Visual-inertial fusion SLAM method for mobile robots based on variational Bayesian

Abstract

The present invention provides a mobile robot visual inertial fusion SLAM method based on variational Bayes, including: S1, obtaining the predicted state vector and covariance matrix of the mobile robot at time k; S2, establishing a model of the observation noise at time k that conforms to the Student's t mixed distribution based on the gamma distribution, and obtaining the likelihood function distribution in a hierarchical Gaussian form based on the visual information, observation noise model and predicted state vector of the mobile robot at time k; S3, constructing a joint posterior probability density function based on the likelihood function distribution, and using the variational Bayes method to jointly solve the posterior joint probability density function to obtain the state vector of the mobile robot at time k. A new distribution is proposed to model the observation noise to reduce the influence of non-stationary thick-tailed noise on the accuracy of the visual and inertial data fusion results, and the variational Bayes is used to realize the joint estimation of the mobile robot state, thereby improving the accuracy of the final output state vector.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent robots, and in particular to a mobile robot vision-inertial fusion SLAM method based on variational Bayes. Background Art

[0002] Simultaneous Localization And Mapping (SLAM) refers to the process in which a mobile robot extracts and combines useful information while operating in an unknown environment, determines its own position and gradually builds a map of the surrounding environment. SLAM has been successfully applied in underwater unmanned vehicles, unmanned driving, and aerial drones, and has received increasing attention.

[0003] The filter based on Bayesian estimation is widely used in the back-end optimization of SLAM. The current state vector is obtained based on the state vector of the mobile robot at the previous moment and the estimated value at the current moment. The earliest SLAM algorithm based on the extended Kalman filter (EKF) needs to use the first-order Taylor expansion to linearize the motion model and observation model in the SLAM algorithm before using the EKF to estimate the robot posture and map feature position. However, the linearization process of the two models will introduce corresponding errors. In addition, the calculation process of the Jacobian matrix will also bring great burden to the algorithm. Later, the unscented Kalman filter (UKF) appeared. Patent CN109141436A used this method to substitute a set of sampling points with certain weights into the motion model and observation model of the SLAM algorithm, avoiding linearization in the calculation of the state information and observation information of the mobile robot. However, when the system dimension gradually increases, non-local sampling will occur in the UKF, which directly leads to the instability or even divergence of the calculation results of SLAM. Then there is a SLAM algorithm based on volumetric Kalman filter (cubature Kalman filter, CKF), which substitutes a set of sampling points with the same weights into a nonlinear model for calculation, but this method discards part of the approximation error, resulting in reduced filtering accuracy. None of the above methods consider the influence of thick-tailed noise. To address this problem, patent CN113971752A uses student t distribution to simulate non-Gaussian noise with wild values ​​to improve estimation accuracy, and approximates the posterior probability density function to a student t probability density function with a fixed degree of freedom parameter. But in practice, thick-tailed noise also has non-stationary characteristics, and not considering the influence of non-stationary characteristics will greatly reduce the estimation accuracy of Kalman filtering. Summary of the invention

[0004] The present invention aims to solve the problem in the prior art that the Kalman filter does not consider the non-stationarity of noise, which affects the final fused data accuracy, and provides a mobile robot visual-inertial fusion SLAM method based on variational Bayesian.

[0005] In order to achieve the above-mentioned purpose of the present invention, according to the first aspect of the present invention, the present invention provides a mobile robot visual-inertial fusion SLAM method based on variational Bayes, characterized in that it includes: step S1, based on the state vector and covariance matrix of the mobile robot at time k-1, and the IMU data and system noise covariance matrix of the mobile robot at time k, obtaining the predicted state vector and one-step prediction error covariance matrix of the mobile robot at time k; step S2, establishing a model of the observation noise at time k that conforms to the Student's t mixed distribution based on the gamma distribution, and obtaining a likelihood function distribution in a hierarchical Gaussian form based on the visual information, observation noise model and predicted state vector of the mobile robot at time k; step S3, constructing a joint posterior probability density function based on the likelihood function distribution, and using the variational Bayes method to jointly solve the posterior joint probability density function to obtain the state vector of the mobile robot at time k.

