CN104185275B - A kind of indoor orientation method based on WLAN - Google Patents

Abstract

The invention discloses a WLAN-based indoor positioning method, which belongs to the technical field of indoor wireless communication and network. The method includes: preprocessing the RSSI data of each AP collected at the sampling point, extracting one-dimensional and two-dimensional vectors as feature vectors respectively; performing cluster analysis on the feature vectors to divide the area to be positioned into multiple positioning sub-regions; Group feature vectors to train their corresponding classification models; based on the classification model combined with the "voting" mechanism, select the sub-region set with the highest number of votes from all sub-regions; use two rounds of positioning to narrow the scope of the sub-region set to improve positioning accuracy. The present invention fully exploits the spatial distribution characteristics of RSSI, solves the problems of large-scale indoor positioning search matching space and high computational complexity; establishes a new positioning model, and solves the problem that the existing WLAN indoor positioning methods cannot effectively learn and adapt Due to non-line-of-sight transmission effects and abnormal RSSI attenuation rules, RSSI signals have nonlinear and non-Gaussian statistical characteristics.

Description

A WLAN-based Indoor Positioning Method

technical field

The invention relates to a positioning method in the field of WLAN indoor positioning, and belongs to the technical field of indoor wireless communication and network.

Background technique

In recent years, with the continuous improvement of people's material living standards, people's demand for location-based services is also increasing day by day, such as in personnel dispatching, asset management, emergency rescue, safety monitoring, safety dispatching, intelligent transportation, map navigation, travel guides, etc. There is a wide demand for positioning; especially in emergency situations, such as emergency rescue, disaster relief emergency command and dispatch and other special application scenarios, positioning information is even more important.

With the in-depth research of ubiquitous computer and distributed communication technology, the rapid development of indoor wireless communication and network technology, derived from WLAN (Wireless Local Area Networks, wireless local area network), Bluetooth (Bluetooth), WSN (wireless sensor network, wireless sensor network) and other indoor positioning methods, as well as indoor positioning methods based on fingerprints and probabilistic methods.

Based on WLAN, Bluetooth, WSN and other positioning technologies, through indoor grid division and a large number of APs (Access Points, Access Points) deployed indoors, the terminal detects multiple APs received in each grid RSSI (Received Signal Strength Indication, Received Signal Strength Indication), because the signal strength of each signal node received at different locations is different, the RSSI of each node received in each grid is used as the characteristic quantity of the network. Finish positioning.

Indoor positioning based on fingerprints, by collecting the RSSI of different APs in the indoor area, and storing the address and coordinates of the corresponding wireless access points in the database, the terminal user measures the surrounding wireless signal strength and compares it with the pre-stored in the database. The appropriate amount of RSSI is used for matching and positioning, so as to obtain the coordinate information of the positioned end user.

The probability method uses the existing training samples on the reference points to obtain the probability distribution of RSSI signals on each reference point. Generally, the Gaussian function is used for probability distribution fitting, and the mean value and bandwidth of the Gaussian probability distribution of each reference point are obtained. The probability method makes full use of the statistical characteristics of the signal distribution, and its positioning accuracy is generally higher than that of the weighted nearest neighbor method.

However, they also have their own problems. The fingerprint-based indoor positioning method, in practical applications, has the disadvantages of large space matching search range, high computational complexity, and large storage space requirements for large-scale indoor positioning, while the indoor positioning method based on the probability method, in In practical applications, the probability distribution of the RSSI signal at a fixed reference point presents non-Gaussian, nonlinear, and multi-modal characteristics, which makes the fitted probability distribution function quite different from the actual probability distribution, resulting in time-consuming positioning problems. large matching error.

Contents of the invention

The technical problem to be solved by the present invention is: to overcome the deficiencies of the prior art, to provide a WLAN-based indoor positioning method, which can not only reduce the matching search range, but also obtain a prediction model in line with the actual situation, which reduces the computational complexity to a certain extent. complexity and time complexity.

The technical problem to be solved in the present invention is: reduce matching search scope, set up a kind of prediction model that conforms to actual situation, provide a kind of indoor positioning method based on WLAN, comprise the following steps and realize:

Step 1: Preprocess the RSSI data of each AP collected at the sampling point, and extract one-dimensional and two-dimensional vectors as feature vectors respectively.

