JP4799120B2 - Intention inference system, method and program using personal behavior characteristics - Google Patents

Description

  The present invention relates to an intention estimation system and method using personal behavior characteristics in a tagged world and a computer program for realizing the system or method.

  Although it is said that the ubiquitous environment is being prepared by the advancement of network technology and the development of embedded computers, in reality, only those who are familiar with small computers and network systems can master the ubiquitous environment. . In order to realize a truly ubiquitous world, when a user unfamiliar with computers and networks behaves according to an intention as usual, the behavior is recognized on the ubiquitous environment side and adapted to the intention of the user. Service must be provided. For example, when a user tries to go out, the service that turns off the air-conditioner that was forgotten to turn off, checks the source of the kitchen fire, and remotely calls the elevator to the floor where the user lives is provided without sending a special command. A system that can do this is desired.

Anind, K, Dey, et al., CybreMinder: A Context-Aware System for Supporting Reminders.Proceedings of the 2nd International Symposium on Handheld and Ubiquitous Computing (HUC2K), 2000. Cory, D, Kidd, et al., The Aware Home: A Living Laboratory for Ubiquitous Computing Research.Proceedings of CoBuild'99, pp. 19 1-198, 1999. Matthias, Grimm, et al., Towards a Framework for Mobile Knowledge Management.Proceedings of the 4th International Conference on Practical Aspects of Knowledge Management, pp.326-338, 2002. Jani, Mantyjarvi, et al., Sensor Signal Data Set for Exploring Context Recognition of Mobile Devices.In Workshop "Benchmarks and a database for context recognition" in conjunction with the 2nd Int. Conf. On Pervasive Computing, pp. 18-23, 2004. Panu, Korpipaa, et al., Bayesian approach to sensor-based context awareness, Personal and Ubiquitous Computing, Vol.7, Issue 2, pp.1 1 3-124, 2003. Aoki Shigeki, et al. Recognition of action patterns with attention to the position and posture of a person. IEICE Technical Report, PRMU2002-82, WIT2002-25, PP.13-18.2002. Mike, Perkowitz, et al., Mining Models of Human Activities from the Web.Proceedings of the 13th international conference on World Wide Web, pp.573-582, 2004. Taketoshi Mori, et al., Sensing Room: Distributed Sensor Environment for Measurement of Human Daily Behavior, First International Workshop on Networked Sensing Systems (INSS2004), pp.40-43, 2004. Diane J. Cook, et al., MavHome: An Agent-Based Smart Home, IEEE International Conference on Pervasive Computing and Communications (PerCom), pp.521-524, 2003. Hiroyuki Yamahara, et al., Behavior Recognition Method Information Science and Technology Forum (FIT) Proceedings Focusing on Partial Order Relations in Individual Habitual Behavior, 2005. Yoichi Motomura, "Bayesian Network Softwares", Japanese Society for Artificial Intelligence Poetry V0117, No5, pp.559-565. 2002. Gregory, F, Cooper, et al., A Bayesian Method for the Induction of Probabilistic Networks from Data.Machine Learning, Vol. 9, pp.309-347., 1992. Satoshi Kaede, et al., Context Aware Operation Reproduction for Safety Driving, 11th International Conference on Human-Computer Interaction (HCI), 2005. Kyoko Fujiwara, et al., Modeling behavior patterns using RFID access history, Proceedings of Information Technology Forum (FIT), 2005. BayoNet http://www.msi.co.jp/BAYONET/

  At present, unique identifiers (ucodes) are allocated to various “things” and “locations” in the real world, and the purpose is to provide context-aware information corresponding to those “things” and “locations”. As a context-aware application, research has been conducted on providing a service according to a user's location and time (Non-Patent Documents 1 and 2). In addition, research that supports users using non-patent devices using a variety of sensors in a ubiquitous environment (Non-Patent Document 3) and research that recognizes user contexts (Non-Patent Document 4) It is. In addition to these human behavior recognition, research has also been conducted to infer current behavior from past behavior (Non-Patent Document 5). In addition, many studies have been conducted to recognize user behavior and use it to propose services according to user behavior. For example, in an environment in which an omnidirectional sensor that can capture 360-degree images at a time is installed in a room, a user's behavior pattern is extracted by an image processing technique from a trajectory that the user moves through the room (Non-Patent Document 6). ). In addition, research is also being conducted to instruct human behavior using the RFID system (Non-patent Document 7).

