JP2023129115A - Position estimation model construction device, position estimation device, method for constructing position estimation model, and program - Google Patents

Abstract

To automatically construct a position estimation model by using not only wireless LAN information with a label but also wireless LAN information without a label as learning data.SOLUTION: A position estimation model construction device includes: an information acquisition unit for acquiring wireless LAN information made of a reception signal intensity of each access point; and a position estimation model learning unit for learning a position estimation model for estimating the positional information from the wireless LAN information by using, as learning data, an aggregate of the wireless LAN information related to the positional information and the positional information and the wireless LAN information acquired by the information acquisition unit.SELECTED DRAWING: Figure 1

Description

The present invention relates to a technique for estimating a position using wireless LAN radio wave information.

Indoor position estimation in indoor environments where it is difficult to use GPS is expected to have many applications, such as indoor navigation and monitoring, and is being actively researched.

Non-Patent Document 1 discloses a method using Wi-Fi (registered trademark) fingerprints, which is the most feasible among indoor position estimation methods. Note that wireless LAN technology such as Wi-Fi (registered trademark) will be referred to as "wireless LAN" hereinafter.

The method disclosed in Non-Patent Document 1 requires a large amount of training data, which is a correspondence between wireless LAN radio wave information and indoor position coordinates where the information is observed, and is therefore expensive to introduce.

On the other hand, in recent years, social networking services (SNS) that upload images taken with smartphone cameras have become popular, and it has become commonplace in our daily lives to take photos of our daily activities and upload them to SNS. . In addition, EXIF makes it easy to add context information such as GPS latitude and longitude (geotag) to photos taken with smartphones.

P. Davidson and R. Piche: A survey of selected indoor positioning methods for smartphones, IEEE Commu-nications Surveys Tutorials, vol.19, no.2, p.1347-1370, 2017, https://ieeexplore.ieee.org /document/7782316/. Zhang et al.: Location semantic labeling of Wi-Fi information using convolutional neural networks and camera images, IEE of Japan 69th Information Systems Research Meeting, IS-17-014 (March 2017). Diederik P. Kingma, Max Welling: Auto-Encoding Variational Bayes, ICLR, 2014, https://arxiv.org/abs/1312.6114.

As a conventional technique for constructing an indoor position estimation system using wireless LAN radio wave information attached to a photograph, there is a technique disclosed in Non-Patent Document 2, for example. However, the conventional technology has a problem in that data without labels (indoor position information) cannot be used for model learning used in indoor position estimation systems.

The present invention has been made in view of the above points, and uses not only labeled wireless LAN information but also unlabeled wireless LAN information as learning data to automatically construct a position estimation model. The purpose is to provide technology that enables this.

According to the disclosed technology, an information acquisition unit that acquires wireless LAN information consisting of received signal strength for each access point;
Estimating location information from the wireless LAN information using a set of pairs of wireless LAN information associated with location information and the location information and the wireless LAN information acquired by the information acquisition unit as learning data. A position estimation model construction device is provided, comprising: a position estimation model learning section that learns a position estimation model;

According to the disclosed technology, there is a technology that makes it possible to automatically construct a position estimation model by using labeled wireless LAN information as well as unlabeled wireless LAN information as learning data. provided.

1 is a configuration diagram of a position estimation model construction device 100. FIG. It is a flowchart which shows the flow of processing. FIG. 2 is a configuration diagram of a position estimation device 200. It is a figure for explaining the outline of a store estimation procedure. FIG. 2 is a diagram for explaining an overview of a position estimation model. FIG. 3 is a diagram showing a network configuration of a submodel. It is a diagram showing an example of the hardware configuration of the device.

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention (this embodiment) will be described below with reference to the drawings. The embodiments described below are merely examples, and embodiments to which the present invention is applied are not limited to the following embodiments.