Necessary preprocessing of the scanned RSSI data includes: deleting data whose RSSI is less than -100dB, and deleting data of non-located APs. The deleting the data of the non-positioned AP refers to deleting the RSSI of the AP that is not suitable for positioning. APs that are not suitable for positioning are characterized by low strength (RSSI less than -95dB) or poor stability (variance greater than 20). Using these APs will increase computational complexity and reduce positioning accuracy, so they are excluded.

Different extraction methods are used to extract a variety of feature vectors that can accurately quantify the distribution of RSSI from the original data. Include the following steps:

(1) Sort all the scanned APs in ascending order according to the MAC address, and mark all the raw data scanned during offline collection according to the corresponding sampling point number on the collection position;

(2) The respective feature vectors can be extracted according to the following two methods:

a. Combine the sorted APs in pairs, that is, divide the APs according to their MAC addresses Each group of APs is expressed as (AP _i , AP _j ) (wherein, 0<i<j≤m, m represents the number of all APs), and the RSSI of the corresponding AP combination is extracted from the original data marked with the sampling points Vector and corresponding sampling points;

b. Each AP is regarded as a group alone, that is, all offline collected data is divided into m groups according to the MAC address of the AP, and each group of APs is represented as AP _i (wherein, 0<i≤m, m represents the number of all APs), from The RSSI one-dimensional vector corresponding to the AP and the corresponding sampling points are extracted from the original data marked with the sampling points.

Step 2: clustering and analyzing the feature vectors, dividing the area to be located into a plurality of positioning sub-areas, and each sub-area reflects a RSSI distribution feature.

The feature vector constructed in step 1 is used as input, and the distance between feature vectors is used as the similarity measure function for cluster analysis. Optionally, the cluster analysis adopts the X-means algorithm that can automatically discover the number of clusters. The X-means clustering algorithm improves the K-means algorithm. In the initial operation of the algorithm, there is no need to pre-specify the number of clusters K, just specify a value range of K [K1,K2] (K1<K2), the algorithm will Find an optimal cluster number K within the range of , and realize cluster division. Guided by the Bayesian information criterion, the X-means algorithm continuously traverses the cluster centers of different clusters to represent different signal characteristics, and the signal characteristics reflect the aggregation phenomenon of signal distribution in a certain area.

Step 3: For each group of feature vectors combined with the clustering results, train the respective classification models; based on the classification model combined with the "voting" mechanism, select the sub-region set with the highest number of votes from all sub-regions. These include:

In the offline stage, for the feature vectors constructed by the two construction methods proposed in step 2, respectively train the Support Vector Machine (SVM) classification model corresponding to each feature vector of each construction method. SVM is based on the VC dimension theory of statistical learning and the principle of structural risk minimization. SVM trades off the classification accuracy (correctness of classification for a specific sample) and classification ability (no error classification for any sample) in order to obtain the best generalization ability of the classifier. As the input of the SVM classifier, the eigenvalue is an abstract description of the data, so the selection of the eigenvalue is very important. Whether it can accurately reflect the characteristics of the data to be classified will directly affect the final classification effect.

In the online stage, the classification feature vector is extracted from the real-time data, and the corresponding SVM classification model trained in the offline stage is read. According to the support vector polynomial expansion item value, the probability that the vector to be classified corresponds to different regions is calculated, combined with the "voting" mechanism Select the region set R with the highest number of votes from all regions.

Specific operations in the online positioning phase include:

(1) Read the trained SVM classification model and calculate the support vector polynomial expansion item value;

(2) read the currently collected RSSI, and extract the classification feature vector;

(3) map the classification feature vector to a high-dimensional space by the polynomial kernel function, and calculate the probability that the vector to be classified corresponds to different regions according to the polynomial expansion item value of the support vector;

(4) For each AP group (AP _i , AP _j ), judge whether each divided sub-area meets the conditions. If there are multiple sub-areas that meet the conditions, the SVM model believes that the current device may be in the union of these sub-areas Inside;

The eligible area means that when the predicted probability of an AP group (AP _i , AP _j ) in a certain sub-area is greater than a certain threshold ε (0<ε<1), the area is considered to be eligible;

(5) Combine the "voting" mechanism to select the region set R with the highest number of votes from all regions, and the specific steps include:

If the sample data of the AP group (AP _i , AP _j ) is identified as being in a certain area after SVM prediction, the number of votes in this area will be increased by 1. Geometrically, the region with the most coverage is selected as a coarse-grained positioning region, and the number of votes in each region should be between 0 and between.

