JP2016097228A - Behavior classification system, behavior classification device, and behavior classification method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To classify a person's behavior based on measurement data of a sensor mounted on a predetermined region of the person.SOLUTION: A behavior classification system holds model information including a rule to discriminate a person's behavior based on a physical feature amount calculated from an acceleration of a predetermined region of the person, and includes an acceleration sensor mounted on the predetermined region of the person, and an analysis part for classifying the person's behavior based on the physical feature amount calculated from the acceleration measured by the acceleration sensor and the model information.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a technique for classifying a person's behavior based on data measured by a sensor attached to the person.

  In recent years, in exercise guidance for preventing lifestyle-related diseases, methods of increasing exercise such as walking and running and non-motor physical activity such as housework in daily life different from exercise have attracted attention worldwide. In order to maintain efficient exercise guidance and exercise motivation, it is necessary to visualize changes in daily behavior in more detail.

  Patent Document 1 discloses a technique for displaying operation contents such as a golf swing form on a display using a wristband type sensor.

US Patent Application Publication No. 2011/0054782 Specification

  The display of the operation content by the wristband type sensor shown in Patent Document 1 is based on the acceleration and gyro data, estimating the position and locus of the hand to which the sensor is added, the estimated hand locus, and the golf swing The body and leg forms defined in advance are displayed together, and it is impossible to estimate anything other than hand movements.

  By attaching a plurality of sensors similar to the above to parts such as legs and waist, it is also possible to estimate and display the movement of parts other than the hand, but it involves the burden of wearing multiple sensors, It is difficult to measure for a long time in daily life.

  In order to solve the above-described problems, the behavior classification system of the present invention holds model information including rules for discriminating the behavior of the person based on the physical feature amount calculated from the acceleration of a predetermined part of the person. An acceleration sensor attached to the predetermined part of the person, and an analysis unit for classifying the person's behavior based on the physical feature amount calculated from the acceleration measured by the acceleration sensor and the model information, It is characterized by having.

  According to one aspect of the present invention, various actions of a person can be classified based on measurement data of a sensor attached to a predetermined part of the person.

It is a block diagram which shows the main structures of the action classification system of Example 1 of this invention. It is explanatory drawing which shows the external appearance of the sensor device of Example 1 of this invention, and typical arrangement | positioning of an acceleration sensor. It is explanatory drawing which shows the typical process sequence of estimation of the part state and action content classification which the server of Example 1 of this invention performs. It is a flowchart which shows the typical process sequence which the server of Example 1 of this invention performs. It is a flowchart which shows the typical process sequence which the server of Example 1 of this invention performs in order to produce | generate a site | part state discrimination | determination model. It is explanatory drawing which shows two types of typical physical feature-values which can be calculated from the triaxial acceleration data of the arm which can be measured with the sensor device of Example 1 of this invention. It is explanatory drawing which shows the typical site | part state which can be discriminate | determined from the triaxial acceleration data of the arm which can be measured with the sensor device of Example 1 of this invention. It is explanatory drawing which shows the typical method in which the server of Example 1 of this invention detects the change point of action from site | part state data, and divides | segments action content. It is explanatory drawing which shows the structural example of the physical feature-value data which the server of Example 1 of this invention hold | maintains. It is explanatory drawing of the 1st example of the site | part state discrimination | determination model in Example 1 of this invention. It is explanatory drawing of the 2nd example of the site | part state discrimination | determination model in Example 1 of this invention. It is explanatory drawing which shows the 1st structural example of the site | part state data which the server of Example 1 of this invention hold | maintains. It is explanatory drawing which shows the 2nd structural example of the site | part state data which the server of Example 1 of this invention hold | maintains. It is explanatory drawing which shows the structural example of the physical feature-value data which the server of Example 1 of this invention hold | maintains. It is explanatory drawing which shows the example of the classification | category of the action content by the server of Example 1 of this invention. It is explanatory drawing which shows the example of a data structure of the action classification data which the server of Example 1 of this invention hold | maintains. It is explanatory drawing which shows the structural example of the screen output in order that the server of Example 1 of this invention may display an action record. It is explanatory drawing which shows the 1st structural example of the screen output in order that the server of Example 1 of this invention may display the result of action analysis. It is explanatory drawing which shows the 2nd structural example of the screen output in order that the server of Example 1 of this invention may output the result of action analysis. It is explanatory drawing of the typical example of the conventional method which estimates action content from the data of a sensor. It is a block diagram which shows the main structures of the action classification system of Example 2 of this invention. It is a block diagram which shows the main structures of the action classification system of Example 3 of this invention. It is explanatory drawing which shows the typical process sequence of the site | part state estimation and action content estimation which the server of Example 3 of this invention performs. It is explanatory drawing which shows the data structural example of the action classification data which the server of Example 3 of this invention hold | maintains.

  Embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 is a block diagram illustrating a main configuration of the behavior classification system according to the first embodiment of the present invention.

  The behavior classification system according to the present embodiment includes a sensor device 1 worn by a user, a PC 2 or a smartphone 3 that communicates with the sensor device 1, and a server 5 that can communicate with the PC 2 or the smartphone 3 via a network 4. Yes. The sensor device 1 transmits the measured sensor data to the server 5 via the PC 2 or the smartphone 3. The server 5 analyzes the sensor data, estimates the movement of each part such as the user's limb, and classifies the action content from the content of the estimated movement of the part. The PC 2 or the smartphone 3 can download the analyzed result from the server 5 and allow the user to browse.

  The sensor device 1 mainly includes a microcomputer (MCU: Micro Control Unit) 10, an acceleration sensor 11, a memory 12, an input unit 13, a display unit 14, and a communication unit 15. The acceleration sensor 11 always measures acceleration due to user movement or the like at a predetermined frequency (typically, about 20 to 1000 times per second). The microcomputer 10 reads the acceleration data measured by the acceleration sensor 11 by controlling communication and records it in the memory 12.

  The sensor device 1 is attached to a predetermined part of the user's body. In the present embodiment, an example of attaching to the wrist portion of the user's forearm will be described.

