Visual SLAM (Simultaneous Localization and Mapping) back-end optimization method based on data adaptive iterative reweighted robust kernel

Abstract

The invention belongs to the field of robot and computer vision, and particularly relates to a visual SLAM (Simultaneous Localization and Mapping) back-end optimization technology based on a data self-adaptive iterative reweighted robust kernel. Comprising the following steps: S1, a front-end data reading and processing module: a robot carries a camera to collect an image, performs feature extraction and feature matching on the image, and generates an initial camera pose and map points; s2, a rear-end optimization module: according to the initial camera pose and map point generated at the front end, calculating a median absolute deviation in an adaptive strategy, calculating a robustness parameter, then substituting the median absolute deviation and the robustness parameter into a target function, simplifying the target function by using an iterative reweighting strategy, and finally solving the optimal camera pose and map point by using an L-M algorithm; s3, the loopback detection module detects loopback according to whether the robot reaches a previous position or not, and transmits information to a rear end for processing; and S4, constructing a map module: constructing an environment map according to the optimized camera pose and map points.

Description

Technical Field

[0001] The present invention relates to the field of robotics and computer vision, and in particular to a visual SLAM backend optimization technology based on a data-adaptive iterative reweighted robust kernel, which is used to improve the accuracy and robustness of a VSLAM system in complex scenes. Background Art

[0002] Visual Simultaneous Localization and Mapping (VSLAM) technology is one of the core technologies for robot navigation and environmental perception. Its development has benefited from its wide application in fields such as mobile robots, self-driving cars, and virtual / augmented reality (VR / AR). VSLAM estimates the camera's pose in real time and constructs a three-dimensional environmental map through image data collected by the camera, providing key support for the perception and decision-making of autonomous mobile devices in complex environments. With the rapid development of artificial intelligence and machine learning technologies, the research and application of VSLAM have made significant progress, bringing unprecedented possibilities to robotics and intelligent systems. However, despite many breakthroughs in technology, the complexity of the real world still poses severe challenges to VSLAM. Dynamic changes, noise interference, data loss, and the presence of outliers in the real environment will significantly affect the performance of the VSLAM system. Especially in dynamic scenes, such as densely populated urban streets, indoor environments with rapidly changing lighting conditions, or areas with scarce texture features, VSLAM systems often find it difficult to maintain accurate mapping of the environment and efficient estimation of pose. These problems not only affect the accuracy and robustness of the system, but also limit its widespread deployment in complex and changing scenes.

[0003] In order to improve the mapping accuracy of the visual simultaneous localization and mapping (VSLAM) system, back-end optimization is an indispensable part. Back-end optimization reduces error accumulation and ensures the global consistency of the system output by processing the preliminary pose and map information obtained from the front-end. In the early days of VSLAM technology, back-end optimization usually uses filter-based methods such as extended Kalman filtering (EKF) and particle filtering. These methods perform well in low-dimensional and small-scale environments, but their application is gradually limited due to their sensitivity to outliers and noise, and the difficulty in maintaining global consistency. Especially in large-scale scenes or complex environments, these filters are often difficult to meet actual needs.

[0004] In recent years, nonlinear optimization methods have gradually become the core technology of advanced VSLAM frameworks, and representative systems include ORB-SLAM2 and ORB-SLAM3. These frameworks effectively improve the robustness and accuracy of the system by modeling the positioning and mapping problems as nonlinear minimization problems, using reprojection errors as cost function targets, and combining optimization algorithms such as Levenberg-Marquardt (LM) and Gauss-Newton. This graph optimization-based method can not only process large-scale data, but also achieve global consistency through loop detection, becoming the mainstream solution for current VSLAM systems.

[0005] To further enhance the adaptability and robustness of the system in different scenarios, advanced VSLAM frameworks including ORB-SLAM3 and CCM-SLAM have introduced robust kernel functions in backend optimization. Robust kernel functions effectively improve the performance of the system in dynamic or complex environments by suppressing the influence of outliers on the optimization results. Common kernel functions such as Huber kernel, Cauchy kernel, and Tukey kernel achieve optimization through fixed robust parameters. This fixed parameter design performs well in some scenarios, but in the diverse scenarios of the real world, fixed parameter methods often have difficulty coping with different noise patterns and outlier distributions.