Step 4: Use two-wheel positioning to narrow the scope of the area set and improve the positioning accuracy. Specifically include:

(1) Read the trained SVM classification model and calculate the support vector polynomial expansion item value;

(2) Read the currently collected RSSI, extract the classification feature vector, and standardize the classification feature;

(3) Map the classification feature vector to a high-dimensional space through the polynomial kernel function, and calculate the probability that the vector to be classified corresponds to different regions according to the polynomial expansion item value of the support vector, and select the coarse-grained positioning region obtained in step 3 therefrom The probability of each region in R;

(4) For each AP _i , judge whether each divided sub-area meets the conditions, and this sub-area is a subset of the coarse-grained positioning area R obtained in step 3. If there are multiple sub-areas that meet the conditions, then SVM The model believes that the current device may be in the union of these sub-areas;

The eligible area means that when the predicted probability of AP _i in a certain area is greater than a certain threshold ε (0<ε<1), the area is considered to be eligible;

(5) Select the region set R' with the highest number of votes from R in combination with the "voting" mechanism. The specific steps include: if the sample data of AP _i is identified as being in a certain region after SVM prediction, the number of votes in this region is increased by 1. Geometrically, the region that is covered the most times is selected as a fine-grained positioning region, and the number of votes in each region should be between 0 and m.

The beneficial effects of the technical solution provided by the present invention are: the present invention fully exploits and utilizes the spatial distribution characteristics of RSSI, reduces the positioning area deviation caused by improper area division; establishes a new positioning model, and solves the problem that the existing WLAN indoor positioning method cannot Effectively learn and adapt to the nonlinear and non-Gaussian statistical characteristics of RSSI signals caused by non-line-of-sight transmission effects, multipath propagation effects, and abnormal RSSI attenuation laws, as well as large-scale indoor positioning. The search and matching space is too large and the calculation is complicated high-level issues.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

Fig. 1 is the realization flow chart of the method of the present invention;

Fig. 2 is the clustering flowchart of the inventive method;

Fig. 3 is another kind of clustering flowchart of the method of the present invention;

Fig. 4 is the training flowchart of the inventive method;

Fig. 5 is a kind of coarse-grained positioning flowchart of the method of the present invention;

Fig. 6 is a fine-grained positioning flow chart of the method of the present invention.

detailed description

The specific implementation of the present invention will be further described below in conjunction with the flowchart and specific examples.

Fig. 2 is a clustering flowchart of the method of the present invention, which is part of the offline stage. Specifically, the following steps may be included:

201. At each calibration point, use a smart phone to scan the surrounding AP signals at high frequency, and the scanned data format is shown in Table 1. It should be noted that the number of data pieces collected by each calibration point is not fixed, and varies with the length of collection time. If the current location fails to collect the RSSI of the corresponding AP, fill it with -100dB.

Table 1 Scanned data format

Calibration point number 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. Extract all APs from the collected data, and sort them in ascending order of MAC addresses. The purpose is that the SVM algorithm used in the positioning stage is related to the order of the vectors, so a vector order must be artificially determined. In this embodiment of the present invention, MACs of APs are sorted in ascending order as a sorting method.

203. Combine the sorted APs into two groups (AP _i , AP _j ) (wherein, 0<i<j≤m, m represents the number of all APs), that is, the APs are divided into Group. The two-dimensional RSSI vector corresponding to the AP combination is extracted from each RSSI data marked with the sampling point as the classification original data. As shown in Table 2 and Table 3.

Table 2 Extraction data format

Calibration point number AP1 AP2 1 -85 -97 1 -83 -92 2 -70 -73 ... ... ...

