A program, apparatus, system, and method for estimating the number of targets while considering the interaction between features.

Abstract

To provide a target number estimation program for performing estimation of the target number with the reduced influence of an error related to observation carried out by positioning means.SOLUTION: The program is for estimating the number of targets located in a plurality of zones in a certain area that is a positioning unit of first positioning means, and causes a computer to function as: unevenly distributed feature quantity generation means that generates an unevenly distributed feature quantities including, as a unit feature quantity, a zone feature quantity that is a feature quantity related to an event or a thing related to stay or movement of the targets in the zones; frequency information generation means that generates, from the unevenly distributed feature quantities, an interaction feature quantity reflecting the relationship between the unit feature quantities by using a learned neural network algorithm, and by using the interaction feature quantity, generates frequency information related to the frequency at which the combination of the events related to the unit feature quantities occurs; and target number determination means that calculates the probability that the targets are located in the zones from the frequency information, and determines information related to the number of targets located in the zones from the probability and observation values of the number of targets.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a technique for estimating the distribution of objects, such as population distribution, in a predetermined area.

Background Art

[0002] Understanding the population distribution in a predetermined area, or the flow of people in the sense of a distribution that changes over time, is very important for marketing, the efficient operation of transportation networks (such as avoiding traffic jams), urban planning including the maintenance of public facilities, and further for infectious disease countermeasures.

[0003] Conventionally, in a people flow data providing service, in each area (hereinafter also referred to as a mesh) constituting the area covered by a base station, the people flow is estimated using the number of terminals (number of users) observed by GPS (Global Positioning System). In reality, not all people in this area own a terminal, and also the users who provide GPS positioning information to the base station are limited. Therefore, an extrapolation is performed on the number of terminals (number of users) observed in each area to estimate how many people are present in each area.

[0004] On the other hand, the location registration information (connection sector information) recorded as a communication connection result between a base station and a terminal, that is, information about the terminals (users) present in the covered area, also naturally includes information on the number of terminals (number of users) present in this area. Here, the observation of terminals (users) using this location registration information is generally large in scale and is suitable for estimating the people flow in that regard. However, the interval between base stations may reach, for example, several km (kilometers), and generally, the position resolution of the observation results by the base stations becomes low.

[0005] To address the problems with location registration information, for example, the user number estimation process disclosed in Patent Document 1 estimates the number of users using data that links acquired GPS logs and location registration information with user identifiers (IDs). Specifically, using such data, a probability map P(M|B) is constructed through training, which consists of the probability that a user observed at base station B (registered at base station B) exists within mesh M. This probability map P(M|B) is then used to estimate the number of users in each mesh.

[0006] For example, if base station B1 is observed 100 times and users are observed in mesh M1 30 times, the probability P(M1|B1) is set to 0.3 (=30 / 100). In this case, if 500 users are observed at base station B1 at a certain time, it can be estimated that 150 of them (=500 × 0.3) were located within mesh M1. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 2021-005167 [Overview of the project] [Problems that the invention aims to solve]

[0008] However, in conventional technologies, including the technology disclosed in Patent Document 1 mentioned above, the error in GPS observations remains a major problem.

[0009] In reality, while GPS can provide observation results with high positional resolution, the scale of observable devices (users) is usually quite small. Therefore, it is not uncommon for only one device to be observed in each area, or even none at all. In this case, the number of people after magnified estimation can vary greatly depending on whether or not a device is observed by chance. Thus, GPS observations suffer from large errors due to the small sample size (hereinafter also referred to as sampling error).

[0010] For example, in the user number estimation process described in Patent Document 1, the number of observations per base station and the number of observations per mesh are used as training data for training the probability map P(M|B). As a result, the probability map P(M|B) is constructed independently for each base station-mesh pair. Furthermore, considering the need to grasp population bias and fluctuations according to time of day, the probability map P(M|B) must also be constructed independently for each time of day.

[0011] In this regard, in urban areas where a certain number of users provide GPS positioning information to base stations, observational data related to base station-mesh combinations can be obtained with considerable frequency for each time period. However, in rural areas with a small user base, sampling errors are easily introduced into the observational data related to the mesh, resulting in large errors in the probability values ​​of the constructed probability map P(M|B). In other words, the smaller the population size being estimated, the larger the error in the estimated number of users becomes.

[0012] Therefore, the present invention aims to provide a target number estimation program, apparatus, system, and method that can estimate the number of targets while suppressing the influence of errors related to observation by positioning means. [Means for solving the problem]

[0013] According to the present invention, a target number estimation program estimates the number of targets located within a plurality of zones included in an area which is a unit of location of an target identified by a first positioning means, A localized feature generation means that generates localized features that include, as unit features, a localized feature quantity which is a feature quantity relating to events or things related to the stay or movement of the subject in the area, a first identification feature quantity as an identifier for the area or the first positioning means, and / or a second identification feature quantity as an identifier for the area; From the unevenly distributed features, the trained model incorporates the interactions between those unit features. Machine LearningUsing an algorithm, interaction features are generated that reflect the relationships between the unit features, and these interaction features From the result obtained by applying the learned corresponding transformation matrix operator to it , a frequency information generation means that generates frequency information relating to the frequency at which combinations of events related to each of the unit features occur, A means for determining the number of targets that determines information relating to the number of targets located within the area, based on the frequency information, and the probability that the target is located within the area, based on the probability and the observed number of targets. to make the computer work 、 The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated at each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded eccentric feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction features as the interaction features, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which operates the computer to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable, and the corresponding transformation matrix operator is applied to the linear regression feature which is information relating to the frequency of occurrence of combinations of events related to each of the unit features, and the computer operates to generate the interaction feature as the CIN interaction feature, the DNN interaction feature, and the linear regression feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm. That is, versus A quadrant estimation program is provided.

[0014] As one embodiment of the target number estimation program according to the present invention, the uneven feature generation means generates the uneven feature which further includes attribute feature quantities, which are feature quantities relating to the attributes of the target, as unit feature quantities. The object number determination means may also calculate the probability of an object being located within the area and possessing the attribute from the frequency information, and determine information regarding the number of objects located within the area from the probability of an object being located within the area and the observed number of objects being located within the area and possessing the attribute.

[0015] Furthermore, in another embodiment of the target number estimation program according to the present invention, the uneven feature generation means generates the uneven feature which further includes a time feature, which is a feature related to a time interval, as a unit feature. The object number determination means may also preferably calculate the probability of object existence within the area, which is the probability that an object located within the area during the time interval is located within the area, from the frequency information, and determine information regarding the number of objects located within the area from the probability of object existence within the area and the observed number of objects located within the area during the time interval.

[0016] Furthermore, in yet another embodiment of the target number estimation program according to the present invention, the uneven feature generation means generates the uneven feature which further includes, as unit features, an attribute feature which is a feature related to the attribute of the target and a time feature which is a feature related to the time interval, The object number determination means may also preferably calculate the probability of presence within the area, which is the probability that an object located within the area and possessing the attribute is located within the area during the given time interval, from the frequency information, and determine information regarding the number of objects located within the area from the probability of presence within the area and the observed number of objects located within the area and possessing the attribute during the given time interval.

[0017] Furthermore, in another embodiment of the target number estimation program according to the present invention, it is also preferable that the biased feature generation means generates biased features that further include, as unit features, a first identification feature as an identifier of the area or the first positioning means for a certain time interval, a previous first identification feature as an identifier of the area or the first positioning means for a time interval prior to the certain time interval, and / or a subsequent first identification feature as an identifier of the area or the first positioning means for a time interval later than the certain time interval.

