JP7526752B2 - Target distribution estimation program, device, system and method for performing expanded estimation based on actual number measurements - Google Patents

Description

The present invention relates to a technology for estimating the distribution of objects, such as population distribution, in a given area.

Understanding population distribution (or people flow, including distribution that changes over time) in a given area as high-resolution data is extremely important for marketing, efficient operation of transportation networks (avoiding congestion, etc.), urban planning including the development of public facilities, and even infectious disease countermeasures.

Increasing the resolution of population distribution (people flow) data here means reducing the size of the regional mesh (hereafter abbreviated as mesh), which is the unit for tabulating population (people flow). In this regard, the positioning data sets from mobile devices (owned by users) that have traditionally been used as the source data for population distribution are data sets relating to points in latitude and longitude coordinates, for example, and in that sense have the highest resolution. However, it is natural that such positioning data sets alone cannot provide information on the positions of all people in the target area.

Therefore, in practice, population distribution data must be generated by performing an expansion estimate to the total population from the mobile device positioning data. However, if an expansion estimate is simply performed from only the mobile device positioning data, for example, for meshes where no positioning data exists, the error in the estimate from the expansion estimate will be large (because no matter how many times you multiply 0 people, it still remains 0 people). Furthermore, particularly in urban areas, there is a good chance that GPS (Global Positioning System) positioning data originally contains large errors due to the influence of densely packed buildings, etc.

As a result, in order to generate population distribution (people flow) data that accurately reflects the overall distribution of all people, it has been necessary to set the mesh size large, i.e., low resolution, and to avoid situations where each mesh does not contain positioning data or does not (does) have positioning data that it should (should not) have. Figure 6 shows a schematic diagram of how the precision (accuracy) of a population distribution heat map generated using a smaller mesh (high-resolution mesh) from sparse terminal location information (positioning data) is significantly reduced.

As an example of a technology to achieve high resolution of population distribution (people flow), which has been difficult to achieve, Non-Patent Document 1 discloses a technology that models the spatial correlation of people flow and the influence of external factors such as weather, and uses AI (artificial intelligence) to estimate a high-resolution people flow map while using the total population as a constraint.

Non-Patent Document 2 also discloses a technology that generates high-resolution people flow data by combining an AI model that achieves super-resolution in the spatial direction with an AI model that achieves super-resolution in the temporal direction.

Furthermore, Non-Patent Document 3 discloses a technology that combines (a) smartphone location information and (b) beacon location information to generate detailed people flow data. In this technology, (a) smartphone location information is positioned as "low resolution data that can tell us the volume of people flow (population) in a certain wide area," and (b) beacon location information is positioned as "data for only a portion of users, but from which detailed location information can be obtained," and the (b) beacon location information data is expanded and estimated so that the total population is the number of (a) smartphone location information pieces.

Fan Zhou, Liang Li, Ting Zhong, Goce Trajcevski, Kunpeng Zhang, Jiahao Wang, “Enhancing Urban Flow Maps via Neural ODEs”, Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (IJCAI2020) Main track. Pages 1295-1302 , 2020 Zong, Z., Feng, J., Liu, K., Shi, H., & Li, Y., “DeepDPM: Dynamic Population Mapping via Deep Neural Network”, Proceedings of the AAAI Conference on Artificial Intelligence, 33(01) ), pp.1294-1301. 2019 "Tomorrow's Congestion Forecast Portal", [online], [searched March 10, 2022], Internet <URL: https://prtimes.jp/main/html/rd/p/000000004.000090522.html>

However, the conventional techniques disclosed in the above-mentioned non-patent documents 1 to 3 still have problems with increasing the resolution of population distribution (people flow) data.

For example, the techniques disclosed in Non-Patent Documents 1 and 2 are both premised on the acquisition of high-resolution correct data on people flow in the estimation target area. In other words, high-resolution correct data is required to train an AI model. However, in reality, such correct data is rarely acquired, and therefore this premise is often unrealistic. For example, a group of GPS positioning data from a smartphone provides only the location information of "users who have a contract with the target telecommunications carrier, own a GPS-equipped model, and agree to the provision of GPS positioning data" out of the location information of the entire population that should be considered as correct, and it is clearly not possible to use this as correct data on people flow (population distribution).

In addition, the technology disclosed in Non-Patent Document 3 uses an interpolation process to make up for the gaps in the beacon location information (location data) (b), which is only collected spot-wise, and there are concerns that the accuracy of the estimated location may decrease. Furthermore, the smartphone location information (a) is considered to be "data that indicates the volume of people moving around (population)," and for example, a simple, uniform, expanded estimate is applied to this information. As a result, there are concerns that the accuracy of the estimated number of people may decrease.

The present invention aims to provide a target distribution estimation program, device, system, and method that can estimate the distribution of a target with high resolution and higher precision (accuracy) without relying on ground-truth data related to the actual distribution of the target.

According to the present invention, there is provided an object distribution estimation program for causing a computer to function to estimate a distribution of a plurality of objects in a predetermined area, the program comprising:
a presence distribution determining means for determining a presence distribution as a distribution of the number of objects present in the predetermined area from a positioning result obtained during a predetermined period for a terminal attached to the object;
a correct answer magnification coefficient determination means for determining a correct answer magnification coefficient for each measurement point from measurement results obtained from a plurality of real number measurement means capable of actually measuring the number of the target at each measurement point within the specified area and the presence number distribution;
a feature generating means for generating, for each of a plurality of points within the predetermined region, a feature characterizing the point or an area including the point;
an expansion coefficient estimation means for inputting the feature quantity to an expansion coefficient estimation model constructed using the correct expansion coefficient, and estimating an expansion coefficient of the point from an output of the expansion coefficient estimation model;
There is provided an object distribution estimation program that causes a computer to function as object presence distribution determination means for determining the number of objects present at each of a plurality of points within the specified area based on the presence distribution and the estimated expansion coefficient.

In the object distribution estimation program according to the present invention, the feature generating means
A process of generating a feature quantity of the location or the area including the location from information related to a specific thing or phenomenon related to the location or the area including the location based on the acquired map information of the specified area;
It is also preferable to generate the features by performing at least one of the following processes: a process of generating features of the location or the area including the location from attribute information of a target attached to a terminal located at the location or in an area including the location, the attribute information being estimated from history information of the positioning result or obtained from registered information of the target; and a process of treating a code representing the area including the location as a word, treating time-series data of the code representing the movement history of the target as a sentence, and generating features of the area using a natural language processing model that has learned similarities between words.

In addition, as an embodiment of the target distribution estimation program according to the present invention,
The presence number distribution determination means determines a normal presence number distribution as a distribution of numbers normally present from the positioning results accumulated during a first predetermined period, and determines an estimated presence number distribution as a distribution of numbers present at or immediately before the estimation time of the distribution from the positioning results acquired during a second predetermined period shorter than the first predetermined period and immediately before the estimation time of the distribution, and the positioning results accumulated during the first predetermined period;
The correct answer expansion coefficient determination means determines a correct answer expansion coefficient for the measurement point from the measurement result and the normal presence distribution;
It is also preferable that the object presence distribution determining means determines the number of the object present at each of a plurality of points within the specified area based on the estimated presence number distribution and the estimated expansion coefficient.

Furthermore, in the above embodiment, it is also preferable that the presence distribution determination means sets, as the first predetermined period, a period type related to a preset month, day, day of the week, weekday/holiday, and/or time period, which corresponds to the period type to which the estimated time point of the distribution belongs.

In the above embodiment, it is also preferable that the presence distribution determination means performs kernel density estimation using information including the terminal's position and time, which is the positioning result accumulated during the first specified period, to determine a normal presence distribution that is a function of the position and the time within the first specified period or within the period of the set period type.

