CN105865448A - Indoor positioning method based on IMU - Google Patents

Abstract

The invention discloses an indoor positioning method based on IMU. The indoor positioning method is technically characterized by comprising the following steps: S1, according to acquired data about acceleration of human motion, segmenting valid gait by using a finite-state machine and carrying out step calculation and step length estimation; S2, according to data of the angular velocity (measured with a gyroscope), acceleration (measured with an accelerometer) and magnetic field intensity (measured with a magnetic field sensor) of the valid gait, calculating the motion direction of current gait by using a complementary filtering model; and S3, carrying out particle filtering according to the step length, motion direction estimate results and map information and adjusting parameters of a step length model and the complementary filtering model according to attributes of effective viable particles. According to the invention, through usage of the finite-state machine, invalid gait information can be removed, so calculation amount of data is reduced; the complementary filtering model is utilized to fuse a plurality of sensor information so as to provide more accurate direction measurement results; a particle filtering method is employed for dynamic adjustment of parameters of a gait model and the complementary filtering model, so step length estimation errors and direction measurement errors are reduced; and in a word, the method provided by the invention can inhibit error accumulation in the process of dead reckoning and improve precision of indoor positioning.

Description

A kind of indoor orientation method based on IMU

Technical field

The present invention relates to indoor positioning field, be specifically related to a kind of indoor dead reckoning localization method based on IMU.

Background technology

Service based on positional information becomes more and more important, and major applications is all to use based on satellite at present Global positioning system.But, under indoor environment, satellite-signal can be blocked by building causes blind area, a large amount of position Occur, and the precision of current satellite positioning method temporarily cannot meet the requirement of indoor positioning.Therefore, build one stable, Indoor locating system becomes study hotspot in recent years in high precision.

Indoor positioning technologies can be broadly divided into two classes at present: based on location fingerprint method and dead reckoning method.Based on position Put in the indoor positioning technologies of fingerprint, need indoor environment is carried out substantial amounts of hardware modification, to arrange WiFi node, therefore its Hardware cost is higher.On the other hand, owing to wireless signal is easily affected by transmission environment, produce such as multipath effect, non-line-of-sight and pass Broadcast, the situation such as signal fadeout and noise jamming, therefore its positioning precision is disturbed significantly, needs irregularly for this Updating location fingerprint data base, even revise signal propagation model, the later maintenance for indoor locating system brings higher skill Art cost.And indoor positioning technologies of based on dead reckoning is unrelated with external environment, and only with being moved through of moving target self Cheng Youguan, therefore has important value and status in indoor locating system.And MEMS (Micro-electro Mechanical Systems, MEMS) technology develop rapidly promoted further this method indoor positioning field should With.

In the indoor positioning for pedestrian is applied, dead reckoning method based on IMU has obtained studying widely and answering With.The method is mainly by the movable information during various sensor acquisition pedestrian movement, such as acceleration, angular velocity and ring Magnetic field, border etc., and then the information such as the step number of pedestrian movement, step-length, direction can be obtained.In the case of known initial position, can Recursively to estimate the position of pedestrian.But the most generally there is following problem in the method: the multiformity of the first, motion gait Cause meter step and the inaccurate problem of step-size estimation；The second, the error in data that various sensor noise brings, particularly in motion Direction estimation aspect；3rd, dead reckoning self with deviation accumulation problem.

Summary of the invention

Present invention is primarily targeted at, ask for step-size estimation during dead reckoning and orientation measurement are coarse Topic, utilizes particle filter method, is actively modified step-length model and the deviation of directivity, to alleviate error during dead reckoning The problem of accumulation, improves the overall precision of indoor positioning.

