Visual inertia SLAM positioning method and system in low-texture dynamic environment

Abstract

The invention provides a visual inertia SLAM positioning method and system in a low-texture dynamic environment, and belongs to the technical field of robot positioning. The method comprises the following steps: acquiring binocular vision image data and inertial measurement data of an external environment, and performing preprocessing; grid division is carried out on each frame of image after preprocessing, and feature points are extracted; tracking features among continuous frames by adopting a feature point error detection method based on FPG, establishing a tracking relationship between feature points of a current frame and a previous frame, and identifying a mistracked feature point pair; and based on the tracking relationship between the feature points, judging the motion state of the feature points by adopting a two-stage dynamic feature recognition method, and recognizing the dynamic feature points. According to the method, the problem that feature points of a traditional feature extraction algorithm in a low-texture area are sparse or are not uniformly distributed is solved, and the feature extraction success rate of a system in a weak-texture environment is improved. Feature point tracking errors are reduced, and the accuracy of feature tracking is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of robot positioning, and particularly relates to a visual-inertial SLAM positioning method and system in a low-texture dynamic environment. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] Simultaneous Localization and Mapping (SLAM) plays a key role in enabling robots and autonomous systems to navigate and understand their surroundings in real-time, and is indispensable in applications such as autonomous driving. Visual-Inertial SLAM (VINS) has received extensive attention due to its ability to fuse camera and inertial measurement unit (IMU) measurement data for stable state estimation. However, ensuring accurate and stable positioning in dynamic and low-texture environments remains a major challenge, especially in scenarios where moving objects and low-texture surfaces coexist.

[0004] To address the challenges posed by dynamic objects, existing dynamic SLAM methods can be broadly divided into semantic segmentation-based methods and geometric constraint-based methods. While the above methods have improved the performance of SLAM systems in dynamic environments to some extent, semantic segmentation-based methods require pre-trained networks, resulting in high computational costs and domain dependence. Geometric constraint-based methods often rely on a single geometric constraint, which has certain limitations.

[0005] In addition to dynamic objects, the low-texture characteristics of scenes pose additional challenges for autonomous driving. Existing methods such as point-line-based visual-inertial odometry (VIO) systems integrate innovative feature fusion and prediction matching mechanisms to ensure reliable long-line tracking. However, these methods often rely on point-line feature fusion or depth-based methods, failing to adequately balance feature extraction and feature matching. The use of point-line feature fusion can result in uneven distribution of feature points or insufficient feature recognition, increasing the risk of feature matching errors and thus affecting the positioning accuracy of the algorithm. Although depth features improve matching accuracy, their high computational cost makes them difficult to implement in real-time. SUMMARY

[0006] To overcome the shortcomings of the above existing technologies, the present application provides a visual-inertial SLAM positioning method and system in a low-texture dynamic environment, which utilizes grid-based feature extraction technology and feature point group (FPG)-based stable tracking methods to overcome the challenges of extracting and tracking feature points in low-texture scenes. In addition, dynamic feature recognition is achieved by using epipolar constraint and inter-frame distance consistency constraint in two stages, effectively solving the problem of dynamic target recognition in complex environments and enhancing the robustness of the system in complex dynamic environments.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a visual-inertial SLAM localization method for low-texture dynamic environments; A visual-inertial SLAM localization method for low-texture dynamic environments includes: Acquire binocular visual image data and inertial measurement data of the external environment, and perform preprocessing; Each preprocessed image frame is divided into grids, and feature points are extracted. An FPG-based feature point error detection method is used to track features between consecutive frames, establish the tracking relationship between feature points of the current frame and the previous frame, and identify mistracked feature point pairs. Based on the tracking relationship between feature points, a two-stage dynamic feature recognition method is used to determine the motion state of feature points and identify dynamic feature points. Based on the retained static feature points and inertial measurement data, the device pose is calculated, and the device pose is used as the final positioning result.

[0008] As a further technical solution, the acquisition of binocular visual image data and inertial measurement data of the external environment, and the preprocessing thereof, includes: The environment is captured simultaneously by two cameras, left and right, to form a binocular image pair; The acquired binocular image pairs are subjected to image denoising and correction.