Further, in the above embodiment, the abundance distribution determining means
It is also preferable to perform one or both of a process of determining the normal presence count distribution by assigning a greater weight to the positioning result the more recent the positioning time point related to the positioning result is within a first specified period, and a process of determining the estimated presence count distribution by assigning a greater weight to the positioning result the more recent the positioning time point related to the positioning result is within the first specified period and the second specified period.

In the above embodiment, the abundance distribution determining means
It is also preferable to perform one or both of the following processes: a process of setting, for each positioning result accumulated during a first specified period, at least one presence point corresponding to the positioning result, the sum of which is weighted to 1, within an area having a size that is a monotonically increasing function of the positioning error included in the positioning result or determined from the positioning result, and determining the normal presence number distribution from the presence point; and a process of setting, for each positioning result acquired during the first specified period and the second specified period, at least one presence point corresponding to the positioning result, the sum of which is weighted to 1, within an area having a size that is a monotonically increasing function of the positioning error included in the positioning result or determined from the positioning result, and determining the estimation-time presence number distribution from the presence point.

Further, in the above embodiment, the abundance distribution determining means
determining a subperiod presence number distribution as a distribution of the number of objects present at a representative time point of each subperiod from the positioning result acquired in the subperiod for each subperiod among a plurality of subperiods included in the first predetermined period and the second predetermined period;
It is also preferable to extrapolate the sub-period presence count distribution at the representative time point to the presence count distribution at the estimation time point, and determine the presence count distribution obtained by extrapolation as the presence count distribution at the estimation time.

According to the present invention, there is also provided an object distribution estimation device for estimating a distribution of a plurality of objects in a predetermined area, comprising:
a presence distribution determining means for determining a presence distribution as a distribution of the number of objects present in the predetermined area from a positioning result obtained during a predetermined period for a terminal attached to the object;
a correct answer magnification coefficient determination means for determining a correct answer magnification coefficient for each measurement point from measurement results obtained from a plurality of real number measurement means capable of actually measuring the number of the target at each measurement point within the specified area and the presence number distribution;
a feature generating means for generating, for each of a plurality of points within the predetermined region, a feature characterizing the point or an area including the point;
an expansion coefficient estimation means for inputting the feature quantity to an expansion coefficient estimation model constructed using the correct expansion coefficient, and estimating an expansion coefficient of the point from an output of the expansion coefficient estimation model;
There is provided an object distribution estimation device having an object presence distribution determination means for determining the number of objects present at each of a plurality of points within the specified area based on the presence distribution and the estimated expansion coefficient.

According to the present invention, there is further provided an object distribution estimation system for estimating a distribution of a plurality of objects in a predetermined area, comprising:
a presence distribution determining means for determining a presence distribution as a distribution of the number of objects present in the predetermined area from a positioning result obtained during a predetermined period for a terminal attached to the object;
a correct answer magnification coefficient determination means for determining a correct answer magnification coefficient for each measurement point from measurement results obtained from a plurality of real number measurement means capable of actually measuring the number of the target at each measurement point within the specified area and the presence number distribution;
a feature generating means for generating, for each of a plurality of points within the predetermined region, a feature characterizing the point or an area including the point;
an expansion coefficient estimation means for inputting the feature quantity to an expansion coefficient estimation model constructed using the correct expansion coefficient, and estimating an expansion coefficient of the point from an output of the expansion coefficient estimation model;
An object distribution estimation system is provided, the object distribution estimation system having an object presence distribution determination means for determining the number of objects present at each of a plurality of points within the specified area based on the presence distribution and the estimated magnification coefficient.

According to the present invention, there is further provided a computer-implemented object distribution estimation method for estimating a distribution of a plurality of objects in a predetermined area, the method comprising:
A step of determining a presence distribution as a distribution of the number of objects present in the predetermined area from positioning results obtained during a predetermined period for a terminal attached to the object;
determining an expansion coefficient of a correct answer for each measurement point from measurement results obtained from a plurality of real number measurement means capable of measuring the number of the target at each measurement point within the specified area and the distribution of the number of targets present; and generating, for each of the plurality of points within the specified area, a feature quantity that characterizes the point or an area including the point;
a step of inputting the feature quantity into an expansion coefficient estimation model constructed using the correct expansion coefficient, and estimating an expansion coefficient of the point from an output of the expansion coefficient estimation model;
and determining a count of the objects at each of a plurality of locations within the predetermined area based on the count distribution and the estimated expansion factor.

The subject distribution estimation program, device, system, and method of the present invention make it possible to estimate the distribution of subjects with high resolution and greater precision (accuracy) without relying on ground-truth data relating to the actual distribution of the subjects.

1 is a functional block diagram showing a functional configuration of an embodiment of a target distribution estimation device according to the present invention. 11A and 11B are schematic diagrams for explaining another embodiment of the normal presence distribution determination process according to the present invention. 13 is a schematic diagram for explaining still another embodiment of the normal presence distribution determination process according to the present invention. FIG. 11 is a schematic diagram for explaining another embodiment of the estimation-time existence distribution determination process according to the present invention. FIG. 1 is a schematic diagram showing an overview of an embodiment of a target distribution estimation method according to the present invention; FIG. 1 is a schematic diagram for explaining a conventional problem in generating a population distribution heat map using a high-resolution mesh.

The following describes in detail an embodiment of the present invention with reference to the drawings.

[Target distribution estimation device]
FIG. 1 is a functional block diagram showing the functional configuration of an embodiment of an object distribution estimation device according to the present invention.

The communication equipment device 1 shown in FIG. 1 is of course a device that performs the functions of communication equipment (e.g., a relay function), but it is also one embodiment of an object distribution estimation device according to the present invention. That is, the communication equipment device 1 is a device that estimates the distribution of multiple objects in a specified area (e.g., the entire country of 47 prefectures), or in this embodiment, the distribution of a large number of people (i.e., population distribution), and outputs population distribution data (e.g., a population distribution graph or a population heat map) as the estimation result.

Specifically, the communication equipment 1 includes:
(a) "terminal positioning result" which is a positioning result for a terminal 2 (e.g., a carried smartphone) attached to a human (target), for example, (a1) GPS positioning data from a GPS (Global Positioning System) function installed in the terminal 2, (a2) base station positioning data by a base station positioning method using a base station 3 or a smart pole 4 equipped with a base station function, and (a3) at least one of positioning data by a wireless LAN (Local Area Network) access point when the terminal 2 is equipped with a wireless LAN communication function such as Wi-Fi (registered trademark), and
(b) “Actual measurement results” obtained from a plurality of real measurement means capable of measuring the number of people (targets) at a plurality of measurement points within a specified region (for example, the entire country). In this embodiment, number of people (count number) data obtained from smart poles 4 is obtained, and the population distribution in the specified region (the entire country) is estimated based on these data.

In order to perform such estimation, the communication equipment 1
(A) a presence count distribution determination unit 111 that determines a “presence count distribution” as a distribution of the number of objects present in a predetermined area (e.g., nationwide) from “terminal positioning results” acquired during a predetermined period;
(B) a correct answer expansion coefficient determination unit 112 that determines a correct answer expansion coefficient for a measurement point based on the actual measurement results and the presence distribution at the measurement points;
(C) a feature amount generating unit 113 that generates, for each of a plurality of points within a predetermined region (nationwide), a “feature amount” that characterizes the point or an area including the point;
(D) an enlargement coefficient estimation unit 115 that inputs the “feature value” to an “enlargement coefficient estimation model” constructed using the “correct enlargement coefficient” and estimates the “enlargement coefficient” of the point from the output of the “enlargement coefficient estimation model”;
(E) A target presence distribution determination unit 116 that determines the number of targets (people) present (the number of people present) at each of a plurality of points within a predetermined region (nationwide) based on the "number of people distribution" and the estimated "magnification coefficient." Here, in this embodiment, the target presence distribution determination unit 116 of (E) generates and outputs population distribution data (for example, a population distribution graph or a population heat map) as an estimation result from the determined number of people present at each point.

