CN104185275A - Indoor positioning method based on WLAN - Google Patents

Abstract

The invention discloses an indoor positioning method based on WLAN, and belongs to the field of indoor wireless communication and network technologies. The method includes the steps that RSSI data, collected by sampling points, of all APs are preprocessed, and one-dimensional vectors and two-dimensional vectors are extracted from the RSSI data and respectively serve as feature vectors; clustering analysis is carried out on the feature vectors, and an area to be positioned is divided into multiple positioning subareas; classification models corresponding to groups of the feature vectors are respectively trained in terms of the feature vectors in each group; the subarea set with the largest number of votes is selected from all the subareas on the basis of the classification models and in combination with a voting mechanism; the method of two rounds of positioning is adopted to narrow the set range of the subareas, and positioning accuracy is improved. According to the method, spatial distribution characteristics of RSSI are fully mined and utilized, and the problems of wide-range indoor positioning, too large searching and matching space and high computational complexity are solved; a novel positioning model is established, and the problems that in an existing WLAN indoor positioning method, the nonlinear and non-Gaussian statistical properties of RSSI signals, caused by non-line-of-sight transmission effect, RSSI attenuation law abnormity and the like, cannot be learned and adapted to are solved.

Description

A kind of indoor orientation method based on WLAN

Technical field

The localization method that the present invention relates to a kind of WLAN indoor positioning field, belongs to indoor wireless communication and networking technology area.

Background technology

In recent years, along with improving constantly of people's living standard, people also grow with each passing day to the demand of location-based service, as the widespread demand to location in all many-sides such as personal scheduling, asset management, emergency relief, security monitoring, security dispatching, intelligent transportation, digital map navigation, travel guides; Particularly that, under the special applications scenes such as emergency relief, disaster relief emergency command scheduling, it is particularly important that locating information more seems in reply emergency.

Along with the further investigation of general fit calculation machine and Distributed Communication Technology, indoor wireless communication and network technology develop rapidly, derive (the Wireless Local Area Networks based on WLAN, WLAN (wireless local area network)), Bluetooth (bluetooth), the indoor positioning modes such as WSN (wireless sensor network, wireless sensor network), and indoor orientation method based on fingerprint and probabilistic method.

Based on WLAN, Bluetooth, the location technology of WSN etc., by carrying out grid division indoor, and at a large amount of AP of indoor deployment (Access Point, access points), terminal detects RSSI (the Received Signal Strength Indication of the multiple AP that receive in each grid, received signal strength indicator), the signal strength signal intensity difference that each signal node receiving due to diverse location is sent, the characteristic quantity using the RSSI of each node receiving in each grid as this network is to complete location.

Based on the indoor positioning of fingerprint, by gathering the RSSI of different AP in room area, and the address of corresponding WAP (wireless access point) and coordinate are stored in database, terminal use measures wireless signal strength around, it is mated to location in right amount with pre-stored RSSI in database, thus the terminal use's that obtains being positioned coordinate information.

Probabilistic method is utilized the existing training sample in reference point, show that the RSSI signal probability in each reference point distributes.The general Gaussian function that adopts carries out Probability Distribution Fitting, draws average and the bandwidth of the gaussian probability distribution of each reference point.Probabilistic method takes full advantage of the statistical nature of signal distributions, and positioning precision generally wants high compared with weighting nearest neighbor method.

But, their same problems existing separately.Based on the indoor orientation method of fingerprint, in actual applications, for large-scale indoor positioning, Existential Space match search scope is larger, computation complexity is high, memory space requires larger deficiency, and indoor orientation method based on probabilistic method, exist in actual applications the probability distribution of RSSI signal in certain fixing reference point to present non-Gauss, non-linear, multi-modal characteristic, make the probability-distribution function and the actual probability distribution that simulate differ larger, thus larger matching error while causing locating.

