Indoor pedestrian navigation positioning method under multi-constraint condition

Abstract

The invention relates to an indoor pedestrian navigation positioning method under a multi-constraint condition, and belongs to the field of indoor pedestrian navigation positioning. According to the method, on the basis of a foot zero-speed correction algorithm, a Kalman filtering correction algorithm based on limb course change consistency is provided by combining a constraint condition that a difference value between a foot course variation and a waist course variation is relatively small in a movement process of each step; meanwhile, in consideration of stability of relative information output characteristics of building same-floor air pressure height measurement and stability constraint of a building geometric structure in an indoor environment, a constraint algorithm based on air pressure relative height change and an improved heuristic course drift elimination algorithm based on a building structure are provided. According to the invention, an improved hierarchical zero-speed interval detection algorithm is designed for stationary motion and non-stationary motion, and the recognition accuracy of a single zero-speed detection interval is improved. And meanwhile, through an improved multi-constraint condition algorithm, relatively high navigation and positioning precision can be provided for indoor pedestrians in a relatively long time, accumulated errors of pedestrian navigation and positioning are reduced, and a high-precision navigation and positioning function is provided for the pedestrians in a three-dimensional space environment.

Description

Technical Field

[0001] The present invention belongs to the field of indoor pedestrian navigation and positioning, and particularly relates to an improved indoor pedestrian navigation and positioning method under multiple constraint conditions. Technical Background

[0002] In recent years, with the development of society, pedestrian navigation technology has received increasing attention and research from researchers in the fields of disaster relief, medical search and rescue, public safety, fitness consumption, anti-terrorism, etc. Traditional pedestrian navigation and positioning methods mainly rely on the Global Navigation Satellite System (GNSS). However, in real life, pedestrians do not always stay in an environment with good GNSS signals. When pedestrians are in buildings, viaducts, forests with poor GNSS signals, or even in underground parking lots, indoors, tunnels, etc. without satellite signals, the Global Navigation Satellite System will not be able to provide accurate navigation and positioning functions for pedestrians. When performing special tasks indoors, pedestrians usually have multiple motion modes such as running, going upstairs and downstairs, which will affect the dynamic performance of wearable indoor pedestrian navigation methods. Especially in special application environments such as rescue, it is necessary to monitor the physical state of special operation personnel in real time for the command and dispatch center to perceive their motion modes and trajectories, so as to facilitate decision-makers to adjust task strategies in a timely manner according to the physical conditions and positions of the operation personnel. Therefore, it is urgent to carry out research on navigation methods to support indoor pedestrians in the satellite failure environment to meet the actual application needs.

[0003] In recent years, scholars have carried out research on indoor pedestrian navigation based on wearable inertial sensors (inertial measurement unit, IMU). Aiming at the high-precision navigation requirements of indoor pedestrians with inertial sensors as the core, the existing navigation and positioning technologies have problems such as large errors during long-term navigation and relatively relying on external active sensing information for navigation. Therefore, it is necessary to study the wearable inertial sensor network navigation technology for indoor pedestrians, so as to enhance the ability and stability of the indoor pedestrian navigation system and meet the requirements of long-term high-precision and high-reliability positioning of the indoor pedestrian navigation system in the indoor environment. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an improved indoor pedestrian navigation and positioning method under multiple constraint conditions in view of the deficiencies of the prior art. This method can meet the long-term high-precision navigation and positioning functions of indoor pedestrians with four motion modes of walking, running, going upstairs and downstairs, and improve the accuracy and reliability of the indoor pedestrian navigation system.

[0005] The technical solution to be solved by the present invention is achieved through the following technical solutions. The present invention is a method for indoor pedestrian navigation and positioning under multiple constraint conditions, characterized in that the steps are as follows: An inertial sensor is fixedly installed on the instep, both legs and the waist of the human body to form a four-node inertial sensor network, and the navigation and positioning method includes the following steps:

[0006] Step 1: Collect data of the four-node inertial sensor network on the instep, both legs and the waist of the human body; in the navigation and positioning system, set the motion type for the navigation and positioning system, and divide the pedestrian motion mode into two categories: steady motions including walking, going upstairs and going downstairs, and non-steady motions including running;

[0007] Step 2: Design an improved hierarchical zero-velocity interval detection algorithm for the two different motion types;

[0008] Step 3: Analyze the inertial sensor data of the pedestrian's motion, and use the quaternion method to calculate the position, velocity and attitude of the pedestrian;

[0009] Step 4: Establish a Kalman filtering algorithm based on zero-velocity correction, where the Kalman filter is used to estimate the position, velocity and attitude errors to correct the cumulative errors in the pedestrian navigation and positioning process;

[0010] Step 5: Establish an error correction algorithm based on the consistency of limb heading change. When the pedestrian's foot is in the zero-velocity interval, correct the heading error in the pedestrian navigation and positioning process;

[0011] Step 6: Establish a constraint algorithm based on the change of relative barometric height to correct the height error in the pedestrian navigation and positioning process;

[0012] Step 7: Establish an improved heuristic heading drift elimination algorithm based on the building structure. When the pedestrian moves along the dominant direction of the indoor building, correct the heading error in the pedestrian navigation and positioning process;

[0013] Step 8: Establish a Kalman filtering algorithm based on the inequality constraint of the maximum single-step length. Since the distance between two adjacent zero-velocity intervals of a single foot during the human body movement should not exceed the maximum step length, constrain and correct the position, velocity and attitude of the navigation system.

[0014] For the method for indoor pedestrian navigation and positioning under multiple constraint conditions described in the present invention, its further preferred technical solution is:

[0015] In the said Step 1, preset the motion mode for the pedestrian: steady motion or non-steady motion, and then collect the required data, mainly including foot acceleration, foot angular velocity, waist angular velocity and waist barometric pressure value;

[0016] In step 2, during a complete motion gait cycle of a pedestrian's movement, the motion state of a single foot is generally divided into the following four stages: stationary, foot leaving the ground stage, aerial swing, and foot landing stage; during the foot landing process, it enters the zero-velocity stage, where the speed approaches zero and the sum of the acceleration modulus values is approximately the acceleration due to gravity; the foot movement is simplified into two processes: the zero-velocity process and the non-zero-velocity process.

[0017] The following four judgment conditions are used to detect the zero-velocity interval: acceleration modulus value, angular velocity modulus value, acceleration mean square deviation, and angular velocity mean square deviation.

[0018]

[0019] In the formula: k is the sliding window size, f represents acceleration, ω represents angular velocity; x, y, z represent the three axes of the accelerometer and gyroscope sensors; ε a1 , ε a2 respectively represent the set acceleration modulus value threshold and acceleration mean square deviation threshold, and ε ω1 , ε ω2 respectively represent the set angular velocity modulus value threshold and angular velocity mean square deviation threshold.

[0020] For the two motion modes of input smooth motion and non-smooth motion, analyze the inertial data characteristics of the two major motions and set different zero-velocity detection methods respectively; when the motion mode is smooth motion, the zero-velocity detection comprehensive judgment scheme is as follows:

[0021]

[0022] When the motion mode is non-smooth motion, the following is used for zero-velocity discrimination:

[0023]

[0024] Continuously save the zero-velocity detection values Z j (k n-k+1 ),..., Z j (k n-1 ), Z j (k n ), where j = 1 or 2. When there are k consecutive points with zero-velocity detection equal to 1, it is considered that this is a real zero-velocity interval point, and the judgment formula is:

[0025]

[0026] For a method for indoor pedestrian navigation and positioning under multiple constraint conditions according to the present invention, a further preferred technical solution is: in step 3, the specific content of the pedestrian's position, speed, and attitude solution is:

[0027] Position solution differential equation:

[0028]

[0029] Velocity solution differential equation:

[0030]

[0031] Attitude solution differential equation:

[0032]

[0033] Where: L represents longitude, λ represents latitude, h represents the moving altitude, and R m represents the radius of curvature of each point on the ellipsoidal meridian; is the transformation matrix between the navigation system and the body system, f represents the specific force, g is the gravitational acceleration, w en =[0 0] T , v is the velocity vector in the three directions of "east-north-up", Λ is the quaternion matrix, θ is the pitch angle, γ is the roll angle, and ψ is the heading angle.