  As described above, many studies are currently underway to recognize human movements, movement paths, and locations, and to infer their intentions. They try to infer intentions using behavioral features common to everyone. However, each person has behavioral characteristics. Therefore, it is virtually impossible to create a behavior pattern that can be applied to all humans, and it is difficult to recognize each person's behavior with high accuracy when such a pattern is created.

  In other words, each human being has habits and customs. The above-described researches in Non-Patent Documents 1 to 7 acquire human positions and movement histories, and do not recognize or infer user behavior based on individual characteristics. For example, if you are going out after deciding whether you want to clean or go out of the entrance to the user at the entrance, turn off your forgotten items and appliances when you leave the living room, and turn the elevator when you leave the entrance. It is not possible to provide various detailed services such as calling to the floor. In order to provide such a fine-grained service, it is necessary to recognize and infer behavior in consideration of the habits and habits of the user.

  In connection with this, research has already been carried out in which various sensors are installed in the user's living space to recognize or guess the user's behavior (Non-Patent Documents 8 and 9). However, since existing research recognizes and estimates user behavior using a wide variety of sensors, high costs are required.

  Therefore, the present invention realizes the provision of fine-tuned services by extracting the characteristics of individual behavior using the behavior that the user has performed in the past and estimating the intention of the user using the characteristics. And it aims at providing the said service at low cost.

An intention inference system using personal behavior characteristics according to the present invention is:
Provided in each of a plurality of objects touched by a user such as an electrical appliance or a lighting switch, a bag, a mobile phone, a wallet, etc. existing in a space such as an office space, a school, a store space, or a home, and identifies the plurality of objects A plurality of RFID tags each pre-recorded information for,
A portable server including an RFID reader carried by a user moving in the space;
A processing unit that receives information on the RFID tag from the portable server and performs processing;
The portable server reads the information of the RFID tag with the RFID reader when the user moves within the readable range of the RFID tag,
The processing unit accumulates the object touched or used by the user based on the information of the RFID tag read by the RFID reader as a record of the user's action, and the accumulated record of the user's action is created in advance. The user's intention is estimated by comparing with the behavior pattern.

The present invention is provided in each of a plurality of objects touched by a user, such as a switch such as an electrical product or a lighting, a bag, a mobile phone, a wallet, etc. existing in a space such as an office space, a school, a store space, and a home. A user based on a record of actions accumulated by a system including a plurality of RFID tags each pre-recording information for identifying an object, and a portable server including an RFID reader carried by the user moving in the space A method of guessing the intention of
In the portable server, when a user moves within the RFID tag readable range, reading the RFID tag information with the RFID reader;
Accumulating an object touched or used by the user based on information of the RFID tag read by the RFID reader as a record of the user's behavior;
Estimating the user's intention by comparing the accumulated record of the user's behavior with a behavior pattern created in advance.

Preferably, further, generating a behavior pattern template using a Bayesian network from behavior records accumulated for a plurality of users;
And refining the behavior pattern template generated based on the accumulated record of the user's behavior, and generating a behavior pattern for the user as behavior characteristics reflecting personal habits and habits.

The program according to the present invention is for causing a computer to execute the above method.
The program according to the present invention is recorded on a recording medium, for example.
Examples of the medium include EPROM devices, flash memory devices, flexible disks, hard disks, magnetic tapes, magneto-optical disks, CDs (including CD-ROMs and Video-CDs), DVDs (DVD-Videos, DVD-ROMs, DVD-s). RAM), ROM cartridge, RAM memory cartridge with battery backup, flash memory cartridge, nonvolatile RAM cartridge, and the like.

  A medium is a medium in which information (mainly digital data, a program) is recorded by some physical means, and allows a processing device such as a computer or a dedicated processor to perform a predetermined function. is there.

  In the embodiment of the present invention, a low-cost 13.56 MHz band proximity RFID tag is attached to an object existing in the user's living space. A user carries a portable small server to which an RFID reader is connected. Objects touched by the user are recorded in chronological order using the RFID system. This record is a personal action log. The user's intention is estimated by comparing the action log with an action pattern indicated by an object frequently touched by the user who acts with a specific intention. Therefore, a method for generating a template for creating a pattern for recognizing an individual's behavior in advance from the behavior history of a large number of people, and by applying an individual's behavior history to the template, The following describes how to refine behavior patterns.