[0006] Especially when the scene complexity and dynamic changes increase, choosing the right robust kernel function and its parameters becomes a major challenge. On the one hand, the types and scope of application of existing robust kernel functions are limited, and it is difficult to cover all possible application scenarios; on the other hand, the multi-parameter coupling problem significantly increases the complexity and difficulty of parameter adjustment, which limits the practical application of robust kernel functions. For example, when dealing with non-Gaussian noise or complex dynamic scenes, a robust kernel with fixed parameters may not be able to strike a balance between accuracy and robustness. Therefore, finding a method that can adaptively adjust the parameters of the robust kernel function according to different scenarios is of great significance to the existing back-end optimization research. Summary of the invention

[0007] In order to solve the above problems, the present invention provides a visual SLAM back-end optimization method based on a data-adaptive iteratively reweighted robust kernel (DA-IRRK, data-adaptive iteratively reweighted robust kernel-based approach for back-end optimization in visual SLAM).

[0008] The present invention adopts the following technical scheme: a visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel, comprising:

[0009] S1: Front-end data reading and processing module: collects images through the camera carried by the robot, extracts and matches features of the images, and generates initial camera poses and map points;

[0010] S2: Backend optimization module: After this module is turned on, in order to reduce computing resources and realize real-time operation of the robot system, after collecting images, several representative frames in the image frames will be selected as key frames as optimization objects. At the same time, during the SLAM operation, if no loop is detected, the key frames of the local map (local key frames) will be optimized (i.e. local optimization). After the loop is detected, the key frames of the entire map (global key frames) will be optimized (i.e. global optimization); this module first calculates the median absolute deviation (MAD) in the adaptive strategy based on the initial camera pose and map points generated by the front-end data reading and processing module, and brings the robustness parameters obtained by the median absolute deviation (MAD) back to the constructed objective function, simplifies the objective function using the iterative reweighting strategy, and finally uses the Levenberg-Marquardt (LM) algorithm to solve the optimal camera pose and map points; the key frame refers to an image with clear pixels after repetition is eliminated;

[0011] S3: Loop detection module: detects loops based on whether the robot has reached the starting position. If a loop is detected, the information will be passed to the back-end optimization module for processing;

[0012] S4: Map building module: After the optimization is completed, the required environment map is built according to the optimal camera pose and map points.

[0013] Step S1: The front-end data reading and processing module includes the following steps:

[0014] S11: Collect image data in the environment and obtain input through a camera (such as a monocular, binocular or RGB-D camera);

[0015] S12: Perform feature extraction and filtering on the input image to extract key points and descriptors to reduce noise interference;

[0016] S13: Use geometric constraints to perform feature matching, eliminate false matches, and generate initial camera poses and map points.

[0017] Step S2: The backend optimization module includes the following steps:

[0018] S21: Calculate the median absolute deviation MAD of the adaptive strategy according to the initial camera pose and map points generated by the front-end data reading and processing module, and then calculate the robustness parameters in the robust kernel function through the median absolute deviation to achieve adaptive adjustment of the robustness parameters;

[0019] S22: construct the objective function, bring the robustness parameter back into the objective function, and simplify it through iterative reweighting strategy;

[0020] S23: Optimize the pose and map points of the current frame through the Levenberg-Marquardt (LM) algorithm until the reprojection error converges or reaches a preset threshold.