[0034] In the present invention, the "east-north-up" mentioned refers to the EN geographical coordinate system of east-north-up, that is: the X-axis points to the east, the Y-axis points to the north, and the Z-axis is perpendicular to the local horizontal plane and points upward along the local vertical line.

[0035] In the step 4, the Kalman filter algorithm model based on zero-velocity correction is established as:

[0036] Construct an 18-dimensional Kalman filter state variable, specifically as follows:

[0037]

[0038] Among them, is the platform angle error, δV E , δV N , δV U are the velocity errors, δL, δλ, δh are the position errors, ε bx , ε by , ε bz are the gyro random constants, ε rx , ε ry , ε rz are the gyro first-order Markov process random errors, is the accelerometer first-order Markov process random error;

[0039] The system state equation is:

[0040]

[0041] Wherein, A is the system matrix, X is the system state vector, G is the system noise matrix, and W is the system white noise; after the pedestrian's foot enters the zero-speed state, theoretically, it has the characteristics of zero speed and unchanged position, thus satisfying:

[0042] ΔV = V INS -V zupt = 0

[0043] ΔP = P INS -P zupt = 0

[0044] V zupt = [0 0 0] T

[0045] P zupt = P first

[0046] Based on the inertial / zero-speed integrated navigation system, which adopts the method of combining position and speed, the speed V INS and position P INS output by the inertial navigation solution at the current moment are subtracted from the speed V zupt and position P zupt corresponding to the zero-speed moment as the observation quantity, and P first is the position at the previous moment within the current zero-speed interval;

[0047] When the zero-speed interval is detected, the observation information is Z ZUPT (t), and the observation equation is:

[0048]

[0049] Wherein, N zupt (t) is the observation noise.

[0050] For a method for indoor pedestrian navigation and positioning under multiple constraint conditions described in the present invention, a further preferred technical solution is: in step 5, an error correction algorithm based on the consistency of limb heading change is established. For each step walked, the change amount Δψ waist of the waist heading angle and the change amount Δψ head of the foot heading angle are set to be equal, and the difference Δψ waist between Δψ head and Δψ 1 is used to construct a virtual observation quantity:

[0051] The observation information is:

[0052]

[0053] Wherein, respectively represent the average values of the waist heading angles at the current step and the previous step, respectively represent the average foot heading angles of the current step and the previous step; the observation equation of the error correction algorithm based on the consistency of limb heading changes is as follows:

[0054]

[0055] In the formula, ψ is the heading angle and θ is the pitch angle;

[0056] The observation equation of the combined system is:

[0057] Z WF (t) = [Δψ 1 == H WF (t)X(t) + N WF (t)

[0058] In the formula, N WF (t) is the observation noise.

[0059] For a method for indoor pedestrian navigation and positioning under multiple constraint conditions described in the present invention, a further preferred technical solution is: in step 6, a constraint algorithm based on the change of relative barometric height is established.

[0060] The observation equation after introducing barometric height is:

[0061] Z baro (t) = [(h baro - h baroinit ) - (h SINS - h SINSinit )] = H baro (t)X(t) + N baro (t)

[0062] In the formula, h baroinit is the initial height calculated by the barometer, h SINS is the height calculated in real time by pure inertia, and h SINSinit is the initial height of pure inertia;

[0063] The state equation is:

[0064] H baro (t) = [0 1×8 1 0 1×9

[0065] In step 7, an improved heuristic heading drift elimination algorithm based on building structure is established; record the heading angles of three consecutive steps of movement as ψ k-2 , ψ k-1 and ψ k , the average value of these three heading angles is ψ k,m , the variance is σ k,ψ , and there are the following judgment conditions:

[0066] ​

[0067] In the formula, TH ψ,m is the mean threshold for straight-line judgment; when the maximum difference between the heading angles in these three steps and their average value is less than the set threshold, it is considered that the current is in straight-line motion; otherwise, it is in non-straight-line motion;

[0068]

[0069] In the formula, TH σ is the variance threshold for straight-line judgment; when the variance of the heading angles in these three steps is less than the set threshold, it is considered that the current is in straight-line motion; otherwise, it is in non-straight-line motion;

[0070] When Z 4 (k) and Z 5 (k) are both detected as 1 at the same time, it means that the pedestrian is in straight-line motion:

[0071]

[0072] On the premise of determining that the pedestrian is currently in straight-line motion, in order to facilitate the judgment of whether the pedestrian is moving straight along the dominant direction, the current heading angle ψ k is projected into the interval through the remainder function mod, and the calculation formula is:

[0073]

[0074] When |ζ k | > TH ζ,k , it means that the pedestrian is currently moving straight along the dominant direction, and the heading angle is corrected for error; otherwise, the heading angle is not corrected; among them, TH ζ,k represents the judgment threshold angle for straight-line motion in the dominant direction;

[0075] When ζ k > 0, it means that the pedestrian is moving along the right side close to the dominant direction; when ζ k < 0, it means that the pedestrian is moving along the left side close to the dominant direction;

[0076] Every time a step is taken, the difference Δψ k between the current heading angle ψ 2 and the adjacent main heading angle is used to construct the heading angle observation quantity, and an improved heuristic heading drift elimination algorithm is proposed; the observation information is:

[0077]

[0078] In the formula, represents the main heading angle at the current moment;

[0079] The observation equation is:

[0080]

[0081] The state equation is:

[0082]

[0083] where N ψ,iHDE (t) is the observation noise.

[0084] For the indoor pedestrian navigation and positioning method under multiple constraint conditions according to the present invention, a further preferred technical solution is: in the step 8, aiming at the maximum value of the distance between two adjacent landings of a single foot, that is, the maximum step length d of human movement max , a correction method based on the maximum step length of a single foot with multiple constraints is designed; during the movement process, the distance d between two adjacent zero-velocity intervals k can be regarded as the step length of this step, where k represents the k-th step, and this distance should not exceed the maximum step length, and must satisfy d k ≤d max ; on the basis of zero-velocity correction, a single-foot distance inequality constraint Kalman filter system is constructed; this system constructs two sets of state quantities of adjacent zero-velocity intervals into a new 36-dimensional state quantity, and only corrects the state quantity of the current zero-velocity interval point, and then corrects the position, velocity and attitude at the current moment;

[0085] The system state quantity is as follows:

[0086] Y = [Y k-1 Y k

[0087] where Y k-1 and Y k are respectively:

[0088]

[0089]

[0090] The single-foot maximum step length inequality constraint Kalman filter navigation model is as follows:

[0091]

[0092] where is the system matrix, Y is the system state vector, is the system noise matrix, and W′ is the system white noise;

[0093] The single-foot distance error during the pedestrian movement process satisfies the following relationship:

[0094]

[0095] ​where \(g(Y)\) is the objective function, \(L\) s represents the distance constraint model, \(\delta d = d - d\) max , \(\delta d\) represents the single - foot step - length error, and \(d\) is the calculated single - foot step - length;

[0096]

[0097] P ic is a positive - definite projection matrix. Generally P m is the covariance matrix of;

[0098] When at time \(k\) it satisfies , it is considered that the single - foot step - length at this time is \(d\) max , and a state variable The re - organized filtering model constraint model is:

[0099]

[0100] In the formula, is the linearized single - foot step - length equation, and its expression is as follows:

[0101]

[0102] At this time, the state variable should satisfy the following constraint equations:

[0103]

[0104] In the formula, \(Z\) ic is the observed quantity, and the updated state variable and the error covariance matrix are expressed as follows:

[0105]

[0106] Among them:

[0107] \(\mu=P\) ic \(\kappa\) T (\(\kappa P\) ic \(\kappa\) T ) -1 \(\kappa\)

[0108] After obtaining the state variable and error covariance matrix of the current zero - velocity point under the re - organized filtering model, the navigation parameters such as position, velocity, and attitude are corrected.