The present invention solves existing technical problem is that and realizes by the following technical solutions:

A kind of indoor orientation method based on IMU relies on a smart mobile phone based on Android platform, described intelligence Embedded in mobile phone three-dimensional accelerometer, three-dimensional gyroscope, three-D magnetic field sensor；Smart mobile phone is detected by described built-in sensors Acceleration, angular velocity and the direction of motion during individual sports, then uses particle filter side during dead reckoning Method carries out location estimation；In position fixing process, by the step-length attribute of effective particle and direction attribute are carried out statistics and analysis, The deviation of directivity of the step-length model during dead reckoning can be modified.

In technique scheme, described indoor orientation method based on IMU comprises the steps:

The segmentation of step S1, gait and step-size estimation: according to the human motion acceleration information gathered, use finite state machine Effective gait is split and completes meter step and step-size estimation；

Step S2, angular velocity (gyroscope), acceleration (acceleration transducer) and magnetic field intensity according to effective gait (magnetic field sensor) data, use complementary filter model to calculate the direction of motion of current gait；

Step S3, carry out particle filter according to step-length, direction of motion estimated result and cartographic information, and according to effectively depositing The Attribute tuning step-length model of bioplast and the parameter of complementary filter model.

And, described step S1 comprises the following steps:

Step S1.1, acceleration is carried out data prediction；

Kinestate is judged, and enters effective gait under kinestate by step S1.2, use finite state machine Row time division；

Step S1.3, acceleration information and step-length model according to effective gait carry out step-size estimation.

And, the concrete methods of realizing of described step S1.1 is: first according to the formula (1) the acceleration information meter to three axles Calculate resultant acceleration, and use sliding window averaging method to be filtered, with smooth jittering noise.The reason using resultant acceleration is main It is: forward acceleration, lateral acceleration when the measured value on acceleration transducer three axle is pedestrian movement and vertically accelerate Degree sum of component on each direction of principal axis, uses quadratic sum numerical value can enable the mobile phone that meter step algorithm is applicable on diverse location Put.

$a=\sqrt{a_x^2+a_y^2+a_z^2}---\left(1\right)$

And, the concrete methods of realizing of described step S1.2 is: according to the threshold value of the setting in finite state machine to currently Kinestate judges, and splits effective sensor data when target is kept in motion.At finite state machine In, the implication of each state is respectively as follows:

S0: represent resting state；

S1: represent that preliminary exercise state, i.e. target are likely to be at kinestate；

S2, S3: enter peak state and leave peak state；

S4, S6: enter valley state and leave valley state；

S7: gait done state；

S5: tolerate for noise.

Only when acceleration information input makes state machine arrive state S7, just think and there occurs effective gait row For, and the sensing data between state S1～S7 is effectively recorded data as current gait.

Additionally, in state transition graph, Thr is motion detection threshold；Thr_ppFor peak threshold；Thr_npFor valley threshold； Thr_negThreshold value is terminated for gait.Thr and Thr_negIt is generally fixed near 9.8, for judging beginning and the end of gait；Thr_pp And Thr_npIt is generally required to set according to gait feature, and it it is the key factor of impact meter step precision.Use finite state machine Method be that a little it combines peak detection algorithm and the advantage of zero passage detection algorithm, to the various vibrations in motor process Noise has a higher tolerance, and can recognize that " pseudo-gait " that human body occurs in the course of action such as stand up, raise one's hand is existing As.

And, the concrete methods of realizing of described step S1.3 is: for effective gait, record between state machine S2～S3 The maximum a of acceleration_max；In minima a recording acceleration between state machine S4～S6_min, then according to formula (2) institute Step-length L of current gait estimated by the meter step model given.

$L=K*\left(\right(a_{max}-a_{\min })+\sqrt[4]{a_{max}-a_{\min }})---\left(2\right)$

In general, model parameter K is difficult to accurately calculate.To this end, the particle filter mistake that the present invention is in step s3 Cheng Zhong, is made K obey a normal distribution, is then estimated the value of K further by the attribute of survival particle.