Summary of the invention

The technical problem to be solved in the present invention is: what deficiency that overcomes prior art, a kind of indoor orientation method based on WLAN is provided, can reduce match search scope, the forecast model that can obtain tallying with the actual situation again, has reduced computation complexity and time complexity to a certain extent.

The technical problem to be solved in the present invention is: reduce match search scope, set up a kind of forecast model tallying with the actual situation, a kind of indoor orientation method based on WLAN is provided, comprise the following steps realization:

Step 1: the RSSI data preliminary treatment of each AP that sampled point is collected, therefrom extracts a peacekeeping bivector respectively as characteristic vector.

The RSSI data that scan are carried out to necessary preliminary treatment to be comprised: delete the data of be less than-100dB of RSSI, delete the data of non-location AP.The data of the non-location of described deletion AP refer to, delete the RSSI of the AP that is unsuitable for location.What be unsuitable for locating AP is characterized as intensity too low (be less than-95dB of RSSI) or less stable (variance is greater than 20), uses these AP can increase computation complexity, reduces positioning precision, is therefore got rid of.

Adopt Different Extraction Method from initial data, to extract the multiple characteristic vector that can accurately quantize the RSSI regularity of distribution.Comprise the following steps:

(1) by all AP that scan according to MAC Address ascending sort, all initial data that scan when off-line is gathered are according to sampled point numbering corresponding on its collection position mark;

(2) can extract characteristic vector separately according to following two kinds of methods:

A. by the AP combination of two after sequence, be divided into according to MAC Address by AP group, every group of AP is expressed as (AP _i, AP _j) (wherein, 0<i<j≤m, m represents the number of all AP), from mark the initial data of sampled point, extract RSSI vector and the corresponding sampled point of corresponding A P combination;

B. each AP, separately as one group, is divided into m group by all off-line image data according to the MAC Address of AP, and every group of AP is expressed as AP _i(wherein, 0<i≤m, m represents the number of all AP), from mark the initial data of sampled point, extract RSSI one-dimensional vector and the corresponding sampled point of corresponding A P.

Step 2: to characteristic vector cluster analysis, area to be targeted is divided into multiple locators region, every sub regions has reflected a kind of RSSI distribution characteristics.

Taking the characteristic vector of constructing in step 1 as input, carry out cluster analysis using the distance between characteristic vector as measuring similarity function.Optionally, cluster analysis adopts the X-means algorithm that can automatically find clusters number.X-means clustering algorithm has improved K-means algorithm, in the time of the initial computing of algorithm, need not specify in advance cluster numbers K, only need to specify the span [K1 of a K, K2] (K1<K2), algorithm, by find an optimum cluster numbers K in the scope of specifying, is realized clustering.X-means algorithm is taking bayesian information criterion as guidance, and the cluster centre that constantly travels through inhomogeneity bunch represents different signal characteristics, and signal characteristic has reflected the clustering phenomena of signal distributions in a certain region.

Step 3: in conjunction with cluster result, train respectively corresponding disaggregated model separately for every stack features vector; From all subregions, choose in conjunction with " ballot " mechanism the subregion set that poll is the highest based on disaggregated model.Comprising:

Off-line phase, for two kinds of characteristic vectors that building method is constructed that propose in step 2, trains respectively the corresponding SVMs of every feature vectors (Support Vector Machine, the SVM) disaggregated model of every kind of building method.SVM is that the VC that is based upon statistical learning ties up on (VC dimension) theory and structural risk minimization (structural risk minimization) principle basis.SVM is by nicety of grading (to the classification correctness of specific sample) and classification capacity (arbitrary sample is carried out to inerrancy classification) are traded off, to making grader obtain best Generalization Ability.Characteristic value, as the input of svm classifier device, is the abstractdesription to data, therefore characteristic value choose extremely importantly, can reflect accurately that data characteristics to be sorted will directly affect final classifying quality.