[0109] A method for indoor pedestrian navigation and positioning under multiple constraint conditions according to the present invention, a further preferred technical solution thereof is: calculating LSTM network parameters based on SSA through training set data of the legs, and inputting test data of the leg IMU into the model for motion mode identification to provide a basis for subsequent navigation; optimizing the parameters of the LSTM algorithm: using the sparrow search algorithm to optimize the LSTM algorithm.

[0110] Virtual sparrows search for food. The mathematical expression of a sparrow population composed of q sparrows is:

[0111]

[0112] In the formula, s q,d where q is the sparrow number, d is the dimension of the variable to be optimized, and s q,d is the d-dimensional state quantity of the qth sparrow. In optimizing the LSTM classification, s q,d represents the d-dimensional eigenvalue in the qth step;

[0113]

[0114] In the formula, M S represents the fitness of all sparrows, and m([s q,1 s q,2 …s q,d ) represents the fitness value m of the qth sparrow s ; in optimizing the LSTM classification, m([s q,1 s q,2 ...s q,d ) represents all eigenvalues in the qth step;

[0115] The SSA optimizes the following four parameters of the LSTM network: the number of hidden layers, the maximum number of training epochs, the initial learning rate, and the L2 parameter; the fitness function formula for the sparrow search algorithm to optimize the long short-term memory network is:

[0116] fitness = E train + E test

[0117] In the formula, E train represents the training set error error rate, and E test represents the test set error error rate; the fitness function is the sum of the prediction error rates of the LSTM for the training set and the test set.

[0118] A method for indoor pedestrian navigation and positioning under multiple constraint conditions according to the present invention, a further preferred technical solution thereof is: the process of optimizing the LSTM algorithm based on SSA is as follows:

[0119] Step 1: Input the modal identification training set data, including multi-modal motion feature values and modal flag bits;

[0120] Step 2: Initialize the LSTM model;

[0121] Step 3: Calculate the initial fitness value;

[0122] Step 4: Initialize SSA, and update the parameters of the discoverer, joiner, and warning person;

[0123] Step 5: Iteratively optimize the global optimal fitness value;

[0124] Step 6: End the iterative optimization, and output the LSTM parameters optimized by SSA: the number of hidden layers, the maximum training period, the initial learning rate, and the L2 parameter.

[0125] Compared with the prior art, the method of the present invention has the following beneficial effects:

[0126] The improved indoor pedestrian navigation and positioning method under multiple constraints of the present invention, based on the Zero Velocity Update (ZUPT) algorithm, combines the constraint condition that the difference between the foot heading change amount and the waist heading change amount is relatively small in each movement process, and proposes a Kalman filter correction algorithm based on the consistency of limb heading changes; at the same time, considering that in the indoor environment, the relative information output characteristics of the same floor air pressure height measurement of the building are stable and the building geometric structure has stable constraints, a constraint algorithm based on the relative change of air pressure height and an improved heuristic drift elimination (iHDE) algorithm based on the building structure are proposed.

[0127] In addition, the present invention further improves the accuracy and reliability of the indoor pedestrian navigation system through the air pressure relative elevation constraint and the single-foot maximum step length inequality constraint. The long-term high-precision navigation and positioning function of indoor pedestrians is realized.

[0128] The present invention designs an improved hierarchical zero-speed interval detection algorithm for steady motion and non-steady motion, which improves the recognition accuracy of a single zero-speed detection interval. At the same time, through the improved multi-constraint condition algorithm, it can provide a relatively high navigation and positioning accuracy for indoor pedestrians for a long time, reduce the cumulative error of pedestrian navigation and positioning, and provide a high-precision navigation and positioning function for pedestrians in a three-dimensional space environment.

[0129] The present invention is applicable to pedestrian navigation in four motion modes, namely walking, going upstairs, going downstairs, and running. Four micro-inertial devices are used for navigation, which are portable and wearable, and have extremely high engineering application and commercial value. An improved hierarchical zero-velocity interval detection algorithm is designed for both steady and non-steady motions. The indoor pedestrian navigation method based on multiple constraint conditions corrects the errors of the multi-motion-mode pedestrian navigation and positioning system under the condition of no active information, and has high accuracy over a long period of time, thus having certain engineering value. BRIEF DESCRIPTION OF THE DRAWINGS

[0130] Figure 1 is the overall schematic block diagram of the improved indoor pedestrian navigation method under multiple constraint conditions of the present invention;

[0131] Figure 2 is the flow chart of optimizing LSTM parameters by SSA;

[0132] Figure 3 is the schematic diagram of the main motion heading angle;

[0133] Figure 4 is the flow chart of correcting the main heading of indoor pedestrians;

[0134] Figure 5 is the schematic diagram of the horizontal motion of a pedestrian's single foot;

[0135] Figure 6 is the schematic diagram of the appearance and installation configuration of the MTW inertial sensor; among them, the appearance of the sensor and the direction of the body system are as Figure 6 (a) shown; the IMU is installed on the foot, leg, and waist, and the wearing positions are as Figure 6 (b) shown;

[0136] Figure 7 is the schematic diagram of the indoor three-dimensional trajectory and two-dimensional projection;

[0137] Figure 8 is the two-dimensional navigation result of pure zero-velocity correction based on motion mode identification;

[0138] Figure 9 is the comparison diagram of the two-dimensional navigation results under the zero-velocity correction / heading change / iHDE constraint conditions based on SSA / LSTM; among them: (a) the combined navigation result based on ZUPT / WF; (b) the combined navigation result based on ZUPT / iHDE; (c) the combined navigation result based on ZUPT / WF / iHDE;

[0139] Figure 10 is the comparison diagram of the navigation results under different mode identification algorithms; among them: (a) the navigation result based on the DCNN algorithm; (b) the navigation result based on the PCA / SVM algorithm; (c) the navigation result based on the LSTM algorithm;

[0140] Figure 11 3D navigation result graph of barometric-aided correction under zero velocity correction / heading change / iHDE constraints based on SSA / LSTM, where (a) has barometric-aided correction and (b) has no barometric-aided correction;

[0141] Figure 12 It is a comparison graph of navigation results under different modal identification algorithms. Specific implementation manners

[0142] The implementation process of the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0143] Example 1, referring to Figure 1 -2,

[0144] In the present invention, the relevant definitions are described as follows:

[0145] Global Navigation Satellite System (GNSS);

[0146] Inertial Measurement Unit (IMU);

[0147] Zero Velocity Update (ZUPT);

[0148] Improved Heuristic Drift Elimination (iHDE);

[0149] Sparrow Search Algorithm (SSA)

[0150] Deep Convolution Neural Network (DCNN);

[0151] Principal Component Analysis (PCA);

[0152] Support Vector Machine (SVM);

[0153] Long Short Term Memory (LSTM);

[0154] Root Mean Squared Error (RMSE).

[0155] Waist-foot (WF)-based limb heading change consistency algorithm.

[0156] Based on the indoor pedestrian navigation method of the multi-node inertial sensing network (recorded in Zhengchun Wang, ZhiXiong, Li Xing, et al. A Method for Autonomous Multi-Motion Modes Recognition and Navigation Optimization for Indoor Pedestrian [J]. Sensors, 2022, 22(13), 5022.), since the parameters of the LSTM algorithm (recorded in Ding C, Jia Y, Cui G, et al. Continuous Human Activity Recognition through Parallelism LSTM with Multi-Frequency Spectrograms [J]. Remote Sensing, 2021, 13(21): 4264.) are mainly set according to empirical values or multiple experiments, and these parameters may not necessarily reflect the optimal effect of the data set. In order to improve the accuracy of the LSTM algorithm for motion mode identification, it is necessary to optimize the parameters of the LSTM algorithm; in addition, for the pedestrian navigation algorithm without other active sensing information assistance indoors, relying solely on the ZUPT algorithm (recorded in Elwell J. Inertial navigation for the urban warrior [C] / / Digitization of the Battle space IV. SPIE, 1999, 3709: 196-204.), the navigation error is still prone to divergence during long-term navigation. Therefore, an improved indoor pedestrian navigation method with multiple constraint conditions based on zero velocity update / heading change / iHDE / barometric change assistance / single-foot maximum step length inequality constraint is proposed. The framework of the designed indoor pedestrian navigation method is shown in Figure 1 .