  The sensor data recorded in the memory 12 is stored in the PC 2 or the smartphone 3 automatically at a timing at which the microcomputer 10 controls the communication unit 15 and can communicate with the PC 2 or the smartphone 3 by wireless or wired, or at any timing of the user. Can be sent to. Data to be transmitted can be transmitted without being read by another user or the like by being encrypted by a known function of the communication unit 15 or the microcomputer 10. The data measured by the acceleration sensor 11 is not only transmitted to the outside, but also analyzed by a simple analysis function recorded in advance in the microcomputer 10 as a program, and the resulting step count or calorie consumption, for example, is displayed. It can be displayed on the part 14.

  The PC 2 and the smartphone 3 can communicate with the sensor device 1 and a server 5 connected to a network 4 such as the Internet. The PC 2 and the smartphone 3 can transfer the sensor data received from the sensor device 1 to the server 5, and can display and operate published information recorded in the server 5.

  The server 5 includes a CPU 51, a communication unit 52, a memory 54, and a database 70. The memory 54 stores a WEB display program 53 and an analysis program 60. The server 5 analyzes the sensor data of the sensor device 1 transmitted from the PC 2 or the smartphone 3 by the analysis program 60, converts it into information on the state of the human part or the action content, and records it in the database 70. The server 5 can also generate an algorithm to be analyzed and the rule itself.

  The database 70 is configured by a storage device such as a hard disk drive. The analysis program 60 and the WEB display program 53 may be stored in this storage device, and a part or all of them may be copied to the memory 54 as necessary.

  The CPU 51 can execute and calculate the processing of the analysis program 60 and the WEB display program 53. That is, in the following description, the processing executed by these programs is actually executed by the CPU 51 in accordance with instructions described in these programs. In other words, when the CPU executes these programs, a functional module that executes processing to be described later is realized. The functional module of the present embodiment is realized by the CPU 51 executing the program as described above, but this is an example, and may be realized by another means, for example, a dedicated logic circuit.

  The communication unit 52 can communicate with other devices such as the PC 2 or the smartphone 3 via the network 4 and transmit / receive data to / from them. The acceleration data received from the sensor device 1 via the PC 2 or the smartphone 3 is recorded in the database 70.

  The WEB display program 53 can open the data recorded in the database 70 to the user of the sensor device 1 via the network 4. The analysis program 60 includes a physical feature amount calculation program 61, a part state estimation program 62, an action content classification program 63, a part state determination model generation program 64, and a part state determination model selection program 65. The database 70 includes acceleration data 80, physical feature data 71, part state data 72, action classification data 73, part state determination model data 74, user profile data 75, and part / physical quantity teacher data 76.

  The physical feature amount calculation program 61 is a program that processes the acceleration data 80 measured and transmitted by the sensor device 1, calculates a feature amount related to human movement or behavior, and records it in the physical feature amount data 71. Only necessary features can be extracted from a large amount of time-series data.

  The part state estimation program 62 is a program that estimates the movement of each part of the user at that time based on the physical feature data 71 and records it in the part state data 72. From the movement characteristics of the sensor device 1, Regarding the movement of the part where the sensor device 1 is not directly attached, the statistical characteristics of the combination of the part states, the distribution bias, and the characteristics and the distribution bias, which are recorded in the part state determination model data 74 in advance. It can be estimated by a dividing algorithm. Further, the part state estimation program 62 receives attribute information such as the gender, physical characteristics such as weight and height of the user who transmitted the acceleration data 80 that is the basis of analysis, and lifestyle characteristics such as hobbies and occupations. Among the user profile data 75 to be included, the corresponding part state determination model data 74 is selected and used by referring to a part or all of the part state determination model selection program 65 selected. This is because the statistical relationship between the movement of the sensor device 1 and the movement and state of each part may vary depending on physical characteristics and lifestyle.

  The action content classification program 63 is a program for classifying the action contents into an arbitrary number based on the part state data 72 and recording the action contents in the action classification data 73, without defining the name or content of the action. The daily action content can be classified only by the combination of the state of each part.

  The part state determination model generation program 64 is a program for generating part state determination model data 74 based on the part / physical quantity teacher data 76 acquired in advance. The part / physical quantity teacher data 76 includes motion data indicating the movement of each part of a person in daily life, and acceleration data of the sensor device 1 recorded at the same time, or physical feature values calculated from the acceleration data. Yes.

  For example, before starting the operation of the behavior classification system of the present embodiment, a plurality of measurement target users wearing the sensor device 1 behave in the same manner as in daily life, and the sensor device 1 has a plurality of times in between. Get acceleration data. At the same time, motion data indicating movements of a plurality of parts of the measurement target user is acquired by a known technique, for example, motion capture which is a known product. The acceleration data acquired in this way or the physical feature amount calculated therefrom and the motion data are associated with each other based on the acquired time, and are recorded as the part / physical quantity teacher data 76.

  The part state discriminating model generation program 64 finds statistical features and distribution bias of physical feature amounts corresponding to each part state based on the relational data between the part state and the physical feature quantity, and the part state based on the statistical feature Are generated and recorded in the part state determination model data 74.

  FIG. 2 is an explanatory diagram illustrating an appearance of the sensor device 1 according to the first embodiment of the present invention and a typical arrangement of the acceleration sensor 11.

  The direction of the triaxial acceleration to be measured (X, Y, Z) is defined by the arrangement of the acceleration sensor 11. Although only one sensor device 1 is shown in FIG. 1 as an example, the server 5 actually processes acceleration data from a plurality of sensor devices 1 attached to different users. The analysis program 60 can perform the same calculation in each sensor device 1 when the arrangement of the acceleration sensor 11 of each sensor device 1 is the same. On the other hand, when the arrangement of the acceleration sensor 11 of each sensor device 1 is different, the sensor device 1 or the server 5 that has received the acceleration data from the sensor device 1 executes a coordinate system conversion process for unifying the orientation. There is a need to.

  FIG. 3 is an explanatory diagram illustrating a typical processing procedure of part state estimation and action content classification executed by the server 5 according to the first embodiment of this invention.

  Specifically, FIG. 3 shows a procedure for processing the acceleration data 80 in the order of the physical feature amount calculation program 61, the part state estimation program 62, and the action content classification program 63. In the feature quantity extraction processing 110, the physical feature quantity calculation program 61 cuts out acceleration data in a predetermined time unit and performs an operation. The section to be cut out depends on the time for estimating the part state. For example, the physical feature quantity calculation program 61 calculates acceleration data every minute if the time unit for estimating the part state is a minute unit, and calculates acceleration data every second if the unit is seconds.

