU measurements, and GNSS observation data to ensure the correspondence between multi-source data at the same time. The coordinate system alignment includes establishing the extrinsic parameter calibration relationship between the camera coordinate system, IMU coordinate system, and GNSS coordinate system, unifying them to an inertial coordinate system or navigation coordinate system. The data preprocessing includes zero-bias calibration and noise filtering of IMU data, distortion correction and feature extraction of camera image data, and differential correction and outlier removal of GNSS-RTK data to improve the stability and accuracy of fusion optimization.

4. The GNSS / VIO fusion high-precision positioning method based on graph optimization according to claim 2, characterized in that, Step three specifically involves feature extraction from the monocular image, inter-frame matching to obtain camera motion constraints, and pre-integration of the IMU data to obtain the preliminary pose estimation method for the camera in the local coordinate system. The feature extraction method uses the Tomasi corner detection algorithm to extract image feature points and combines it with KLT sparse optical flow to achieve feature point tracking. To address the difficulty in extraction caused by excessively low or high lighting conditions, adaptive histogram equalization is introduced to enhance contrast. During the tracking process, the fundamental matrix and RANSAC method are used to remove false matches and outliers. When the number of effective feature points is insufficient to meet a set threshold, corner point extraction is performed again to supplement them. A feature point spacing threshold is set to ensure uniform distribution of features in the image, and a priority is set based on the frequency of feature point occurrences, prioritizing the retention of feature points with longer durations. The IMU pre-integration process first performs time alignment and noise filtering on the original measurement values, and corrects the accelerometer and gyroscope biases; then, between two keyframes, the continuous acceleration and angular velocity observations are discretely integrated to accumulate the pre-integrated quantities, including displacement increment, velocity increment and attitude increment. The method involves introducing noise and the Jacobian matrix of the bias during the integration process to establish an error state propagation model. Finally, the pre-integration results are used to establish a constraint relationship with the keyframe state variables and are added as an IMU factor in graph optimization, thereby achieving effective constraints on the relative motion between adjacent keyframes.

5. The GNSS / VIO fusion high-precision positioning method based on graph optimization according to claim 3, characterized in that, Step four specifically involves: establishing a graph optimization fusion framework, combining visual, IMU, and GNSS-RTK data to establish a visual reprojection factor, an IMU pre-integration factor, and a GNSS-RTK factor.

6. The visual reprojection factor is constructed based on the error between the observed and estimated positions of camera image feature points to optimize the camera pose and the trajectory of 3D points; the IMU pre-integration factor is constructed based on the pre-integration results of IMU acceleration and angular velocity data to construct motion constraints between adjacent frames; the GNSS-RTK factor is constructed based on the matching error between GNSS-RTK observation data and camera pose to provide absolute positioning constraints.

7. The GNSS / VIO fusion high-precision positioning method based on graph optimization according to claim 4, characterized in that, Step five specifically involves: performing graph optimization using a sliding window approach. During the optimization process, VIO provides high-frequency local constraints, and GNSS-RTK provides low-frequency global constraints. The final output includes optimized camera / carrier pose, velocity, and IMU bias, among other state variables. This sliding window method reduces computational complexity and improves real-time performance by setting a fixed window size and performing optimization calculations only on the current timeframe and a few historical frames. The VIO high-frequency local constraint provides relative motion constraints between the camera and the IMU in a short time through the IMU pre-integration factor and the visual reprojection factor. The aforementioned GNSS-RTK low-frequency global constraint provides location information constraints over a longer time scale within a window through the GNSS-RTK factor, ensuring global consistency. The method described above minimizes global error by jointly optimizing multiple constraint factors, thereby obtaining accurate optimization results for state variables such as camera / carrier pose, velocity, and IMU bias.

8. The GNSS / VIO fusion high-precision positioning method based on graph optimization according to claim 5, characterized in that, Step six specifically involves comparing the optimized pose results with the ground truth to evaluate the positioning accuracy and robustness, and verifying the algorithm's continuous positioning capability in a GNSS signal obstruction environment.

Images