On-line stage, extract characteristic of division vector from real time data, read the corresponding svm classifier model that off-line phase trains, according to described support vector polynomial expansion item value, calculate the probability of vector to be sorted corresponding to zones of different, from All Ranges, choose in conjunction with " ballot " mechanism the set of regions R that poll is the highest.

The concrete operations of online positioning stage comprise:

(1) read the svm classifier model training, calculate support vector polynomial expansion item value;

(2) read the current RSSI collecting, extract characteristic of division vector;

(3) by polynomial kernel function, characteristic of division DUAL PROBLEMS OF VECTOR MAPPING is arrived to higher dimensional space, and calculate the probability of vector to be sorted corresponding to zones of different according to described support vector polynomial expansion item value;

(4) for each AP group (AP _i, AP _j), whether every sub regions that judgement marks off is eligible, if exist multiple subregions eligible, SVM model thinks that current device may be in the union of these subregions;

Described qualified region refers to, as AP group (AP _i, AP _j) in the time that the prediction probability of a certain subregion is greater than a certain threshold epsilon (0< ε <1), just think that this region is qualified;

(5) from All Ranges, choose in conjunction with " ballot " mechanism the set of regions R that poll is the highest, concrete steps comprise:

If AP group (AP _i, AP _j) sample data be identified as in a certain region through SVM prediction, this region poll adds 1.Choose from geometrically showing as the locating area that coarseness is used as in the maximum region of capped number of times, the poll in each region should arrive 0 between.

Step 4: adopt two-wheeled location to dwindle set of regions scope, improve positioning precision.Specifically comprise:

(1) read the svm classifier model training, calculate support vector polynomial expansion item value;

(2) read the current RSSI collecting, extract characteristic of division vector, and characteristic of division is carried out to standardization;

(3) by polynomial kernel function, characteristic of division DUAL PROBLEMS OF VECTOR MAPPING is arrived to higher dimensional space, and calculate the probability of vector to be sorted corresponding to zones of different according to described support vector polynomial expansion item value, the probability of regional in the coarseness locating area R therefrom obtaining in selecting step three;

(4) for each AP _iwhether every sub regions that judgement marks off is eligible, and this subregion is the subset of the coarseness locating area R that obtains in step 3, if exist multiple subregions eligible, SVM model thinks that current device may be in the union of these subregions;

Described qualified region refers to, works as AP _iin the time that the prediction probability in a certain region is greater than a certain threshold epsilon (0< ε <1), just think that this region is qualified;

(5) from R, choose in conjunction with " ballot " mechanism the set of regions R ' that poll is the highest, concrete steps comprise: if AP _isample data be identified as in a certain region through SVM prediction, this region poll adds 1.Choose the maximum region of capped number of times and be used as fine-grained locating area from geometrically showing as, the poll in each region should be between 0 to m.

The beneficial effect of technical scheme provided by the invention is: the abundant digging utilization of the present invention the spatial distribution characteristic of RSSI, reduced because of region divide the improper locating area deviation causing; Set up novel location model, solve in existing WLAN indoor orientation method, cannot effectively learn and adapt to RSSI signal due to non-linear, the non-Gaussian statistics characteristic that reason causes such as non line of sight transmission effects, multipath transmisstion effect and RSSI attenuation law be abnormal, and large-scale indoor positioning, search package space is excessive, the problems such as computation complexity height.

Brief description of the drawings

In order to be illustrated more clearly in the technical scheme in the embodiment of the present invention, below the accompanying drawing of required use during embodiment is described is briefly described, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skill in the art, do not paying under the prerequisite of creative work, can also obtain according to these accompanying drawings other accompanying drawing.

Fig. 1 is the inventive method realization flow figure;

Fig. 2 is the cluster flow chart of the inventive method;

Fig. 3 is the another kind of cluster flow chart of the inventive method;

Fig. 4 is the training flow chart of the inventive method;

Fig. 5 is a kind of coarseness positioning flow figure of the inventive method;

Fig. 6 is a kind of fine granularity positioning flow figure of the inventive method.

