CN105657653A - Indoor positioning method based on fingerprint data compression - Google Patents

Abstract

The invention discloses an indoor positioning method based on fingerprint data compression, and belongs to the field of indoor positioning technology. The indoor positioning method based on the fingerprint data compression can be used for removing invalid AP, calculating the amplitude of a receiving signal of all AP at first, calculating a target positioning co-occurrence rate according to the signal amplitude of each AP, and extracting a plurality of AP having the maximal contribution rates to cooperatively work, and the indoor positioning method is used for compressing data to improve the computational efficiency of a positioning algorithm and is used for optimizing the calculation result of the positioning algorithm and improving the precision in combination with a position estimation method.

Description

Indoor positioning method based on fingerprint data compression

Technical Field

The invention belongs to the field of indoor positioning technology, and particularly relates to an indoor positioning technology based on fingerprint data compression.

Technical Field

Along with the development of society and science and technology, position information gradually occupies an important position in daily life and science and technology research and development, and accurate acquisition of target position information also prompts better and faster realization of a plurality of practical functions, and the requirements of people on positioning technology are increasing day by day. The indoor positioning field is also receiving more and more attention, and the application is extremely wide, mainly in families, commercial hospitals and the like. However, in the field of indoor positioning, the traditional GPS positioning has unsatisfactory use effect in indoor positioning due to its own limitations, so that various indoor positioning methods are proposed, and among them, the WIFI-based fingerprint indoor positioning technology has good effect.

The location fingerprint identification technology is a more mainstream indoor positioning method. The method mainly comprises the steps of carrying out data acquisition and data description on a positioning space, describing the position information of a positioning environment by using an RSSI (received signal strength) sequence of each AP (wireless access) access point, and gathering the information of different positions into a fingerprint database. And finally, matching the RSSI data measured by the user mobile terminal with the data in the fingerprint database, and selecting the position with the best similarity as an estimated position. The whole process is divided into an off-line training phase and an on-line positioning phase, as shown in fig. 1. In the off-line positioning stage, a position fingerprint database is required to be established, positioning system deployment personnel traverse all positions in a positioning environment, RSSI values from different AP access points are collected at each reference position, and MAC addresses, RSSI values and position information of all AP points form a related ternary array to be stored in the database. In the on-line positioning stage, when a positioning user is in a positioning area, RSSI values of all AP access points are collected in real time, and the MAC addresses and the RSSI values form a binary array which is used as input data of a position matching algorithm, and the position is estimated by using a specific matching algorithm.

However, the indoor positioning technology also meets the development bottleneck at present, and two major problems mainly hindering the progress of the indoor positioning technology are: namely how to improve the algorithm precision and the algorithm efficiency of indoor positioning. In view of the two technical difficulties, a compression method and a filtering model aiming at fingerprint data are provided to improve the efficiency and the precision of an indoor positioning algorithm.

In general, the present invention provides an indoor positioning technique for fingerprint data compression. The WIFI indoor positioning technology based on the position fingerprints is used in the algorithm level, and the traditional method has certain limitations in the operation efficiency and the algorithm precision. After the technology is used, tests show that the method can play a role in compressing data after the main feature vector is extracted, so that the arithmetic efficiency of the algorithm can be greatly improved. And then, a filtering model which is modeled according to the motion condition of the positioning object is added to optimize the positioning result, so that the algorithm precision can be greatly improved under the condition that the modeling of the motion state of the moving object is more accurate.

Disclosure of Invention

The invention aims to further improve the operation efficiency and the operation precision of the traditional indoor positioning algorithm, provides an indoor positioning method aiming at fingerprint data compression, can extract main AP points for matching work by removing invalid AP points, plays a role of compressing data to improve the operation efficiency of the positioning algorithm, and optimizes the result calculated by the positioning algorithm by utilizing filtering to improve the precision.