  The main feature amounts calculated by the physical feature amount calculation program 61 are predetermined values of the average value, the maximum value, the minimum value, and the variance value of the amplitude 111 of each axis, the frequency 112, and the arm rotation angles θ113 and φ114. Statistical values such as average, maximum, minimum, variance, etc. of the time interval of, but not limited to, use general statistical values calculated from triaxial acceleration or signal processed values as feature quantities Can do.

  Here, the arm rotation angle θ113 is a pitch angle indicating the amount of movement of raising and lowering the arm about the shoulder or elbow, and the arm rotation angle φ114 is a movement for rotating the forearm about the longitudinal direction of the forearm, that is, the wrist It is a roll angle indicating the amount of movement to twist. The physical feature amount is recorded in the physical feature amount data 71 of the database 70.

  The part state estimation program 62 estimates the part state using the part state determination model data 74 from the physical feature quantity data 71 in the same time interval or close to each other, and the arm state 121, the elbow angle 122, the thigh State 123 and posture 124 are estimated. Since the data necessary for estimation of each part state is different, the part state estimation program 62 extracts only necessary data from the physical feature amount data 71. Further, the part state determination model selection program 65 selects a model that matches the user profile data 75 of the user to be estimated from the part state determination model data 74, and the part state estimation program 62 uses it. Part states such as arm state 121, elbow angle 122, thigh state 123, and posture 124 are recorded in part state data 72 of database 70.

  The action content classification program 63 classifies the action contents by referring to the part state data 72 of a predetermined section and clustering it into a set of data having similar characteristics of the combination of the states of the parts. In the example of FIG. 3, it is classified into action A131 to action D134. Of the data included in the part state data 72, clustering may be performed on all users' data without limiting the conditions, or the data may be limited to a group of users with similar profiles or limited to one user. Then, clustering may be performed. As for the time interval, for example, clustering may be performed for a time interval within a short period such as one week, or a time interval within a long period such as one year or several years may be targeted. The wider the target, the more general classification, but the possibility of similar classification results increases. On the other hand, for example, when targeting only a time section within a short period, data of one user, or data of a plurality of users having similar profiles, the possibility that the result is obtained as a clear classification increases, Will disappear.

  The part state determination model data 74 is generated by the part state determination model generation program 64 from the part / physical quantity teacher data 76 acquired in advance. The part / physical quantity teacher data 76 is a database that associates the part state acquired by the actual user with the physical feature quantity, and the part state discriminating model generation program 64 performs statistical features or distributions by a known process such as machine learning. A rule or a threshold value that is discriminated based on the bias of is recorded as part state discrimination model data 74.

  FIG. 4 is a flowchart illustrating a typical processing procedure executed by the server 5 according to the first embodiment of this invention.

  The processing shown in FIG. 4 is performed by the physical feature quantity calculation program 61, the part state determination model selection program 65, the part state estimation program 62, and the action content classification program 63 in order, and the time based on the acceleration sensor data. The movement and state of each part are estimated, and the action content is classified using the movement and state of each part.

  The server 5 starts the analysis process from step S100 after the server 5 receives the data measured by the sensor device 1. In step S101, the physical feature quantity calculation program 61 reads acceleration sensor data for a period to be analyzed from the received acceleration sensor data. The target period is, for example, an unanalyzed period in the newly received acceleration sensor data of one user. Next, in process S <b> 102, the physical feature quantity calculation program 61 calculates the read acceleration sensor data, calculates the physical feature quantity, and records the result in the physical feature quantity data 71 of the database 70.

  In process S103, the part state determination model selection program 65 reads out user profile data 75 including body information and lifestyle information of the user wearing the sensor device 1 that has measured the data read out in process S102. The part state discriminating model most suitable for the read user profile (for example, the part state discriminating model generated based on the part / physical quantity teacher data acquired from the group of users having a profile similar to the read user profile) Is selected from the part state determination model data 74. An ID for identifying a user corresponding to the data is attached to the measured acceleration sensor data, and is similarly added to the physical feature data 71 and the part state data 72, which are analysis results.

  In the process S104, the part state estimation program 62 reads out the physical feature quantity of the target section to be analyzed from the physical feature quantity data 71, determines the part state by the part state discrimination model selected in the process S103, and obtains the result for each part. Every time, it is recorded in the part state data 72.

  In the process S105, the action content classification program 63 reads each part state from the part state data 72 in order to detect a break of the action contents, and the action change point where the change per time of the state of each part is the largest. Detect as. At this time, for example, when observing the behavior on a daily basis, if the behavior break is made too short (for example, in seconds), it is difficult for a person to understand the behavior for each break later. It is desirable to set the delimiter to an appropriate length that is not too short (for example, in units of one minute). In the process S106, the action content classification program 63 sets an action section (that is, a time zone in which each action is estimated to be performed) by dividing the action contents based on the change point detected in the process S105. To do. In the process S107, the action content classification program 63 classifies the action contents by clustering the divided action sections according to the vectors of the respective part states, and records them in the action classification data 73. In the process S108, the typical process of the present invention is terminated.

  FIG. 5 is a flowchart illustrating a typical processing procedure executed by the server 5 according to the first embodiment of this invention to generate the part state determination model.

  Specifically, FIG. 5 shows a part state determination model generation process that is necessary in advance for executing the process for classifying the behavior shown in FIG. 4, and starts from process S200. In the generation of the part state determination model, the time in which both the motion data of the part for which the determination model is to be generated in advance and the acceleration data measured by the sensor device 1 at the same time as the motion data can be statistically determined Only the length is necessary.

  In process S201, the part state determination model generation program 64 reads motion data of a period to be analyzed that has been measured in advance. In process S202, the part state determination model generation program 64 reads acceleration data 80 having the same time length starting from the same time as the motion data acquired in process S201.

  In the process S203, the part state determination model generation program 64 reads out the profile data 75 of the user who is the target for measuring the motion data to be analyzed and the acceleration data. The user profile data 75 is attribute information indicating the physical information, lifestyle, and the like of the user who has measured motion data and acceleration data.

