Visual slam feature matching method, system, device and readable storage medium

Abstract

The application provides a visual SLAM feature matching method, system and device and a readable storage medium. The method comprises the following steps: step 1, preprocessing binocular camera image data; step 2, pre-integrating inertial measurement unit (IMU) measurement values; step 3, error analysis on the pre-integration; step 4, predicting pixel coordinates; and step 5, determining an adaptive threshold. The adaptive algorithm provided by the application can effectively reduce the search radius of feature matching, provide regional constraints for feature points, improve the accuracy of image feature matching, provide a relatively accurate initial value of a pose for a SLAM system, and further improve the overall pose accuracy of the system.

Description

Technical Field

[0001] This invention relates to an image feature matching algorithm, and more specifically, to a visual SLAM feature matching method, system, apparatus, and readable storage medium. Background Technology

[0002] As a front-end technology of Visual SLAM (Simultaneous Localization and Mapping), the main task of visual odometry is to estimate camera motion based on matching information from neighboring images. Feature matching is the most critical step, and its accuracy directly affects the precision of the VSLAM system. Commonly used feature matching algorithms include brute-force matching and fast nearest neighbor matching. However, when the camera moves too fast, image blurring, dynamic scenes, lighting changes, and the local characteristics of image features themselves lead to widespread feature mismatches, becoming a major bottleneck restricting the performance improvement of visual SLAM.

[0003] To address the high feature mismatch rate, most scholars both domestically and internationally employ a fusion technique of VSLAM and Inertial Measurement Unit (IMU). One scholar proposed an improved ASIFT feature matching method, utilizing motion information provided by the IMU to create a virtual camera view and estimate the essential matrix, thereby enhancing the robustness of feature matching. Wu Bin and Wang Xuri established a mathematical relationship between the epipolar lines of pixels in consecutive frames using epipolar constraints, and combined this with IMU-predicted camera pose to obtain the epipolar lines of feature points, effectively eliminating mismatched feature points. Qi Guan proposed a feature matching algorithm that alternates between global and local search operations with automatic switching. It uses a camera projection model and IMU pre-integration results to predict the possible regions where feature points may appear, and performs search matching on local regions. Similarly, the visual-inertial SLAM system ORB-SLAM3, proposed by Campos C and Elvira R, also introduces inertial constraints in the feature matching stage. It uses the IMU to predict camera pose, projects map points obtained from left-right matching or triangulation of the binocular cameras onto the image, and predicts pixel positions. Inertial navigation system (INS) solutions can provide regional constraints for feature point matching and improve the accuracy of feature matching, so they have been widely used.

[0004] However, due to inherent errors in IMU measurements, in inertial navigation-assisted image feature matching, the feature matching of the entire image is typically transformed into local region feature matching with the predicted position as the origin and a fixed radius. Furthermore, to accommodate IMUs of varying accuracy, a common practice is to set a large search radius to avoid missing feature points. This can lead to wasted computational resources when the IMU accuracy is high, increasing the mismatch rate of image features and reducing the pose estimation accuracy of the SLAM system. Summary of the Invention

[0005] To address the shortcomings of the prior art, this invention provides an adaptive inertial navigation-assisted visual SLAM feature matching method and system. Based on inertial unit error, it automatically adjusts the image feature matching search radius according to the error propagation law, which not only improves the accuracy of feature point matching but also reduces the amount of computation and time required for matching, thereby improving the efficiency of feature matching.

[0006] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0007] The first aspect of this invention provides a visual SLAM feature matching method based on adaptive inertial navigation assistance, comprising the following steps;

[0008] Step 1: Preprocess the stereo camera image data;

[0009] Step 2: Pre-integrate the IMU measurements;

[0010] Step 3: Perform error analysis on the pre-integration;

[0011] Step 4: Predict pixel coordinates;

[0012] Step 5: Determine the adaptive threshold.

[0013] A second aspect of the present invention provides a visual SLAM feature matching system based on adaptive inertial navigation assistance, comprising:

[0014] The first processing module is used to preprocess the image data from the binocular camera.