An example of a space covered by the system according to the embodiment of the invention is shown in FIG.
FIG. 1 shows a plan view of a room in an apartment, and each of a plurality of objects such as a personal computer PC, a table, a television receiver TV switch, a kitchen kitchen range knob, and a front shoe box placed in this room. RFID tags are embedded in the vicinity of them. The user carries the portable server Pocket Assistant, and stores the RFID tag information of each object, that is, the presence / absence of the user's access to each object and the order (time) thereof as action records (behavior log). A server (processing unit) may be provided as necessary. The server communicates with the mobile server via a wireless LAN.

  The world in Fig. 1 is called Tagged Wor1d. Tagged World is a world where a close RFID tag of 13.56MHz band is embedded in an object existing in the home to recognize human behavior, and a user with an RFID reader acts in it. Since the proximity RFlD system with a recognition distance of about 2cm is used, the object touched by the user can be recorded accurately. A server that includes a RFID reader and is carried by a user is called a Pocket Assistant.

  The RFID tag attached to the object touched by the user in daily life is read by the RFID reader connected to the Pocket Assistant, and these logs are accumulated in the Pocket Assistant as a record of the user's actions.

  Furthermore, Pocket Assistant not only accumulates the access history to RFID Tag, but also uses this accumulated behavior log to recognize the user's habits and habits from the personal behavior log, Recognize and guess the intention of the user's action based on it. Pocket Assistant, which was able to guess the user's intention, can provide a detailed service according to the user's behavior by transmitting the guessed intention to the ubiquitous environment. Note that this processing may be performed by a server (processing unit).

  In the existing researches of Non-Patent Documents 8 and 9, high costs are required to recognize and infer user actions using a wide variety of sensors. However, in Tagged World, RFID tags in the 13.56 MHz band with a low unit price are required. Since it is installed in a living environment and the user only carries a small portable server to which an RFID reader is connected, the environment can be constructed at a lower cost than in existing research.

  FIG. 2 is an explanatory diagram (functional block diagram) of functions incorporated when constructing Tagged Wor1d. These functions are realized by, for example, a computer system and software.

In FIG. 2, D1 is a behavior log of a large number of people prepared in advance. For a large number of people, for example, action logs are recorded when actions such as “going out”, “visitor”, and “returning home” are performed. D2 is a behavior pattern template generated based on the behavior logs D1 of many people. The action pattern template D2 includes information such as node name, parent node name, probability value, etc. consisting of three levels of act / action / behavior and order order act name, order destination act name, order vs. number of points. Information. D3 is a behavior pattern that is generated from the behavior pattern template D2 and refined (modified) based on the actual personal behavior log D4 of the user who is the target of inference. The user's intention is estimated by comparing the user's actual personal action log D4 with the action pattern D3. The action pattern D3 includes the same data as D2, but includes information in which the probability value / point number of the pattern is updated based on D4. D4 is a personal action log of the user actually collected by Pocket Assistant (see FIG. 4). The personal action log D4 includes information on an object that has been operated or approached, information on the time, order, and the like.
We consider the order of acts as a connection between two acts and define this as an ordered pair. This ordered pair does not need to be a series of acts, thereby expressing the order of the acts flexibly. FIG. 9 shows an example of ordered pair set elements and data. An order pair ID is provided as an identifier of one order pair, and the order pair type is used to determine whether this order pair is common to all people or individual-specific, and the personal habit is extracted. In order to reduce the amount of future calculations, the number of grouping acts in order and the number of grouping acts in order are defined by grouping meaningful sets of acts and simplifying the set, and defining the number of acts belonging to the group. The order source act name and the order destination act name indicate the connecting order of the two acts, and this can be replaced with the order source group name and the order destination group name. The number of ordered pairs is the number of points to be added when this ordered pair appears. This is used for comparing the ordered pair set for each statistical action prepared in advance with the ordered pair set of the actual action log, and adding to the score according to the number of points of the appearing ordered pair. If this score exceeds a certain threshold, the action is determined.

Reference numeral 10 denotes an action pattern template generation function for statistically mining action logs D1 of a large number of people and extracting an action pattern template D2 that becomes a standard action pattern.
Reference numeral 11 denotes an action pattern refinement function that recalculates an individual action log based on the action pattern D3 and extracts an individual action pattern.
Reference numeral 12 denotes a pattern matching function for matching a behavior pattern from an individual behavior log D4 based on the behavior pattern D3.
Reference numeral 13 denotes an action log collection function that accumulates RFID tag information read by the RFID Reader in order of time.

