Visual inertial navigation fusion method and system based on improved sliding window strategy
By fusing the inertial navigation unit IMU in the ORB-SLAM2 framework and adopting an improved sliding window optimization algorithm, the problem of the monocular camera SLAM system lacking absolute scale information in complex scenarios is solved, and the accuracy and robustness of posture changes are improved.
Patent Information
- Application Number
- CN202411138725.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-12-06
AI Technical Summary
The pure visual SLAM system of a monocular camera is prone to positioning failure in complex scenarios such as low texture, repeated structure, and high dynamics, mainly due to the lack of absolute scale information.
Based on the ORB-SLAM2 framework, the inertial navigation unit IMU is integrated, and through the improved sliding window optimization algorithm, a marginalized computing strategy is adopted to tightly couple non-linear optimization of vision and inertial data to achieve joint initialization and state consistency of vision and inertia sexual alignment.
It improves the accuracy of calculating pose changes between adjacent frames, and significantly improves the positioning accuracy and robustness in complex scenarios.
Smart Images

Figure CN119085639A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of simultaneous positioning and map construction, and in particular to a visual inertial navigation fusion method and system based on an improved sliding window strategy. Background Art
[0002] Simultaneous Localization and Mapping (SLAM) is widely recognized as a key technology for robots to achieve highly intelligent and autonomous capabilities. Simply put, a mobile robot placed in an unfamiliar environment uses its various sensors to estimate its position and posture and generate a map of the scene. Purely visual SLAM systems rely entirely on camera-generated image information for both positioning and mapping. These low-cost, contactless solutions are currently widely used in consumer robotics, drone navigation, augmented reality, and other fields.
[0003] Traditional classic pure vision SLAM systems are prone to positioning failure in complex scenes such as low texture, repetitive structure, and high dynamics. One of the reasons for the above problem is that the monocular camera pure vision SLAM system lacks absolute scale information. Summary of the Invention
[0004] In order to solve the above problems, the purpose of the present invention is to provide a visual inertial navigation fusion technology based on an improved sliding window strategy to solve the problem that the monocular camera pure vision SLAM system lacks absolute scale information.
[0005] To achieve the above technical objectives, this application provides a visual inertial navigation fusion method based on an improved sliding window strategy, comprising the following steps:
[0006] Based on the ORB-SLAM2 framework, the inertial navigation unit (IMU) is integrated and the visual and inertial data are jointly initialized.
[0007] Through an improved sliding window optimization algorithm and a marginalization calculation strategy, tightly coupled nonlinear optimization is performed on the visual reprojection error, IMU measurement residual and closed-loop detection residual information about point and line features.
[0008] Preferably, in the process of fusing the inertial navigation unit IMU, a pre-integration operation is performed on the IMU.
[0009] Preferably, during the joint initialization process, the point and line feature visual information and the IMU data are loosely coupled and jointly initialized.
[0010] Preferably, during the loosely coupled joint initialization process, the monocular camera pure vision is initialized, the first frame of the image frame is selected as the reference coordinate system, and the image frame is converted to the IMU coordinate system;
[0011] Based on the conversion results and the relative pose changes between adjacent visual frames, the IMU gyroscope bias estimation objective function is constructed using the least squares method.
[0012] Based on the IMU gyroscope bias estimation objective function, the IMU displacement and velocity of adjacent key frames obtained by pre-integration operation of the IMU are used to construct state variables based on the IMU's velocity, gravity, and scale information. The visual image frame is aligned with the IMU information through the nonlinear least squares method to complete the joint initialization of vision and inertia.
[0013] Preferably, in the process of tightly coupled nonlinear optimization, the camera trajectory point and line feature information and IMU measurement information are combined into state variables, and a calculation strategy of image frame marginalization is adopted. Before entering the sliding window calculation each time, it is determined whether the current frame is a key frame. If the key frame conditions are met, it enters the calculation queue, and the previous key frame state variables are removed from the window.