Embodiment

Below in conjunction with flow chart and specific embodiment, specific embodiments of the present invention are described further.

Fig. 2 is the cluster flow chart of the inventive method, and this flow process belongs to a part for off-line phase.Specifically can comprise the steps:

201, use smart mobile phone high frequency sweep periphery AP signal at each school punctuate, the data format scanning is as shown in table 1.It should be noted that the data number that each school punctuate gathers is fixing, because of acquisition time length different.If current location fails to collect the RSSI of corresponding AP, fill up with-100dB.

Table 1 scan-data form

School punctuate numbering	AP1	AP2	AP3	AP4	AP5	AP6	AP7
								1	-85	-97	-63	-100	-100	-90	-72
1	-83	-92	-65	-100	-98	-85	-69
								2	-70	-73	-95	-82	-63	-100	-100
……	……	……	……	……	……	……	……

202, all AP of extracting data from gathering, according to MAC Address ascending sort.Its object is, the SVM algorithm that positioning stage uses is sequentially relevant with vector, therefore must determine artificially that a kind of vector puts in order.In embodiments of the present invention, use the MAC ascending order of AP to arrange as sort method.

203, the AP combination of two after sequence is become to (AP _i, AP _j) (wherein, 0<i<j≤m, m represents the number of all AP), be divided into according to MAC Address by AP group.From mark each RSSI extracting data of sampled point go out the two-dimentional RSSI vector of corresponding A P combination, as classification initial data.As table 2, shown in table 3.

Table 2 extracts data format

School punctuate numbering	AP1	AP2
			1	-85	-97
1	-83	-92
			2	-70	-73
……	……	……

Table 3 extracts data format

School punctuate numbering	AP2	AP3
			1	-97	-63
1	-92	-65
			2	-73	-95
……	……	……

The vector of 204, constructing taking step 203 is input, and using the distance between vector as measuring similarity function, employing can find that the X-means algorithm of clusters number carries out cluster analysis automatically.Record respectively each two-dimentional AP and combine the dividing condition to whole locating area.

The specific implementation process of X-means algorithm cluster analysis is as follows:

Step1. specify clusters number k scope [k _min, k _max], and initialization k=k _min.The scope of K is selected according to the size in reality region to be measured, and the scope of every sub regions is at 200m ²to 700m ², calculate [k with the method _min, k _max];

Step2. in the set of eigenvectors EV extracting, choose at random k AP number of combinations strong point u from step 202 ₁, u ₂, u ₃... u _kas initial cluster center; Set of eigenvectors EV, as table 2, shown in table 3, selects k characteristic vector as initial center;

Step3. for each the AP number of combinations strong point x in set of eigenvectors EV ⁱ, judge the class bunch under it according to similarity, wherein s (arg ₁, arg ₂) be similarity computing function;

Step4. repeat above process, all data points be all assigned to the most similar class bunch, thus by all AP group data points all Preliminary division in corresponding class bunch;

Step5. for each class bunch, recalculate its cluster centre, wherein, c ⁽ⁱ⁾represent this data point x ⁱtype under preliminary identification; c ⁽ⁱ⁾=j refers to: if data point x ⁱbelong to class bunch j, (c ⁽ⁱ⁾=j)=1, otherwise (c ⁽ⁱ⁾=j)=0; This center represents the weighted average center position of each class bunch;

Step6. calculation criterion function, wherein x ⁱthe data point of data centralization, u _jit is the cluster centre of class bunch j; K refers to the number of cluster centre;

If Step7. criterion function no longer changes and turns to Step8, illustrate that this cluster result is stable; Otherwise jump to Step3, re-start cluster;

Step8. carry out Further Division and calculate the bayesian information criterion BIC before and after dividing gathering each class bunch _pre, BIC _post; Bayesian information criterion (Bayesian Information Criterions, BIC) be an important component part of bayesian theory, can evaluate the different models in same data set based on posterior probability, be suitable as that to choose complexity lower and data set is described to the reference frame of good model.