  In Tagged Wor1d, a plurality of action pattern templates D2 necessary for recognizing a user's action and estimating a user's intention corresponding thereto are prepared in advance. This behavior pattern template D2 is created by the Tagged Wor1d service provider based on the behavior log D1 of a large number of people using the behavior pattern template generation function 10 in order to infer the user's behavior necessary for providing the service. .

  The action log D4 used in the tagged world is acquired by the action log collection function 13. When the user carrying the Pocket Assistant touches the object, the action log collection function 13 reads the RFID tag attached to the object by the RFID reader connected to the Pocket Assistant and stores it as a personal action log.

For example, as shown in the flowchart of FIG. 3, Pocket Assistant sends an RFID read command to the RFID reader (S1), and when the RFID tag falls within the reading range, the RFID tag is read (S2).
In more detail, in the case of a passive tag without a built-in battery, it is as follows.
-Send the command [Continuous reading of tag ID] from the host system such as a computer to the reader / writer.
• The tag enters the wireless field generated by the reader / writer.
・ Power is supplied to the tag by radio signal.
A command is transmitted via a radio signal, and the tag receives the command.
-The tag interprets the command and transmits the ID data that it has with a radio signal.
-The reader / writer receives the ID data and sends it to the computer.
-The computer checks the ID data against an internal table (or database on the network) and executes a specific operation.

  The behavior pattern refinement function 11 refines the behavior pattern template D3 in order to provide services based on the behavioral characteristics of each individual, and reflects the personal habits and habits using the user's daily behavior log. A personal behavior pattern D3 is generated as a characteristic. For example, as shown in the flowchart of FIG. 5, an action pattern template D2 is generated from the prepared action logs of a large number of people using a Bayesian network (S11), and the generated action pattern template D2 is used as a user action. Refinement is performed based on the log D4, and an action pattern reflecting the action characteristics of the user is generated.

  Then, the personal action log D4 collected by the pattern matching function is checked against the action pattern D3 (S4 in FIG. 3), the intention of the user is estimated (S5), and the service provision is notified to the ubiquitous platform (same as above). S6).

  However, it is impractical to perform detailed pattern matching each time the user touches an object because the amount of power held by the Pocket Assistant carried by the user is limited. Therefore, the pattern matching function consists of a two-level matching mechanism. The first stage is a collation using the action pattern according to the embodiment of the present invention, and the second stage is a known collation (Non-Patent Document 10) focusing on the order of objects touched by the user. The former is intended to be a filter that reduces the events that are the detailed reference. In the latter, the order relationship of the objects is examined, and a detailed collation is performed as a result of the two stages of work.

Here, the action log collection function and the pattern matching function are always activated while the user spends their daily life, whereas the action pattern refinement function 11 is activated when the user finishes his / her daily life. is there.
Hereinafter, the action pattern template generation function that is a basis for generating the action pattern will be described in detail.

<Action pattern template generation function>
1. Method for Estimating Action Intention The purpose of the present invention is to estimate the intention of action by providing attention to the characteristics of the action that the user has, and to provide a service corresponding thereto. However, behavioral characteristics vary widely among individuals. Therefore, it is practically impossible to create a rule that can handle all the actions of the user to be collated. Therefore, in the embodiment of the present invention, a behavior pattern template to be collated is generated using the behavior logs of a large number of people, and the behavior pattern reflecting the behavior characteristics is generated by refining it with the personal behavior log. To do.

  A Bayesian network based on the K2 algorithm is used to analyze behavior patterns. In this algorithm, the calculation amount becomes enormous due to the possibility of generating a bidirectional link and the possibility of a large number of links between nodes, and it is difficult to analyze the behavior in real time. Therefore, the present invention solves this problem by constructing a network in which the node types of the Bayesian network are classified into three layers of act, action, and behavior. In addition, since the Bayesian network does not consider the order between nodes, a method of statistically patterning the order of acts is introduced. Therefore, when refining an action pattern with an individual action log, the inter-node probability value of the Bayesian network is updated, and a set of order patterns of relatively frequent acts is extracted from the individual action log to Extract the pattern.

  Humans achieve their intentions by accumulating actions one by one. For example, the intention of the action of “going out” is composed of operations such as “preparation of luggage”, “switching off electrical appliances”, and “opening doors”. Each of these actions is necessary to achieve this intent “go out”, and there is a dependency between this intent and action. Therefore, we propose a method for estimating human intentions using Bayesian Network (hereinafter referred to as BN), which assumes that human movements touch each object.

