Location estimation model and method using area features, and route prediction program and device

Abstract

To provide a position estimation model that estimates the position of a target taking into account the features in an area.SOLUTION: The present position estimation model estimates, from time-series position information of a target in the area to be examined, the subsequent position information of the target. The position estimation model causes a computer to function as: (a) a district feature estimation unit that generates, using a district feature estimator having been trained by learning data including the time-series data of a district feature quantity and the data of a subsequent district feature quantity, a district feature quantity that comes after this time-series district feature quantity, on the basis of a time-series district feature quantity that pertains to a plurality of unit districts corresponding to each of the time-series position information of the target; and (b) a position information determination unit that generates feature corresponding district information that pertains to a unit district corresponding to a subsequent district feature quantity on the basis of the district feature quantity and each of district feature quantities pertaining to the unit district, and determines the subsequent position information of the target on the basis of this feature corresponding district information.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a technique for estimating the position of an object. [Background technology]

[0002] When designing and updating urban transportation networks, and when deciding and improving transportation policies, it is important to predict the travel behavior of people and automobiles.In recent years, in predicting this travel behavior, methods have been proposed that use activity-based models that model user travel behavior in detail, instead of the conventional method of predicting travel demand based only on future daytime and nighttime population.

[0003] For example, Non-Patent Document 1 discloses a technology for predicting destinations and the occurrence of travel itself based on the attractiveness of the destination (building statistics) and travel costs (traffic conditions). This technology defines a series of actions from home to a main destination and then back home as a tour, and by incorporating this tour into an activity-based model, it aims to reproduce the user's daily behavior in urban transportation.

[0004] In many cases, models for estimating travel behavior, including the model disclosed in Non-Patent Document 1, have been constructed based on data acquired through questionnaire surveys. In contrast, in recent years, vigorous research has been conducted into constructing models of people's travel behavior from smartphone location information, which can be acquired in large quantities.

[0005] For example, Non-Patent Document 2 discloses a technology that uses GPT-2, a deep learning algorithm developed in natural language processing, to model human travel behavior based on smartphone location information, generate travel trajectories according to the travel history up to the present, and perform highly accurate travel behavior simulations.

[0006] In this technology, a large number of sequences of location information are regarded as text, and GPT-2 trained in this way is used. After encoding the sequences of location information into fixed-length vectors, the next location in this sequence is predicted. Therefore, for users in a certain area, based on their previous movement trajectories, it is possible to distinguish, for example, whether they will pass through the area as it is or return home at the end of the day. Specifically, in Non-Patent Document 2, an experiment was actually conducted in Kyoto to determine whether the user would return home from the obtained movement trajectories of the user, and it is stated that an experimental result was obtained indicating that the proportion of users determined to return home generally matched the actual data of this proportion.

Prior Art Documents

Non-Patent Documents

[0007]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0008] The movement behaviors of the objects as described above are expected to be affected by environmental changes, especially changes in land use such as redevelopment of areas or construction of large commercial facilities. In fact, local governments, developers, etc. have a strong need to grasp the magnitude of such impacts and predict changes in the flow of people, the number of visitors, and the surrounding traffic conditions.

[0009] For example, in the past, local governments and other organizations were predicated on taking long-term, fixed policies such as road construction and changes to regular bus routes, and therefore placed importance on predicting the total volume of traffic and pedestrian flow and locations of congestion. On the other hand, in policies related to dynamically provided transportation services such as on-demand taxis and ride-sharing taxis, which are currently attracting attention, it is important to predict the travel behavior of individual users and estimate and understand detailed travel demand.

[0010] It is known that a user's travel behavior is usually strongly influenced by the environment, such as land use, depending on the place of residence and lifestyle, etc. However, until now, there have been few predictions of travel behavior that take such environments into consideration.

[0011] For example, in the technology disclosed in Non-Patent Document 1, the way a user has traveled to reach their current location is not taken into consideration at all when estimating the next location. As a result, it is extremely difficult to predict travel behavior that takes into consideration the influence of the environment that each user has passed through and the environment of their current location. Furthermore, since the behavior is estimated on a tour-by-tour basis, it is only possible to predict travel behavior such as "returning home at night." In other words, it is difficult to predict travel behavior that passes through a target area for sightseeing or transportation purposes, for example.

[0012] Furthermore, research into predicting user travel behavior using language models used in natural language processing, such as the technology disclosed in Non-Patent Document 2, has only just begun, and the reality is that predictions that take into account the influence of the environment, such as buildings, have not yet been made.

[0013] Therefore, an object of the present invention is to provide a location estimation model and a location estimation method that estimate the location of an object by taking into account the characteristics of an area, and a route prediction program and a route prediction device that predict the route of an object using such a location estimation model. [Means for solving the problem]

[0014] According to the present invention, there is provided a location estimation model for estimating future location information of an object from time-series location information relating to the position of the object in a certain area, the model comprising: A zone feature value representing a feature of each unit zone within the area is set, a zone feature estimation unit that generates zone features that are subsequent to the zone features in the time series, based on zone features in the time series that are zone features related to a plurality of unit zones corresponding to each of the time series location information, using a zone feature estimator trained with learning data including time series data of the zone features and data of subsequent zone features; a location information determination unit that generates feature corresponding area information, which is information related to the unit area corresponding to the subsequent area feature amount, based on the subsequent area feature amount and each of the area feature amounts related to the unit area, and determines subsequent location information of the object based on the feature corresponding area information; A location estimation model is provided that allows a computer to function as a

[0015] In one embodiment of the position estimation model according to the present invention, the position information determination unit: a unit area estimation unit that generates, based on the time-series unit areas corresponding to each of the time-series location information, history-corresponding area information, which is information related to a unit area that comes after the unit area in the time series, by using a unit area estimator trained with learning data including time-series data of the unit area and data of the unit area that comes after it; and determining subsequent position information of the object based on the feature corresponding area information and the history corresponding area information.

[0016] In another embodiment of the location estimation model according to the present invention, the unit area estimator comprises an embedding unit trained to convert input data into an embedding representation vector; the location information determination unit performs an embedding process on area feature map information consisting of area feature values ​​of each unit area within the area, using the same parameters as those of the embedding unit, and generates the feature corresponding area information using the area feature values ​​related to each embedded expression related to the unit area in the area feature map information that has been subjected to the embedding process and the area feature values ​​to be generated; It is also preferable that the generated feature corresponding area information is information of embedded expressions related to the unit area corresponding to the subsequent area feature.

[0017] Furthermore, it is also preferable that the unit area estimator according to the present invention is constructed using a plurality of modules including a self-attention layer and a fully-connected neural network layer in a Transformer.