In the traditional fingerprint positioning method, RSSI values from a plurality of AP points are collected on each fingerprint point, but the contributions to the positioning work of a plurality of AP points are not completely the same, the RSSI values of some AP points have a decisive effect on the positioning work, and the RSSI values of some AP points have little influence on the positioning work. The RSSI data of a plurality of AP points can greatly increase the dimensionality of the fingerprint database matrix, thereby greatly increasing the calculation burden of the fingerprint positioning algorithm. Therefore, a fingerprint data compression technology is provided, and information is processed, compressed and feature extracted based on a variable covariance matrix. According to the method, the data feature extraction and compression are realized by taking the direction of the sample point with the maximum change in the space, namely the direction with the maximum variance, as a discrimination vector according to the position distribution of the sample point in a multi-dimensional mode space. From the viewpoint of probability statistics, the larger the variance of a random variable is, the larger the amount of information contained in the random variable is, and if the variance of a random variable is 0, the amount of information contained in a constant is also 0. The main AP component is a variable obtained by linearly combining (or mapping) m variables of the original data, and the variation makes the variance of the changed variable be the largest (the first principal component), and the principal components are linearly independent of each other, and from the first principal component, the principal components are sorted according to the variance (corresponding eigenvalues are arranged according to the order of magnitude). For a characteristic value of λ_iMain component of (A) (-)_iThe variance of the principal component, and the contribution rate of the principal component is η_i，η_i＝λ_i/(λ₁+λ₂+...+λ_p). In modeling, λ_iThe effect of the smaller principal component and corresponding AP point is considered noise and is not introduced into the model. Therefore, the main component AP points in the fingerprint database are reduced, so that the purpose of reducing the dimension is achieved, the calculation amount is reduced, and the calculation efficiency is increased.

The filtering model is mainly established by inputting the previous predicted coordinates into a model modeled according to the motion state of the target, fitting the predicted coordinates and the motion state, and finally making an optimal estimation on the coordinate information. The method is a continuous prediction and correction process, the algorithm does not need to store a large amount of predicted coordinate data, and when new predicted coordinate data are obtained, a new parameter filtering value can be calculated at any time, so that the observation result can be processed in real time conveniently, and therefore the filtering model can be applied to dynamic indoor positioning processing. In indoor positioning application, the filter model can be used for reducing errors of position prediction and improving prediction accuracy of a motion trail of a moving object. The method inputs real-time predicted coordinate data through a direct recursion method, and outputs an optimal coordinate estimation value with minimum mean square error under a linear system. The key of the successful application of the algorithm lies in whether the modeling of the system dynamic model and the observation model is accurate, which requires that the motion condition of the object is known clearly, and if the modeling is accurate, the precision of the positioning algorithm can be effectively improved to a certain extent.

To verify the superiority of this indoor positioning technique for fingerprint data compression, we performed two experiments. The experimental site map and the actual sampling path are shown in fig. 2.

The positioning algorithm adopted in the first experiment is a minimum neighbor method which is unified into a plurality of methods used in the current positioning algorithm, the evaluation standards mainly refer to average error, 90% positioning error, maximum positioning error and calculation time, and the superiority of the algorithm is verified by comparing the results of the indoor positioning algorithm aiming at fingerprint data compression. The results of the experiments are shown in Table 1.

In the results it can be seen that the average error distance is somewhat improved with the positioning algorithm for fingerprint data compression and most notably that the computation time is reduced by about 2/3. This fully demonstrates that the accuracy and efficiency of the algorithm can be improved using this technique. For the reason, because the RSSI signal is inevitably affected by multipath effect, shadowing effect, and other factors during the propagation process, the RSSI value of an AP received at the same position may fluctuate to different degrees over time, and the average values of the RSSI signals between several adjacent fingerprint points are likely to be close under the statistical characteristics of the RSSI signal. At the moment, the main positioning characteristics with stability are extracted by applying an indoor positioning technology analysis method aiming at fingerprint data compression, so that the possible situations are solved, the data dimensionality is also compressed, the complexity is reduced for the following algorithm operation, and the positioning precision and efficiency can be improved.