  BN is a method that estimates the probability distribution of a variable you want to know by modeling the target using a network and evaluates the certainty of each possible state by probability. Use graph structures to represent particularly complex dependencies. BN is a probability model that is represented by an acyclic directed graph in which variables having dependencies are connected as directed nodes with conditional probabilities associated with them, and the paths that follow the links do not circulate.

2. Action pattern template generation using K2 algorithm
When using BN, the network structure used when estimating the probability distribution of the variable you want to know greatly affects the estimation results. Therefore, there is a known K2 algorithm (Non-Patent Document 12) as a method for automatically deriving a Network structure from statistical data. The K2 algorithm selects a parent node on which a certain node depends by using the appearance frequency of a set of values that two nodes take. For this reason, there is a possibility that a bidirectional link representing a dependency relationship between nodes may appear, and there is a possibility that an acyclic directed graph is not configured.

  Therefore, when a bi-directional link appears in the BN constructed using the K2 algorithm, a method for converting the constructed BN, which is a circular directed graph, to an effective BN for inferring human intentions is being studied (non-patented). Reference 13). Since the space to be estimated by Non-Patent Document 13 is a limited space such as in a private car, the number of nodes constituting the BN is small. When guessing the intention of a human subject to the present invention, it is necessary to construct a BN that has many human actions as nodes. The K2 algorithm used in the method of Non-Patent Document 13 performs a brute force search to determine a parent node of an arbitrary node when determining the structure of a BN. Therefore, enormous calculation time is required to determine the BN structure. Also, assuming that the number of both links that appear in the BN constructed with the K2 algorithm is n, the maximum number of BN candidates used when estimating human intentions is 2 ^ n (2 to the power of n). . Therefore, it is impossible to store all of them in the Poceket Assistant, which has a limited amount of memory. Therefore, using the method of Non-Patent Document 13 when estimating the intention of an operation performed by a user at home causes inconvenience in terms of calculation time and memory constraints.

3. Model Generation Method Using Modeled Intent The inventor proposes a method (Non-patent Document 14) for modeling human behavior when creating a behavior pattern necessary for inferring an intention from human motion. is doing. In the modeling method shown in Non-Patent Document 14 (hereinafter, behavior modeling method), the minimum unit of human action that can be sensed when the user touches an object is called an act. In addition, a sequence of acts performed to achieve the purpose of opening the door is called an action. Furthermore, a collection of actions habitually performed in a specific situation is called a behavior, which indicates an intention of an action performed by the user. So far, we have described the estimation target as the intention of the action performed by the user, but in order to maintain consistency with the behavior modeling method, the estimation target is expressed as a behavior from now on. Here, it is assumed that the occurrence probability of each action depends on the occurrence probability of each act constituting the action. Similarly, any behavior is assumed to depend on the probability of occurrence of each action that constitutes it. In this way, if the structure of the BN that guesses the behavior is determined based on the behavior modeling method, the structure of the BN is limited to three layers, the calculation time required to determine the BN structure is greatly reduced, and Pocket Assistant has a memory. The restriction problem is also solved.

  In TaggedWor1d, service providers need to acquire basic data for constructing BN with a large number of people as subjects, and build behavior pattern templates that express the behavior that must be inferred when providing each service. There is. The conditional probability table (hereinafter referred to as CPT) possessed by each node constituting the BN for estimating behavior shown in FIG. 6 is generated using the acquired action logs of a large number of people.

  The structure is determined by this behavior modeling method, and a CPT is created in which each node expresses a dependency relationship with its parent node from the behavior logs of many people. The BN constructed in this way is used as an action pattern template to guess the behavior. Individual action logs are added to this action pattern template, and the dependency of each act, action, and behavior that expresses the BN obtained by recalculating CPT as an individual action pattern expresses the human action characteristics. This behavior pattern can be used to evaluate the first stage to infer behavior.

The model generation method in the patent document expresses a set of acts as an action and a set of actions as a behavior, and forms an action pattern model by a Bayesian network. In addition, objects with RFID tags (door knobs, etc.) and actions correspond one-on-one to achieve a three-layer network structure of actions, actions, and behaviors, and generate action pattern templates.
Since the order generation method is not included in the model generation method of the patent document, almost similar network structures are configured with very similar behavior patterns such as going out and visiting customers. Therefore, in the template generation method shown in Non-Patent Document 14, the order is defined and configured between the actions constituting the action. Further, there are a plurality of objects constituting the act, and the plurality of objects are grouped using the expressions “and” and “or” to generate an action pattern template.

