CN105044662A - Fingerprint clustering multi-point joint indoor positioning method based on WIFI signal intensity - Google Patents

Abstract

The invention relates to a fingerprint clustering multi-point joint indoor positioning method based on the WIFI signal intensity, which is characterized in that a signal fingerprint at a reference point is collected, the fingerprint classification number and an initial clustering center feature vector are determined by using a distance matrix based dynamic virtual point initial clustering center selection method, fingerprints in a fingerprint database are classified through an initial clustering method, then cluster soft classification is realized on the basis by using a log-likelihood probability expectation maximizing method, clusters are enabled to be mutually overlapped, and probability model parameters of the fingerprints are acquired. In positioning, reference points participating in positioning are acquired by using a signal vector, the probability model parameters and a class feature fingerprint vector which are received in real time, and finally, positioning coordinates are estimated by using a multi-reference-point joint positioning method. According to the invention, a clustering optimized fingerprint positioning algorithm is adopted, thereby greatly improving the fingerprint matching efficiency and the positioning precision, reducing the cost of wireless positioning, and having high market competitiveness.

Description

A kind of fingerprint cluster multi-point joint indoor orientation method based on WIFI signal intensity

Technical field

The present invention relates to wireless communication technology field, be specifically related to the localization method of the fingerprint cluster multi-point joint indoor locating system under WIFI signal intensity level.

Background technology

WIFI is the wireless network standards based on IEEE802.11 standard, is that in people's daily life, the terminal such as PC, handheld mobile device wirelessly accesses one of main way of Internet core net.Along with the development of wireless technology, nowadays in the social scene of the large-scale life such as market, house, office, WIFI access point almost all covers, and its development trend will finally realize the comprehensive covering of living region.Under this trend, utilize WIFI to carry out indoor positioning convenient all the more, and WIFI signal overcome gps signal cannot have effect spread shortcoming in indoor, utilize existing WIFI access point completely, without the need to arranging great deal of nodes as bluetooth location, without the need to wiring, with low cost.Fingerprint location technology is a kind of localization method based on not finding range, typical localization method is as TOA, TDOA, AOA, RSSI etc., because indoor environment exists serious multiple scattering, often there is comparatively big error in the estimation of above-mentioned parameter, positioning performance is often not ideal; And this imparametrization localization method of fingerprint location is without the need to estimating environmental parameter, effectively can resist indoor multipath to propagate, greatly enhance the precision of indoor positioning, generally to receive the fingerprint of the signal containing environmental error as this reference point in reference point before location, then at positioning stage, the signal received and fingerprint base are contrasted, thus realize location estimation.

Summary of the invention

Technical matters: the object of this invention is to provide a kind of WIFI cover indoor environment in request location mobile subscriber provide the method and system of location estimation to realize.Build storehouse by off-line, process implementation that fingerprint cluster is mated with tuning on-line.

Technical scheme: for realizing object of the present invention, the present invention relates to a mobile terminal and come collection signal, a server to set up fingerprint base and to realize localization method, under simultaneously the present invention need be applied in the indoor environment of many WIFI access point, wherein server selection software comprises collection signal off-line and builds library facility and localization method practical function.

The present invention mainly comprises: the optimization cluster fuzzy partitioning clustering method under a kind of point of the dynamic virtual based on distance matrix initial cluster center selection algorithm, and a kind of based on class participation multi-point joint dynamic positioning method, idiographic flow is as follows:

(1) area to be targeted divides by mode in a grid formation that arrange a reference point at interval of 1 meter, and each lattice point is 1 reference point, divides N number of reference point altogether.

(2) terminal of loading server positioning software is held, successively at reference point place collection signal, by signal intensity vector stored in fingerprint database.

(3) carry out distance matrix calculating according to fingerprint base vector of samples, realize fingerprint classification according to clustering algorithm, obtain the class distribution matrix of category feature vector and reference point received signals fingerprint.

(4) the current WIFI list of mobile subscriber's automatic acquisition to be positioned and respective signal value, sends Location Request to server, and sends WIFI signal vector.

(5) server mates with all kinds of proper vector after obtaining band positioning signal vector, obtains class label.Select local positioning region according to class label, utilize local locating method to calculate estimated position to be positioned.

(6) estimated position is returned mobile subscriber to be positioned by server.