[0015] The second processing module is used to pre-integrate the IMU measurement values;

[0016] The third processing module is used to perform error analysis on the pre-integration;

[0017] The fourth processing module is used to predict pixel coordinates;

[0018] The fifth processing module is used to determine the adaptive threshold;

[0019] The first processing module, the second processing module, the third processing module, the fourth processing module, and the fifth processing module are connected in sequence to complete the visual SLAM feature matching method based on adaptive inertial navigation assistance.

[0020] A third aspect of the present invention provides a visual SLAM feature matching device, comprising:

[0021] Memory; and

[0022] A processor coupled to the memory is configured to execute the adaptive inertial navigation-assisted visual SLAM feature matching method based on instructions stored in the memory.

[0023] A fourth aspect of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned adaptive inertial navigation-assisted visual SLAM feature matching method.

[0024] This invention has outstanding substantive features and significant progress compared to the prior art. Specifically, this invention has the following beneficial effects and advantages:

[0025] This invention proposes an adaptive threshold inertial measurement unit (IMU)-assisted image feature matching algorithm. Based on IMU error, it automatically adjusts the image feature matching search radius according to the error propagation law, thereby efficiently assisting visual SLAM feature matching using IMU data of varying precision. The proposed adaptive algorithm effectively reduces the feature matching search radius, provides regional constraints for feature points, improves the accuracy of image feature matching, provides more accurate initial pose values ​​for the SLAM system, and ultimately enhances the overall pose accuracy of the system. Attached Figure Description

[0026] Figure 1 Flowchart of inertial navigation-assisted image feature matching.

[0027] Figure 2 Diagram of an imaging model for a binocular camera.

[0028] Figure 3 Flowchart for determining the adaptive feature matching search radius. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings.

[0030] Explanation of the technical name:

[0031] Feature points are representative pixels in the pixel coordinate system. They are representative points selected from the image, and their corresponding coordinates are called pixel coordinates.

[0032] A map point is a three-dimensional point in space that corresponds to a feature point.

[0033] The carrier refers to the platform that carries the camera and inertial navigation system.

[0034] Example 1

[0035] like Figure 1-3 As shown, this embodiment provides a visual SLAM feature matching method based on adaptive inertial navigation assistance, including the following steps;

[0036] Step 1: Preprocess the stereo camera image data;

[0037] Methods for preprocessing stereo image data include:

[0038] Step 1-1, based on the principle of similar triangles, the depth value z of the map point in the stereo camera coordinate system is obtained as follows:

[0039]

[0040] Where f is the focal length of the camera and b is the baseline of the binocular camera;

[0041] Steps 1-2: Substitute the depth value z of the feature point in the stereo camera coordinate system into the pinhole camera projection model to obtain the coordinates P of the map point in the camera coordinate system. l for

[0042]

[0043] In the formula, (u l ,v l ) represents the pixel coordinates of map point P in the left-eye camera image plane coordinate system, and c x ,c y It's the camera's internal parameters.

[0044] Step 2: Pre-integrate the IMU measurements;

[0045] Between two consecutive image frames i and j, there are multiple frames of IMU data. The relative motion increment between the two frames can be obtained by pre-integrating the IMU measurements.

[0046] Methods for pre-integrating IMU measurements include:

[0047]

[0048]

[0049]

[0050] In the formula, ΔR ij ,Δv ij ,Δp ij These are the relative attitude, velocity, and position of two adjacent image frames i and j, obtained through pre-integration calculation. Δt represents the measurements of the gyroscope and accelerometer at time k, the zero bias, and the discrete Gaussian white noise, respectively; Δt is the sampling time interval of the inertial measurement unit (IMU); Exp(·) is the exponential mapping that maps the rotation vector to a rotation matrix in the Lie group space.