Therefore, the invention relates to an indoor positioning method based on fingerprint data compression, which comprises the following steps:

step 1: acquiring training samples, recording the position information of the M samples, and acquiring RSSI values of all APs acquired by corresponding samples;

step 2: calculating the average RSSI intensity of the M samples;

and step 3: calculating the difference value between each sample and the average RSSI intensity;

and 4, step 4: constructing a covariance matrix according to the difference value obtained in the step 3;

and 5: calculating eigenvalues and corresponding eigenvectors of the covariance matrix to obtain N eigenvalues and corresponding eigenvectors, wherein N represents the total number of the APs;

step 6: solving the maximum P characteristic values and the corresponding characteristic vectors to ensure that the P characteristic values can meet the required total contribution rate threshold, wherein the total contribution rate meets the following formula:

$\mathrm{\φ}=\frac{\underset{i=1}{\overset{p}{\Σ}}{\mathrm{\λ}}_i}{\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\λ}}_i}\≥a$

wherein,denotes the total contribution, λ_iRepresenting a characteristic value, a representing a total contribution rate threshold value set in advance according to requirements; establishing a feature space w for the obtained P feature values;

and 7: projecting the covariance matrix obtained in the step 4 into the feature space w obtained in the step 6 to obtain compressed data of M p-dimensional matching vectors;

and 8: in the positioning stage, the RSSI value of each AP received by the target is collected, the RSSI average intensity calculated in the step (2) is subtracted by the RSSI average intensity, and then the RSSI average intensity is projected to a w in a characteristic space to obtain a target vector;

and step 9: and (4) matching the target vector obtained in the step (8) with the compressed data obtained in the step (7) to obtain the best matched vector, wherein the position corresponding to the vector is the target position.

Further, after the target position is positioned, the current time position is further accurate according to the position information of the target at the previous time.

Further, the covariance matrix method constructed in step 4 is:

$C=\frac{1}{M}\underset{i=1}{\overset{M}{\Σ}}{d}_i{d}_i^T=\frac{1}{M}{\mathrm{AA}}^T,A=(d_1,d_2,...,d_M)$

wherein: c is a covariance matrix, d_iAnd (4) representing the difference value between the ith sample obtained in the step (3) and the average RSSI strength, wherein M represents the total number of samples.

Further, in step 5, a singular value decomposition method is used to obtain N eigenvalues and corresponding eigenvectors.

Further, the matching method in step 9 is to calculate a 2-norm of the target vector and each matching vector in the compressed data, wherein a position corresponding to the matching vector with the minimum 2-norm is a target position.

The positioning algorithms adopted in the second experiment are unified into a positioning algorithm subjected to data compression processing, and the evaluation criteria mainly refer to average error, 90% positioning error and maximum positioning error. The method is mainly used for verifying the optimization effect of the filtering model on the positioning algorithm. The experimental results are shown in fig. 3, 6, 7, 8 and table 2.

As is apparent from the results, comparing fig. 4, fig. 5, fig. 7 and fig. 8, it can be seen that the difference between the nearest neighbor algorithm only and the nearest neighbor algorithm after data compression is not great, and although some points can be well predicted, too many predicted points are duplicated and the operation trajectory cannot be well predicted. As can be seen from fig. 6, 7 and 8, after filtering is used, the number of the repetition points is small, and the sampling trajectory can be well tracked. Observing fig. 3, it can be seen that the positioning effect is improved obviously after the filtering is used. In the previous work, a positioning algorithm after data compression is used, the algorithm efficiency is greatly improved, the algorithm precision is improved only by a small degree, and the effect is not obvious. After the filter model is used on the positioning algorithm, as shown in table 2, the accuracy of the positioning algorithm is significantly improved. The average error distance is only close to 1m, and the precision of 90 percent of positioning error is within 1.9 m. The maximum positioning error is less than 3.4 m.

