Movement / stay estimation model and method incorporating area characteristics, and movement trajectory prediction device and program

Abstract

To provide a movement and stay estimation model which can estimate information related to movements and stays of an estimation target, the information thus also enabling prediction of a route, with consideration for an area condition.SOLUTION: The model causes a computer to function as: an area feature and position encoding unit into which area feature information related to things and events which can influence or relate to movements and stays of an estimation target in a research target area, and information related to a time series of the movements or the stays of the estimation target are input and which produces area feature embedded information having the area feature information embedded therein and position embedded information having information related to a position in the time series embedded therein; an area movement and stay encoding unit into which the area feature embedded information and the position embedded information are input and which produces area movement and stay embedded information having information related to the movements and the stays of the estimation target in the research target area embedded therein; and a movement and stay estimation unit into which the area movement and stay embedded information is input and which produces information related to posterior movements and stays.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a technique for estimating the manner in which an object (for example, a person) stays or moves using sensing data. [Background technology]

[0002] In recent years, in the field of transportation policy, the Activity Based Model (ABM), which estimates human activities and predicts the resulting travel behavior of those activities, has been attracting attention. Specifically, ABM is a travel behavior prediction model that simulates the activities of individual people based on human attributes such as place of residence, age, gender, and occupation, as well as traffic conditions and regional characteristics, and then aggregates and predicts the overall traffic volume that may result from these.

[0003] For example, the transportation demand forecasting technology disclosed in Non-Patent Document 1 uses this type of ABM to forecast not only the transportation demand that already exists, but also the transportation demand that will arise only when new facilities or transportation systems are built.

[0004] Here, ABM training requires learning data that includes activity details, and this activity details are obtained from survey results such as the Ministry of Land, Infrastructure, Transport and Tourism's Person Trip (PT) Survey, which collects detailed transportation data (trip data) such as the starting point, destination, purpose, means of transportation used, and time of activity through questionnaires.

[0005] Therefore, such surveys are essential for building ABM. For example, the reality is that ABM cannot be built using only location measurement information from communication devices such as smartphones, which are expected to be used for traffic demand forecasting.

[0006] In response to this, Non-Patent Document 2 discloses a location prediction technology that applies BERT (Bidirectional Encoder Representations from Transformers), a well-known natural language processing model. In this technology, BERT is pre-trained with a large amount of movement trajectory data to extract a multidimensional vector containing the meaning of a user's movement trajectory, and this multidimensional vector is used to predict the user's future location. Here, the pre-training uses a task of estimating masked positions and times in a time series of user positions, instead of a word fill-in-the-blank problem in natural language processing.

[0007] Incidentally, the above-mentioned BERT is explained in detail in Non-Patent Document 3. [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] Ryo Kanamori, Takayuki Morikawa, Toshiyuki Yamamoto, and Tomio Miwa, "Development of an Integrated Transportation Demand Forecasting Model for the Formulation of a Comprehensive Transportation Strategy," Journal of the Japan Society of Civil Engineers, Vol. 65, No. 4, pp. 503-518, 2009 [Non-patent document 2] Lin, Y., Wan, H., Guo, S., & Lin, Y. “Pre-training Context and Time Aware Location Embeddings from Spatial-Temporal Trajectories for User Next Location Prediction”, Proceedings of the AAAI Conference on Artificial Intelligence, 35(5), pp. 4241-4248. 2021,<https: / / ojs.aaai.org / index.php / AAAI / article / view / 16548> [Non-patent document 3] Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova, “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding”, NAACL-HLT (1) 2019: 4171-4186, 2019,<https: / / aclanthology.org / N19-1423.pdf> Summary of the Invention [Problem to be solved by the invention]

[0009] However, conventional location prediction techniques that do not rely on the results of questionnaire surveys, such as those disclosed in Non-Patent Document 2, are difficult to apply to transportation demand prediction for transportation policies, for example.

[0010] Specifically, the location prediction model that applies BERT in Non-Patent Document 2 is constructed by learning only the time-series relationships between locations, such as, for example, if you have traveled from point A to point B, it is likely that you will travel to point C at the next point. In other words, the basic premise of location prediction is that the intra-area conditions of the prediction target area, such as the distribution of transportation means and the locations of various facilities, remain the same. Therefore, it is impossible to predict the location of traffic congestion when these intra-area conditions, which are important in traffic demand prediction, change, for example, the location of traffic congestion that would occur if a hospital located at point C were to be relocated to point D.

[0011] Furthermore, the location prediction model applying BERT in Non-Patent Document 2 only uses time-series information on the user's location as input, so it is possible to predict the user's future destination, but it is not possible to predict the route the user will take to reach the destination. As a result, it is impossible to predict information about individual routes, such as the usage volume of each road or railway line, which is important in traffic demand prediction.

[0012] Therefore, an object of the present invention is to provide a movement / stay estimation model and method that can estimate information related to the movement or stay of an estimation target, which information also enables route prediction, while taking into account in-area conditions, and a movement trajectory prediction device and program that can predict the movement trajectory of the estimation target using the estimation results. [Means for solving the problem]

[0013] According to the present invention, there is provided a movement and stay estimation model that causes a computer to function to estimate information related to a subsequent movement or stay of a certain object in a certain area, the movement and stay estimation model comprising: an area feature position encoding unit that receives as input area feature information generated from information related to things or events that may affect or are related to the movement or stay of the subject in the area and information related to the time series of the movement or stay of the subject, and generates area feature embedded information in which the area feature information is embedded and position embedded information in which information related to the position in the time series is embedded, or generates area feature position embedded information in which information combining the area feature information and the information related to the position in the time series is embedded; The area feature embedding information and the location embedding information are input, or Area feature embedding information and location Embedded information and an area movement stay encoding unit that receives as input information combining the above or receives as input the area feature position embedding information, and generates area movement stay embedding information in which information related to the movement or stay of the target in the area is embedded; a movement and stay estimation unit that receives the area movement and stay embedding information as an input and generates information related to the subsequent movement or stay; A travel and stay estimation model is provided that allows a computer to function as a

[0014] Here, it is also preferable that the above-mentioned area characteristic information according to the present invention is generated from area thing and event information including at least one of information regarding the number, size or content of transportation or means of travel in each of the multiple sections that make up the area, information regarding the number, size or content of public facilities or public events, information regarding the number, size or content of commercial / service providing facilities or commercial / service providing events, information regarding the number, size or content of residential facilities, and information regarding the number, size or content of business facilities.