[0051] Step 3: Perform error analysis on the pre-integration;

[0052] Methods for error analysis of pre-integration include:

[0053] Step 3-1: Due to the complex dependency between the pre-integrated quantity and the IMU noise, it is difficult to directly derive the pre-integrated information matrix using the error propagation law. Therefore, the IMU noise can be separated from the pre-integrated quantity, and the relationship between the pre-integrated measurement noise and the IMU noise is obtained as follows:

[0054]

[0055]

[0056]

[0057] In the formula, The attitude error δR in the pre-integrated quantity of the inertial measurement unit (IMU) from time i to j. ij The corresponding Lie algebra; δv ij ,δp ij These represent the errors in velocity and position in the pre-integrated quantities of the inertial measurement unit (IMU) from time i to j, respectively. Let be the discrete noise of the gyroscope and accelerometer at time k, respectively, and be the product of their continuous noise and the square root of the sampling frequency, with units of rad / s and m / s, respectively. 2 ;

[0058] Step 3-2: Since the error expressions for the three pre-integral state variables are quite complex, direct calculation would waste resources on the computing platform. After derivation and simplification, the recursive formula for the pre-integral error is obtained as follows:

[0059]

[0060] In the formula, The attitude error δR in the pre-integrated quantity of the inertial measurement unit (IMU) from time i to j-1. i,j-1 The corresponding Lie algebra; δv i,j-1 ,δp i,j-1 These represent the errors in velocity and position in the pre-integrated quantity of the inertial measurement unit (IMU) from time i to j-1, respectively. The relative attitude of the carrier from time j to j-1; The accelerometer measurement at time j-1; The accelerometer bias is zero at time i, and here we define that the bias between two adjacent image frames remains unchanged. This is the inverse of the right Jacobian matrix; These are the discrete noises of the gyroscope and accelerometer at time j-1, respectively.

[0061] Step 3-3: According to the error propagation law, obtain the covariance ∑ of the IMU pre-integral quantity between two adjacent image frames. ij for:

[0062]

[0063]

[0064]

[0065]

[0066] In the formula, ∑ i,j-1 Let be the covariance of the pre-integral quantity of the inertial measurement unit (IMU) from time i to j-1. We define the initial time as when the pre-integral quantity of the IMU has no error, i.e., ∑ ii It is a zero matrix; Let be the covariance of the noise of the inertial measurement unit (IMU) measurement at time j-1. The IMU covariance at any given time is defined as a fixed value, and the noise η is calibrated by the gyroscope and accelerometer. gd ,η ad constitute.

[0067] Step 4: Predict pixel coordinates;

[0068] Methods for predicting pixel coordinates:

[0069] The rotation matrix of the carrier between two adjacent image frames is calculated using the measurements from the inertial measurement unit and the pre-integration principle. Translation vector The relative attitudes of two adjacent image frames i and j are obtained by pre-integration calculation using an inertial measurement unit (IMU). and location

[0070] Based on the carrier's pose in world coordinates at the previous image frame time and the relative pose between two adjacent images, the predicted pose of the carrier in world coordinates at the current image frame time is:

[0071]

[0072]

[0073] in, These represent the attitude, velocity, and position of the carrier in the world coordinate system at the time of the previous image frame, respectively; ΔT is the camera's sampling interval; and g is Earth's gravity.

[0074] The pose of the current image frame's carrier system in the world coordinate system is converted to the pose of the carrier system relative to the world coordinate system as follows:

[0075]

[0076]

[0077] As the camera pose predicted by pre-integration of the inertial measurement unit (IMU);

[0078] Using the camera-to-carrier frame transformation matrix calibrated by visual inertial SLAM, the predicted pose of the camera relative to the world frame at the current image frame time is:

[0079]

[0080]

[0081] Among them, R cb ,t cb These are the transformation matrix and translation vector between the camera system and the vehicle system, respectively;

[0082] When the camera model is a pinhole camera model, the pixel coordinates of the map points obtained using the stereo camera model on the current frame image are:

[0083] P c =R cw P w +t cw

[0084]

[0085] Among them, f x ,f y ,c x ,c y P is the camera's intrinsic parameter. w P represents the three-dimensional coordinates of a map point in the world coordinate system. c U represents the coordinates of the map point in the current frame's camera system, and u and v represent the corresponding pixel coordinates of the map point.