[0021] Step S21 includes:

[0022] S211: Calculate the current reprojection error based on the initial camera pose and map points generated by the front end:

[0023] e ij (χ) = z ij -f ij (x)

[0024] The above formula represents the three-dimensional map point χ in the real environment mj , the observation value z on the camera's i-th two-dimensional image ij The projection value f calculated on the corresponding two-dimensional image ij (·), where χ mj It refers to the 3D map point in the real environment projected onto the m-th frame of the 2D image. The value of m ranges from 1 to i. The observation value z ij Expressed in pixel coordinates, corresponding to χ mj The projection in the i-th frame; where i ranges from 1 to N, N is the number of image frames captured by the camera; j ranges from 1 to M, M is the number of map points that can be seen from the position of the i-th frame;

[0025] S212: Calculate the median absolute deviation MAD value and the corresponding robustness parameter τ according to the reprojection error obtained from the current initial camera pose and map points;

[0026]

[0027] In the formula, σ MAD (e ij (χ)) represents the adaptive threshold, median(x) represents the median of the data set consisting of the reprojection error, and x v represents the vth reprojection error value in the set, MAD(x) refers to the median of the absolute values ​​of the deviations between the reprojection error data and their medians, C represents the scaling factor, is MAD (e ij (χ)), and T is the vector form of τ.

[0028] The step S22 comprises:

[0029] Taking the Huber kernel function as an example, first of all, the Huber kernel function and its first-order derivative function are as follows:

[0030]

[0031] Among them, τ represents the robustness parameter and e represents the error.

[0032] When using the Huber kernel function to solve the visual SLAM backend optimization problem, the objective function constructed is as follows:

[0033]

[0034] Deriving the objective function takes the following form:

[0035]

[0036] in, Defined as l(e ij (χ)), l′(e ij (χ)) represents the influence function, represents the weight function.

[0037] The step S23 comprises:

[0038] In the Levenberg-Marquardt algorithm, the projection equation f ij (χ) First, perform Taylor expansion:

[0039]

[0040] Where Δχ represents the increment of camera pose and map points, and J is The matrix form of

[0041] Then substitute it into the reprojection error formula and the derivative formula of the objective function:

[0042]

[0043] Finally, the incremental matrix form of the camera pose and map points is obtained, and the optimal solution of the camera pose and map points is obtained through iterative update:

[0044] Below is the matrix form of the increment Δχ in the LM algorithm:

[0045] Δχ=(J T WJ+λ·diag(J T WJ)) -1 J T WΔE

[0046] Where W is the weight function ωij The matrix form of (Δχ), λ is the Lagrange multiplier, ΔE is Δe ij The matrix form of .

[0047] The loop detection in step S3 includes the following steps:

[0048] S31: performing loop detection while S13 is being performed to determine whether the current frame forms a loop with the image frame captured at the previous starting position;

[0049] S32: If a loop is detected, all camera poses and map points are input into the S2 backend optimization module for optimization to reduce drift errors.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The present invention proposes a method based on data adaptive iterative reweighted robust kernel (DA-IRRK) for back-end optimization of VSLAM. In this method, the kernel function is used as the objective function of the back-end optimization problem. The robust parameters in the kernel function are adaptively updated according to the reprojection error reflected by the median absolute deviation (MAD). The proposed method shows high robustness in different scenarios. In addition, the method solves the constructed back-end optimization problem through an iterative reweighted update process.

[0052] 2. The proposed method has been implemented in a variety of VSLAM frameworks, including ORB-SLAM3, JORB-SLAM and CCM-SLAM, to demonstrate its effectiveness in vision-only SLAM, multi-sensor fusion SLAM and collaborative SLAM. The method was tested on indoor and outdoor datasets and compared with other robust kernel methods and the advanced multi-convex combination maximum correlation entropy criterion (MCMCC) method.

[0053] 3. The present invention analyzes the performance difference between this method and other methods from the perspective of reprojection error statistics, providing new insights into the application of adaptivity in VSLAM backend optimization problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flow chart of the present invention;

[0055] Figure 2 It is a schematic diagram of the reprojection process of the present invention;

[0056] Figure 3 Schematic diagram of the relationship between the reprojection error of the present invention and the robust parameter τ in the DA-IRRK method in EuRoC_V101 and EuRoC_V203;

[0057] Figure 4 A schematic diagram showing the relationship between the reprojection error of the present invention and the robust parameter τ in the Huber method in EuRoC_V101 and EuRoC_V203;

[0058] Figure 5 A schematic diagram showing the relationship between the reprojection error of the present invention and the robust parameter τ in the DA-IRRK method in KITTI_01 and KITTI_09;