[0015] Furthermore, as one embodiment of the movement and stay estimation model according to the present invention, the movement and stay estimation model causes a computer to function as a movement and stay time point encoding unit that receives information related to a time series of the movement or stay of the target as input and generates time point embedded information in which information related to a time point in the time series is embedded; The area movement stay encoding unit receives the area feature embedding information, the location embedding information, and the time point embedding information as input, or The area feature embedding information, the location embedding information, and the time point Embedded information and Alternatively, it is also preferable to input information that combines the area feature position embedding information and the time point embedding information, or to input information that combines the area feature position embedding information and the time point embedding information.

[0016] Furthermore, in the movement and stay estimation model according to the present invention, the information relating to the time series of the movement or stay of the subject includes information relating to the positions and times of movement or stay during a certain period, and information relating to the positions and times of movement or stay during an estimation period that follows the certain period; It is also preferable that the movement / stay estimation unit generates, as the information relating to the subsequent movement or stay, information relating to the likelihood of information relating to the position and time of movement or stay during the estimation period.

[0017] Furthermore, as another embodiment of the movement and stay estimation model according to the present invention, the area feature location encoding unit has an area feature encoding unit that receives the area feature information as an input and generates area feature embedded information in which the area feature information is embedded, and a movement and stay location encoding unit that receives information related to a time series of the movement or stay and generates the location embedded information, The area movement stay encoding unit receives the area feature embedding information and the location embedding information as input, or Area feature embedding information and location Embedded information and It is also preferable to input information that combines the above.

[0018] In an embodiment including the area feature encoding unit, the area feature information is area graph information in which information relating to things or events that may affect or are related to the movement or stay of the subject in a target area among the multiple areas that make up the area is used as a node, and a matrix that represents the setting status of links between the target area and other surrounding areas is used as a graph adjacency matrix, It is also preferable that the area feature encoding unit includes a graph convolution network (GCN) that receives the area graph information as an input.

[0019] It is also preferable that the graph adjacency matrix has matrix elements that reflect the presence or absence of specified means of transportation, movement, distribution, or communication between the target area and other surrounding areas, or the traffic volume, movement volume, distribution volume, or communication volume per unit time.

[0020] Furthermore, as yet another embodiment of the movement and stay estimation model according to the present invention, it is also preferable that the area characteristic position encoding unit generates, from the received area characteristic information, a category value that represents a predetermined category and that expresses the characteristics of a target section among multiple sections that make up the area, and generates area characteristic position embedded information in which information combining the category value and information related to the position in the time series is embedded.

[0021] Furthermore, the movement and stay estimation model according to the present invention is based on information related to the time series of the movement or stay of the target having the masked area and the correct area. Features It is also preferable that the model is trained by a training process in which training data includes data including information on the movement and stay of a person, and the training process is performed with a loss equal to the sum of an error from a correct answer in the output from a masked location estimation unit that is provided for training together with the movement and stay estimation unit subsequent to the area movement and stay encoding unit and that solves the problem of guessing data for a masked location, and an error from a correct answer in the output from the movement and stay estimation unit.

[0022] According to the present invention, there is also provided an area feature generating means for generating the area feature information from information relating to things or events that may affect or are related to the movement or stay of the subject; a movement / stay time series information generating means for generating information relating to a time series of the movement or stay of the target from the acquired location information of the terminal; The above-mentioned travel and stay estimation model and, This travel and stay estimation model a movement / stay prediction means for determining information related to the subsequent movement or stay using the above information; a movement trajectory prediction means for predicting information relating to the movement trajectory of the target based on the information relating to the subsequent movement or stay determined sequentially; A trajectory prediction device is provided, comprising:

[0023] According to the present invention, there is further provided an area feature generating means for generating the area feature information from information relating to things or events that may affect or are related to the movement or stay of the subject; a movement / stay time series information generating means for generating information relating to a time series of the movement or stay of the target from the acquired location information of the terminal; The above-mentioned travel and stay estimation model The area characteristic information and information relating to the time series of the movement or stay of the target are provided as input information to the , information regarding subsequent movements or stays to A travel and stay prediction means for determining a movement trajectory prediction means for predicting information relating to the movement trajectory of the target based on the information relating to the subsequent movement or stay determined sequentially; have Movement trajectory prediction Device is provided.

[0024] According to the present invention, there is also provided a computer-implemented movement and stay estimation method for estimating information related to a subsequent movement or stay of an object in an area, the method comprising: a step of using area feature information generated from information related to things or events that may affect or are related to the movement or stay of the subject in the area and information related to the time series of the movement or stay of the subject as input, and generating area feature embedded information in which the area feature information is embedded and location embedded information in which information related to the location in the time series is embedded, or generating area feature location embedded information in which information combining the area feature information and the information related to the location in the time series is embedded; The area feature embedding information and the location embedding information are input, or Area feature embedding information and location Embedded information and a step of generating area movement / stay embedded information in which information related to the movement or stay of the target in the area is embedded, using information combining the above and the above or the area feature position embedded information as input; a step of generating information relating to the subsequent movement or stay using the area movement / stay embedding information as an input; A travel dwell estimation method is provided, which has: [Effects of the Invention]

[0025] According to the movement and stay estimation model and method of the present invention, it is possible to estimate information related to the movement or stay of an estimation target, which information also enables route prediction, while taking into account in-area conditions. Furthermore, according to the movement trajectory prediction device and program of the present invention, it is possible to predict the movement trajectory of the estimation target using the estimation results. [Brief explanation of the drawings]

[0026] [Figure 1]1 is a schematic diagram illustrating an embodiment of a movement / stay estimation model according to the present invention and an embodiment of a movement trajectory prediction device according to the present invention. [Figure 2] FIG. 2 is a schematic diagram for explaining one embodiment of movement stay time-series information (input to a movement stay position encoding unit) according to the present invention. [Figure 3] 5 is a schematic diagram for explaining an area feature encoding process in the area feature encoding unit according to the present invention; FIG. [Figure 4] FIG. 1 is a schematic diagram illustrating an embodiment of a training process for a movement and stay estimation model according to the present invention. [Figure 5] FIG. 10 is a schematic diagram for explaining another embodiment of the movement and stay estimation model according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0027] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0028] [Movement and Stay Estimation Model] FIG. 1 is a schematic diagram showing an embodiment of a movement / stay estimation model according to the present invention and an embodiment of a movement trajectory prediction device according to the present invention.