[0014] Preferably, in the process of obtaining state variables, state variables are generated based on the number of key frames in the sliding window, the number of point features in the image key frames, the number of line features in the image key frames, the inverse depth of the three-dimensional space feature points, the orthogonal representation of the three-dimensional space line features, the IMU position information, velocity information, rotation information of the kth image frame in the sliding window, and the accelerometer zero drift information and the gyroscope zero drift information.
[0015] Preferably, after removing older key frame state variables from the window, edge probability calculation is used to transfer the removed variable information to the remaining variables to ensure the integrity of the information.
[0016] The present invention discloses a visual inertial navigation fusion system based on an improved sliding window strategy, comprising:
[0017] A model building module is used to integrate the inertial navigation unit (IMU) based on the ORB-SLAM2 framework and perform joint initialization of visual and inertial data;
[0018] The nonlinear optimization module is used to perform tightly coupled nonlinear optimization on the visual reprojection error, IMU measurement residual, and closed-loop detection residual information about point and line features through an improved sliding window optimization algorithm and a marginalization calculation strategy.
[0019] Preferably, the model building module is also used to perform pre-integration operations on the IMU and loosely couple the point and line feature visual information with the IMU data for joint initialization.
[0020] Preferably, the nonlinear optimization module is also used to combine the camera trajectory point and line feature information and the IMU measurement information into state variables, and adopt a calculation strategy for marginalizing the image frame. Before entering the sliding window calculation each time, it will determine whether the current frame is a key frame. If the key frame conditions are met, it will enter the calculation queue, and at the same time, the previous key frame state variables will be removed from the window; and the edge probability calculation will be used to pass the variable information to the remaining variables to ensure the integrity of the information.
[0021] The present invention discloses the following technical effects:
[0022] The present invention has higher accuracy in calculating posture changes between adjacent frames. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0024] Figure 1 It is a schematic diagram of the algorithm framework of the present invention;
[0025] Figure 2 This is a schematic diagram of the vision and IMU data calculations described in the present invention;
[0026] Figure 3 This is a schematic diagram of the sliding window backend optimization of the present invention;
[0027] Figure 4 It is the state factor graph of the SLAM system described in the present invention. DETAILED DESCRIPTION
[0028] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.
[0029] like Figure 1-4 As shown, the present invention provides a visual inertial navigation fusion technology based on an improved sliding window strategy, integrates the inertial navigation unit IMU on the basis of the ORB-SLAM2 framework, and proposes a monocular vision and inertial navigation fusion algorithm based on point and line features. First, the IMU is pre-integrated to complete the joint initialization of visual and inertial data. Secondly, an improved sliding window optimization algorithm is proposed to achieve tightly coupled nonlinear optimization of visual reprojection error, IMU measurement residual, and closed-loop detection residual information about point and line features. Finally, the EuRoC open source dataset is used to conduct algorithm experiments, and the absolute pose error and relative pose error are compared with the ORB-SLAM2 algorithm, successfully verifying the effectiveness of the method proposed by the present invention, which specifically includes the following processes:
[0030] Based on the ORB-SLAM2 framework, this paper proposes a monocular vision-inertial navigation fusion algorithm that introduces point and line features. The algorithm framework is as follows: Figure 1 As shown. The algorithm framework of the present invention mainly includes input data, joint initialization of vision and IMU, and back-end optimization. First, the system accepts input data of images and IMU. The image data contains point features and line features. Since the frequency of the visual sensor and the inertial navigation unit are inconsistent, the IMU needs to be pre-integrated. Secondly, since the monocular camera and the inertial navigation unit need to be aligned through consistency before they can be fused, the point and line feature visual information and the IMU data are loosely coupled and jointly initialized. Finally, an improved sliding window back-end optimization algorithm is designed. By adopting a marginalized computing strategy, that is, only maintaining a fixed-size time window, the system only performs optimization calculations on the key frames and landmarks within the window, thereby controlling the calculation scale and avoiding the problem of decreased computing efficiency caused by the increasing number of key frames over time.
