CN104507097A - Semi-supervised training method based on WiFi (wireless fidelity) position fingerprints - Google Patents

Abstract

The invention relates to a semi-supervised training method based on WiFi (wireless fidelity) position fingerprints. The method comprises the following steps that an off-line training route in a target region is determined; continuous sampling on sample points in the target region is carried out; the linear interpolation of position information is carried out on the sampling points obtained through sampling according to the position coordinates of starting and ending points of the off-line training route, the timestamp of each sampling point; target region signal intensity distribution modeling is carried out according to the sample points obtained after the steps of the continuous sampling and the linear interpolation; original sample point data and obtained target sample point data build a database. Compared with the prior art, the semi-supervised training method has the advantages that two evaluation indexes including off-line training time and positioning precision are adopted, the off-line training process is evaluated by a fixed point measuring method, the evaluation value of the respective off-line training process is obtained, a signal intensity distribution model is built according to the original sampling data and a target region topological structure, and the fast and accurate database building is realized.

Description

A kind of semi-supervised training method based on WiFi location fingerprint

Technical field

The present invention relates to wireless communication technology field, especially relate to a kind of semi-supervised training method based on WiFi location fingerprint.

Background technology

Along with the fast development of smart mobile phone and wireless network, the location Based service such as such as public safety, first aid, freight transportation obtain paying close attention to more widely.Location-based service under outdoor conditions can provide good result by GPS (Global Position System).Location-based service under indoor conditions has based on localization methods such as WiFi, bluetooth, terminal built-in motion sensors according to signal kinds, and WiFi signal is comparatively stable and acquisition is convenient, is one of indoor positioning signal of comparatively main flow at present.Fingerprint location system based on WiFi generally comprises off-line training step---is used for setting up fingerprint location database, and the tuning on-line stage---carry out current location according to current RSS data and database and calculate two steps.Off-line training step wherein affects the precision of final location, is the key of fingerprint navigation system.

The training method of WiFi fingerprint location system is generally using low, portable strong, the factor such as positioning precision is stable consuming time as standard.Traditional off-line training method is divided into three steps: target area gridding, grid point RSSI gather and the foundation of target area fingerprint base.Traditional off-line training method carries out fingerprint base foundation by the mode of the true position data and RSSI vector that record all reference points, and what we looked is full supervised training.Full supervised training is intended to be undertaken building storehouse by a large amount of data acquisitions, improves coverage and fineness that database describes target area information, to obtaining hi-Fix result.But in full supervised training in a large number the problems such as storehouse efficiency is on the low side, system transplantation sexual deviation of building caused with work consuming consuming time still need badly and research and solve.

Through finding the retrieval of prior art document, consider in the article " WiFi-SLAM Using Gaussian Process Latent Variable Models " (using the WiFi-SLAM algorithm of Gaussian process latent variable model) that the people such as B Ferris, D Fox delivers in IJCAI meeting in 2007 that in target area, signal strength signal intensity proposes a kind of by a small amount of signal strength data in target area according to the regularity of spatial distribution and signal intensity profile according to the relative stability of time, sets up the algorithm of target area topological diagram and signal intensity profile model.But there are two weak points in it: 1) do not cause positioning precision lower containing actual position information in database; 2) the target area topological diagram only carried out according to RSS is drawn, can not the reflection cartographic information of entirely accurate.

Summary of the invention

Object of the present invention be exactly in order to overcome above-mentioned prior art exist defect and a kind of semi-supervised training method based on WiFi location fingerprint is provided, storehouse original sample number is built in employing and positioning precision two evaluation indexes are evaluated off-line training step, achieves off-line training process fast and accurately.

For accurate evaluation off-line training quality, present invention employs training time and positioning precision as evaluation index.The former is the parameter of reaction off-line training efficiency, and the latter is the important parameter of reflection off-line training quality.Training time refers in off-line training step, have recorded the temporal summation that crude sampling process is used; Positioning precision refers at positioning stage, and according to current database, the result positioned in one-point measurement method and actual position are with the departure of rice.