[0006] In order to achieve the above-mentioned objectives of the present invention, according to the same inventive concept, the present invention provides a computer-readable storage medium, in which at least one instruction, at least one program, a code set or an instruction set is stored, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement the variational Bayesian-based mobile robot visual-inertial fusion SLAM method as described in the first aspect of the present invention.

[0007] In order to achieve the above-mentioned objectives of the present invention, based on the same inventive concept, the present invention provides a mobile robot, including a mobile robot body and a visual sensor, an IMU sensor and a processor arranged on the mobile robot body, the processor being connected to the visual device and the IMU sensor respectively, and the processor executing the steps of the mobile robot visual-inertial fusion SLAM method based on variational Bayes described in the first aspect of the present invention.

[0008] The above technical solution: proposes a new distribution (i.e., a Student-t mixed distribution based on gamma distribution) to model observation noise. This distribution uses gamma distribution as the mixed density function and scales the Student-t distribution to cope with the non-stationary behavior of the noise, thereby reducing the impact of non-stationary thick-tailed noise on the accuracy of visual and inertial data fusion results, thereby improving the estimation accuracy of the Kalman filter, and introduces information iterative update based on variational Bayes in the process of fusing IMU data and visual sensor measurement data, and uses variational Bayes to realize the joint estimation of the mobile robot state, thereby further improving the accuracy of the final state vector obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 1 is a flow chart of a mobile robot visual-inertial fusion SLAM method based on variational Bayes in Example 1 of the present invention;

[0010] Figure 2 It is a schematic diagram of the data fusion process of Example 1 of the present invention in an application scenario. DETAILED DESCRIPTION

[0011] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0012] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0013] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0014] Example 1

[0015] This embodiment discloses a mobile robot visual-inertial fusion SLAM method based on variational Bayesian, such as Figure 1 As shown, including:

[0016] Step S1, based on the state vector x of the mobile robot at time k-1 k-1||k-1 and the covariance matrix P k-1|k-1 , as well as the IMU data of the mobile robot at time k and the system noise covariance matrix Q k , obtain the predicted state vector x of the mobile robot at time k k|k-1 and the one-step forecast error covariance matrix P k|k-1 .

[0017] IMU data is provided by IMU sensors, which include accelerometers and gyroscopes. The accelerometer provides the acceleration of the mobile robot during movement, and the gyroscope provides the angular velocity of the mobile robot during movement. After obtaining the IMU data, the IMU data is pre-integrated. The pre-integration process includes removing noise from the acceleration value measured by the accelerometer and integrating it once to obtain the velocity, and integrating it twice to obtain the relative displacement; the pre-integration process also includes removing noise from the angular velocity value measured by the gyroscope and integrating it once to obtain the angular change of the mobile robot during movement.

[0018] In step S1, the predicted state vector x of the mobile robot at time k is obtained using the IMU data of the mobile robot at time k and the constructed volume points. k|k-1 , specifically: X k|k-1 =F k-1 X k-1||k-1 , where F k-1 represents the state vector conversion relationship between the mobile robot at time k-1 and time k obtained based on IMU data. Preferably, F k-1 Obtained through pre-integration of IMU data; X k-1||k-1 represents the state vector of the mobile robot at time k-1; X k|k-1 =[P bk , V bk , Q bk ], P bk and V bk They are the three-dimensional position and velocity information of the mobile robot, Q bk A quaternion representing the conversion from the IMU coordinate system to the world coordinate system.

[0019] In step S1, the predicted state vector X of the mobile robot at time k is used k|k-1 and the system noise covariance matrix Q k The one-step prediction error covariance matrix of the mobile robot at time k is calculated as follows:

[0020] Among them, P k-1|k-1 represents the covariance matrix of the mobile robot at time k-1, Q k represents the system noise covariance matrix of the mobile robot at time k. Then the prior probability density function is: p(X k |Y k-1 )=N(X k ;X k|k-1 P k|k-1 ).