Wherein for Clustering Model corresponding to clusters number k, the computing formula of bayesian information criterion: wherein, EV is the set of the characteristic vector that extracts in step 202; R is the number of the characteristic vector that comprises in EV, and characteristic vector number equals the number of the combination of the RSSI in all positions that AP group collects herein; P represents number of parameters, is called Schwarz criterion, and its computing formula is p=k+kd in the present invention, and wherein, d is the dimension of characteristic vector in EV, i.e. d=2; can regard the punishment to Clustering Model complexity as; clustering Model M _kmaximum posteriori log-likelihood estimation on characteristic vector set EV, its computing formula is as shown in the formula institute

$\hat{l}\left(\mathrm{EV}\right)=-\frac{R}{2}\log \left(2\mathrm{\&pi;}\right)-\frac{R\&CenterDot;d}{2}\log \left({\widehat{\mathrm{\&sigma;}}}^2\right)-\frac{R-k}{2}$

Wherein, ${\widehat{\mathrm{\&sigma;}}}^2=\frac{1}{R-k}{\mathrm{\&Sigma;}}_i{({\mathrm{ev}}_i-u_{\left(i\right)})}^2,$ U _(i)for the cluster centre of class bunch i;

If Step9. BIC _pre> BIC _post, whether highly than original score watch results model, mark height is accepted division, turns to Step10, otherwise makes k=k+1 and jump to Step8;

If Step10. k > k _max, need to re-start cluster, turn to Step7; Otherwise make k=k+1 and jump to Step2, calculate the cluster situation that increases a class;

Step11. choose the dividing mode of BIC maximum as cluster result;

Suppose that M is the model set that different clusters number k are corresponding, has be best Clustering Model.Each type is expressed as a kind of signal characteristic, and signal characteristic has reflected the clustering phenomena of signal distributions in a certain region.

Fig. 3 is the another kind of cluster flow chart of the inventive method, and this flow process belongs to a part for off-line sample phase.Specifically can comprise the steps:

301, use smart mobile phone high frequency sweep periphery AP signal at each school punctuate, the data format scanning is as shown in table 1.It should be noted that the data number that each school punctuate gathers is fixing, because of acquisition time length different.If current location fails to collect the RSSI of corresponding AP, fill up with-100dB.

302, AP is divided into m group (m represents the number of all AP) according to MAC Address.From mark each RSSI extracting data of sampled point go out the one dimension RSSI vector of corresponding A P, as classification initial data.As table 4, shown in table 5.

Table 4 extracts data format

School punctuate numbering	AP1
		1	-85
1	-83
		2	-70
……	……

Table 5 extracts data format

School punctuate numbering	AP2
		1	-97
1	-92

2	-73
		……	……

The vector of 303, constructing taking step 302 is input, carry out cluster analysis using the distance between vector as measuring similarity function, cluster analysis adopts the X-means algorithm that can automatically find clusters number, clustering method and step 203 are similar, and difference is that the bivector that the characteristic vector in step 203 is combined by AP changes the one-dimensional vector of single AP into.Signal mode has reflected the clustering phenomena of signal distributions in a certain region, records respectively the dividing condition of each AP to whole locating area.It should be noted that, for same locating area, different AP are different to the division meeting in this region, its reason is the spatially wide apart of deployed position of these AP, be subject to non line of sight transmission effects, multipath transmisstion effect and RSSI attenuation law extremely different, therefore may produce difference to division result.

Fig. 4 is the training flow chart of the inventive method, and this flow process belongs to a part for off-line sample phase.Specifically can comprise the steps:

401, extract the two-dimentional RSSI vector of AP combination of two at each calibration point, as table 2, shown in table 3, and the numbering of calibration point is replaced with to the classification numbering after corresponding cluster.