[0018] It is also preferable that the location information determination unit according to the present invention determines subsequent location information of the object based on the feature corresponding area information and the history corresponding area information using a neural network algorithm, an attention mechanism, or a Hadamard product.

[0019] Furthermore, it is also preferable that the area feature estimator according to the present invention is constructed using a transformer encoder and decoder or a fully connected neural network algorithm.

[0020] It is also preferable that the location information determination unit according to the present invention generates the feature corresponding area information using an inner product or attention mechanism based on the subsequent area feature and each of the area feature values ​​related to the unit area.

[0021] Furthermore, it is also preferable that the area feature amount according to the present invention includes a quantity related to an object that exists or occurs in the corresponding unit area and that may affect the movement or stay of the object. For example, it is also preferable that the area feature amount includes a quantity related to at least one of the type, use, size, and number of buildings or structures that exist in the corresponding unit area.

[0022] Furthermore, it is also preferable that the location estimation model according to the present invention can determine the subsequent location information of the target without retraining the estimator used, even if the area feature map information consisting of area features of each unit area within the area is changed.

[0023] The present invention also provides a route prediction program that causes a computer to function as route prediction means for predicting a route including subsequent position information of an object based on time-series position information relating to the position of the object in a certain area, using the position estimation model described above.

[0024] According to the present invention, there is further provided a route prediction device having route prediction means for predicting a route including subsequent position information of an object based on time-series position information relating to the position of the object in a certain area, by using the above-described position estimation model.

[0025] According to the present invention, there is further provided a position estimation method for estimating future position information of an object from time-series position information relating to the position of the object in a certain area, the method comprising: A zone feature value representing a feature of each unit zone within the area is set, generating, based on the time-series area features that are area features related to a plurality of unit areas corresponding to each of the time-series location information, area features that are subsequent to the time-series area features, using an area feature estimator trained with learning data including time-series data of the area features and data of subsequent area features; generating feature-corresponding area information, which is information related to the unit area corresponding to the subsequent area feature, based on the subsequent area feature and each of the area feature amounts related to the unit area, and determining subsequent position information of the object based on the feature-corresponding area information; A computer-implemented location estimation method is provided, comprising: [Effects of the Invention]

[0026] The location estimation model and location estimation method of the present invention can estimate the location of an object by taking into account the characteristics of an area. Furthermore, the route prediction program and route prediction device of the present invention can predict the route of an object using such a location estimation model. [Brief explanation of the drawings]

[0027] [Figure 1] FIG. 1 is a schematic diagram illustrating an embodiment of a location estimation model according to the present invention. [Figure 2] FIG. 10 is a schematic diagram illustrating another embodiment of a location estimation model according to the present invention. [Figure 3] 10 is a histogram showing the calculation results of cos similarity in an embodiment of inner product processing in the position information determination means according to the present invention. [Figure 4] 1 is a schematic diagram for explaining the policy and procedure of a location estimation process in a location estimation model according to the present invention. FIG. [Figure 5] FIG. 10 is a schematic diagram illustrating yet another embodiment of a location estimation model according to the present invention. [Figure 6] FIG. 10 is a schematic diagram illustrating yet another embodiment of a location estimation model according to the present invention. [Figure 7] FIG. 10 is a schematic diagram illustrating yet another embodiment of a location estimation model according to the present invention. [Figure 8] 1 is a functional block diagram showing a functional configuration of an embodiment of a route prediction device according to the present invention; DETAILED DESCRIPTION OF THE INVENTION

[0028] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0029] [Location estimation model] FIG. 1 is a schematic diagram showing one embodiment of a location estimation model (location estimation model 1) according to the present invention.

[0030] The location estimation model 1 of this embodiment shown in FIG. 1 is a machine learning model that estimates the "later location information" of a "target" (location estimation target) from "time-series location information" relating to the target's location in a survey area (e.g., Tokyo or the entire country). In this embodiment, the "target" can be, for example, a human being, such as a pedestrian or a public transport user, or a vehicle traveling on a road, and is an entity that moves independently or autonomously while being influenced (acted upon) by the surrounding environment, including the land use patterns and buildings. Incidentally, the "target" can also be an aircraft such as a drone, a bicycle, a motorcycle, an electric wheelchair, a yacht, a boat, a ship, or even an animal being surveyed for habitat survey.

[0031] Furthermore, the "time-series location information" of the "target" can be information generated based on positioning results obtained, for example, from a user's mobile device equipped with a GPS (Global Positioning System) positioning function or a communication terminal equipped with a GPS positioning function installed in a vehicle, after obtaining permission to use the location information. For example, it may be latitude and longitude information every five minutes from 7:00 to 12:00. Furthermore, this "time-series location information" may include location information generated from a base station positioning method, a wireless LAN (Local Area Network) access point positioning method, positioning results using a smart pole, or even survey results.

[0032] In addition, in this embodiment, the estimated "later location information" is location information that is later than the time period related to the "time-series location information," and in the above example, it could be, for example, latitude and longitude information at 12:05, or latitude and longitude information every 5 minutes from 12:05 to 13:00.

[0033] Furthermore, in the area to be surveyed, an "area feature amount" that represents the characteristics of each unit area (mesh in this embodiment) within this area is set. In this embodiment, this "area feature amount" includes amounts of things that exist or occur within the corresponding unit area (mesh) and that may affect the movement or stay of the "target," particularly the purpose of the movement or stay.

[0034] Incidentally, hereinafter, "area feature quantity" refers to the quantity related to the number of buildings belonging to each classification when buildings (including, for example, commercial facilities and stations) existing within a mesh are classified by type (private houses, apartment buildings, offices, complex buildings, etc.), use (industry) (food and beverage, retail, education, medical, etc.), and size (number of floors, floor area, etc.), but it is not limited to this. For example, it may include a quantity related to at least one of the type, use, size, and number of buildings or structures, and may also include a quantity related to at least one of the type, use, size, and number of various events (occurring during the survey period) and various POIs (Points of Interest).

[0035] Under the above-described settings, the location estimation model 1 is specifically as follows: (A) a zone feature estimation unit 11 that generates a "later zone feature" that is a zone feature that is later (e.g., immediately after) the "time-series zone feature" (time period) using a "zone feature estimator" based on "time-series zone feature" that is a zone feature related to a plurality of unit zones corresponding to each of the "time-series location information"; (B) Based on the generated "later area feature" and each of the area feature values ​​related to the unit area (mesh) (using the inner product (cosine similarity) of the two in Figure 1), "feature-corresponding area information" is generated, which is information related to the unit area (mesh) corresponding to the "later area feature." Based on this "feature-corresponding area information," the "later position information" of the target (in Figure 1, "mesh m^ at t=i+1") is calculated. i+1 a location information determination unit 12 that determines the location information (unit area) It is a machine learning model that allows computers to function.