[0018] The present invention also provides a target number estimation program for estimating the number of targets located within a plurality of zones included in an area which is a unit of location of an target identified by a first positioning means, A means for generating a localized feature that generates localized feature quantities that include, as unit features, localized feature quantities which are features related to events or things related to the stay or movement of the subject in the area, and attribute feature quantities which are features related to the attributes of the subject. From the unevenly distributed features, the trained model incorporates the interactions between those unit features. Machine Learning Using an algorithm, interaction features are generated that reflect the relationships between the unit features, and these interaction features From the result obtained by applying the learned corresponding transformation matrix operator to it , a frequency information generation means that generates frequency information relating to the frequency at which combinations of events related to each of the unit features occur, A means for determining the number of objects located within an area determines information relating to the number of objects located within an area, based on the frequency information, and the probability of an object being located within an area and possessing the attribute being located within that area. to make the computer work 、 The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated at each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded eccentric feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction features as the interaction features, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which operates the computer to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable, and the corresponding transformation matrix operator is applied to the linear regression feature which is information relating to the frequency of occurrence of combinations of events related to each of the unit features, and the computer operates to generate the interaction feature as the CIN interaction feature, the DNN interaction feature, and the linear regression feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm. That is, versus A quadrant estimation program is provided.

[0019] The present invention further provides a target number estimation program for estimating the number of targets located within a plurality of zones included in an area which is a unit of location of an target identified by a first positioning means, A means for generating a localized feature that generates localized features that include, as unit features, localized features that are features relating to events or things related to the stay or movement of the subject in the area, and localized features that are features relating to a time interval. From the unevenly distributed features, the trained model incorporates the interactions between those unit features. Machine Learning Using an algorithm, interaction features are generated that reflect the relationships between the unit features, and these interaction features From the result obtained by applying the learned corresponding transformation matrix operator to it , a frequency information generation means that generates frequency information relating to the frequency at which combinations of events related to each of the unit features occur, A means for determining the number of objects located within an area determines information relating to the number of objects located within an area, based on the frequency information, and the probability of an object being located within an area within a given time interval, which is the probability of an object being located within an area within that time interval, based on the probability of an object being located within an area and the observed number of objects located within that area within that time interval. to make the computer work 、 The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated at each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded eccentric feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction features as the interaction features, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which operates the computer to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable, and the corresponding transformation matrix operator is applied to the linear regression feature which is information relating to the frequency of occurrence of combinations of events related to each of the unit features, and the computer operates to generate the interaction feature as the CIN interaction feature, the DNN interaction feature, and the linear regression feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm. That is, versus A quadrant estimation program is provided.

[0020] Furthermore, in the target number estimation program according to the present invention described above, it is also preferable that the uneven feature generation means generates the uneven feature which further includes as a unit feature a first positioning feature which is a feature related to the installation environment, specifications, or performance of the first positioning means. It is also preferable that the uneven feature generation means generates the uneven feature which further includes as a unit feature a positional relationship feature which is a feature related to the positional relationship between the area and the point related to the first positioning means.

[0023] Furthermore, in the target number estimation program according to the present invention described above, the neural network algorithm and the transformation matrix operator It is also preferable that the model be trained using training data that includes a localized feature that includes the localized feature in the area identified by the second positioning means as a unit feature, and frequency information as ground truth data relating to the frequency at which combinations of events related to each of the unit features included in the localized feature occur.

[0024] Furthermore, it is preferable that the target is a human being, including the user of the terminal, the first positioning means is a base station, the observed number of targets is the observed number of terminals connected to the base station, and the second positioning means is a GPS (Global Positioning System) positioning means installed in the terminal.

[0025] The present invention also provides an object number estimation device for estimating the number of objects located within a plurality of zones included in an area which is a unit of location of an object identified by a first positioning means, A localized feature generation means that generates localized features that include, as unit features, a localized feature quantity which is a feature quantity relating to events or things related to the stay or movement of the subject in the area, a first identification feature quantity as an identifier for the area or the first positioning means, and / or a second identification feature quantity as an identifier for the area; From the unevenly distributed features, the trained model incorporates the interactions between those unit features. Machine Learning Using an algorithm, interaction features are generated that reflect the relationships between the unit features, and these interaction features From the result obtained by applying the learned corresponding transformation matrix operator to it , a frequency information generation means that generates frequency information relating to the frequency at which combinations of events related to each of the unit features occur, A means for determining the number of targets that determines information relating to the number of targets located within the area, based on the frequency information, and the probability that the target is located within the area, based on the probability and the observed number of targets. to have death, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated at each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded eccentric feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction features as the interaction features, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which operates the computer to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable, and the corresponding transformation matrix operator is applied to the linear regression feature which is information relating to the frequency of occurrence of combinations of events related to each of the unit features, and the computer operates to generate the interaction feature as the CIN interaction feature, the DNN interaction feature, and the linear regression feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm. That is, versus A quadrant estimation device is provided.

[0026] According to the present invention, the present invention further provides an object number estimation system for estimating the number of objects located within a plurality of zones included in an area which is a unit of location of an object identified by a first positioning means, A localized feature generation means that generates localized features that include, as unit features, a localized feature quantity which is a feature quantity relating to events or things related to the stay or movement of the subject in the area, a first identification feature quantity as an identifier for the area or the first positioning means, and / or a second identification feature quantity as an identifier for the area; From the unevenly distributed features, the trained model incorporates the interactions between those unit features. Machine Learning Using an algorithm, interaction features are generated that reflect the relationships between the unit features, and these interaction features From the result obtained by applying the learned corresponding transformation matrix operator to it , a frequency information generation means that generates frequency information relating to the frequency at which combinations of events related to each of the unit features occur, A means for determining the number of targets that determines information relating to the number of targets located within the area, based on the frequency information, and the probability that the target is located within the area, based on the probability and the observed number of targets. to have death, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated at each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded eccentric feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction features as the interaction features, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which operates the computer to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable, and the corresponding transformation matrix operator is applied to the linear regression feature which is information relating to the frequency of occurrence of combinations of events related to each of the unit features, and the computer operates to generate the interaction feature as the CIN interaction feature, the DNN interaction feature, and the linear regression feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm. That is, subject A number estimation system is provided.

[0027] The present invention further provides a method for estimating the number of objects located within a plurality of zones that are included in an area which is a unit of location of an object identified by a first positioning means, A step of generating a localized feature that includes, as a unit feature, an area feature which is a feature related to events or things related to the stay or movement of the subject in the area, a first identification feature as an identifier for the area or the first positioning means, and / or a second identification feature as an identifier for the area; From the unevenly distributed features, the trained model incorporates the interactions between those unit features. Machine Learning Using an algorithm, interaction features are generated that reflect the relationships between the unit features, and these interaction features From the result obtained by applying the learned corresponding transformation matrix operator to it , a step of generating frequency information relating to the frequency of occurrence of each combination of events related to the unit feature, The steps include: calculating the probability that the target is located within the area based on the frequency information; and determining information regarding the number of targets located within the area based on the probability and the observed number of targets. to have death, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated at each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded eccentric feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction features as the interaction features, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which operates the computer to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable, and the corresponding transformation matrix operator is applied to the linear regression feature which is information relating to the frequency of occurrence of combinations of events related to each of the unit features, and the computer operates to generate the interaction feature as the CIN interaction feature, the DNN interaction feature, and the linear regression feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm. That is, Ko A method for estimating the number of targets, performed by a computer, is provided. [Effects of the Invention]

[0028] According to the object number estimation program, apparatus, system, and method of the present invention, it is possible to estimate the object number while suppressing the influence of errors related to observation by positioning means. [Brief explanation of the drawing]

[0029] [Figure 1] This is a functional block diagram showing the functional configuration of one embodiment of the target number estimation device according to the present invention. [Figure 2] This schematic diagram, including a table, illustrates one embodiment of statistical processing of the number of unique users for generating training data that includes unevenly distributed features. [Figure 3] This is a schematic diagram illustrating another embodiment of the localized feature according to the present invention. [Figure 4] This is a schematic diagram illustrating one embodiment of the interaction feature generation process according to the present invention. [Figure 5]This is a schematic diagram illustrating yet another embodiment of the interaction feature (and frequency information) generation process according to the present invention. [Modes for carrying out the invention]

[0030] Embodiments of the present invention will be described in detail below with reference to the drawings.