[0157] The inventor proposed the following algorithm: First, based on indoor pedestrian navigation in a multi-node inertial sensing network, the parameters of the LSTM network based on SSA are calculated through the training set data of the legs, and the test data of the leg IMU is input into the model for motion mode identification, providing a basis for subsequent navigation. Second, in an indoor environment, based on the limb heading change consistency correction algorithm based on the ZUPT / WF principle, taking advantage of the characteristics of stable output features of relative building same-floor air pressure height measurement and stable constraints of the building geometric structure, a constraint algorithm based on relative air pressure height change and an improved heuristic heading drift elimination algorithm based on the building structure are proposed. Finally, according to the gait characteristics of pedestrians during movement, combined with the single-foot step length of adjacent steps, a navigation method based on the inequality constraint Kalman filter of the single-foot maximum step length is proposed to further correct the pedestrian navigation system.

[0158] 1 Algorithm Description and System Model

[0159] 1.1 Long Short-Term Memory Method Optimized by Sparrow Search Algorithm

[0160] Generally, the parameters of the LSTM algorithm are set according to empirical values or multiple experiments, and these parameters may not necessarily reflect the optimal effect of the data set. To improve the accuracy of the LSTM algorithm for motion mode identification, it is necessary to optimize the parameters of the LSTM algorithm. In recent years, swarm intelligence is an algorithm or distributed problem-solving method inspired by the social behavior mechanisms of arbitrary animal groups, which is a parameter optimization method with high stability and fast convergence speed. The swarm intelligence optimization algorithm simulates the foraging or other behaviors of animal groups such as bird flocks, fish schools, and insects. These animal groups search for food in a certain cooperative manner, and each member in the group continuously exchanges information, learns the experience of itself and other members in the group, and continuously adjusts and changes the search direction.

[0161] The sparrow search algorithm was proposed by Xue Jiankai in 2020. This algorithm is inspired by the foraging behavior and risk avoidance behavior of sparrows and solves the global optimization problem by simulating some behaviors of sparrows. Compared with other optimization algorithms, it has the advantages of strong optimization ability, short iteration time, and high prediction accuracy. Therefore, the inventor selects the sparrow search algorithm to optimize the LSTM algorithm.

[0162] In the experiment of simulating the sparrow search algorithm, virtual sparrows search for food. The mathematical expression of the sparrow population composed of q sparrows is:

[0163]

[0164] In the formula, s q,d where q is the sparrow number, d is the dimension of the variable to be optimized, s q,dThat is, the d-dimensional state quantity of the q-th sparrow.

[0165] In optimizing the LSTM classification, s q,d represents the d-dimensional eigenvalue in the q-th step.

[0166]

[0167] In the formula, M S represents the fitness of all sparrows, and m([s q,1 s q,2 ... s q,d ) represents the fitness value m of the q-th sparrow s . In optimizing the LSTM classification, m([s q,1 s q,2 ... s q,d ) represents all eigenvalues in the q-th step.

[0168] The entire foraging sparrow population is divided into three categories: discoverers, joiners, and early warning birds. The identities of the discoverers and joiners can change dynamically, but the proportions of the discoverers and joiners in the entire population are fixed. The role of the discoverers is to search for areas with abundant food and provide foraging directions and areas for the entire population; the joiners obtain food based on the information provided by the discoverers; when the sparrow population encounters a predator or other danger, the early warning birds send out alarm signals, and when the signal exceeds the threshold, the discoverers will lead the population to other safe foraging areas.

[0169] The present invention selects the following four parameters of the SSA-optimized LSTM network: the number of hidden layers, the maximum number of training epochs, the initial learning rate, and the L2 parameter. The fitness function formula for designing the sparrow search algorithm to optimize the long short-term memory network is

[0170] fitness = E train +E test (3)

[0171] In the formula, E train represents the training set error rate, and E test represents the test set error rate. The fitness function is the sum of the prediction error rates of the LSTM for the training set and the test set. The lower the error rate, the better the LSTM model. Based on the above modeling, the flowchart for optimizing the LSTM parameters based on the SSA is as Figure 2 , and the algorithm process is as follows:

[0172] Step 1: Input the modal identification training set data, including multi-modal motion eigenvalues and modal flag bits;

[0173] Step 2: Initialize the LSTM model;

[0174] Step 3: Calculate the initial fitness value fitness ini ;

[0175] Step 4: Initialize SSA, and update the parameters of the discoverer, joiner, and warning detector;

[0176] Step 5: Iteratively optimize the global optimal fitness value fitness opt ;

[0177] Step 6: End the iterative optimization, and output the LSTM parameters optimized by SSA: the number of hidden layers, the maximum number of training epochs, the initial learning rate, and the L2 parameter.

[0178] 1.2 Improved Heuristic Heading Drift Elimination Algorithm Based on Building Structure

[0179] Considering that the heading error is the main factor directly affecting the navigation accuracy, and relying solely on the constraints of human motion characteristics cannot meet the long-term high-precision requirements. Therefore, aiming at the characteristic that the heading angle of the indoor pedestrian inertial navigation system is prone to divergence, an improved heuristic heading drift elimination algorithm (iHDE) is studied according to the indoor building structure. By classifying the pedestrian motion trajectory through the straight-line judgment and main path judgment during the pedestrian motion process, the straight-line judgment condition is added, thereby reducing the misdetection of straight-line motion, and different heading correction methods are adopted for different types of trajectories to avoid over-correcting the heading angle, improving the robustness of the heading correction algorithm.

[0180] The method of the present invention classifies the indoor motion trajectory of pedestrians into two categories: straight-line motion and non-straight-line motion. Among them, the straight-line motion is divided into straight-line motion in the dominant direction and straight-line motion in the non-dominant direction. The straight-line motion in the dominant direction is corrected by feeding back the difference between the heading angle obtained by inertial solution and the main heading angle into the navigation system to correct the pedestrian heading error.

[0181] (1) Setting of the dominant direction

[0182] Generally, based on the characteristics of the indoor building with four orientations of east, west, south, and north, the dominant direction is set to the following four reference angles: 0°, 90°, 180°, 270°, as the main heading of pedestrian motion, and the angle interval Δ h is set to 90°, as Figure 3 shown. If too many dominant directions are set, it is easy to cause confusion and lead to incorrect correction of the navigation system.

[0183] (2) Judgment of straight-line motion

[0184] Since there is a certain error in the single-step heading calculated when a person is moving, relying solely on the matching of the single-step heading and the dominant direction is likely to result in a large error. Therefore, based on practical experience, the average value of the heading angles of multiple consecutive movement steps is usually used as the matching value. The present invention selects continuous multi-step movement as the basis for determining straight-line movement. When the current movement belongs to straight-line movement, the heading angle is used as a reference value to correct the heading error of the navigation system.

[0185] (3) Navigation modeling based on iHDE

[0186] This part of the content first studies the method for determining the straight-line movement of pedestrians. Based on the straight-line movement of pedestrians, the determination of the main heading and the correction of the heading angle error are carried out. The flow chart is as Figure 4 shown.

[0187] a) Determination of straight-line movement

[0188] The traditional HDE algorithm determines whether it is going straight by calculating the average value of the heading angles of consecutive m steps. However, a single judgment condition is prone to misjudgment. The inventor proposes an improved method based on the combination detection of the average value and variance of the heading angles of consecutive m steps, which reduces the misjudgment probability of straight-line movement.