(Summary of embodiment)
In this embodiment, an indoor commercial facility such as a shopping mall is assumed, and a wireless LAN indoor position estimation system (position estimation model) in such an environment is automatically constructed using photos taken by a terminal such as a smartphone. However, each photo is assumed to have wireless LAN radio wave information observed by the smartphone when the photo was taken as metadata.

Since anyone can easily take photos with a smartphone, it is possible to easily collect the information necessary to build a wireless LAN indoor position estimation system using methods such as crowdsourcing.

There is a technique disclosed in Non-Patent Document 2 as a study for automatically constructing a wireless LAN indoor position estimation system using wireless LAN radio wave information attached to such photographs. The technique disclosed in Non-Patent Document 2 uses a wireless LAN indoor position estimation method using linear discriminant analysis. On the other hand, in this embodiment, we propose an indoor position estimation method based on deep learning, and it is possible to learn a deep learning model by additionally using wireless LAN radio wave information without labels (indoor position coordinates). It is.

In addition, in this embodiment, indoor position estimation assuming an indoor commercial facility is explained as an example, but the application of the technology according to the present invention is not limited to this, and can be applied to any place. , it is also applicable to facilities. Furthermore, the technology of the present invention can also be applied to outdoor position estimation.

(Environment assumed in this embodiment)
As described above, in this embodiment, an indoor environment such as a shopping mall that is composed of a plurality of stores is assumed. It is assumed that the floor map of the environment is known. The floor map includes location information for each store (information on the area occupied by the store) and information on the store name. It is also assumed that multiple wireless LAN access points are installed in the environment.

Furthermore, it is assumed that a plurality of camera images taken at a plurality of points within the store environment are provided. However, wireless LAN information w _n is attached to the n-th camera image i _n . w _n is a N _A dimensional vector, where N _A is the number of access points in the environment. That is, w _n is a vector whose elements are the received signal strength of each access point. In other words, a certain element in w _n given to a camera image is the reception strength of a signal transmitted from a certain access point at the point where the camera image was taken. Hereinafter, it is assumed that "wireless LAN information" is a vector whose elements are the received signal strength of each access point.

Furthermore, in this embodiment, a series of unlabeled wireless LAN information can be used as additional learning data. This is assumed to be obtained from a smartphone or the like of a certain user walking in the environment, and is expressed as follows.

W ^u =[w ₁ ^u ,w ₂ ^u ,...,w _t ^u ,...]
However, w _t ^u is the wireless LAN information obtained at time t.

(Device configuration, operation overview)
In this embodiment, the position estimation model construction device 100 constructs (learns) a position estimation model. FIG. 1 shows a configuration example of a position estimation model construction device 100 in this embodiment.

As shown in FIG. 1, the location estimation model construction device 100 includes a store estimation section 110, a store information DB (database) 120, an information acquisition section 130, a location estimation model learning section 140, and a location estimation model DB 150. An overview of the operation of the position estimation model construction device 100 will be explained with reference to FIG.

In S1 (step 1), the store estimation unit 110 performs store estimation. The store estimation result is stored in the store information DB 120 as store information. It is assumed that the store information DB 120 stores information necessary for learning the position estimation model in addition to store information.

In S2, the position estimation model learning unit 140 uses the store information read from the store information DB 120 and the wireless LAN information W ^u acquired by the information acquisition unit 130 to learn a position estimation model. The learned position estimation model is stored in the position information model DB 150.

Note that the "model" in this embodiment is a neural network model, and when stored in a storage unit such as a DB, it is stored as data consisting of weight parameters and the like.

Alternatively, the store estimation section 110 may not be included in the position estimation model construction device 100. In that case, store information obtained in advance is stored in the store information DB 120.

The position estimation model constructed (learned) by the position estimation model construction device 100 is used for position estimation using wireless LAN information. FIG. 3 shows the configuration of a position estimation device 200 that performs such position estimation.