[0021] Step S2, establish a model of the observation noise at time k that conforms to the Student's t mixed distribution based on the gamma distribution, and obtain the likelihood function distribution in the form of a hierarchical Gaussian based on the visual information of the mobile robot at time k, the observation noise model and the predicted state vector.

[0022] Model the observation noise at time k with a Student's t mixed distribution based on the gamma distribution:

[0023]

[0024] Right now

[0025] Among them, v k represents the observation noise of the mobile robot at time k, Indicates v k The initial value of β k represents the shape parameter, β k The distribution function of in, represents the first constant real number, y represents the second constant real number, I m represents the m-order identity matrix; R k represents the scale matrix; u k represents the first degree of freedom, u k The distribution function is p(u k )=G(u k ; a k , b k ), a k , b k They belong to u k The first real parameter and the second real parameter of the gamma distribution; δ k represents the second degree of freedom; k represents the first auxiliary parameter, λ k The distribution function of k )=G(λ k ; δ k / 2,δ k / 2);ξ k represents the second auxiliary parameter, ξ k The distribution function is p(ξ k )=G(ξ k ;u k / 2,u k / 2); N represents Gaussian distribution; G represents gamma distribution, and GaSTM represents Student-t mixture distribution based on gamma distribution.

[0026] A new distribution is proposed: Student-t mixture distribution based on gamma distribution This distribution has a location parameter Shape parameter βk , scale matrix R k , the first and second different degrees of freedom parameters u k and δ k , which uses the gamma distribution as the mixing density function and performs scale mixing in Student-t steps to cope with the non-stationary characteristics of noise.

[0027] In step S2, the step of obtaining a likelihood function distribution in a hierarchical Gaussian form based on the visual information of the mobile robot at time k, the observation noise model and the predicted state vector includes:

[0028] Step S21, obtaining the visual state information Y of the mobile robot at time k based on the visual information of the mobile robot at time k k , and Y k Meet Y k =H k X k|k-1 +v k ; Among them, H k Represents the observation matrix, which represents the visual state information Y of the mobile robot at time k k With the predicted state vector X k|k-1 Specifically, the ORB (Oriented FAST and Rotated BRIEF) feature of the visual information is extracted according to the visual information of the robot at time k, and the visual state information Y is obtained according to the ORB feature. k The process of obtaining visual state information is prior art and will not be described in detail here. k Including the three-dimensional position and velocity information of the mobile robot;

[0029] Step S22, establish the likelihood function distribution as:

[0030] p(Y k |X k|k-1 )=GaSTM(Y k ;H k X k|k-1 , β k , R k ,u k , δ k );

[0031] The likelihood function satisfies the GaSTM distribution, and the likelihood function distribution is converted into a hierarchical Gaussian form:

[0032]

[0033] Step S3, constructing a joint posterior probability density function based on the likelihood function distribution, and using the variational Bayesian method to jointly solve the posterior joint probability density function to obtain the state vector of the mobile robot at time k.

[0034] According to the prior probability density function p(X k |Y k-1 )=N(X k ;X k|k-1 , P k|k-1 ) and the hierarchical Gaussian likelihood function distribution to obtain the joint posterior probability density function p(X k ,λ k ,ξ k ,u k , β k , Y k ):

[0035]

[0036] That is: p(X k ,λ k ,ξ k ,u k , β k , Y k )≈q(X k )q(λ k )q(ξ k )q(u k )q(β k ); (1)

[0037] Among them, q(X k ) represents the state vector X of the mobile robot at time k k The approximate probability density function of q(λ k ) represents the first auxiliary parameter λ k The approximate probability density function of q(ξ k ) represents the second auxiliary parameter ξ k The approximate probability density function of q(u k ) represents the first degree of freedom u k The approximate probability density function of q(β k ) represents the shape parameter β k The approximate probability density function of . Therefore, we only need to solve q(X k ),q(λ k ),q(ξ k )、q(u k ),q(β k ) can obtain the final fused robot state vector X at time k k .