[0009] As a further technical solution, each frame of the preprocessed image is divided into grids, and feature points are extracted, including: Candidate feature points are searched and extracted using the FAST wrapper function on image pyramids at different levels; corner responses are preserved and edge responses are removed using Harris corner detection. The normalized image is used to describe the orientation of the feature points; the orientation of the feature points is used to generate the BRIEF feature point description vector.

[0010] As a further technical solution, the process of using an FPG-based feature point error detection method to track features between consecutive frames, establish the tracking relationship between feature points of the current frame and the previous frame, and identify mistracked feature point pairs includes: Calculate the FPG value of the target point in the feature point group of the current frame and the FPG value of the target point in the feature point group of the previous frame corresponding to the current frame; Calculate the average FPG value of all feature points within the feature point group in the current frame and the previous frame respectively: The difference between the FPG value of the target point in the current frame and the FPG value of the target point in the previous frame is calculated, and the difference between the average FPG value of the current frame and the average FPG value of the previous frame is calculated. When two differences satisfy the following relationship, the feature point is determined to be a mistracking point;

[0011] In the formula, The FPG value of the target point in the current frame; The FPG value of the target point in the previous frame; This is the average FPG value of all points within the feature point group in the current frame; It is the average FPG value of all points in the feature point group in the previous frame; This is an adjustable parameter.

[0012] As a further technical solution, the two-stage dynamic feature recognition method includes: In the first stage, the distance between the feature point and the epipolar line is calculated through epipolar constraints to screen the first type of dynamic feature points; In the second stage, by constraining the consistency of inter-frame correlation distance and adaptively adjusting the threshold based on IMU motion data, a second type of dynamic feature points are selected.

[0013] As a further technical solution, the distance between the feature point and the epipolar line calculated by epipolar constraint is as follows:

[0014] In the formula, This is the distance from the midpoint of the epipolar constraint to the epipolar line; , , represents the parameters of the epipolar line in the image q; The basic matrix; and Let be the pixel coordinates of the i-th point in image q; Let be the second coordinate of a point in image q; Let be the homogeneous coordinates of a point in image p.

[0015] As a further technical solution, the step of filtering the second type of dynamic feature points by combining inter-frame correlation distance consistency constraints with IMU motion data adaptively adjusting the threshold includes: The inter-frame association distance of matched feature points between two frames is:

[0016] In the formula, Inter-frame correlation distance; and It is the first in the previous frame The position coordinates of each point and It is the first in the current frame The position coordinates of each point; The adaptive threshold adjustment is confirmed based on the theoretical displacement and safety factor of static feature points on the image. Dynamic feature points are identified by combining inter-frame correlation distance and adaptive threshold adjustment.

[0017] A second aspect of the present invention provides a visual inertial SLAM system for low-texture dynamic environments.

[0018] A visual-inertial SLAM system for low-texture dynamic environments includes: The image data acquisition module is configured to acquire binocular visual image data and inertial measurement data of the external environment, and perform preprocessing. The feature point extraction module is configured to perform grid division on each preprocessed frame of the image and extract feature points. The mistracked feature point identification module is configured to: use an FPG-based feature point error detection method to track features between consecutive frames, establish the tracking relationship between the feature points of the current frame and the previous frame, and identify mistracked feature point pairs; The dynamic feature point recognition module is configured to: determine the motion state of feature points and identify dynamic feature points based on the tracking relationship between feature points using a two-stage dynamic feature recognition method.

[0019] The pose estimation and localization module is configured to: calculate the device pose based on the retained static feature points and inertial measurement data, and use the device pose as the final localization result.

[0020] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a visual-inertial SLAM localization method in a low-texture dynamic environment as described in the first aspect of the present invention.

[0021] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a visual-inertial SLAM localization method in a low-texture dynamic environment as described in the first aspect of the present invention.

