CN111757258A - An adaptive positioning fingerprint database construction method in complex indoor signal environment - Google Patents

Abstract

A self-adaptive positioning fingerprint database construction method under a complex indoor signal environment belongs to the technical field of indoor positioning. According to the AP laying position and the space structure, the system adopts a one-to-many support vector machine algorithm to perform partition operation on a target area so as to accurately determine the area range of signal change. A multivariate Gaussian mixture model based on mutual interference between signals is established by using the coupling relation between the signals in the narrow and small subareas so as to improve the reduction of the positioning precision caused by signal fluctuation. When the indoor environment changes, the self-adaptive updating algorithm based on the partition multivariate Gaussian mixture model can judge the credibility of fingerprint data of each partition, and updates the model parameters of the partitions with larger signal fluctuation by the self-adaptive algorithm, so that the coupling degree between the model and the existing environment is improved. The experimental result shows that the method can utilize relatively small amount of sample data to construct a stable and maintainable indoor signal distribution model, and compared with other algorithms, the positioning accuracy is improved to a certain extent.

Description

An adaptive positioning fingerprint database construction method in complex indoor signal environment

technical field

The invention relates to an offline fingerprint database construction and optimization technology for indoor positioning, and belongs to the technical field of indoor positioning.

Background technique

Global satellite navigation systems have been widely used to provide people with location services in outdoor environments, but the lack of signals also makes the system unable to function in complex indoor environments. The widespread deployment of WiFi facilities and the popularization of smartphones have made indoor positioning systems based on Received Signal Strength (RSS) values closely watched by a large number of researchers. However, since the original design of wireless communication facilities is not to provide people with indoor navigation, how to reduce the impact of environmental fluctuations on the positioning caused by the uncertain interference of wireless signals has become a difficulty that existing research has to face and solve.

The positioning system based on the existing wireless facilities, its common commercial devices do not have the function of self-editing, and can only provide the indoor general RSS measurement value, which makes the traditional method based on the difference of arrival time and the difference of arrival distance, unable to directly translate applications. The fingerprint positioning algorithm using the matching operation between the measured signal and the actual position makes up for the lack of time and space characteristics of the signal sent by the wireless facility, which can effectively realize the mapping between the signal environment and the actual scene, and provide an indoor positioning based on the RSS value. possible.

SUMMARY OF THE INVENTION

For large indoor scenes, the above method will consume a lot of manpower and material resources, and is easily affected by environmental factors. The mapping relationship between the constructed offline fingerprint database and the signal distribution in the actual scene will also weaken due to time changes, and the offline fingerprint database needs to be continuously revised. In order to reduce the accumulation of mapping errors caused by time accumulation.

Technical scheme of the present invention:

A method for constructing an adaptive positioning fingerprint database in a complex indoor signal environment, including three stages: an offline stage, an online positioning stage, and an online fingerprint database update stage. The specific steps are as follows:

Step 1. Divide the target area grid and build an offline fingerprint library

其中n∈{1,2,...,N_k}，

为2×N_k维位置矩阵表示RP位置。对于各RP位置

相应的区域标志为

其中

且

i≠k，表示参考点位于k分区。在

收集到的来自M_k个AP的RSS样本值为

其中

表示在

处收集到的来自第m个AP的RSS样本值，m∈{1,2,...,M_k}。The target area is composed of K divided areas Ω _k , k∈{1,2,...,K}, the area Ω _k is divided into N _k grids, and the geometric center of each grid is taken as the reference point

where _n∈ {1,2,...,Nk},

represents the RP position as a 2×N _k -dimensional position matrix. For each RP location

The corresponding region flags are

in

and

i≠k, indicating that the reference point is located in the k partition. exist

The collected RSS sample values from M _k APs are

in

expressed in

RSS sample values from the mth AP collected at m∈{1,2,..., _Mk }.

Step 2. Build the SVM partition model in the online positioning stage

The support vector machine probability classification model is set up in a one-to-many manner. For each pre-set partition, the two-class identification is performed according to whether the target is located in this partition, and the SVM model of each partition is constructed through the training data. For a given K partitions, K SVM models are established, and the measurement signals of all reference APs in each partition are taken to form the current observation data r=[r ₁ , r ₂ ,...,r _M ], where M is received by the target The number of APs in the area, and the value of the unreceived signal is -100db. For the current observation data, each partition SVM model gives the distribution probability p(y ^k =1|r) of whether the target is located in the corresponding area, where y ^k is the partition identifier, indicating that the target is located in the partition Ω _k , k∈{1, 2,...,K}. According to the distribution probability p (y ^k =1|r) given by the SVM model of each partition, a preliminary judgment is made on the partition where the target is located, and it is used as the first-level judgment basis.