The signal strength signal intensity that locating terminal receives is continuous print in the distribution in space.In transmission of wireless signals model, the distance of signal strength signal intensity distance access point (AP) is negative correlation, when terminal moves to the direction deviating from AP, the signal strength signal intensity received will diminish gradually, otherwise, when terminal moves in face of the direction of AP, the signal strength signal intensity received will grow gradually; And introduce space block and interchannel interference impact after, signal strength signal intensity shows as irregular distribution in target area, but still is consecutive variations.

Object of the present invention can be achieved through the following technical solutions:

Based on a semi-supervised training method for WiFi location fingerprint, it is characterized in that, comprise the following steps:

Step one, according to target area topological structure, carries out training route for target area and divides, and determine the off-line training route in target area;

Step 2, in units of the off-line training route determined, to the continuous sampling carrying out sample point in target area;

Step 3, gathers to above the linear interpolation that gained sample point carries out positional information according to off-line training route terminal position coordinates and each sampled point timestamp;

Step 4, according to the sample point obtained after above-mentioned steps continuous sampling and linear interpolation, carries out target area signal intensity profile modeling;

Step 5, the target sample point data building database that original sample point data and step 4 are obtained;

Step 6, according to database and current signal strength data, selected location algorithm positions.

In bag step one by target area to cover more, few overlapping off-line training route in target area to be determined for principle.

Continuous sampling described in step 2 is specially:

In units of the off-line training route determined, mobile terminal at the uniform velocity moves on training route, carries out continuous sampling simultaneously, and record the terminal position coordinates of current route to the timestamp of each access point signals intensity received and each sampling.

In the continuous sampling of step 2, the terminal sample packages of training route is containing location coordinate information, timestamp information and signal strength information, and the sample of other points only comprises timestamp information and signal strength information.

Spatial linear interpolation described in step 3 is specially:

To determine and under the relatively stable condition of translational speed at route, according to the sampling time stamp information of arbitrary sample point in terminal location coordinate information, terminal sampling time stamp information and route, this sample point position coordinate is calculated, for sample point specifically:

$L_i=L_{\mathrm{start}}+\frac{t_i-t_{\mathrm{staet}}}{t_{\mathrm{end}}-t_{\mathrm{start}}}\×(L_{\mathrm{end}}-L_{\mathrm{start}}),t_{\mathrm{start}}\≤t_i\≤t_{\mathrm{start}}$

Wherein L _start, L _endfor the actual training route terminal position coordinates recorded, t _start, t _endfor the terminal signal strength signal intensity sampling time of physical record stabs, t _ifor the sampling time stamp of certain valid signal strengths sampled point in training route, L _ifor the calculated value of this position coordinates.

Signal intensity profile modeling described in step 4 adopts Gaussian process modeling, specific as follows:

1) by sample information typing, according to sample information computation modeling kernel matrix K:

$K(i,j)={\mathrm{\σ}}_f^2\exp (-\frac{1}{2l^2}{|L_i-L_j|}^2)$

Wherein, σ _frepresentation signal tension variance, l representation signal Strength Space correlation yardstick, l is larger, and equidistant sample point signal strength signal intensity correlation is less, otherwise equidistant sample point signal strength signal intensity correlation is larger, L _i, L _jrepresent the locus coordinate of two sample points in a former n sample point;

2) according to target area topological diagram and former input amendment, distribution situation in region chooses target sample point, and described target sample point is arbitrfary point in target area, calculates the sample point of signal strength data according to distributed model and position coordinates;

3) impact point signal strength data is calculated according to kernel matrix K and aiming spot coordinate data:

$\mathrm{rs}{s}_{*}=K_{*}^t{(k+{\mathrm{\σ}}_S^{2}I)}^{-1}\mathrm{rss}$

Wherein, rss _*represent impact point signal strength signal intensity, k _*represent the correlation matrix of impact point and former n input point position coordinates:

$k_{*}\left(i\right)={\mathrm{\σ}}_f^2\exp (-\frac{1}{2l^2}{|L_{*}-L_i|}^2)$

σ _srepresent Gauss's observation noise, I is unit matrix, and rss represents the signal strength data matrix of a former n input point, L _*for aiming spot coordinate.