[0059] Figure 6 A schematic diagram showing the relationship between the reprojection error of the present invention and the robust parameter τ in the Huber method in KITTI_01 and KITTI_09;

[0060] Figure 7 Schematic diagram of the reprojection error distribution of the global key frame in EuRoC of the present invention;

[0061] Figure 8 Schematic diagram of the reprojection error distribution of local key frames at three different time points in EuRoC of the present invention;

[0062] Fig. 9 Schematic diagram of the reprojection error distribution of the global key frame in KITTI of the present invention;

[0063] Fig.10 Schematic diagram of the reprojection error distribution of local key frames at three different time points in KITTI of the present invention;

[0064] Fig.11 A schematic diagram showing the relationship between the reprojection error of the present invention and the robust parameter τ in the DA-IRRK method in EuRoC_V201 and EuRoC_V202;

[0065] Fig.12 A schematic diagram showing the relationship between the reprojection error of the present invention and the robust parameter τ in the Huber method in EuRoC_V201 and EuRoC_V202;

[0066] Fig.13 A schematic diagram showing the relationship between the reprojection error of the present invention and the robust parameter τ in the DA-IRRK method in KITTI_06 and KITTI_07;

[0067] Fig.14 A schematic diagram showing the relationship between the reprojection error of the present invention and the robust parameter τ in the Huber method in KITTI_06 and KITTI_07;

[0068] Fig.15It is a schematic diagram of the local mapping reprojection error distribution of sequences 06 and 07 in the JORB-SLAM of the present invention;

[0069] Fig.16 Framework diagram of the optimization method described in the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0071] like Figure 1 The figure shows a general SLAM framework. The embodiment of the present invention provides a visual SLAM back-end optimization method based on a data-adaptive iteratively reweighted robust kernel. First, the image is collected through the first two modules, and then the image is preprocessed in the front-end module. The obtained initial camera pose and map points are passed to the back-end module, and finally the updated camera pose and map points are passed to the mapping module to obtain the required map. The embodiment of the present invention mainly involves (1) calculating the median absolute deviation MAD of the adaptive strategy according to the initial camera pose and map points generated by the front-end data reading and processing module, and calculating the robustness parameters in the robust kernel function through the median absolute deviation to achieve adaptive adjustment of the robustness parameters; (2) constructing the objective function, bringing the robustness parameters back to the objective function, and simplifying it through the iterative reweighting strategy; (3) optimizing the pose and map points of the current frame through the Levenberg-Marquardt algorithm until the increments of the camera pose and map points converge or reach a preset threshold. The specific steps are as follows:

[0072] A visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel, comprising:

[0073] S1: Front-end data reading and processing module: collects images through the camera carried by the robot, extracts and matches features of the images, and generates initial camera poses and map points;

[0074] S11: Collect environmental image data in real time through a camera (such as a monocular, binocular or RGB-D camera), and the input image may include continuous frame data in a static or dynamic environment;

[0075] S12: Perform feature extraction and filtering on the input image to extract key points and descriptors to reduce noise interference;

[0076] Step S12 includes:

[0077] S121: Feature detection: Detect significant key points in the image through algorithms (such as ORB or SIFT);

[0078] S122: Descriptor calculation: Generate the corresponding descriptor according to the feature point for feature matching;

[0079] S123: Filtering and denoising: Remove low-quality feature points, reduce the impact of noise on matching, and enhance adaptability to dynamic scenes.

[0080] S13: Use geometric constraints (such as fundamental matrix or homography matrix) to perform feature matching, eliminate false matches, and generate initial camera poses and map points.

[0081] Step S13 includes:

[0082] S131: Matching point pair generation: Find the potential corresponding feature point pairs between images through descriptor matching;

[0083] S132: False matching elimination: Combined with the RANSAC algorithm, the erroneous matching point pairs are eliminated and the feature point pairs with high confidence are retained;

[0084] S133: Preliminary pose estimation: Calculate the camera pose transformation of the current frame relative to the reference frame based on high-quality matching point pairs, and generate preliminary sparse map points in combination with disparity information.