Using the partition operation based on probability SVM, the reference point observation data obtained in step 1 is divided into training set and test set, and the partition judgment model SVM is trained. The probability value that the target is located in the corresponding partition is obtained through K SVM models, and the second-level judgment basis is set to overcome the problem of misjudgment of the test points near the partition handover. Select two partitioned regions with the highest probability in the determined region, namely p(y ⁱ =1|r) and p(y ^j =1|r), i,j∈{1,2,...,K}, And p(y ⁱ =1|r)>p(y ^j =1|r), the difference is expressed as:

Δy _p =p(y ⁱ =1|r)-p(y ^j =1|r), (1)

When Δy _p >Δy, it means that the influence of the i partition on the test point is much greater than that of the j partition, and the reference point is determined in the i partition, where Δy is the secondary judgment threshold. For Δy _p < Δy, both regions are determined as target regions, and the corresponding region matching operations are performed respectively, and the probability average of the target positions obtained by the respective partitions is performed to obtain the final position estimate.

Step 3. Build a multivariate Gaussian mixture model MVGMM

The multivariate Gaussian mixture model MVGMM is established by using the correlation of different AP signals. By continuously increasing the number of Gaussian elements, the probability density function of different parameters is used to weight and approximate the joint distribution between the RP positions in the partition and the acquired AP signals to make up for The coupling relationship between AP signals is ignored in the construction of traditional fingerprint database. The probability distribution function of the multivariate Gaussian mixture model is expressed as:

Among them, C represents the number of constituent elements, p _N (x|μ _c , P _c ) represents a Gaussian composition with a mean of μ _c and a covariance of P _c , and the sum of the weights w _c is 1.

Based on equation (2), the _multivariate Gaussian mixture model is expressed as:

为多元高斯混合模型的组成元素权重，

为元素均值，

为元素协方差，C_k为组成元素数量。Among them, y ^k =1 indicates that the target is in the kth partition, r indicates the RSS value of each AP signal received at the reference point x,

is the element weight of the multivariate Gaussian mixture model,

is the element mean,

is the element covariance, and C _k is the number of constituent elements.

协方差为

的多元高斯函数，拟合分区Ω_k内N_k个参考点

与相应分区内所获取各AP信号RSS值

的联合概率分布。利用k-means算法将样本数据聚类为C_k个初始簇，取各簇平均值与协方差用于初始化EM算法参数且各簇初始权重设置相同。联合概率分布的对数似然形式表示为：The MVGMM model uses the EM algorithm to estimate the model parameters. The MVGMM model selects C _k means as

The covariance is

The multivariate Gaussian function is fitted to N _k reference points in the partition Ω _k

and the RSS value of each AP signal obtained in the corresponding partition

The joint probability distribution of . The sample data is clustered into C _k initial clusters by the k-means algorithm, and the mean and covariance of each cluster are taken to initialize the parameters of the EM algorithm, and the initial weights of each cluster are set the same. The log-likelihood form of the joint probability distribution is expressed as:

隐变量γ_c,n表示第n个样本值属于第c个高斯组成元素的概率。where z ^k =[x ^k ; R ^k ],

The latent variable γc _,n represents the probability that the nth sample value belongs to the cth Gaussian component.

属于第c个高斯组成元素的概率表示为：From this, the observed data are determined

The probability of belonging to the c-th Gaussian component is expressed as:

均值

和协方差

分别为：Calculate the weight of the multivariate Gaussian function by formula (5)

mean

and covariance

They are:

Repeating the clustering and EM estimation process for different values of C _k The offline fingerprint library is denoted as (w ^k ; μ ^k ; P ^k ), where k∈{1,2,..,K}.

Step 4. Locate the target in the sub-domain

通过分区判断模型获得目标在各分区的分布概率p(y_k＝1|r^now)，k∈{1,2,...,K}。The target location in the area is estimated through the offline fingerprint database. If the RSS value of M APs accepted by the target at the current moment is

The distribution probability p(y _k =1|r ^now ) of the target in each partition is obtained through the partition judgment model, k∈{1,2,...,K}.