[0086] Step 5: Determine the adaptive threshold;

[0087] Methods for determining adaptive thresholds:

[0088] Step 5-1: The covariance of the relative pose between images can be obtained from the pre-integration error of the inertial measurement unit (IMU) and the error propagation law. By utilizing the principle of inertial navigation-assisted feature matching, the pixel value corresponding to each map point can be obtained, thus acquiring the mean square error of the predicted pixel position. Since the function used in the feature matching process is pose-dependent, the function is linearized from the perspective of Lie algebra during error propagation. In the process of propagating the pre-integrated relative pose error of the inertial measurement unit (IMU) to the pose error of the carrier relative to the world frame at the current image frame time, it is necessary to consider the relative attitude. and location Linearization, therefore, calculations are performed with respect to relative attitude. and location The derivative:

[0089]

[0090]

[0091] Step 5-2, calculate the covariance of the carrier's pose relative to the world frame at the current image frame time:

[0092]

[0093]

[0094] in, for The corresponding Lie algebra, for Right-multiply by the inverse of the Jacobian matrix, Obtain the covariance of relative pose through IMU pre-integration;

[0095] Step 5-3, similarly, is based on the recursive formula for the pre-integral error and the covariance ∑ of the IMU pre-integral values ​​between two adjacent image frames. ij The covariance of the camera pose relative to the world system can be obtained from the pose covariance of the current image frame. The camera pose covariance is then transmitted to the corresponding pixel of each map point using the pinhole projection model of the camera to obtain the pixel covariance.

[0096] The relationship between the adaptive feature matching search radius and the pixel covariance is as follows:

[0097]

[0098] r = 3max{σ u ,σ v}

[0099] Where σ0 is the unit weight mean error, σ u ,σ v denoted as the mean square error of pixel coordinates u and v, respectively, and r is the search radius for feature matching, taking the limit error of pixel coordinates.

[0100] Experimental verification

[0101] To verify the feasibility and effectiveness of the method described in this embodiment, experimental analysis was conducted using different scene sequences from two public datasets, EuRoC and TUM. The EuRoC dataset was collected by a miniature UAV equipped with an MT9V034 camera and an ADIS16448 inertial measurement unit, with its flight attitude captured in real-time using a laser tracker or Vicon motion capture system as the UAV's actual trajectory. The TUM dataset was collected by a mobile vehicle equipped with a uEye camera and a BMI160 inertial measurement unit, with the actual trajectory provided by the OptiTrack motion capture system.

[0102] This experiment selected seven representative sequences from two publicly available datasets: MH01, V101, V201, Corridor1, Magistrale1, Room1, and Slides1. For each sequence, the specific scene, distance, and noise levels from the accelerometer and gyroscope are provided, as shown in Table 1.

[0103] Table 1. Sequences selected in the experiment

[0104]

[0105] To verify the usability of the method in this embodiment, this experiment compared the absolute pose error of different image sequences in different datasets before and after processing the output pose using the ORB-SLAM3 system before and after the improvement, as shown in Table 2. The data here is the average of the root mean square of the absolute pose error of each sequence in 10 evaluations.

[0106] Table 2. Root mean square (m) of the absolute pose error of the overall pose for each sequence.

[0107]

[0108] As shown in Table 2, in the selected image sequences, the absolute trajectory error of the improved ORB-SLAM3 system according to the method described in this embodiment is better than that of the original algorithm, and the overall pose accuracy of the improved system is improved by an average of 24.32%. This proves that the improved ORB-SLAM3 system can run well in different scene sequences, has strong robustness, and also improves the positioning accuracy. This is because the ORB-SLAM3 system uses a nonlinear optimization method in the back end. Even though the final output pose has undergone three optimizations, it still has a strong correlation with the given initial pose. The initial pose here is provided by the tracking thread. Therefore, the method proposed in this embodiment directly improves the accuracy of the tracking thread's pose estimation, provides a more accurate initial value for the SLAM system, and thus improves the overall positioning accuracy of the SLAM system.