[0029] The movement and stay estimation model 1 of this embodiment shown in FIG. 1 is a model that estimates information (later movement and stay information) related to future movement or stay of an estimation target (in this embodiment, a user (a mobile terminal carried by the user)) in an area that is a survey target (survey target area), that is, route information of one future trip in this embodiment, and specifically, (A) an area feature position encoding unit 11 that receives as input (a) "area feature information" generated from "area thing / event information" relating to things or events that may affect or are related to the movement or stay of a user (or a mobile terminal carried by the user) in a survey target area, and (b) "movement / stay time series information" which is information relating to the time series of the movement or stay of a target user, and generates (c) "area feature embedded information" emb_ar in which the "area feature information" is embedded, and (d) "position embedded information" emb_p in which information relating to the position in the time series of the movement or stay of the user is embedded; (B) an area movement stay encoding unit 14 that receives the generated “area feature embedded information” emb_ar and “location embedded information” emb_p as input, or receives information combining these embedded information as input, and generates “area movement stay embedded information” z in which information related to the movement or stay of the estimated target user in the survey target area is embedded; (C) a movement and stay estimation unit 15 that receives the generated “area movement and stay embedded information” z as an input and generates route information (later movement and stay information) for one future trip of the estimation target user; It is a machine learning model that allows computers to function.

[0030] Here, the "area thing and event information" that is the basis for generating the "area feature information" in (A) above is, for example, the following in each of the multiple zones (mesh in this embodiment) that make up the survey target area: (A1) Information on the number, scale, or content of transportation or means of travel (for example, the number of train stations, the number of bus stops, the number of highway interchanges, whether express trains stop at existing train stations or not, the number of stops per day at existing bus stops, and the length of roads with a specified width or greater, etc.); (A2) Information regarding the number, scale, or content of public facilities or public events (e.g., the number of schools and the number of students, the number of hospitals and the number of beds, etc.), (a3) Information on the number, size, or content of commercial and service-providing facilities or commercial and service-providing events (e.g., the number of commercial facilities by size, etc.), (A4) Information on the number, size, or content of residential facilities (for example, the number and number of apartment buildings, the number of ordinary houses, etc.), and (A5) Information on the number, size, or content of business facilities (e.g., number of business establishments by size, etc.) The information includes at least one of the above, preferably information relating to each of (A1) to (A5).

[0031] Incidentally, such "area thing and event information" may be information compiled in advance by collecting and aggregating relevant information for each area (mesh) from map data, timetable data, etc., including the area to be surveyed. Also, "area feature information" may be a feature quantity that reflects such "area thing and event information," such as a multidimensional feature vector.

[0032] In this way, the movement and stay estimation model 1 is capable of estimating route information for one future trip (future movement and stay information) by taking into consideration the "area characteristic information," which is the in-area condition that may affect or be related to the movement and stay of the user, as well as the "movement and stay time-series information" of the estimation target user. Furthermore, as will be described in detail later, the movement trajectory prediction device 9 equipped with this movement and stay estimation model 1 is also capable of predicting the route of the estimation target user.

[0033] As a result, the movement and stay estimation model 1 (and the movement trajectory prediction device 9 equipped with it) can also be applied to traffic demand prediction for transportation policies, for example. For example, the movement and stay estimation model 1 takes into account not only "movement and stay time series information" including information on the relationship between movement and stay locations over time, but also "area characteristic information" which is the conditions within the area that may affect or are related to the user's movement and stay, so it is possible to predict the location of traffic congestion when the conditions within the area change, which is important in traffic demand prediction, for example, the location of traffic congestion that would occur if a hospital that was at point C was relocated to point D.

[0034] Furthermore, the movement and stay estimation model 1 (or the movement trajectory prediction device 9 equipped with it) can also predict the movement trajectory of the estimation target user. In other words, it can not only predict the destination to which the estimation target user will head in the future, but also predict the route the estimation target user will take to reach the destination. As a result, it becomes possible to predict information about individual routes, such as the usage volume of each road and each railway line, which is important in traffic demand prediction.

[0035] [Model configuration, movement and stay estimation method] The configuration of the movement stay estimation model 1 of this embodiment will be described in more detail below. Also in Fig. 1, the movement stay estimation model 1 includes an area feature position encoding unit 11 including an area feature encoding unit 111 and a movement stay position encoding unit 112, a movement stay time point encoding unit 12, an integration unit 13, an area movement stay encoding unit 14, and a movement stay estimation unit 15 as functional components (embodied by a program installed in a computer). In other words, it is a model that causes a computer to function as these functional components.

[0036] <Area feature position encoding section> 1, the area characteristic position encoding unit 11 of this embodiment includes an area characteristic encoding unit 111 and a movement and stay position encoding unit 112.

[0037] (Moving position encoding section) Of these, in this embodiment, the movement / stay position encoding unit 112 receives as input movement / stay time series information generated from the terminal position information of the user to be estimated, and generates position-embedded information emb_p in which information related to the position of the user to be estimated in the time series of movement or stay is embedded.

[0038] Here, the movement and stay position encoding unit 112 uses the embedding part of BERT (Bidirectional Encoder Representations from Transformers), a well-known natural language processing model, specifically a one-layer fully connected one-hot encoder Ω, as in the position prediction technology disclosed in Non-Patent Document 2. The embedding part of this BERT outputs a multidimensional vector in which the meaning of the input sentence (word string) is embedded (representing the meaning), and the movement and stay position encoding unit 112 inputs movement and stay time series information, for example, (u1, p1, t1), (u1, p2, t2), (u1, p3, t3), ... When you enter Ω(p1), Ω(p2), Ω(p3), ··· The position-embedded information emb_p is output, in which the position part of the input movement / stay time-series information is embedded. Here, for example, (u1, p1, t1) is data indicating that user 1 (u1) is at position p1 (p1 is, for example, a corresponding mesh code as described later) at time t1.

[0039] FIG. 2 is a schematic diagram for explaining one embodiment of the movement / stay time-series information (input to the movement / stay position encoding unit 112) according to the present invention.

[0040] According to the route diagram shown in the upper part of Fig. 2, in a given day, user 1 travels along the following route: Position A (home) → Position B → Position C (α Station) → Position D (β Station) → Position E (work) → Position F → Position G → ... These positions are the results of measuring the position of the mobile terminal carried by user 1 using a base station positioning method or a GPS (Global Positioning System). This measured position may also be expressed discretely, and may be displayed, for example, using the mesh code of the mesh that includes the measured position, or the base station identifier (ID) of the base station that includes the mobile terminal under its control.

[0041] In any case, for user 1 who took the route shown in Figure 2 on a certain day, the terminal location information shown below is A, A, A, B, C, D, E, E, E, F, G, H, I, I, I, J, ... Here, each capital alphabet letter separated by a comma and arranged in order represents a measurement position measured every predetermined time (for example, 15 minutes).

[0042] Next, a known movement and stay determination process is performed on such terminal location information to extract trips (movements), and the locations are expressed in a discrete form (for example, in a mesh code representation) to obtain time-series information on the movement and stay of user 1 on a certain day, for example A, A, A, [STAY_END], A, B, C, D, E, [TRIP_END], E, E,... Here, [STAY_END] indicates that the stay ended at this position (time), and [TRIP_END] indicates that the trip (movement) ended at this position (time).