As shown in FIG. 3, the position estimation device 200 includes an information acquisition section 210, a position estimation section 220, an output section 230, and a position estimation model DB 240.

The position estimation model DB 240 stores learned position estimation models. The information acquisition unit 210 acquires wireless LAN information and inputs it to the position estimation unit 220. The position estimation unit 220 inputs the wireless LAN information into the position estimation model read from the position estimation model DB 240, thereby obtaining position information (information on the position where the wireless LAN information is measured) as an output from the position estimation model. and output it. The output unit 230 outputs the position information.

The location information output by the output unit 230 may be information indicating which store the point at which the wireless LAN information was measured is (that is, a store name, etc.), or may be two-dimensional coordinates. or other information.

Note that by adding the output unit 230 to the position estimation model construction apparatus 100, the position estimation model construction apparatus 100 may be used as the position estimation apparatus 200. Furthermore, the position estimation model construction device 100 may be referred to as a position estimation device, and the position estimation model learning section 140 may be referred to as a position estimation section.

The position estimation device 200 may be a device (eg, a server on a network) different from the terminal that measures wireless LAN information, or may be a terminal that measures wireless LAN information.

In addition, the position estimation model learning unit 140 of the position estimation model construction device 100 and the position estimation unit 220 of the position estimation device perform AP embedding processing, which will be described later, to input wireless LAN information into the position estimation model and perform position estimation. They have the same functionality in terms of performing. The position estimation model learning unit 140 differs from the position estimation unit 220 in that it has a loss calculation function and a parameter update function.

The processing operations in the position estimation model construction device 100 will be described in more detail below.

(S1: Store estimation)
First, the store estimation process executed by the store estimation unit 110 will be described. In the store estimation process, it is estimated in which store an image taken by a camera (an image with wireless LAN information) was taken. Since the store estimation process itself executed by the store estimation unit 110 is a conventional technique disclosed in Non-Patent Document 2, only an outline of the processing contents will be described. FIG. 4 shows an outline of the processing contents.

The store estimation process performed by the store estimation unit 110 is based on the idea that if a store in the environment is a branch of a chain store, its interior will be similar to other branches. To do this, first, we use an image search engine to collect web images related to the interior of a chain store based on the name of the store included in the environment. Then, using this image as learning data, a convolutional neural network (image classifier) that estimates store names from the captured images is trained.

However, non-chain stores and stores for which many web images could not be found are classified as "others". Then, the class of each image captured by a smartphone or the like is estimated by an image classifier. Since the location information of each store is included in the floor map, the rough location information where each image was taken can be obtained (excluding images classified as other). Finally, outlier removal is performed using wireless LAN information added to images classified as the same store.

Through the store estimation process in S1, it is possible to obtain information that associates images, wireless LAN information, and stores (that is, location information). This information is called store information. Store information is stored in store information DB 120.

(S2: Location estimation model learning)
Next, the position estimation model learning process executed by the position estimation model learning section 140 will be described.

The position estimation model is a model that takes wireless LAN information as input and estimates the store where the wireless LAN information was observed. The position estimation model learning unit 140 learns a position estimation model by using the image obtained in S1 (the wireless LAN information attached to it) and the estimated store information as learning data. The position estimation model of this embodiment is composed of a plurality of neural network models (submodels) in order to cope with nonlinearity in the wireless LAN space.

FIG. 5 shows an overview of the position estimation model. As shown in FIG. 5, each submodel configuring the position estimation model receives wireless LAN information as input and outputs position information of the point where the wireless LAN information was measured.

As explained below, the position estimation model uses the correlation between the signal strengths of nearby access points in order to deal with unstable wireless LAN radio field strength. Therefore, a convolutional neural network is used for each submodel that makes up the position estimation model, and the received signal strengths of multiple access points are convoluted and processed. Therefore, in the input wireless LAN information (received signal strength vector), nearby access points need to be located in close dimensions.