[0036] Here, the "area feature estimator" in (A) above is an estimator trained with learning data including time-series data of area feature amounts and subsequent correct answer data of area feature amounts, as will be explained in detail later. Also, in this embodiment, each of the area feature amounts relating to the unit area (mesh) in (B) above is extracted and used from "area feature map information" (an updatable map in FIG. 1) consisting of the area feature amounts of each unit area (mesh) within the survey target area.

[0037] In this way, the location estimation model 1 can determine the "future location information" of the object, i.e., estimate the future location of the object, by taking into account the area feature quantities (in this embodiment, area feature map information (updatable map)) of each unit area (mesh) that are characteristics within the area.

[0038] In this embodiment, the "area characteristic map information" (updatable map) is composed of area characteristic quantities, i.e., building statistical vectors (described later), which are quantities related to the number of buildings in each classification when buildings are classified by type, use, and size. Naturally, this information must be updated, for example, when a new commercial facility or station is constructed within the area. However, in this embodiment (as described later in detail), even if the "area characteristic map information" (updatable map) is changed, the location estimation model 1 can determine the "future location information" of the target location estimation without retraining the estimator (area characteristic estimator 11a in FIG. 1). In other words, it is possible to quickly estimate the impact of changes in the environment (building distribution, land use pattern) on the future location of a person, whose location is being estimated, and thus on people flow, without retraining the estimator.

[0039] Incidentally, location estimation model 2 (Figure 2), location estimation model 3 (Figure 5), location estimation model 4 (Figure 6), and location estimation model 5 (Figure 7), which will be explained in detail later, are also machine learning models that cause a computer to function as an area feature estimation unit (21, 31, 41, 51) and a location information determination unit (22, 32, 42, 52), and achieve the same effects as those achieved by location estimation model 1 described above.

[0040] [Model configuration, location estimation method] Next, a detailed description will be given of a location estimation model 2 as one embodiment that includes the functional configuration of the above-described location estimation model 1. Fig. 2 is a schematic diagram showing another embodiment (location estimation model 2) of a location estimation model according to the present invention.

[0041] In the embodiment described below, the area to be surveyed is divided into N (e.g., 50,000) meshes (unit areas). Here, the meshes may be, for example, regional meshes of the Japanese Geodetic System, and the data for each mesh m may be, for example, a regional mesh code of JIS X 410. Furthermore, hereinafter, a point in time (time) related to time-series data will be expressed as t = k. Here, the point in time (time) k refers to a point in time (time) that has elapsed a time (Δt × k) from a reference point in time (reference time), where Δt is the unit time.

[0042] As shown in Figure 2, location estimation model 2 is (A) a district feature estimation unit 21 that uses a district feature estimator 21a to perform estimation processing of district feature quantities (building statistical vectors, which will be described later in this embodiment); (B) a location information determination unit 22 having a unit area estimation unit 221 that performs estimation processing of a mesh (unit area) using a unit area estimator 221a; These are functional components that are realized by the execution of programs installed in the computer. In other words, it is a model that makes the computer function as these functional components.

[0043] Here, the area feature estimation unit 21 in (A) above may include an area feature estimator 21a, or may be able to use an external area feature estimator 21a (for example, installed on an external server). Also, the unit area estimation unit 221 in (B) above may include a unit area estimator 221a, or may be able to use an external unit area estimator 221a (for example, installed on an external server). Here, the fact that the estimator used may be an included one or an external, available one also applies to the above-mentioned location estimation model 1 (FIG. 1), and location estimation model 3 (FIG. 5), location estimation model 4 (FIG. 6), and location estimation model 5 (FIG. 7) described later.

[0044] The area feature estimation unit 11 and area feature estimator 11a of the above-mentioned location estimation model 1 (FIG. 1) have the same functional configuration as the area feature estimation unit 21 and area feature estimation unit 21a, which will be described later. Similarly, the location information determination unit 12 of the location estimation model (FIG. 1) calculates the feature corresponding area vector S(m i+1 ) is generated.

[0045] <Area feature estimation unit> As shown in FIG. 2, the area feature estimation unit 21 (A1) “Time-series location data” {X t A plurality of mesh code data {m t The "time series building statistical vector data" {x_m t |t=0, 1, , i} into the trained region feature estimator 21 a; (A2) The output of this area feature estimator 21a is the "time-series building statistical vector data" {x_m t |t=0, 1, . . . , i}, in this embodiment, the building statistical vector x_mi+1 , and remove it.

[0046] Here, the building statistical vector x_m, which is the area feature in this embodiment, is the number of buildings present in mesh m, and is a vector whose vector elements are the number of buildings belonging to each classification (when classified by type, use, and size). If the number of building classifications is C, the building statistical vector is a C-dimensional vector, and is generated as a fixed-length vector by setting a predetermined number of building classifications in advance. For example, the "building statistical vector of Mesh X" shown in Figure 4 (used in the explanation below) is a four-dimensional vector whose vector elements are the number of buildings (5, 3, 2, 0) in each classification of four (C = 4) types of buildings (private housing, commercial facilities, office buildings, and stations) set in advance.

[0047] The "time-series building statistical vector data" {x_m t |t=0, 1, ··, i} can also be regarded as a matrix (tensor) of dimension C×(i+1), as shown in Figure 2.

[0048] Furthermore, the area feature estimator 21a of this embodiment: (a) Time-series building statistical vector (area feature) data corresponding to the acquired time-series location data; (b) The correct answer data of the building statistical vector (area feature) that was obtained after that (in this embodiment, immediately after that) It is an estimator for time series data trained with training data containing multiple (large number) sets of

[0049] Specifically, as the area feature estimator 21a, for example, Non-Patent Document 3: Zeng, Ailing, et al. "Are Transformers Effective for Time Series Forecasting?", arXiv preprint arXiv:2205.13504v1, 2022<https: / / doi.org / 10.48550 / arXiv.2205.13504> Disclosed in (a) Transformer encoder and decoder Autoformer, Informer, FEDformer, etc., built using (b) DLinear, NLinear, etc., which are constructed using a fully-connected neural network algorithm. can be adopted.

[0050] The Transformer Encoder is the encoding part of the Transformer, which has been widely used in the field of natural language processing since its proposal in 2017. It consists of multiple modules, each of which includes a multi-head self-attention layer and a fully connected neural network layer. Autoformer, Informer, and FEDformer (a) above are models that have been devised to reduce the complexity of the multi-head self-attention layer when handling time-series data.