<Experimental result>
Hereinafter, the operation of the embodiment of the invention will be described based on specific examples.
For example, assuming a room in an apartment building a Tagged World. In addition to the living space, the space includes a kitchen space in which household appliances such as a gas stove, a sink, and a refrigerator are installed as objects, a entrance space in which shoe boxes and actual doors are installed as objects, and a toilet space. Proximity RFID tags are affixed to objects such as household electrical appliances, lighting switches, bags, mobile phones, and wallets that are installed in the space that the user touches when living. The user carries Pocket Assistant. The antenna of the RFID reader is fixed to the finger of the right hand of the user, for example, and reads the ID number of the RFID tag attached to the object.

  When a user touches an object, Pocket Assistant reads the ID number of the RFID tag read by the RFlD reader, the time stamp at that time, the experiment number, and the number of readings per second in the storage area of Pocket Assistant as an action log. Write. The action log is stored offline in the database server. The BN that is the behavior pattern template is created from the behavior log stored in the database server using a known Bayesian network construction support environment (Non-patent Document 15). The probability calculation of the behavior is performed using the created BN.

  FIG. 7 shows the probability of the behavior pattern template when exiting the living room and exiting the entrance using the behavior pattern template created in advance for the behavior of “going out”, “visitor”, and “returning home” for 17 men and 4 women. The calculation shows that the template shows the user's behavior as going out.

  FIG. 7 shows that almost all the action logs are concentrated in the range of 30% -50% for "visitor" and "return home", whereas 2% of "outing" are 30% -50% and 60% -80%. Divided into two ranges. As shown in FIG. 7, when the subject actually performs the “outing” behavior, the probability of going out is clearly higher (60% -80%) than when the subject performs the other behavior. In FIG. 7, for example, by setting the threshold value to 50%, it is possible to determine that the user's action is “going out”. At this time, “visiting” and “returning home” are erroneously determined as “going out”. Almost disappear. Therefore, it was shown that the behavior pattern template created by this method can guess the human behavior.

  However, if the threshold for detecting behavior is set to 50%, only about 49.2% of behavior can be detected. On the other hand, when the threshold is set to 35%, 96.5% of going out can be detected, but 45.8% of visitors and returning homes are also detected at the same time. For this reason, in order to properly call the pattern matching function at the second stage of the pattern matching function, it is necessary to determine a threshold value that allows the filter to pass as many desired behaviors as possible and to pass as much as possible the behavior other than the intended one. .

  In order to verify the behavior pattern refinement function, is it possible to generate a behavior pattern that reflects the behavior characteristics of subject A by adding subject A's behavior log to the created behavior pattern template and reconstructing the CPT? I investigated.

  In order to verify the effectiveness of the behavior pattern refinement function, the behavior of the subject A is estimated using the refined behavior pattern and the behavior pattern template before refinement. The results were compared using behavioral going out and behavioral return. The probability value indicated by the action pattern template and the probability value indicated by the sophisticated action pattern increased by an average of 9.69% when going out and an average of 5.32% when returning home. This factor was added only when the behavior going out was recalculated for each node constituting the BN (because only the behavior going out case was used and the CPT was recalculated for each node constituting the BN. ) The overall probability value increased as an initial value. However, the probability value when the subject goes out is significantly improved compared to the probability value when the subject goes home.

  In addition, when the subject performed the behavior of going out and going home, it was verified how the behavior detection ratio before and after refinement changes each time the threshold is changed (see FIG. 8). If the threshold is set to 33%, it is possible to detect 75.OO% of going out before refinement, but 100.00% of returning homes are falsely detected as going out. On the other hand, it is possible to detect 100% going out after refinement, but it is also falsely detected that going home is 100% going out. This is due to the fact that the probability value has risen both after going out and going home after the refinement described above. The point of interest in FIG. 8 is the difference in the ratio determined to go out when the threshold is fixed. When the threshold is set to 43%, 96.17% of going out is detected, and 16.67% of returning home is falsely detected as going out, and it is better after refining the action pattern template using subject A's action log. It was shown that many outings were detected and more homecoming were excluded. From the above, it has been found that if the threshold value is dynamically changed, the technique according to the embodiment of the present invention works as a good filter for filtering out the target behavior.