[0031] [Target Quantity Estimation Device / System] Figure 1 is a functional block diagram showing the functional configuration of one embodiment of the target number estimation device according to the present invention. Note that functional components not related to the target number estimation process are omitted in this functional block diagram.

[0032] The communication equipment device 1 shown in Figure 1 is, of course, a device that performs the function of communication equipment (for example, a relay function), but it is also an embodiment of the target number estimation device according to the present invention. That is, in this embodiment, the communication equipment device 1 determines the number of targets in each of the multiple areas (for example, many meshes of 250m (meters) on each side) that make up a predetermined region (for example, all 47 prefectures) by augmented estimation, the number of people in this embodiment, estimates the population distribution in the predetermined region (nationwide), and is a device that can output population distribution data (for example, a population distribution graph or a population heat map) as an estimation result.

[0033] In this embodiment, the communication equipment device 1 uses the base station 3 as a first positioning means to observe the number of terminals 2 that are connected to (location registered with) one base station 3, i.e., the number of users of terminals 2, as location registration information (connection sector information). This observed number of users is then used as the observed number of people (targets) located within the base station area covered by this base station 3.

[0034] The number of terminals 2 (users) located within this base station area and registered is typically the number under conditions where the user base is considerably large. However, the base stations 3 are spaced apart, for example, over several kilometers, and the base station area is generally quite large. As a result, the location resolution of the observation results of terminals 2 (users) by base stations 3 is usually very low.

[0035] Under these circumstances, in this embodiment, the communication equipment 1 further utilizes the GPS (Global Positioning System) positioning means, which is a second positioning means installed in the terminal 2, to observe people (targets) located within each mesh. Specifically, it acquires the GPS positioning information of the terminal 2 transmitted from the terminal 2 of a user who has authorized the provision of location information to the base station 3, and performs observation within each mesh.

[0036] Here, GPS positioning information typically has high positional resolution. However, generally, the number of observable GPS-enabled users is quite small. Therefore, it is not uncommon for only one terminal 2 to be observed in each mesh, or for none to be observed at all. As a result, GPS positioning information usually has a large sampling error, which tends to increase the error in the population estimate after scaling.

[0037] In order to resolve the problems of location registration information and GPS positioning information described above and to more suitably estimate the number of people (targets) within a mesh (area), the communication equipment 1 in this embodiment is configured as follows: (A) A localized feature generation unit 111 that generates localized features that include, as unit features, a "localized feature quantity" which is a feature quantity relating to events or things related to the stay or movement of a person (object) in a certain mesh (area), a "first identification feature quantity" which is an identifier (ID) of a certain base station area or a certain base station 3 (first positioning means), and / or a "second identification feature quantity" which is an ID of the certain mesh (area); (B) A frequency information generation unit 112 generates "interaction features" that reflect the relationships between unit features, using a trained "neural network algorithm" that incorporates the interactions between unit features from the generated "uniform features," and uses these "interaction features" to generate "frequency information" relating to the frequency at which combinations of events related to each unit feature occur. (C) From the generated "frequency information," the "probability of presence within a certain area," which is the probability that a person (target) located within a certain base station area is located within a certain mesh (area), is calculated, and the target number determination unit 113 determines information related to the number of people (targets) located within a certain mesh (area) from this "probability of presence within a certain area" and the observed value (acquired by base station 3) of the number of people (targets) located within a certain base station area. It has.

[0038] Conventionally, most techniques for estimating the population within a mesh, such as the technique disclosed in Patent Document 1, use information such as base station IDs as explanatory variables to obtain information related to the mesh, such as the mesh ID, as the objective variable. In this case, the training data used to train the probability map or estimation model includes mesh ID information where GPS users are located (exist) as ground truth data. As a result, the sampling error inherent in this mesh ID information increases the error in estimating the population within the mesh, which is a problem.

[0039] In contrast, communication equipment 1, unlike conventional equipment, uses features related to the mesh (area) (area features and second discriminant features) as explanatory variables and derives "frequency information" as the target variable for estimating the population (number of targets) within the mesh (area) from a "neural network algorithm" that incorporates the interaction between unit features.

[0040] In this way, when estimating the population (number of targets) within a mesh (area), the communication equipment device 1 replaces the conventional "problem of finding the mesh (where the user is located)" with "the problem of deriving the frequency of occurrence of the mesh (where the user is located) using mesh information." As a result, it is possible to avoid including mesh ID information (where the GPS user is located), which has a large sampling error, in the ground truth data, and to suppress the estimation error of the population (number of targets) within the mesh. In other words, the communication equipment device 1 makes it possible to estimate the population (number of targets) while suppressing the influence of errors related to observation by positioning means. For example, it becomes possible to improve the accuracy of population (number of targets) estimation in rural areas where the user base is small and therefore the number of authorized users is very small.

[0041] Furthermore, the "uniform features" used as explanatory variables, which include the mesh information described in (A) above, are converted into D-dimensional embedding representation vectors in this embodiment, as will be explained later. These are further converted into "interaction features" that reflect the relationships between unit features using a "neural network algorithm." These features (representation vectors) are the result of learning the representation of the mesh (area), and can be shared knowledge that is transferable and translatable among all target meshes. For example, they can include information such as "the population is high in commercial facilities adjacent to train stations" or (if time features and attribute features, which will be discussed later, are also adopted) "there are many housewives in commercial facilities in the evening."

[0042] In this embodiment, the target for which the number is to be determined is "humans," and in the following explanation, the target will also be "humans" to estimate the population distribution (human flow). However, the target of the present invention is not limited to "humans." That is, a wide variety of things can be the target of the present invention as long as they can be observed by the second positioning means to determine whether or not they are located (exist) within the area of ​​the first positioning means. For example, it is also possible to use a "vehicle" equipped with a communication terminal as the target of the present invention, with the first positioning means being a base station and the second positioning means being an in-vehicle GPS.

[0043] Incidentally, at least one of the above-mentioned (A) uneven feature generation unit 111, (B) frequency information generation unit 112, and (C) target number determination unit 113 can be a functional component of another device, and the target number estimation function can be performed by the entirety of multiple devices (for example, multiple servers). In this case, the entirety of these multiple devices constitutes the target number estimation system according to the present invention. The target number estimation device (communication equipment device 1) of this embodiment will be described in more detail below.

[0044] [Device Functional Configuration, Target Number Estimation Program / Method] Similarly, according to the functional block diagram in Figure 1, the communication equipment device 1, as one embodiment of the target number estimation device according to the present invention, includes a communication interface unit 101, a mesh information storage unit 102, a user information storage unit 103, a GPS positioning information storage unit 104, a base station information storage unit 105, a unique user (UU) statistics storage unit 106, a user interface (UI) unit 107, and a processor memory (a processing system equipped with memory functions).