In this way, the communication equipment device 1 determines the "correct answer expansion coefficient" using the "actual number measurement result" obtained from the real number measurement means (smart pole 4), estimates the "expansion coefficient" using the "expansion coefficient estimation model" constructed from this "correct answer expansion coefficient", and thereby determines the target presence count (number of people present) at each point.

Even in a situation where the number of measurement points is limited, by setting a sufficient number of points in a specified region (nationwide) in advance, it is possible to obtain high-resolution target distribution data (population distribution data). In addition, since the "correct answer expansion coefficient" is determined from the "actual measurement results" and the "presence distribution," it is possible to improve the precision (accuracy) of the "expansion coefficient" related to the output of the "expansion coefficient estimation model" by setting an appropriate "feature value." Furthermore, this "expansion coefficient" is different from the conventional simple and uniform one such as (population)/(number of users authorized to provide location information), and is a more suitable one that reflects the characteristics (that affect the actual number of people) at that point.

As described above, the communication equipment device 1 makes it possible to estimate the distribution of targets (people) with high resolution and higher precision (accuracy) without relying on ground-truth data relating to the actual distribution of targets (people).

The smart pole 4, which is the real number measurement means in this embodiment, is installed, for example, along a sidewalk and is equipped with a camera for analyzing people flow and an image recognition function, and is capable of counting the number of people (objects) present around the installation point (for example, passing through the installation point) and outputting, for example, the number of people (count number, real number measurement result) for each measurement time interval. In addition, the smart pole 4 may also have a base station function for a mobile phone communication network, an access point function for a wireless LAN such as Wi-Fi (registered trademark), or even a digital signage function. Incidentally, it is also possible to obtain "real number measurement results" using real number measurement means other than smart poles, such as human presence sensors and cameras installed at each measurement point.

In this embodiment, the target for which distribution is sought is "people", and in the following description, population distribution (people flow) is estimated for "people", but the target according to the present invention is not limited to "people". In other words, as long as the target is equipped with a terminal 2 that can obtain "terminal positioning results" and the number of people can be measured by some real number measurement means, various things can be used as the target according to the present invention. For example, if at least some automobiles are equipped with terminals 2, and traffic counters (equipped with cameras and image recognition functions), vehicle detection sensors (installed along roads), and even surveillance cameras are installed at multiple points within a specified area, at least one of these can be used as the real number measurement means, and "automobiles" can be the target according to the present invention.

[Device functional configuration, target distribution estimation system, program, and method]
Similarly, according to the functional block diagram of FIG. 1, a communication equipment 1 as an embodiment of an object distribution estimation device according to the present invention has a communication interface unit 101, a positioning result storage unit 102, an actual measurement result storage unit 103, a map information storage unit 104, an attribute information storage unit 105, a user interface unit 106, and a processor/memory (a calculation processing system with a memory function).

Here, this processor memory stores one embodiment of the target distribution estimation program according to the present invention, and has computer functions, and performs the target distribution estimation process by executing this target distribution estimation program. For this reason, the communication equipment device 1 may be a device dedicated to the target distribution estimation process other than communication equipment, and further, it may be a cloud server, a non-cloud server, a personal computer (PC), a notebook or tablet computer, or even a mobile terminal such as a smartphone, equipped with the target distribution estimation program according to the present invention.

The processor memory further includes an existence distribution determination unit 111 including a normal existence distribution determination unit 111a and an estimated existence distribution determination unit 111b, a correct answer expansion coefficient determination unit 112, a feature generation unit 113, a model construction unit 114, an expansion coefficient estimation unit 115, a target existence distribution determination unit 116, and a communication control unit 121. The functional components described above can be considered as functions realized by executing a target distribution estimation program stored in the processor memory. The process flow shown by connecting the functional components of the communication equipment 1 in FIG. 1 with arrows can also be understood as one embodiment of the target distribution estimation method according to the present invention.

Incidentally, at least one of the existence number distribution determination unit 111, the correct answer expansion coefficient determination unit 112, the feature generation unit 113, the model construction unit 114, the expansion coefficient estimation unit 115, and the target existence distribution determination unit 116 may be a functional component of another device, and the target distribution estimation function may be performed by a plurality of devices (e.g., a plurality of servers) as a whole. In this case, the plurality of devices as a whole constitutes the target distribution estimation system according to the present invention. Furthermore, among these, the model construction unit 114 may be handled by a device separate from the target distribution estimation device/system, and the target distribution estimation device/system may acquire a trained expansion coefficient estimation model from this separate device.

Below, we will omit a description of the communication control function (e.g., relay function) in the communication equipment 1, and only explain the target distribution estimation processing function according to the present invention. Also in FIG. 1, the positioning result storage unit 102 acquires the terminal positioning results (GPS positioning data, etc.) of the above-mentioned terminal 2 from the terminal 2 via the communication interface unit 101 and communication control unit 121, and stores and manages them. Here, the terminal positioning results are positioning data only for users who have contracted with the communication carrier and have consented to the provision of terminal position information, and are data only for a portion of the total population that is to be the target of the distribution to be estimated.

The real number measurement result storage unit 103 acquires the above-mentioned real number measurement result, which in this embodiment is the number of people (count number) data output from the smart pole 4, from the smart pole 4 via the communication interface unit 101 and the communication control unit 121, and stores and manages it. Here, this real number measurement result is the result of accurately measuring the number of all people present around the measurement point where the smart pole 4 is installed (e.g. passing through the measurement point), but the number (number of measurement points) is limited (it is not acquired for any point).

<Existence number distribution determination means: Normal existence distribution determination unit>
Similarly, in the functional block diagram of FIG. 1, the normal presence distribution determination unit 111a of the presence distribution determination unit 111 stores terminal positioning results (positioning data) accumulated during a first predetermined period (for example, one month) in a positioning result storage unit The terminal positioning results are acquired from 102, and a "normal presence number distribution" is determined from the acquired terminal positioning results as a normal (i.e., normal, peacetime) distribution of the number of people present.

Here, the reason why the terminal positioning results accumulated over a long first predetermined period (for example, one month) are used is that the terminal positioning results are not the entire population but only the positioning data of some users, as described above, and in particular in sparsely populated suburbs and rural areas, positioning data cannot be obtained at many points, resulting in a population of zero at those points. In other words, by setting a long first predetermined period (for example, one month), if it is possible to measure even one person per month, for example, it is possible to set the number of people per day to a finite value exceeding zero, such as about 0.03 (≒1/30). As a result, it is possible to more accurately grasp the shading of, for example, the number of people at each point. Incidentally, this can be said to be a similar method to taking a camera picture in the dark by lengthening the exposure time (light accumulation time) to capture more light into the camera sensor and obtain a camera image with more accurate shading.

The normal presence distribution determination unit 111a regards the acquired (accumulated) positioning data group as a point group in the presence distribution space, and in this embodiment, determines the normal presence number distribution CE A (x), which is a function of a point (e.g., latitude and longitude) x in a specified region (nationwide), using kernel density estimation (KDE), which is a known _{method for estimating the entire distribution from a finite number of sample points. Specifically, the normal presence number distribution CE A (x) is expressed by the following formula (1): CE A ( x} ) = P _A (x) × V _A (x)
Here, P _A (x)=(nh) ^-1 Σ _i=1 ⁿ K((xx _i )/h)
is determined by.