Therefore, in S2, the position estimation model learning unit 140 performs a process of determining the order of access points in the vector (referred to as "AP embedding"), and then performs learning of the position estimation model (submodel group).

<Embed AP>
First, we will explain AP embedding. The position estimation model learning unit 140 performs a process of determining the order of access points in wireless LAN information (wireless LAN signal vector) as AP embedding process. Assuming that there are multiple wireless LAN signal vectors, the nth wireless LAN signal vector is expressed as follows.

w _n =[s _n ¹ ,s _n ² ,…,s _n ^j ,…,s _n ^N_A )] ^T ,
However, s _n ^j is the signal strength from the jth access point. As described above, in the arrangement of elements of a wireless LAN signal vector, elements corresponding to nearby access points need to be located in a close arrangement (for example, the j-th and j+1-th).

This problem can be viewed as a problem in which the order of access points is determined given the physical proximity of the access points. In this embodiment, the physical proximity (distance) of two access points is expressed as the similarity of the signal strengths of the access points (the reciprocal of the correlation of signal strengths or the average error) in the wireless LAN information added to the image group. demand.

The position estimation model learning unit 140 of this embodiment solves this problem by applying it to the traveling salesman problem. The traveling salesman problem is a combination of, given a set of cities and the travel cost between each two cities, finding the route that visits all the cities exactly once and returns to the starting point with the minimum total travel cost. It is an optimization problem.

The cities are access points, the distance (cost) between cities is the estimated physical distance between the access points, and the order of cities in the obtained route is used as the order of access points in w _n .

Here, c _ab is the distance between cities (access points) a and b. x _ab is a binary variable, and when it is 1, the route includes the path between cities a and b. If it is 0, it is not included. The position estimation model learning unit 140 finds a route that minimizes the travel distance by solving the following optimization problem.

上記の式（全５行）の１行目は、後述する制約とともに、全ての都市（アクセスポイント）をちょうど一度ずつ巡り出発地に戻る巡回路のうちで総移動コストが最小のものを求めることを意味する。

The first line of the above equation (total of 5 lines), along with the constraints described later, calculates the route with the minimum total travel cost among the routes that visit all cities (access points) exactly once and return to the starting point. means.

The 3rd and 4th lines are constraints that only pass through the city (access point) once. Q represents a subset of cities (subroute), and the constraint on the fifth line is for the finally obtained route to pass through all cities.

<Configuration of position estimation model>
As described above, the position estimation model in this embodiment is composed of a plurality of sub-models. Each sub-model has an area and is responsible for that area. Specifically, a sub-model is prepared for each store (area), and the sub-model is trained using learning data obtained from that store and surrounding stores. In other words, the submodel is in charge of that store and surrounding stores. Surrounding stores are stores whose centers are at a distance of top k (nearest k) from the center of the target store. Here, there is no learning data obtained from stores that are not chain stores (because only the others class is estimated in the store estimation in S1). Therefore, for stores that are not chain stores, learning is performed based on learning data (wireless LAN information) virtually generated by sub-models of surrounding stores.

FIG. 6 shows an example of the configuration of the sub-model. The submodel is designed based on a convolutional variational autoencoder (CVAE) (Non-Patent Document 3), and includes an encoder 10 and a decoder 20. Note that using CVAE as a submodel is one example. Autoencoders other than CVAE may be used as submodels.

As shown in FIG. 6, the encoder 10 includes Conv1D11, Conv1D12, Flatter13, and Dense14. The decoder 20 includes Dense 21, Reshape 22, DeConv1D 23, and DeConv1D 24.

The input to the encoder 10 is wireless LAN information, and the wireless LAN information in the order of access points obtained by AP embedding is used as input. The encoder 10 outputs two-dimensional coordinates corresponding to the measurement position of the input wireless LAN information, and also calculates the mean and variance (standard deviation) when the latent representation of the learning data (wireless LAN information) is assumed to be a normal distribution. Output.