[0109] Example 2

[0110] This embodiment provides a visual SLAM feature matching system based on adaptive inertial navigation assistance, including:

[0111] The first processing module is used to preprocess the image data from the binocular camera.

[0112] The second processing module is used to pre-integrate the IMU measurement values;

[0113] The third processing module is used to perform error analysis on the pre-integration;

[0114] The fourth processing module is used to predict pixel coordinates;

[0115] The fifth processing module is used to determine the adaptive threshold;

[0116] The first processing module, the second processing module, the third processing module, the fourth processing module, and the fifth processing module are connected in sequence to complete the visual SLAM feature matching method based on adaptive inertial navigation assistance described in Embodiment 1.

[0117] For the specific implementation method of the system in this embodiment, please refer to the method described in Embodiment 1, which will not be repeated here.

[0118] Example 3

[0119] This embodiment provides a visual SLAM feature matching device, including:

[0120] Memory; and

[0121] A processor coupled to the memory is configured to execute the adaptive inertial navigation-assisted visual SLAM feature matching method described in Example 1 based on instructions stored in the memory.

[0122] The memory may include, for example, system memory, fixed non-volatile storage media, etc. System memory may store, for example, the operating system, application programs, the boot loader, and other programs.

[0123] The device may also include input / output interfaces, network interfaces, and storage interfaces. These interfaces, as well as the memory and processor, can be connected via, for example, a bus. The input / output interfaces provide connection interfaces for input / output devices such as monitors, mice, keyboards, and touchscreens. The network interfaces provide connection interfaces for various networked devices. The storage interfaces provide connection interfaces for external storage devices such as SD cards and USB flash drives.

[0124] Example 4

[0125] This embodiment provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the adaptive inertial navigation-assisted visual SLAM feature matching method described in Embodiment 1.

[0126] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-non-transitory readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer program code.