[0022] The above one or more technical solutions have the following beneficial effects: (1) This invention employs a gridded feature extraction method, dividing the image into a uniform 64×48 grid and forcibly extracting a specified number of feature points within each grid, thus ensuring a balanced spatial distribution of feature points. This processing method effectively overcomes the problem of sparse or unevenly distributed feature points in low-texture areas (such as white walls, monochrome ground, etc.) in traditional feature extraction algorithms, and improves the feature extraction success rate of the system in weak-texture environments.

[0023] (2) The present invention adopts a feature point error detection method based on FPG to calculate the relative positional relationship (FPG value) between feature points in the same grid. For the corresponding feature point groups in the current frame and the previous frame, their FPG values ​​are calculated and the differences are compared. This can effectively identify mismatched feature point pairs, accurately track features between consecutive frames, minimize feature point tracking errors, and improve the accuracy of feature tracking.

[0024] (3) This invention achieves efficient identification and elimination of dynamic features by employing a two-stage dynamic feature recognition method that combines epipolar constraints and inter-frame correlation distance dual geometric constraints. The first stage uses epipolar constraints, effectively identifying most dynamic features by calculating the projection distance between feature points and the epipolar line. For dynamic features moving along the epipolar direction that are difficult to detect by epipolar constraints, the second stage introduces inter-frame correlation distance consistency constraints, adaptively adjusting the discrimination threshold in conjunction with IMU motion data. The dual constraint mechanism improves the overall recognition accuracy of dynamic features.

[0025] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0027] Figure 1 This is a flowchart of the method in the first embodiment.

[0028] Figure 2 This is a schematic diagram of the feature point tracking process between adjacent frames in the first embodiment.

[0029] Figure 3 This is a schematic diagram of polar line constraints in the first embodiment.

[0030] Figure 4 This is a schematic diagram of the object moving along the polar line in the first embodiment.

[0031] Figure 5This is a comparison of inter-frame correlation distances under different camera motion states in the first embodiment.

[0032] Figure 6 In the first embodiment, the trajectory of this method is compared with that of other methods in the VIODE dataset.

[0033] Figure 7 This is a system structure diagram of the second embodiment. Detailed Implementation

[0034] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0035] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0036] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0037] Example 1 This embodiment discloses a visual inertial SLAM localization method in a low-texture dynamic environment; like Figure 1 As shown, a visual-inertial SLAM localization method for low-texture dynamic environments includes: Step S1: Acquire binocular visual image data and inertial measurement data of the external environment, and perform preprocessing. Step S2: Divide each preprocessed frame image into a grid and extract feature points; Step S3: The feature point error detection method based on FPG is used to track features between consecutive frames, establish the tracking relationship between the feature points of the current frame and the previous frame, and identify the mistracked feature point pairs. Step S4: Based on the tracking relationship between feature points, a two-stage dynamic feature recognition method is used to determine the motion state of the feature points and identify dynamic feature points. Step S5: Based on the retained static feature points and inertial measurement data, calculate the device pose and use the device pose as the final positioning result.

[0038] To more clearly illustrate this embodiment, a visual-inertial SLAM localization method in a low-texture dynamic environment can be specifically described as follows: In step S1, for binocular vision image data, environmental images are simultaneously acquired by two cameras (left and right) to form a binocular image pair, and the acquired binocular image pair is subjected to image denoising and correction.

[0039] For inertial measurement (IMU) data, it is first aligned with the image data to ensure time synchronization. Furthermore, raw IMU measurements often contain static biases, so initial optimization estimation is required to remove these biases. Additionally, preliminary filtering of the IMU data is performed to eliminate potential outliers or abrupt changes.

[0040] Since extracting a sufficient number of feature points in scenes with weak texture or sparse features is challenging, this embodiment employs a grid-based feature extraction method. Specifically, in step S2, each frame of the image is divided into several grids, and feature points are extracted as evenly as possible within each grid to ensure a balanced spatial distribution of feature points, thereby guaranteeing the stability of feature point extraction. Specifically, the FAST encapsulation function is used to extract candidate feature points. To overcome the edge response that may be generated by FAST feature points, Harris corner detection is used to retain corner responses and discard edge responses.