Here, _x1 , _x2 , ..., _xn are n pieces of positioning data as sample points, and for example, _xi (i = 1, 2, ..., n) can be _xi = (latitude, longitude). h is a bandwidth (smoothing parameter). Also, P _A (x) is the probability density distribution of the presence of a human (target) in the first predetermined period. Furthermore, K is a kernel function, and can be, for example, a standard Gaussian function (with a mean of zero and a variance of 1) of the following equation (2).
(2) K(x')=(2π) ^-0.5 exp(-x' ² /2)

Furthermore, in the above formula (1), V _A (x) is a volume determined using the number of positioning data groups accumulated in a first predetermined period, and is a so-called "conversion rate" when converting the probability density value (the value of P _A (x)) into a number of people (number of objects). In other words, the normal presence number distribution _CEA (x) is the presence probability density distribution P _A (x) multiplied by this volume V _A (x), and can be regarded as the distribution of the number of people (number of objects) related to the positioning data. Also, the value of the normal presence number distribution _CEA (x) at point x (normal presence number) is the expected value of the number of people (related to the positioning data) at point x.

By determining the normal population distribution CE _A (x) in this way, it is possible to obtain the number of people at any point x in a given region (nationwide). Incidentally, KDE is implemented as a standard function model in many numerical analysis and programming platforms such as MATLAB (registered trademark), which are used by many engineers and researchers in this field. Of course, if a function allows the entire distribution to be estimated from a finite number of sample points, that function may be used instead of KDE to determine the normal population distribution.

In addition, it is also preferable that the normal presence distribution determination unit 111a sets, as the first predetermined period, a period type related to at least one of a preset month, date, day of the week, weekday/holiday, and time period, which corresponds to the period type to which the population distribution estimation time belongs. For example, if the population distribution estimation time is "between 8:00 on Wednesday", the first predetermined period may be set to "between 8:00 on Wednesday in the past M months (M is a positive integer)". This makes it possible to reflect circumstances specific to the period type, such as the number of people commuting to work or school being the highest in a certain area during the "between 8:00 on Wednesday", in the normal presence number distribution, thereby making it possible to perform a more accurate (precise) population distribution estimation. Note that a normal presence number distribution related to each required period type may be prepared in advance.

Furthermore, the normal presence distribution determination unit 111a determines the positioning data x _i (i=1, 2, . . . , n) as follows:
(3) x _i = (latitude, longitude, time)
may be adopted to determine the normal presence distribution CE _A (x) in the three-dimensional presence distribution space. Here, the range (domain) of "time (point in time)" may be the entire first predetermined period, or may be a period of a preset period type. This makes it possible to obtain the normal presence distribution CE _A (x) that is a function of the position (latitude, longitude) and the time (point in time) within the first predetermined period or within the period of the preset period type.

For example, by setting the range (domain) of "time (point in time)" to the period "between 8:00 and 8:00 on Wednesdays" in a first specified period (e.g., August), it is possible to obtain a normal presence number distribution that can continuously express the difference in the number of targets due to differences in time (point in time) within this period. In this case, it is not necessary to prepare a large number of normal presence number distributions in advance to match the period type at the time of estimating the population distribution; for example, it is often sufficient to prepare just one normal presence number distribution whose range (domain) includes the period of the expected period type.

In addition, it is also preferable that the normal presence distribution determination unit 111a determines the normal presence distribution CE _A ( _x ) by assigning a larger weight to the positioning data x _i (i=1, 2,..., n) whose positioning time point is newer within the first predetermined period. Hereinafter, this embodiment will be described with reference to Fig. 2 (a schematic diagram for explaining another embodiment of the normal presence distribution determination process).

2, when the normal presence distribution determiner 111a determines the normal presence number distribution CE A (x) from the positioning data x _i (i = 1, 2, ..., n) accumulated during the first predetermined period _M (the period from time (point in time) t0 to time (point in time) t1), the more recent the positioning time (positioning time point) tm of the positioning data x _i , the more weight is assigned to this positioning data x _i . Specifically, when inputting the positioning data (as sample points) into the KDE model to generate and determine the normal presence number distribution CE _A (x) of the above formula (1), the following formula (4) W _T = (tm - t0) / (t1 - t0) is applied to the positioning data x _i (i = 1, 2, ..., n).
Alternatively, data W _T ×x _i multiplied by a weight W _T calculated by the following formula may be input. This makes it possible to obtain a normal population distribution that is closer to the actual situation, and to improve the accuracy of the final population distribution estimation. Note that the weight is not limited to that in formula (4) above, and various other weights can be used as long as they are monotonically increasing functions of the positioning time (positioning point in time) tm.

Furthermore, it is also preferable that the normal presence distribution determination unit 111a, for each piece of positioning data xi (i = 1, 2, ..., n) (terminal positioning result) accumulated during the first predetermined period, sets W presence points (sample points) corresponding to the positioning data, W (W is a positive integer) with "weights" assigned to them that sum to 1, within an area (e.g., a circular area with the positioning error as its radius) whose size is a monotonically increasing function of the positioning error contained in or determined from the positioning data, and determines the normal presence number distribution from the presence points (sample points).

The positioning error here may be the one that is normally assigned to the acquired positioning data. Alternatively, an appropriate positioning error may be determined in advance depending on whether the acquired positioning data is GPS positioning data, base station positioning data, or Wi-Fi (registered trademark) positioning data. This embodiment will be described below with reference to FIG. 3 (a schematic diagram for explaining yet another embodiment of the normal presence distribution determination process).

As shown in FIG. 3, in this embodiment, the normal presence distribution determination unit 111a sets a circular area for one piece of positioning data, centered on the position (latitude, longitude) of the data and with the positioning error as its radius, and generates and arranges W presence points (sample points) each assigned a weight of 1/W within this circular area (for example, 100 presence points with a weight of 0.01 are generated and arranged). Here, it is also preferable to arrange the W presence points (sample points) set so that the closer they are to the center of the circular area, the higher the density. Furthermore, it is also preferable to assign weights so that the values of each weight are not uniform, but rather the closer to the center of the circular area, the higher the weight is assigned (the sum is 1).

In any case, if the positioning data group is then input into a KDE model to determine the normal abundance distribution, it is possible to improve the precision (accuracy) of the determined normal abundance distribution by increasing the number of reliable sample points. Here, the smaller the positioning error, the smaller the correct sample points are thought to be distributed within a smaller circular area (further closer to the center), so by increasing the number of sample points using the method described above, it is expected that the precision (accuracy) of the normal abundance distribution will be improved. Note that the presence points (sample points) within the circular area can be set using known sampling methods other than those described above.

<Existence number distribution determination means: Estimation existence distribution determination unit>
Returning to the functional block diagram of FIG. 1, the estimation-time presence distribution determiner 111b of the presence distribution determiner 111 is
(a) the positioning results (positioning data) accumulated during the first predetermined period, and (b) a second predetermined period that is shorter than the first predetermined period and immediately precedes the distribution estimation time point. (For example, the current time is the most recent distribution estimation time, i.e. the past hour)
from the positioning result storage unit 102, and from the acquired terminal positioning result, a "distribution of numbers present at the time of estimation" is determined as a distribution of numbers present at or immediately before the distribution estimation time.

As mentioned above, this estimated presence distribution is generated and determined from the positioning data of only a portion of users, not the entire population, but is close to the presence distribution at the time of distribution estimation (e.g., the current time), and is therefore suitable as the presence distribution to which the magnification factor should be multiplied thereafter. Here, since the positioning data around the time of distribution estimation (the current time) alone is too few sample points, a certain period (e.g., the most recent hour) is set as the second specified period depending on the positioning interval and amount of data, and the positioning data for this period (and also the first specified period) is used.

In this embodiment, the estimated presence distribution determination unit 111b regards the group of positioning data accumulated in the first specified period and the group of positioning data acquired in the second specified period as point groups in a presence distribution space, and uses KDE to determine the estimated presence distribution CE _B (x), which is a function of point (e.g., latitude and longitude) x in a specified region (nationwide), in a similar manner to the procedure for determining the normal presence distribution CE _A (x) described above.