[0031] 1. When achieving tightly coupled fusion of vision and inertial navigation units, it is necessary to address the noise characteristics of the IMU. Pre-integration can significantly reduce the cumulative impact of IMU measurement noise between visual frames, improving fusion efficiency and accuracy. By integrating the IMU measurement data between two visual observations and taking the noise model into account, a concise and easy-to-process intermediate state representation is generated, effectively reducing the amount of computation during the optimization process and improving system robustness.
[0032] IMU can measure the rotational angular velocity and acceleration information of the current sensor coordinate system in three directions. Since the acquisition frequency of IMU is much higher than that of camera, if the IMU is tightly coupled with the camera's observation data, the IMU data needs to be integrated to align with the camera's sampling frequency. Figure 2 shown.
[0033] The IMU measurement values between the k+1th and kth time frames are processed to infer the position, speed, and rotation information of the mobile robot at the next moment, as described below:
[0034]
[0035] Where, Indicates location information. represents speed information, Δt represents time interval, represents the estimated value of the accelerometer measurement, Indicates rotation information, represents the estimated value of the gyroscope measurement, and EXP() represents the natural exponential function.
[0036] The following is the differential equation form of the kinematic model:
[0037]
[0038] Between two adjacent image frames, the quaternion multiplication operator is By integrating the IMU measurement information at the k+1th and kth moments, the continuous form can be obtained as shown in (3).
[0039]
[0040] Where g w represents the acceleration due to gravity, Indicates the accelerometer measurement bias, Indicates the gyroscope measurement bias;
[0041] It can be seen that the IMU measurement state information is affected by the data of the previous moment and before. When the value of the first frame changes, the integral needs to be recalculated, which leads to a serious increase in the amount of data calculation and cannot meet the real-time requirements. Therefore, the present invention adopts the IMU pre-integration operation, so that the pose estimation between two adjacent frames is only related to the data at different moments, and does not depend on the initial value of the previous frame, so as to avoid the state value of the previous frame affecting the integral calculation. By multiplying and simplifying, we can get the increment of position information, velocity information, and rotation information between two image frames:
[0042]
[0043] Therefore, the pre-integration term can be organized into the following form:
[0044]
[0045] Through IMU pre-integration, the relative pose transformation between two image frames is obtained, allowing constraints to be established and optimized. Even if the initial state or deviation changes, there is no need to recalculate the entire integration process; only the original state can be updated, greatly improving computational efficiency and meeting real-time performance requirements.
[0046] 2. Since the monocular camera and the inertial navigation unit need to be initialized before they can be integrated, the present invention adopts a loosely coupled approach to perform visual-inertial joint initialization.
[0047] First, initialize the pure vision of the monocular camera and select the first frame of the image frame as the reference coordinate system. The conversion from the image frame to the IMU coordinate system is:
[0048]
[0049] Among them, the parameters marked with c are in the camera coordinate system, and the parameters marked with b are in the IMU coordinate system.
[0050] The initial moment of the camera coordinate system is the parameter marked with c0, and the initial moment of the IMU coordinate system is the parameter marked with b0. and are the translation and rotation from the camera to the IMU coordinate system, respectively, and s is the scale information factor.
[0051] According to the relative pose changes between adjacent visual frames, the IMU gyroscope bias estimation objective function is constructed, which is expressed by the least squares method:
[0052]
[0053] in, This is the rotation information obtained during IMU pre-integration, and B is represented by the set of key frames in the sliding window.
[0054] in, is the rotation information obtained during the previous IMU pre-integration, and B is the set of key frames in the sliding window.
[0055] The state variable constructed by the IMU's velocity, gravity, and scale information is:
[0056]
[0057] Where, represents the gravity, and s represents the scale information.
[0058] The IMU displacement of adjacent key frames can be obtained through IMU pre-integration and speed
[0059]
[0060] Finally, the least squares problem can be constructed:
[0061] in,
[0062]
[0063] Where I represents the identity matrix.