[0041] The FAST encapsulation function and Harris corner detection are used to search for candidate feature points on image pyramids of different levels. Normalized images are used to describe the orientation of feature points, and BRIEF feature point description vectors are generated using the feature point orientations.

[0042] In step S3, after a sufficient number of feature points are extracted, the features must be accurately tracked between consecutive frames to improve the accuracy of relative pose transformation estimation. To minimize feature point tracking errors, this embodiment employs an FPG-based feature point error detection method to track features between consecutive frames, establish the tracking relationship between feature points in the current frame and the previous frame, and identify mistracked feature point pairs.

[0043] Specifically, based on the pixels of the input image, the current frame is divided into... Each grid contains 100 pixels. Based on the feature points identified from this grid, a [structure / process] is formed. Feature point group FPG. For example... Figure 2 As shown, optical flow is used to separate points within the same feature point group (FPG) in the current frame. Track the corresponding point in the previous frame By analyzing the relative positional relationship between a feature point and other feature points within the same feature point group (FPG), mistracked feature point pairs can be identified.

[0044] The specific steps of using an FPG-based feature point error detection method to track features between consecutive frames, establish the tracking relationship between feature points in the current frame and the previous frame, and identify mistracked feature point pairs are as follows: First, calculate the target point in the previous frame. and The coordinates of the target point are multiplied by the outer product of the coordinates of all other feature points within the same feature point group (FPG). These products are then summed and averaged; this result is defined as the FPG value, which represents the relative positional relationship between the target point and other feature points within its FPG, as shown in the following formula:

[0045] in, The FPG value of the target point in the current frame; and Representing the target points respectively x and y coordinates and Let x and y represent the x and y coordinates of other points within the FPG, respectively, and let n represent the number of feature points in the feature point group. From the above formula, we can obtain... The FPG value is obtained to obtain The relative positional relationship between it and other feature points in its feature point group.

[0046] Furthermore, calculate the values ​​in the previous frame that are related to... Corresponding feature points The FPG values ​​are shown below:

[0047] In the formula, The FPG value of the target point in the previous frame; and Representing the target points respectively x and y coordinates and These represent the x and y coordinates of other points within the FPG in the previous frame, respectively, and the variable n represents the number of feature points in the feature point group.

[0048] Through calculation and The tracking relationships between feature points are constructed. To determine whether a feature point is being mistracked, the average FPG value of all feature points within the feature point group in the current frame and the previous frame is calculated based on the constructed tracking relationships, as shown in the following formula:

[0049]

[0050] In the formula, It is the average FPG value of all points within the feature point group in the current frame. It is the average FPG value of all points in the feature point group in the previous frame.

[0051] The FPG value of the target point in the current frame is subtracted from the FPG value of the target point in the previous frame, and the average FPG value of the current frame is subtracted from the average FPG value of the previous frame. When the two differences satisfy the following relationship, the feature point is determined to be a mistracking point.

[0052] In the formula, This is an adjustable parameter.

[0053] Using the above method, after extracting feature points using a grid-based approach and detecting mistracking based on FPG, the feature points can be tracked stably, but their motion state still cannot be determined. Therefore, in step S4, a two-stage dynamic feature recognition method is adopted, combining inter-frame correlation distance consistency and epipolar constraints, to determine the motion state of the feature points and identify dynamic feature points.

[0054] Specifically, in the first stage, for feature points that are stably tracked between two adjacent frames, This indicates the camera's position at the first moment. This indicates the camera's position at the second moment. Indicates the feature points being tracked. The projection point on the second frame image plane must be located on the epipolar line. Above. (As shown) Figure 3 The polar constraint is illustrated in Figure (a). The polar constraint is expressed by the following formula:

[0055] In the formula, The polar constraint value of the feature point; This is the polar line corresponding to that point; For image middle Homogeneous coordinates of a point; Let i be the homogeneous coordinates of point i in image p; It is the basic matrix.