And, described step S2 comprises the following steps:

Step S2.1, by gyroscope integration obtain direction of motion θ_g；

Step S2.2, obtain direction of motion θ by magnetic field sensor and acceleration transducer_m。

Step S2.3, merge θ by complementation model_gAnd θ_mInformation, it is thus achieved that direction of motion θ of correction_c。

And, the concrete methods of realizing of described step S2.1 is: according to formula (3), have described step S1.2 segmentation Three-axis gyroscope data in effect gait are integrated, and result are added in the direction of motion of a upper gait, obtain mesh The absolute direction θ of mark motion_g。

θ_g(t)=θ_g(t-1)+Δθ_t (4)

And, the concrete methods of realizing of described step S2.3 is: utilize the complementary filter model described in formula (5), to institute The result stating step S2.1 and step S2.2 is for further processing, and obtains the absolute movement direction θ of target travel_c。

θ_c=α * θ_g+(1-α)*θ_m (4)

The main thought of complementary filter model is: gyroscope dynamic response characteristic is excellent, when resolving attitude angle, due to gyro The impact of instrument low frequency wonder, after integration, low-frequency excitation can produce bigger error；The attitude angle that accelerometer resolves can be moved During the impact of dither, output angle is carried the High-frequency Interference of bigger component.The two has complementary special on frequency domain Property, use complementary filter to both Data Fusion of Sensors, the precision of attitude angle measurement and dynamic response can be improved Performance.

And, described step S3 comprises the following steps:

Step S3.1, generate each particle property according to step-length and direction estimation result and noise model；

Step S3.2, all particles are carried out dead reckoning；

The invalid particle of step S3.3, according to the map information deletion: region can not be arrived (such as body of wall, floating district for entering Territory etc.) particle so that it is weight is zero and lost efficacy, and other particles are effective particle；

Step S3.4, judge whether to need to carry out resampling: when number of effective particles is less than the 2/3 of primary number, Illustrate to need to carry out resampling, perform described step S3.5, otherwise directly estimate the position of target according to effective particle；

Step S3.5, according to described step S3.4 judge need to carry out resampling time, use sequential importance resampling (SIR) method carries out resampling to particle；

Step S3.6, after carrying out resampling according to described step S3.4, step-length attribute and direction attribute to particle enter Row statistics and analysis, to revise step error and the deviation of directivity.

And, the concrete methods of realizing of described step S3.1 is: according to step-length and direction estimation result and noise model (formula 5,6) generates each particle property.

$\begin{array}{c}{L}_t^i=L_t+n_L^i,n_L^i~N_{0,{\mathrm{\σ}}_L}\\ {\mathrm{\θ}}_t^i={\mathrm{\θ}}_t+{\mathrm{\θ}}_{drift}+n_{\mathrm{\θ}}^i,n_{\mathrm{\θ}}^i~N_{0,{\mathrm{\σ}}_{\mathrm{\θ}}}\end{array}---\left(5\right)$

Wherein,For i-th particle in the step-length of t, including: L_tThe step estimated according to step-length model for t Long result；It is that an obedience is distributed the most very muchStep-length noise, and σ_LRatio for fixed step size；For i-th particle In the direction of t, including: direction of motion value θ of estimation_t, deviation of directivity θ_driftAnd a Normal DistributionDirection noise, and σ_θIt is set to the fixed proportion of adjacent gait direction change.

$\begin{array}{c}{K}_t^i=K_t+n_K^i,n_{\mathrm{\θ}}^i~N_{0,{\mathrm{\σ}}_K}\\ {\mathrm{\α}}_t^i={\mathrm{\α}}_t+n_{\mathrm{\α}}^i~N_{0,{\mathrm{\σ}}_{\mathrm{\α}}}\end{array}---\left(6\right)$

Wherein,For i-th particle in the step-length model parameter of t, including: estimated value part K_tAnd random value PartFor i-th particle in the complementation model parameter of t, including: estimated value part α_tWith random value part

And, the concrete methods of realizing of described step S3.6 is: during described step S3.3, a part of particle because Flying in the face of facts and be deleted, the particle of another part survival is then by random reproduction, to keep number of particles constant.In this process In, the Discrete Distribution situation of each property value of particle changes, and its statistical value can preferably arrange model parameter.