[0043] It is also preferable to use, as the movement / stay determination process, the method disclosed in the non-patent document: Kobayashi Nao, Ishizuka Hiroki, Minamigawa Atsunobu, Muramatsu Shigeki, Ono Tomohiro, "Proposal of a Movement / Stay State Estimation Method Suitable for Mobile Phone Communication History," Information Processing Society of Japan Transactions on Database, Vol. 10, No. 1, pp. 13-23, 2017. Here, a trip (movement) may be extracted as a section between one stay and a subsequent stay.

[0044] The following describes movement and stay time-series information as input information when estimating next trip information (later movement and stay information) related to the next possible trip (movement) of the estimation target user using this movement and stay estimation model 1. This explains what form of movement and stay time-series information can be input when estimating the movement route (next trip information) that user 1 will take after commuting to work based on the movement route (actual trip information) that user 1 took to arrive at work.

[0045] Specifically, such movement and stay time series information is a set of trip information relating to an actual trip (movement) of the estimation target user and a candidate (candidate next trip information) of next trip information, which is information relating to a trip (movement) that may occur next, for example, (u1, p1, t1), (u1, p2, t2), ..., (u1, p5, t5), [TRIP_END] [SEP] (u1, p6, t6), (u1, p7, t7), ···, (u1, p9, t9), [TRIP_END] Here, for example, (u1, p1, t1) is data indicating that user 1 (u1) is at position p1 (p1 is, for example, the corresponding mesh code) at time t1. Also, [SEP] (the second line above) is a symbol that separates the actual trip information (the first line above) from the candidate next trip information (the third line above).

[0046] The movement and stay estimation model 1 (movement and stay estimation unit 15) of this embodiment inputs movement and stay time series information including such candidate next trip information, and outputs the likelihood that the next trip to occur will be the trip related to this candidate next trip information, and determines the one with the highest likelihood among the multiple input candidate next trip information as the information related to the estimated (likely to occur next) next trip (later movement and stay information).

[0047] Here, the candidate next trip information can be, for example, randomly selected from the user's trip information collected up to that point, trip information whose starting point location matches or is separated within a predetermined range from the final location of the trip information before [SEP] (for example, the location at the time of estimation). While it is also possible to generate trip information completely randomly and use it as candidate next trip information, it is also preferable to generate trip information whose trajectory is within a predetermined travel distance range from the final location of the trip information before [SEP] (for example, the location at the time of estimation), and use it as candidate next trip information.

[0048] Furthermore, it is also preferable to perform at least one of the following filtering processes on the movement / stay time-series information described above in order to remove noise, particularly during training of the movement / stay estimation model 1. (a) If the total time interval between one "stay" and the "stays" occurring consecutively thereto is shorter than a predetermined time, these "stays" are merged into one "stay." (b) For positions represented discretely (e.g., displayed as mesh codes), if the same values ​​are repeated, the duplicates are deleted (e.g., if the values ​​are A,A,A,B,B,B,C,C,C, then they are changed to A,B,C). (c) If the result of the deletion in (b) above is that the "movement" remains in the same position (for example, the same mesh), this "movement" is deleted as noise.

[0049] Although one embodiment of the movement / stay time series information input to the movement / stay position encoding unit 112 has been described above, the movement / stay time series information is not limited to this. For example, the movement / stay time series information may be a time series of "stay" and / or "movement" for a predetermined period, for example, ..., [TRIP_END], ..., [STAY_END], ...[TRIP_END] [SEP] ..., [STAY_END], ...[TRIP_END] Here, the candidate information after [SEP] can be set in various forms, for example, only "stay".

[0050] (Area feature encoding part) Returning to FIG. 1, the area feature encoding unit 111 (a) among a plurality of pre-defined meshes (areas) constituting the survey target area, area feature information generated from area object event information in a mesh (target mesh) including the time-series positions of the movement or stay of the estimated target user is input; (b) Generate area feature embedding information emb_ar in which the area feature information of this target mesh is embedded.

[0051] FIG. 3 is a schematic diagram for explaining the area feature encoding process in the area feature encoding unit 111. As shown in FIG.

[0052] According to FIG. 3(A), first, area object and event information, such as the above-mentioned information, for each mesh constituting the survey area is obtained from map data, timetable data, etc. (A1) Number of transportation or means of transportation, (A2) Number of public facilities or public events, (A3) Number of commercial or service-providing facilities or commercial or service-providing events, (A4) Number of residential facilities, (A5) Number of business facilities, has been compiled.

[0053] The area feature information of each mesh is generated from such area fact and event information. In this embodiment, the area feature information of each mesh is, for example, ([Number of train stations], [Number of trains stopping at each station per day], [Number of schools], [Number of large retail stores], ...) Here, it is preferable that the numerical value in [·] is a value normalized to the range of 0 to 1. In any case, in this embodiment, such area feature information for each mesh is stored and managed in the area feature database (DB) 111a.

[0054] Here, in this embodiment, the area feature encoding unit 111: (a) Identify the mesh code including the travel route (related to the actual trip information) from the collected travel and stay time series information, (b) Specify and input each identified mesh code to the area feature DB 111a to obtain area feature information for each target mesh (related to actual trip information); (c) The area feature information of each target mesh is input to a multi-layer perceptron (MLP) for compression, and the area feature embedding information emb_ar is output. That is why.

[0055] As another embodiment of the area feature encoding process, as shown in FIG. 3(B), the area feature encoding unit 111 is provided with a known graph convolution network (GCN) that receives graph information as input, extracts and outputs the feature amount of the graph, (a) The area feature information is the area graph information in which the area event information in the target mesh (area) is the "node" and the matrix representing the link setting status between the target mesh and other surrounding meshes (areas) is the "graph adjacency matrix." (b) Input area graph information (area feature information) to the above GCN and output area feature embedding information emb_ar as a feature representing the input graph. It is also preferable.

[0056] The GCN used here is described in, for example, the following non-patent document: Thomas N. Kipf, Max Welling, "Semi-Supervised Classification with Graph Convolutional Networks", Published as a conference paper at ICLR (Poster) 2017,<https: / / arxiv.org / pdf / 1609.02907.pdf> The model can be that disclosed in

[0057] Also in this embodiment, the area graph information (area feature information) for each mesh may be stored and managed in the area feature DB 111a. The area feature encoding unit 111 specifies and inputs to this area feature DB 111a the mesh code of each target mesh (including the actual movement route) identified based on the acquired movement and stay time-series information, acquires the area graph information of each target mesh (related to actual trip information), inputs the acquired area graph information of each target mesh to the GCN for feature quantification processing, and outputs the area feature embedding information emb_ar.

[0058] In addition, the graph adjacency matrix of a certain target mesh included in the area graph information (area feature information) can be a matrix in which a link is established between the target mesh and a geographically adjacent mesh, and the matrix element corresponding to the adjacent mesh to which the link is established is set to a predetermined value, for example, 1.