根据多元高斯分布的条件概率准则，获得给定观测数据下目标位置的后验概率分布：Use the current measurement value r ^now to select the AP signal measurement value that is valid for target positioning in the corresponding partition Ω _k

According to the conditional probability criterion of multivariate Gaussian distribution, the posterior probability distribution of the target position under the given observation data is obtained:

Based on the obtained offline fingerprint database (w ^k ; μ ^k ; P ^k ), we have:

Combining the partition probability with equation (9), the posterior distribution probability of each reference point of the target in the partition Ω _k is obtained as:

The distribution weight of each reference point in the partition Ω _k is updated as:

Among them, λ is the normalization factor, then the distribution weight of each reference point in the partition Ω _k as the target position is:

In summary, the location of the target is estimated as

Step 5. Update the fingerprint database

对分区模型进行参数更新。基于现有多元高斯混合模型，通过采集数据对模型参数进行自适应调整，以推导出与现有环境具有更加紧密耦合关系的MVGMM模型。与EM算法一致，MVGMM模型参数的自适应过程也是一个两步估计。首先，通过k-means算法将新增观测数据

聚类为C_k组簇，并通过聚类所得数据权值

均值

协方差

初始化各高斯组成成分。利用式(5)-(8)分别计算新增观测数据的统计权重

均值

与协方差

When the information entropy obtained in the partition Ω _k is less than the partition threshold, N _k data are newly collected by

Parameter update for the partition model. Based on the existing multivariate Gaussian mixture model, the model parameters are adaptively adjusted by collecting data to derive a MVGMM model with a more tightly coupled relationship with the existing environment. Consistent with the EM algorithm, the adaptation process of the MVGMM model parameters is also a two-step estimation. First, the new observation data will be added through the k-means algorithm.

Clustering into C _k groups of clusters, and data weights obtained by clustering

mean

Covariance

Initialize each Gaussian component. Use equations (5)-(8) to calculate the statistical weights of the newly added observation data respectively

mean

with covariance

The MVGMM model parameters corresponding to the partition Ω _k are adaptively optimized using the statistical parameters of the newly added data, namely:

ρ∈{w,m,v}用于平衡新旧数据对模型参数的影响程度，λ为归一化因子，r^ρ为参数的固有相关因子。in,

ρ∈{w,m,v} is used to balance the influence of old and new data on model parameters, λ is the normalization factor, and r ^ρ is the inherent correlation factor of the parameters.

Step 6. Update model parameters

The model parameters updated in the adaptive process belong to the second and third steps, and are updated in real time.

Beneficial effects of the present invention: According to the AP laying position and spatial structure, the system adopts a one-to-many support vector machine algorithm to perform a partition operation on the target area, so as to accurately determine the area range of the signal change. Using the coupling relationship between signals in a narrow partition, a multivariate Gaussian mixture model based on mutual interference between signals is established to improve the positioning accuracy drop caused by signal fluctuations. When the indoor environment changes, the adaptive update algorithm based on the partitioned multivariate Gaussian mixture model can judge the reliability of the fingerprint data of each partition, and use the adaptive algorithm to update the model parameters of the partitions with large signal fluctuations, so as to improve the model and the accuracy of the model. The degree of coupling between existing environments.

Description of drawings

Figure 1 is a flow chart of the method of the present invention.

Figure 2 shows the experimental scene diagram.

Figure 3 is a comparison diagram of the effect of partitions, in which (a) is the situation of regional division, and (b) is the effect diagram of regional division after practice.

Figure 4 is a comparison diagram of the construction effect of RSS fingerprint map, in which (a) is the real measurement value, (b) is the GP-type estimation, and (c) is the MVGMM model estimation.

FIG. 5 is a comparison diagram of the fitting effect of AP3 signals in partition one.

FIG. 6 is a comparison diagram of trajectory prediction of target motion.

Figure 7 is a boxplot of trajectory estimation error.

Figure 8 is a comparison diagram of the error accumulation function.

Figure 9. Comparison of AP3 data fitting effects before and after the fingerprint database update. Among them, (a) is the AP3 signal strength attenuation value, (b) is the AP3 signal prediction error.

Detailed ways

In view of the large amount of sampling data and high maintenance cost in large indoor scenes, the method accurately maintains the area through partition operation, and proposes a partitioned Multivariate Gaussian Mixture Model (MVGMM) according to the coupling relationship between signals in the partition to improve the The fit of the signal distribution. The model divides the target area according to the location of the signal access point (AP) and the physical connection structure, and realizes the partition operation through the one-to-many support vector machine model. In a relatively narrow partition area, the multivariate Gaussian mixture model is established by using the mutual interference between the signals to strengthen the fitting degree of the signals, and finally achieve the effect of improving the partition positioning accuracy. When the environment changes, the algorithm uses the information entropy as the partition data update criterion, and responds to the impact of the partition change on the fingerprint database in time to reduce the maintenance cost. Therefore, in indoor positioning applications, a small amount of data can support the construction of an efficient and maintainable fingerprint database.

The target localization system based on the multivariate Gaussian mixture model can accurately locate the target with low acquisition cost. The algorithm uses the measurement signals obtained in traditional WiFi facilities as training data to construct the distribution of AP signals in the target area. Compared with the traditional target positioning process, the system consists of three modules, including an offline stage, an online positioning stage, and an online fingerprint database update stage, as shown in Figure 1.