  If the association between the movement of the arm and other parts that are not directly measured depends on the profile, generating a discriminant model for each group of users with similar profiles increases the accuracy of action classification. User profile data 75 can be used to raise. However, for example, when a very high accuracy is not required, or when it is considered that a sufficiently accurate classification is possible without considering the profile, the part state determination model generation program 64 uses the user profile data 75. A discrimination model that is common to all users may be generated without acquiring. In any case, it is possible to classify with high accuracy by generating a discrimination model based on actually measured data.

  In process S204, the part state determination model generation program 64 sets a time unit to be analyzed. The time unit is, for example, a minute unit or a second unit, and a time unit suitable for the movement of the part to be determined may be set, or may be set according to the granularity of the part state to be obtained. For example, when the action to be finally classified is determined by the minute unit, it is desirable to record the result of determining the part state by the minute unit in the database 70.

  In process S205, the part state determination model generation program 64 calculates a physical feature amount in units of time set in process S204. In the process S206, the part state determination model generation program 64 reads the part state defined by the time unit set in the process S205 from the motion data. The part state and physical feature obtained in steps S205 and S206 are used as teacher data 76. In process S207, the part state discrimination model generation program 64 generates a discrimination model by a statistical discriminant analysis method such as known machine learning. The part state determination model generation program 64 records the determination model generated in step S207 in the part state determination model data 74 in association with the user profile information acquired in step S203.

  FIG. 6 is an explanatory diagram illustrating two types of typical physical feature quantities that can be calculated from the three-axis acceleration data of the arm that can be measured by the sensor device 1 according to the first embodiment of the present invention.

  The number of zero crossings 202 is the total number of times that the acceleration waveform 201 has changed beyond a predetermined threshold per unit time. The pitch angle (θ) 211 is an angle generated by raising and lowering the sensor device 1 with respect to the ground (in other words, the longitudinal direction of the forearm, for example, the angle formed by the horizontal plane and the longitudinal direction of the forearm), and the roll angle (φ). Reference numeral 212 denotes a rotation angle of the forearm about the longitudinal direction of the forearm.

  Specifically, the orientation of the sensor device 1 with respect to the direction of gravitational acceleration (ie, the vertical direction) is calculated from the triaxial acceleration data measured by the sensor device 1. For example, when accelerations in the X, Y, and Z axis directions (see FIG. 2) at time t are x (t), y (t), and z (t), respectively, pitch angle θ (t) and roll at time t The angle φ (t) is calculated by equations (1) and (2), respectively.

  FIG. 7 is an explanatory diagram illustrating a typical part state that can be discriminated from the triaxial acceleration data of the arm that can be measured by the sensor device 1 according to the first embodiment of the present invention.

  When the thigh state is exemplified as the part state, the standing state 221 and the sitting state 222 are defined as the part states corresponding to the thigh angle, and the part state estimation program 62 determines the thigh part state at each time. In daily operations, the usage of the arm in the standing position and the sitting position is greatly different, so that statistical determination can be made from the three-axis acceleration data of the sensor device 1 attached to the arm. Moreover, when the part state of an elbow is illustrated, the elbow extension state 223, the elbow bending state 224, etc. are defined. Also in this example, based on the difference in how to use the arm according to the state of the elbow, the part state of the elbow can be statistically determined. That is, from only the acceleration data of the arm, only the pitch angle 211 in FIG. 6 is known, but depending on the direction and strength of the change in acceleration, the change in acceleration caused by moving the arm by bending the elbow and the shoulder without bending the elbow It can be distinguished from the change in acceleration caused by moving the arm around the axis.

  FIG. 8 is an explanatory diagram showing a typical method in which the server 5 according to the first embodiment of the present invention detects a behavior change point from the part state data 72 and divides the behavior content.

  FIG. 8 shows an example in which the thigh, elbow, and arm region states are used. In this case, the thigh state 231, the elbow state 232, and the arm state 233 each have a continuous or discrete value. For example, g (t) indicating the thigh state 231 may be a discrete value indicating either standing or sitting, h (t) indicating the elbow state 232, and k ( t) may be discrete values indicating the elbow and arm angles, respectively.

  The action content classification program 63 calculates a total change of the thigh state 231, the elbow state 232, and the arm state 233 as the change amount 234. A value f (t) indicating the change amount 234 is calculated by, for example, Expression (3). This is a value obtained by multiplying a change amount of a value indicating each part state by a predetermined weighting coefficient a, b, c, and totaling them. The amount of change in the value indicating the part state is the difference between the part state estimated at time t and the part state estimated at the previous time ti, where i is a time interval for executing part state estimation, for example. It may be.

f (t) = a (g (t) -g (t-i)) + b (h (t) -h (t-i)) + c (k (t) -k (t-i)) (3)

  The action content classification program 63 detects a point where the amount of change 234 exceeds a predetermined threshold 235 as a change point 236 (S105). When the content of the actual action changes, there is a high possibility that the part state also changes. Therefore, it is expected that the action can be appropriately divided based on the change amount of the part state. The threshold value 235 is desirably set so that the interval between the change points 236 is always a certain length or more. Here, the certain length is suitable for a length that allows a person to recognize what he / she was doing later, for example, 10 minutes or more. Actions 238, 239 and the like divided by the change points 236 are sections for classifying the action contents (that is, time zones where each action is estimated to have been performed).

  FIG. 9 is an explanatory diagram illustrating a configuration example of the physical feature amount data 71 held by the server 5 according to the first embodiment of this invention.

  One row of the physical feature amount data 71 indicates one hour unit data of one user. A user ID 301 is an ID for identifying a user. The date and time 302 indicates the actual date and time when the data is measured, and is a time unit suitable for determining the part state. For example, when determining the part state in minutes, it is desirable to calculate the physical feature value in minutes as shown by the date 302. Zero crosses 303, 304, and 305 respectively indicate the number of zero crosses of the X, Y, and Z axes calculated by the method shown in FIG. As physical quantities that are characteristic of the acceleration waveform, other statistical values such as maximum, minimum, average, and variance in time units of amplitude and differential value can be applied. The maximum angles 306 and 307 respectively indicate the maximum values of the pitch angle (θ) 211 and roll angle (φ) 212 during the unit time. In addition, statistical values such as minimum values, average values, and variances of measured values can be applied as physical feature amounts.

  FIG. 10 is an explanatory diagram of a first example of the part state determination model according to the first embodiment of the present invention.