[0038] When the posterior joint probability density function is jointly solved using the variational Bayesian method, X k , k , k 、uk , β k The optimal solution satisfies:

[0039]

[0040] Among them, the collection θ represents the set Θ k Any element in Θ k (-θ) represents the set θ k The subset consisting of all elements except θ in C θ represents a constant that is independent of the element θ; E[·] represents the expected value.

[0041] In step S3, the step of jointly solving the posterior joint probability density function using the variational Bayesian method to obtain the fusion state vector of the mobile robot at time k includes:

[0042] Step S31, parameter initialization:

[0043]

[0044] Step S32, using the variational Bayesian method to iteratively update the joint posterior probability density function, and obtaining X k , k , k 、u k , β k If the iteration stop condition is reached, the state vector X of the mobile robot at time k obtained after the last iteration is output. k and the covariance matrix P k|k .

[0045] Further preferably, in step S32, for the i+1th iteration, execute:

[0046] Step S321, substitute formula 1 into formula 2, let θ = X k , we can get:

[0047]

[0048] From the above formula, we can infer that q (i+1) (X k ) obeys Gaussian distribution: q (i+1) (X k ) indicates X k The approximate probability density function in the i+1th iteration. represents the state vector of the mobile robot at time k in the i+1th iteration, X k|k-1 represents the predicted state vector of the mobile robot at time k, Pk|k-1 represents the one-step prediction error covariance matrix of the mobile robot at time k, represents the Kalman gain of the i+1th iteration, E (i) [λ k ] represents λ k The expected value at the i-th iteration, E (i) [β k ] indicates β k The expected value at the i-th iteration, E (i) [ξ k ] represents ξ k The expected value at the i-th iteration, represents the covariance matrix of the mobile robot at time k in the i+1th iteration.

[0049] Step S322, substitute formula 1 into formula 2, and let θ = λ k , we can get:

[0050]

[0051] From the above formula, we can analyze that it is impossible to (i+1) (λ k ) is directly analyzed and approximate points are processed to obtain the optimal solution:

[0052]

[0053] in, E (i) [ξ k ] represents ξ k The expected value at the i-th iteration, tr{} represents the trace of the matrix; n 1 、n 2 、n 3 are the third auxiliary parameter, the fourth auxiliary parameter, and the fifth auxiliary parameter respectively; E (i) [(Y k -H k X k )(Y k -H k X k ) T ] indicates (Y k -H k X k )(Y k -H k X k ) T The expected value at the i-th iteration; express The expected value at the i-th iteration; m represents the identity matrix I m The order of .

[0054] Step S323, substitute formula 1 into formula 2, let θ = ξ k , we can get:

[0055]

[0056] From the above formula, we can infer that q (i+1) (ξ k ) follows a gamma distribution, so ξ k The approximate probability density function in the i+1th iteration is:

[0057] q (i+1) (ξ k )=G(ξ k ; A (i+1) , B (i+1) );

[0058] in, A (i+1) represents the value of the sixth auxiliary parameter A in the i+1th iteration, B (i+1) represents the value of the seventh auxiliary parameter B in the i+1th iteration, C (i+1) represents the value of the eighth auxiliary parameter C in the i+1th iteration, E (i) [u k ] means u k The expected value at the i-th iteration, E (i) [(Y k -H k X k -λ k β k )(Y k -H k X k -λ k β k ) T ] indicates (Y k -H k X k -λ k β k )(Y k -H k X k -λ k β k ) T The expected value at the i-th iteration.

[0059] Step S324, substitute formula 1 into formula 2, let θ = u k , we can get:

[0060]

[0061] From the above formula, we can infer that q(i+1) (u k ) obeys the gamma distribution, and obtains u k The approximate probability density function in the i+1th iteration is:

[0062]

[0063] in, a k (0) Indicates a k The initial value of Indicates a k The value of b in the i+1th iteration k (0) Indicates b k The initial value of Indicates b k The value at the i+1th iteration.