Table 3 Extraction data format

Calibration point number AP2 AP3 1 -97 -63 1 -92 -65 2 -73 -95 ... ... ...

204. Taking the vector constructed in step 203 as an input, using the distance between vectors as a similarity measurement function, and using the X-means algorithm that can automatically find the number of clusters to perform cluster analysis. The division of each two-dimensional AP combination to the entire positioning area is recorded separately.

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

Step1. Specify the range of clustering number k [k _min ,k _max ], and initialize k=k _min . The range of K is selected according to the size of the actual area to be tested, and the range of each sub-area is from 200m ² to 700m ² , and [k _min ,k _max ] is calculated by this method;

Step2. Randomly select k AP combined data points u ₁ , u ₂ , u ₃ ... u _k from the eigenvector set EV extracted in step 202 as the initial clustering center; the eigenvector set EV is shown in Table 2 and Table 3 As shown, select k eigenvectors as the initial center;

Step3. For each AP combination data point x ⁱ in the eigenvector set EV, determine the cluster it belongs to according to the similarity, Where s(arg ₁ , arg ₂ ) is the similarity calculation function;

Step4. Repeat the above process, and assign all data points to the most similar clusters, so that all AP group data points are initially divided into corresponding clusters;

Step5. For each cluster, recalculate its cluster center, Among them, c ⁽ⁱ⁾ indicates the type that the data point x ⁱ is preliminarily determined to belong to; c ⁽ⁱ⁾ = j means: if the data point x ⁱ belongs to the cluster j, then (c ⁽ⁱ⁾ = j) = 1, Otherwise (c ⁽ⁱ⁾ = j) = 0; the center represents the weighted average center point position of each cluster;

Step6. Calculate the criterion function, Where x ⁱ is the data point in the data set, u _j is the cluster center of cluster j; k refers to the number of cluster centers;

Step7. If the criterion function no longer changes and turns to Step8, it means that the clustering result has been stabilized; otherwise, skip to Step3 and re-cluster;

Step8. Further divide the clusters that have been clustered and calculate the Bayesian Information Criterion BIC _pre and BIC _post before and after the division; Bayesian Information Criterions (BIC) is an important component of Bayesian theory In part, different models on the same data set can be evaluated based on the posterior probability, which is suitable as a reference for selecting a model with lower complexity and better description of the data set.

Among them, for the clustering model corresponding to the number of clusters k, the calculation formula of Bayesian information criterion: Wherein, EV is the set of feature vectors extracted in step 202; R is the number of feature vectors contained in EV, where the number of feature vectors is equal to the number of RSSI combinations at all positions collected by the AP group; p represents the number of parameters, which is called the Schwarz criterion, and its calculation formula is p=k+k d in the present invention, wherein, d is the dimension of the feature vector in the EV, i.e. d=2; Can be seen as a penalty for the complexity of the clustering model; is the maximum a posteriori logarithmic likelihood estimate of the clustering model M _k on the eigenvector set EV, and its calculation formula is as follows

in, u _(i) is the cluster center of cluster i;

Step9. If BIC _pre > BIC _post , check whether the resulting model is higher than the original score. If the score is high, accept splitting and turn to Step10, otherwise set k=k+1 and skip to Step8;

Step10. If k>k _max , you need to re-cluster and turn to Step7; otherwise set k=k+1 and skip to Step2 to calculate the clustering situation of adding a class;

Step11. Select the largest division method of BIC as the clustering result;

Assuming that M is a model set corresponding to different cluster numbers k, then we have is the best clustering model. Each type is expressed as a signal feature, which reflects the aggregation phenomenon of signal distribution in a certain area.

Fig. 3 is another clustering flowchart of the method of the present invention, which is part of the offline sampling stage. Specifically, the following steps may be included:

301. At each calibration point, use a smart phone to scan the surrounding AP signals at high frequency, and the scanned data format is shown in Table 1. It should be noted that the number of data pieces collected by each calibration point is not fixed, and varies with the length of collection time. If the current location fails to collect the RSSI of the corresponding AP, fill it with -100dB.

302. Divide the APs into m groups according to the MAC addresses (m represents the number of all APs). The one-dimensional RSSI vector corresponding to the AP is extracted from each RSSI data marked with the sampling point as the classification original data. As shown in Table 4 and Table 5.