[0085] S2: Backend optimization module: According to the reprojection error distribution (environmental characteristics), the median absolute deviation (MAD) is calculated to construct a cost function that minimizes the reprojection error. The Huber kernel function is used as an example to illustrate this.

[0086] S21: Calculate the current reprojection error based on the initial camera pose and map points obtained in S1, such as Figure 2 As shown, it is the principle diagram of reprojection error;

[0087] e ij (χ) = z ij -f ij (x)

[0088] The above formula represents the three-dimensional map point χ in the real environment mj , the observation value z on the camera's i-th two-dimensional image ij The projection value f calculated on the corresponding two-dimensional image ij (·), where χ mj It refers to the 3D map point in the real environment projected onto the m-th frame of the 2D image. The value of m ranges from 1 to i. The observation value z ij Expressed in pixel coordinates, corresponding to χ mjThe projection in the i-th frame; where i ranges from 1 to N, N is the number of image frames captured by the camera; j ranges from 1 to M, M is the number of map points that can be seen from the position of the i-th frame.

[0089] The robustness parameter τ is calculated according to the median absolute deviation (MAD) strategy, as Figure 3 , 5, 11, 13 are respectively the robustness parameter τ values ​​obtained by the adaptive algorithm proposed in this invention in different data sets, Figure 4 , 6, 12, and 14 are the robustness parameter τ values ​​obtained by the Huber kernel function in different data sets.

[0090]

[0091] In the formula, σ MAD (e ij (χ)) represents the adaptive threshold, median(x) represents the median of the data set consisting of the reprojection error, and x v represents the vth reprojection error value in the set, MAD(x) refers to the median of the absolute value of the deviation between the reprojection error data and their median, C represents the scaling factor, and according to the efficiency value of the normal model is 95%, C is usually taken as 1.345, is MAD (e ij (χ)), and T is the vector form of τ.

[0092] S22: Iteratively solving the objective function through a reweighted update process;

[0093] Taking the Huber kernel function as an example, first of all, the Huber kernel function and its first-order derivative function are as follows:

[0094]

[0095] Among them, τ represents the robustness parameter and e represents the error.

[0096] When using the Huber kernel function to solve the visual SLAM backend optimization problem, the objective function constructed is as follows:

[0097]

[0098] In the general algorithm of SLAM, in order to optimize the reprojection error, the minimum value of the objective function is usually solved, that is, the reprojection error is minimized. At this time, the objective function needs to be derived, which is the following form:

[0099]

[0100] in, Defined as l(e ij (χ)), l′(e ij (χ)) represents the influence function, represents the weight function.

[0101] After simplifying the objective function using the above iterative reweighting strategy, the objective function can be further optimized using the Levenberg-Marquardt (LM) algorithm.

[0102] S23: Optimizing the pose and map points of the current frame through the LM algorithm until the reprojection error converges or reaches a preset threshold;

[0103] In the LM algorithm, the projection equation f ij (χ) First, perform Taylor expansion,

[0104]

[0105] Where Δχ represents the updated camera pose and map point, and J is The matrix form of

[0106] Then substitute it into the reprojection error formula and the derivative formula of the objective function.

[0107]

[0108] Finally, the incremental matrix form of the camera pose and map points is obtained, and the optimal solution of the camera pose and map points is obtained through iterative update:

[0109] The following is the increment in the LM algorithm Δ The matrix form of χ is:

[0110] Δχ=(J T WJ+λ·diag(J T WJ)) -1 J T WΔE

[0111] Where W is the weight function ω ij The matrix form of (Δχ), λ is the Lagrange multiplier, ΔE is Δe ij The matrix form of .