$\begin{array}{c}{\mathrm{\θ}}_{drift}=\frac{1}{N_s}\underset{i=1}{\overset{N_s}{\Σ}}{n}_{\mathrm{\θ}}^i\\ K_t=K_{t-1}+\frac{1}{N_s}\underset{i=1}{\overset{N_s}{\Σ}}{n}_K^i\\ {\mathrm{\α}}_t={\mathrm{\α}}_{t-1}+\frac{1}{N_s}\underset{i=1}{\overset{N_s}{\Σ}}{n}_{\mathrm{\α}}^i\end{array}---\left(7\right)$

Compared to the shortcoming and defect of prior art, the method have the advantages that for tradition IMU indoor positioning In method, the parameter of step-length model arranges the deviation accumulation problem that inaccurate and orientation measurement is brought, and the present invention is filtered by particle Fluctuation state adjusts step-length model parameter and complementary filter parameter, it is achieved that adaptive step and more accurately orientation measurement knot Really, various sensor information is incorporated, it is possible to be effectively improved indoor position accuracy.Additionally, by limited shape during meter step Sensing data is split by state machine, it is possible to identify the sensing data of invalid gait, decreases unnecessary data and calculates And analysis.

Accompanying drawing explanation

The indoor orientation method flow chart based on IMU of Fig. 1 present invention.

Fig. 2 judges and the finite state machine status transition diagram of gait segmentation for kinestate.

Fig. 3 is for the complementary filter model algorithm flow chart of direction estimation

S1.1, finite state machine in Fig. 1；S1.2, meter step and step-size estimation；S2, orientation measurement；S3, particle filter and ginseng Number feedback.

Detailed description of the invention

Understandable for enabling the above-mentioned purpose of the present invention, feature and advantage to become apparent from, real with concrete below in conjunction with the accompanying drawings The present invention is further detailed explanation to execute mode, and these accompanying drawings are the schematic diagram of simplification, and this is described the most in a schematic way The basic structure of aspect, therefore it only shows the composition relevant with the present invention.

The smart mobile phone of the use of the present invention is Samsung Galaxy S6 (G9200), and built-in sensor model number is MPU6500 six axle sensor (acceleration transducer+gyroscope) of InvenSense company and the AK09911 of AKM company, carry Operating system be Android OS 5.0.

Data acquisition is originated:

In android system, developer uses various sensing data for convenience, there has been provided sense number in a large number According to API.The key data used in the present invention has: acceleration, TYPE_ACCELEROMETER provide；Angular velocity, by TYPE_GYROSCOPE provides；Rotating vector, is provided by TYPE_ROTATION_VECTOR.Wherein rotating vector is a kind of virtual Sensor, refers to the mobile phone coordinate system rotating vector relative to world coordinate system, and its value is by acceleration transducer and magnetic field sensor The data provided calculate and obtain.

Indoor positioning flow process:

First pass through method initial position of labelling trip people on map of artificial demarcation.In order to simplify in position fixing process Problems of Coordinate Rotation, in implementation process by pedestrian with read mobile phone screen state handheld mobile phone carry out gait motion, now The sensing of mobile phone coordinate system x-axis i.e. pedestrian's direction of motion.Under other states, can be by calculation of motion vectors at mobile phone coordinate Value in system carries out coordinate rotation.

After initialized target position, described finite state machine carry out meter step and split with gait.At finite state machine In, the setting of each threshold value is described as follows: owing to smart mobile phone receives the impact of acceleration of gravity, under remaining static, and closes The value becoming acceleration is 9.8 (1 G), therefore Thr and Thr_negBe generally fixed near 9.8, for judge gait beginning and Terminate, in implementation process, respectively by Thr and Thr_negIt is set to 10.0 and 9.7；Thr_ppAnd Thr_npRepresent peak threshold respectively With valley threshold, it is to need to arrange according to the motion amplitude of gait in general, or carries out dynamically by the way of study Set, in the implementation process of the present invention, be respectively set to 11.0 and 9.0.