These two-dimensional coordinates are the estimated coordinates of the position where the input wireless LAN information was measured. CVAE can learn its parameters based on the constraint that the latent representation follows a normal distribution. This is because the distribution of signal strength in wireless LAN information is considered to follow a locally simple distribution. This assumption also facilitates the generation of virtual wireless LAN information for non-chain stores.

The submodel decoder 20 restores the input wireless LAN information. By using this decoder 20, it is possible to (i) generate wireless LAN information from position coordinates and (ii) learn a model from unlabeled data (wireless LAN information to which no position information is attached).

<Learning of position estimation model>
The location estimation model learning unit 140 performs learning of the submodel of the location estimation model using learning data (pairs of wireless LAN information and store (location information)) obtained from the store that the submodel is in charge of and surrounding stores. . During learning, the position estimation model learning unit 140 updates the parameters of the sub-model (CVAE) so as to minimize the following function.

ただし、θ_eとθ_dはそれぞれ、エンコーダ１０のパラメータ集合とデコーダ２０のパラメータ集合である。N_Tは学習データ（無線LAN情報と店舗のペア）の数である。L_r( )は入力と出力の無線LAN情報の再構成誤差（MSE）である。L_g( )は正則化項であり、学習データの潜在表現の分布と標準正規分布とのKLダイバージェンスである。L_p ( )は、無線LAN情報によるエンコーダ１０の位置推定誤差である。

However, θ _e and θ _d are a parameter set for the encoder 10 and a parameter set for the decoder 20, respectively. N _T is the number of learning data (wireless LAN information and store pairs). L _r ( ) is the reconstruction error (MSE) of input and output wireless LAN information. L _g ( ) is a regularization term and is the KL divergence between the distribution of the latent representation of the training data and the standard normal distribution. L _p ( ) is the position estimation error of the encoder 10 based on the wireless LAN information.

The position estimation error is an error between the two-dimensional coordinates estimated by the encoder 10 and the correct answer data (correct answer store). If the two-dimensional coordinates are within the area of the correct store, the error is 0; otherwise, the error is the distance between the two-dimensional coordinates and the boundary closest to the two-dimensional coordinates among the boundaries of the area of the correct store.

If a sequence of unlabeled wireless LAN information (W ^u described above) observed in the vicinity of the area covered by the submodel is given, the submodel is additionally trained using that data. The sequence of wireless LAN information is assumed to be obtained while walking. Among the series, a series of wireless LAN information having a close Euclidean distance to the wireless LAN information used for learning the submodel is used as additional learning data. The Euclidean distance being close means, for example, that the Euclidean distance is less than or equal to a threshold value. During training, CVAE parameters are updated to minimize the following function.

ただし、N'_Tは追加学習データの数である。L_u( )は歩行時に得られた追加学習データのための制約である。歩行中に得られた系列における隣接する無線LAN情報は、得られた位置が近いと考えられる。そこで、隣接する無線LAN情報の推定２次元座標の距離がしきい値以下なら、L_u( )はそれらのペアに対して0を返す。そうでない場合は、その２次元座標間の距離を返す。

However, _N'T is the number of additional learning data. L _u ( ) is a constraint for additional learning data obtained during walking. Adjacent wireless LAN information in a series obtained while walking is considered to have a nearby location. Therefore, if the distance between the estimated two-dimensional coordinates of adjacent wireless LAN information is less than or equal to the threshold, L _u ( ) returns 0 for those pairs. Otherwise, returns the distance between the two-dimensional coordinates.

(Position estimation)
The position estimating device 200 uses a learned position estimating model (a set of sub-models) to estimate the store where the wireless LAN information has been collected. First, a submodel is selected for estimating the store of wireless LAN information. Here, it is assumed that the position estimation model DB 240 of the position estimation device 200 stores learning data used in learning together with the learned position estimation model. In the following processing, learning data is also read out from the position estimation model DB 240 and used.