The offline stage mainly establishes a complete fingerprint database by locating the position of each reference point (Reference Point, RP) in the area and the received RSS signal value, and applies it to the matching operation in the online stage. In the online positioning stage, the partition model is used to determine the target area, and the conditional probability criterion of the multivariate Gaussian function is used to calculate the posterior probability of the target position conditioned on the current observation data, and estimate the target position through WKNN. In the online fingerprint database update stage, the information entropy calculated by the posterior distribution probability is used as the measure of the fingerprint database update, and a feedback scheduling fingerprint database based on the partitioned multivariate Gaussian mixture model is established.

details as follows:

Build an offline fingerprint library

Fingerprint Collection Scheme

其中n∈{1,2,...,N_k}，

为2×N_k维位置矩阵表示RP位置。对于各RP位置

相应的区域标志为

其中

且

i≠k，表示参考点位于k分区。在

收集到的来自M_k个AP的RSS样本值为

其中

表示在

处收集到的来自第m个AP的RSS样本值，m∈{1,2,...,M_k}。The target area consists of K partitioned areas Ω _k , k∈{1,2,...,K}, and the area Ω _k is divided into N _k meshes, and the geometric center of each mesh is taken as the reference point

where _n∈ {1,2,...,Nk},

represents the RP position as a 2×N _k -dimensional position matrix. For each RP location

The corresponding region flags are

in

and

i≠k, indicating that the reference point is located in the k partition. exist

The collected RSS sample values from M _k APs are

in

expressed in

RSS sample values from the mth AP collected at m∈{1,2,..., _Mk }.

Partition Model Construction

Considering the accuracy and efficiency of regional classification, setting the support vector machine probability classification model in a one-to-many manner can effectively solve the classification problem of partitions set according to AP location and physical connectivity [14]. For each pre-set partition, the two-class identification is carried out according to whether the target is located in this partition, and the SVM model of each partition is constructed by training data. For a given K partitions, K SVM models are established, and the measurement signals of all reference APs in each partition are taken to form the current observation data r=[r ₁ , r ₂ ,...,r _M ], where M is received by the target The number of APs in the area, and the value of the unreceived signal is -100db. For the current observation data, the SVM model of each partition can give the distribution probability p(y ^k =1|r) of whether the target is located in the corresponding area, where y ^k is the partition identifier, indicating that the target is located in the partition Ω _k , k∈{1 ,2,...,K}. Through the distribution probability p(y ^k =1|r) given by the SVM model of each partition, a preliminary judgment can be made on the partition where the target is located, and it can be used as a first-level judgment basis

The algorithm adopts the partition operation based on probability SVM, divides the observation data of reference points obtained in the offline phase into training set and test set, and trains the partition judgment model. Through K SVM models, the probability value of the target located in the corresponding partition can be obtained, but the signal distribution at the junction of the partition is complex, which is easy to cause errors in the judgment of the partition model. Therefore, by setting a secondary judgment basis, the problem of misjudging the test points near the partition handover can be overcome. Select two partitioned regions with the highest probability in the determined region, namely p(y ⁱ =1|r) and p(y ^j =1|r), i,j∈{1,2,...,K}, And p(y ⁱ =1|r)>p(y ^j =1|r), the difference can be expressed as

Δy _p =p(y ⁱ =1|r)-p(y ^j =1|r), (1)

When Δy _p >Δy, it means that the influence of the i partition on the test point is much greater than that of the j partition, and the reference point can be determined in the i partition, where Δy is the secondary judgment threshold. For Δy _p < Δy, both regions are determined as target regions, and corresponding region matching operations can be performed respectively, and the probability average of the target positions obtained by each partition is performed to obtain the final position estimate

Multivariate Gaussian Mixture Model Construction

The radiation model of the antenna is not uniform in direction due to the attenuation and reflection of the electromagnetic signal by the walls in the narrow area. The signal strength values of different APs at a certain time are also correlated with each other due to wall refraction, personnel occlusion, and channel problems. Therefore, simply sampling the value of a single AP will split the correlation between the AP signals, resulting in a large error in model construction.

Considering the mutual interference between the signals in the narrow partition after division, a multivariate Gaussian mixture model can be established by using the correlation of different AP signals, and by continuously increasing the number of Gaussian elements, the probability density function of different parameters can be used to weight and approximate the RP position in the partition. The joint probability density distribution between the acquired AP signals makes up for the neglect of the coupling relationship between AP signals in common work. The probability distribution function of the multivariate Gaussian mixture model can be expressed as

Among them, C represents the number of constituent elements, p _N (x|μ _c , P _c ) represents a Gaussian composition with a mean value of μ _c , a covariance of P _c , and the sum of the weights w _c is 1.