Beneficial effect: the present invention has the following advantages:

A. equipment investment is decreased.Utilizing existing WIFI access point, without the need to redeploying localizing environment, reducing cost;

B. improve location efficiency.Without the need to measuring environmental parameter, decrease location previous work;

C. positioning precision is added.Strengthen the analysis and treament to fingerprint characteristic, take the fingerprint vector and the feature of positional information effectively, and positioning precision is improved greatly;

D., a kind of new algorithm of fingerprint location is proposed.By the analysis to fingerprint cluster process, making off-line to build in storehouse final step cluster boundary can fuzzy partitioning, this principle proposes a kind of new location algorithm, is optimized matching algorithm during location, and consider the select permeability of reference point, make location algorithm reach optimum.

Accompanying drawing explanation

Fig. 1 is based on the fingerprint cluster multi-point joint indoor orientation method process flow diagram of WIFI signal intensity.

Fig. 2 off-line builds the storehouse stage, and handheld terminal is collection signal schematic diagram in each reference point.

Fig. 3 positioning stage schematic diagram.

The lower reference point distribution plan of initial hardening point of Fig. 4 classification.S1 represents locating area.

The lower reference point distribution plan of Fig. 5 fuzzy partitioning classification.S2 represents locating area.

Reference point distribution plan in a class under Fig. 6 fuzzy partitioning.S3 represents locating area.

Embodiment

For the analysis of fingerprint clustering method in (3) and implementation process as follows:

1, a kind of point of the dynamic virtual based on distance matrix initial cluster center system of selection

According to flow process (1) described mode, the terminal of holding loading server positioning software receives the signal intensity vector r from M WIFI access point in N number of reference point successively _i, vector is made up of M signal intensity component, namely

$r_i=(r_{i,{\mathrm{AP}}_1,}{r}_{i,{\mathrm{AP}}_2}\·\·\·r_{i,{\mathrm{AP}}_M}),$

Reference point set is

U＝{u|u≤N，u∈N ^*}

Because a certain WIFI signal value of pointing out can fluctuate in time in practice, so should gather multiple signal in same reference point, namely gather T signal intensity vector in each reference point, the signal intensity vector of collection is specifically expressed as:

$r_i\left(t\right)=(r_{i,{\mathrm{AP}}_1}\left(t\right),r_{i,{\mathrm{AP}}_2}\left(t\right),\·\·\·r_{i,{\mathrm{AP}}_M}\left(t\right)),t=1,2,\·\·\·,T,i=1,2,...,N.$

To T sampling filter process, average as fingerprint stored in fingerprint database, then each reference point place fingerprint is

$r_i=\frac{1}{T}{\mathrm{\Σ}}_{t=1}^T{r}_i\left(t\right),i=\mathrm{1,2},...,N.$

Set up reference point raw range matrix D _{n × N}:

$\left[\begin{array}{ccccc}0& D_{\mathrm{1,2}}& \·\·\·& D_{1,N-1}& D_{1,N}\\ D_{\mathrm{2,1}}& 0& \·\·\·& D_{1,N-2}& D_{2,N}\\ \·& \·& \·& \·& \·\\ \·& \·& \·& \·& \·\\ \·& \·& \·& \·& D_{N-1,N}\\ D_{N,1}& D_{N,2}& \·\·\·& D_{N,N-1}& 0\end{array}\right]$

Wherein, distance element D _{i, j}=|| r _i-r _j|| ², D _{i, j}=D _{j, i}, this distance matrix is symmetric matrix.

Dynamic virtual point initial cluster center system of selection embodiment based on distance matrix is as follows:

1) for Distance matrix D _{n × N}if, distance element D _{i, j}< ε, then by i, j point deletion in former reference point set U.

2) interpolation represents i, and the received signals fingerprint of the virtual reference point v of j, v point is upgrade set U=U ∪ v _g-i, j}, g=1,2 ....

3) upgrade current reference point number N, upgrade Distance matrix D _{n × N}, now the dimension of distance matrix reduces, repeats 1,2, if produce without new virtual point, and, increase and upgrade ε=ε+θ, repeat 1,2.

4) repeat 1,2,3, if ε > is γ, then stop iteration.

5) now reference point set U is initial cluster center set.Corresponding received signals fingerprint is initial cluster center characteristic fingerprint vector { m _j}={ m ₁, m ₂..., m _k.