  As a result of conducting the verification experiment, it was shown that the behavior pattern template created by the method according to the embodiment of the present invention can infer human behavior. It was also shown that the behavioral characteristics of an individual can be reflected by using a behavioral pattern that refines the behavioral pattern template using the behavioral log of the individual.

  However, as shown in FIG. 8, it is necessary to set an appropriate threshold value in order to detect more desired behaviors and to exclude behaviors other than more objectives. Need to be set. In order to detect behavior with Pocket Assistant, it is preferable to automatically set a threshold value. Furthermore, it is preferable to consider the order in which the user touches the object.

  According to an embodiment of the present invention, a proximity RFID Tag is pasted on an object existing in a user's living space, and the user carries a small personal server to which an RFID reader is connected and moves in the living space. Based on the above, it is possible to recognize the user behavior based on the individual behavior characteristics by the small server, to guess the intention, and to refine the template created for guessing the intention. It is possible to generate a template for creating a behavior pattern in consideration of individual behavior characteristics from a large number of people's behavior logs.

  The present invention is not limited to the above embodiments, and various modifications can be made within the scope of the invention described in the claims, and these are also included in the scope of the present invention. Needless to say.

It is explanatory drawing of the space covered with the system which concerns on embodiment of invention. It is explanatory drawing (function block diagram) of the function integrated when constructing Tagged Wor1d which concerns on embodiment of invention. It is a schematic process flowchart of the system which concerns on embodiment of invention. It is explanatory drawing of the action log which concerns on embodiment of invention. It is explanatory drawing of the production | generation of the action pattern template and action pattern which concern on embodiment of invention. It is explanatory drawing of each node which comprises the Bayesian network (BN) which guesses behavior. It is an example of the estimation result of the intention by the system which concerns on embodiment of invention. It is explanatory drawing of the behavior (intention) detection ratio before and behind the refinement | purification which concerns on embodiment of invention. It is explanatory drawing of the order pair set definition file element which concerns on embodiment of invention.

Explanation of symbols

10 Action Pattern Template Generation Function 11 Action Pattern Refinement Function 12 Pattern Matching Function 13 Action Log Collection Function

Claims (3)