Table 4 extract data format

Calibration point number AP1 1 -85 1 -83 2 -70 ... ...

Table 5 extract data format

Calibration point number AP2 1 -97 1 -92

2 -73 ... ...

303. Take the vector constructed in step 302 as input, and use the distance between the vectors as a similarity measure function to perform cluster analysis. The cluster analysis adopts the X-means algorithm that can automatically find the number of clusters. The clustering method is similar to step 203. , the difference is that the feature vector in step 203 is changed from a two-dimensional vector of AP combination to a one-dimensional vector of a single AP. The signal mode reflects the aggregation phenomenon of signal distribution in a certain area, and records the division of each AP to the entire positioning area. It should be noted that for the same positioning area, different APs will divide the area differently. The reason is that the deployment locations of these APs are far apart in space and are affected by non-line-of-sight transmission effects, multipath propagation effects and The anomalies of the RSSI attenuation laws are different, so it may cause differences in the division results.

Fig. 4 is a training flowchart of the method of the present invention, which is a part of the off-line sampling stage. Specifically, the following steps may be included:

401. Extract two-dimensional RSSI vectors of pairwise combinations of APs at each calibration point, as shown in Table 2 and Table 3, and replace the numbers of the calibration points with corresponding clustered category numbers.

402. Extract the one-dimensional RSSI vector of a single AP at each calibration point, as shown in Table 4 and Table 5, and replace the number of the calibration point with the corresponding clustered category number.

403. Perform SVM training on the vectors obtained in steps 401 and 402 respectively, and calculate classification feature values of the support vector machine. The calculation of classification feature values provides data support for the subsequent judgment of the validity of the initialization range and the narrowing of the positioning range. The classification features selected in this embodiment are the RSSI vectors obtained in steps 401 and 402 .

Fig. 5 is a coarse-grained positioning flow chart of the method of the present invention, which is part of the online phase. Specifically, the following steps may be included:

501. Load the trained SVM classification model of each two-dimensional AP group (AP _i , AP _j ), read the currently collected RSSI, sort in ascending order according to the MAC addresses of the APs, and then extract each group of AP combinations as a classification vector. It should be noted that the SVM algorithm is related to the vector order, so a vector arrangement order must be artificially determined. In this example, use AP's MAC ascending order as the sorting method.

502. Use the corresponding SVM model to predict the classification vector formed by combining the APs extracted in step 501, and calculate the probability of each group of APs in each region under its corresponding region division mode. Because the current location may be at the edge of multiple areas, or because one or some APs in the two-dimensional AP combination are subject to non-line-of-sight transmission effects, multipath propagation effects, and RSSI attenuation abnormalities, etc., RSSI fluctuations may occur. There may be situations where multiple areas meet the requirements, and you can select them in the following ways:

The feature vectors extracted by each AP sampling may have multiple prediction results that meet the requirements after being predicted by the corresponding SVM model, and each prediction result corresponds to a sub-region divided by the AP group into the region to be located. In the above formula, s represents the number of prediction results that meet the conditions, which means that the _SVM model believes that the current device may be in several sub-areas; within the area. Area(AP _i , AP _j ) represents the area where the current location determined by the AP group (AP _i , AP _j ) is located, which means that the SVM model believes that the current device may be in the union of these sub-areas. The selection method of sub-regions that meet the requirements is that if the predicted probability of the current feature vector in a certain sub-region is not less than a certain threshold ε (0<ε<1), the sub-region is considered to meet the requirements. In this example, ε=1/n is selected, and n represents the number of sub-areas divided by the AP combination.

503. For the Areas (AP _i , AP _j ) of all AP groups (AP _i , AP _j ) obtained in step 502, calculate positioning results in a "voting" manner. If the sample data of a certain AP combination is predicted to be in a certain area after step 502, the number of votes in this area is increased by 1. Traverse the Area (AP _i , AP _j ) of all AP combinations and vote, select the area with the most votes as the coarse-grained positioning area for positioning, and the number of votes in each area ranges from 0 to between. Geometrically, the area covered by Area(AP _i , AP _j ) is selected as the coarse-grained positioning area. If the number of votes in all areas is less than a certain threshold ξ, it is considered that the positioning fails, and the positioning ends; if there are multiple areas with the largest number of votes and greater than ξ, the union of these areas is calculated as the coarse-grained positioning area for positioning. Since normal positioning requires at least 3 APs to participate in the calculation, in this example, the value of ξ is 4.