2, the method for preliminary fingerprint cluster

Initial cluster center { m is established in based on the dynamic virtual point initial cluster center system of selection of distance matrix _k| k=1,2..., K}, class number K.Next will according to initial cluster center by the received signals fingerprint of all reference point according to the raw range matrix D between it _{n × N}be divided into K regional area, namely realize received signals fingerprint higher for similarity to be divided into a class.Classification the target that meets be:

${\mathrm{\&Sigma;}}_{k=1}^K{\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&rho;}}_{\mathrm{ik}}{\left(d_{\mathrm{ik}}\right)}^2---\left(1\right)$

Wherein hypotaxis degree ρ _ikvalue represent received signals fingerprint r _iwhether belong to a kth class, d _ikrepresent r _iwith initial cluster center m _kdistance, then:

${\mathrm{\&rho;}}_{\mathrm{ik}}=\left\{\begin{array}{cc}1,& d_{\mathrm{ij}}={\min }_{l\&Element;K}\left(d_{\mathrm{il}}\right)\\ 0,& d_{\mathrm{ij}}\&NotEqual;{\min }_{l\&Element;K}\left(d_{\mathrm{il}}\right)\end{array}\right.$

The concrete implementing procedure realizing fingerprint cluster is:

1) calculating reference point received signals fingerprint r successively _i(i=1,2 ..., N) and cluster centre m _k(k=1,2 ..., K) distance, work as d _ik=min _{l ∈ K}(d _il) time, make ρ _ik=1, i.e. r _ibelong to class k, otherwise ρ _ij=0, i.e. r _ido not belong to class k.Record is about ρ _iksubordinate matrix Q _{n × K}.

2) regional area k (kth class) reference point set omega is set up _k(k ∈ K), according to subordinate matrix Q _{n × k}if, ρ _ik=1, then reference point i ∈ Ω _k.Calculate new distance center

3) if | m _{k (new)}-m _k| > τ, then use m _{k (new)}replace the distance center m of last iteration _k, repeat 1,2.Ω is upgraded in 2 _k(k ∈ K).

4) if | m _{k (new)}-m _k|≤τ, then cluster centre m _kconvergence, stops repeating.

5) now m _{k (new)}for the final characteristic fingerprint of regional area k, reference point set omega _kthe final classification reference point set that (k ∈ K) is regional area k.

3, maximize log-likelihood probability expectation method and realize cluster fuzzy partitioning

Because clustering algorithm in 2 exists the problem of hard plot in practical operation, i.e. the degree of membership of certain reference point one class non-zero namely 1.And the reference point two class degree of membership be on class separatrix is more or less the same, if it is assigned in a certain class utterly, the target that undesirable similar degree in the class is large as far as possible, unfavorable to classification.The fuzzy partitioning of cluster can be realized by setting up likelihood probability Density Distribution model, degree of membership is determined according to probability, so be in the borderline reference point of class, to two classes or more multiclass has degree of membership, its value is between 0 to 1, and the classification situation describing reference point with this probability properlyer more effectively can realize class object.Perform by two steps fingerprint classification border fuzzy partitioning, one is the expectation obtaining eligible probability density function, and two is the parameter values under asking the expectation of maximization conditional probability density function maximum, comprises degree of membership probability.Output for fingerprint cluster: category feature fingerprint vector m _j, and reference point cluster array:

$\begin{array}{cccc}\mathrm{cluster}1& u_a& u_b& \&CenterDot;\\ \mathrm{cluster}2& u_m& u_n& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \mathrm{clusterK}& u_p& u_q& \&CenterDot;\end{array}$

Using the input as cluster fuzzy partitioning algorithm.Due to the various probability Distribution Model of gaussian probability distribution function matching more preferably, so here, adopt gauss hybrid models for each class reference point signal distributions model, so in Zone Full internal reference examination point, signal distributions model is:

$P\left(R\right|\mathrm{\&Phi;})={\mathrm{\&Sigma;}}_{k=1}^K{\mathrm{\&alpha;}}_k\mathrm{\&phi;}\left(r\right|{\mathrm{\&Phi;}}_k)---\left(2\right)$

${\mathrm{\&Sigma;}}_{k=1}^K{\mathrm{\&alpha;}}_k=1$

For a kth local class locating area signal distributions model be:

Wherein gaussian probability distribution form is:

$\mathrm{\&phi;}\left(r_j\right|{\mathrm{\&phi;}}_k)=\frac{1}{\sqrt{2\mathrm{\&pi;}{S}_k}}\exp (-\frac{{(r-m_k)}^2}{2{S_k}^2})---\left(3\right)$

For a reference point u _ithe attribute possessed comprises (r _i, b _i), wherein r _ifor fingerprint vector, b _ifor participation vector, indicate current reference point received signals fingerprint and whether belong to a kth class.B _icannot directly be drawn by observation, the complete likelihood function under model is:

It is very high that the complete likelihood function of direct use solves computation complexity, can take the logarithm to this to this likelihood function, multiplication transferred to addition to reduce complexity, be convenient to solve.Then log-likelihood function is:

The expectation of log-likelihood function is:

E[logP(r _i，b _i|Φ)|r，Φ](6)

A wherein jth observation signal r _jto the participation Υ of a kth local class locating area signal distributions model _jk, by its conditional expectation approximate evaluation be:

Find out that the participation of reference point is numerically equal to the fingerprint vector of this reference point by a kth posterior probability that local class locating area signal distributions model is specified thus.

Build the parameter vector Φ of gaussian probability distributed model _k={ α _k, m _k, S _k ², for solving parameter, master mould being converted into log-likelihood and expecting form, expecting that in order to make new iterative model (6) are maximum, to m in (6) _kand S _k ²ask local derviation respectively:

${\hat{m}}_k=\frac{{\mathrm{\&Sigma;}}_{j=1}^N{\widehat{\mathrm{\&gamma;}}}_{\mathrm{jk}}{r}_j}{{\mathrm{\&Sigma;}}_{j=1}^N{\widehat{\mathrm{\&gamma;}}}_{\mathrm{jk}}}---\left(8\right)$

${{\hat{S}}_k}^2=\frac{{\mathrm{\&Sigma;}}_{j=1}^N{\widehat{\mathrm{\&gamma;}}}_{\mathrm{jk}}{({\mathrm{\&gamma;}}_j-m_k)}^2}{{\mathrm{\&Sigma;}}_{j=1}^N{\widehat{\mathrm{\&gamma;}}}_{\mathrm{jk}}}---\left(9\right)$

${\widehat{\mathrm{\&alpha;}}}_k=\frac{{\mathrm{\&Sigma;}}_{j=1}^N{\widehat{\mathrm{\&gamma;}}}_{\mathrm{jk}}}{N}---\left(10\right)$

Expect that fuzzy partitioning specific implementation process is as follows based on maximization log-likelihood probability:

1) output of fingerprint cluster is utilized, i.e. the characteristic fingerprint vector { m of regional area k _k| the reference point set omega of k ∈ K}, regional area k _kreference point fingerprint vector { the r of (k ∈ K) and correspondence thereof _j| j ∈ Ω _kgenerate the initial parameter of gauss hybrid models, order sub-model parameter by formula (8), (9), (10) calculating k class:

2) a jth observation signal r is upgraded according to formula (7) _jto the participation of a kth local class locating area signal distributions model

The wherein received signals fingerprint r of reference point j _jthe posterior probability being subordinate to kth class Gauss sub-model is:

$P\left(r_j\right|{\mathrm{\&Phi;}}_k)={\mathrm{\&alpha;}}_k\mathrm{\&phi;}\left(r_j\right|{\mathrm{\&Phi;}}_k)=\frac{{\mathrm{\&alpha;}}_k}{\sqrt{2\mathrm{\&pi;}}{S}_k}\exp (-{\frac{(r_j-m_k)}{2{S_k}^2}}^2)---\left(11\right)$

The received signals fingerprint r of reference point j _jto the posterior probability of Gauss's sub-model of all classes be:

$P\left(r_j\right|\mathrm{\&Phi;})={\mathrm{\&Sigma;}}_{k=1}^K{\mathrm{\&alpha;}}_k\mathrm{\&phi;}\left(r\right|{\mathrm{\&Phi;}}_k)={\mathrm{\&Sigma;}}_{k=1}^K{\mathrm{\&alpha;}}_k\frac{1}{\sqrt{2\mathrm{\&pi;}}{S}_k}\exp \left(\frac{{(r_j-m_k)}^2}{2{S_k}^2}\right)---\left(12\right)$

3) according to the participation after renewal the sub-model parameter of k class is recalculated with formula (8), (9), (10):