[0056] However, dynamic feature points may not satisfy epipolar constraints in some aspects. For example... Figure 3 As shown in Figure (b), It is one of the tracked feature points. Indicates the position at the first moment. This indicates the position at the second moment. exist The projection onto the polar surface is represented by a green dot, which ideally should coincide with the blue dot on the polar line; however, the green dot is significantly different from the polar line. For the polar line... The distance between the projected feature point and the epipolar line is:

[0057] In the formula, The distance between the feature point and the epipolar line; , , These are the parameters of the epipolar lines in the image; , For image midpoint The coordinates. Furthermore, due to the movement of key feature points, errors may occur in the estimation of the fundamental matrix, such as... Figure 3 The blue and green dashed lines at the bottom of Figure (b) illustrate this. This error can also serve as a constraint to distinguish dynamic objects.

[0058] Furthermore, when dynamic feature points move along the epipolar line, the epipolar constraint distance may still be within the threshold range, leading to inaccurate identification of these moving dynamic feature points. For example... Figure 4 As shown in Figure (a), the feature points are from... Move to Their projection points in the current frame all lie on the epipolar line. In this case, it is difficult to identify dynamic feature points using epipolar constraints.

[0059] In the second stage, for objects moving along the epipolar direction, the inter-frame association distance between feature points is evaluated. For feature points matched between two frames, , The inter-frame association distance of matching feature points between two frames is calculated using the following formula:

[0060] In the formula, Inter-frame association distance for the same feature point; and It is the first in the previous frame The position coordinates of each point and It is the first in the current frame The position coordinates of each point; When the object is stationary and only the camera is moving, the inter-frame correlation distance between static feature points remains consistent across consecutive frames, while the distance between dynamic feature points and static feature points differs significantly.

[0061] Figure 4 The human movement pattern shown in Appendix (b) is along the polar direction. Therefore, in a real environment, according to... Figure 4 The motion pattern shown in attached figure (b) is used to capture a set of examples, such as... Figure 5 As shown, this illustrates the consistency constraints of inter-frame association distance.

[0062] When the object is stationary and only the camera is moving, the inter-frame correlation distance between static feature points remains consistent across consecutive frames, while the distance between dynamic and static feature points differs significantly. To clearly represent the inter-frame correlation distance between static and dynamic objects under different camera motion states, we focus only on a pair of dynamic feature points located on the dynamic object (a person moving to the right). and And two pairs of static feature points located on static objects (switches) in the scene. and , and To facilitate comparison of inter-frame correlation distances, the position of the corresponding object in the current frame is projected onto the image of the previous frame (as shown in the light orange and light green shaded areas in the figure).

[0063] Based on the above analysis, it is clear that the inter-frame correlation distance between static feature points remains consistent regardless of camera movement. In contrast, the inter-frame correlation distance of dynamic feature points deviates from that of static feature points due to their own motion.

[0064] Therefore, a threshold is used in the second stage. To identify dynamic feature points, this threshold The system can adaptively adjust based on the camera's movement speed. Specifically, it first calculates the camera's linear and angular velocities using IMU data. Then, based on the camera's motion, it calculates the theoretical displacement of static feature points on the image. To reduce the impact of noise and errors, a safety factor λ is introduced. Threshold. The motion state of an object moving along the epipolar line can be determined by the consistency constraint of the inter-frame correlation distance. This is determined by both the theoretical displacement and the safety factor λ.

[0065] The proposed two-stage method can accurately determine the motion state of feature points by utilizing epipolar constraints and consistency constraints of inter-frame correlation distance, thereby effectively eliminating dynamic feature points.

[0066] In step S5, the device pose is calculated based on the retained static feature points and inertial measurement data, and the device pose is used as the final positioning result.

[0067] Specifically, based on the raw IMU data (acceleration and angular velocity) between consecutive image frames, pre-integration is performed after removing the IMU bias to obtain the relative position, velocity, and rotation between the two frames, and IMU constraint terms are constructed to provide motion priors for subsequent optimization.

[0068] Based on the static feature point tracking relationship between image frames, combined with camera intrinsic parameters and initial pose estimation, the reprojection error of each feature point is calculated; and it is constructed as a visual observation constraint term in the optimization problem.