Fig. 6 is a fine-grained positioning flow chart of the method of the present invention, which is part of the online phase. Specifically, the following steps may be included:

601. Load the previously trained SVM classification model of each AP, read the currently collected RSSI, form the RSSI into a one-dimensional classification vector, and use the previously trained SVM classification model of each AP to predict it , respectively calculate the probability of each AP in each area under its corresponding area division mode, and select the probability of each area in the coarse-grained positioning area R obtained in the previous step. Because the current location may be at the edge of multiple areas, or because the AP is subject to non-line-of-sight transmission effects, multipath propagation effects, and RSSI fluctuations due to abnormal RSSI attenuation rules, multiple areas may meet the requirements. Circumstances can be selected as follows:

The feature vectors extracted by each AP sampling may have multiple prediction results that meet the requirements after being predicted by the corresponding SVM model, and each prediction result corresponds to a sub-region divided by the AP into the region to be located. In the above formula, s represents the number of prediction results that meet the conditions, which means that the _SVM model believes that the current device may be in several sub-areas; In the region, the region must be a subset of the coarse-grained positioning region R obtained in step 4. Area(AP _i ) represents the area where the current location determined by AP _i is located, that is, it represents the union of these sub-areas that the SVM model believes that the current device may be in. The selection method of sub-regions that meet the requirements is that if the predicted probability of the current feature vector in a certain sub-region is not less than a certain threshold ε (0<ε<1), the sub-region is considered to meet the requirements. In this example, ε=1/n is selected, and n represents the number of sub-areas divided by the AP.

602. Calculate the Area (AP _i ) of all AP _i in step 601 by "voting", that is, if the sample data of a certain AP is predicted to be in a certain area after step 601, the number of votes for this area is increased by 1. Traverse all AP's Area (AP _i ) and vote, select the area with the most votes as the fine-grained positioning area for positioning, and the number of votes in each area is between 0 and m. Geometrically, it is expressed as selecting the area covered by Area (AP _i ) the most times as a coarse-grained positioning area. If the number of votes in all areas is less than a certain threshold ξ, it is considered that the positioning has failed, and the positioning ends; if there are multiple areas with the largest number of votes and greater than ξ, then find the union of these areas, and find the coordinates and radius of their center points as the final targeting area. Since normal positioning requires at least 3 APs to participate in the calculation, in this example, the value of ξ is 4.

Parts not described in detail in the present invention belong to the well-known technology in the art.

The above are only some specific implementations of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims (4)

1. An indoor positioning method based on WLAN is characterized by comprising the following steps:

the method comprises the following steps: preprocessing RSSI data of each AP acquired by a sampling point, and extracting a one-dimensional vector and a two-dimensional vector from the preprocessed RSSI data as characteristic vectors respectively;

step two: performing clustering analysis on the characteristic vectors, and dividing a region to be positioned into a plurality of positioning sub-regions;

step three: respectively training corresponding classification models for each group of feature vectors in combination with clustering results; selecting a subregion set with the highest vote number from all subregions based on a classification model and a 'voting' mechanism;

step four: the sub-area set range is reduced by two-wheel positioning, and the positioning precision is improved;

the first step of extracting one-dimensional and two-dimensional vectors from the preprocessed data as feature vectors respectively comprises the following steps:

(1) sequencing all the scanned APs in an ascending order according to the MAC addresses;