[0189] From the empirical information, m is usually taken as 3. Denote the heading angles of three consecutive steps of movement as ψ k-2 、ψ k-1 and ψ k , the average value of these three heading angles is ψ k,m , and the variance is σ k,ψ . There are the following judgment conditions:

[0190]

[0191] In the formula, TH ψ,m is the mean threshold for straight-line judgment. When the maximum value of the difference between the heading angle and its average value in these three steps is less than the set threshold, it is considered that the current is in straight-line movement; otherwise, it is in non-straight-line movement.

[0192]

[0193] In the formula, TH σ is the variance threshold for straight-line judgment. When the variance of the heading angles in these three steps is less than the set threshold, it is considered that the current is in straight-line movement; otherwise, it is in non-straight-line movement.

[0194] Through analysis, it is proposed that when Z 1 (k) and Z 2 (k) are both detected as 1, it indicates that the pedestrian is in straight-line movement.

[0195]

[0196] b) Determination of the dominant direction

[0197] On the premise that it is determined that the pedestrian is currently moving straight, in order to facilitate the determination of whether the pedestrian is moving straight along the dominant direction, the current heading angle ψ k is projected onto the interval through the modulo function mod, and the calculation formula is

[0198]

[0199] When ζ k | > TH ζ,k it indicates that the pedestrian is currently moving straight along the dominant direction, and the heading angle is corrected for errors. Otherwise, the heading angle is not corrected. Among them, TH ζ,k represents the judgment threshold angle for straight movement along the dominant direction.

[0200] When ζ k > 0, it indicates that the pedestrian is moving along the right side close to the dominant direction; when ζ k < 0, it indicates that the pedestrian is moving along the left side close to the dominant direction.

[0201] c) Navigation modeling for movement along the dominant direction based on straight movement

[0202] For each step taken, the difference Δψ k between the current heading angle ψ and the mean value of the adjacent main heading angles 2 is used to construct the heading angle observation quantity, and an improved heuristic heading drift elimination algorithm is proposed. The observation information is shown in formula (8).

[0203]

[0204] In the formula, represents the main heading angle at the current moment.

[0205] The observation equation is shown in formula (9).

[0206]

[0207] The state equation is shown in formula (10).

[0208]

[0209] In the formula, θ is the pitch angle.

[0210] To sum up, the observation equation of the combined system is shown in formula (11).

[0211]

[0212] 1.3 Relative elevation constraint algorithm based on air pressure

[0213] In order to accurately complete indoor rescue tasks, the height position information of the pedestrian's location plays an important role in the indoor system. Therefore, it is necessary to perform real-time positioning and tracking of the pedestrian's three-dimensional position.

[0214] Under normal circumstances, the height relative to sea level is less than 11,000 meters. According to the calculation relationship between air pressure and height, the calculation formula for air pressure and height is:

[0215]

[0216] In the formula, p is the measured air pressure value at the current position, h baro is the calculated altitude, in meters.

[0217] The measured value of the barometer is relatively stable. Using the air pressure value output by the barometer, the altitude can be obtained conveniently and quickly. The error of measuring the air pressure value will not accumulate over time, and the cost of the barometer is low. However, the output value of the barometer is affected by environmental factors such as wind speed, temperature, humidity, longitude and latitude. The air pressure value output by the barometer may not necessarily reflect the true value.

[0218] During the navigation of indoor pedestrians, as long as they know their relative height, they can successfully complete the operation tasks. The barometer can provide reference information on the relative height for pedestrian navigation and can suppress the divergence of the height channel. Therefore, the present invention selects the barometric differential altimetry for height assistance in the pedestrian navigation system.

[0219] By real-time air pressure calculation of height h baro and the air pressure height h init at the starting point, the relative height information Δh can be obtained by taking the difference.

[0220] Δh = h baro - h init (13)

[0221] The observation equation after introducing the air pressure height is shown in Equation (14).

[0222] Z baro (t) = [(h baro - h baroinit ) - (h SINS - h SINSinit )] = H baro (t)X(t) + N baro (t) (14)

[0223] In the formula, h baroinit is the initial height calculated by the barometer, h SINS is the height calculated by pure inertial real-time solution, and h SINSinit is the pure inertial initial height.

[0224] The state equation is shown in Equation (15).

[0225] H baro (t)=[0 1×8 10 1×9 (15)

[0226] 1.4 Navigation method based on inequality-constrained Kalman filter using the maximum single-foot step length

[0227] For the above-mentioned foot-mounted inertial navigation algorithm assisted by zero-velocity correction / heading change / iHDE / barometric pressure change, the error is mainly constrained from the perspective of single-foot measurement features, and the heading of the pedestrian navigation system is corrected to a certain extent. However, there is a problem that the position error obtained by inertial navigation calculation is too large, resulting in an unreasonable situation where the single-foot step length exceeds the normal range.

[0228] Therefore, in order to further increase the constraint information, considering that there is a maximum value for the distance between two consecutive landings of a single foot, that is, the maximum step length d max of human movement, a correction method based on multiple constraints of the maximum single-foot step length is designed. During the movement process, the distance d k between two adjacent zero-velocity intervals can be regarded as the step length of this step, where k represents the kth step. This distance should not exceed the maximum step length and must satisfy d k ≤ d max . Therefore, based on zero-velocity correction, the present invention constructs a single-foot distance inequality-constrained Kalman filter system to constrain and correct the position, velocity, and attitude of the navigation system, further improving the accuracy of the pedestrian navigation system.

[0229] (1) Principle of single-foot maximum step length constraint

[0230] Figure 5 The following figure shows a schematic diagram of the constraint model of the right foot in the stepping state. In the horizontal xoy plane, point A(p k-1,x , p k-1,y ) is a point in the previous zero-velocity interval, and point B(p k,x , p k,y ) is a point in the current zero-velocity interval. The relative position of two consecutive steps of the right foot is constrained within a circle with point A as the center and d max as the radius. When d k > d max , that is, point B is outside the circle, this method is needed to constrain and correct the navigation parameters of point B to obtain a new position coordinate point B'.

[0231] (2) System state model

[0232] The system constructs two sets of state variables in adjacent zero-velocity intervals into a new 36-dimensional state variable, only correcting the state variable at the current zero-velocity interval point, and then correcting the position, velocity, and attitude at the current moment.

[0233] The system state variables are as follows:

[0234] Y = [Y k-1 Y k (16)

[0235] In the formula, Y k-1 and Y k Refer to formulas (17) and (18), and are respectively

[0236]

[0237] The single-foot maximum step length inequality constraint Kalman filter navigation model is as follows:

[0238]

[0239] In the formula, A′ is the system matrix, Y is the system state vector, G′ is the system noise matrix, and W′ is the system white noise.

[0240] (3) Single-foot maximum step length inequality constraint measurement model

[0241] The single-foot distance error during the pedestrian movement process satisfies the following relationship:

[0242]

[0243] In the formula, g(Y) is the objective function, and L s represents the distance constraint model. For δd = d - d max , δd represents the single-foot step length error, and d is the single-foot step length obtained by calculation.

[0244]

[0245] P ic is a positive definite projection matrix, and generally where P m is the covariance matrix of.

[0246] When the kth moment satisfies it is considered that the single-foot step length at this time is d max , so a new state variable is obtained The constraint model of the recombined filter model is

[0247]

[0248] In the formula, is The linearized single-footstep equation has the following expression:

[0249]

[0250] The state variables at this time should satisfy the following constraint equations:

[0251]

[0252] In the formula, Z ic is the observed quantity. Solving equation (26) gives the updated state variables and the error covariance matrix which are expressed as follows:

[0253]

[0254] where

[0255] μ = P ic κ T (κP ic κ T ) -1 κ (29)

[0256] After obtaining the state variables and the error covariance matrix of the current zero-velocity point under the reorganized filtering model, the navigation parameters such as position, velocity, and attitude are corrected.

[0257] 2. Verification and analysis 2

[0258] 2.1 Experimental conditions

[0259] To verify the effectiveness of the navigation method proposed in the present invention, its effect is verified through experiments. This experiment uses the MTW inertial sensor set of Xsens Company in the Netherlands. The appearance of the sensor and the direction of the body coordinate system are as shown in Figure 6 (a). The IMUs are installed on the foot, leg, and waist, and the wearing positions are as shown in Figure 6 (b).