Based on equation (2), the _multivariate Gaussian mixture model can be expressed as

为多元高斯混合模型的组成元素权重，

为元素均值，

为元素协方差，C_k为组成元素数量。Among them, y ^k =1 indicates that the target is in the kth partition, r indicates the RSS value of each AP signal received at the reference point x,

is the element weight of the multivariate Gaussian mixture model,

is the element mean,

is the element covariance, and C _k is the number of constituent elements.

协方差为

的多元高斯函数，拟合分区Ω_k内N_k个参考点

与所获取各AP信号RSS值

的联合概率分布。利用k-means算法将样本数据聚类为C_k个初始簇，取各簇平均值与协方差用于初始化EM算法参数且各簇初始权重设置相同。联合概率分布的对数似然形式可表示为In order to improve the fitting effect of the MVGMM model to the sample data collected in the target area, the EM algorithm is used to estimate the model parameters. The MVGMM model selects C _k means as

The covariance is

The multivariate Gaussian function is fitted to N _k reference points in the partition Ω _k

and the obtained RSS value of each AP signal

The joint probability distribution of . The sample data is clustered into C _k initial clusters by the k-means algorithm, and the mean and covariance of each cluster are taken to initialize the parameters of the EM algorithm, and the initial weights of each cluster are set the same. The log-likelihood form of the joint probability distribution can be expressed as

隐变量γ_c,n表示第n个样本值属于第c个高斯组成元素的概率。where z ^k =[x ^k ; R ^k ],

The latent variable γc _,n represents the probability that the nth sample value belongs to the cth Gaussian component.

属于第c个高斯组成元素的概率可表示为From this, the observed data are determined

The probability of belonging to the c-th Gaussian component can be expressed as

均值

和协方差

分别为：The weight of the multivariate Gaussian function can be calculated by formula (5)

mean

and covariance

They are:

Repeat the clustering and EM estimation process for different values of _Ck :

初始均值

初始协方差

分区Ω_k内参考点位置

与相应采样值

c∈{1,2,..,C_k}(1) Input: the number of Gaussian constituent elements C _k , the initial weight of the Gaussian constituent elements

initial mean

initial covariance

Position of reference point in partition Ω _k

with the corresponding sample value

c∈{1,2,..,C _k }

(2) Calculated based on formula (5)

(3) Based on formulae (6) to (8), calculate w _c ,

(4) Repeat steps (2) to (3) until

均值

协方差

(5) Output: Gaussian composition element weights

mean

Covariance

Based on the above analysis, the offline fingerprint library can be expressed as (w ^k ; μ ^k ; P ^k ), where k∈{1,2,..,K}.

Online Phase: Targeting

通过分区判断模型可得目标在各分区的分布概率p(y_k＝1|r^now)，k∈{1,2,...,K}。The collected sample data is used to build an offline fingerprint database about the target area, through which the target position in the area can be estimated. If the RSS value of M APs accepted by the target at the current moment is

Through the partition judgment model, the distribution probability p(y _k =1|r ^now ) of the target in each partition can be obtained, k∈{1,2,...,K}.

根据多元高斯分布的条件概率准则，可得给定观测数据下目标位置的后验概率分布Using the current measurement value r ^now , the AP signal measurement value that is valid for target positioning in the corresponding partition Ω _k can be selected

According to the conditional probability criterion of multivariate Gaussian distribution, the posterior probability distribution of the target position under the given observation data can be obtained

Among them, based on the obtained offline fingerprint database (w ^k ; μ ^k ; P ^k ), we can get

Combining the partition probability with equation (9), the posterior distribution probability of each reference point of the target in the partition Ω _k can be obtained as

The distribution weight of each reference point in the partition Ω _k can be updated as

Among them, λ is the normalization factor, then the distribution weight of each reference point in the partition Ω _k as the target position is:

In summary, the location of the target is estimated as

Online Phase: Fingerprint Database Update

对分区模型进行参数更新。基于现有多元高斯混合模型，通过采集数据可对模型参数进行自适应调整，以推导出与现有环境具有更加紧密耦合关系的MVGMM模型。与EM算法一致，MVGMM模型参数的自适应过程也是一个两步估计^[13]。首先，通过k-means算法将新增观测数据

聚类为C_k组簇，并通过聚类所得数据权值

均值

协方差

初始化各高斯组成成分。利用式(5)-(8)可分别计算新增观测数据的统计权重

均值

与协方差

When the information entropy obtained in the partition Ω _k is less than the partition threshold, N _k data are newly collected by

Parameter update for the partition model. Based on the existing multivariate Gaussian mixture model, the model parameters can be adaptively adjusted by collecting data to derive the MVGMM model with a more tightly coupled relationship with the existing environment. Consistent with the EM algorithm, the adaptation process of the MVGMM model parameters is also a two-step estimation ^[13] . First, the new observation data will be added through the k-means algorithm.