[0045] Here, this processor memory stores one embodiment of the target number estimation program according to the present invention and also has computer functionality, and performs target number estimation processing by executing this target number estimation program. Furthermore, the communication equipment device 1 may be a device dedicated to target number estimation processing that is not a communication device, and it may also be a cloud server, non-cloud server, personal computer (PC), notebook or tablet computer, or mobile terminal such as a smartphone, equipped with the target number estimation program according to the present invention.

[0046] Furthermore, the above-mentioned processor memory includes a localized feature generation unit 111, a frequency information generation unit 112 including an interaction feature generation unit 112a, a target number determination unit 113 including a mesh existence probability calculation unit 113a, a communication control unit 121, and an input / output control unit 122. The functional components described above can be understood as functions realized by executing the target number estimation program stored in the processor memory. Also, the processing flow shown by connecting the functional components of the communication equipment device 1 with arrows in Figure 1 can be understood as one embodiment of the target number estimation method according to the present invention.

[0047] <Universal feature generation means> The following description will omit the explanation of the communication control function (e.g., relay function) in the communication equipment device 1, and will only explain the target number estimation processing function according to the present invention. As shown in the functional block diagram of Figure 1, the uneven feature generation unit 111 in this embodiment is (a) Time features, which are features related to the time interval, (i) The first distinguishing feature as an identifier (ID) of the base station area or base station 3, (c) Second discriminant feature as a mesh identifier (ID), (e) Attribute features, which are features related to human attributes, (e) Area features, which are features related to events or things that are related to the stay or movement of people in a mesh. (c) The first positioning feature quantity, which is a feature quantity related to the installation environment, specifications, or performance of base station 3, and (k) Positional relationship features are features related to the positional relationship between the mesh and the base station location. Generate a localized feature that includes as a unit feature.

[0048] First, the time features in (a) above are features that represent the time intervals involved in the estimation (or in the training data), and can be 48 (=2×24) dimensional vectors that represent date and time features such as "weekdays around 3 PM" or "holidays around 4 PM". Alternatively, they can be 148 (=7×24) dimensional vectors that represent date and time features such as "Mondays around 7 AM" or "Saturdays around 12 PM".

[0049] Next, the first discriminant feature in (i) above is a feature that serves as an identifier for the base station area or base station 3 related to the estimation (or in the training data), and can be, for example, a one-hot vector having dimensions equal to the total number of base stations 3 in the target area (e.g., all 47 prefectures). Also, the second discriminant feature in (iii) above is a feature that serves as an identifier for the mesh related to the estimation (or in the training data), and can be, for example, a one-hot vector having dimensions equal to the total number of meshes in the target area (e.g., all 47 prefectures).

[0050] Furthermore, the attribute features in (e) above are features that represent the attributes of a human being involved in the estimation (or in the training data), and may be one-hot vectors that represent the characteristics of (user) attributes such as "gender," "age group," and "prefecture of residence" (which can be obtained as registration information for the user of terminal 2). For example, the vector part representing "prefecture of residence" can be a 47-dimensional one-hot representation part.

[0051] Here, the attribute information used to generate these attribute features can be, for example, attribute information (associated with a user ID) received (via the communication interface unit 101 and the communication control unit 121) from a contract information management server or a survey results management server managed by the telecommunications carrier related to terminal 2, and stored in the user information storage unit 103. Alternatively, it may be attribute information of the terminal 2 user estimated from the communication log of terminal 2 using known technology.

[0052] Furthermore, the area features described in (e) above can be POIs (Points of Interest), structures, buildings, and features related to means of transportation such as railways in the mesh related to estimation (or in the training data), as well as features related to information about the mesh obtained from road map information, (past) population survey results such as censuses, and hazard maps. For example, as area features, (a) The number of roads with a width of ** meters or more included in the mesh, (b) The number of houses, apartment buildings, businesses and commercial facilities included in the mesh, (c) Number of bus stops included in the mesh, (d) Number of railway stations included in the mesh, (e) Nighttime population in the mesh, and (f) Inundation risk in the relevant mesh shown on the flood hazard map One-hot vectors can be generated to represent these values. For example, numerical values ​​such as the number of roads or the nighttime population can be represented by dividing the possible range of values ​​into categories such as [0-1][1-5][5-10]... or [0-10][10-100][100-500]... and then creating a one-hot representation with dimensions equal to the number of categories.

[0053] Here, the various information used to generate such regional features can be various information received from management servers for various map information (including road and railway maps), census results management servers, hazard map management servers, etc., and stored in the mesh information storage unit 102.

[0054] Furthermore, the first positioning feature in (k) above is a feature related to the installation environment, specifications, or performance of base station 3 (in the training data) related to estimation. For example, as the first positioning feature, (a) The type of location where the base station 3 is installed, for example, whether it is outdoors (1) or indoors (0), (b) The radio output of the base station 3 (e.g., in kilowatts (kW)), and (c) The orientation (azimuth angle) and altitude (e.g., elevation) of the antenna of the base station 3. A one-hot vector representing this can be generated. This one-hot vector can also be a vector of one-hot representations that represent the category to which the corresponding number belongs.

[0055] Here, the information used to generate such a first positioning feature can be information relating to base stations 3 that the communication equipment device 1 has previously stored and managed in the base station information storage unit 105. Alternatively, it may be information relating to each base station 3 obtained from each base station 3 as appropriate.

[0056] Furthermore, the positional relationship feature described in (k) above is a feature relating to the relationship between the location of the mesh (in the estimation or training data) and the installation location of base station 3. For example, as a positional relationship feature, (a) the distance between the center of the mesh and the installation location of the base station 3, and (b) The orientation (azimuth angle) of the installation position of base station 3 as seen from the center of the mesh. A one-hot vector representing this can be generated. This one-hot vector can also be a one-hot representation vector representing the category to which the corresponding numerical value belongs. Furthermore, this positional relationship feature can also be generated using information stored and managed, for example, in the mesh information storage unit 102 or the base station information storage unit 105.

[0057] Here, when estimating the number (population) of people with a certain attribute who exist within a certain mesh of a certain base station area during a certain time period (time interval), the localized features input to the neural network algorithm are the set of features (a) to (g) above that correspond to these "certain time period (time interval)," "certain base station area," "certain mesh," and "certain attribute."

[0058] In summary, the localized feature X in this embodiment is given by the following formula (1) X={T, B, M, A, fm, fb, fmb} The quantity is represented by the following: Here, T is the time feature of (a) above, B is the first discriminant feature of (b) above, M is the second discriminant feature of (c) above, A is the attribute feature of (d) above, fm is the area feature of (e) above, fb is the first positioning feature of (f) above, and fmb is the positional relationship feature of (g) above.

[0059] Furthermore, the localized feature X is not limited to including all of (a) to (g) above. For example, it may include at least the regional feature fm of (e) above, and one or both of the first discriminant feature B of (b) above and the second discriminant feature M of (c) above. It may also include at least the regional feature fm of (e) above and the time feature T of (a) above. Moreover, the localized feature X may also include at least the regional feature fm of (e) above and the attribute feature A of (d) above. For example, it may include at least the regional feature fm of (e) above, the time feature T of (a) above, and the attribute feature A of (d) above. Furthermore, these localized features may also include the first positioning feature fb of (f) above, and may also include the positional relationship feature fmb of (g) above.

[0060] Next, we will explain in detail the process of generating training data containing the ubiquitous feature X for constructing the "neural network algorithm," using Figure 2. Figure 2 is a schematic diagram including a table showing one embodiment of the statistical processing of the number of unique users (UU) for generating training data containing the ubiquitous feature X.