[0064] Step S325, substitute formula 1 into formula 2, and let θ = β k , we can get:

[0065]

[0066] From the above formula, we can infer that q (i+1) (β k ) obeys Gaussian distribution, and obtains β k The approximate probability density function in the i+1th iteration is:

[0067]

[0068] in, Represents β k The value at the i+1th iteration, Represents the eighth auxiliary parameter K β The value at the i+1th iteration, Indicates the ninth auxiliary parameter P β The value at the i+1th iteration.

[0069] Step S326, obtain the expected values ​​of each parameter in the (i+1)th iteration. These expected values ​​can be used as input values ​​for the next iteration or output values ​​after the iteration stop condition is met. The expected values ​​are:

[0070]

[0071] E (i+1) [ξ k ]=A (i+1) / B (i+1) ;

[0072] E(i+1) [u k ] = a (i+1) / b (i+1) ;

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] Step S327, determine whether If satisfied, As the fusion state vector X of the mobile robot at time k k Output, will As the covariance matrix P of the mobile robot at time k k|k Output, ε represents the preset iteration stop threshold. If it is not satisfied, set i=i+1 and return to execute steps S321 to S327.

[0079] In view of the problem that the Kalman filter proposed in the background technology does not consider the non-stationary characteristics of noise, which affects the accuracy of the final fused data, this embodiment proposes a Student-t mixed Kalman filter SLAM algorithm based on gamma distribution. First, a new distribution - Student-t mixed distribution based on gamma distribution is proposed. This distribution has a location parameter Shape parameter β k , scale matrix R k , the first and second different degrees of freedom parameters u k and δ k , which uses the gamma distribution as the mixing density function and performs scale mixing in Student-t steps to cope with the non-stationary characteristics of noise.

[0080] In this embodiment, the main process of the SLAM method can be divided into: first, using the state vector X at time k-1 k-1|k-1 And IMU data, the state vector X at time k k and the one-step forecast error covariance matrix P k|k-1 Second, use GaSTM to measure the observed noise v kThe model is built, and the likelihood function distribution in the form of hierarchical Gaussian is derived using two auxiliary parameters. Thirdly, the posterior joint probability density function is jointly solved using the variational Bayesian method, and the expected values ​​of the state vector, auxiliary parameters, first degree of freedom and shape parameters are iterated in sequence, ensuring the accuracy of the fusion of IMU and visual sensor measurement data and improving the positioning accuracy of SLAM. This method introduces the iterative update of information based on variational Bayesian in the fusion process of IMU and visual sensor measurement data, and uses GaSTM to model the observation noise, which is used to reduce the impact of non-stationary thick-tailed noise on the accuracy of the fusion result, and uses variational Bayesian to realize the joint estimation of the mobile robot state, thereby improving the accuracy of the final state information.

[0081] In an application scenario of this embodiment, Figure 2 The figure shows a schematic diagram of the data fusion process for this application scenario. First, the information at the next moment is predicted based on the IMU data and the current moment information, that is, the information at moment k is predicted using the information at moment k-1. Then, the visual sensor data is used to obtain the observed value, and the observation noise is modeled using GaSTM. Finally, variational Bayes is used to obtain the final result.

[0082] Example 2

[0083] This embodiment is based on the same inventive concept as Embodiment 1 and discloses a computer-readable storage medium, in which at least one instruction, at least one program, code set or instruction set is stored. The at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement the variational Bayesian-based mobile robot visual-inertial fusion SLAM method provided in Embodiment 1.

[0084] Example 3

[0085] Based on the inventive concept of Example 1, this embodiment discloses a mobile robot, including a mobile robot body and a visual sensor, an IMU sensor and a processor arranged on the mobile robot body, the processor is connected to the visual sensor and the IMU sensor respectively, and the processor executes the steps of the mobile robot visual-inertial fusion SLAM method based on variational Bayes provided in Example 1. The visual sensor is preferably, but not limited to, a camera.