  Specifically, FIG. 10 shows an example of distributions 241 and 242 in the physical feature amount space of the physical feature amounts in the standing state 221 and the sitting state 222. These are the discrimination models for discriminating between the standing state 221 and the sitting state 222 shown in FIG.

  In a person's daily life, there is a statistical bias in the physical feature amount indicating the movement of the arm because the restrictions on how to use the hand are different between standing and sitting. Based on this bias, the physical feature is divided into distributions 241 and 242 by a known machine learning discriminant analysis method, so that whether the thigh is in the standing state 221 or the sitting state 222 is determined based on the physical feature. Is possible. Similar discrimination is possible for other parts. In addition, since this discriminant model also depends on the user's lifestyle and physique, the discriminant model is generated by dividing the user into groups of users having similar profiles based on user profile information that can be known in advance. Is optimal.

  FIG. 11 is an explanatory diagram of a second example of the part state determination model according to the first embodiment of the present invention.

  Specifically, FIG. 11 shows a transition model between states 251, 252, and 253, each of which is an upper arm state. These are discrimination models for discriminating the direction of the upper arm, and are necessary for estimating the elbow bending state 224 and the elbow extension state 223 shown in FIG.

  Usually, the angle of the upper arm cannot be known only from the angle of the acceleration sensor attached to the forearm (in this embodiment, the wrist). The transition model shown in FIG. 11 defines the transition from the state of each upper arm to the state of another upper arm according to the direction and strength of the forearm swing specified from the measured acceleration. The direction is estimated. In the daily movement, this transition model starts from the downward state 252 based on the fact that the upper arm is downward with high probability when the arm is downward, and the state 251 depends on the strength of the upward swing. A transition to a state (that is, a state where the upper arm faces upward) or a state 253 (that is, a state where the upper arm faces the horizontal direction (horizontal direction)) can be distinguished. The part state discriminating model data 74 may include a rule for discriminating the part state at a certain point of time based on the part state at a point before that as described above. In some cases, it is possible to determine the state of other parts by the same method. In this way, by generating a discrimination model that takes into account the probability of state transition, it becomes possible to classify actions with high accuracy.

  FIG. 12 is an explanatory diagram illustrating a first structure example of the part state data 72 held by the server 5 according to the first embodiment of this invention.

  FIG. 12 shows an example of a data structure for a part that is particularly suitable for discrimination in minutes. One row of the part state data 72 shows data for each user's time unit. The user ID 311 is information for identifying the user who wears the sensor device 1, and is added to the measurement value data when measured by the sensor device 1, and the feature amount and part calculated from the measurement value after analysis It is carried over to values such as state. The date and time 312 is information added based on the date and time when the acceleration is measured by the sensor device 1. The thigh state 313 indicates the discrimination result between the standing state 221 and the sitting state 222 shown in FIG. 7 in binary (for example, one is 0 and the other is 1). Similarly, the posture state 314 indicates whether the body is in a sleeping state or standing state (in other words, the orientation of the whole body) in binary. By similarly creating a discrimination model from data on the relationship between each part and arm in daily life, items of part states other than the above thigh state 313 and posture state 314 can be increased.

  FIG. 13 is an explanatory diagram illustrating a second structure example of the part state data 72 held by the server 5 according to the first embodiment of this invention.

  FIG. 13 shows an example of a data structure for a part particularly suitable for discrimination in seconds. One row of the part state data 72 shows data for each user's time unit. The user ID 321 is information for identifying the user wearing the sensor device 1, as with the user ID 311 in FIG. 12, is added when measured by the sensor device 1, and is inherited after analysis. Date and time 322 is information added based on the date and time when the sensor device 1 measured acceleration. The elbow state 323 indicates the determination result of the elbow bending state 224 and the elbow extension state 223 shown in FIG. The upper arm state 324 indicates the state of the upper arm with discrete values (for example, by assigning values such as 0, 1, and 2 to upward, lateral, and downward, respectively), and a discrimination result based on the discrimination model example shown in FIG. It is. By creating a discrimination model in the same manner from data on the relationship between each part and arm in daily life, items of part states other than the elbow state 323 and the upper arm state 324 can be increased.

  FIG. 14 is an explanatory diagram illustrating a configuration example of the physical feature amount data 71 held by the server 5 according to the first embodiment of this invention.

  One row of the physical feature data 71 shows data for each user for each time unit. FIG. 14 shows physical data in units of seconds that are particularly necessary when performing site state determination in units of seconds. The example of the feature-value is shown.

  The user ID 331 is an ID for identifying the user. The date and time 332 indicates the actual date and time when the data is measured, and is a time unit suitable for determining the part state. Amplitudes 333, 334, and 335 respectively represent average values per unit of acceleration intensity of the X, Y, and Z axes calculated by the method illustrated in FIG. As physical quantities that are characteristic of the acceleration waveform, other statistical values such as the number of zero crossings and the maximum, minimum, average, and variance in the time unit of the differential value can be applied. The average angles 336 and 337 indicate the average values of the pitch angle 211 and the roll angle 212 during the unit time, respectively. In addition, statistical values such as the minimum value, maximum value, and variance of the pitch angle 211 and the roll angle 212 can be applied as physical feature amounts.

  FIG. 15 is an explanatory diagram illustrating an example of action content classification by the server 5 according to the first embodiment of this invention.

  The action content classification program 63 performs change point detection using a plurality of part states recorded in the part state data 72, and for each action section delimited by the change points, a plurality of user's corresponding to the action section are detected. A part state vector having elements of part state values (for example, thigh state, posture state, elbow state, upper arm state, etc.) is generated. FIG. 15 shows an example of the distribution in the part state vector space of the part state vectors of the plurality of action sections of the plurality of users generated as described above, for example.

  Since the statistical features and distribution bias of the part state values differ depending on the content of the user's action, the action content classification program 63 can classify the action of each action section by clustering the part state vector. Yes (S107). In the example of FIG. 15, the action content is classified into action classifications 261 to 264 by clustering. In the action sections classified into the same cluster, since the bias of the part state is similar to each other, it is estimated that the same kind of action (having the same or similar contents) was performed. As a specific method of clustering, any known method can be used, and thus detailed description thereof is omitted.

  By classifying action contents based on such part states, it is possible to comprehensively grasp the action contents according to the characteristics of the part state without defining a name for each action and considering comprehensiveness. Can do. Also, the name of the actual action content can be inferred from the part state.