[0127] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0128] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0129] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive inertial navigation system aided visual SLAM feature matching method, characterized in that, Comprising the following steps; Step 1, pre-processing binocular camera image data; The method for pre-processing binocular image data comprises: Step 1-1, according to the principle of similar triangles, the depth value z of the map point in the binocular camera coordinate system is z=fb / d ， d=u l －u r wherein, f is the focal length of the camera, b is the baseline of the binocular camera; Step 1-2, the depth value z of the feature point in the binocular camera coordinate system is substituted into the pinhole camera projection model to obtain the coordinates of the map point in the camera coordinate system P l To ， In the formula, u l , v l is the pixel coordinate of the map point P in the left camera image plane coordinate system, c x , c y is the camera intrinsic parameter; Step 2, pre-integrating inertial measurement unit IMU measurement values; The method for pre-integrating inertial measurement unit IMU measurement values comprises: ， where, Δ R ij , Δ v ij , Δ p ij , are the relative pose, velocity and position of two image frames at time i, j respectively, which are obtained by pre-integration algorithm; are the measurements, biases and discrete white Gaussian noise of the gyroscope and accelerometer at time k respectively; ∆t is the sampling time interval of the inertial measurement unit (IMU); Exp(·) is the exponential mapping, which maps the rotation vector to the rotation matrix in the Lie group space. Step 3, error analysis on the pre-integration; The method for error analysis on the pre-integration comprises: Step 3-1, the relationship between the pre-integration measurement noise and the inertial measurement unit IMU noise is: ， wherein is the attitude error in the i to j moment IMU pre-integral δR ij corresponding Lie algebra; δv ij , δ p ij are the velocity and position errors in the i to j moment IMU pre-integral, respectively; η k gd , η k ad are the discrete noises of the gyro and accelerometer at the k moment, respectively, which is the product of the continuous noise and the square root of the sampling frequency, units are rad / s, m / s 2 ; Step 3-2, the recursive formula of the pre-integration quantity error is ， wherein is the inertial measurement unit IMU pre-integration quantity of the attitude error at time instant i to j-1 δR i,j-1 corresponding Lie algebra; δv i,j-1 , δp i,j-1 are the errors of the velocity and position, respectively, in the inertial measurement unit IMU pre-integration quantity at time instant i to j-1; is the relative attitude of the carrier at time instant j to j-1; is the accelerometer measurement value at time instant j-1; b i a is the accelerometer bias at time instant i, where it is defined that the bias between two adjacent image frames is constant; J r j-1 is the inverse of the right Jacobian matrix; η j-1 gd , η j-1 ad are the discrete noises of the gyroscope and accelerometer, respectively, at time instant j-1; Step 3 - 3, Covariance of IMU pre-integration between two consecutive image frames, Σ ij For ， ， wherein∑ i,j-1 is the covariance of the IMU pre-integration from time i to j-1, define the initial IMU pre-integration error-free, i.e.∑ ii is a zero matrix;∑ d j-1 is the covariance of the IMU measurement noise at time j-1, define the IMU covariance at any time to be a constant value, and is composed of gyroscope and accelerometer calibration noise η gd , η ad ​ Step 4, predicting pixel coordinates; Step 5, determining an adaptive threshold. 2.The method of claim 1, wherein, The method for predicting pixel coordinates: a rotation matrix of the carrier between two adjacent image frames is calculated using the measurement values of the inertial measurement unit and a pre-integration principle and a translation vector , respectively, are the relative poses and positions and , respectively, which are obtained by pre-integration of the inertial measurement unit IMU for the two image frames at times i and j According to the pose of the carrier in the world coordinate system at the time of the previous image frame and the relative pose between the two adjacent images, the pose of the carrier in the world coordinate system at the time of the current image frame is predicted as: ， wherein, are the pose, velocity and position of the vehicle in the world coordinate system at the time of the previous image frame, respectively, AT is the sampling interval of the camera, and g is the earth gravity. The pose of the carrier in the world coordinate system at the time of the current image frame is converted into the pose of the world system relative to the carrier system as: ， The pose of the camera relative to the world system at the time of the current image frame is predicted by using the camera system and the carrier system conversion matrix of the visual inertial SLAM calibration as: ， wherein, R cb , t cb are the camera system and the translation vector of the transformation matrix of the camera system and the translation vector of the vehicle system, respectively, is the camera pose predicted by the inertial measurement unit (IMU) pre-integration. When the camera model is a pinhole camera model, the corresponding pixel coordinates of the map point on the current frame image obtained by using the binocular camera model are: ， wherein, f x , f y , c x , c y is an intrinsic parameter of the camera, P w is a three-dimensional coordinate of the map point in the world coordinate system, P c is a coordinate of the map point in the camera system of the current frame, u , v is a corresponding pixel coordinate of the map point. 3.The method of claim 2, wherein, The method for determining an adaptive threshold: Step 5-1, Compute derivatives with respect to relative pose and position ​ ， Step 5-2, calculating the covariance of the pose of the carrier relative to the world system at the time of the current image frame: (13) wherein, is corresponding Lie algebra, is inverse of the right Jacobian matrix, is the covariance of the relative pose obtained by IMU pre-integration; Step 5-3, calculating the adaptive feature matching search radius: ， where σ0is the unit weight error, σ u ,σ v are the errors of pixel coordinates u, v, respectively, and r is the search radius of feature matching, which is taken as the limit error of pixel coordinates.

4. An adaptive inertial-aided visual SLAM feature matching system, comprising: Comprise: A first processing module for pre-processing binocular camera image data; A second processing module for pre-integrating inertial measurement unit IMU measurement values; A third processing module for error analysis on the pre-integration; A fourth processing module for predicting pixel coordinates; A fifth processing module for determining an adaptive threshold; The first processing module, the second processing module, the third processing module, the fourth processing module and the fifth processing module are connected in sequence to complete the visual SLAM feature matching method based on adaptive inertial auxiliary of any one of claims 1-3. 5.A visual SLAM feature matching device, comprising: a memory; and a processor coupled to the memory, the processor being configured to execute the visual SLAM feature matching method based on adaptive inertial auxiliary of any one of claims 1-3 based on instructions stored in the memory. 6.A non-transitory computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the visual SLAM feature matching method based on adaptive inertial auxiliary of any one of claims 1-3.

Images