[0064] By solving the nonlinear least squares problem, we can obtain the IMU's velocity information and the magnitude and direction of gravity. On this basis, we align the gravity information with the IMU information, thus completing the joint initialization of vision and inertia.
[0065] 3. As the system's visual information of point and line features and IMU data continues to increase, the amount of computation increases, leading to reduced real-time performance. Therefore, this invention adopts a sliding window backend optimization algorithm. Only a fixed-size time window is maintained, and a marginalized computing strategy is adopted. The system only performs optimization calculations on the keyframes and landmarks within this window, thereby controlling the computational scale and avoiding the problem of decreased computational efficiency caused by the increasing number of keyframes over time. By dividing the time window into multiple subwindows and performing calculations within each subwindow, the real-time performance of the system is guaranteed.
[0066] The sliding window operation process is as follows Figure 3 First, the camera trajectory point and line feature information and IMU measurement information are combined into the state variable χ:
[0067]
[0068] Among them, n k is the number of key frames in the sliding window, np is the number of point features in the image keyframe, n l is the number of line features in the image keyframe; λ represents the inverse depth of the three-dimensional space feature point, y represents the orthogonal representation of the three-dimensional space line feature; x k It contains the IMU position information, velocity information, rotation information, accelerometer zero drift information and gyroscope zero drift information of the kth image frame in the sliding window.
[0069] After constructing the state variables, it is necessary to adopt a calculation strategy for marginalizing the image frames. Before entering the sliding window calculation each time, it will determine whether the current frame is a key frame. If the key frame conditions are met, it will enter the calculation queue. At the same time, the previous key frame state variables are removed from the window. This process removes no longer important variables from the state vector of the optimization problem, thereby controlling the overall calculation scale of the system.
[0070] The present invention uses marginal probability calculation to pass the information of the eliminated variables to the remaining variables, thereby ensuring the integrity of the information. Assuming that △x is marginalized, the following incremental linear equation is obtained:
[0071] HΔx=b (14)
[0072] After expanding formula (14), it has the following form:
[0073]
[0074] Since it is difficult to solve the inverse matrix, we can only solve the inverse matrix for the diagonal matrix. Therefore, we use the Schur elimination method to operate the above formula to obtain:
[0075]
[0076] After finishing, we can get:
[0077]
[0078] The new incremental equation can be obtained from the above formula:
[0079]
[0080] From formula (18), we can know that the equation Δx a The parameter information of Δx is retained. b Parameters are related. A factor graph within the sliding window can be constructed as Figure 4 As shown. Therefore, the system state estimation objective function can be written as:
[0081]
[0082] The residual information is shown in formula (20).
[0083]
[0084] In the objective function, the prior residual information obtained after marginalization is r p And the Jacobian matrix is J p ; ρ is a robust kernel function used to achieve outlier suppression; r1 represents the pre-integration residual information of the inertial navigation unit IMU, r2 represents the visual reprojection residual information from point features, r3 represents the visual reprojection residual information from line segment features, and r4 represents the residual information from closed-loop detection.
[0085] Finally, the objective function can be solved in a nonlinear way, and the optimal estimate of the system state can be obtained.
[0086] For the visual-inertial fusion SLAM algorithm experiment:
[0087] The EuRoC dataset classifies the difficulty of the dataset according to the complexity of its scene and the trajectory situation, and is divided into three levels: simple, medium and difficult. The experiment of the present invention selects 7 different levels of sequences in the EuRoC dataset, including MH_01_easy, MH_03_medium, V1_01_easy, V1_02_medium, V1_03_difficult, V2_01_easy, and V2_02_medium, and runs the classic open source framework ORB-SLAM2 algorithm and the algorithm of the present invention respectively. Each dataset sequence is tested three times and the average value is recorded. The EVO evaluation and analysis tool is used to compare the accuracy of the running results with the true trajectory. The visual inertial system formed by the algorithm of the present invention on the basis of the fusion of point and line features makes the pose trajectory always consistent with the true value trajectory, especially in complex scenes such as smooth walls and weak textures, with better robustness performance.