[0112] S3: Loop detection module: detects loops based on whether the robot has reached the starting position. If a loop is detected, the information will be passed to the back-end optimization module for processing;

[0113] S31: While executing S13 feature matching, start the loop detection module to determine in real time whether the current frame forms a loop with the image frame collected at the starting position. By comparing the image features or posture similarity between the current frame and the image frame collected at the starting position, when the similarity exceeds a preset threshold, it is considered that a loop is established;

[0114] S32: Once a loop is detected, the camera pose information and map point information are passed to the S2 backend optimization module, and the pose and map points of all image frames are optimized using optimization methods to reduce trajectory drift errors and ensure the consistency of pose and map points. During the optimization process, the pose is adjusted based on the objective function constructed based on the adaptive robust kernel function to enhance the robustness of the system.

[0115] S4: Map building module: After the optimization is completed, the required environment map is built according to the optimal camera pose and map points.

[0116] The evaluation tool EVO is used to evaluate the obtained trajectory. The evaluation index is the root mean square error RMSE corresponding to the absolute pose error APE. The proposed algorithm is compared with other methods based on robust functions (Huber, MCMCC, Cauchy, Tukey) through EuRoC datasets and KITTI datasets. At the same time, in order to verify the versatility of the proposed method, it is verified in different SLAM frameworks (ORB-SLAM3, JORB-SLAM, CCM-SLAM) and multi-sensor combinations (vision + inertial navigation).

[0117] The specific experimental process includes the following steps:

[0118] (1) The proposed MAD adaptive strategy is implanted into the Huber kernel function, the robustness parameters are updated in real time, and then the objective function is solved using the iterative reweighting strategy.

[0119] (2) The proposed method was experimented on the visual SLAM framework ORB-SLAM3 using the EuRoC dataset and the KITTI dataset. The experimental results are as follows:

[0120] Table 1 Experimental results on the EuRoC dataset

[0121]

[0122] Table 2 Experimental results on the KITTI dataset

[0123]

[0124] Table 3 Experimental results of multi-sensor combination (vision + inertial navigation) on EuRoC dataset

[0125]

[0126] Table 4 Experimental results on JORB-SLAM

[0127]

[0128] Table 5 Experimental results on CCM-SLAM

[0129]

[0130] By comparing the experimental results, whether in indoor and outdoor scenes of pure visual frameworks, or in multiple modes such as multi-sensor combinations and different visual SLAM frameworks, the proposed method is more accurate than other algorithms in most cases, showing good robustness. In the optimization process, in order to reduce computing resources and achieve real-time operation, key frames will be selected as optimization objects after the image is collected. At the same time, during the SLAM operation process, if no loop is detected, the key frames (local key frames) of the local map will be optimized (local optimization). After the loop is detected, the key frames (global key frames) of the entire map will be optimized (global optimization). Therefore, in the present invention, by analyzing the distribution of reprojection errors of global key frames and local key frames of the KITTI dataset and the EuRoC dataset, the root causes of different experimental results of various algorithms are deeply understood, such as Figure 7 is the reprojection error distribution of the global keyframe of the EuRoC dataset, such as Figure 8 is the reprojection error distribution of the local keyframes of the EuRoC dataset, such as Fig. 9 is the reprojection error of the global keyframe of the KITTI dataset, such as Fig.10 is the reprojection error distribution of local keyframes of the KITTI dataset.

Claims

1. A visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel, characterized in that: include: S1: Front-end data reading and processing module: collects images through the camera carried by the robot, extracts and matches features of the images, and generates initial camera poses and map points; S2: Backend optimization module: After collecting images, several representative frames in the image frames will be selected as key frames for optimization. At the same time, during the SLAM operation, if no loop is detected, the key frames of the local map, namely the local key frames, will be optimized. After the loop is detected, the key frames of the entire map, namely the global key frames, will be optimized. This module first calculates the median absolute deviation MAD in the adaptive strategy based on the initial camera pose and map points generated by the front-end data reading and processing module, and then brings the robustness parameters obtained by the median absolute deviation MAD back to the constructed objective function, uses the iterative reweighting strategy to simplify the objective function, and finally uses the Levenberg-Marquardt algorithm to solve the optimal camera pose and map points. S3: Loop detection module: detects loops based on whether the robot has reached the previous position. If a loop is detected, the information will be passed to the back-end optimization module for processing; S4: Map building module: After the optimization is completed, the required environment map is built according to the optimal camera pose and map points.