4) 2,3 are repeated, until $|{\hat{m}}_{k\left(\mathrm{new}\right)}-{\hat{m}}_k|<{\mathrm{\&omega;}}_1,|{{\hat{S}}_{k\left(\mathrm{new}\right)}}^2-{{\hat{S}}_k}^2|<{\mathrm{\&omega;}}_2,|{\widehat{\mathrm{\&alpha;}}}_{k\left(\mathrm{new}\right)}-{\widehat{\mathrm{\&alpha;}}}_k|<{\mathrm{\&omega;}}_3$

5) stop parameter calculating, algorithm stops.The now participation of a N number of reference point K class and the parameter Φ of k Gauss's sub-model _k(k ∈ K) trend is stable.Record the participation vector of N number of reference point the parameter of k Gauss's sub-model ${\mathrm{\&Phi;}}_k=\{{\hat{m}}_{k\left(\mathrm{new}\right)},{{\hat{S}}_{k\left(\mathrm{new}\right)}}^2,{\widehat{\mathrm{\&alpha;}}}_{k\left(\mathrm{new}\right)}\},(k\&Element;K).$

For select the methods analyst in local positioning region according to class label in flow process (5) and implementation process as follows:

1, the method in local positioning region is selected according to class label

The object of this method is the regional area that the signal vector obtained according to terminal during location finds terminal place fast in whole region, position by the reference point terminal near terminal, and general way is to each reference point matched signal one by one in whole region, efficiency is very low, and easily because signal error of floating causes larger matching error, very high to the stability requirement of signal, the inventive method is according to the fingerprint classification method proposed before, the signal that mobile terminal obtains can be mated with all kinds of characteristic fingerprint, this signal is put under local class, positioned by the reference point terminal in the class of local, greatly reduce fingerprint matching number of times, improve location efficiency and speed.Specific implementation method is as follows:

1) handheld terminal obtains the WIFI signal vector of current anchor point is r _locati.n。R is calculated respectively according to formula (7) _locati.Nto the participation Υ of k class _{location, k}, wherein Gauss's sub-model parameter is by the output Φ expecting soft stroke based on maximization log-likelihood probability _kspecify, then current anchor point signal vector r _locationparticipation vector

2) according to b _locationthe overall participation vector obtained Υ _k∈ (0,1), selects in i ∈ N, participates in the reference point of locating.Concrete grammar is: for b _location, set up locating area collection according to right retrieval, will only to locating area collection middle class k has the reference point of participation to find out, and sets up position reference point set

For utilize in flow process (5) analysis of local locating method and implementation process as follows:

1, based on class participation multi-point joint localization method

Local positioning region is selected according to class label, can find out that the received signals fingerprint of these reference point has identical category feature with positioning signal vector from selected process, from geographic position, these reference point are close with locating point position, the position of anchor point so can be estimated by these reference point, matching error risk is divided by the matching way of comprehensive multiple reference point, can avoid the high matching error risk produced by single reference point.If so adopt average reference degree for each selected reference point, so inequitable for the higher reference point of matching degree, so the method based on matching degree should be adopted, give higher reference degree by reference point high for matching degree, and suitably reduce the reference degree of the relatively poor reference point of matching degree.Because signal intensity and distance are logarithm nonlinear relationship in practice, so negative exponential function should be adopted to embody participation.

The coordinate of anchor point is estimated as:

${\hat{x}}_p=\frac{1}{n}{\mathrm{\&Sigma;}}_{m\&Element;{\mathrm{\&Gamma;}}_{\mathrm{location}}{k}_m{x}_m---\left(13\right)}$

${\hat{x}}_p=\frac{1}{n}{\mathrm{\&Sigma;}}_{m\&Element;{\mathrm{\&Gamma;}}_{\mathrm{location}}{k}_m{y}_m---\left(14\right)}$

Wherein n is location reference point number, k _mfor reference point m is to the reference degree of anchor point, computing method are:

As follows based on class participation multi-point joint localization method process:

1) by formula (15) respectively compute location reference point concentrate the reference degree k of n reference point _m.