Compared with prior art, the present invention has following beneficial effect:

1) adopt continuous sampling and linear interpolation to replace traditional discrete sampling method, greatly reduce off-line training step required time.

2) sample data obtained according to continuous sampling and linear interpolation carries out signal intensity profile modeling, and calculate arbitrfary point signal strength data in target area according to distributed model, reduce off-line training step work consuming on the one hand, can ensure that the description of database to target area signal intensity profile is more excellent simultaneously, positioning stage accuracy is guaranteed.

Accompanying drawing explanation

Fig. 1 is workflow diagram of the present invention;

Fig. 2 is the semi-supervised training route map in target area in embodiments of the invention, and wherein dotted line represents sampling route, and solid line represents wall in region or partition, and stain represents sampled point;

Fig. 3 is spatial linear interpolating method schematic diagram of the present invention;

Fig. 4 is the single access point signals intensity distribution in target area in embodiments of the invention, x, and y-axis represents two-dimensional space reference axis, and z-axis represents signal strength signal intensity, unit dBm;

Fig. 5 is the target area single access point signals intensity targets point schematic diagram in embodiments of the invention, and dotted line represents sampling route, and solid line represents wall in region or partition, and stain represents sampled point, and triangle is impact point distribution signal point.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

Below in conjunction with accompanying drawing, embodiments of the invention are elaborated: the present embodiment is implemented under premised on technical solution of the present invention, give detailed execution mode and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

In the present embodiment, target area structure as shown in Figure 2, is covered by 11 access points in region.Terminal use uniform motion Real-time Collection sample point signal strength signal intensity back and forth between selected route.Use the method for the invention and existing conventional exercises method to sample respectively to build storehouse and position, finally obtain the sample number shown in table 1 and positioning precision comparison diagram.

As shown in Figure 1, the present embodiment comprises the steps:

Step one: according to target area topological structure, carries out training route for target area and divides, in this example, target area is divided into seven routes.

Step 2: carry out continuous sampling to train in units of route, in this example, back and forth repeats continuous sampling totally 10 times, and namely every bar training route comprises 20 groups of continuous sampling data.

Step 3: measure to above the linear interpolation that gained sample point carries out positional information according to off-line training route terminal position coordinates and each sampled point timestamp, obtain effective sample point 862 in this example altogether.

Step 4: the sample point obtained according to abovementioned steps continuous sampling and linear interpolation, carries out target area signal intensity profile modeling.Carried out signal intensity profile modeling respectively to 11 of target area access points in this example, the spatial distribution of each access point signals intensity is as shown in table 1.

Table 1

Step 5, target sample point data building database original sample point data and step 4 obtained, generates 1000 target sample points of coverage goal region trapezoidal shape in target area for each access point in this example.

Step 6, according to database and current signal strength data, selected location algorithm positions.Select K-Nearest-Neighbor (KNN) algorithm as location algorithm in this example.

After described KNN algorithm refers to obtain current signal strength vector in this example, signal strength signal intensity vector in itself and database is contrasted, K sample point minimum with current vectorial Euclidean distance in select-out signal intensity vector space, carries out weighting summation after filtering, obtains current positioning result.

The concrete formula calculating Euclidean distance in signal strength signal intensity vector space is:

$d_{\mathrm{ij}}=\sqrt{{(\mathrm{rs}{s}_{i1}-\mathrm{rs}{s}_{j1})}^2+{(\mathrm{rs}{s}_{i2}-\mathrm{rs}{s}_{j2})}^2+...+{(\mathrm{rs}{s}_{\mathrm{in}}-\mathrm{rs}{s}_{\mathrm{in}})}^2}$

Wherein d _ijrefer to the signal strength signal intensity vector space Euclidean distance of sample point i and sample point j, rss _ijrefer to that sample point i receives the signal strength values of access point j.