More specifically, the estimation-time presence distribution determination unit 111b sets W _A and W _B as weights that satisfy the condition "W _A + W _B = 1" that is appropriately set in advance, and sets T _A and T _B as the length of a first predetermined period (e.g., 720 hours (= 24 hours/day × 30 days)) and the length of a second predetermined period (e.g., 1 hour), respectively, and
(a) multiplying each point (positioning data) of the positioning data group (point group) accumulated during a first predetermined period by a weight W _A /T _A ;
(b) multiplying each point (positioning data) of the positioning data group (point cloud) acquired during the second predetermined period by a weight W _B /T _B ;
Using the weighted positioning data set ( _x1 , _x2 , ..., _xn' ), the following equation (5) CE _B (x) = P _B (x) × V _B (x) is calculated, which is similar to the above equation (1).
Here, P _B (x)=(n'h) ^-1 Σ _i=1 ^n' K((xx _i )/h)
The distribution of the number of individuals present at the time of estimation, CE _B (x), is calculated as follows:

In the above formula (5), P _B (x) is the presence probability density distribution of humans (objects) in the first and second predetermined periods. V _B (x) is a volume determined using the number of positioning data groups acquired in the second predetermined period, and is a so-called "conversion rate" when converting the probability density value (the value of P _B (x)) into the number of people (number of objects). In other words, the presence number distribution CE _B (x) at the time of estimation is the presence probability density distribution PB ( _x ) multiplied by this volume V _B (x), and can be regarded as the distribution of the number of people (number of objects). The value of the presence number distribution CE _B (x) at the time of estimation at point x is the expected value of the number of people at point x. Incidentally, the presence number distribution CE _B (x) at the time of estimation is used to estimate the number of people (number of objects) at the time of distribution estimation after this, so in this embodiment, it is calculated by multiplying it by V _B (x) (as a volume close to the time of distribution estimation).

Moreover, the estimation time presence distribution determination unit 111b
(a) for each subperiod in a plurality of subperiods (e.g., three one-hour periods in the most recent three hours) included in a second predetermined period (e.g., the most recent three hours), a "subperiod presence count distribution" is determined as a presence count distribution at a representative time point of the subperiod (e.g., a midpoint of the subperiod) from the positioning data (terminal positioning results) acquired in the subperiod;
(b) It is also preferable to extrapolate the "sub-period presence distribution" at the representative time to the presence distribution at the distribution estimation time, and determine the presence distribution obtained by extrapolation as the presence distribution at estimation time. Note that in the above (a), it is also possible to set a plurality of sub-periods from among a first predetermined period (e.g., one month) and a second predetermined period. Hereinafter, this embodiment will be described with reference to FIG. 4 (a schematic diagram for explaining another embodiment of the presence distribution determination process at estimation time).

In the example shown in FIG. 4, a first subperiod is set, which includes the distribution estimation time t(0) and has a representative time (the midpoint of the subperiod in FIG. 4) of t(-1) (<t(0)), and a second subperiod is set, which has a representative time t(-2) (<t(-1)). Here, the presence count distribution CE _-2 (x) (generated using KDE) in the second subperiod is considered to be data at approximately time t(-2), while the presence count distribution CE _-1 (x) (generated using KDE) in the first subperiod is considered to be data at approximately time t(-1). Therefore, the presence count distribution at the distribution estimation time t(0) is approximately calculated using an extrapolation process from the presence count distributions at these two times, and this is used as the presence count distribution at the estimation time CE _B (x).

Specifically, for example, the following formula (6) CE _B (x) = CE _-1 (x) + (t(0) - t(-1)) × ((CE _-1 (x) - CE _-2 (x)) / (t(-1) - t(-2)))
Using this, the estimation-time presence distribution CE _B (x) can be generated. Of course, even if there are three or more subperiods, the estimation-time presence distribution CE _B (x) can be generated in a similar manner using known extrapolation processing. Also, instead of linear extrapolation processing, it is possible to use higher-order extrapolation processing (such as spline extrapolation). In any case, such extrapolation processing makes it possible to obtain the estimation-time presence distribution CE _B (x) that is closer to the distribution estimation time t(0), which also makes it possible to improve the accuracy of the final population distribution estimation.

Furthermore, in the same manner as described for the normal presence distribution CE _A (x) using FIG. 2, when the estimation-time presence distribution CE _B (x) is determined from the positioning data of the first and second predetermined periods, it is preferable that the more recent the positioning time (positioning time) of the positioning data is (the closer it is to the distribution estimation time), the greater the weighting is given to this positioning data. Specifically, when inputting the positioning data of the first predetermined period (multiplied by the weight W _A /T _A ) and the positioning data of the second predetermined period (multiplied by the weight W _B /T _B ) into the KDE model to generate and determine the estimation-time presence distribution CE _B (x), data obtained by further multiplying these positioning data by the weight W _T ' similar to the above formula (4) may be input. This makes it possible to obtain the presence distribution at the time of estimation that is closer to the distribution estimation time, and to improve the accuracy of the final population distribution estimation.

Furthermore, in the same manner as described for the normal presence count distribution CE _A (x) using Fig. 3, it is also preferable that the estimation-time presence distribution determiner 111b sets, for each positioning data of the first and second predetermined periods, W'(W' is a positive integer) presence points corresponding to the positioning data, W' presence points (sample points) to which "weights" are assigned such that the sum of the W' presence points is 1, within an area (for example, a circular area with the positioning error as a radius) whose width is a monotonically increasing function of the positioning error contained in the positioning data or determined from the positioning data, and determines the estimation-time presence count distribution from the W' presence points (sample points). Increasing the number of highly reliable sample points in this manner makes it possible to improve the precision (accuracy) of the determined estimation-time presence count distribution.

The method of determining the presence count distribution CE _B (x) at the time of estimation is not limited to the above. For example, it is possible to determine the presence count distribution CE _B (x) at the time of estimation in the same manner as described above using only the positioning data for the second specified period. However, by also using the positioning data for the first specified period as described above, it is possible to derive a more appropriate presence count distribution CE _A (x) as a presence count distribution that reflects the distribution trend at the time of distribution estimation and the normal distribution trend.

Here, we summarize the abundance distribution determination process in the abundance distribution determination unit 111 of this embodiment described above. Figure 5 is a schematic diagram showing an overview of one embodiment of the target distribution estimation method according to the present invention.

As shown in FIG. 5, in this embodiment, the abundance distribution determination unit 111
(a) determining a normal presence distribution CE _A (x) using KDE from positioning data (terminal location information) for a first predetermined longer period;
(b) From the positioning data (terminal position information) in the first predetermined period and a shorter second predetermined period (including the distribution estimation time point), an estimated presence distribution CE _B (x) is determined using KDE.

Here, in this embodiment, the normal presence count distribution CE _A (x) is then used in the answer expansion coefficient determination unit 112 to determine the answer expansion coefficient, while the estimated presence count distribution CE _B (x) is then used in the target presence distribution determination unit 116 to generate target presence distribution data (e.g. a high-resolution population distribution graph or a high-resolution population heat map).

<Correct answer expansion coefficient determination means>
Returning to the functional block diagram of FIG. 1, the correct answer enlargement coefficient determination unit 112
(a) actual count data c(x) (actual measurement results) during a specified measurement period obtained from smart poles 4 (actual measurement means) installed at multiple measurement points x within a specified region (e.g., nationwide); ,
(b) A correct answer expansion coefficient R _C (x) for the measurement point is determined from the presence count distribution determined by the presence count distribution determination unit 111, which in this embodiment is the normal presence count distribution _CEA (x).