[0051] DLinear (b) above is a model that decomposes the input time series data into trend and seasonal components and processes each component individually with a single-layer linear neural network.NLinear, on the other hand, normalizes the input time series data by subtracting the final value of the input from the time series data (normalization in the time direction), and then processes it with a single-layer linear neural network.

[0052] In any case, the building statistical vector x_m i+1 is the result of estimating the building environment that the user will experience at time t=i+1 from the (mesh) building environment that the user has experienced between time t=0 and time t=i. In this embodiment, this building statistical vector x_mi+1 is normalized so that the norm is 1 and output in preparation for the subsequent inner product process (cos similarity calculation process).

[0053] <Position information determination section> As also shown in FIG. 1, the position information determination unit 22 of this embodiment (B1) The building statistical vector x_m generated by the area feature estimation unit 21 i+1 The normalized vector of the (future area feature) and the normalized vector of the building statistics vector x_m(j) (j=1, 2, . . . , N (N is the number of meshes)) of each mesh extracted from the updatable map (in Figure 2, an N × C dimensional matrix (tensor)) are taken as the inner product to obtain the N-dimensional feature correspondence area vector S(m i+1 ), Furthermore, the unit area estimation unit 221 (of the location information determination unit 22) (B2) "Time-series location data" {X t The "time-series mesh code data" {m t |t=0, 1, . . . , i}, the unit area estimator 221a and the decoder 221b are used to generate "time-series mesh code data" {m t |t=0, 1, . . . , i}, in this embodiment, the mesh m i+1 The history corresponding area vector P(m i+1 ), Furthermore, the position information determination unit 22 (B3) The generated feature corresponding area vector S(m i+1 ) and the generated history corresponding area vector P(m i+1 ), a position estimator 22a is used to generate an N-dimensional area estimation vector, and from this area estimation vector, the subsequent position information of the object of position estimation (the user of the GPS-equipped terminal), in this embodiment, the mesh m estimated at time t=i+1, is calculated. ^ i+1 Determine.

[0054] Here, the feature corresponding area vector S(m i+1 ), each of the N vector elements is a normalized vector of the building statistics vector x_m(j) (j=1, 2, . . . , N) for the corresponding mesh (out of N meshes) and a vector of the building statistics vector x_m i+1 The inner product value (cos similarity) of the normalized vector of the feature corresponding area vector S(m i+1 ) is information that represents the degree of similarity between the estimated building statistical vector (as the building statistical vector at time t=i+1) and the building statistical vector of each mesh, and is information that indicates which mesh's building statistical vector is most similar to the estimated building statistical vector.

[0055] Incidentally, in the location estimation model 1 (Fig. 1) described above, the mesh with the largest inner product value (cos similarity), that is, the feature corresponding area vector S(m i+1 ) is the mesh corresponding to the vector element with the maximum value at time t=i+1. ^ i+1 It can be said that:

[0056] Here, an example of inner product processing (cosine similarity calculation processing) in the position information determination unit 22 will be described. Fig. 3 is a histogram showing the calculation results of the cosine similarity (inner product value) in this example.

[0057] In this example, a large number of time-series mesh code data were generated every 30 minutes over eight days from a large number of time-series location data acquired from users of GPS-equipped devices whose locations were to be estimated. A portion of this data was used as training data to construct a DLinear-based area feature estimator 21a. The mesh code used was a quarter-area mesh code of approximately 250 m x 250 m. Buildings were classified by type (individual house, apartment building, business, complex, etc.), purpose (industry) (e.g., food and beverage, retail, education, medical, etc.), and size (number of floors, floor area, etc.). Using commercially available map information, a building statistical vector was generated for each mesh, with the number of buildings in each category representing the number of buildings present within that mesh as vector elements.

[0058] Next, a portion of the generated mesh code data over a period of eight days was used as evaluation data, and the corresponding building statistical vectors over a period of seven days were input into the constructed area feature estimator 21a, and the cosine similarity (inner product value) between the building statistical vector (estimated building statistical vector) obtained as the output and the corresponding building statistical vector for the remaining one day (correct building statistical vector) was calculated.

[0059] As a result, as shown in the histogram in Figure 3, it was found that many of the obtained estimated building statistical vectors were highly similar to the correct building statistical vector. Here, the median value of the calculated cos similarity reached 0.71. Therefore, it can be understood that the feature corresponding area vector (feature corresponding area information) with such cos similarity (inner product value) as a vector component is suitable information for use in subsequent position estimation.

[0060] Returning to Figure 2, the unit area estimator 221a in (B2) above is a machine learning model trained using learning data including time series data of mesh codes generated based on positioning results obtained from mobile terminals with GPS positioning functions owned by many users, and correct answer data of the mesh codes that actually occurred after that (in this embodiment, immediately after that).

[0061] This unit area estimator 221a can be configured to perform autoregression using multiple modules including a self-attention layer and a fully connected neural network layer in a transformer (for example, using a transformer encoder or a transformer decoder). In natural language processing, autoregression refers to a process in which the probability of a next word occurring is determined based on previously occurring words. Specific examples of such unit area estimator 221a include BERT, GPT-2 used in Non-Patent Document 2, and GPT-3. Of these, GPT-2 and GPT-3 are language models configured with a transformer decoder. GPT-3 in particular has a large-scale structure in which 96 modules including the self-attention layer and fully connected neural network layer are stacked as decoder modules.

[0062] Incidentally, the unit area estimator 221a can also be constructed using an RNN (Recurrent neural network) algorithm such as LSTM (Long Short Term Memory). However, by using a transformer-based estimator as the unit area estimator 221a as in this embodiment, it becomes easier to handle longer-term contexts (location histories) with high accuracy.

[0063] In any case, the unit area estimator 221a calculates the "time series position data" {X tUsing input |t=0, 1, . . . , i}, it outputs an internal state vector as a discrete representation of the estimated position at time t=i+1, i.e., a (fixed-length) history-corresponding area hidden vector with a preset hidden vector dimension. Incidentally, in the case of GPT-2, the history-corresponding area hidden vector for t=0, 1, . . . , i, i+1 is output by default, so the history-corresponding area hidden vector corresponding to time t=i+1 can be extracted from this and output. In this case, to reflect the internal state at time t=0, 1, . . . , i in the estimation result, it is also possible to perform pooling on these history-corresponding area hidden vectors (from t=0 to i+1) and output the result as the history-corresponding area hidden vector for time t=i+1.