  FIG. 16 is an explanatory diagram illustrating a data structure example of the action classification data 73 held by the server 5 according to the first embodiment of this invention.

  The behavior of one action section of one user is recorded in one line of the action classification data 73, and the time width of the action section delimited by the change point of the part state and the ID of the classification are recorded. And The user ID 351 is an ID for identifying the user. The start time 352 and the end time 353 indicate the start and end dates and times of one section of the action, respectively. The action 354 records the ID or name of the classified action.

  FIG. 17 is an explanatory diagram illustrating a configuration example of a screen that the server 5 according to the first embodiment of the present invention outputs in order to display an action record.

  A screen 401 shown in FIG. 17 displays the daily action contents of the user based on the part state data 72 and the action classification data 73, and the contents of the movement in each action section are concretely indicated by the part state data. The candidate of the actual action assumed from there is displayed, and the user can input the actual action content and record the result. The action content display 402 color-codes the action section divided by the change in the part state according to the result of classifying the action (for example, by displaying the action section estimated to be the same kind of action in the same color), Can be displayed in series. Such color coding is an example of a display method, and various methods can be employed such as, for example, distinguishing the content of an action by a pattern, a symbol, or a character as long as the result of classifying the action can be read. Thereby, the user can easily recognize the result of classifying the behavior.

  The action content display 403 specifically indicates the contents of one action section selected from the plurality of action sections displayed on the action content display 402. For example, the action content display 403 shows the movement of the user according to the shape of the person specified from the part state and FIG. 403A showing the position in the vector space of the part state that the action classification of the selected action section is characterized FIG. 403B, an action name candidate 403C, and a form 403D in which the user inputs an actual action name can be shown.

  The name of the action entered by the user in the form 403D is recorded in a database or the like, and can be referred to as a name candidate in the same action classification of another user from the user himself / herself. By this screen 401, various actions of the user can be comprehensively visualized, and input of correct names can be promoted and collected.

  For example, in FIG. 403A, a cluster including the part state vector of the selected action section among the parts state vector clusters generated in step S107 is displayed. In FIG. 403B, the state of each part of the user specified from the part state vector belonging to the cluster is displayed. For example, as illustrated in FIG. 403B, the shape of the user's body, such as the sitting position, the upper arm facing sideways (horizontal), and the forearm horizontal, estimated from the state of each part of the identified user may be illustrated. . Action name candidates 403 </ b> C display action name candidates estimated from the shape of the user's body. For example, if the sitting position, the upper arm is horizontal, and the forearm is horizontal, “PC work”, “car driving”, and the like are listed as candidates.

  The user refers to these displays, recalls the action he / she actually performed during the time zone of this action section, and selects the name corresponding to the actual action if it is displayed as a candidate To form 403D. When the name corresponding to the actual action is not displayed as a candidate, the user can input the name in the form 403D.

  The name input in this manner can be displayed as a candidate name 403C for another action section displayed after the next time. Specifically, when the part state vector of the other action section belongs to the same cluster as the part state vector of the action section whose name has been previously input (that is, classified into the same kind of action), the input is performed. The name can be displayed as a name candidate 403C.

  For example, the Web display program 53 creates and outputs screen data for displaying the screen 401, the server 5 transmits the screen data to the PC 2 or the smartphone 3 via the network 4, and the PC 2 or the smartphone 3 The screen 401 may be displayed based on the screen data. In this case, the PC 2 or the smartphone 3 is used as a display device. Furthermore, the input to the form 401D may be performed on the PC 2 or the smartphone 3. In this case, the PC 2 or the smartphone 3 is used as an input device. The same applies to analysis screens 411 and 421 described below.

  FIG. 18 is an explanatory diagram illustrating a first configuration example of a screen that the server 5 according to the first embodiment of the present invention outputs in order to display the result of the behavior analysis.

  The analysis screen 411 shown in FIG. 18 is an example of a screen viewed by a user of the sensor device 1 or a leader who gives health advice to the user of the sensor device 1, and is recorded separately from the acceleration data measured by the sensor device 1. It is characterized by analyzing behavioral contents highly related to daily weight gain and displaying the results. The analysis screen 411 includes a related action ratio display 412 and action content displays 413 and 414. The action content display 413 indicates a time zone in which action contents highly related to weight gain appear at a high frequency (for example, a predetermined value or more) in daily life. By combining the display of this time zone and the ratio display 412, it is possible to encourage the instructor or the user to recall the actual action content of the user at that time.

  In addition, the action content display 414 can indicate the movement of the estimated action content in the form of a person, and can provide a candidate for a name corresponding to the shape and advice such as prompting increase or decrease in the time during which the action is performed. . The action shape classification program 63 and the like can specify the human shape and the corresponding name candidate in the same manner as described with reference to FIG.

  The advice corresponding to the estimated action content can be generated from the correlation between the change in weight and the length of time of action related to weight gain. For example, the candidates for the name of the action B of the user shown in FIG. 18 are “PC work” and “driving a car”, and the daily weight increase of the user recorded so far and the time when the action B is performed are shown. It is desirable that the correlation with the length is higher (or higher than a predetermined threshold) compared with the correlation with the time of other actions, the correlation is a positive correlation, and the user's weight is not increased. Information indicating action B having a high correlation (in the example shown, the shape of a person indicating the movement of action B) is displayed, and a message prompting to reduce the time for performing action B is displayed as advice. The Conversely, when there is a negative correlation between the weight gain and the length of time during which the action B is performed, or when it is desired to increase the weight of the user, the time for performing the action B is increased. A prompt message appears.

  In the above example, the action content having a high correlation with weight gain is a combination of one of the classified actions and the length of time for which the action is performed (for example, the length of time for which the action B is performed). However, you may specify the action content with a high correlation with a weight gain using the other parameter which shows the content of the classified action. The other parameter indicating the content of the classified action may be such a parameter when, for example, a parameter indicating the intensity of a person's movement when performing the classified action can be acquired. If it is possible to count the number of times the classified action has been executed, that number may be used.

  FIG. 19 is an explanatory diagram illustrating a second configuration example of a screen that the server 5 according to the first embodiment of the present invention outputs in order to display the result of the behavior analysis.