2) by the x coordinate of formula (13), (14) difference compute location point with y coordinate

3) complete anchor point coordinate to estimate

Claims (3)

1. the fingerprint cluster multi-point joint indoor orientation method based on WIFI signal intensity, it is characterized in that, comprise collection signal off-line and build library facility and location algorithm practical function, described function is by realizing based on the optimization cluster fuzzy partitioning method under the dynamic virtual point initial cluster center selection algorithm of distance matrix and based on class participation multi-point joint dynamic positioning method, and idiographic flow is as follows:

(1) area to be targeted divides by mode in a grid formation that arrange a reference point at interval of 1 meter, and each lattice point is 1 reference point, divides N number of reference point altogether;

(2) terminal of loading server positioning software is held, successively at reference point place collection signal, by signal intensity vector stored in fingerprint database;

(3) carry out distance matrix calculating according to fingerprint base vector of samples, realize fingerprint classification according to clustering algorithm, obtain the class distribution matrix of category feature vector and reference point received signals fingerprint;

(4) the current WIFI list of mobile subscriber's automatic acquisition to be positioned and respective signal value, sends Location Request to server, and sends WIFI signal vector;

(5) server mates with all kinds of proper vector after obtaining band positioning signal vector, obtains class label, selectes local positioning region, utilize local locating method to calculate estimated position to be positioned according to class label;

(6) estimated position is returned mobile subscriber to be positioned by server.

2. the method for claim 1, it is characterized in that, three parts are comprised for the optimization cluster fuzzy partitioning method under the point of the dynamic virtual based on the distance matrix initial cluster center selection algorithm that fingerprint classification in flow process (3) adopts: based on the dynamic virtual point initial cluster center system of selection of distance matrix, preliminary fingerprint cluster method, maximize log-likelihood probability expectation method and realize cluster fuzzy partitioning method, concrete grammar is as follows:

1) a kind of point of the dynamic virtual based on distance matrix initial cluster center system of selection, according to flow process (1) described mode, the terminal of holding loading server positioning software receives the signal intensity vector r from M WIFI access point in N number of reference point successively _i, vector is made up of M signal intensity component, namely

Reference point set is

U＝{u|u≤N，u∈N ^*}

Each reference point gathers T signal intensity vector, and the signal intensity vector of collection is specifically expressed as:

To T sampling filter process, average as fingerprint stored in fingerprint database, then each reference point place fingerprint is:

Set up reference point raw range matrix D _{n × N}:

Wherein, distance element D _{i, j}=|| r _i-r _j|| ², D _{i, j}=D _{j, i}, this distance matrix is symmetric matrix; A kind of some initial cluster center system of selection of the dynamic virtual based on distance matrix process is as follows:

A. for Distance matrix D _{n × N}if, distance element D _{i, j}< ε, then by i, j point deletion in former reference point set U;

B. add and represent i, the received signals fingerprint of the virtual reference point v of j, v point is upgrade set U=U ∪ v _g-i, j}, g=1,2 ...;

C. upgrade current reference point number N, upgrade Distance matrix D _{n × N}, now the dimension of distance matrix reduces, repeats 1,2, if produce without new virtual point, and, increase and upgrade ε=ε+θ, repeat 1,2;

D. repeat 1,2,3, if ε > is γ, then stop iteration;

E. now reference point set U is initial cluster center set; Corresponding received signals fingerprint is initial cluster center characteristic fingerprint vector { m _j}={ m ₁, m ₂..., m _k;

2) in based on the dynamic virtual point initial cluster center system of selection of distance matrix, initial cluster center { m is established _k| k=1,2..., K}, class number K; Next will according to initial cluster center by the received signals fingerprint of all reference point according to the raw range matrix D between it _{n × N}be divided into K regional area, namely realize received signals fingerprint higher for similarity to be divided into a class; Classification the target that meets be:

Wherein hypotaxis degree ρ _ikvalue represent received signals fingerprint r _iwhether belong to a kth class, d _ikrepresent r _iwith initial cluster center m _kdistance, then:

The method of preliminary fingerprint cluster is as follows:

A. calculating reference point received signals fingerprint r successively _i(i=1,2 ..., N) and cluster centre m _k(k=1,2 ..., K) distance, work as d _ik=min _{l ∈ K}(d _il) time, make ρ _ik=1, i.e. r _ibelong to class k, otherwise ρ _ij=0, i.e. r _ido not belong to class k; Record is about ρ _iksubordinate matrix Q _{n × K};

B. regional area k (kth class) reference point set omega is set up _k(k ∈ K), according to subordinate matrix Q _{n × K}, or ρ _ik=1, then reference point i ∈ Ω _k; Calculate new distance center