After effective gait being detected, estimate the step-length of current gait first by step-length model；Meanwhile, complementation is used The direction of motion is estimated by Filtering Model.In implementation process, the parameter of step-length model and complementary filter model is all as grain The property value of son exists, and needs to carry out randomization.Wherein, step-length model parameter K is set to 1, and normal distribution ginseng time initial Number σ_KIt is set to 0.1；Complementation model parameter alpha is set to 0.98, Parameters of Normal Distribution σ_αIt is set to 0.01, and need to meet.Need It is noted that all parameters are arranged when initial the most based on experience value, during particle filter subsequently, some particles Disappearing because of dead, the distribution situation of parameter value can change, now each parameter value is modified can time position system System can adapt to different gait feature and environment, obtains preferably locating effect.

Finally, pedestrian position estimation and parameter feedback are carried out by particle filter method.In implementation process, number of particles N needs to be configured according to computing capability.Owing to the particle filter implementation process of the present invention is to carry out on the server, therefore will N is set to bigger 2000.If needing to be placed on smart mobile phone this process completes, need to consider the calculating of cell phone processor Ability, arranges a less N value.

Parameter offering question in the implementation process of the present invention is mainly illustrated by said process, and idiographic flow can be tied Close the algorithm steps provided in summary of the invention.It is pointed out that the one skilled in the art's concrete reality to the present invention Any change that the mode of executing is done is all without departing from the scope of claims of the present invention.Correspondingly, the claim of the present invention Scope be also not limited only to described detailed description of the invention.

Claims (4)

1. an indoor dead reckoning method based on IMU, it is characterised in that the method comprises the following steps:

The segmentation of step S1, gait and step-size estimation: according to the human motion acceleration information gathered, use finite state machine to having Effect gait carries out splitting and complete meter step and step-size estimation；

Step S2, angular velocity (gyroscope), acceleration (acceleration transducer) and magnetic field intensity (magnetic field according to effective gait Sensor) data, use complementary filter model to calculate the direction of motion of current gait；

Step S3, carry out particle filter according to step-length, direction of motion estimated result and cartographic information, and according to grain of effectively surviving The Attribute tuning step-length model of son and the parameter of complementary filter model.

A kind of indoor dead reckoning method based on IMU, it is characterised in that in step sl, institute State gait segmentation to comprise the following steps with step-size estimation method:

Step S1.1, data prediction: calculate total acceleration amplitude according to formula (1) and 3-axis acceleration data, and use cunning Window averaging method is filtered；

$a=\sqrt{a_x^2+a_y^2+a_z^2}---\left(1\right)$

Step S1.2, kinestate judge: judge current motion state according to the threshold value of the setting in finite state machine, And when target is kept in motion, effective sensor data are split.In finite state machine, the implication of each state is divided It is not:

S0: represent resting state；

S1: represent that preliminary exercise state, i.e. target are likely to be at kinestate；

S2, S3: enter peak state and leave peak state；

S4, S6: enter valley state and leave valley state；

S7: gait done state；

S5: tolerate for noise.

Only when acceleration information input makes state machine arrive state S7, just think the effective gait behavior that there occurs, and Sensing data between state S1～S7 is effectively recorded data as current gait.

Step S1.3, the step-size estimation that the currently active gait carried out according to the step-length model described by formula (2):

$L=K*\left(\right(a_{max}-a_{min})+\sqrt[4]{a_{max}-a_{min}})---\left(2\right)$

Wherein, L is the step-length result estimated；K is model parameter；a_max、a_minIt is respectively the acceleration in the currently active gait Big value and minima.