[0061] As shown in Figure 2, the localized feature generation unit 111 in this embodiment is (a) Location registration information (connection sector information) is obtained from each base station 3, and aggregation processing is performed to generate the "location registration information aggregation result". (b) Obtain the "user attribute information" received from the contract information management server and the survey results management server managed by the telecommunications carrier related to terminal 2 and stored in the user information storage unit 103, (c) Based on the user's location registration information and GPS positioning results collected over a predetermined period (e.g., 3 months), the prefecture where the user spent the longest continuous period at night (e.g., 6pm to 3am) is identified and designated as the user's prefecture of residence. The prefecture where the user spent the longest continuous period during the daytime (e.g., 7am to 7pm) is identified and designated as the user's prefecture of work. A "home / workplace determination result" is generated for each user, summarizing these determinations.

[0062] Next, the localized feature generation unit 111 links the information from (a) to (c) above using the user ID to generate the "location registration user attribute linking result". Finally, using this "location registration user attribute linking result", it counts the number of users (unique users, UU) that belong to the same base station ID, the same date and time (time interval), and the same age, gender, and prefecture of residence, and compiles these count results to generate the "time period, base station, and user attribute-specific UU count result".

[0063] Here, the "UU count results by time period, base station, and user attribute" are temporarily stored and updated in the UU statistics storage unit 106, and are read out and used as appropriate when generating training data. Specifically, a time feature T is generated from the date and time period in the read-out "UU count results by time period, base station, and user attribute," an attribute feature A is generated from age, gender, and prefecture of residence, and a first discriminant feature B and a first positioning feature fb are generated from the base station ID. In other words, a large number of combinations of {T, A, B, fb} are generated, each corresponding to the number of UU users.

[0064] Next, (a) Each of the above numerous combinations of {T, A, B, fb}, (b) Second discriminant feature M and area feature fm generated from information relating to meshes (including positioning points) contained in a large amount of GPS positioning information prepared for training data (transmitted from base station 3 and stored and managed by GPS positioning information storage unit 104) and Combinations can be generated by linking them with user IDs, and then a spatial relationship feature fmb can be added to these combinations to generate a localized feature X for training data. The ground truth data can be the frequency of occurrence of combinations {T, B, M, A, fm, fb, fmb} calculated from the corresponding number of unique users.

[0065] (Other embodiments relating to eccentric features) Figure 3 is a schematic diagram illustrating another embodiment of the localized feature X according to the present invention.

[0066] As shown in Figure 3(A), terminal 2, which is located within the base station area of ​​base station B2 at a given time, may have moved into this base station area from an adjacent base station area using a main road that passes through this base station area. Furthermore, it is quite possible that it will then move from this base station area to an adjacent base station area using this main road. Thus, a base station (current base station) at a given time period (time interval), and by extension the corresponding mesh, can be better identified by the base station in the previous time period (previous base station) and the base station in the subsequent time period (sequential base station).

[0067] Therefore, in this embodiment, as shown in Figure 3(B), not only the base station ID of the current base station, but also the base station ID of the immediately preceding base station and the base station ID of the immediately following base station are added to the table items to generate the "UU count result," and as a result, the following equation (2) X={T, B pre , B, B post , M, A, ···} A localized feature X represented by is generated and adopted. Here B pre This is the first prior discriminant feature, which represents the base station area or base station ID for a time interval prior to a certain time interval (in this case, the time interval corresponding to T) (for example, the immediately preceding time interval). Also, B post This then becomes the first discriminant feature, representing the base station area or base station ID for a time interval following a certain time interval (corresponding to the time interval T) (for example, the time interval immediately following).

[0068] By training a "neural network algorithm" using such unevenly distributed features X and then performing population estimation, the likelihood of obtaining more accurate estimation results increases. Incidentally, previously, the first discriminant feature B pre and thereafter, the first discriminant feature B postIt is also possible to adopt a biased feature X that includes only one of the two.

[0069] <Means for generating frequency information> Returning to the functional block diagram in Figure 1, in this embodiment, the interaction feature generation unit 112a of the frequency information generation unit 112 generates interaction features p+ that reflect the relationships between unit features (T, B, M, A, fm, fb, fmb) from the localized feature X (={T, B, M, A, fm, fb, fmb}) using a trained neural network algorithm, in this embodiment the CIN (Compressed Interaction Network) algorithm, which incorporates the interactions between unit features (T, B, M, A, fm, fb, fmb).

[0070] The process of generating interaction features p+ using the CIN algorithm described above will be explained using Figure 4 below. Incidentally, CIN and xDeepFM, which will be explained later, are referenced in Non-Patent Literature 1: Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, Guangzhong Sun, “xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems”, Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining.<https: / / doi.org / 10.48550 / arXiv.1803.05170> This is explained in detail in 2018.

[0071] Figure 4 is a schematic diagram illustrating one embodiment of the interaction feature generation process according to the present invention.

[0072] According to Figure 4, the interaction feature generation unit 112a (of the frequency information generation unit 112) first, (a) For each unit feature amount (T, B, M, A, fm, fb, fmb) of the input uneven feature amount X, embedding processing (for example, a process of compressing dimensions by applying a conversion matrix operator) is performed to generate an embedding vector e of D dimensions corresponding to each unit feature amount i (i = 1, 2, ···, m, where m is the number of unit feature amounts).

[0073] Next, as also shown in FIG. 4, the action feature amount generation unit 112a (b) combines the generated embedding vectors e i as one row (or column) to generate a first feature matrix x0 of m (= H0) × D dimensions, (c) From the n-th feature matrix x(n - 1) of H(n - 1) × D dimensions and the first feature matrix x0 of m × D dimensions, an n-th feature tensor zn of H(n - 1) × m × D dimensions is generated, and for this n-th feature tensor zn, filtering processing is performed while sliding along the dimension axis (D-axis) corresponding to the dimension of the unit feature amount to generate a plurality of (Hn pieces) of feature maps, and from these plurality of (Hn pieces) of feature maps, a process of generating an (n + 1)-th feature matrix xn of Hn × D dimensions is repeated until the value of n increases by 1 from 1 to a predetermined integer value T, (d) For each of the plurality of (Hn pieces) of feature maps generated at each value of n, the results of performing pooling processing (={p1, p2, ···, pT}) are added together to generate an interaction feature amount p+ (= Σ n=1 T pn). Here, the processing in (b) to (d) is processing by the CIN algorithm.

[0074] In this way, the CIN algorithm is an algorithm that incorporates the interaction between unit feature amounts (T, B, M, A, fm, fb, fmb) in the form of feature maps, and as a result, it is possible to generate an interaction feature amount p+ that reflects the relationship between unit feature amounts (T, B, M, A, fm, fb, fmb).

[0075] Finally, the frequency information generation unit 112 uses the interaction feature p+ generated as described above to generate frequency information y^, which indicates the frequency of occurrence of each combination of events related to the unit feature (T, B, M, A, fm, fb, fmb), using the following formula (3) y^=σ(W CIN '(p+)) This is determined by the following: Here, σ is the sigmoid function, and W CIN ' is a transformation matrix operator. Note that the CIN algorithm and transformation matrix operator W described above are also relevant. CIN In this embodiment, the training of ' is carried out using the training data explained with reference to Figure 2.

[0076] The above describes the process of generating interaction features (frequency information) using CIN. However, as another embodiment, it is also possible to generate interaction features, and thus frequency information, by using a fully-connected deep neural network (DNN) algorithm instead of the CIN algorithm.