[0069] Within the current sliding window, a nonlinear minimization problem is constructed, incorporating IMU pre-integration factors and visual reprojection factors. By solving this optimization problem, state variables such as pose, velocity, and IMU bias are jointly estimated for each frame within the sliding window. The pose of the current frame is extracted from the sliding window optimization results as the final localization result, including position and orientation in 3D space (rotation matrix or quaternion), for use in navigation, mapping, or further tasks.

[0070] Furthermore, to verify the effectiveness of the system of this invention, a comprehensive comparison was made between this invention and classic SLAM systems (such as VINS-Fusio and ORB-SLAM3) and state-of-the-art visual-inertial dynamic SLAM systems (such as Dyna-VINS and EPLVINS). Experimental results are shown in Table 1 and... Figure 6 As shown. In the quantitative experiments, the accuracy of each algorithm in estimating the trajectory was evaluated using absolute trajectory error (ATE), and the degradation rate was calculated.

[0071] As shown in Table 1, this invention achieved the best performance across nine sequences and exhibited the lowest ATE in all three high-dynamic scenes. Compared to other methods, this invention improved accuracy by up to 85.39%. This improvement is attributed to the invention's consideration of feature point extraction and matching in low-texture scenes and its robust handling of dynamic objects through composite geometric constraints, effectively mitigating the interference from low-texture and dynamic objects.

[0072] Table 1 Comparison of ATE using different methods in the VIODE dataset

[0073] Figure 6 Figures (a), (b), and (c) illustrate the trajectory estimation results of the Dyna-VINS method for dynamic sequences in urban daytime high-dynamic scenes, urban nighttime non-dynamic scenes, and dynamic sequences in parking lots, respectively. The estimated trajectories of these sequences using this method are shown below. Figure 6 As shown in Figures (d), (e), and (f), although Dyna-VINS achieved significant trajectory estimation accuracy as the best-performing comparative method, Figure 7 This method outperforms Dyna-VINS in all three sequences. Notably, in the parking lot sequence, even in dimly lit conditions with a large number of moving vehicles, this invention effectively matches features and removes dynamic features.

[0074] Example 2 This embodiment discloses a visual inertial SLAM system for low-texture dynamic environments; like Figure 7As shown, a visual-inertial SLAM system for low-texture dynamic environments includes: The image data acquisition module is configured to acquire binocular visual image data and inertial measurement data of the external environment, and perform preprocessing. The feature point extraction module is configured to perform grid division on each preprocessed frame of the image and extract feature points. The mistracked feature point identification module is configured to: use an FPG-based feature point error detection method to track features between consecutive frames, establish the tracking relationship between the feature points of the current frame and the previous frame, and identify mistracked feature point pairs; The dynamic feature point recognition module is configured to: determine the motion state of feature points and identify dynamic feature points based on the tracking relationship between feature points using a two-stage dynamic feature recognition method; The pose estimation and localization module is configured to: calculate the device pose based on the retained static feature points and inertial measurement data, and use the device pose as the final localization result.

[0075] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0076] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a visual-inertial SLAM localization method for a low-texture dynamic environment as described in Example 1.

[0077] Example 4 The purpose of this embodiment is to provide an electronic device.

[0078] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in a visual-inertial SLAM localization method for a low-texture dynamic environment as described in Embodiment 1.

[0079] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0080] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0081] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A visual-inertial SLAM localization method for low-texture dynamic environments, characterized in that, include: Acquire binocular visual image data and inertial measurement data of the external environment, and perform preprocessing; Each preprocessed image frame is divided into grids, and feature points are extracted. An FPG-based feature point error detection method is used to track features between consecutive frames, establish the tracking relationship between feature points of the current frame and the previous frame, and identify mistracked feature point pairs. Based on the tracking relationship between feature points, a two-stage dynamic feature recognition method is used to determine the motion state of feature points and identify dynamic feature points. Based on the retained static feature points and inertial measurement data, the device pose is calculated, and the device pose is used as the final positioning result.

2. The visual-inertial SLAM localization method in a low-texture dynamic environment as described in claim 1, characterized in that, The acquisition of binocular visual image data and inertial measurement data of the external environment, and the preprocessing thereof, include: The environment is captured simultaneously by two cameras, left and right, to form a binocular image pair; The acquired binocular image pairs are subjected to image denoising and correction.