Clustering into C _k groups of clusters, and data weights obtained by clustering

mean

Covariance

Initialize each Gaussian component. Using equations (5)-(8), the statistical weights of the newly added observation data can be calculated separately

mean

with covariance

The MVGMM model parameters corresponding to the partition Ω _k are adaptively optimized using the statistical parameters of the newly added data, namely:

ρ∈{w,m,v}用于平衡新旧数据对模型参数的影响程度，λ为归一化因子，r^ρ为参数的固有相关因子。新增数据更能体现当前环境内信号的分布状况，在自适应更新过程中，如下所示，平衡因子

ρ∈{w,m,v}更依赖于新增观测数据。in,

ρ∈{w,m,v} is used to balance the influence of old and new data on model parameters, λ is the normalization factor, and r ^ρ is the inherent correlation factor of the parameters. The newly added data can better reflect the distribution of signals in the current environment. During the adaptive update process, as shown below, the balance factor

ρ∈{w,m,v} is more dependent on the newly added observation data.

均值

协方差

新增分区Ω_k内采样数据

c∈{1,2,..,C_k}；(1) Input: the number of original Gaussian constituent elements C _k , the weight of the original Gaussian constituent elements

mean

Covariance

Added sampling data in partition Ω _k

c∈{1,2,.., _Ck };

(2) Based on formula (5), calculate

基于公式(15)至(17)计算

(3) Calculated based on formulae (6) to (8)

Calculated based on equations (15) to (17)

(4) Repeat steps (2) to (3) until

均值

协方差

(5) Output: The updated Gaussian component weights

mean

Covariance

Illustrate according to what is contained in the claims

Example 1: Person Location Detection

To detect the location information of indoor personnel in real time, it is necessary to equip each target person with a smart bracelet, and determine the target position through the AP signal received by the smart bracelet. However, because the indoor environment is complex, and AP devices are affected by time and change. The SMVGMM algorithm is used to fit the indoor distribution of AP signals, which can improve the positioning accuracy and reduce the impact of environmental factors. In the online update stage, the model parameters can be updated through the adaptive update algorithm to reduce the impact of time changes on the offline fingerprint database.

The test scene is a ring corridor environment on a certain floor in the C area of a college, and the WiFi routers evenly laid by the mobile operator in the college are selected as the AP signal source. Since the AP signal source is mainly laid in the corridor, and the one-sided corridor area is relatively open, the difference of the signal affected by the wall is small, so the one-sided corridor area with relatively combined physical structure and small difference in receiving AP signals is divided into corresponding partitions , the test area can be divided into K=4 partitions. The RP location adopts a grid topology and is arranged in the form of a corridor width and center. The adjacent RPs are separated by 1 meter, with a total of 368 RP points. The layout of AP signal sources and RP points is shown in Figure 2. According to the stability of AP signal sources in each partition, the number of AP signal sources in areas 1-4 is selected to be {4, 5, 4, 4} respectively. In order to reduce the impact of device differences on the positioning algorithm, the experiment uses a uniform model of smart phone for signal collection.

Collect the signal strength values of a total of 12 APs selected for all partitions at all reference points, the sampling interval is 1.2s, and the total collection is 4.8s (to avoid data buffering caused by the frequency of the device), constructed according to the process described in Section 3.1 Offline fingerprint library. During the test phase, the experimenter walked around the test area with the same smartphone, and then traveled to the test point to obtain real-time observation data through operations, and marked the current position. During the test, a total of 184 test points were obtained, with an interval of 1 meter.

After the target area is divided, the collection reference point can be identified as a zone, as shown in (a) in Figure 3, a zone model is constructed by sampling data and zone identification. Based on the partition model, the test data can be partitioned, and the partition effect is shown in (b) in Figure 3. The algorithm divides the test points that need to activate the second-level discrimination criteria into area 5 to represent the complex signal area.

In order to verify the effect of the algorithm on the construction of the RSS fingerprint map, Figure 4 shows the real numerical map of the AP3 signal in the target area and the estimated effect of the MVGMM model and the GP model. After fitting the joint distribution of the reference point position and the AP signal in the target area by using the MVGMM model, the RSS distribution of the AP3 signal in the area can be obtained by adjusting the corresponding position using formulas (7) and (8). In the experiment, the RSS measurement value of AP3 signal was extracted from the observation data of the reference point, 50% of the test value was selected as the training data, the model was trained, and the remaining 50% of the measurement value was used to verify the RSS distribution effect. Since AP facilities are not used in different partitions and AP3 mainly acts in partition one, Figure 5 shows the fitting effect of the two models on AP3 signals at some reference points in partition one. By comparing the actual measured value and the model estimated value, it can be seen that the MVGMM model can better fit the distribution of AP signals in the indoor environment through multiple Gaussian components, especially in partition one.