[0077] Specifically, the D-dimensional embedding vector e is generated from each unit feature (T, B, M, A, fm, fb, fmb) of the ubiquitous feature X. i The set {e1, e2, e..., m} (i=1,2,···, m) m From}, the nodes of the input layer are constructed, then the weighted sum of these nodes is taken to form the nodes of the first hidden layer, and so on, this node weighting process is repeated up to a predetermined number of hidden layers, and the weighted sum of the nodes of the last hidden layer is taken to form the nodes of the output layer, and from these nodes, interaction features x that reflect the relationships between unit features are obtained. DNN It can generate [this].

[0078] In this case, the frequency information generation unit 112 generates the interaction feature x DNN Using this, frequency information y^, which indicates the frequency of occurrence of each combination of events related to each unit feature (T, B, M, A, fm, fb, fmb), is expressed by the following equation (4) y^=σ(W DNN '(x DNN )) This is determined by W. DNN ' is also a transformation matrix operator. Furthermore, the fully connected DNN (FC-DNN) algorithm and the transformation matrix operator W described above are also relevant. DNN In this embodiment, training can also be performed using the training data explained with reference to Figure 2.

[0079] Furthermore, as yet another embodiment of the interaction feature (frequency information) generation process, interaction features, and thus frequency information, can also be generated using DNN algorithms known in the field of product and service recommendation technology, instead of CIN algorithms or fully connected deep DNN algorithms. Examples of such recommendation DNN algorithms include Non-Patent Document 1, "Matrix Factorization," [online], [Retrieved September 18, 2022], Internet.<https: / / developers.google.com / machine-learning / recommendation / collaborative / matrix> Examples of algorithms for Matrix Factorization disclosed in [publication name] include [publication name].

[0080] In this Matrix Factorization, the product UV is the product of the embedding matrix U, which represents the characteristics of the user group, and the embedding matrix V, which represents the characteristics of the product / service. T The model is trained to approach the feedback matrix A, making it possible to acquire embedded representations of each user and product / service capable of generating suitable recommendation information. Here, the product UV T This quantity precisely reflects the relationship between the user and the product / service.

[0081] Furthermore, as another embodiment of the interaction feature (and frequency information) generation process, it is also possible to use xDeepFM disclosed in Non-Patent Document 1 mentioned above in the explanation of CIN. This embodiment will be explained below with reference to Figure 5.

[0082] Figure 5 is a schematic diagram illustrating yet another embodiment of the interaction feature (and frequency information) generation process according to the present invention.

[0083] According to the embodiment shown in Figure 5, the interaction feature generation unit 112a (of the frequency information generation unit 112) is, (a) D-dimensional embedding vector e, corresponding to each unit feature (T, B, M, A, fm, fb, fmb), generated from the ubiquitous feature X. i Using (i=1,2,···,m), the pre-trained CIN algorithm generates interaction features (CIN interaction features) p+, (b) Similarly, the above D-dimensional embedding vector e i Using (i=1,2,···,m), the pre-trained FC-DNN algorithm generates the interaction features (DNN interaction features) x DNN Generate, (c) Each unit feature (T, B, M, A, fm, fb, fmb) of the eccentric feature X is used as an explanatory variable, and the transformation matrix operator W (described later) is used. LINEAR A pre-trained linear regression analysis process, which uses "linear regression features that become information equivalent to frequency information by applying the process" as the target variable, is applied to the localized feature X to generate linear regression feature a.

[0084] Next, the frequency information generation unit 112, (d) Using the generated features from (a) to (c) above, frequency information y^, which indicates the frequency of occurrence of each combination of events related to the unit features (T, B, M, A, fm, fb, fmb), is calculated using the following formula (5) y^=σ(W CIN (p+)+W DNN (x DNN )+W LINEAR (a)) W CIN ,W DNN ,W LINEAR : Transformation matrix operator It will be determined by [the relevant factor]. Here, the processes described in (a) to (d) above are performed by xDeepFM.

[0085] Thus, in this embodiment, it is possible to generate frequency information as a result of incorporating the relationships between unit features (T, B, M, A, fm, fb, fmb) in a more multifaceted way. The above-mentioned CIN algorithm, FC-DNN algorithm, linear regression analysis process (linear regression equation), and transformation matrix operator W are used. CIN , W DNN and W LINEAR This training can also be carried out using the training data explained with Figure 2.

[0086] <Method for determining the number of targets> Returning to the functional block diagram of Figure 1, the mesh-based existence probability calculation unit 113a of the target number determination unit 113 is, in this embodiment, (a) Frequency information y^ indicating the frequency of occurrence of each combination of events related to the unit feature quantities (T, B, M, A, fm, fb, fmb) generated by the frequency information generation unit 112. T,B,M,A Obtain, (b) Frequency information y^ T,B,M,A Therefore, the probability of a terminal 2 (or its user) located within the base station area of ​​B during the time period of T being located within the mesh (area) of M, P(M|T, B), is given by the following equation: (6) P(M|T, B) = Σ A' (y^ T,B,M,A' ) / Σ M' Σ A' (y^ T,B,M',A' ) It is calculated by Σ. Here, A' (y^ T,B,M,A' ) is the frequency information (y^) calculated for all set attributes (A'). T,B,M,A' This is the sum of A' in ). Also, Σ M' Σ A' (y^ T,B,M',A' ) is frequency information (y^) calculated for all set meshes (M') and all set attributes (A'). T,B,M',A'This is the sum of M' and A' in ). Incidentally, P(M|T, B) calculated by the above equation (6) is what is known as a probability map.

[0087] Next, the target number determination unit 113 calculates the probability of presence within the area P(M|T, B) and the observed value UU of the number of terminals 2 located within the base station area related to B during the time period related to T. T B Therefore, the number of people located within the mesh area related to M during the time period related to T (information related to the number of subjects) n^ T M The following equation (7) n^ T M =( α T ) -1 ×Σ B' P(M|T, B')×UU T B' It is calculated by [method].

[0088] Here, α in equation (7) above T This α is the ratio of the number of users who have authorized the provision of GPS positioning information (authorized users) to the population of a designated area (for example, all 47 prefectures) during the time period related to T. In other words, it is the probability that a person located within the designated area is an authorized user. T This can be a pre-set value or an estimated value based on survey results, and if the carrier is a telecommunications company, the actual value can be obtained. In any case (α T ) -1 This is the so-called scaling factor. Also, Σ B' P(M|T, B')×UU T B' This is calculated for all configured base station (area) IDs (B') (e.g., nationwide) and P(M|T, B') × UU T B' This is the sum of B' in the given context.

[0089] Furthermore, the following formula is used to describe the modification method. (8) P(M|T, B, A)=y^ T,B,M,A / ΣM' (y^ T,B,M',A ) Using the probability of existence within the area P(M|T, B, A) calculated by the following equation, (9) n^ T M,A =( α T ) -1 ×Σ B' P(M|T, B',A)×UU T B',A The number of people (information related to the number of targets) that are located within the mesh area related to M during the time period related to T and possess the attributes related to A is n^ T M,A It is also possible to calculate this.

[0090] Here, the probability of presence within the area P(M|T, B, A) calculated by equation (8) above is the probability that terminal 2 (or its user) located (exists) within the area related to B and possessing the attributes related to A is located (exists) within the mesh (area) related to M. Also, in equation (9) above, UU T B',A This represents the observed number of terminals 2 (unique users) located within the base station area related to B' during the time period related to T and possessing attributes related to A.