(2) one-dimensional and two-dimensional vectors are extracted as feature vectors according to the following two methods:

a. combining the sequenced APs pairwise, and dividing the APs into the groups according to the MAC addressesGroups, each group of APs denoted As (AP)_i,AP_j) Wherein, 0<i<j is less than or equal to m, m represents the number of all APs, and a vector formed by combining the APs is extracted from the preprocessed data to serve as a feature vector;

b. each AP is independently used as one group, namely all off-line collected data are divided into m groups according to the MAC address of the AP, and each group of APs is expressed as an AP_iWherein, 0<i is less than or equal to m, m represents the number of all APs, and a vector formed by the APs is extracted from the preprocessed data to be used as a feature vector;

the third step, the concrete implementation process includes an off-line stage and an on-line stage;

in the off-line stage, a Support Vector Machine (SVM) classification model corresponding to each feature Vector of each construction method is trained respectively aiming at the feature vectors constructed by the two construction methods provided in the step one;

in the online stage, a classification feature vector is extracted from real-time data, an SVM classification model trained in the offline stage is read, the probability that the vector to be classified corresponds to different regions is calculated according to support vector polynomial expansion term values, and a region set R with the highest vote number is selected from all the regions by combining a voting mechanism;

the voting mechanism refers to if a group of APs (APs) is present_i,AP_j) Is determined to be in a certain area through SVM prediction, thenAdding 1 to the ticket number of the region; EV (AP) traversing all AP groups_i,AP_j) Voting, selecting the region with the most votes as the coarse-grained location region, wherein the votes of each region should be 0 to 0EV is a feature vector set;

and step four, adopting two rounds of positioning to reduce the range of the region set, and specifically realizing the following steps:

(1) reading a trained SVM classification model, and calculating a support vector polynomial expansion term value;

(2) reading the currently acquired RSSI, extracting a classification feature vector, and standardizing classification features;

(3) mapping the classified characteristic vectors to a high-dimensional space through a polynomial kernel function, calculating the probability of the vectors to be classified corresponding to different regions according to the support vector polynomial expansion term values, and selecting the probability of each region in the coarse-grained positioning region R obtained in the third step;

(4) for each AP_iJudging whether each divided sub-region meets the condition, wherein the sub-region is a subset of the coarse-grained positioning region R obtained in the third step, and if a plurality of sub-regions meet the condition, the SVM model considers that the current equipment is possibly in a union of the sub-regions;

(5) combining a ' voting ' mechanism to select the region set R ' with the highest vote number from the R, the method specifically comprises the following steps: if AP_iThe sample data of the AP is predicted by the SVM and is determined to be in a certain area, the number of votes in the area is added with 1, the area with the largest number of votes is selected as a positioning fine-grained positioning area according to the positioning area votes of each AP, and the number of votes in each area is between 0 and m.

2. The WLAN based indoor positioning method of claim 1, wherein: the first step of preprocessing the RSSI data of each AP collected by the sampling points comprises the following steps: deleting data with too low RSSI, deleting data of non-positioning AP, and filling up RSSI data which is not scanned;

the step of deleting the data with the low RSSI refers to deleting the data with the RSSI intensity lower than a certain threshold value; the data of the non-positioning AP is deleted, namely the RSSI of the AP which is not suitable for positioning is deleted, and the characteristic that the RSSI is not suitable for positioning is too low, namely the RSSI is less than-95 dB or the stability is poor, namely the variance is more than 20.

3. The WLAN based indoor positioning method of claim 1, wherein: in the second step, the feature vector is subjected to cluster analysis, and the region to be positioned is divided into a plurality of positioning sub-regions, and the specific steps are as follows: and (4) taking the characteristic vectors constructed in the step one as input, taking the distance between the characteristic vectors as a similarity measurement function to perform cluster analysis, wherein the cluster analysis adopts an X-means algorithm capable of automatically finding the cluster number.

4. The WLAN based indoor positioning method of claim 1, wherein: the specific operation of the online positioning stage comprises:

(1) reading a trained SVM classification model, and calculating a support vector polynomial expansion term value;

(2) reading the currently acquired RSSI and extracting a classification feature vector;

(3) mapping the classified feature vectors to a high-dimensional space through a polynomial kernel function, and calculating the probability of the vectors to be classified corresponding to different regions according to the support vector polynomial expansion term values;

(4) for each AP group (AP)_i,AP_j) And judging whether each divided sub-region meets the condition, and if a plurality of sub-regions meet the condition, the SVM model considers that the current equipment is possibly in the union of the sub-regions.