[0260] The sampling frequencies of both the accelerometer and the gyroscope are 100 Hz, and the sampling frequency of the barometer is 50 Hz. The mass of a single sensor is 27 grams. The performance parameters of the MTW are shown in Table 1.

[0261] Table 1 MTW sensor parameters

[0262] / Range Bias stability Accelerometer <![CDATA[±160m / s 2 > / Gyroscope ±1200deg / s 20degh Barometer 300 - 1000mBar 100Pa / year

[0263] 2.2 Experimental design

[0264] The experimental site is selected in Buildings 1 and 2 of the School of Automation of a certain university, and the two buildings are internally connected.

[0265] The reference position coordinate diagram selected for the experiment is as shown in Figure 7 Figure. Starting from the coordinate point (0, 0), the overall direction is counterclockwise movement. When the experiment ends, the experimenter returns to the starting point. The specific experimental content is as follows: The starting point is set at the position ① (0, 0) on the fifth floor of Building 1. Move two circles counterclockwise. The second circle is for running at ② - ③. Go downstairs from position ⑤ to position ⑥ on the third floor. Move from ⑥ to ⑦, and then move two circles counterclockwise, where running is done at ⑧ - ⑨. Go upstairs from position ⑥ on the third floor to position ⑤ on the fifth floor. Move from ⑤ to ①, and then move one circle counterclockwise back to position ① on the fifth floor. The unmentioned movement is in the walking mode. The whole movement takes about 21 minutes and 2 seconds.

[0266] To evaluate the positioning error, project the movement trajectory onto a two - dimensional plane. Set the coordinates of the actual physical lengths at four turning points on the experimental path (0, 0), (56.05, 0), (56.05, 47.81), (0, 47.81) as reference points to calculate the error. The total length of the four straight lines in the corridor is 207.72 meters, and the vertical distance from the third - floor floor to the fifth - floor floor is 7.54 meters. The initial relative height of the experiment is set to 0 meters.

[0267] Evaluating the navigation error includes two points:

[0268] (1) Start - end point error: It refers to that for a closed movement trajectory, the starting point and the end point should theoretically coincide, representing the distance deviation between the starting point and the end point at the same position.

[0269] (2) Root Mean Squared Error (RMSE): As a further criterion for evaluating the error of the pedestrian navigation system, using the four points ①②③④ in Figure 7 (b) as reference points, the calculation method of RMSE is formula (30):

[0270]

[0271] In the formula, n is 20, X obs,i is the position estimated by the algorithm, X ref,i is the reference point position, and RMSE is the final positioning error.

[0272] 2.3 Experimental Results and Analysis

[0273] (1) Experimental verification of the modal identification algorithm of the long - short - term memory network optimized by the sparrow search algorithm

[0274] When counting statistics, taking the number of right - foot movements as the benchmark, each swing is counted as one step, and there is a zero - speed interval between two steps. During walking, going upstairs, going downstairs, and running, the actual number of steps is 1106, and the algorithm detects 1110 steps. The data of the zero - speed interval is reduced by half compared with the number of steps in modal identification, and the comprehensive recognition rate of the number of movement steps is 99.64%.

[0275] The statistical results of the multi - motion modal identification algorithm are shown in Table 2, and the number of single - leg movement steps identified is shown in Table 2. Since the proportion of the walking part is relatively high among the four motions, the gap in the overall recognition rate is relatively small. Generally speaking, the motion modal identification accuracy of the method based on SSA - optimized LSTM proposed in this invention is the highest.

[0276] Table 2 Statistical table of results based on multi - motion modal identification

[0277]

[0278]

[0279] (2) Experimental verification of navigation error under the zero - speed correction method based on motion modal identification

[0280] Figure 8 The two - dimensional projection of the pedestrian navigation result based on pure zero - speed correction of motion modal identification is shown. It can be seen that after multiple behaviors of entering and leaving rooms and turning, the navigation result is not ideal.

[0281] (3) Experimental verification of navigation error under the zero - speed correction / heading change / iHDE / barometric pressure change method based on motion modal identification

[0282] ① Algorithm verification of zero - speed correction / heading change / iHDE constraint based on SSA / LSTM

[0283] Figure 9 The two - dimensional projection diagram of the navigation result under the zero - speed correction / heading change / iHDE constraint based on SSA / LSTM is shown. Figure 9 (a) is the combined navigation result based on ZUPT / WF. Compared with the navigation result based on pure inertia, the heading divergence is converged to a certain extent, but there is still a large gap from the real motion trajectory; Figure 9 (b) is the combined navigation result based on ZUPT / iHDE. The heading of the straight - line motion along the dominant direction of the indoor building is well corrected, but the position information deviates from the real trajectory; Figure 9 (c) is the combined navigation result based on ZUPT / WF / iHDE. The navigation result is improved on the basis of the two methods of ZUPT / WF and ZUPT / iHDE, but there is still a certain deficiency in the position information.

[0284] Table 3 is a statistical table of two-dimensional navigation results under zero-speed correction / heading change / iHDE constraints based on SSA / LSTM. It can be seen that the error of the ZUPT / WF / iHDE combined algorithm is the smallest.

[0285] Table 3 Error Table of Two-Dimensional Navigation Results under Zero-Speed Correction / Heading Change / iHDE Constraints Based on SSA / LSTM

[0286]

[0287] ② Algorithm verification under zero-speed correction / heading change / iHDE constraints based on different motion mode identification methods. The navigation results of the algorithms based on DCNN (recorded in Ronao C A, Cho S B. Human activity recognition with smartphone sensors using deep learning neural networks[J]. Expert systems with applications, 2016, 59: 235 - 244.) and PCA / SVM (recorded in Zhou B, Wang H, Hu F, et al. Accurate recognition of lower limb ambulation mode based on surface electromyography and motion data using machine learning[J]. Computer Methods and Programs in Biomedicine, 2020, 193: 105486.) are compared as Figure 10 , Table 4 is a statistical table of navigation errors under different mode identification algorithms. The navigation results based on the DCNN algorithm are as Figure 10 shown, and the navigation results based on the PCA / SVM algorithm are as Figure 10 (b) shown. The error of the SSA / LSTM mode identification algorithm proposed in the present invention is lower than the other two algorithms.

[0288] Table 4 Statistical Table of Navigation Errors under Different Mode Identification Algorithms

[0289]

[0290]

[0291] ③Verification of the navigation algorithm for zero-speed correction / heading change / iHDE condition with barometric pressure assisted correction based on SSA / LSTM motion mode identification

[0292] Figure 11 It is a three-dimensional navigation result diagram with or without barometric pressure assisted correction under the constraints of zero-speed correction / heading change / iHDE based on SSA / LSTM. Figure 11 (a) Compared with (b), in the elevation direction, the navigation data obtained by inertial solution will gradually diverge without the constraint of other information. When there is barometric pressure assistance, the position information in the elevation direction is relatively stable, and the calculated relative height between the fifth floor and the third floor is 7.32 meters; when there is no barometric pressure assistance, the height divergence is more obvious, and the calculated relative height between the fifth floor and the third floor is 22.90 meters. The actual relative height of the fifth floor relative to the third floor is 7.54 meters. The relative height obtained by the barometric pressure assisted solution adopted by the present invention is closer to the true relative height and can better assist in correcting the elevation information.

[0293] (4) Experimental verification of the multi-constraint integrated navigation method based on mode identification

[0294] Compare the experimental results of the multi-constraint integrated navigation method based on mode identification proposed in the present invention with the bipedal distance inequality constraint method based on mode identification. The two-dimensional projection results of the two navigation methods are as follows Figure 12 shown, among which, the two-dimensional projection result of the multi-constraint integrated navigation method based on mode identification proposed in the present invention is as follows Figure 12 shown in (a), and the error statistics of the positioning result are shown in Table 5. The error from the starting point to the ending point is 3.83 meters. The two-dimensional projection result of the bipedal distance inequality constraint method based on mode identification is as follows Figure 12 shown in (b), and the error from the starting point to the ending point is 4.89 meters. The error of the multi-constraint method based on mode identification proposed in the present invention is lower than that of the bipedal inequality constraint method based on mode identification. Although the inertial sensors used in the bipedal distance inequality are more than those of a single foot, the operation result is essentially the mean of two inertial sensors in a certain form. On the basis of simplifying the sensor configuration, the single-foot multi-constraint integrated navigation effect based on mode identification is better than the bipedal inequality effect based on mode identification.