[0064] Furthermore, the decoder 221b in (B2) above can be constructed by, for example, a fully connected (single layer) neural network algorithm, and uses the history corresponding area hidden vector at time t=i+1 generated as above as input, and generates an N (N is the number of meshes) dimensional history corresponding area vector P(m i+1 ) is output. Here, this history corresponding area vector P(m i+1 ) is the vector element of the mesh m at time t=i+1. i+1 The likelihood is as follows.

[0065] Furthermore, the position estimator 22a in (B3) above may be constructed using, for example, a fully connected DNN (Deep Neural Network) algorithm. Specifically, the position estimator 22a calculates the feature corresponding area vector S(m i+1 ) and the historical corresponding area vector P(m i+1 ) and outputs an N-dimensional area estimation vector (N is the number of meshes). Here, each vector element of this area estimation vector is the mesh m at time t=i+1. i+1The likelihood of this is the likelihood that the building environment has been passed through up to now, and it also takes into account the building environment that has been passed through up to now. In other words, the area estimation vector is the likelihood of the next mesh (at time t=i+1) in each mesh that is estimated taking into account the building environment (similarity with) that is predicted from the building environment history up to now and the next mesh (likelihood that it is) that is predicted from the history of the previous mesh. i+1 This information expresses the possibility that this is the case.

[0066] Here, the location information determination unit 22 determines the mesh corresponding to the vector element (likelihood) having the maximum value in the area estimation vector generated in this way as the mesh m estimated at time t=i+1. ^ i+1 It can be said that:

[0067] Next, we will explain the training of the above-mentioned location estimation model 2. As is clear from the above explanation, the location estimation model 2 has many parameters to be trained, and depending on the design, it is highly likely that training in so-called End2End (collectively from the input terminal to the output terminal) will be extremely difficult.

[0068] Therefore, pre-training may be performed individually on the estimators and decoders (21a, 221a, and 221b, 22a) to be used. For example, it is preferable to pre-train the unit area estimator 221a and decoder 221b employing GPT-2 using the method disclosed in Non-Patent Document 2, and further pre-train the area feature estimator 21a using the method disclosed in Non-Patent Document 3 (Zeng, Ailing, et al., "Are Transformers Effective for Time Series Forecasting?"). In this case, the training process (model construction process) can be completed by finally fine-tuning the entire location estimation model 2 (neural network). Note that training for the previously described location estimation model 1 (FIG. 1), and the later-described location estimation model 3 (FIG. 5), location estimation model 4 (FIG. 6), and location estimation model 5 (FIG. 7) can also be performed in a similar manner to the above-described method.

[0069] Furthermore, in the model training described above, the building statistical vectors of the relevant meshes are naturally used as learning data. However, even if the building statistical vectors and, in turn, the updatable map are updated in response to changes (increases or decreases) in the number of buildings in the survey area or changes in their type or size, the already constructed location estimation model 2 (and 1, 3 to 5) does not need to be retrained. In other words, without conducting new training using new learning data related to the updated building statistical vectors (updatable maps), the location information determination unit 22 can simply handle the updated updatable map (building statistical vectors) to perform location estimation processing. This makes it possible to quickly estimate the impact of changes in the building environment on, for example, the future location of a target person and, in turn, future pedestrian flow, without relying on retraining, which involves a very large processing load.

[0070] Incidentally, the fact that location estimation model 2 (and 1, 3 to 5) does not need to be retrained even when the updatable map is updated means that location estimation model 2 (and 1, 3 to 5) is designed under the assumption that "the location with the building distribution that the surveyed user will pass through is determined by the building environment that the user has passed through up to that point." In other words, even if the building environment changes (even if the updatable map is updated), the purpose (meaning) of the user's movement is maintained for the user, and the user's movement pattern (movement intention), such as "I tend to move to a commercial area after going to an office district," is assumed to remain unchanged.

[0071] 4 is a schematic diagram for explaining the policy and procedure of the location estimation process in location estimation model 2 according to the present invention. The policy and procedure explained here also generally apply to location estimation model 3 (FIG. 5), location estimation model 4 (FIG. 6), and location estimation model 5 (FIG. 7), which will be described later.

[0072] According to the example shown in FIG. 4, the location estimation model 2 (FIG. 2) (a) User movement trajectory (time-series location information) acquired by observation by a base station of device 2, a smartphone equipped with GPS positioning function: Mesh 1 → Mesh 2 → Mesh 3 → Mesh 4 → Mesh 5 The unit area estimator 221a (Fig. 2) takes in the above data and outputs a "discrete representation of the user's movement trajectory, including a discrete representation of the mesh at the next time point," from which "information relating to the mesh at the next time point estimated from the movement trajectory" (history-corresponding area information) is generated.

[0073] In addition, location estimation model 2 is (b) Building statistical vectors for each mesh generated from the building distribution (land use type) information within each mesh. The area feature estimator 21a takes in the data and generates a time-series building statistical vector (residential area → station → office → commercial area → station) corresponding to the user's movement trajectory (Mesh 1 → Mesh 2 → Mesh 3 → Mesh 4 → Mesh 5). This is input to the area feature estimator 21a, which outputs the "building statistical vector at the next time point," and generates from this "information relating to the mesh at the next time point estimated from the building environment that has been passed through" (feature-corresponding area information).

[0074] Finally, location estimation model 2 estimates the mesh for the next time point: Mesh 6, taking into consideration both the generated "information relating to the mesh for the next time point estimated from the movement trajectory" (history corresponding area information) and the generated "information relating to the mesh for the next time point estimated from the building environment that has been passed through" (feature corresponding area information).

[0075] In this way, location estimation model 2 can estimate the user's location more accurately by considering not only the user's previous movement trajectory but also the building environment that the user has passed through. In other words, it is possible to estimate the user's location more accurately by considering not only the "spatial constraints" of the user's location but also the "semantic constraints" of the user's location from the user's perspective. For example, it is possible to estimate the user's location by incorporating the user's purpose (intention), such as "going to a commercial area after the office (to do some shopping after work)."

[0076] Other embodiments of the location estimation model FIG. 5 is a schematic diagram showing yet another embodiment (location estimation model 3) of the location estimation model according to the present invention.

[0077] The position estimation model 3 of this embodiment shown in Fig. 5 is a model obtained by partially modifying the above-mentioned position estimation model 2 (Fig. 2). Specifically, the position estimation model 2 integrates an N-dimensional (N is the number of meshes) feature corresponding area vector and an N-dimensional history corresponding area vector to estimate the mesh m ^ i+1In contrast to this, the location estimation model 3 of this embodiment integrates the feature corresponding area hidden vector of the "hidden vector dimension" and the history corresponding area hidden vector of the "hidden vector dimension" to determine and output the estimated mesh m ^ i+1 Here, the "hidden vector dimension" refers to the dimension of the internal state vector (hidden vector) of the unit area estimator (321a).