Specifically, in this embodiment, the correct answer expansion coefficient determination unit 112 determines the correct answer expansion coefficient R _C (x) at the point x according to the following formula (7): R _C (x)=c(x)/CE _A (x)
Here, CE _A (x) is a normal existence number distribution. That is, the correct answer expansion coefficient determination unit 112 defines the expansion coefficient of the "correct answer" as the ratio of the count number (which can be said to be the number of correct answers (total number of people) at the measurement point x) to the existence number (related to only some users), and obtains the correct answer expansion coefficient according to this definition. In addition, this makes it possible to obtain a correct answer expansion coefficient R _C (x) for each of multiple (many) measurement points x, and by using these (many) correct answer expansion coefficients R _C (x) as learning data, it becomes possible to construct (train) an expansion coefficient estimation model thereafter.

As mentioned above, the actual count data c(x) must be considered to indicate the correct number of people (total number of people) at point x. Therefore, the actual count data at a measurement point where, for example, multiple underground passages and above-ground walkways overlap vertically, or where multiple underground (express) roads and above-ground (express) roads overlap vertically, must be handled with care, for example by excluding that data or adding data that is in a vertical relationship.

Here, it is also preferable that the measurement period of the actual count data c(x) and the first predetermined period related to the normal abundance distribution CE _A (x) are set to the same length. Alternatively, the first predetermined period (period related to input to the KDE) may be set arbitrarily, and the value of the volume V _A (x) in the above formula (1) may be adjusted to match this first predetermined period and the measurement period of c(x). For example, if the measurement period of c(x) is 10 days,
(a) the first predetermined period may be ten days; or
(b) When the first predetermined period is 30 days, the input to the KDE may remain the same for 30 days, but the value of the volume V _A (x) may be the value for 10 days out of the 30 days, or the volume value for 30 days may be multiplied by 1/3 to obtain V _A (x).

<Feature Generation Means>
1, the feature generator 113 generates, for each of a plurality of points x in a predetermined region (e.g., the whole country), a feature F(x) that characterizes the point or an area including the point. In constructing (training) an expansion coefficient estimation model to be described later, the feature F(x) at the measurement point x where the correct expansion coefficient R _C (x) exists is generated as learning data.

The "expansion coefficient" derived by the expansion coefficient estimation model is a quantity that reflects the proportion of "users who have contracted with a telecommunications carrier and have consented to the provision of their terminal location information" to the total population, as described above. Whether or not to "contract with a telecommunications carrier" and whether or not to "convert their terminal location information" is up to the thinking and judgment of each individual (individual target), that is, it is highly dependent on the attribute information of the individual (target). Therefore, it is considered appropriate to use the feature quantity F(x), which is the explanatory variable for deriving the "expansion coefficient", as a quantity that expresses the attribute information of individuals present at point x (or in its vicinity), or a quantity that interacts with the attribute information to affect the degree to which individuals are present there.

The following describes the features generated by the feature generator 113: (a) features based on map information, (b) features based on target attribute information, and (c) features taking into account the context of movement. Of course, the features according to the present invention are not limited to these, and various features can be used as long as they are quantities that express the attribute information of individuals or quantities that interact with the attribute information to affect the degree to which individuals are present.

(Features based on map information)
It is also preferable that the feature generating unit 113 generates a feature F(x) of the point x or an area including the point x (a circular area having a predetermined radius centered on the point x) from information related to a specific thing or phenomenon related to the point x or an area including the point x (for example, a circular area having a predetermined radius centered on the point x) based on map information of a predetermined region (for example, the whole country). Here, as the map information, for example, information distributed from the map information management server 5 and stored in the map information storage unit 104 may be used.

Specifically, a multidimensional vector whose elements are the number of POIs (Points of Interest) in each (predefined) category included in an area (circular region) including point x, for example, the following formula (8): F(x) = (number of office properties, number of commercial properties, number of transportation-related properties,
Number of residential properties, number of public facility properties, etc.)
can be used as the feature F(x).

For example, even if two locations are spatially close to each other, the attributes of the people present in an office and a nursery school are significantly different, and as a result, the magnification coefficients are also thought to be significantly different. Therefore, as in formula (8) above, by using as a feature what kind of place point x or the area (circular area) containing point x is from the perspective of a POI, it becomes easier to more appropriately derive the magnification coefficient of point x. In addition to features such as those in formula (8) above, it is possible to adopt features based on various map information as long as they are quantities related to things or events that exist or occur at point x or the area (circular area) containing point x, and interact with the attribute information of individuals to affect the degree to which individuals are present there.

(Features based on target attribute information)
The feature amount generating unit 113 generates attribute information of a user (an object to which the terminal 2 is attached) who owns a terminal 2 located at the point x or an area including the point x (for example, a circular area having a predetermined radius centered on the point x),
It is also preferable to generate a feature quantity F(x) of the point x or an area (circular area) including the point x from (a) attribute information estimated from historical information of positioning data (terminal positioning results) or (b) attribute information acquired from registered information of the user (target) (e.g., subscriber information held by a telecommunications carrier). In other words, the type of person present (staying) in the point x or the area (circular area) including the point x is characterized.

As the attribute information of (a) above, for example, time series data of the user's movement and stay can be generated from historical information of positioning data using a publicly known movement/stay determination method (*), and the regional mesh (a unit area obtained by dividing a specified region (the whole country) into a number of unit areas, hereafter abbreviated as mesh) in which the user has the longest stay time (highest probability of stay) during a set "nighttime" (for example, a time period from 1:00 to 5:00) can be used as the residential area of the user's attribute information.
(*) For example, the method disclosed in non-patent document: Nao Kobayashi, Hiroki Ishizuka, Atsunori Minamikawa, Shigeki Muramatsu, Tomohiro Ono, “Proposal of a method for estimating movement and stay status suitable for mobile phone communication history,” IPSJ Transactions on Information Processing Society of Japan, Database, vol. 10, no. 1, pp. 13-23, 2017, can be used.

In this case, specifically, the feature generating unit 113 may assume that the selection of a communication carrier and the presence or absence of location information permission are related to the regionality of the residence (address, location of home) of the visitor (user), and generate a feature (vector) with elements of the breakdown of the residence (home) area for a group of users staying in an area (circular area) including point x. For example, the following formula (9) F(x) = (..., proportion of users whose residence is in Tokyo, ...,
Percentage of users who live in Kanagawa Prefecture, ...)
can be used as the feature F(x).

The attribute information in (b) above may be received from an attribute information management server 5 (managed by a telecommunications carrier, for example) and stored in the attribute information storage unit 105. Incidentally, the residence information in the above formula (9) can also be obtained from the attribute information storage unit 105 (attribute information management server 5).

Specifically, the feature generating unit 113 may estimate movement and stay from the history information of the positioning data using the above-mentioned known technology, and then generate a feature (vector) with the ratio of user attributes of each (pre-set) attribute category for a group of users staying in an area (circular area) including point x as an element. For example, gender and age are adopted as attribute categories, and the following formula (10) F(x)=(proportion of men in their teens, proportion of men in their twenties, ...,
Percentage of women in their teens, ..., Percentage of women in their 70s, ...)
can be used as the feature F(x).

(Features that take into account the context of movement)
The feature generation unit 113 regards mesh codes, which are codes representing meshes (unit areas obtained by dividing a specific region (the whole country) into multiple unit areas), as words, and regards the time series data of the mesh codes representing the movement history of a user (target) as sentences.The feature generation unit 113 is a natural language processing model that learns the similarities between words and predicts the next word from a word string, and may use a natural language processing model trained on the history of some masked positioning results to obtain an embedded representation (distributed representation) of a mesh including point x, and use this as the feature F(x).