[0086] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0087] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A mobile robot visual-inertial fusion SLAM method based on variational Bayes, characterized in that: include: Step S1, based on the state vector and covariance matrix of the mobile robot at time k-1, and the IMU data and system noise covariance matrix of the mobile robot at time k, obtain the predicted state vector and one-step prediction error covariance matrix of the mobile robot at time k; Step S2, establishing a model of the observation noise at time k that conforms to the Student's t mixed distribution based on the gamma distribution, and obtaining a likelihood function distribution in a hierarchical Gaussian form based on the visual information of the mobile robot at time k, the observation noise model and the predicted state vector; Step S3, constructing a joint posterior probability density function based on the likelihood function distribution, and using a variational Bayesian method to jointly solve the joint posterior probability density function to obtain the state vector of the mobile robot at time k; In step S2, the observation noise model of the mobile robot at time k is: Among them, v k represents the observation noise of the mobile robot at time k, Indicates v k The initial value of β k represents the shape parameter, β k The distribution function of in, represents the first constant real number, y represents the second constant real number, I m represents the m-order identity matrix; R k represents the scale matrix; u k represents the first degree of freedom, u k The distribution function is p(u k )=G(u k ; a k , b k ), a k 、b k They belong to u k The first real parameter and the second real parameter of the gamma distribution; δ k represents the second degree of freedom; k represents the first auxiliary parameter, λ k The distribution function of k )=G(λ k ; δ k / 2,δ k / 2);ξ k represents the second auxiliary parameter, ξ k The distribution function is p(ξ k )=G(ξ k ;u k / 2,u k / 2); N represents Gaussian distribution; G represents gamma distribution, and GaSTM represents Student-t mixed distribution based on gamma distribution; In step S2, the step of obtaining a likelihood function distribution in a hierarchical Gaussian form based on the visual information of the mobile robot at time k, the observation noise model and the predicted state vector includes: Based on the visual information of the mobile robot at time k, the visual state information Y of the mobile robot at time k is obtained k , and Y k Meet Y k =H k X k|k-1 +v k ; Among them, H k Represents the observation matrix, which represents the visual state information Y of the mobile robot at time k k With the predicted state vector X k|k-1 The mapping relationship of The likelihood function distribution is established as: p(Y k |X k|k-1 )=GaSTM(Y k ;H k X k|k-1 ,β k ,R k ,u k ,δ k ); Convert the likelihood function distribution into hierarchical Gaussian form: X k represents the state vector of the mobile robot at time k; In step S3, the joint posterior probability density function is: p(X k ,l k ,x k ,u k ,b k ,Y k )≈q(X k )q(λ k )q(ξ k )q(u k )q(β k ); Among them, q(X k ) represents the state vector X of the mobile robot at time k k The approximate probability density function of q(λ k ) represents the first auxiliary parameter λ k The approximate probability density function of q(ξ k ) represents the second auxiliary parameter ξ k The approximate probability density function of q(u k ) represents the first degree of freedom u k The approximate probability density function of q(β k ) represents the shape parameter β k The approximate probability density function of ; In step S3, the step of jointly solving the joint posterior probability density function using the variational Bayesian method to obtain the fusion state vector of the mobile robot at time k includes: Step S31, parameter initialization: Step S32, using the variational Bayesian method to iteratively update the joint posterior probability density function, and obtaining X k , k , k 、u k , β k If the iteration stop condition is reached, the state vector X of the mobile robot at time k obtained after the last iteration is output. k and the covariance matrix P k|k ; In step S32, for the i+1th iteration, execute: step S321, obtain X based on the variational Bayesian method k The approximate probability density function in the i+1th iteration is: in, represents the state vector of the mobile robot at time k in the i+1th iteration, X k|k-1 represents the predicted state vector of the mobile robot at time k, P k|k-1 represents the one-step prediction error covariance matrix of the mobile robot at time k, represents the Kalman gain of the i+1th iteration, E (i) [λ k ] represents λ k The expected value at the i-th iteration, E (i) [β k ] indicates β k The expected value at the i-th iteration, E (i) [ξ k ] represents ξ k The expected value at the i-th iteration, represents the covariance matrix of the mobile robot at time k in the i+1th iteration; Step S322, obtaining λ based on the variational Bayesian method k The optimal solution in the i+1th iteration is: in, E (i) [ξ k ] represents ξ k The expected value of the i-th iteration, tr{} represents the trace of the matrix; n1, n2, n3 are the third auxiliary parameter, the fourth auxiliary parameter, and the fifth auxiliary parameter respectively; E (i) [(Y k -H k X k )(Y k -H k X k ) T ] indicates (Y k -H k X k )(Y k -H k X k ) T The expected value at the i-th iteration; express The expected value at the i-th iteration; m represents the identity matrix I m The order of Step S323, obtaining ξ based on the variational Bayesian method k The approximate probability density function in the i+1th iteration is: in, A (i+1) represents the value of the sixth auxiliary parameter A in the i+1th iteration, B (i+1) represents the value of the seventh auxiliary parameter B in the i+1th iteration, c (i+1) represents the value of the eighth auxiliary parameter C in the i+1th iteration, E (i) [u k ] means u k The expected value at the i-th iteration, E (i) [(Y k -H k X k -λ k β k )(Y k -H k X k -λ k β k ) T ] indicates (Y k -H k X k -λ k β k )(Y k -H k X k -λ k β k ) T The expected value at the i-th iteration; Step S324, obtaining u based on the variational Bayesian method k The approximate probability density function in the i+1th iteration is: in, a k (0) Indicates a k The initial value of Indicates a k The value of b in the i+1th iteration k (0) Indicates b k The initial value of Indicates b k The value at the i+1th iteration; Step S325, obtaining β based on the variational Bayesian method k The approximate probability density function in the i+1th iteration is: in, Represents β k The value at the i+1th iteration, Represents the eighth auxiliary parameter K β The value at the i+1th iteration, Indicates the ninth auxiliary parameter P β The value at the i+1th iteration; Step S326, obtaining the expected value: Step S327, determine whether If satisfied, As the fusion state vector X of the mobile robot at time k k And output, As the covariance matrix P of the mobile robot at time k k|k And output, ε represents the iteration stop threshold. If it is not satisfied, set i=i+1 and return to execute steps S321 to S327.