Based on the obtained MVGMM model and test data, the patent (abbreviated as: SMVGMM) is compared with the traditional WKNN algorithm and GP algorithm respectively, and the positioning accuracy of the algorithm is analyzed. The user travels in a circle in the target area, and the position estimation comparison diagram of the three algorithms is shown in Figure 6. It can be seen from Figure 6 that the target travel trajectory prediction obtained by the patent is smoother, and compared with the GP algorithm, its full-time positioning accuracy is improved. Figure 7 shows the box plots of the error values of the three algorithms at each test point. As can be seen from the figure, for the prediction of target trajectory, compared with the GP algorithm, the patent's full-time positioning accuracy is improved by more than 20%, which further confirms the smoothness of the target predicted trajectory obtained by the patent.

Due to the huge difference in positioning accuracy between the WKNN algorithm and the other two algorithms, Figure 8 only shows the cumulative probability comparison of the position estimation error between the patent and the GP algorithm. It can be seen from the figure that the initial error accumulation speed of the patent is slower than that of the GP algorithm, and the overall effect is better than that of the GP algorithm. On the other hand, it shows that the patent comprehensively improves the traditional algorithm through the correlation between AP signals in a narrow partition. positioning effect.

In order to verify the effect and value of the online update of the fingerprint database, the data was collected in the target area twice (with a one-week interval), and the parameters of the MVGMM model constructed from the original data were adaptively updated using the data collected for the second time. Through the test data collected twice, the fitting effect of the MVGMM model on the AP3 signal was compared before and after the parameter update. Figure 9 shows the fitting status of the model to the AP3 signal before and after the parameter update, and its fitting error. It can be seen from the figure that the test data collected twice are quite different in area 4, and the fitting effect of the model to the latest test data after parameter update is better than that of the previous model, especially in area 4.

Claims (1)

1. A self-adaptive positioning fingerprint database construction method under a complex indoor signal environment is characterized by comprising three stages: the method comprises the following steps of an off-line stage, an on-line positioning stage and an on-line fingerprint database updating stage, wherein the specific steps are as follows:

step one, dividing target area grids and constructing an offline fingerprint database

The target region is divided into K regions omega_kK ∈ {1, 2.., K } component, which is the region Ω_kDivision into N_kEach grid, the geometric center of each grid is taken as a reference point

Where N ∈ {1, 2., N_k}，

Is 2 × N_kThe dimension position matrix represents the RP position; for each RP location

The corresponding region is marked as

Wherein

And is

i ≠ k, meaning that the reference point is located in the k partition; in that

Collected from M_kRSS sample value of AP

Wherein

Is shown in

The collected RSS sample value from the mth AP, M ∈ {1,2_k}；

Step two, constructing SVM partition model of online positioning stage

Setting a probability classification model of a support vector machine in a one-to-many mode, carrying out two-classification identification on each preset partition according to whether a target is located in the partition, and constructing an SVM model of each partition through training data; for given K partitions, K SVM models are set, and measurement signals of all reference APs in each partition are taken to form current observation data r ═ r₁,r₂,...,r_M]Wherein M is the number of APs in the target received area, and the value of the unreceived signal is-100 db; aiming at the current observation data, each partitioned SVM model gives the distribution probability p (y) of whether the target is positioned in the corresponding region^k1| r), wherein y^kIdentify for partition, indicate that the target is located in partition Ω_kInner, K ∈ {1, 2.. multidot.K }, distribution probability p (y) given by each partitioned SVM model^k1 r), performing primary judgment on a partition where the target is located, and taking the partition as a primary judgment basis;

dividing the reference point observation data obtained in the step one into a training set and a test set by adopting partition operation based on a probability SVM, and training a partition judgment model SVM; the probability value of the target in the corresponding partition is obtained through K SVM models, and the problem of misjudgment of the test point near the partition handover is solved by setting a secondary judgment basis; selecting 2 subarea areas with the maximum probability in the judged area, namely p (y)ⁱ1| r) and p (y)^j1| r), i, j ∈ {1,2ⁱ＝1|r)＞p(y^j1| r), the difference being expressed as:

Δy_p＝p(yⁱ＝1|r)-p(y^j＝1|r)， (1)

when Δ y_pWhen the value is more than delta y, the influence of the partition i on the test point is far larger than that of the partition j, and the reference point is determined in the partition i, wherein the delta y is a secondary determination threshold value; for Δ y_pIf the difference is less than delta y, both the areas are judged as target areas, corresponding area matching operation is respectively carried out, and the target positions obtained by respective subareas are subjected to probability averaging to obtain the final position estimation

Step three, constructing a multivariate Gaussian mixture model (MVGMM)

Establishing a multivariate Gaussian mixture model (MVGMM) by utilizing the correlation of different AP signals, and compensating the neglect of the coupling relation between the AP signals in the traditional fingerprint library construction work by continuously increasing the number of Gaussian elements and utilizing the probability density function weighting and approximate simulation of different parameters to simulate the joint distribution condition between the RP position in the subarea and the obtained AP signals; the probability distribution function of the multivariate gaussian mixture model is expressed as:

wherein C represents the number of constituent elements, p_N(x|μ_c,P_c) Represents the mean value of μ_cCovariance of P_cOf the weight w_cThe sum of (a) and (b) is 1;

based on equation (2), using partition omega_kThe posterior probability of the inner RP location and RSS signal value joint distribution represents the multivariate gaussian mixture model as:

wherein, y^k1 indicates that the target is in the kth partition, r indicates the RSS value of each AP signal received at reference point x,

is the weight of the constituent elements of the multivariate Gaussian mixture model,

is the average value of the elements and is the average value of the elements,

is the element covariance, C_kIs the number of constituent elements;

the MVGMM adopts an EM algorithm to estimate model parameters; MVGMM model selection C_kMean value of

Covariance of

Fitting the partition omega_kInner N_kA reference point

With the RSS value of each AP signal acquired in the corresponding partition

A joint probability distribution of (a); clustering sample data into C by using k-means algorithm_kTaking the mean value and covariance of each cluster to initialize EM algorithm parameters, wherein the initial weight settings of each cluster are the same; the log-likelihood form of the joint probability distribution is expressed as:

wherein z is^k＝[x^k；R^k]，

Hidden variable gamma_c,nRepresenting the probability that the nth sample value belongs to the c-th gaussian component element;

thereby determining observation data

The probability of belonging to the c-th gaussian component element is expressed as:

calculating the height of the multiple element by the formula (5)Weight of a gaussian function

Mean value

Sum covariance

Respectively as follows:

for different C_kThe value repeat clustering and EM estimation process offline fingerprint library is represented as (w)^k；μ^k；P^k) K ∈ {1, 2.., K }, and step four, positioning the target in the subarea

Estimating the target position in the area through an offline fingerprint database; if the target at the current moment accepts the RSS values of M APs as

Obtaining the distribution probability p (y) of the target in each subarea through a subarea judgment model_k＝1|r^now)，k∈{1,2,...,K}；

Using the current measured value r^nowSelecting corresponding sub-region omega_kAP signal measurement value effective for internal target positioning

Obtaining the posterior probability distribution of the target position under given observation data according to the conditional probability criterion of the multivariate Gaussian distribution:

based on the resulting off-line fingerprint library (w)^k；μ^k；P^k) Obtaining:

combining the partition probability with the formula (9) to obtain the target in the partition omega_kThe posterior distribution probability of each internal reference point is as follows:

target is located in partition Ω_kThe distribution weight of each reference point in the image is updated as follows:

wherein, λ is normalization factor, then the division is Ω_kThe distribution weight of the internal reference points as the target positions is

In summary, the position of the target is estimated as

Step five, updating the fingerprint database

When dividing into omega_kWhen the entropy of the obtained information is less than the partition threshold value, newly collecting N_kData of a person

Updating parameters of the partition model; based on the existing multivariate Gaussian mixture model, model parameters are adaptively adjusted through data acquisition so as to deduce an MVGMM model which has a closer coupling relation with the existing environment; consistent with the EM algorithm, the self-adaptive process of the MVGMM model parameters is also a two-step estimation; firstly, newly-added observation data is added through a k-means algorithm

Cluster as C_kClustering, and obtaining data weight by clustering

Mean value

Covariance

Initializing each Gaussian composition; respectively calculating the statistical weight of the newly added observation data by using the formulas (5) to (8)

Mean value

And covariance

Dividing omega by using statistic parameter of new data_kThe corresponding MVGMM model parameters are subjected to self-adaptive optimization, namely:

wherein,

used for balancing the influence degree of new and old data on model parameters, wherein lambda is a normalization factor, and r^ρIs an intrinsic correlation factor of the parameter;

sixthly, updating the model parameters

And (5) carrying out real-time updating on the model parameters obtained by updating in the self-adaptive process in the second step and the third step.

Images