[0059] As a modification, it is also preferable that this graph adjacency matrix be a matrix having matrix elements that reflect the presence or absence of specified means of transportation, movement, distribution, or communication between a certain target mesh and other meshes surrounding it (adjacent or within a specified range), or the traffic volume, movement volume, distribution volume, or communication volume per unit time.

[0060] For example, in a graph adjacency matrix, the matrix element corresponding to another mesh containing another railway station on a railway line including a railway station present in the target mesh may be set to a predetermined value, for example, 1. Alternatively, the number (or standard value) of "other railway stations on the railway line" included in this other mesh may be set as the matrix element. Furthermore, the matrix element corresponding to another mesh containing another railway station on a railway line including a railway station present in this target mesh may be the number (or standard value) of trains running between these two stations per unit time (for example, 1 hour or 24 hours).

[0061] Also, in the graph adjacency matrix, for example, the traffic volume (for example, the number of cars passing through) per unit time (for example, 1 hour or 24 hours) on the road passing between the target mesh and its adjacent mesh (its standard value) may be set as the matrix element corresponding to this adjacent mesh. In any case, by inputting the graph adjacency matrix as described above into the GCN, the area feature encoding unit 111 can generate area feature embedded information emb_ar in which the geographical relationships between meshes and the relationships related to mutual entry and exit and exchange are also embedded.

[0062] Furthermore, when using the graph adjacency matrix described above, the area graph information may include not only the node information (area thing event information) of the target mesh, but also the node information (area thing event information) of other surrounding meshes in which non-zero finite values ​​are set as matrix elements of the graph connectivity matrix. This allows more information useful for estimating the movement and stay of the estimation target to be reflected in the area feature embedding information emb_ar.

[0063] <Movement Stay Time Encoding> Returning to FIG. 1, the movement stay time point encoding unit 12 receives the movement stay time series information generated from the terminal location information as input, and generates time point embedded information emb_t in which information relating to a time point in the time series of this information is embedded.

[0064] In this embodiment, the movement / stay time encoding unit 12 uses a temporal embedding module Φ of the location prediction technology disclosed in Non-Patent Document 2. This temporal embedding module Φ uses, for example, the following as movement / stay time-series information: (u1, p1, t1), (u1, p2, t2), (u1, p3, t3), ... When is entered, Φ(t1), Φ(t2), Φ(t3), ··· The time-point embedded information emb_p is output, in which the time-point portion of the input movement / stay time series information is embedded. Here, for example, Φ(t1) is expressed by the following equation: (1) Φ(t1) =[cos(w1×t1), sin(w1×t1), ..., cos(wd×t1), sin(wd×t1)] In the above formula (1), w1,...,wd are learning parameters, and d is half the number of vector dimensions of the generated multidimensional feature vector (here, the time-point embedding information emb_t).

[0065] By generating time embedding information emb_t using such a temporal embedding module Φ and performing a movement stay estimation process, it becomes possible to obtain a more accurate movement stay estimation result that takes into account information that is very important for movement stay estimation, such as the length of stay along the movement route and the movement speed. However, this movement stay time encoding unit 12 can be omitted, although this advantage cannot be enjoyed.

[0066] <Integration Department> In this embodiment, the integration unit 13 shown in FIG. (a) Area feature embedding information emb_ar output from the area feature encoding unit 111; (b) location embedding information emb_p output from the movement stay location encoding unit 112; (c) The time embedding information emb_t output from the travel time encoding unit 12 The integration unit 13 generates a result of taking a (vector) sum or a (vector) concatenation of the above, or a result of compressing the result of taking a (vector) sum or the result of concatenation using a multi-layer perceptron (MLP), and outputs this embedded information as integrated information to the area movement stay encoding unit 14. Alternatively, the integration unit 13 can output the embedded information (a) to (c) above as is to the area movement stay encoding unit 14.

[0067] <Area Movement and Stay Encoding Section> Also in Figure 1, the area movement stay encoding unit 14 uses the output from the above-mentioned integration unit 13 as input and generates area movement stay embedded information z in which information related to the movement or stay of the estimated target user in the survey target area is embedded.

[0068] Specifically, in this embodiment, the area movement stay encoding unit 14 uses transformer encoder layers of the position prediction technology disclosed in Non-Patent Document 2. The transformer encoder layers are an application of the BERT transformer encoder, and each of the N layers in the transformer encoder layers is a layer configured from a multi-head self-attention module and a position-wise fully connected feed-forward network.

[0069] For example, if the input output from the integration unit 13 is z'=(z'1, z'2, z'3, . . .), the output of the kth layer is expressed by the following equation, where the processing in the layer is Trans( . ). (2) (h(k)1, h(k)2, h(k)3,...)=Trans(h(k)1, h(k)2, h(k)3,...) Here (h(0)1, h(0)2, h(0)3,...)=(z'1, z'2, z'3,...) The area change stay embedding information z, which is the final output of the transformer encoder layers (area change stay encoding unit 14), is expressed by the following formula: (3) z=(z1, z2, z3,...)=(h(N)1, h(N)2, h(N)3,...) It is calculated as follows.

[0070] Such area movement and stay embedding information z is a representation vector similar to the word distributed representation vector in which the meaning of a sentence, which is a word sequence in BERT, is embedded, and is a multidimensional representation vector in which information on the activity intention and activity content (which also depends on the conditions within the area) of the estimated target user's movement and stay in the survey area is embedded.

[0071] <Movement and Stay Estimation Unit> Also in Figure 1, the movement and stay estimation unit 15 receives the area movement and stay embedded information z output from the area movement and stay encoding unit 14 as input, and generates and outputs information related to the future movement or stay of the estimated user, in this embodiment, route information for one trip at the next point in time (future).

[0072] Specifically, in this embodiment, the movement / stay estimation unit 15 outputs the likelihood that the candidate next trip information included in the movement / stay time series information input to the movement / stay estimation model 1 is information related to the actual next trip. That is, the movement / stay time series information input to the model 1 is (a) Trip information relating to actual trips (movements) of the target user during a certain period; and (b) Candidate next trip information, which is information related to trips (movements) that may occur in an estimated period after the certain period. The movement stay estimation unit 15 generates and outputs information relating to the likelihood of the candidate next trip information (b) above based on the area movement stay embedded information z taken from the area movement stay encoding unit 14.

[0073] Here, the movement / stay estimation unit 15 is configured to apply NSP (Next Sentence Prediction) in BERT, which determines whether two passed sentences are consecutive or not, and can have the configuration of, for example, a fully-connected feed-forward neural network (FFN).