[0078] As also shown in FIG. 5, the location estimation model 3 of this embodiment is (A) an area feature estimation unit 31 that performs estimation processing of a building statistical vector (area feature amount) using an area feature estimator 31a; (B) a location information determination unit 32 having a unit area estimation unit 321 that performs estimation processing of a mesh (unit area) using a unit area estimator 321a; These are functional components that are realized by executing a program installed in the computer.

[0079] Of these, the area feature estimation unit 31 and area feature estimator 31a in (A) above have the same functional configuration as the area feature estimation unit 21 and area feature estimator 31a in the location estimation model 2 (FIG. 2) described above. Also, the unit area estimator 321a in (B) above has the same functional configuration as the unit area estimator 221a in the location estimation model 2 (FIG. 2). Meanwhile, unlike the unit area estimation unit 221 (FIG. 2), the unit area estimation unit 321 in (B) above does not have a decoder, and generates and outputs a history-corresponding area hidden vector (history-corresponding area information) in the hidden vector dimension.

[0080] Furthermore, the unit area estimator 321a of this embodiment (as well as the unit area estimator 221a of the embodiment shown in FIG. 2) has an embedding unit 321ae trained to convert input data into an embedded representation (distributed representation) vector. Here, the location information determination unit 32 of this embodiment (B1) For the updatable map (area feature map information), i.e., for the column vectors (when the building statistics vector is a row vector) in the updatable map, an embedding process is performed using the same parameters as the embedding unit 321ae to generate an embedded updatable map of (the number of hidden vector dimensions) × C dimensions; (B2) Building statistical vectors (row vectors) for each embedded representation of a mesh in the embedded updateable map (area feature map information) and the building statistical vector x_m estimated by the area feature estimation unit 31 i+1 (Subsequent area feature) (in this embodiment, the normalized inner product of the two is taken) to generate a hidden vector-dimensional feature corresponding area hidden vector (feature corresponding area information), (B3) Based on the generated feature-corresponding area hidden vector and the history-corresponding area hidden vector generated by the unit area estimation unit 321, a location estimator 32a constructed by, for example, a fully connected DNN algorithm is used to generate an area estimation vector, and the estimated mesh m ^ i+1 Determine and output.

[0081] In this embodiment, the feature corresponding area hidden vector (feature corresponding area information) in (B2) is the building statistical vector x_m i+1 The embedding parameters are information about the mesh corresponding to the area feature (later). The embedding parameters obtained by pre-training the unit area estimator (e.g., GPT-2 or GPT-3) 321a can be used as the "same parameters as the embedding unit 321ae" in (B2) above. The embedded updateable map obtained using such embedding parameters can be considered as a (compressed) projection of the updateable map onto the internal state space of the unit area estimator (e.g., GPT-2 or GPT-3) 321a.

[0082] As described above, in the position estimation model 3, the feature corresponding area information (feature corresponding area hidden vector) and history corresponding area information (hidden vector of history corresponding area) handled as input to the position estimator 32a are hidden vector-dimensional information (e.g., approximately 500 dimensions) that is fewer than N dimensions (e.g., approximately 50,000 dimensions) (N is the number of meshes). In contrast, the position estimation model 2 (FIG. 2) handles N-dimensional feature corresponding area information (feature corresponding area vector) and history corresponding area information (history corresponding area vector) as input to the position estimator 22a. Therefore, according to the position estimation model 3, by reducing the number of dimensions of the handled vectors, the burden of calculation processing within the model, particularly in the position estimator 32a, can be further reduced, and further, higher generalization performance can be achieved.

[0083] FIG. 6 is a schematic diagram showing yet another embodiment (location estimation model 4) of the location estimation model according to the present invention.

[0084] The position estimation model 4 of this embodiment shown in Fig. 6 is a model obtained by partially modifying the above-mentioned position estimation model 3 (Fig. 5). Specifically, the position information determination unit 42 of the position estimation model 4 replaces the position estimator (32a) employed in the position information determination unit 32 (Fig. 5) with (a) an attention unit 42a constructed with an attention mechanism; (b) a decoder 42b constructed by, for example, a fully connected (one-layer) neural network algorithm; It is equipped with:

[0085] The attention mechanism is a mechanism that determines which part of the input data should receive particular attention, and is currently a widely used technology in the field of machine learning. For details about this attention mechanism, see, for example, the following non-patent document: Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, AN, Kaiser, L. ukasz, & Polosukhin, I., "Attention is All You Need," Advances in Neural Information Processing Systems (Vol. 30), 2017,<https: / / proceedings.neurips.cc / paper / 2017 / file / 3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf> This is explained in detail in

[0086] The position information determination unit 42 uses the attention unit 42a and the decoder 42b to determine the estimated mesh m based on the feature-corresponding area hidden vector and the history-corresponding area hidden vector. ^ i+1 The area feature estimation unit 41 and the area feature estimator 41a of the location estimation model 4 have the same functional configuration as the area feature estimation unit 31 and the area feature estimator 31a of the location estimation model 3 (FIG. 5) described above. The unit area estimator 321a also has the same functional configuration as the unit area estimator 321a of the location estimation model 3 (FIG. 5). Furthermore, the location information determination unit 42, other than the attention unit 42a and decoder 42b described above, has the same functional configuration as the parts other than the location estimator 32a of the location estimation model 3 (FIG. 5).

[0087] As also shown in FIG. 6, the attention unit 42a of this embodiment functions as a first weighting mixer, a scaling unit, a softmax function unit, and a second weighting mixer. First, the first weighting mixer inputs the feature-corresponding area hidden vector and the history-corresponding area hidden vector as Query and Key, respectively, and outputs a function value vector (additive attention function) related to the weighted sum of the two, or a function value vector (multiplicative attention function) related to the weighted product of the two. Here, this weighting can be interpreted as representing the importance of each vector element in the history-corresponding area hidden vector (Key) relative to the feature-corresponding area hidden vector (Query). Next, the scaling unit normalizes the output vector from the first weighting mixer using the number of dimensions of the history-corresponding area hidden vector (Key).

[0088] Then, the softmax function unit applies a known softmax function to maximize the maximum value of the vector elements in this normalized output vector, and outputs a vector whose elements sum to 1. Finally, the second weighting mixer unit outputs a function value vector related to the weighted sum of the output vector from the softmax function unit and the history-corresponding area hidden vector input as Value, or a function value vector related to the weighted product of both. Here, this output vector is a new feature-corresponding area hidden vector that takes into account the importance of each vector element in the history-corresponding area hidden vector (Value) relative to the feature-corresponding area hidden vector (Query).