A kind of indoor dead reckoning method based on IMU, it is characterised in that in step s 2, fortune Dynamic direction determining method comprises the following steps:

Step S2.1, according to formula (3), the three-axis gyroscope data in effective gait of described step S1.2 segmentation are amassed Point, and result is added in the direction of motion of a upper gait, obtain the absolute direction θ of target travel_g(t)=θ_g(t-1)+Δ θ_t。

Step S2.2, according to acceleration information obtain Smartphone device inclination angle result, then in conjunction with magnetic field sensor data Obtain the absolute direction θ of target travel_m。

Step S2.3, utilize the complementary filter model described in formula (4), the result of described step S2.1 and step S2.2 is made into One step processes, and obtains the absolute movement direction θ of target travel_c。

θ_c=α * θ_g+(1-α)*θ_m (4)

A kind of indoor dead reckoning method based on IMU, it is characterised in that in step s3, grain Son filtering dead reckoning method comprises the following steps:

Step S3.1, generate each particle property according to step-length and direction estimation result and noise model (formula 5).

$\begin{array}{c}{L}_t^i=L_t+n_L^i,n_L^i~N_{0,{\mathrm{\σ}}_L}\\ {\mathrm{\θ}}_t^i={\mathrm{\θ}}_t+{\mathrm{\θ}}_{drift}+n_{\mathrm{\θ}}^i,n_{\mathrm{\θ}}^i~N_{0,{\mathrm{\σ}}_{\mathrm{\θ}}}\end{array}---\left(5\right)$

Wherein,For i-th particle in the step-length of t, including: L_tThe step-length knot estimated according to step-length model for t Really；It is that an obedience is distributed the most very muchStep-length noise, and σ_LRatio for fixed step size；For i-th particle when t The direction carved, including: direction of motion value θ of estimation_t, deviation of directivity θ_driftAnd a Normal DistributionSide To noise, and σ_θIt is set to the fixed proportion of adjacent gait direction change.

Step S3.2, according to formula (6) with all particles to be carried out dead reckoning:

$\begin{array}{c}{x}_t^i=x_{t-1}^i+L_t^i{\mathrm{cos\θ}}_t^i\\ y_t^i=y_{t-1}^i+L_t^i{\mathrm{sin\θ}}_t^i\end{array}---\left(6\right)$

The invalid particle of step S3.3, according to the map information deletion.

According to the interior architecture thing cartographic information provided, for entering the grain that can not arrive region (such as body of wall, floating region etc.) Son so that it is weight is zero and lost efficacy, and other particles are effective particle.

Step S3.4, judge whether to need to carry out resampling.

When number of effective particles is less than the 2/3 of primary number, illustrates to need to carry out resampling, perform described step S3.5, Otherwise directly according to the position of formula (7) estimation target:

${\hat{S}}_t=\underset{i=1}{\overset{N}{\Σ}}{s}_t^i---\left(7\right)$

Step S3.5, according to described step S3.4 judge need to carry out resampling time, use sequential importance resampling (SIR) Method carries out resampling to particle.

Step S3.6, after carrying out resampling according to described step S3.4, step-length attribute and direction attribute to particle are united Meter and analysis, to revise step error and the deviation of directivity.

During described step S3.3, a part of particle is deleted because of flying in the face of facts, and the particle of another part survival Then by random reproduction, to keep number of particles constant.In this process, the Discrete Distribution situation of each property value of particle becomes Changing, its statistical value can preferably arrange model parameter.

$\begin{array}{c}{\mathrm{\θ}}_{drift}=\frac{1}{N_s}\underset{i=1}{\overset{N_s}{\mathrm{\Σ}}}{n}_{\mathrm{\θ}}^i\\ K_t=K_{t-1}+\frac{1}{N_s}\underset{i=1}{\overset{N_s}{\mathrm{\Σ}}}{n}_K^i\\ {\mathrm{\α}}_t={\mathrm{\α}}_{t-1}+\frac{1}{N_s}\underset{i=1}{\overset{N_s}{\mathrm{\Σ}}}{n}_{\mathrm{\α}}^i\end{array}---\left(8\right)$