[0074] As described above, the movement and stay estimation unit 15 can generate information related to the likelihood of candidate next trip information (later movement and stay information) based on the area movement and stay embedded information z in which information on the activity intention and activity content (which also depends on the in-area conditions) of the estimation target user's movement and stay is embedded. In other words, it is possible to obtain more accurate movement and stay estimation results that also take into account the in-area conditions.

[0075] [Training of movement and stay estimation model 1]

[0076] FIG. 4 is a schematic diagram for explaining an embodiment of the training process in the movement and stay estimation model 1. In FIG.

[0077] 4, in the training process of the movement stay estimation model 1 of this embodiment, specifically in pre-training for setting initial values ​​of learning parameters, a movement stay estimation unit 15 and a masked part estimation unit 16 are provided in the subsequent stage of the area movement stay encoding unit 14. This masked part estimation unit 16 is configured to apply MLM (Masked Language Model) for solving a problem in BERT in which data in a masked part is assigned (a word fill-in-the-blank problem), and can have the same FFN configuration as the above-mentioned NSP (movement stay estimation unit 15).

[0078] Here, the training (learning) data is, for example, (u1, p1, t1), (u1, p2, t2), ..., (u1, p5, t5), [TRIP_END] [SEP] (u1, p6, t6), (u1, [MASK], [MASK]), ···, (u1, p9, t9), [TRIP_END] For example, if the actual trip information in the first row is followed by the next trip information in the third row (with some masked areas), then the training data above will be the correct answer data.

[0079] In addition, device location information from smartphones and other devices often contains missing data or superimposed noise, which can lead to trip information that is not actually continuous being selected as the correct next trip information, resulting in the creation of training data that becomes noisy during pre-training. Therefore, if location data cannot be collected continuously for the previous "stay," for example, because the positioning measurement time span exceeds a predetermined threshold, it is preferable to exclude the corresponding trip information from the training data.

[0080] In the pre-training of this embodiment, a large amount of such training data and the (correct) area Features The information is input to the travel and stay estimation model 1 with the training specifications shown in Figure 4, (a) Error from the correct answer (1 or 0) in the output of the movement / stay estimation unit 15 (NSP), (b) Error from the correct value in the output of the mask location estimation unit 16 (MLM) The combined (e.g., weighted) amount is taken as the loss, and the initial values ​​of the learning parameters in Model 1 are determined according to the backpropagation method. By carrying out such pre-training, the area movement / stay encoding unit 14 (transformer encoder) can output area movement / stay embedding information z that more appropriately reflects information on activity intentions and activity contents that also depend on the conditions within the area during movement / stay.

[0081] [Another embodiment of the movement and stay estimation model] FIG. 5 is a schematic diagram for explaining another embodiment of the movement and stay estimation model according to the present invention.

[0082] Like the movement stay estimation model 1 shown in Fig. 1, the movement stay estimation model 2 of the embodiment shown in Fig. 5 is a model that estimates and outputs the likelihood of subsequent movement stay information, for example, candidate next trip information, from area feature information in the survey target area and movement stay time-series information of the estimation target user. Furthermore, the movement stay time point encoding unit 22, area movement stay encoding unit 24, and movement stay estimation unit 25 of this movement stay estimation model 2 each have the same functional configuration as the functional configuration units of the same names in the movement stay estimation model 1 (Fig. 1).

[0083] In addition, training (pre-training) of the moving stay estimation model 2 can be performed in the same way as for the moving stay estimation model 1 (FIG. 4). Furthermore, the moving stay time encoding unit 22 of the moving stay estimation model 2 can also be omitted, just like the moving stay time encoding unit 12 of the moving stay estimation model 1.

[0084] However, the area feature position encoding unit 21 of the movement stay estimation model 2 differs from the area feature position encoding unit 11 of the movement stay estimation model 1 in that the area feature processing unit (category value generation unit 211) is provided before the movement stay position encoding unit 213. As a result, instead of outputting the area feature embedding information emb_ar and the position embedding information emb_p, it outputs "area feature position embedding information" emb_ar_p in which information combining the area feature information and the position information in the movement stay time-series information is embedded.

[0085] In response to this, the integration unit 23 (a) generates information that combines the area characteristic position embedded information emb_ar_p and the time point embedded information emb_t and outputs the combined information, or (b) each of these embedded information, to the area movement stay encoding unit 24.

[0086] Specifically, in the movement and stay estimation model 2 of this embodiment, the area characteristic position encoding unit 21 (A) A category value generation unit 211 that generates, based on the received area feature information, a category value representing a preset category, which is a category value that expresses the characteristics of a target mesh (including the time-series position of the movement or stay of the estimation target user) among multiple meshes that make up the survey target area; (a) a mesh code unit 212 that generates time-series information of a location (mesh code) from the received movement and stay time-series information; (c) a movement / stay position encoding unit 213 that receives the category value of (a) above and the mesh code time series information of (b) above, and generates and outputs area characteristic position embedding information emb_ar_p; It is equipped with:

[0087] Here, in this embodiment, the category value generation unit 211 (a) above first receives area feature information (related to the movement or stay of the user to be estimated) such as the stay time, movement speed, movement direction, date and time, home location, and workplace location of the user to be estimated.

[0088] Next, the category value generation unit 211 uses the received area feature information to calculate the following for each target mesh (including the time-series positions of the movement or stay of the estimation target user): (a) Stay time category value (for example, if the stay time in the target mesh is less than 15 minutes, it is "A", if it is between 15 minutes and 60 minutes, it is "B", and if it is 60 minutes or more, it is "C"); (b) Speed ​​category value (for example, if the average moving speed in the target mesh is less than 5 km / h, it is "X", if it is between 5 km / h and 20 km / h, it is "Y", and if it is 20 km / h or more, it is "Z"); (c) Movement direction category value (for example, if the direction of the movement path going out of the target mesh is north (0 degrees or more and less than 45 degrees and 315 degrees or more and less than 360 degrees in the left-handed north reference azimuth), then "K", if east (45 degrees or more and less than 135 degrees), then "E", if south (135 degrees or more and less than 225 degrees), then "S", if west (225 degrees or more and less than 315 degrees), then "T", if there is no movement path going out, then "O"), (d) Home location category value (for example, the mesh code of the mesh including the home location), (e) Workplace location category value (for example, the mesh code of the mesh including the workplace location), (f) Regular holiday category value (for example, if the day in the time series within the target mesh is a weekday, then "W", if it is a holiday, then "H"), and (g) Time zone category value (for example, if the time zone in the time series within the target mesh is daytime (from 6 o'clock to 19 o'clock), then "D", if it is nighttime (from 20 o'clock to 5 o'clock the next day), then "N") Generate at least one of them, for example, all of them.

[0089] Next, the movement stay position encoding unit 213 in the above (c) receives the generated category values and the mesh code time series information in the above (a), and creates "mesh code time series information with additional information" from these information, and generates area feature position embedding information emb_ar_p from this mesh code time series information.