402, extract the one dimension RSSI vector of single AP at each calibration point, as table 4, shown in table 5, and the numbering of calibration point is replaced with to the classification numbering after corresponding cluster.

403, to step 401,402 vectors that obtain carry out respectively SVM training, calculate the characteristic of division value of SVMs.The follow-up judgement initialization scope that is calculated as of characteristic of division value provides Data support with the validity of dwindling orientation range.The characteristic of division that the present embodiment is chosen is exactly step 401, the 402 RSSI vectors that obtain.

Fig. 5 is a kind of coarseness positioning flow figure of the inventive method, and this flow process belongs to a part for on-line stage.Specifically can comprise the steps:

501, load the each two-dimentional AP group (AP training _i, AP _j) svm classifier model, read the current RSSI collecting, according to the MAC Address ascending sort of AP, then extract every group of AP and combine as class vector.It should be noted that SVM algorithm is relevant to vector order, therefore must determine that a kind of vector puts in order artificially.In this example, use the MAC ascending order of AP to arrange as sort method.

502, the class vector that AP step 501 being extracted is combined to form, uses corresponding SVM model to predict it, obtains respectively every group of AP probability at regional under its corresponding region partition mode.Because current location may be in the edge in multiple regions, or because certain or some AP in this two dimension AP combination causes RSSI fluctuation owing to being subject to the reasons such as non line of sight transmission effects, multipath transmisstion effect and RSSI attenuation law be abnormal, may there is all satisfactory situations of multiple regions, can choose in accordance with the following methods:

$\mathrm{Area}({\mathrm{AP}}_i,{\mathrm{AP}}_j)=\underset{k=1}{\overset{s}{\&cup;}}{\mathrm{area}}_k$

Each AP characteristic vector extracting of sampling may have multiple predicting the outcome and meet the requirements, each sub regions that predicts the outcome and corresponding to this AP group, area to be targeted is marked off after corresponding SVM model prediction.In above formula, s represents the qualified number that predicts the outcome, and represents that SVM qualification of model current device may be in several sub regions; Area _krepresent k qualified region, represent SVM qualification of model current device may in which subregion.Area (AP _i, AP _j) represent AP group (AP _i, AP _j) region at determined current location place, represent that SVM model thinks that current device may be in the union of these subregions.The choosing method of satisfactory subregion is, is not less than certain threshold epsilon (0< ε <1) if dope current characteristic vector at the probability of certain sub regions, just thinks that this subregion meets the requirements.In this example, the ε=1/n choosing, n represents that this AP combines the subregion number marking off.

503, to all AP group (AP that obtain in step 502 _i, AP _j) Area (AP _i, AP _j) employing " ballot " mode compute location result.May be in a certain region if sample data process step 502 prediction of certain AP combination is identified as, this region poll adds 1.Travel through the Area (AP of all AP combinations _i, AP _j) and ballot, the maximum region of selected poll is as the coarseness locating area of location, and the poll in each region arrives 0 between.Choose (the AP by Area from geometrically showing as _i, AP _j) region that degree of covering is maximum is used as the locating area of coarseness.If the poll of All Ranges is all less than a certain threshold xi, think and locate unsuccessfully, finish location; If the poll that has multiple regions at most and be greater than ξ, is asked the union in these regions, as the coarseness locating area of location.Participate in calculating because normal location at least needs 3 AP, therefore in this example, ξ value is 4.