[0088] Next, we further analyze the errors in detail and obtain the quantitative comparison results of the absolute pose error APE and the relative pose error RPE, as shown in Tables 1 and 2.
[0089] Table 1 Quantitative comparison of absolute pose errors of the proposed algorithm (unit: meter)
[0090]
[0091] Table 2 Quantitative comparison of relative pose errors of the proposed algorithm (unit: meter)
[0092]
[0093] Table 3 shows the parameters and improvement percentages of the absolute pose errors of the algorithm of the present invention and the ORB-SLAM2 algorithm.
[0094] Table 3 Absolute posture error improvement of the algorithm of the present invention (unit: meter)
[0095] APE ORB-SLAM2 Algorithm of the present invention Percentage increase MAX 0.1793 0.1070 40.31% Mean 0.0574 0.0432 24.62% Median 0.0529 0.0425 19.73% RMSE 0.0640 0.0470 26.64% SSE 10.6349 6.2325 41.40% SD 0.0261 0.0182 30.42%
[0096] The parameters and improvement percentages of the relative pose errors of the algorithm of the present invention and the ORB-SLAM2 algorithm are shown in Table 4.
[0097] Table 4 Relative posture error improvement of the algorithm of the present invention (unit: meter)
[0098]
[0099]
[0100] According to the above experimental data, from the perspective of absolute pose error, in the simple and medium-level environments of the EuRoC data set, the absolute pose error APE of the algorithm designed by the present invention can achieve an improvement of about 25% to 41% compared to ORB-SLAM2, among which RMSE and SD in the medium-difficulty environment increased by 34.64% and 25.31% respectively. The local fluctuation in the difficult and complex environment is slightly larger, and the performance of several absolute error indicators is slightly lower than that of ORB-SLAM2. This is because while emphasizing accuracy, the robustness is reduced, resulting in the inability to successfully track in some more extreme cases. In addition, from the perspective of relative pose error RPE, all indicators are better than ORB-SLAM2, and can achieve an improvement of about 70% to 90%. The RMSE and SD of the three levels of environment are increased by 78% and 74.48% respectively, indicating that the algorithm of the present invention is more accurate in calculating the posture changes between adjacent frames.
[0101] This paper first addresses the shortcoming of a monocular camera-based pure vision SLAM system lacking absolute scale information. By integrating an inertial navigation unit (IMU) with the ORB-SLAM2 framework, this paper proposes a monocular vision and inertial navigation fusion algorithm based on point and line features. Secondly, due to the inconsistent frame rates of vision and IMU, the IMU is pre-integrated and then initialized together with the visual-inertial system. Simultaneously, this paper proposes an improved sliding window optimization algorithm to achieve efficient tightly coupled nonlinear optimization of the visual reprojection error, IMU measurement residual, and closed-loop detection residual information regarding point and line features. Finally, the algorithm of this paper is experimented with the EuRoC open source dataset, and the absolute pose error and relative pose error are compared with the ORB-SLAM2 algorithm. The algorithm designed in this paper achieves an improvement of approximately 25% to 41% in terms of the absolute pose error (APE) compared to ORB-SLAM2, and an improvement of approximately 70% to 90% in terms of the relative pose error (RPE), thus successfully verifying the effectiveness of the algorithm proposed in this paper.
[0102] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0103] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.
[0104] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.
Claims
1. A visual inertial navigation fusion method based on an improved sliding window strategy, It is characterized in that The following steps are involved: Based on the ORB-SLAM2 framework, the inertial navigation unit IMU is integrated and the visual and inertial data are jointly initialized; Through an improved sliding window optimization algorithm and a marginalization calculation strategy, tightly coupled nonlinear optimization is performed on the visual reprojection error, IMU measurement residual and closed-loop detection residual information about point and line features.