[0090] Here, the "mesh code time series information with additional information" uses <·| as the mesh code time series information and |·> as the additional information. For example, for a certain target mesh, <5000000>003|-|AYK> <5000000>003|-|5000000123>-|5000000987> <5000000>003|-|WD> For example, |AYK> in the first row is additional information indicating that the estimated target user's stay time in this target mesh is less than 15 minutes, the average movement speed is between 5 km / h and 20 km / h, and the direction of movement is north.

[0091] Also, |5000000123> in the second row indicates that the home location is within the mesh with mesh code 5000000123, and |5000000987> indicates that the workplace location is within the mesh with mesh code 5000000987. Furthermore, |WD> in the third row indicates that the day related to the time series within the target mesh is a weekday and that the time period is daytime.

[0092] Here, the additional information in the first row (|AYK> in the above) is a behavioral feature B associated with a total of 45 (=3×3×5) types of category value sequences, which are a combination of, for example, three types of "A", "B", and "C", three types of "X", "Y", and "Z", and five types of "K", "E", "S", "T", and "O". The additional information in the second row (|5000000123>-|5000000987> in the above) is an individual feature P associated with two types of category value sequences, each consisting of n types of category values ​​and m types of category values, where the number of meshes that can be determined as home and workplace is n and m, respectively. The additional information in the third row (|WD> in the above) is a time feature T associated with a total of 4 (=2×2) types of category value sequences, which are a combination of, for example, two types of "W" and "H", and two types of "D" and "N".

[0093] The functional configuration of the area characteristic position encoding unit 21 has been described above, but in this embodiment, the category values ​​(which are the source of additional information) generated by the category value generating unit 211 correspond to subwords in BERT. Similarly, the mesh code time-series information generated by the mesh coding unit 212 corresponds to words (sequences of words) in BERT.

[0094] Here, BERT uses the WordPiece algorithm to split a word (e.g., "doing") into a word (e.g., "do") and a subword (e.g., "##ing", where ## represents a subword), making it easier to understand the meaning of the word and suppressing the increase in the dimensionality of the word representation vector, thereby achieving efficient training and reducing computational costs.

[0095] Furthermore, the "mesh code time series information with additional information" <·|-|·> in the first and third rows above corresponds to the word information in BERT: Word-Subword. Also, the "mesh code time series information with additional information" <·|-|·>-|·> in the second row above corresponds to word information with two subwords: Word-Subword-Subword. Furthermore, the area feature position encoding unit 21 applies the transformer encoder of BERT to generate area feature position embedded information emb_ar_p from the "mesh code time series information with additional information" described above.

[0096] In this way, the area characteristic position encoding unit 21 of this embodiment can fully utilize the functions of BERT to generate area characteristic position embedded information emb_ar_p that can accurately estimate future movement and stay information (e.g., the likelihood of candidate next trip information).

[0097] [Trajectory prediction device, Trajectory prediction program / method] Returning to FIG. 1, the following will describe a movement trajectory prediction device 9 that is equipped with the movement and stay estimation model 1 and predicts the future movement trajectory of a prediction target user. Mobile stay The following description will be given assuming that the estimation model 1 is installed, but it is of course also possible to install the movement and stay estimation model 2 shown in FIG.

[0098] The trajectory prediction device 9 of this embodiment shown in the lower part of FIG. (A) an area feature generation unit 94 that generates area feature information from area object event information acquired via an input unit 91 from an external server, such as a map data management server, and stored and managed in an area object information storage unit 92; (B) a movement / stay time series information generating unit 95 that generates a plurality of pieces of movement / stay time series information (including mutually different pieces of candidate next trip information) for the prediction target user from terminal location information acquired via an input unit 91 from an external device, for example, a mobile data management server, and stored and managed in terminal location information 93; (C) A movement stay prediction unit 96 that inputs these multiple pieces of movement stay time series information and area feature information into a movement stay estimation model 1 and determines the candidate next trip information (included in the movement stay time series information) with the highest likelihood output from this model 1 as predicted next trip information (later movement stay information), and that sequentially determines next trip information (later movement stay information) that is consecutive in time series by repeatedly performing a process of predicting the above-mentioned next trip information from actual trip information including the initial position of the user to be predicted, and then predicting the next next trip information using this next trip information as the actual trip information; (D) a movement trajectory prediction unit 97 that predicts movement trajectory information (for example, for a certain day) of the prediction target user based on the sequentially determined next trip information (subsequent movement and stay information); (E) an output unit 98 having a communication function and transmitting the predicted movement trajectory information to an external information processing device, or having a display and displaying the predicted movement trajectory information; It has the following characteristics.

[0099] In this way, the movement trajectory prediction device 9 can predict the movement trajectory of the user to be predicted, taking into consideration not only movement and stay time-series information but also area characteristic information (conditions within the area). This makes it possible to predict, for example, the movement trajectories of a number of users in a certain urban area, including their respective routes, taking into consideration information related to the behavior and activity details of the users, and to predict the traffic behavior of the entire urban area, similar to ABM (activity model).

[0100] That is, the movement trajectory prediction device 9 can also perform traffic demand prediction for traffic measures equivalent to ABM without relying on questionnaire surveys. Furthermore, it is also possible to predict the location of traffic congestion when the conditions within the area change, which is important in traffic demand prediction, for example, the location of traffic congestion that would occur if a hospital located at point C were to be relocated to point D.

[0101] Incidentally, the area feature generation unit 94, movement / stay time-series information generation unit 95, movement / stay prediction unit 96, and movement trajectory prediction unit 97 described above in (A) to (D) are main functional components that implement an embodiment of a movement trajectory prediction method according to the present invention, and can also be considered as functions of a processor / memory that stores an embodiment of a movement trajectory prediction program according to the present invention. Therefore, the movement trajectory prediction device 9 may be a device dedicated to movement trajectory prediction, or it may 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 movement trajectory prediction program according to the present invention.

[0102] As explained in detail above, according to the present invention, it is possible to estimate subsequent movement and stay information by taking into consideration not only the movement and stay time-series information of the target to be estimated, but also area characteristic information, which is an in-area condition that may affect or be related to the movement and stay of the target. This also makes it possible to clarify the route of the target to be estimated and predict its movement trajectory.

[0103] Furthermore, for example, by applying the movement trajectory prediction method according to the present invention, accurate and detailed traffic behavior predictions can be made in urban areas, leading to the development of resilient traffic and transportation infrastructure and efficient urban planning that solves the challenges brought about 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.