For example, according to the technology disclosed in the non-patent document: Kobayashi Akihiro, Uesaka Daisuke, Takeda Naoto, Minamikawa Atsunobu, “Representation Learning Technology for Traffic Demand Forecasting Using Location Information Big Data”, Civil Engineering Planning Research Lectures, Vol. 64, CD: 8 pages in total, 2021, BERT (Bidirectional Encoder Representations from Transformers), which is known as a natural language processing model, is adopted, a large amount of time-series data of mesh codes representing the user's movement history is prepared, some of the mesh codes in the time-series data (sentences) are masked, and BERT is trained to predict this, thereby acquiring an embedding representation (distributed representation) of each mesh (word) through unsupervised (self-supervised) learning.

This mesh embedding representation can be a multidimensional vector in which meshes with more similar features have closer values, and can therefore be used as the feature F(x) of the present invention. For example, when set to 512 dimensions, a feature F(x) of the form F(x) = (f1, f2, ..., f512) can be generated.

Above, we have explained (a) features based on map information, (b) features based on target attribute information, and (c) features taking into account the context of movement. However, two or all of these may be combined (e.g., joined) to generate the feature F(x) for point x.

<Model building method>
Also in the functional block diagram of FIG. 1, a model construction unit 114 constructs an expansion coefficient estimation model (regression model) that estimates an expansion coefficient R(x)^, which is a response variable, using the feature quantity F(x) generated by the feature quantity generation unit 113 as an explanatory variable, in other words, the following equation (11): R(x)^=f(F(x))
An enlargement coefficient estimation model f(·) represented by the following formula is constructed (trained). As an algorithm for constructing this enlargement coefficient estimation model f(·), various machine learning algorithms such as a known multiple regression model, a LightGBM (Light Gradient Boosting Machine), and a multi-layer perceptron (MLP), can be used.

Furthermore, when training the scale-up coefficient estimation model f(.), a large amount of learning data including pairs of feature amount F(x) for each measurement point x (of the smart pole 4) and the scale-up coefficient R _C (x) (determined by the scale-up coefficient determination unit 112) is prepared, and training is performed to minimize the error (loss function) of the estimated scale-up coefficient R(x)^ from the scale-up coefficient R _C (x). Incidentally, when the feature amount (taking into account the context of movement) derived using the BERT in (c) above is used, it is also preferable to perform fine tuning on the pre-trained BERT model here to improve the estimation accuracy of the scale-up coefficient R(x)^.

<Expansion factor estimation means>
Also in the functional block diagram of FIG. 1, the magnification coefficient estimation unit 115 inputs the feature generated by the feature generation unit 113 to the constructed (trained) magnification coefficient estimation model, and estimates the magnification coefficient from the output of this magnification coefficient estimation model. As described above, the trained magnification coefficient estimation model f(·) is not limited to the measurement point (of the smart pole 4), but can input the feature F(x) of any point x and output (estimate) the magnification coefficient R(x)^ of the arbitrary point x. Incidentally, even in the case of the feature F(x) derived using the BERT (c) above (taking into account the context of movement), a sufficiently small mesh (for example, a rectangle of 20 m x 20 m) can be set, and the feature F(x) at approximately any point x can be obtained.

Therefore, in this embodiment, the magnification coefficient estimation unit 115 comprehensively estimates the magnification coefficient (magnification coefficient map) R(X)^ for all points X = (x1, x2, ..., xm) in a specified region (e.g., the entire country) that are to be made high-resolution. This makes it possible to generate highly accurate population distribution data, for example, a high-resolution population heat map, as will be explained later. Note that all points X = (x1, x2, ..., xm) can also be, for example, a collection of representative points (e.g., central points) of each mesh in a high-resolution population heat map.

<Target presence distribution determining means>
Also in the functional block diagram of Figure 1, the object presence distribution determination unit 116 determines the number of people (objects) present, i.e., the number of people (population), at each of a number of points within a specified area (e.g., the entire country), based on the estimated presence distribution CE _B (x) determined by the presence distribution determination unit 111 and the magnification coefficient R(x)^ estimated by the magnification coefficient estimation unit 115.

Specifically, in this embodiment, the target presence distribution determination unit 116 calculates the number of people (count number) C(x) at each point x of the multiple set points by the following formula (12): C(x) = CE _B (x) × R(x)^
In this way, the target presence distribution determination unit 116 can determine the number of people (count number) C(x) at any point x by multiplying the estimated presence number distribution CE _B (x), which is capable of calculating the number of people (count number) at any point x, even though it is a number of people (count number) determined from the positioning data of some users, by the magnification coefficient R(x)^ at the arbitrary point x, which represents the ratio between the total number of people and the number of users for which positioning data is obtained.

The number of people C(x) calculated and determined here is, so to speak, the estimated count that would be obtained as the measurement value if smart poles 4 (actual number measuring means) were installed at point x (where smart poles 4 are not actually installed). In other words, it is an estimate of the actual number of people at point x (actual number measurement estimate).

The target presence distribution determination unit 116 can calculate and determine the number of people (count number) C(x) for any (multiple) points x in a specified area (for example, nationwide), and can generate, for example, a high-resolution population distribution graph. It can also calculate and determine the number of people (count number) C(X) for all points to be highly resolved, for example, the representative points (for example, central points) X = (x1, x2, ..., xn) of each mesh, and can generate, for example, a high-resolution population heat map.

The population distribution data (number of people (count number) C(X)), high-resolution population distribution graph, and high-resolution population heat map generated by the target presence distribution determination unit 116 may be displayed on the user interface unit 106 having a display unit and presented to the operator/manager of the device 1. They may also be transmitted to an external information processing device via the communication control unit 121 and the communication interface unit 101 and used there.

Here, the above-described processes in the model construction unit 114, the expansion coefficient estimation unit 115, and the target presence distribution determination unit 116 of this embodiment will be briefly summarized. As shown in Fig. 5, the model construction unit 114 constructs an expansion coefficient estimation model (regression model) by performing a training process in which the error between the correct expansion coefficient R _C (x) determined by the correct expansion coefficient determination unit and the estimated expansion coefficient R(x)^ is set as a loss.

5, the expansion coefficient estimation unit 115 inputs the feature value F(x) at any (many) points x in a specific region (for example, the whole country) to the constructed expansion coefficient estimation model, and determines the expansion coefficient R(x)^ from the output of the expansion coefficient estimation model. Finally, the target presence distribution determination unit 116 uses the expansion coefficient R(x)^ and the presence count distribution _CEB (x) at the time of estimation determined by the presence count distribution determination unit 111 to calculate and determine the number of people (count number) C(x) at any (many) points x, and further generates a high-resolution population distribution graph or a high-resolution population heat map from this.

As described above in detail, according to the present invention, a correct answer magnification coefficient is determined using the real number measurement results obtained from the real number measurement means, and the number of targets present at each point is determined using the magnification coefficient estimation model constructed using this correct answer magnification coefficient. Here, even in a situation where the number of measurement points is limited, by setting a sufficient number of multiple points within a specified area, it is possible to obtain high-resolution target distribution data.

In addition, since the ground truth magnification coefficient is determined from the actual number measurement results and the abundance distribution, it is possible to improve the precision (accuracy) of the magnification coefficient related to the output of the magnification coefficient estimation model by setting appropriate features. As a result, according to the present invention, it is possible to estimate the distribution of a high-resolution target with higher precision (accuracy) without relying on ground truth data related to the target's actual distribution.

Furthermore, for example, the target distribution estimation method according to the present invention can be applied to perform accurate and detailed prediction of people flow and even traffic behavior in urban areas, to develop resilient traffic and transportation infrastructure, and to implement efficient urban planning to solve the challenges brought about by urbanization. In other words, the present invention can contribute to Goal 9 "Develop 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.

A person 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 ideas and viewpoints of the present invention. The above description is merely an example of an embodiment and does not impose any unnecessary restrictions. The present invention is limited only by the scope of the claims and their equivalents.