[0091] Thus, by using a localized feature X that includes attribute feature A as a unit feature, the probability of a terminal 2 (user) located within a certain base station area and possessing attribute A can be calculated (probability map), which is the probability that the terminal 2 is located within a certain mesh and possesses attribute A. This, in turn, makes it possible to calculate the number of people located within a specific mesh and possessing attribute A (information related to the number of targets). Similarly, by using a localized feature X that includes time feature T as a unit feature, the probability of a terminal 2 (user) located within a certain base station area during a time period (time interval) related to T can be calculated (probability map), which is the probability that the terminal 2 is located within a certain mesh and possesses attribute A. This, in turn, makes it possible to calculate the number of people located within a specific mesh during a time period (time interval) related to T (information related to the number of targets). Of course, it is also possible to calculate the population within a specific mesh (information related to the number of targets) even by using a localized feature X that does not include attribute feature A or time feature T.

[0092] Similarly, in the functional block diagram of Figure 1, the target number determination unit 113, in this embodiment, uses the calculation results of the number of people (population) located within a specific mesh, as described above, to generate a high-resolution population distribution graph or a high-resolution population heatmap that represents the population of each mesh (e.g., a 250m square mesh) in a predetermined region (e.g., all 47 prefectures). It is also possible to generate population distribution graphs or population heatmaps that specify time periods (time intervals) or attributes. For example, it is possible to generate a "population heatmap of single men on weekday evenings."

[0093] Furthermore, the results of the population calculation within the mesh, population distribution graphs, and population heatmaps may be output to the UI unit 107 equipped with a display via the input / output control unit 122 and displayed. Alternatively, the results may be transmitted to an external information processing device via the communication control unit 121 and the communication interface unit 101 and used by that device.

[0094] As described in detail above, in this invention, frequency information for estimating the number of targets within an area is derived from a neural network algorithm that incorporates interactions between unit features, using area-related features as explanatory variables. This makes it possible to estimate the number of targets while suppressing the influence of errors related to observation by positioning means.

[0095] Furthermore, by applying the target number estimation method according to the present invention, it is possible to conduct accurate and detailed predictions of pedestrian flow and even traffic behavior in urban areas, thereby developing resilient transportation infrastructure and implementing efficient urban planning to solve the challenges brought about by urbanization. In addition, based on the predicted pedestrian flow information, it becomes possible to implement effective infectious disease control measures tailored to each region and area. In other words, according to the present invention, it becomes possible to contribute to achieving United Nations-led Sustainable Development Goals (SDGs) Goal 9, "Build resilient infrastructure, promote sustainable industrialization and expand innovation," Goal 11, "Make cities inclusive, safe, resilient and sustainable," and Goal 3, "Ensure healthy lives and promote well-being for all at all ages."

[0096] Various modifications, alterations, and omissions within the scope of the technical concept and viewpoint of the present invention can be easily made to the various embodiments of the present invention described above by those skilled in the art. The above description is merely illustrative of embodiments and does not impose unnecessary restrictions. The present invention is limited only to what is limited by the claims and their equivalents. [Explanation of symbols]

[0097] 1. Communication equipment (target number estimation device) 101 Communication Interface Section 102 Mesh Information Storage Unit 103 User Information Storage Unit 104 GPS positioning information storage unit 105 Base station information storage section 106 Unique User (UU) Statistics Storage Section 107 User Interface (UI) Section 111 Unevenly distributed feature generator 112 Frequency Information Generation Unit 112a Interaction Feature Generation Unit 113 Target number determination unit 113a Mesh Probability Calculation Unit 121 Communication Control Unit 122 Input / Output Control Unit 2 terminals 3,B1,B2,B3 Base station

Claims

1. A target number estimation program that estimates the number of targets located within a plurality of zones included in an area which is a unit of location of an target identified by a first positioning means, A localized feature generation means that generates localized features that include, as unit features, a localized feature quantity which is a feature quantity relating to events or things related to the stay or movement of the subject in the area, a first identification feature quantity which is an identifier for the area or the first positioning means, and / or a second identification feature quantity which is an identifier for the area; Frequency information generation means that generates interaction features that reflect the relationships between the unit features, using a pre-trained machine learning algorithm that incorporates the interactions between the unit features from the unevenly distributed features, and generates frequency information relating to the frequency of occurrence of combinations of events related to each of the unit features from the result of applying a pre-trained corresponding transformation matrix operator to the interaction features, A means for determining the number of targets that determines information relating to the number of targets located within the area, based on the frequency information, and the probability that the target is located within the area, based on the probability and the observed number of targets. to make the computer work, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated for each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded localized feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction feature as the interaction feature, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which causes the computer to function to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable and the corresponding transformation matrix operator is applied to the linear regression feature which contains information about the frequency of occurrence of combinations of events related to each of the unit features, and wherein the computer is functioning to generate the set of the CIN interaction feature, the DNN interaction feature, and the linear regression feature as the interaction feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm That is A target number estimation program characterized by the following features.

2. The aforementioned uneven feature generation means generates the uneven feature which further includes attribute feature quantities, which are feature quantities relating to the attributes of the target, as unit feature quantities. The object number determination means calculates the probability of an object being located within the area and possessing the attribute from the frequency information, and determines information regarding the number of objects located within the area from the probability of an object being located within the area and the observed number of objects being located within the area and possessing the attribute. The target number estimation program according to feature 1.

3. The aforementioned uneven feature generation means generates an uneven feature that further includes a time feature, which is a feature related to a time interval, as a unit feature. The object number determination means calculates the probability of an object being located within the area during a given time interval from the frequency information, and determines information regarding the number of objects located within the area from the probability of an object being located within the area during a given time interval and the observed number of objects located within the area during that time interval. The target number estimation program according to feature 1.

4. The aforementioned uneven feature generation means generates the uneven feature which further includes, as unit features, an attribute feature which is a feature related to the attribute of the target and a time feature which is a feature related to the time interval. The object number determination means calculates the probability of an object being located within the area and possessing the attribute from the frequency information, and determines information relating to the number of objects located within the area from the probability of an object being located within the area and possessing the attribute in that time interval. The target number estimation program according to feature 1.

5. The target number estimation program according to claim 1, characterized in that the biased feature generation means generates biased features that further include, as unit features, a first identification feature as an identifier of the area or the first positioning means for a certain time interval, a previous first identification feature as an identifier of the area or the first positioning means for a time interval prior to the certain time interval, and / or a subsequent first identification feature as an identifier of the area or the first positioning means for a time interval later than the certain time interval.

6. A target number estimation program that estimates the number of targets located within a plurality of zones included in an area which is a unit of location of an target identified by a first positioning means, A means for generating a localized feature that generates localized feature quantities that include, as unit features, localized feature quantities which are features related to events or things related to the stay or movement of the subject in the area, and attribute feature quantities which are features related to the attributes of the subject. Frequency information generation means that generates interaction features that reflect the relationships between the unit features, using a pre-trained machine learning algorithm that incorporates the interactions between the unit features from the unevenly distributed features, and generates frequency information relating to the frequency of occurrence of combinations of events related to each of the unit features from the result of applying a pre-trained corresponding transformation matrix operator to the interaction features, A means for determining the number of objects located within an area determines information relating to the number of objects located within an area, based on the frequency information, and the probability of an object being located within an area and possessing the attribute being located within that area. to make the computer work, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated for each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded localized feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction feature as the interaction feature, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which causes the computer to function to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable and the corresponding transformation matrix operator is applied to the linear regression feature which contains information about the frequency of occurrence of combinations of events related to each of the unit features, and wherein the computer is functioning to generate the set of the CIN interaction feature, the DNN interaction feature, and the linear regression feature as the interaction feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm That is A target number estimation program characterized by the following features.