[0104] 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]

[0105] 1, 2 Movement and Stay Estimation Model 11, 21 Area feature position encoding section 111 Area feature encoding part 111a Area Characteristics Database (DB) 112, 213 Moving and Staying Position Encoding Section 12, 22 Travel time encoding section 13, 23 Integrated Department 14, 24 Area Moving Stay Encoding Section 15, 25 Movement and Stay Estimation Section 16 Mask location estimation section 211 Category Value Generation Unit 212 Mesh cord section 9. Trajectory prediction device 91 Input section 92 Area Object and Event Information Storage Unit 93 Terminal location information storage unit 94 Area feature generation unit 95 Movement and Stay Time Series Information Generation Unit 96 Travel and Stay Prediction Department 97 Trajectory Prediction Unit 98 Output section

Claims

1. A movement and stay estimation model that causes a computer to estimate information related to a future movement or stay of an object in an area, comprising: an area feature position encoding unit which receives as input area feature information generated from information relating to things or events related to or that may affect the movement or stay of the subject in the area, and information relating to a time series of the movement or stay of the subject, and generates area feature embedding information in which the area feature information is embedded, and position embedding information in which information relating to a position in the time series is embedded, or generates area feature position embedding information in which information combining the area feature information and the information relating to a position in the time series is embedded; an area movement stay encoding unit which receives as input the area feature embedding information and the position embedding information, or receives as input information combining the area feature embedding information and the position embedding information, or receives as input the area feature position embedding information, and generates area movement stay embedding information in which information related to the movement or stay of the target in the area is embedded; a movement and stay estimation unit that receives the area movement and stay embedding information and generates information related to the subsequent movement or stay; A movement and stay estimation model characterized by causing a computer to function as follows.

2. The movement and stay estimation model according to claim 1, characterized in that the area characteristic information is generated from area thing and event information including at least one of information related to the number, size or content of transportation or means of travel in each of a plurality of sections that make up the area, information related to the number, size or content of public facilities or public events, information related to the number, size or content of commercial / service facilities or commercial / service providing events, information related to the number, size or content of residential facilities, and information related to the number, size or content of business facilities.

3. making the computer function as a movement / stay time point encoding unit that receives information related to a time series of the movement or stay of the object as input and generates time-embedded information in which information related to a time point in the time series is embedded; The area movement stay encoding unit receives the area feature embedding information, the position embedding information, and the time embedding information as input, or receives information combining the area feature embedding information, the position embedding information, and the time embedding information as input, or receives the area feature position embedding information and the time embedding information as input, or receives information combining the area feature position embedding information and the time embedding information as input. The movement and stay estimation model according to claim 1 or 2.

4. The information on the time series of the movement or stay of the object includes information on the location and time of the movement or stay during a certain period of time and information on the location and time of the movement or stay during an estimated period after the certain period of time; The movement and stay estimation model according to any one of claims 1 to 3, characterized in that the movement and stay estimation unit generates, as the information related to the subsequent movement or stay, information related to the likelihood of information related to the location and time of the movement or stay during the estimation period.

5. the area feature position encoding unit has an area feature encoding unit that receives area feature information as an input and generates area feature embedded information in which the area feature information is embedded, and a movement / stay position encoding unit that receives information related to a time series of the movement or stay and generates the position embedded information, The area movement stay encoding unit receives the area feature embedding information and the location embedding information as input, or receives information combining the area feature embedding information and the location embedding information as input. The movement and stay estimation model according to any one of claims 1 to 4.

6. The area characteristic information is area graph information in which information on things or events that may affect or are related to the movement or stay of the subject in a target area among multiple areas constituting the area is used as a node, and a matrix expressing the setting status of links between the target area and other surrounding areas is used as a graph adjacency matrix, The area feature encoding unit includes a graph convolution network (GCN) that receives the area graph information as an input. The movement and stay estimation model according to claim 5 .

7. The movement and stay estimation model according to claim 6, characterized in that the graph adjacency matrix has matrix elements reflecting the presence or absence of a specified means of traffic, movement, distribution or communication between the target area and other surrounding areas, or the traffic volume, movement volume, distribution volume or communication volume per unit time.

8. The movement and stay estimation model according to any one of claims 1 to 4, characterized in that the area characteristic position encoding unit generates, from the received area characteristic information, a category value that represents a preset category and expresses a characteristic of a target area among a plurality of areas that constitute the area, and generates area characteristic position embedding information in which information combining the category value and information related to the position in the time series is embedded.

9. The moving stay estimation model according to any one of claims 1 to 8, characterized in that the moving stay estimation model is a model trained by a training process in which training data includes information related to a time series of movement or stay of the target having a masked area and correct area feature information, and the training process assumes a loss equal to the sum of an error from a correct answer in an output from a masked area estimation unit for solving a problem of guessing data for a masked area, which is provided for training together with the moving stay estimation unit and is provided downstream of the area moving stay encoding unit, and an error from a correct answer in an output from the moving stay estimation unit.

10. An area feature generating means for generating the area feature information from information on things or events that may affect or are related to the movement or stay of the subject; a movement / stay time series information generating means for generating information related to a time series of the movement or stay of the target from the acquired terminal location information; A movement and stay estimation model according to any one of claims 1 to 9; A movement and stay prediction means for determining information related to the subsequent movement or stay using the movement and stay estimation model; a movement trajectory prediction means for predicting information related to the movement trajectory of the target based on the information related to the subsequent movement or stay determined in sequence; A movement trajectory prediction device comprising:

11. An area feature generating means for generating the area feature information from information on things or events that may affect or are related to the movement or stay of the subject; a movement / stay time series information generating means for generating information related to a time series of the movement or stay of the target from the acquired terminal location information; a movement and stay prediction means for providing the area feature information and information related to a time series of the movement or stay of the target as input information to the movement and stay estimation model according to any one of claims 1 to 9, and determining information generated by the movement and stay estimation unit as information related to the subsequent movement or stay; a movement trajectory prediction means for predicting information related to the movement trajectory of the target based on the information related to the subsequent movement or stay determined in sequence; A movement trajectory prediction device comprising:

12. 1. A computer-implemented movement and residence estimation method for estimating information related to a future movement or residence of an object in an area, comprising: a step of taking as input area feature information generated from information relating to things or events that may affect or are related to the movement or stay of the subject in the area, and information relating to a time series of the movement or stay of the subject, and generating area feature embedded information in which the area feature information is embedded, and location embedded information in which information relating to the location in the time series is embedded, or generating area feature location embedded information in which information combining the area feature information and the information relating to the location in the time series is embedded; a step of receiving the area feature embedding information and the location embedding information, or receiving information combining the area feature embedding information and the location embedding information, or receiving the area feature location embedding information, and generating area movement / stay embedded information in which information related to the movement or stay of the target in the area is embedded; A step of generating information related to the subsequent movement or stay using the area movement / stay embedding information as an input; A movement and stay estimation method comprising:

Images