The overall process steps are shown in fig. 9.

Drawings

FIG. 1 is a schematic diagram of a location fingerprint positioning method

Fig. 2 is an experimental site map and an actual sampling path.

Fig. 3 is a CDF graph showing comparison of accuracy of algorithms using a nearest neighbor algorithm only, a data compression positioning algorithm, and a filter model.

Fig. 4 is a comparison of an actual sampling trajectory with a trajectory of a positioning result using only the minimum neighbor method.

FIG. 5 is a comparison of an actual sampling trace with a trace of a positioning result after processing using a fingerprint data compression positioning algorithm.

FIG. 6 is a comparison of an actual sampling trajectory to the trajectory of an algorithm optimized using a filter model.

FIG. 7 is a simulation of X-axis coordinates under several methods and comparison with the X-axis coordinates of an actual sampling trajectory.

FIG. 8 is a simulation of Y-axis coordinates under several methods and comparison with the actual sampling trajectory Y-axis coordinates.

Fig. 9 is a flow chart of an indoor positioning technique based on a fingerprint data compression positioning algorithm.

Detailed description of the invention

Step A: setting training samples (position fingerprints), collecting RSSI (received signal strength indicator) values of all APs by each position fingerprint training sample P, and finally combining the position information L of the APs_p＝(X_p,Y_p) Obtain the sample set T { (S)₁,L₁),(S₂,L₂),...,(S_p,L_p)},WhereinWhich represents the mean value of the RSSI data collected N times received from the d AP at fingerprint point P. Let X be [ S ]₁,S₂,...,S_p]For convenience of explanation, if there are 81 APs and 200 fingerprint points, there is X ═ X (X)₁,X₂,...,X₂₀₀)^TWherein X is_i＝(S₁,S₂,...,S₈₁)。

And B: calculating the average RSSI strength:

and C: calculating a difference d_i，d_i＝X_i-ψ,i＝1,2,...,200，

Step D: construction of covariance matrix (and C in the above-mentioned theory)_xSince the intensity mean value is not 0 in the position fingerprint model, the formula will change

$C=\frac{1}{200}\underset{i=1}{\overset{200}{\Σ}}{d}_i{d}_i^T=\frac{1}{200}{\mathrm{AA}}^T,A=(d_1,d_2,...,d_{200})$

Step E: obtaining eigenvalue and eigenvector u 'of covariance matrix'_dConstructing a full feature space w '═ u'₁,u'₂,...,u'₈₁]. Because of the large amount of computation, it is generally considered to process this matrix using singular value decomposition to find the eigenvalue eigenvectors.

Step F: and according to the requirement on the contribution rate, the first P maximum eigenvalues and the corresponding eigenvectors are obtained. Namely, the function of reducing the dimension is realized, and some required main AP points are selected for the next matching. Wherein, the contribution ratio refers to the sum ratio of the selected characteristic values to the sum of all the characteristic values, namely:wherein λ_iIn the author experiment, a is 96%, that is, 96% of energy is projected on the first p eigenvector sets by the training sample, and the first p is 6 largest eigenvalues, that is, RSSI data of the 6 APs with the largest information amount is taken. Setting the eigenvectors of the original covariance matrix A as u respectively according to the arrangement of the corresponding eigenvalues from large to small₁,u₂,...,u_d＝81The feature space w ═ u can be derived₁,u₂,...,u₆]。

Step G: projecting the covariance matrix obtained in step D to the featuresIn the feature space w, the light source is arranged,in this case, Z includes vectors equal to the number of AP points, i.e., 81. After projection processing, the correlation among all vectors in Z is basically eliminated, the first P principal components occupying most information of X are contained in Z, and data compression is completed.