If c. | m _{k (new)}-m _k| > τ, then use m _{k (new)}replace the distance center m of last iteration _k, repeat 1,2; Ω is upgraded in 2 _k(k ∈ K);

If d. | m _{k (new)}-m _k|≤τ, then cluster centre m _kconvergence, stops repeating;

E. now m _{k (new)}for the final characteristic fingerprint of regional area k, reference point set omega _kthe final classification reference point set that (k ∈ K) is regional area k;

3) method that log-likelihood probability expectation method realizes cluster fuzzy partitioning is maximized,

The fuzzy partitioning of cluster can be realized by setting up likelihood probability Density Distribution model, determining degree of membership according to probability; Perform by two steps, one is the expectation obtaining eligible probability density function, and two is the parameter values under asking the expectation of maximization conditional probability density function maximum, and comprise degree of membership probability, in Zone Full internal reference examination point, signal distributions model is:

For a kth local class locating area signal distributions model be:

Wherein gaussian probability distribution form is:

For a reference point u _ithe attribute possessed comprises (r _i, b _i), wherein r _ifor fingerprint vector, b _ifor participation vector, b _i=(γ _i1, γ _i2..., γ _iK), γ _k∈ (0,1) indicates current reference point received signals fingerprint and whether belongs to a kth class; b _icannot directly be drawn by observation, the complete likelihood function under model is:

Take the logarithm to this likelihood function, then log-likelihood function is converted to:

The expectation of log-likelihood function is:

A jth observation signal r _jto the participation γ of a kth local class locating area signal distributions model _jk, by its conditional expectation approximate evaluation be:

The participation of reference point is numerically equal to the fingerprint vector of this reference point by a kth posterior probability that local class locating area signal distributions model is specified;

Build the parameter vector Φ of gaussian probability distributed model _k={ α _k, m _k, S _k ², master mould is converted into log-likelihood and expects form, to m in (6) _kand S _k ²ask local derviation respectively:

Process is as follows:

A. the output of fingerprint cluster is utilized, i.e. the characteristic fingerprint vector { m of regional area k _k| the reference point set omega of k ∈ K}, regional area k _kreference point fingerprint vector { the r of (k ∈ K) and correspondence thereof _j| j ∈ Ω _kgenerate the initial parameter of gauss hybrid models, order sub-model parameter by formula (8), (9), (10) calculating k class:

B. a jth observation signal r is upgraded according to formula (7) method _jto the participation of a kth local class locating area signal distributions model

The wherein received signals fingerprint r of reference point j _jthe posterior probability being subordinate to kth class Gauss sub-model is:

The received signals fingerprint r of reference point j _jto the posterior probability of Gauss's sub-model of all classes be:

C. according to the participation after renewal the sub-model parameter of k class is recalculated with formula (8), (9), (10):

D. 2,3 are repeated, until

E. stop parameter calculating, algorithm stops; The now participation of a N number of reference point K class and the parameter Φ of k Gauss's sub-model _k(k ∈ K) trend is stable; Record the participation vector of N number of reference point the parameter of k Gauss's sub-model

3. the method for claim 1, is characterized in that, comprises and selectes local positioning region according to class label, then position method according to class participation multi-point joint as follows for local locating method in flow process (5):

1) local positioning region is selected according to class label,

Process is as follows:

A. handheld terminal obtains the WIFI signal vector of current anchor point is r _location; R is calculated respectively according to formula (7) _locationto the participation γ of k class _{location, k}, wherein Gauss's sub-model parameter by maximize log-likelihood probability expectation method realize cluster fuzzy partitioning output specify, then current anchor point signal vector r _locationparticipation vector

B. according to b _locationthe overall participation vector obtained middle choosing then participates in the reference point of locating; Concrete grammar is: for b _location, set up locating area collection according to Θ _locationright retrieval, will only to locating area collection Θ _locationmiddle class k has the reference point of participation to find out, and sets up position reference point set

2) based on class participation multi-point joint location,

After selecting local positioning region according to class label, estimated the position of anchor point by these reference point, the coordinate of anchor point is estimated as:

Wherein n is location reference point number, k _mfor reference point m is to the reference degree of anchor point, computing method are:

Process is as follows:

A. by formula (15) respectively compute location reference point concentrate the reference degree k of n reference point _m;

B. the x coordinate of formula (13), (14) difference compute location point is pressed with y coordinate

C. complete anchor point coordinate to estimate