  That is, FIG. 19 is another example of a screen that analyzes the relationship between information measured by a device other than the measuring device such as the sensor device 1 and the action content and displays the result, as in FIG. Subjective information that is difficult to measure with a sensor, such as fatigue, is collected from the user in a questionnaire format, and the relationship between the collected result and the action content measured by the sensor device 1 can be shown.

  The analysis screen 421 includes a behavior ratio display 422, a behavior content display 424, a behavior time and subjective information change graph 423, and a subjective information input form 425. The action ratio display 422 and the action content display 424 are the same as the action ratio display 412 and the action content display 414 shown in FIG. The user can input information indicating the current degree of fatigue by operating the input form 425. The input form 425 can collect information indicating the degree of fatigue in a question format in addition to an input in an analog format as shown in the figure.

  The change graph 423 displays the time change of the action having the highest correlation with the input subjective information and the change of the subjective information together. By viewing this, the user can intuitively grasp the relationship between subjective information (for example, the degree of fatigue) and behavior.

  The degree of weight gain and fatigue shown in the examples of FIGS. 18 and 19 are examples of parameters indicating changes in the user's health condition, and the server 5 uses the same method as described above for other parameters (for example, blood pressure). May analyze the correlation with the action content and output the result. In the above example, the analysis of the correlation between the parameter indicating the health condition and the action content and the generation of the advice based on the result are performed by the action content classification program 63 in S107, but these processes are performed after other functions modules after S107. (For example, a dedicated program not shown) may be performed.

  FIG. 20 is an explanatory diagram of a typical example of a conventional method for estimating action content from sensor data.

  This method is to add only the movement of the arm after setting constraint conditions such as a golf swing in advance. In step 81, three-axis acceleration, gyro, and the like are measured. In step 82, a physical quantity such as a speed is estimated in step 82. In step 84, the wrist trajectory is estimated by matching the speed and angle. Since golf swing conditions are set in advance in step 85, positions other than the movable wrist and arm are fixed with a golf foam. In step 86, the trajectory of the position of the movable wrist is added to the form of the golf swing, and the movement of the arm and the golf club accompanying the movement of the wrist is added. The movement of the arm and golf club can be easily estimated from the position of the wrist.

  On the other hand, according to the first embodiment of the present invention shown in FIG. 1 to FIG. 19, for example, physical characteristics of wrist movement that can be measured by a wristband sensor (strength of swing, angle, etc.) The movement of the extremities that are not directly connected (for example, the upper arm or the thigh) is estimated using a statistical pattern acquired in advance in daily life. Furthermore, the action content is classified by a statistical set of estimated limb movement combinations and limb movement combinations in daily life. Thus, using only data of a sensor that can be easily worn in daily life such as a wristband type sensor, for example, not only a specific action such as a golf swing shown in FIG. Various action contents can be classified in detail, and the action contents at each time can be estimated. For this reason, in exercise guidance for lifestyle-related disease prevention, it is possible to visualize the action content to be improved and the result of improving the action.

  The sensor device 1 according to the first embodiment has a function of transmitting data acquired by measuring acceleration, and does not perform calculation for behavior classification, display of the calculation result, or the like. Therefore, the sensor device 1 can be easily reduced in size, weight, and power consumption, thereby reducing the burden on the person wearing the sensor device 1 for a long time. Further, the PC 2 and the smartphone 3 of the first embodiment have functions as an acceleration data relay device, a behavior classification result display device, and a behavior content input device, and do not perform computation for behavior classification. It can be realized by an easy small computer. Since the server 5 according to the first embodiment communicates with the sensor device 1, the PC 2, the smartphone 3, and the like via the network 4, it is not necessary to carry the server 5 itself, so a computer having sufficient calculation capability is used. It is possible to execute a process for classifying a person's action at high speed.

  Next, a second embodiment of the present invention will be described. Except for the differences described below, each part of the behavior classification system according to the second embodiment has the same functions as the parts denoted by the same reference numerals as those of the first embodiment shown in FIGS. Description of is omitted.

  FIG. 21 is a block diagram illustrating the main configuration of the behavior classification system according to the second embodiment of the present invention.

  The functional modules and data included in the smartphone 3 according to the second embodiment are the same as those included in the server 5 according to the first embodiment, and thus detailed description thereof is omitted. However, the smartphone 3 has a display unit 55, and the display unit 55 displays screens as shown in FIGS. 17 to 19 based on the screen data created by the WEB display program 53. When the display unit 14 of the sensor device 1 has a sufficient display capability, the screen data created by the WEB display program 53 is transmitted from the smartphone 3 to the sensor device 1, and the display unit 14 based on the screen data is displayed in FIG. Etc. may be displayed.

  In the present embodiment, the PC 2 or any other type of terminal device may be used instead of the smartphone 3. In that case, the configuration of the PC 2 and the like is the same as that of the smartphone 3 shown in FIG. For example, when the smartphone 3 can be sufficiently downsized, the smartphone 3 and the sensor device 1 may be integrated and attached to a person.

  According to the second embodiment, as in the first embodiment, in addition to the effect of classifying and visualizing human actions, the data measured by the sensor device 1 is analyzed without being sent to the server 5 via the Internet. Therefore, an effect of excellent response speed can be obtained.

  Next, a third embodiment of the present invention will be described. Except for the differences described below, each part of the behavior classification system according to the second embodiment has the same functions as the parts denoted by the same reference numerals as those of the first embodiment shown in FIGS. Description of is omitted.

  FIG. 22 is a block diagram illustrating the main configuration of the behavior classification system according to the third embodiment of the present invention.

  The analysis program 60 of the third embodiment is obtained by adding an action content determination model generation program 66, an action content determination model selection program 67, and an action content estimation program 68 to the analysis program 60 of the first embodiment shown in FIG. . The database 70 of the third embodiment is obtained by adding behavior content determination model data 77 and behavior / part teacher data 78 to the database 70 of the first embodiment shown in FIG.

  The action content classification program 63 according to the first embodiment classifies the action contents into actions that do not specify the meaning based on the characteristics of the part state. That is, in the first embodiment, it is possible to specify a plurality of action sections performing the same kind of action, but it is not possible to specify what the action is. On the other hand, in the third embodiment, the action content estimation program 68 does not simply classify the action contents, but has a name and makes a meaningful action such as deskwork or housework by determining it as a meaningful action. Can be classified into different types.