[0089] In this embodiment, the decoder 42b receives this output vector as input and outputs an N-dimensional area estimation vector (N is the number of meshes). The attention unit 42a can also perform attention processing using the history-corresponding area hidden vector as the query and the feature-corresponding area hidden vector as the memory (key, value), which is the opposite of the input contents described above.

[0090] In this way, according to the attention unit 42a, (a) Feature-corresponding area hidden vector data derived by taking into account the "semantic constraints" of the user position; (b) History-corresponding area hidden vector data derived taking into account the "spatial constraints" of the user's location and Rather than simply integrating these data using a DNN, by determining which data parts to focus on for each data set based on the other data, it becomes possible to generate mesh (unit area) estimation data with higher estimation accuracy by weighting those data parts.

[0091] The attention mechanism described above can also be used in place of the dot product processing in the position information determination unit 42. Specifically, for example, (a) The building statistical vector x_m output from the area feature estimation unit 41 i+1 The normalized vector of the above is used as the query, and each normalized vector of the "building statistical vector with several hidden vector dimensions" extracted (as each row vector) from the embedded updateable map is used as memory (key, value), and attention processing is performed. (b) The results of the attention process for each normalized vector of the “building statistical vector with the number of hidden vector dimensions” are averaged or pooled using the maximum value. (c) The result of the averaging or pooling process may be input to a predetermined decoder, and a feature corresponding area hidden vector (feature corresponding area information) may be generated as the output of this decoder.

[0092] In addition, in the above-mentioned location estimation model 1 (FIG. 1), location estimation model 2 (FIG. 2), location estimation model 3 (FIG. 5), and also in the location estimation model 5 (FIG. 7) described later, it is possible to generate feature corresponding area information by applying the attention processing described above instead of the dot product processing.

[0093] FIG. 7 is a schematic diagram showing yet another embodiment (location estimation model 5) of the location estimation model according to the present invention.

[0094] The location estimation model 5 of this embodiment shown in Fig. 7 is a model obtained by partially modifying the location estimation model 2 (Fig. 2) described above. Specifically, the location information determination unit 22 (Fig. 2) of the location estimation model 2 integrates the feature corresponding area vector and the history corresponding area vector using a location estimator (22a), whereas the location information determination unit 52 of the location estimation model 5 calculates the Hadamard product of the feature corresponding area vector and the history corresponding area vector, and applies a softmax function to the result to generate an area estimation vector.

[0095] As a result, the location estimation model 5 can perform the area estimation vector generation process (location estimation process) with a significantly smaller processing load than, for example, the location estimator 22a (Figure 2), which is a fully connected DNN.

[0096] The area feature estimation unit 51 and the area feature estimator 51a in the location estimation model 5 of this embodiment have the same functional configuration as the area feature estimation unit 21 and the area feature estimator 21a in the location estimation model 2 (FIG. 2) described above. The unit area estimator 521a and the decoder 521b also have the same functional configuration as the unit area estimator 221a and the decoder 221b in the location estimation model 2 (FIG. 2). Furthermore, the location information determination unit 52, other than the parts that perform the Hadamard product application process and the softmax function application process, also has the same functional configuration as the parts other than the location estimator 22a in the location estimation model 2 (FIG. 2).

[0097] [Route prediction device / program] FIG. 8 is a functional block diagram showing the functional configuration of an embodiment of a route prediction device according to the present invention.

[0098] As shown in FIG. 8, the route prediction device 8 of this embodiment (a) an area building information storage unit 802 that stores and manages area building information (including map information) that is information related to buildings within a target area, acquired from an external server, such as a map data management server, via a communication interface unit 801 and a communication control unit 821; (b) an updatable map generating and updating unit 811 that generates and appropriately updates building statistical vectors for each mesh and, in turn, updatable maps using the area building information read from the area building information storage unit 802; (c) a terminal location information storage unit 803 that stores and manages terminal location information acquired via a communication interface unit 801 and a communication control unit 821 from an external mobile data management server that collects and manages (via a base station 3) GPS positioning data of (a user who owns) a terminal 2 equipped with a GPS positioning function; (d) a time-series location information generating unit 812 that generates time-series location information (time-series mesh code data) for learning and location estimation (of a location estimation target) using the terminal location information read from the terminal location information storage unit 803; (e) a model training unit 813 that performs training (pre-training, fine-tuning) using the generated time-series location information for learning and the building statistical vector of the corresponding mesh, and constructs one of the location estimation models 1 to 5; (f) a route prediction unit 814 that determines the next time point's location information (mesh code) (of the location estimation target) using the constructed location estimation model (1, 2, 3, 4, 5) based on the generated time-series location information for location estimation, and further determines the next time point's location information based on the time-series location information including the determined location information, and repeats this process a predetermined number of times to predict the future movement route of the location estimation target, i.e., generates future movement route information (time-series mesh code data); (g) an input / output control unit 822 that displays future travel route information determined or stored and managed in the route information storage unit 804 on a user interface (UI) unit 805 equipped with, for example, a keyboard and a display, in accordance with instructions from the user interface unit 805; The generated future travel route information may be transmitted to an external information processing device via the communication control unit 821 and the communication interface unit 801 and used therein.

[0099] Incidentally, the updatable map generation / update unit 811 (b), the time-series location information generation unit 812 (d), the model training unit 813 (e), the route prediction unit 814 (f), and the input / output control unit 822 (g) are main functional components that implement an embodiment of the route prediction method according to the present invention, and can also be considered to be functions realized by executing a route prediction program in a processor memory that stores an embodiment of the route prediction program according to the present invention. Therefore, the route prediction device 8 may be a device dedicated to route prediction, or it may also be, for example, a cloud server, a non-cloud server device, a personal computer (PC), a notebook or tablet computer, a smartphone, or the like, that is equipped with the route prediction program according to the present invention.

[0100] In any case, the route prediction device 8 can determine the future movement route of the location estimation target by taking into consideration the building statistical vector of each mesh (unit area) that is a characteristic of the target area, i.e., the updatable map (area characteristic map information). For example, by appropriately setting the building statistical vector, it becomes possible to predict with high accuracy the movement route of the location estimation target, taking into consideration the purpose, meaning and movement pattern of the movement.

[0101] As described above in detail, according to the present invention, by using a location estimation model, it is possible to determine the future location information of an object (e.g., the location information at the next time point) by taking into account the area feature amount (e.g., building statistical vector) of each unit area (e.g., mesh), which is a feature within the area, that is, to estimate the future location of the object. Here, even if the area feature amount is updated (e.g., the building environment has changed), it is possible to quickly obtain a new location estimation result that reflects this update without having to retrain the location estimation model.