Fig. 6 is a kind of fine granularity positioning flow figure of the inventive method, and this flow process belongs to a part for on-line stage.Specifically can comprise the steps:

601, load the svm classifier model of the each AP training before, read the current RSSI collecting, this RSSI is formed to the class vector of an one dimension, the svm classifier model of the each AP training before using is predicted it, obtain respectively each AP probability at regional under its corresponding region partition mode, therefrom choose the probability of regional in the coarseness locating area R that previous step obtains.Because current location may be in the edge in multiple regions, or because this AP causes RSSI fluctuation owing to being subject to the reasons such as non line of sight transmission effects, multipath transmisstion effect and RSSI attenuation law be abnormal, may there is all more satisfactory situations of multiple regions, can choose in accordance with the following methods:

$\mathrm{Area}\left({\mathrm{AP}}_i\right)=\underset{k=1}{\overset{s}{\&cup;}}{\mathrm{area}}_k({\mathrm{area}}_k\&SubsetEqual;R)$

Each AP characteristic vector extracting of sampling may have multiple predicting the outcome and meet the requirements, each sub regions that predicts the outcome and corresponding to this AP, area to be targeted is marked off after corresponding SVM model prediction.In above formula, s represents the qualified number that predicts the outcome, and represents that SVM qualification of model current device may be in several sub regions; Area _krepresent k qualified region, represent SVM qualification of model current device may in which subregion, this region must be the subset of the coarseness locating area R that obtains in step 4.Area (AP _i) represent AP _ithe region at determined current location place, represents that SVM model thinks that current device may be in the union of these subregions.The choosing method of satisfactory subregion is, is not less than certain threshold epsilon (0< ε <1) if dope current characteristic vector at the probability of certain sub regions, just thinks that this subregion meets the requirements.In this example, the ε=1/n choosing, n represents the subregion number that this AP marks off.

602, to all AP in step 601 _iarea (AP _i) adopt " ballot " mode to calculate, if the sample data of certain AP is predicted and is identified as in a certain region through step 601, this region poll adds 1.Travel through the Area (AP of all AP _i) and ballot, the maximum region of selected poll is as the fine granularity locating area of location, and the poll in each region is between 0 to m.Choose (the AP by Area from geometrically showing as _i) region that degree of covering is maximum is used as the locating area of coarseness.If the poll of All Ranges is all less than a certain threshold xi, think and locate unsuccessfully, finish location; If the poll that has multiple regions at most and be greater than ξ, is asked the union in these regions, obtain its center point coordinate and radius, as final locating area.Participate in calculating because normal location at least needs 3 AP, therefore in this example, ξ value is 4.

Non-elaborated part of the present invention belongs to techniques well known.

The above; be only part embodiment of the present invention, but protection scope of the present invention is not limited to this, in the technical scope that any those skilled in the art disclose in the present invention; the variation that can expect easily or replacement, within all should being encompassed in protection scope of the present invention.

Claims (7)

1. the indoor orientation method based on WLAN, is characterized in that performing step is as follows:

Step 1: the RSSI data preliminary treatment of each AP that sampled point is collected, therefrom extracts a peacekeeping bivector respectively as characteristic vector;

Step 2: to characteristic vector cluster analysis, area to be targeted is divided into multiple locators region;

Step 3: in conjunction with cluster result, train respectively corresponding disaggregated model separately for every stack features vector; From all subregions, choose in conjunction with " ballot " mechanism the subregion set that poll is the highest based on disaggregated model;

Step 4: adopt two-wheeled location to dwindle subregion range of convergence, improve positioning precision.

2. the indoor orientation method based on WLAN according to claim 1, it is characterized in that: the RSSI data preliminary treatment of each AP that described step 1 collects sampled point, comprise: delete the too low data of RSSI, delete the data of non-location AP, fill up the RSSI data that are not scanned;

The too low data of described deletion RSSI refer to, RSSI intensity is deleted lower than the data of a certain threshold value; The data of the non-location of described deletion AP refer to, delete the RSSI of the AP that is unsuitable for location, and what be unsuitable for locating AP is characterized as intensity too low (being be less than-95dB of RSSI) or less stable (being that variance is greater than 20).