Claims (6)

1., based on a semi-supervised training method for WiFi location fingerprint, it is characterized in that, comprise the following steps:

Step one, according to target area topological structure, carries out training route for target area and divides, and determine the off-line training route in target area;

Step 2, in units of the off-line training route determined, to the continuous sampling carrying out sample point in target area;

Step 3, gathers to above the linear interpolation that gained sample point carries out positional information according to off-line training route terminal position coordinates and each sampled point timestamp;

Step 4, according to the sample point obtained after above-mentioned steps continuous sampling and linear interpolation, carries out target area signal intensity profile modeling;

Step 5, the target sample point data building database that original sample point data and step 4 are obtained;

Step 6, according to database and current signal strength data, selected location algorithm positions.

2. the semi-supervised training method based on WiFi location fingerprint according to claim 1, is characterized in that, in bag step one by target area to cover more, few overlapping off-line training route in target area to be determined for principle.

3. the semi-supervised training method based on WiFi location fingerprint according to claim 1, it is characterized in that, the continuous sampling described in step 2 is specially:

In units of the off-line training route determined, mobile terminal at the uniform velocity moves on training route, carries out continuous sampling simultaneously, and record the terminal position coordinates of current route to the timestamp of each access point signals intensity received and each sampling.

4. the semi-supervised training method based on WiFi location fingerprint according to claim 1, it is characterized in that, in the continuous sampling of step 2, the terminal sample packages of training route is containing location coordinate information, timestamp information and signal strength information, and the sample of other points only comprises timestamp information and signal strength information.

5. the semi-supervised training method based on WiFi location fingerprint according to claim 4, is characterized in that, the spatial linear interpolation described in step 3 is specially:

To determine and under the relatively stable condition of translational speed at route, according to the sampling time stamp information of arbitrary sample point in terminal location coordinate information, terminal sampling time stamp information and route, this sample point position coordinate is calculated, for sample point specifically:

$L_i=L_{\mathrm{start}}+\frac{t_i-t_{\mathrm{start}}}{t_{\mathrm{end}}-t_{\mathrm{start}}}\×(L_{\mathrm{end}}-L_{\mathrm{start}}),t_{\mathrm{start}}\≤t_i\≤t_{\mathrm{start}}$

Wherein L _start, L _endfor the actual training route terminal position coordinates recorded, t _start, t _endfor the terminal signal strength signal intensity sampling time of physical record stabs, t _ifor the sampling time stamp of certain valid signal strengths sampled point in training route, L _ifor the calculated value of this position coordinates.

6. the semi-supervised training method based on WiFi location fingerprint according to claim 1, is characterized in that, the signal intensity profile modeling described in step 4 adopts Gaussian process modeling, specific as follows:

1) by sample information typing, according to sample information computation modeling kernel matrix K:

$K(i,j)={\mathrm{\σ}}_f^2\exp (1-\frac{1}{2l^2}{|L_i-L_j|}^2)$

Wherein, σ _frepresentation signal tension variance, l representation signal Strength Space correlation yardstick, l is larger, and equidistant sample point signal strength signal intensity correlation is less, otherwise equidistant sample point signal strength signal intensity correlation is larger, L _i, L _jrepresent the locus coordinate of two sample points in a former n sample point;

2) according to target area topological diagram and former input amendment, distribution situation in region chooses target sample point, and described target sample point is arbitrfary point in target area, calculates the sample point of signal strength data according to distributed model and position coordinates;

3) impact point signal strength data is calculated according to kernel matrix K and aiming spot coordinate data:

${\mathrm{rss}}_{*}={k_{*}}^T{(K+{\mathrm{\σ}}_s^{2}I)}^{-1}\mathrm{rss}$

Wherein, rss _*represent impact point signal strength signal intensity, k _*represent the correlation matrix of impact point and former n input point position coordinates:

$k_{*}\left(i\right)={\mathrm{\σ}}_f^2\exp (-\frac{1}{2l^2}{|L_{*}-L_i|}^2)$

σ _srepresent Gauss's observation noise, I is unit matrix, and rss represents the signal strength data matrix of a former n input point, L _*for aiming spot coordinate.