That is, the position estimation unit 220 calculates the Euclidean distance between the wireless LAN information acquired by the information acquisition unit 210 and the average of the wireless LAN information used for learning each submodel, and selects the one with the smallest distance. Let the corresponding submodel be the main submodel. Then, a sub-model group of stores near the store for which the main sub-model is in charge is defined as a sub-sub-model group.

The position estimation unit 220 performs AP embedding processing on the wireless LAN information acquired by the information acquisition unit 210, and inputs the wireless LAN information subjected to the AP embedding process to the main submodel and sub submodel group. The position estimation unit 220 outputs the store closest to the average of the two-dimensional coordinates estimated by the encoders of the main sub-model and the sub-sub model group as the final estimation result. Note that the average of the two-dimensional coordinates itself may be output as the estimation result.

(Hardware configuration example)
Both the position estimation model construction device 100 and the position estimation device 200 described above can be realized by, for example, causing a computer to execute a program. This computer may be a physical computer or a virtual machine on the cloud. The position estimation model construction device 100 and the position estimation device 200 are collectively referred to as the “device”.

That is, the device can be realized by using hardware resources such as a CPU and memory built into a computer to execute a program corresponding to the processing performed by the device. The above program can be recorded on a computer-readable recording medium (such as a portable memory) and can be stored or distributed. It is also possible to provide the above program through a network such as the Internet or e-mail.

FIG. 7 is a diagram showing an example of the hardware configuration of the computer. The computer in FIG. 7 includes a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, etc., which are interconnected by a bus BS.

A program for realizing processing by the computer is provided, for example, by a recording medium 1001 such as a CD-ROM or a memory card. When the recording medium 1001 storing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily need to be installed from the recording medium 1001, and may be downloaded from another computer via a network. The auxiliary storage device 1002 stores installed programs as well as necessary files, data, and the like.

The memory device 1003 reads the program from the auxiliary storage device 1002 and stores it when there is an instruction to start the program. The CPU 1004 implements functions related to the device according to programs stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network or the like. A display device 1006 displays a GUI (Graphical User Interface) or the like based on a program. The input device 1007 is composed of a keyboard, a mouse, buttons, a touch panel, or the like, and is used to input various operation instructions. An output device 1008 outputs the calculation result.

(Additional note)
Regarding the above embodiments, the following additional notes are further disclosed.
(Additional note 1)
memory and
at least one processor connected to the memory;
including;
The processor includes:
Obtain wireless LAN information consisting of received signal strength for each access point,
A location estimation model that estimates location information from wireless LAN information is created using a set of pairs of wireless LAN information associated with location information and the location information and the acquired wireless LAN information as learning data. Learning location estimation model construction device.
(Additional note 2)
The processor changes the order of the received signal strengths in the wireless LAN information consisting of the received signal strengths of each access point so that nearby access points are located in a close order based on the traveling salesman problem in which cities are considered as access points. The position estimation model construction device according to Supplementary Note 1, wherein the wireless LAN information in the determined arrangement order is input into the position estimation model.
(Additional note 3)
The position estimation model construction device according to Supplementary Note 1 or 2, wherein the position estimation model has a plurality of sub-models, each of which is assigned to a different area.
(Additional note 4)
According to Supplementary Note 3, each submodel configuring the plurality of submodels includes an encoder that receives wireless LAN information as input and outputs position information, and a decoder that reconstructs the wireless LAN information input to the encoder. Location estimation model construction device.
(Additional note 5)
The processor includes:
Learning the sub-model using the set of sets obtained from the area covered by the sub-model and its surrounding area,
The position estimation model construction device according to Supplementary Note 3 or 4, wherein additional learning of the sub-model is performed using wireless LAN information acquired by the information acquisition unit that is observed in the vicinity of the area covered by the sub-model.
(Additional note 6)
A position estimation device that performs position estimation using a position estimation model learned by the position estimation model construction device according to any one of Supplementary Notes 1 to 5,
memory and
at least one processor connected to the memory;
including;
The processor includes:
Obtain wireless LAN information consisting of received signal strength for each access point,
A position estimation device that estimates position information of a position where the wireless LAN information is observed by using the acquired wireless LAN information as input to the position estimation model.
(Supplementary Note 7)
A method for constructing a position estimation model executed by a computer, the method comprising:
Obtain wireless LAN information consisting of received signal strength for each access point,
Learn a location estimation model that estimates location information from wireless LAN information by using the set of pairs of wireless LAN information associated with location information and the location information and the acquired wireless LAN information as learning data. How to build a location estimation model.
(Supplementary Note 8)
A non-temporary storage medium storing a program executable by a computer to execute a position estimation model construction process,
The position estimation model construction process includes:
Obtain wireless LAN information consisting of received signal strength for each access point,
A location estimation model that estimates location information from wireless LAN information is created using a set of pairs of wireless LAN information associated with location information and the location information and the acquired wireless LAN information as learning data. Learning non-transitory storage medium.