7. A target number estimation program that estimates the number of targets located within a plurality of zones included in an area which is a unit of location of an target identified by a first positioning means, A means for generating a localized feature that generates localized features that include, as unit features, localized features that are features relating to events or things related to the stay or movement of the subject in the area, and localized features that are features relating to a time interval. Frequency information generation means that generates interaction features that reflect the relationships between the unit features, using a pre-trained machine learning algorithm that incorporates the interactions between the unit features from the unevenly distributed features, and generates frequency information relating to the frequency of occurrence of combinations of events related to each of the unit features from the result of applying a pre-trained corresponding transformation matrix operator to the interaction features, A means for determining the number of objects located within an area determines information relating to the number of objects located within an area, based on the frequency information, and the probability of an object being located within an area within a given time interval, which is the probability of an object being located within an area within that time interval, based on the probability of an object being located within an area and the observed number of objects located within that area within that time interval. to make the computer work, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated for each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded localized feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction feature as the interaction feature, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which causes the computer to function to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable and the corresponding transformation matrix operator is applied to the linear regression feature which contains information about the frequency of occurrence of combinations of events related to each of the unit features, and wherein the computer is functioning to generate the set of the CIN interaction feature, the DNN interaction feature, and the linear regression feature as the interaction feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm That is A target number estimation program characterized by the following features.

8. The target number estimation program according to any one of claims 1 to 7, characterized in that the uneven feature generation means generates an uneven feature that further includes a first positioning feature, which is a feature related to the installation environment, specifications, or performance of the first positioning means, as a unit feature.

9. The target number estimation program according to any one of claims 1 to 7, characterized in that the uneven distribution feature generation means generates the uneven distribution feature which further includes as a unit feature a positional relationship feature which is a feature relating to the positional relationship between the area and the point related to the first positioning means.

10. The target number estimation program according to any one of claims 1 to 7, characterized in that the machine learning algorithm and the transformation matrix operator are trained with training data that includes a localized feature which includes the localized feature in the area identified by the second positioning means as a unit feature, and frequency information which is ground truth data relating to the frequency at which combinations of events related to each of the unit features included in the localized feature occur.

11. The subject in question is a human being, including the user of the terminal. The first positioning means is a base station, and the observed number of the target is the observed number of the terminal connected to the base station. The second positioning means is a GPS (Global Positioning System) positioning means installed in the terminal. The target number estimation program according to feature 10.

12. A device for estimating the number of objects located within a plurality of zones included in an area which is a unit of location of an object identified by a first positioning means, A localized feature generation means that generates localized features that include, as unit features, a localized feature quantity which is a feature quantity relating to events or things related to the stay or movement of the subject in the area, a first identification feature quantity which is an identifier for the area or the first positioning means, and / or a second identification feature quantity which is an identifier for the area; Frequency information generation means that generates interaction features that reflect the relationships between the unit features, using a pre-trained machine learning algorithm that incorporates the interactions between the unit features from the unevenly distributed features, and generates frequency information relating to the frequency of occurrence of combinations of events related to each of the unit features from the result of applying a pre-trained corresponding transformation matrix operator to the interaction features, A means for determining the number of targets that determines information relating to the number of targets located within the area, based on the frequency information, and the probability that the target is located within the area, based on the probability and the observed number of targets. It has, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated for each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded localized feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction feature as the interaction feature, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which causes the computer to function to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable and the corresponding transformation matrix operator is applied to the linear regression feature which contains information about the frequency of occurrence of combinations of events related to each of the unit features, and wherein the computer is functioning to generate the set of the CIN interaction feature, the DNN interaction feature, and the linear regression feature as the interaction feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm That is A target number estimation device characterized by the following features.

13. A system for estimating the number of objects located within a plurality of zones that are included in an area which is a unit of location of an object identified by a first positioning means, A localized feature generation means that generates localized features that include, as unit features, a localized feature quantity which is a feature quantity relating to events or things related to the stay or movement of the subject in the area, a first identification feature quantity which is an identifier for the area or the first positioning means, and / or a second identification feature quantity which is an identifier for the area; Frequency information generation means that generates interaction features that reflect the relationships between the unit features, using a pre-trained machine learning algorithm that incorporates the interactions between the unit features from the unevenly distributed features, and generates frequency information relating to the frequency of occurrence of combinations of events related to each of the unit features from the result of applying a pre-trained corresponding transformation matrix operator to the interaction features, A means for determining the number of targets that determines information relating to the number of targets located within the area, based on the frequency information, and the probability that the target is located within the area, based on the probability and the observed number of targets. It has, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated for each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded localized feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction feature as the interaction feature, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which causes the computer to function to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable and the corresponding transformation matrix operator is applied to the linear regression feature which contains information about the frequency of occurrence of combinations of events related to each of the unit features, and wherein the computer is functioning to generate the set of the CIN interaction feature, the DNN interaction feature, and the linear regression feature as the interaction feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm That is A target number estimation system characterized by the following features.

14. A method for estimating the number of objects located within a plurality of zones included in an area which is a unit of location of an object identified by a first positioning means, A step of generating a localized feature that includes, as a unit feature, an area feature which is a feature related to events or things related to the stay or movement of the subject in the area, a first identification feature as an identifier for the area or the first positioning means, and / or a second identification feature as an identifier for the area; The process involves generating interaction features that reflect the relationships between the unit features using a pre-trained machine learning algorithm that incorporates the interactions between the unit features, applying a pre-trained corresponding transformation matrix operator to these interaction features, and generating frequency information related to the frequency of occurrence of each combination of events associated with the unit features. The steps include: calculating the probability that the target is located within the area based on the frequency information; and determining information regarding the number of targets located within the area based on the probability and the observed number of targets. It has, The trained machine learning algorithm is A trained CIN algorithm is used to generate a CIN (Compressed Interaction Network) interaction feature, which is generated by first generating a feature matrix from multiple unit features contained in the embedded eccentric feature, generating an nth feature tensor from the nth feature matrix and the first feature matrix, applying a filter to the nth feature tensor while sliding along the dimensional axis corresponding to the dimension of the unit feature to generate multiple feature maps, and generating the (n+1)th feature matrix from these multiple feature maps, repeating this process until the value of n increases by 1 from 1 until it reaches a predetermined integer value, and then applying a pooling process to each of the multiple feature maps generated for each value of n to generate the CIN (Compressed Interaction Network) interaction feature as the interaction feature. A pre-trained DNN algorithm, or, if necessary, constructs the input layer nodes from a set of embedding vectors generated from each of the multiple unit features contained in the embedded localized feature, constructs the first hidden layer nodes by taking the weighted sum of the nodes in the input layer, then constructs the nth hidden layer nodes by taking the weighted sum of the nodes in the (n-1)th hidden layer, repeats this process until the value of n increases by 1 from 2 until it reaches the set number of hidden layers, constructs the output layer nodes by taking the weighted sum of the nodes in the last hidden layer, and then uses the constructed nodes from the input layer to the output layer to generate the DNN (Deep Neural Network) interaction feature as the interaction feature, A combination of (a) the trained CIN algorithm, (b) the trained DNN algorithm, and (c) the trained linear regression algorithm which causes the computer to function to generate the linear regression feature as an interaction feature by applying a linear regression analysis process to the eccentric feature, where each of the multiple unit features contained in the eccentric feature is an explanatory variable and the corresponding transformation matrix operator is applied to the linear regression feature which contains information about the frequency of occurrence of combinations of events related to each of the unit features, and wherein the computer is functioning to generate the set of the CIN interaction feature, the DNN interaction feature, and the linear regression feature as the interaction feature, and the combination of the trained CIN algorithm, the trained DNN algorithm, and the trained linear regression algorithm That is A method for estimating the number of targets, performed by a computer, characterized by the above.

Images