[0295] Table 5 Error statistics table of different constraint algorithms

[0296]

[0297] Through the above experimental verification and comparative analysis, the effectiveness of the navigation method under multi-constraint conditions based on motion mode identification for indoor pedestrians proposed in the present invention is fully verified, and it can maintain a high navigation accuracy for a long time.

[0298] 3. Discussion

[0299] The present invention proposes an improved multi - motion - modality identification and navigation optimization method applicable to indoor pedestrians. For the four common indoor motion modalities of walking, running, going upstairs, and going downstairs, in the case of the LSTM - based motion - modality identification algorithm, since the parameter settings of LSTM are empirical values and cannot fully and accurately reflect the LSTM network, resulting in the identification accuracy still needing to be improved. Therefore, the present invention proposes an SSA - based LSTM motion - modality identification algorithm. This algorithm uses the training set data of the motion modality and optimizes the parameters of the LSTM network by means of the SSA algorithm, improving the identification accuracy of the motion modality. For the method of correcting pedestrian navigation without relying on external sensing information sources, for the ZUPT / WF algorithm based on the consistency of limb - heading change, during long - term navigation, due to the absence of external heading information for correction, the pedestrian's heading gradually drifts over time, ultimately leading to the gradual divergence of the navigation error. Therefore, the present invention proposes an improved pedestrian navigation method under multiple constraint conditions. This method studies an improved heuristic heading - drift elimination algorithm based on the building structure according to the stable characteristics of the geometric structure of indoor buildings; at the same time, since the relative information output characteristics of the air - pressure height measurement on the same floor of the building are stable, a constraint algorithm based on the change of relative air - pressure height is studied; in addition, based on the gait characteristics of pedestrians during movement and combining the single - foot step length of adjacent steps, a navigation method based on the inequality constraint of the single - foot maximum step length and Kalman filtering is proposed to further constrain the navigation parameters of pedestrians.

[0300] Through experimental data verification and analysis, the improved multi - motion - modality identification and indoor pedestrian navigation method proposed by the present invention achieves a high identification accuracy of pedestrian motion modalities, and in an indoor environment of about 2600m 2 under the condition that the total distance exceeds 1038.6m, the navigation effect is significantly improved compared with the ZUPT / WF method based on motion - modality identification and the biped - distance inequality constraint method. It can provide high navigation accuracy for indoor pedestrians for a long time, which is of great significance for the practical application of pedestrian inertial navigation.

[0301] The present invention proposes an improved indoor pedestrian navigation method based on modality identification and multiple constraint conditions. This method optimizes the parameters of the LSTM network through SSA, improving the identification accuracy of indoor pedestrian motion modalities. On this basis, under the condition of correcting pedestrian navigation parameters without relying on external sensing information sources indoors, a multi - constraint pedestrian navigation method based on the combination of ZUPT / WF / iHDE / air - pressure change assistance and the inequality constraint of the single - foot maximum step length is proposed, realizing the function of long - term high - precision navigation for indoor pedestrians.

Claims

1. A method for indoor pedestrian navigation and positioning under multiple constraints, characterized in that: The steps are as follows: inertial sensors are fixedly installed on the human foot, legs and waist to form a four-node inertial sensor network. The navigation and positioning method includes the following steps: Step 1: Collect data from the four-node inertial sensor network of the human foot, legs and waist; in the navigation and positioning system, set the motion type for the navigation and positioning system, and divide the pedestrian motion mode into two categories: stable motion including walking, going upstairs and going downstairs, and non-steady motion including running; Step 2: Design an improved hierarchical zero-speed interval detection algorithm for two different motion types; Step 3: Analyze the inertial sensor data of pedestrian movement and use the quaternion method to solve the pedestrian's position, speed and posture; Step 4: Establish a Kalman filter algorithm based on zero-speed correction, in which the Kalman filter is used to estimate the position, speed, and attitude errors, and correct the accumulated errors in the pedestrian navigation and positioning process; Step 5: Establish an error correction algorithm based on the consistency of limb heading changes. When the pedestrian's feet are in the zero-speed range, correct the heading error in the pedestrian's navigation and positioning process; Step 6: Establish a constraint algorithm based on the relative height change of air pressure to correct the height error in the pedestrian navigation and positioning process; Step 7: Establish an improved heuristic heading drift elimination algorithm based on building structure to correct the heading error in the pedestrian navigation and positioning process when the pedestrian moves along the dominant direction of the indoor building; Step 8: Establish a Kalman filter algorithm based on the maximum step length inequality constraint of a single foot. Since the distance between two adjacent zero-speed intervals of a single foot during human movement should not exceed the maximum step length, the position, velocity and attitude of the navigation system are constrained and corrected.

2. According to the method for indoor pedestrian navigation and positioning under multiple constraints of claim 1, it is characterized by: In the step 1, a motion mode is pre-set for the pedestrian: steady motion or non-steady motion, and then required data is collected, mainly including foot acceleration, foot angular velocity, waist angular velocity and waist air pressure value; In step 2, during a complete gait cycle of a pedestrian's movement, the movement state of a single foot is generally divided into the following four stages: stationary, foot-off-the-ground stage, swinging in the air, and foot-on-the-ground stage; during the foot-on-the-ground stage, the foot enters the zero-speed stage, the speed is close to zero, and the sum of the acceleration moduli is approximately the gravity acceleration; the foot movement is simplified into two processes: a zero-speed process and a non-zero-speed process; The zero-speed interval is detected using the following four judgment conditions: acceleration modulus, angular velocity modulus, acceleration mean square error, and angular velocity mean square error: Where: k is the sliding window size, f represents acceleration, ω represents angular velocity; x, y, z represent the three axes of the accelerometer and gyroscope sensor; ε a1 , ε a2 They represent the acceleration modulus threshold and acceleration mean square error threshold respectively. ω1 , ε ω2 Respectively represent the set angular velocity modulus threshold and angular velocity mean square error threshold; According to the two motion modes of input, steady motion and non-steady motion, the inertial data characteristics of the two types of motion are analyzed, and different zero-speed detection methods are set respectively; when the motion mode is steady motion, the comprehensive judgment scheme of zero-speed detection is as follows: When the motion mode is non-stationary motion, the following is used to determine the zero speed: Continuously save zero speed detection value Z j (k n-k+1 ),...,Z j (k n-1 ),Z j (k n ), j = 1or2, when there are k consecutive points with zero speed detection of 1, it is considered to be a real zero speed interval point, and the judgment formula is:

3. The method for indoor pedestrian navigation and positioning under multiple constraints according to claim 1, characterized in that: In step 3, the specific contents of the pedestrian position, speed and posture calculation are as follows: Solve the differential equation for the position: Solving the differential equation for velocity: Attitude solving differential equations: Where: L represents longitude, λ represents latitude, h represents movement altitude, R m Represents the radius of curvature of each point on the ellipsoid meridian; is the conversion matrix between the navigation system and the aircraft system, f represents the specific force, g is the gravitational acceleration, w en =[0 0 0] T , v is the velocity vector in the three directions of "northeast sky"; Λ is the quaternion matrix, θ is the pitch angle, γ is the roll angle, and ψ is the heading angle.