Provided in each of a plurality of objects touched by a user such as an electrical appliance or a lighting switch, a bag, a mobile phone, a wallet, etc. existing in a space such as an office space, a school, a store space, or a home, and identifies the plurality of objects A plurality of RFID tags each pre-recorded information for,
A portable server including an RFID reader carried by a user moving in the space;
A processing unit for receiving and processing information of the RFID tag from the portable server;
An action pattern template D2 generated based on an action log D1 in which an action log when a plurality of persons perform a predetermined action on the plurality of objects is recorded;
The action log D4 of the user's individual collected by the portable server and including information on the object that the user has operated or approached, its time and its order,
A behavior pattern D3 generated from the behavior pattern template D2 based on the individual behavior log D4,
The action pattern template D2 is an action that is a minimum unit of human actions that can be sensed when the user touches the object, an action that is a sequence of the acts to be performed to achieve the purpose, and a specific Extract node names, parent node names, probability values, order source act names, order destination act names, and order pairs that describe the order of the actions from the three levels of behavior that is a collection of the actions that are habitually performed in the situation And the number of points vs. points given when
The ordered pair is an order of the act that is a connection of the two acts, and an ordered pair ID is provided as an identifier of the ordered pair, and an ordered pair type that is a type of the extracted ordered pair is provided,
The node is an act of touching the object performed by the user and a plurality of acts necessary for achieving a predetermined intention;
The occurrence probability of each action depends on the occurrence probability of each act constituting the action, and the arbitrary behavior depends on the occurrence probability of each action constituting the action,
The behavior pattern D3 includes data similar to the behavior pattern template D2,
In the portable server, when a user moves within the RFID tag readable range, reading the RFID tag information with the RFID reader;
In the processing unit, the information of the object operated or approached by the user based on the information of the RFID tag read by the RFID reader , the time, and the step of accumulating as a record of the user's action,
In the processing unit, the probability value between the nodes is updated using a three-layer Bayesian network based on the behavior pattern template D2 and the individual behavior log D4, and the order of the actions with relatively high frequency is updated. Extracting the user's personal behavior pattern by extracting a set of patterns, and generating a behavior pattern D3 for the user as behavior characteristics reflecting personal habits and habits;
Intention estimation using personal behavior characteristics, characterized in that the processing unit performs a step of estimating the user's intention by comparing the accumulated record of the user's behavior with the behavior pattern D3. system. Provided in each of a plurality of objects touched by a user such as an electrical appliance or a lighting switch, a bag, a mobile phone, a wallet, etc. existing in a space such as an office space, a school, a store space, or a home, and identifies the plurality of objects A plurality of RFID tags each pre-recorded information for,
A portable server including an RFID reader carried by a user moving in the space;
An action pattern template D2 generated based on an action log D1 in which an action log when a plurality of persons perform a predetermined action on the plurality of objects is recorded;
The action log D4 of the user's individual collected by the portable server and including information on the object that the user has operated or approached, its time and its order,
A method of estimating a user's intention based on a behavior record accumulated by a system including a behavior pattern D3 generated from the behavior pattern template D2 based on the individual behavior log D4 ,
The action pattern template D2 is an action that is a minimum unit of human actions that can be sensed when the user touches the object, an action that is a sequence of the acts to be performed to achieve the purpose, and a specific Extract node names, parent node names, probability values, order source act names, order destination act names, and order pairs that describe the order of the actions from the three levels of behavior that is a collection of the actions that are habitually performed in the situation And the number of points vs. points given when
The ordered pair is an order of the act that is a connection of the two acts, and an ordered pair ID is provided as an identifier of the ordered pair, and an ordered pair type that is a type of the extracted ordered pair is provided,
The node is an act of touching the object performed by the user and a plurality of acts necessary for achieving a predetermined intention;
The occurrence probability of each action depends on the occurrence probability of each act constituting the action, and the arbitrary behavior depends on the occurrence probability of each action constituting the action,
The behavior pattern D3 includes data similar to the behavior pattern template D2,
In the portable server, when a user moves within the RFID tag readable range, reading the RFID tag information with the RFID reader;
Accumulating information on the object that the user has operated or approached based on information of the RFID tag read by the RFID reader , its time, and order as a record of the user's actions;
Based on the behavior pattern template D2 and the individual behavior log D4, a probability value between the nodes is updated using a three-layer Bayesian network, and a set of order patterns of the relatively frequent acts is extracted. Extracting the user's individual behavior pattern and generating a behavior pattern D3 for the user as behavior characteristics reflecting personal habits and habits;
An inference method using an individual's behavior characteristic, comprising: comparing the accumulated behavior record of the user with the behavior pattern D3 to infer the user's intention. Provided in each of a plurality of objects touched by a user such as an electrical appliance or a lighting switch, a bag, a mobile phone, a wallet, etc. existing in a space such as an office space, a school, a store space, or a home, and identifies the plurality of objects A plurality of RFID tags each pre-recorded information for,
A portable server including an RFID reader carried by a user moving in the space;
An action pattern template D2 generated based on an action log D1 in which an action log when a plurality of persons perform a predetermined action on the plurality of objects is recorded;
The action log D4 of the user's individual collected by the portable server and including information on the object that the user has operated or approached, its time and its order,
A program for causing a computer to execute a method for estimating a user's intention based on a behavior record accumulated by a system including a behavior pattern D3 generated from the behavior pattern template D2 based on the individual behavior log D4 Because
The action pattern template D2 is an action that is a minimum unit of human actions that can be sensed when the user touches the object, an action that is a sequence of the acts to be performed to achieve the purpose, and a specific Extract node names, parent node names, probability values, order source act names, order destination act names, and order pairs that describe the order of the actions from the three levels of behavior that is a collection of the actions that are habitually performed in the situation And the number of points vs. points given when
The ordered pair is an order of the act that is a connection of the two acts, and an ordered pair ID is provided as an identifier of the ordered pair, and an ordered pair type that is a type of the extracted ordered pair is provided,
The node is an act of touching the object performed by the user and a plurality of acts necessary for achieving a predetermined intention;
The occurrence probability of each action depends on the occurrence probability of each act constituting the action, and the arbitrary behavior depends on the occurrence probability of each action constituting the action,
The behavior pattern D3 includes data similar to the behavior pattern template D2,
Based on the behavior pattern template D2 and the individual behavior log D4, a probability value between the nodes is updated using a three-layer Bayesian network, and a set of order patterns of the relatively frequent acts is extracted. Extracting the user's individual behavior pattern and generating a behavior pattern D3 for the user as behavior characteristics reflecting personal habits and habits;
The program for performing the step which estimates a user's intention by comparing the accumulated record of the user's action with the action pattern D3 .

Images