2. According to the visual inertial navigation fusion method based on the improved sliding window strategy of claim 1, Features: In the process of fusing the inertial navigation unit IMU, a pre-integration operation is performed on the IMU.
3. According to the visual inertial navigation fusion method based on the improved sliding window strategy of claim 2, Features: During the joint initialization process, the point and line feature visual information is loosely coupled and jointly initialized with the IMU data.
4. According to claim 3, a visual inertial navigation fusion method based on an improved sliding window strategy, Features: In the process of loosely coupled joint initialization, the pure vision of the monocular camera is initialized, the first frame of the image frame is selected as the reference coordinate system, and the image frame is converted to the IMU coordinate system; Based on the conversion results, according to the relative posture changes between adjacent visual frames, the IMU gyroscope bias estimation objective function is constructed according to the least squares method; Based on the IMU gyroscope bias estimation objective function, according to the IMU displacement and velocity of adjacent key frames obtained by pre-integration operation on the IMU, and according to the state variables jointly constructed according to the velocity, gravity and scale information of the IMU, the visual image frame is aligned with the IMU information through the nonlinear least squares method to complete the joint initialization of vision and inertia.
5. According to the visual inertial navigation fusion method based on the improved sliding window strategy of claim 4, Features: In the process of tightly coupled nonlinear optimization, the camera trajectory point and line feature information and IMU measurement information are combined into state variables. The image frame marginalization calculation strategy is adopted. Before entering the sliding window calculation each time, it is judged whether the current frame is a key frame. If the key frame conditions are met, it enters the calculation queue and the previous key frame state variables are removed from the window.
6. According to claim 5, a visual inertial navigation fusion method based on an improved sliding window strategy, Features: In the process of acquiring state variables, the state variables are generated according to the number of key frames in the sliding window, the number of point features in the image key frames, the number of line features in the image key frames, the inverse depth of the three-dimensional space feature points, the orthogonal representation of the three-dimensional space line features, the IMU position information, speed information, rotation information of the kth image frame in the sliding window, and the accelerometer zero drift information and the gyroscope zero drift information.
7. According to claim 6, a visual inertial navigation fusion method based on an improved sliding window strategy, Features: After removing the older key frame state variables from the window, the edge probability calculation is used to pass the removed variable information to the remaining variables to ensure the integrity of the information.
8. A visual inertial navigation fusion system based on an improved sliding window strategy. It is characterized in that include: The model building module is used to integrate the inertial navigation unit IMU based on the ORB-SLAM2 framework and perform joint initialization of visual and inertial data; The nonlinear optimization module is used to perform tightly coupled nonlinear optimization on the visual reprojection error, IMU measurement residual and closed-loop detection residual information about point and line features through an improved sliding window optimization algorithm and a marginalization calculation strategy.
9. According to claim 8, a visual inertial navigation fusion system based on an improved sliding window strategy, Features: The model building module is also used to perform pre-integration operations on the IMU and loosely couple and jointly initialize point and line feature visual information with IMU data.
10. According to claim 9, a visual inertial navigation fusion system based on an improved sliding window strategy, Features: The nonlinear optimization module is also used to form state variables from the point and line feature information of the camera trajectory and the IMU measurement information, and adopts a calculation strategy for marginalizing the image frame. Before entering the sliding window calculation each time, it will determine whether the current frame is a key frame. If the key frame conditions are met, it will enter the calculation queue, and at the same time, the previous key frame state variables that are older will be removed from the window; and the edge probability calculation is used to pass the removed variable information to the remaining variables to ensure the integrity of the information.
Citation Information
Patent Citations
Visual inertial navigation fusion SLAM method based on Runge-Kutta4 improved pre-integration
CN112240768A
Visual inertia SLAM method based on image edge features
CN112749665A
SLAM method and system for fusing IMU and visual data based on point-line visual features
CN118424266A