Step H: and (5) carrying out an online positioning stage, and collecting the RSSI value of each AP at any point in the experimental area. And combining into a sampling vector, subtracting the previously calculated RSSI average intensity by the sampling vector, and projecting the sampling vector into the feature space. This results in a 6 x 1 vector.

Step I: and comparing the vector with the fingerprint array projected to the feature space, namely solving the vector 2 norm of each row of the vector and the fingerprint array projected to the feature space, which correspond to different positions, and finding out the most matched point, namely the coordinate with the minimum vector 2 norm, as the coordinate of the sub-online sampling point.

Step J: and modeling the moving object by using the principle of filter model establishment. The state equation is established as follows:

x(n)＝F(n,n-1)x(n-1)+(n,n-1)v₁(n-1)

namely:

${\left[\begin{array}{cccc}X& Y& V_x& V_y\end{array}\right]}^T=\left[\begin{array}{cccc}1& 0& T& 0\\ 0& 1& 0& T\\ 0& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right]\×\left[\begin{array}{c}X\\ Y\\ V_x\\ V_y\end{array}\right]+\left[\begin{array}{cc}{T}^2/2& 0\\ 0& T^2/2\\ T& 0\\ 0& T\end{array}\right]\×{\left[\begin{array}{c}{W}^x\\ W^y\end{array}\right]}_k$

establishing an observation equation as follows:

$z\left(n\right)=C\left(n\right)x\left(n\right)+v_2\left(n\right)\⇒\left[\begin{array}{c}{Z}_x\\ Z_y\end{array}\right]={\left[\begin{array}{cc}1& 0\\ 0& 1\\ 0& 0\\ 0& 0\end{array}\right]}^T\×\left[\begin{array}{c}X\\ Y\\ V_x\\ V_y\end{array}\right]+{\left[\begin{array}{c}{v}^x\\ v^y\end{array}\right]}_k$

wherein T is the sampling time, X and Y are the sampling real-time coordinates,velocity V_xAnd V_yI.e. the difference in distance of adjacent points along the X-axis and Y-axis, respectively. W^x,W^y,v^x,v^yIdeally, should be zero mean white gaussian noise.

The state covariance matrix Q is:

$Q=E\{G.w_k.{(G.w_k)}^T\}=\left[\begin{array}{cccc}{a}_x^2.T^4/4& 0& a_x^2.T^3/2& 0\\ 0& a_y^2.T^4/4& 0& a_y^2.T^3/2\\ a_x^2.T^3/2& 0& a_x^2.T^2& 0\\ 0& a_y^2.T^3/2& 0& a_y^2.T^2\end{array}\right]$

the observed covariance matrix R is:

$R=\left[\begin{array}{cc}{\mathrm{\σ}}_x^2& 0\\ 0& {\mathrm{\σ}}_y^2\end{array}\right]$

wherein the parameters T, sigma_x、σ_y、a_x、a_xThe method is set according to the surrounding environment, and the concrete modeling steps are as follows:

setting initial conditions: since the initial state of the process equation is not accurately known at the initial time, it is often described using a mean and correlation matrix to represent the estimated unbiased behavior.

P(0)＝E{[x(0)-E[x(0)]][x(0)-E[x(0)]]^H}

The state is predicted in one step, namely: $\hat{x}\left(n\right|{\mathrm{\χ}}_{n-1})=F(n,n-1)\hat{x}(n-1|{\mathrm{\χ}}_{n-1})\∈C^{N*1}$

calculating an innovation process from the observed signal z (n):

$a\left(n\right)=z\left(n\right)-\hat{z}\left(n\right|{\mathrm{\χ}}_{n-1})=z\left(n\right)-C\left(n\right)\hat{x}\left(n\right|{\mathrm{\χ}}_{n-1})\∈C^{M*1}$

one-step prediction error autocorrelation matrix:

P(n,n-1)＝F(n,n-1)P(n-1)F^H(n,n-1)+(n,n-1)Q^H(n,n-1)∈C^N*N

innovation process autocorrelation matrix:

A(n)＝C(n)P(n,n-1)C^H(n)+R∈C^M*M

k gain:

K(n)＝P(n,n-1)C^H(n)A^-1(n)∈C^N*M

and (3) state estimation:

$\hat{x}\left(n\right|{\mathrm{\χ}}_n)=\hat{x}\left(n\right|{\mathrm{\χ}}_{n-1})+K\left(n\right)a\left(n\right)\∈C^{N*1}$

state estimation error autocorrelation matrix:

P(n)＝[I-K(n)C(n)]P(n,n-1)∈C^N*N

the position and real-time motion speed estimated by the matching algorithm in each online positioning stage are used as input, namely an observed value z (n). Repeating the above steps to perform recursive filtering calculation

Step K: the model built by using the filtering principle outputs an optimized coordinate value when a result calculated by using a fingerprint data compression positioning algorithm before input.

TABLE 1

TABLE 2

Claims (5)

1. An indoor positioning method based on fingerprint data compression, the method comprises:

step 1: acquiring training samples, recording the position information of the M samples, and acquiring RSSI values of all APs acquired by corresponding samples;

step 2: calculating the average RSSI intensity of the M samples;

and step 3: calculating the difference value between each sample and the average RSSI intensity;

and 4, step 4: constructing a covariance matrix according to the difference value obtained in the step 3;

and 5: calculating eigenvalues and corresponding eigenvectors of the covariance matrix to obtain N eigenvalues and corresponding eigenvectors, wherein N represents the total number of the APs;

step 6: solving the maximum P characteristic values and the corresponding characteristic vectors to ensure that the P characteristic values can meet the required total contribution rate threshold, wherein the total contribution rate meets the following formula:

$\mathrm{\φ}=\frac{\underset{i=1}{\overset{p}{\Σ}}{\mathrm{\λ}}_i}{\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\λ}}_i}\≥a$

wherein,denotes the total contribution, λ_iRepresenting a characteristic value, a representing a total contribution rate threshold value set in advance according to requirements; establishing a feature space w for the obtained P feature values;

and 7: projecting the covariance matrix obtained in the step 4 into the feature space w obtained in the step 6 to obtain compressed data of M p-dimensional matching vectors;

and 8: in the positioning stage, the RSSI value of each AP received by the target is collected, the RSSI average intensity calculated in the step (2) is subtracted by the RSSI average intensity, and then the RSSI average intensity is projected to a w in a characteristic space to obtain a target vector;

and step 9: and (4) matching the target vector obtained in the step (8) with the compressed data obtained in the step (7) to obtain the best matched vector, wherein the position corresponding to the vector is the target position.

2. The indoor positioning method based on fingerprint data compression as claimed in claim 1, wherein after the target position is located, the current time position is further refined according to the position information of the target at the previous time.

3. The indoor positioning method based on fingerprint data compression as claimed in claim 1, characterized in that the covariance matrix method constructed in step 4 is:

$C=\frac{1}{M}\underset{i=1}{\overset{M}{\Σ}}{d}_i{d}_i^T=\frac{1}{M}{\mathrm{AA}}^T,A=(d_1,d_2,...,d_M)$

wherein: c is a covariance matrix, d_iAnd (4) representing the difference value between the ith sample obtained in the step (3) and the average RSSI strength, wherein M represents the total number of samples.

4. The indoor positioning method based on fingerprint data compression as claimed in claim 1, characterized in that in step 5, a singular value decomposition method is used to obtain N eigenvalues and their corresponding eigenvectors.

5. The indoor positioning method based on fingerprint data compression as claimed in claim 1, characterized in that the matching method of step 9 is to calculate 2 norms of the target vector and each matching vector in the compressed data, wherein the position corresponding to the matching vector with the minimum 2 norms is the target position.