2. The mobile robot visual-inertial fusion SLAM method based on variational Bayesian as claimed in claim 1, characterized in that: In step S1, the predicted state vector of the mobile robot at time k is: X k|k-1 =F k-1 X k-1||k-1 , where F k-1 represents the state vector conversion relationship between the mobile robot at time k-1 and time k obtained based on IMU data, X k-1||k-1 Represents the state vector of the mobile robot at time k-1; The one-step prediction error covariance matrix of the mobile robot at time k is: Among them, P k-1|k-1 represents the covariance matrix of the mobile robot at time k-1, Q k Represents the system noise covariance matrix of the mobile robot at time k.

3. The mobile robot visual-inertial fusion SLAM method based on variational Bayesian as claimed in claim 1, characterized in that: In step S3, when the joint posterior probability density function is jointly solved using the variational Bayesian method, k , k , k 、u k , β k The optimal solution satisfies: Among them, the collection θ represents the set Θ k Any element in Θ k (-θ) represents the set θ k The subset consisting of all elements except θ in C θ represents a constant that is independent of the element θ; E[·] represents the expected value; Y k Represents the visual state information of the mobile robot at time k.

4. A computer-readable storage medium, wherein at least one instruction, at least one program, a code set or an instruction set is stored in the storage medium, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement the variational Bayesian-based mobile robot visual-inertial fusion SLAM method as described in any one of claims 1 to 3.

5. A mobile robot, characterized in that: It includes a mobile robot body and a visual sensor, an IMU sensor and a processor arranged on the mobile robot body, the processor is connected to the visual sensor and the IMU sensor respectively, and the processor executes the steps of the mobile robot visual-inertial fusion SLAM method based on variational Bayes as described in one of claims 1-3.

Images