Although the present embodiment has been described above, the present invention is not limited to such specific embodiment, and various modifications and changes can be made within the scope of the gist of the present invention as described in the claims. It is possible.

10 Encoder 20 Decoder 100 Position estimation model construction device 110 Store estimation unit 120 Store information DB
130 Information acquisition unit 140 Position estimation model learning unit 150 Position estimation model DB
200 Position estimation device 210 Information acquisition unit 220 Position estimation unit 230 Output unit 240 Position estimation model DB
1000 Drive device 1001 Recording medium 1002 Auxiliary storage device 1003 Memory device 1004 CPU
1005 Interface device 1006 Display device 1007 Input device 1008 Output device

Claims (8)

an information acquisition unit that acquires wireless LAN information consisting of received signal strength for each access point;
Estimating location information from the wireless LAN information using a set of pairs of wireless LAN information associated with location information and the location information and the wireless LAN information acquired by the information acquisition unit as learning data. A position estimation model construction device comprising: a position estimation model learning unit that learns a position estimation model; The position estimation model learning unit adjusts the order of received signal strengths in wireless LAN information consisting of received signal strengths of each access point, so that nearby access points are located in a close order, and the city is considered as an access point. The position estimation model construction device according to claim 1, wherein the wireless LAN information in the determined arrangement order is determined by solving a salesman problem and is input into the position estimation model. The position estimation model construction device according to claim 1 or 2, wherein the position estimation model has a plurality of sub-models, each of which is assigned to a different area. 4. Each submodel configuring the plurality of submodels includes an encoder that inputs wireless LAN information and outputs position information, and a decoder that reconstructs the wireless LAN information input to the encoder. Location estimation model construction device. The position estimation model learning unit includes:
Learning the sub-model using the set of sets obtained from the area covered by the sub-model and its surrounding area,
5. The position estimation model construction device according to claim 3, wherein additional learning of the sub-model is performed using wireless LAN information acquired by the information acquisition unit, which is observed in the vicinity of an area covered by the sub-model. A position estimation device that performs position estimation using a position estimation model learned by the position estimation model construction device according to any one of claims 1 to 5,
an information acquisition unit that acquires wireless LAN information consisting of received signal strength for each access point;
A position estimating unit that estimates position information of a position where the wireless LAN information is observed by using the wireless LAN information acquired by the information acquiring unit as input to the position estimation model. A method for constructing a position estimation model executed by a computer, the method comprising:
Obtain wireless LAN information consisting of received signal strength for each access point,
Learn a location estimation model that estimates location information from wireless LAN information by using the set of pairs of wireless LAN information associated with location information and the location information and the acquired wireless LAN information as learning data. How to build a location estimation model. A program for causing a computer to function as each part of the position estimation model construction device according to any one of claims 1 to 5.

Images