2. A visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel according to claim 1, characterized in that: The step S1 comprises: S11: Collect image data in the environment and obtain input through the camera; S12: Perform feature extraction and filtering on the input image to extract key points and descriptors to reduce noise interference; S13: Use geometric constraints to perform feature matching, eliminate false matches, and generate initial camera poses and map points.

3. A visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel according to claim 1, characterized in that: The step S2 comprises: S21: Calculate the median absolute deviation MAD of the adaptive strategy according to the initial camera pose and map points generated by the front-end data reading and processing module, and then calculate the robustness parameters in the robust kernel function through the median absolute deviation to achieve adaptive adjustment of the robustness parameters; S22: construct the objective function, bring the robustness parameter back into the objective function, and simplify it through iterative reweighting strategy; S23: Optimize the pose and map points of the current frame through the Levenberg-Marquardt algorithm until the increments of the camera pose and map points converge or reach a preset threshold.

4. A visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel according to claim 3, characterized in that: The step S21 comprises: S211: Calculate the current reprojection error based on the initial camera pose and map points generated by the front end: e ij (x)=z ij -f ij (x) The above formula represents the three-dimensional map point χ in the real environment mj , the observation value z on the camera's i-th two-dimensional image ij The projection value f calculated on the corresponding two-dimensional image ij (·), where χ mj It refers to the jth 3D map point in the real environment projected onto the mth 2D image. The value of m ranges from 1 to i. The observation value z ij Expressed in pixel coordinates, corresponding to χ mj The projection in the i-th frame; where i ranges from 1 to N, N is the number of image frames captured by the camera; j ranges from 1 to M, M is the number of map points that can be seen from the position of the i-th frame; S212: Calculate the median absolute deviation MAD value and the corresponding robustness parameter τ according to the reprojection error obtained from the current initial camera pose and map points; MAD(x)=median(|x v -median(x)|) T=cσ MAD (e ij (x)) In the formula, σ MAD (e ij (χ)) represents the adaptive threshold, median(x) represents the median of the data set consisting of the reprojection error, and x v represents the vth reprojection error value in the set, MAD(x) refers to the median of the absolute values ​​of the deviations between the reprojection error data and their medians, c represents the scaling factor, is MAD (e ij (χ)), and T is the vector form of τ.

5. A visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel according to claim 4, characterized in that, The step S22 comprises: Taking the Huber kernel function as an example, first of all, the Huber kernel function and its first-order derivative function are as follows: Among them, τ represents the robustness parameter and e represents the error. When using the Huber kernel function to solve the visual SLAM backend optimization problem, the objective function constructed is as follows: Deriving the objective function takes the following form: in, Defined as l(e ij (χ)), l′(e ij (χ)) represents the influence function, represents the weight function.

6. A visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel according to claim 5, characterized in that: The step S23 comprises: In the Levenberg-Marquardt algorithm, the projection equation f ij (χ) First, perform Taylor expansion: Where Δχ represents the increment of camera pose and map points, and J is The matrix form of Then substitute it into the reprojection error formula and the derivative formula of the objective function: e ij (x)=z ij -f ij (x) =z ij -f ij (x k )+f ij (x k )-f ij (x) ≈No ij -ΣJ·δχ Finally, the incremental matrix form of the camera pose and map points is obtained, and the optimal solution of the camera pose and map points is obtained through iterative update: Below is the matrix form of the increment Δχ in the LM algorithm: Δx=(J T WJ+λ·diag(J T WJ)) -1 J T WΔE Where W is the weight function ω ij The matrix form of (Δχ), λ is the Lagrange multiplier, ΔE is Δe ij The matrix form of .

7. A visual SLAM backend optimization method based on data-adaptive iterative reweighted robust kernel according to claim 2, characterized in that: The step S3 comprises: S31: performing loop detection while S13 is being performed to determine whether a loop is formed between the current image frame and the image frame captured at the starting position; S32: If a loop is detected, the camera pose and map points of the global keyframe are input into the S2 backend optimization module for optimization to reduce drift errors.

Images