[0102] Furthermore, the route prediction device and program according to the present invention can determine the target's future travel route information, i.e., estimate the target's future travel route, by taking into consideration the area feature quantities (e.g., building statistical vectors) of each unit area (e.g., mesh) that are characteristics within the area.

[0103] For example, by adopting a building statistical vector appropriately set as an area feature, it is possible to estimate and understand more detailed travel demand according to the building environment of the area where the user lives and the travel pattern (including the user's lifestyle) that reflects the user's purpose. Furthermore, by utilizing such detailed travel demand, it is possible to improve the quality and efficiency of services, for example, by realizing optimal vehicle dispatching and waiting vehicle distribution at all times in dynamically provided transportation services such as on-demand taxis and ride-sharing taxis.

[0104] Furthermore, for example, the route prediction method according to the present invention can be used to accurately and precisely predict the movement routes of automobiles, pedestrians, etc. in urban areas, and ultimately predict their behavior, thereby enabling the development of resilient infrastructure for roads, sidewalks, intersections, and passageways within facilities, as well as the implementation of efficient urban planning to solve traffic and people flow problems caused by urbanization. In other words, the present invention can contribute to Goal 9 "Build resilient infrastructure, promote sustainable industrialization and foster innovation" and Goal 11 "Make cities inclusive, safe, resilient and sustainable" of the Sustainable Development Goals (SDGs) led by the United Nations.

[0105] Those skilled in the art can easily make various changes, modifications, and omissions to the various embodiments of the present invention described above within the scope of the technical idea and perspective of the present invention. The above description is merely an example of an embodiment and does not impose unnecessary restrictions. The present invention is limited only by the scope of the claims and their equivalents. [Explanation of symbols]

[0106] 1, 2, 3, 4, 5 Location estimation model 11, 21, 31, 41, 51 Area feature estimation section 11a, 21a, 31a, 41a, 51a Area feature estimators 12, 22, 32, 42, 52 Location information determination unit 22a, 32a position estimator 221, 321, 421, 521 Unit Area Estimation Section 221a, 321a, 421a, 521a Unit Area Estimator 221b, 42b, 521b decoder 42a Attention section 2. Terminal 3 base station 8. Route Prediction Device 801 Communication interface unit 802 Area Building Information Storage Department 803 Terminal location information storage unit 804 Route Information Storage Unit 805 User Interface (UI) Department 811 Updateable map generation and update unit 812 Time-series location information generation unit 813 Model Training Department 814 Route Prediction Unit 821 Communication Control Unit 822 Input / Output Control Unit

Claims

1. A location estimation model that estimates future location information of an object from time-series location information relating to the position of the object in a certain area, A zone feature value representing a feature of each unit zone within the area is set, a zone feature estimation unit that generates zone features that are subsequent to the zone features in the time series, based on zone features in the time series that are zone features related to a plurality of unit zones corresponding to each of the time series location information, using a zone feature estimator trained with learning data including time series data of the zone features and data of subsequent zone features; a location information determination unit that generates feature corresponding area information, which is information related to the unit area corresponding to the subsequent area feature amount, based on the subsequent area feature amount and each of the area feature amounts related to the unit area, and determines subsequent location information of the object based on the feature corresponding area information; A location estimation model characterized by causing a computer to function as follows.

2. The position information determination unit a unit area estimation unit that generates, based on the time-series unit areas corresponding to each of the time-series location information, history-corresponding area information, which is information related to a unit area that comes after the unit area in the time series, by using a unit area estimator trained with learning data including time-series data of the unit area and data of the unit area that comes after it; and determining subsequent location information of the object based on the feature corresponding area information and the history corresponding area information. The location estimation model of claim 1 .

3. the unit area estimator has an embedding unit trained to convert input data into an embedding representation vector; the location information determination unit performs an embedding process on area feature map information consisting of area feature values ​​of each unit area within the area, using the same parameters as those of the embedding unit, and generates the feature corresponding area information using the area feature values ​​associated with each embedded expression associated with the unit area in the area feature map information that has been subjected to the embedding process and the area feature values ​​to be used; The generated feature-corresponding area information is information on embedded expressions related to the unit area corresponding to the subsequent area feature. The location estimation model of claim 2 .

4. 4. The localization model according to claim 2 or 3, wherein the unit area estimator is constructed to perform autoregression using a plurality of modules including a self-attention layer and a fully-connected neural network layer in a transformer.

5. The location estimation model according to claim 2 or 3, characterized in that the location information determination unit determines subsequent location information of the object based on the feature corresponding area information and the history corresponding area information using a neural network algorithm, an attention mechanism, or a Hadamard product.

6. 4. The localization model according to claim 1, wherein the area feature estimator is constructed using a transformer encoder and decoder or a fully connected neural network algorithm.

7. The location estimation model according to any one of claims 1 to 3, characterized in that the location information determination unit generates the feature corresponding area information using an inner product or an attention mechanism based on the subsequent area feature and each of the area feature values ​​related to the unit area.

8. A location estimation model according to any one of claims 1 to 3, characterized in that the area features include quantities related to things that exist or occur within the corresponding unit area and that may affect the movement or stay of the subject.

9. The location estimation model according to claim 8, wherein the area feature includes a quantity related to at least one of the type, use, size, and number of buildings or structures present in the corresponding unit area.

10. A location estimation model according to any one of claims 1 to 3, characterized in that even if area feature map information consisting of area feature values ​​for each unit area within the area is changed, subsequent location information of the target can be determined without retraining the estimator used.

11. 4. A route prediction program that causes a computer to function as route prediction means for predicting a route including subsequent position information of an object in a certain area, using the position estimation model according to claim 1, from time-series position information relating to the position of the object in the certain area.

12. 4. A route prediction device comprising: a route prediction means for predicting a route including subsequent position information of an object in a certain area, using the position estimation model according to claim 1, based on time-series position information relating to the position of the object in the certain area.

13. A position estimation method for estimating future position information of an object from time-series position information relating to the position of the object in a certain area, comprising: A zone feature value representing a feature of each unit zone within the area is set, generating, based on the time-series area features that are area features related to a plurality of unit areas corresponding to each of the time-series location information, area features that are subsequent to the time-series area features, using an area feature estimator trained with learning data including time-series data of the area features and data of subsequent area features; generating feature-corresponding area information, which is information related to the unit area corresponding to the subsequent area feature, based on the subsequent area feature and each of the area feature amounts related to the unit area, and determining subsequent position information of the object based on the feature-corresponding area information; 1. A computer-implemented method for location estimation, comprising:

Images