3. the indoor orientation method based on WLAN according to claim 1, is characterized in that: described step 1 extracts a peacekeeping bivector respectively as characteristic vector from preprocessed data, comprising:

(1) by all AP that scan according to MAC Address ascending sort;

(2) extract a peacekeeping bivector as characteristic vector according to following two kinds of methods:

A. by the AP combination of two after sequence, AP is divided into according to MAC Address group, every group of AP is expressed as (AP _i, AP _j), wherein, 0<i<j≤m, m represents the number of all AP, goes out vector that these AP constitute as characteristic vector from pretreated extracting data;

B. each AP, separately as one group, is divided into m group by all off-line image data according to the MAC Address of AP, and every group of AP is expressed as AP _i, wherein, 0<i≤m, m represents the number of all AP, the vector that goes out these AP formations from pretreated extracting data is as characteristic vector.

4. the indoor orientation method based on WLAN according to claim 1, it is characterized in that: in described step 2, to characteristic vector cluster analysis, area to be targeted is divided into multiple locators region, concrete steps are: taking the characteristic vector of constructing in step 2 as input, carry out cluster analysis using the distance between characteristic vector as measuring similarity function, cluster analysis adopts the X-means algorithm that can automatically find clusters number.

5. the indoor orientation method based on WLAN according to claim 1, is characterized in that: described step 3, and specific implementation process comprises off-line phase and on-line stage;

Off-line phase, for two kinds of characteristic vectors that building method is constructed that propose in step 2, trains respectively the corresponding SVMs of every feature vectors (Support Vector Machine, the SVM) disaggregated model of every kind of building method;

On-line stage, extract characteristic of division vector from real time data, read the svm classifier model that off-line phase trains, according to described support vector polynomial expansion item value, calculate the probability of vector to be sorted corresponding to zones of different, from All Ranges, choose in conjunction with " ballot " mechanism the set of regions R that poll is the highest;

Described voting mechanism refers to, if AP group (AP _i, AP _j) sample data be identified as in a certain region through SVM prediction, this region poll adds 1; Travel through the EV (AP of all AP groups _i, AP _j) and ballot, the maximum region of selected poll is as the coarseness locating area of location, and the poll in each region should arrive 0 between.

6. the indoor orientation method based on WLAN according to claim 5, is characterized in that: the concrete operations of described online positioning stage comprise:

(1) read the svm classifier model training, calculate support vector polynomial expansion item value;

(2) read the current RSSI collecting, extract characteristic of division vector;

(3) by polynomial kernel function, characteristic of division DUAL PROBLEMS OF VECTOR MAPPING is arrived to higher dimensional space, and calculate the probability of vector to be sorted corresponding to zones of different according to described support vector polynomial expansion item value;

(4) for each AP group (AP _i, AP _j), whether every sub regions that judgement marks off is eligible, if exist multiple subregions eligible, SVM model thinks that current device may be in the union of these subregions.

7. the indoor orientation method based on WLAN according to claim 1, is characterized in that: described step 4, and adopt two-wheeled location to dwindle set of regions scope, specific implementation is:

(1) read the svm classifier model training, calculate support vector polynomial expansion item value;

(2) read the current RSSI collecting, extract characteristic of division vector, and characteristic of division is carried out to standardization;

(3) by polynomial kernel function, characteristic of division DUAL PROBLEMS OF VECTOR MAPPING is arrived to higher dimensional space, and calculate the probability of vector to be sorted corresponding to zones of different according to described support vector polynomial expansion item value, the probability of regional in the coarseness locating area R therefrom obtaining in selecting step three;

(4) for each AP _iwhether every sub regions that judgement marks off is eligible, and this subregion is the subset of the coarseness locating area R that obtains in step 3, if exist multiple subregions eligible, SVM model thinks that current device may be in the union of these subregions;

(5) from R, choose in conjunction with " ballot " mechanism the set of regions R ' that poll is the highest, concrete steps comprise: if AP _isample data be identified as in a certain region through SVM prediction, this region poll adds 1, according to the locating area ballot of each AP, the maximum region of selected poll is as the fine granularity locating area of location, the poll in each region should be between 0 to m.