3. The visual-inertial SLAM localization method in a low-texture dynamic environment as described in claim 1, characterized in that, Each preprocessed frame of the image is divided into a grid, and feature points are extracted, including: Candidate feature points are searched and extracted using the FAST wrapper function on image pyramids at different levels; corner responses are preserved and edge responses are removed using Harris corner detection. The normalized image is used to describe the orientation of the feature points; the orientation of the feature points is used to generate the BRIEF feature point description vector.

4. The visual-inertial SLAM localization method in a low-texture dynamic environment as described in claim 1, characterized in that, The process of using an FPG-based feature point error detection method to track features between consecutive frames, establish the tracking relationship between feature points in the current frame and the previous frame, and identify mistracked feature point pairs includes: Calculate the FPG value of the target point in the feature point group of the current frame and the FPG value of the target point in the feature point group of the previous frame corresponding to the current frame; Calculate the average FPG value of all feature points within the feature point group in the current frame and the previous frame respectively: The difference between the FPG value of the target point in the current frame and the FPG value of the target point in the previous frame is calculated, and the difference between the average FPG value of the current frame and the average FPG value of the previous frame is calculated. When two differences satisfy the following relationship, the feature point is determined to be a mistracking point; In the formula, The FPG value of the target point in the current frame; The FPG value of the target point in the previous frame; This is the average FPG value of all points within the feature point group in the current frame; It is the average FPG value of all points in the feature point group in the previous frame; This is an adjustable parameter.

5. The visual-inertial SLAM localization method for low-texture dynamic environments as described in claim 1, characterized in that, The two-stage dynamic feature recognition method includes: In the first stage, the distance between the feature point and the epipolar line is calculated through epipolar constraints to screen the first type of dynamic feature points; In the second stage, by constraining the consistency of inter-frame correlation distance and adaptively adjusting the threshold based on IMU motion data, a second type of dynamic feature points are selected.

6. The visual-inertial SLAM localization method in a low-texture dynamic environment as described in claim 5, characterized in that, The distance between the feature point and the epipolar line is calculated using epipolar constraints as follows: In the formula, This is the distance from the midpoint of the epipolar constraint to the epipolar line; , , represents the parameters of the epipolar line in the image q; The basic matrix; and Let be the pixel coordinates of the i-th point in image q; Let be the second coordinate of a point in image q; Let be the homogeneous coordinates of a point in image p.

7. The visual-inertial SLAM localization method for low-texture dynamic environments as described in claim 5, characterized in that, The process of filtering second-type dynamic feature points by combining inter-frame correlation distance consistency constraints with IMU motion data to adaptively adjust thresholds includes: The inter-frame association distance of matched feature points between two frames is: In the formula, Inter-frame correlation distance; and It is the first in the previous frame The position coordinates of each point and It is the first in the current frame The position coordinates of each point; The adaptive threshold adjustment is confirmed based on the theoretical displacement and safety factor of static feature points on the image. Dynamic feature points are identified by combining inter-frame correlation distance and adaptive threshold adjustment.

8. A visual-inertial SLAM system for low-texture dynamic environments, characterized in that: include: The image data acquisition module is configured to acquire binocular visual image data and inertial measurement data of the external environment, and perform preprocessing. The feature point extraction module is configured to perform grid division on each preprocessed frame of the image and extract feature points. The mistracked feature point identification module is configured to: use an FPG-based feature point error detection method to track features between consecutive frames, establish the tracking relationship between the feature points of the current frame and the previous frame, and identify mistracked feature point pairs; The dynamic feature point recognition module is configured to: determine the motion state of feature points and identify dynamic feature points based on the tracking relationship between feature points using a two-stage dynamic feature recognition method; The pose estimation and localization module is configured to: calculate the device pose based on the retained static feature points and inertial measurement data, and use the device pose as the final localization result.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the visual-inertial SLAM localization method in a low-texture dynamic environment as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the visual inertial SLAM localization method for low-texture dynamic environments as described in any one of claims 1-7.

Images