4. The method for indoor pedestrian navigation and positioning under multiple constraints according to claim 1, characterized in that: In step 4, the Kalman filter algorithm model based on zero-speed correction is established as follows: Construct an 18-dimensional Kalman filter state quantity, as follows: in, is the platform angle error, δV E , δV N , δV U is the speed error, δL, δλ, δh are the position errors, ε bx , ε by , ε bz is the gyro random constant, ε rx , ε ry , ε rz is the random error of the gyro first-order Markov process, is the random error of the first-order Markov process of the accelerometer; The system state equation is: Where A is the system matrix, X is the system state vector, G is the system noise matrix, and W is the system white noise. After the pedestrian enters the zero-speed state at the foot, theoretically, the speed is zero and the position remains unchanged, thus satisfying: ΔV=V INS -V zupt =0 ΔP=P INS -P zupt =0 V zupt ＝[000] T P zupt =P first Based on the inertial / zero-speed combined navigation system, the position and speed combination method is used to convert the speed V output by the inertial navigation solution at the current moment into INS 、Position P INS The speed V corresponding to the zero speed moment zupt 、Position P zupt Take the difference as the observed quantity, P first is the position at the previous moment in the current zero-speed interval; When the zero speed interval is detected, the observation information is Z ZUPT (t), the observation equation is: Where N zupt (t) is the observation noise.

5. The method for indoor pedestrian navigation and positioning under multiple constraints according to claim 1, characterized in that: In step 5, an error correction algorithm based on the consistency of limb heading changes is established. For each step, the change in waist heading angle Δψ waist and the foot heading angle change Δψ head Set them equal and set Δψ waist and Δψ head The difference Δψ1 constructs the virtual observation: The observation information is: In the formula, Respectively represent the mean waist heading angle of the current step and the previous step, Represent the mean value of the foot heading angle of the current step and the previous step respectively; the observation equation of the error correction algorithm based on the consistency of limb heading change is: Where ψ is the heading angle, θ is the pitch angle; The observation equation of the combined system is: Z WF (t)=[Δψ1]==H WF (t)X(t)+N WF (t) Where N WF (t) is the observation noise.

6. The method for indoor pedestrian navigation and positioning under multiple constraints according to claim 1, characterized in that: In step 6, a constraint algorithm based on the change of air pressure relative to altitude is established. The observation equation after introducing the pressure altitude is: Z baro (t)=[(h baro -h baroinit )-(h SINS -h SINSinit )]=H baro (t)X(t)+N baro (t) In the formula, h baroinit is the initial altitude calculated by the barometer, h SINS is the height calculated by pure inertial real-time solution, h SINSinit is the pure inertial initial height; The state equation is: H baro (t)=[0 1×8 10 1×9 ] In step 7, an improved heuristic heading drift elimination algorithm based on building structure is established; the heading angles of the three consecutive steps are ψ k-2 , k-1 and ψ k The average heading angle of these three steps is ψ k,m , with variance σ k,ψ , there are the following judgment conditions: Where TH ψ,m is the mean threshold for straight line judgment; when the maximum difference between the heading angle in these three steps and its average value is less than the set threshold, it is considered that the current movement is straight; otherwise, it is in non-straight movement; Where TH σ is the variance threshold for straight line judgment; when the heading angle variance in these three steps is less than the set threshold, it is considered to be in straight-line motion; Otherwise, it is in non-straight motion; When Z4(k) and Z5(k) are both detected as 1, it means that the pedestrian is moving straight: Under the premise that the pedestrian is currently moving straight, in order to facilitate the judgment of whether the pedestrian is currently moving straight in the dominant direction, the current heading angle ψ k Projected to In the interval, the calculation formula is: When k |>TH ζ,k When , it means that the pedestrian is currently moving straight in the dominant direction, and the heading angle is corrected, otherwise the heading angle is not corrected; where TH ζ,k Indicates the judgment threshold angle of the dominant direction of straight movement; When k >0, it means that the pedestrian moves along the right side close to the dominant direction; when ζ k When <0, it means that the pedestrian moves along the left side close to the dominant direction; Each time you walk a step, you change the current heading angle ψ k The heading angle observation is constructed by the difference Δψ2 between the adjacent main heading angles, and an improved heuristic heading drift elimination algorithm is proposed; the observation information is: In the formula, Indicates the main heading angle at the current moment; The observation equation is: The state equation is: Where N ψ,iHDE (t) is the observation noise.

7. The method for indoor pedestrian navigation and positioning under multiple constraints according to claim 1, characterized in that: In step 8, there is a maximum value for the distance between two consecutive landings of a single foot, that is, the maximum step length d of human movement. max , a correction method based on multiple constraints of the maximum step length of a single foot was designed; During the motion, the distance d between two adjacent zero-speed intervals k It can be regarded as the step length of this step, where k represents the kth step. This distance should not exceed the maximum step length and must satisfy d k ≤d max ; On the basis of zero-speed correction, a single-leg distance inequality constrained Kalman filter system is constructed; the system constructs two sets of state quantities in adjacent zero-speed intervals into a new 36-dimensional state quantity, and only corrects the state quantity of the current zero-speed interval point, thereby correcting the current position, speed and posture; The system status is as follows: And=[And k-1 AND k ] Where Y k-1 and Y k They are: The single-leg maximum step length inequality constrained Kalman filter navigation model is as follows: In the formula, is the system matrix, Y is the system state vector, is the system noise matrix, W′ is the system white noise; The single-foot distance error during pedestrian movement satisfies the following relationship: In the formula, g(Y) is the objective function, L s Represents the distance constraint model, δd = dd max , δd represents the single-foot step length error, and d is the solved single-foot step length; P ic is a positive definite projection matrix, generally P m for The covariance matrix of When k satisfies When , the single-leg step length is considered to be d max , and obtain a state quantity The constraint model of the reorganization filter model is: In the formula, for The linearized single-foot step length equation is expressed as follows: At this time, the state The following constraint equations should be satisfied: In the formula, Z ic is the observed quantity, and the updated state quantity is obtained And the error covariance matrix It is expressed as follows: in: μ=P ic k T (kP ic k T ) -1 k After obtaining the state quantity and error covariance matrix of the current zero-speed point under the reorganization filter model, the navigation parameters such as position, speed and attitude are corrected.

8. The method for indoor pedestrian navigation and positioning under multiple constraints according to claim 1, characterized in that: The SSA-based LSTM network parameters are obtained by solving the leg training set data, and the test data of the leg IMU is input into the model for motion mode identification, providing a basis for subsequent navigation; the parameters of the LSTM algorithm are optimized: the sparrow search algorithm is used to optimize the LSTM algorithm. The virtual sparrows search for food, and the mathematical expression of the sparrow population composed of q sparrows is: In the formula, s q,d Where q is the sparrow number, d is the dimension of the variable to be optimized, and s q,d That is, the d-dimensional state of the qth sparrow. In optimizing LSTM classification, s q,d represents the d-dimensional eigenvalue in the qth step; Where M S represents the fitness of all sparrows, m([s q,1 s q,2 …s q,d ]) represents the fitness value m of the qth sparrow s ; In optimizing LSTM classification, m([s q,1 s q,2 ...s q,d ]) represents all eigenvalues ​​in the qth step; SSA optimizes the following four parameters of the LSTM network: the number of hidden layers, the maximum training cycle, the initial learning rate, and the L2 parameter; the fitness function formula of the sparrow search algorithm to optimize the long short-term memory network is: fitness=E train +E test In the formula, E train represents the error rate of the training set, E test Represents the error rate of the test set; the fitness function is the sum of the prediction error rates of LSTM for the training set and the test set.

9. The method for indoor pedestrian navigation and positioning under multiple constraints according to claim 8, characterized in that: The process of optimizing the LSTM algorithm based on SSA is as follows: Step 1: Input the modal identification training set data, including multi-modal motion feature values ​​and modal flags; Step 2: Initialize the LSTM model; Step 3: Calculate the initial fitness value; Step 4: SSA initialization, update the parameters of discoverer, joiner and alerter; Step 5: Iterate to find the global optimal fitness value; Step 6: After the iterative optimization is completed, the LSTM parameters optimized by SSA are output: number of hidden layers, maximum training cycle, initial learning rate, and L2 parameter.

Images