1. Communication equipment (target distribution estimation device)
REFERENCE SIGNS LIST 101 Communication interface unit 102 Positioning result storage unit 103 Real number measurement result storage unit 104 Map information storage unit 105 Attribute information storage unit 106 User interface unit 111 Presence number distribution determination unit 111a Normal presence distribution determination unit 111b Estimation time presence distribution determination unit 112 Correct answer expansion coefficient determination unit 113 Feature amount generation unit 114 Model construction unit 115 Expansion coefficient estimation unit 116 Target presence distribution determination unit 121 Communication control unit 2 Terminal 3 Base station 4 Smart pole (real number measurement means)
5 Map information management server 6 Attribute information management server

Claims (11)

A target distribution estimation program for causing a computer to function to estimate a distribution of a plurality of targets in a predetermined area,
a presence distribution determining means for determining a presence distribution as a distribution of the number of objects present in the predetermined area from a positioning result obtained during a predetermined period for a terminal attached to the object;
a correct answer magnification coefficient determination means for determining a correct answer magnification coefficient for each measurement point from measurement results obtained from a plurality of real number measurement means capable of actually measuring the number of the target at each measurement point within the specified area and the presence number distribution;
a feature generating means for generating, for each of a plurality of points within the predetermined region, a feature characterizing the point or an area including the point;
an expansion coefficient estimation means for inputting the feature quantity to an expansion coefficient estimation model constructed using the correct expansion coefficient, and estimating an expansion coefficient of the point from an output of the expansion coefficient estimation model;
A target distribution estimation program that causes a computer to function as a target presence distribution determination means that determines the number of targets present at each of a plurality of points within the specified area based on the presence distribution and the estimated expansion coefficient. The feature generating means
A process of generating a feature quantity of the location or the area including the location from information related to a specific thing or phenomenon related to the location or the area including the location based on the acquired map information of the specified area;
2. The target distribution estimation program according to claim 1, wherein the target distribution estimation program performs at least one of the following processes to generate feature quantities for the target location or the area including the target location from attribute information of the target to which a terminal located at the target location or the area including the target location is attached, the feature quantities being estimated from history information of the positioning results or obtained from registered information of the target; and, wherein the program regards a code representing the area including the target location as a word, regards time-series data of the code representing the target's movement history as a sentence, and generates feature quantities for the area using a natural language processing model that has learned similarities between words. The presence number distribution determination means determines a normal presence number distribution as a distribution of numbers that are normally present from the positioning results accumulated during a first predetermined period, and also determines an estimated presence number distribution as a distribution of numbers that are present at or immediately before the estimation time of the distribution from the positioning results acquired during a second predetermined period that is shorter than the first predetermined period and immediately before the estimation time of the distribution, and the positioning results accumulated during the first predetermined period;
The correct answer expansion coefficient determination means determines a correct answer expansion coefficient for the measurement point from the measurement result and the normal presence distribution,
The object distribution estimation program according to claim 1 or 2, characterized in that the object presence distribution determination means determines the number of objects present at each of a plurality of points within the specified area based on the presence distribution at the time of estimation and the estimated expansion coefficient. The object distribution estimation program according to claim 3, characterized in that the presence distribution determination means sets, as the first predetermined period, a period type related to a preset month, date, day of the week, weekday/holiday, and/or time period, and a period corresponding to the period type to which the estimation time point of the distribution belongs. The object distribution estimation program according to claim 3 or 4, characterized in that the presence distribution determination means performs kernel density estimation using information including the terminal's position and time, which is the positioning result accumulated during a first specified period, to determine the normal presence distribution, which is a function of the position and the time within the first specified period or within a period of a set period type. The abundance distribution determining means
6. The object distribution estimation program according to claim 3, further comprising: a process for determining the normal presence number distribution by assigning a greater weight to the positioning result the more recent the positioning time point related to the positioning result is within a first specified period; and a process for determining the estimation time presence number distribution by assigning a greater weight to the positioning result the more recent the positioning time point related to the positioning result is within the first specified period and a second specified period. The abundance distribution determining means
7. The object distribution estimation program according to claim 3, wherein the program performs one or both of the following processes: for each positioning result accumulated during a first specified period, at least one presence point corresponding to the positioning result, the presence point being assigned a weight whose sum total is 1, is set within an area having a size that is a monotonically increasing function of the positioning error included in the positioning result or determined from the positioning result, and the normal presence number distribution is determined from the presence point; and for each positioning result acquired during the first specified period and the second specified period, at least one presence point corresponding to the positioning result, the presence point being assigned a weight whose sum total is 1, is set within an area having a size that is a monotonically increasing function of the positioning error included in the positioning result or determined from the positioning result, and the estimation-time presence number distribution is determined from the presence point. The abundance distribution determining means
determining a subperiod presence number distribution as a distribution of the number of objects present at a representative time point of each subperiod from the positioning result acquired in the subperiod for each subperiod among a plurality of subperiods included in the first predetermined period and the second predetermined period;
The target distribution estimation program according to any one of claims 3 to 7, characterized in that the sub-period presence distribution at the representative time is extrapolated to the presence distribution at the estimation time, and the presence distribution obtained by extrapolation is determined to be the presence distribution at the estimation time. An object distribution estimation device that estimates a distribution of a plurality of objects in a predetermined area, comprising:
a presence distribution determining means for determining a presence distribution as a distribution of the number of objects present in the predetermined area from a positioning result obtained during a predetermined period for a terminal attached to the object;
a correct answer magnification coefficient determination means for determining a correct answer magnification coefficient for each measurement point from measurement results obtained from a plurality of real number measurement means capable of actually measuring the number of the target at each measurement point within the specified area and the presence number distribution;
a feature generating means for generating, for each of a plurality of points within the predetermined region, a feature characterizing the point or an area including the point;
an expansion coefficient estimation means for inputting the feature quantity to an expansion coefficient estimation model constructed using the correct expansion coefficient, and estimating an expansion coefficient of the point from an output of the expansion coefficient estimation model;
and an object presence distribution determination means for determining the number of said objects present at each of a plurality of points within said specified area based on said presence distribution and the estimated expansion coefficient. An object distribution estimation system for estimating a distribution of a plurality of objects in a predetermined area, comprising:
a presence distribution determining means for determining a presence distribution as a distribution of the number of objects present in the predetermined area from a positioning result obtained during a predetermined period for a terminal attached to the object;
a correct answer magnification coefficient determination means for determining a correct answer magnification coefficient for each measurement point from measurement results obtained from a plurality of real number measurement means capable of actually measuring the number of the target at each measurement point within the specified area and the presence number distribution;
a feature generating means for generating, for each of a plurality of points within the predetermined region, a feature characterizing the point or an area including the point;
an expansion coefficient estimation means for inputting the feature quantity to an expansion coefficient estimation model constructed using the correct expansion coefficient, and estimating an expansion coefficient of the point from an output of the expansion coefficient estimation model;
and an object presence distribution determination means for determining the number of the object present at each of a plurality of points within the specified area based on the presence distribution and the estimated expansion coefficient. 1. An object distribution estimation method implemented by a computer for estimating a distribution of a plurality of objects in a predetermined area, comprising:
A step of determining a presence distribution as a distribution of the number of objects present in the predetermined area from positioning results obtained during a predetermined period for a terminal attached to the object;
determining an expansion coefficient of a correct answer for each measurement point from measurement results obtained from a plurality of real number measurement means capable of measuring the number of the target at each measurement point within the specified area and the distribution of the number of targets present; and generating, for each of the plurality of points within the specified area, a feature quantity that characterizes the point or an area including the point;
a step of inputting the feature quantity into an expansion coefficient estimation model constructed using the correct expansion coefficient, and estimating an expansion coefficient of the point from an output of the expansion coefficient estimation model;
and determining the number of said objects present at each of a plurality of points within said specified area based on said population distribution and the estimated expansion coefficient.

Images