  The action content estimation program 68 uses action content discrimination model data 77 generated in advance in order to discriminate the action content from the part state. Further, the action content discrimination model selection program 67 can select a discrimination model suitable for the user measured by the sensor device 1 based on the user profile data 75. This is because the association between the part state and the action content may differ depending on the user's age, physique, or lifestyle. However, as in the case of the determination of the part state in the first embodiment, the action content may be determined without considering the user profile.

  The action content determination model generation program 66 can generate the action content determination model data 77 using the action / part teacher data 78 in which the pre-measured action and the part state are simultaneously recorded.

  Specifically, for example, when the acceleration data and motion data of a plurality of measurement target users are acquired by the same method as in the first embodiment, the acceleration data and the motion data are further acquired. Record the name indicating the content of the action that was being performed, and based on them, record the information associating the content of the executed action with the part state when the action was executed as action / part teacher data 78 Also good.

  The behavior content determination model generation program 66 calculates a statistical feature or the like of the part state corresponding to the content of each action based on the behavior / part teacher data 78, and sets rules for determining the behavior content based on the statistical characteristics. The action content discrimination model data 77 including is generated. At this time, as in the case of the part state determination model data 74, the action content determination model generation program 66 creates the action content determination model data 77 in association with the user profile data 75 indicating the attributes of the measurement target user. Also good.

  Alternatively, when acquiring acceleration data and motion data of a plurality of measurement target users in the same manner as in the first embodiment, the content of the action that was executed at the time when the acceleration data and motion data were acquired. And recording teacher data including information associating the content of the executed action with the physical feature amount calculated from the acceleration data when the action is executed based on them. Good. The behavior content determination model generation program 66 calculates a statistical feature or the like of the physical feature amount corresponding to the content of each behavior based on the teacher data, and the behavior including a rule for determining the behavior content based on the statistical feature. The content determination model data 77 may be generated. In this case, it is not necessary for the part state estimation program 62 to estimate the state of each part, and the action content estimation program 68 estimates the action content based on the physical feature amount calculated from the acceleration data and the action content determination model data 77. To do.

  The following description describes an example in which the action / part teacher data 78 is used.

  FIG. 23 is an explanatory diagram illustrating a typical processing procedure of part state estimation and action content estimation executed by the server 5 according to the third embodiment of this invention.

  The processing according to the third embodiment is characterized in that the action content estimation program 68 uses the result of the part state estimation to estimate the action contents as a meaningful action content with a name, rather than classifying the action contents without meaning. It is said. The action content determination model selection program 67 selects the action content determination model data 77 based on the user profile data 75. The action content estimation program 68 is based on the part state data 72 and the selected action content determination model data 77, and the likelihood previously defined in the action content determination model data 77 such as desk work 141, housework 142, reading 143, or sport 144 is presumed. It can be determined that the action content 87 is likely.

  FIG. 24 is an explanatory diagram illustrating a data structure example of the action content data 79 held by the server 5 according to the third embodiment of this invention.

  The action content data 79 according to the third embodiment records the action contents estimated in a specific time interval. The start date and time 362 indicates the date and time when the estimated action starts, and the end date and time 363 indicates the date and time when the estimated action ends. The action 364 indicates the estimated action content, and is determined and selected from names predefined in the action / part teacher data 78 or the action content determination model data 77.

  According to the third embodiment described above, in addition to the same effects as those of the first embodiment, it is possible to obtain the effect of improving the convenience of the user by classifying the action content as a meaningful action content having a name. .

  In addition, this invention is not limited to the Example mentioned above, Various modifications are included. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of one embodiment. Further, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.

  Each of the above-described configurations, functions, processing units, processing means, etc. may be realized in hardware by designing a part or all of them, for example, with an integrated circuit. Further, each of the above-described configurations, functions, and the like may be realized by software by interpreting and executing a program that realizes each function by the processor. Information such as programs, tables, and files for realizing each function is a memory, a hard disk drive, a storage device such as an SSD (Solid State Drive), or a computer-readable non-transitory data such as an IC card, an SD card, or a DVD. It can be stored in a storage medium.

  Further, the drawings show control lines and information lines that are considered necessary for explaining the embodiments, and not necessarily all control lines and information lines included in an actual product to which the present invention is applied. Not necessarily. Actually, it may be considered that almost all the components are connected to each other.

1 Sensor device 2 PC
3 Smartphone 4 Network 5 Server 10 MCU
11 Acceleration sensor 12 Memory 13 Input unit 14 Display unit 15 Communication unit 51 CPU
52 Communication Unit 53 Web Display Program 54 Memory 60 Analysis Program 61 Physical Feature Quantity Calculation Program 62 Part State Estimation Program 63 Action Content Classification Program 64 Part State Discrimination Model Generation Program 65 Part State Discrimination Model Selection Program 66 Action Content Discrimination Model Generation Program 67 Behavior content discrimination model selection program 68 Behavior content estimation program 70 Database 71 Physical feature data 72 Site state data 73 Behavior classification data 74 Site state discrimination model data 75 User profile data 76 Site / physical quantity teacher data 77 Behavior content discrimination model data 78 Behavior -Body teacher data 79 Action content data 80 Acceleration data

Claims (15)

Holding model information including rules for determining the behavior of the person based on physical features calculated from the acceleration of a predetermined part of the person;
An acceleration sensor mounted on the predetermined part of the person;
An action classification system, comprising: an analysis unit that classifies the person's action based on a physical feature amount calculated from acceleration measured by the acceleration sensor and the model information. The behavior classification system according to claim 1,
The model information includes part state determination model information including a rule for determining a state of one or more parts other than the predetermined part based on a physical feature amount calculated from acceleration of a predetermined part of the person. ,
The analysis unit
A part state estimation unit that estimates the state of the one or more parts of the person based on physical feature amounts calculated from acceleration measured by the acceleration sensor and the part state determination model information;
A behavior content classification unit that classifies the behavior of the person based on the estimated state of one or more parts of the person. The behavior classification system according to claim 2,
The physical corresponding to the state of each part based on the value indicating the state of the one or more parts of the person measured at a plurality of times and the acceleration of the predetermined part of the person measured at the same time A behavior classification system further comprising a part state discriminating model generating unit that generates the part state discriminating model information by calculating a statistical feature of the feature amount. The behavior classification system according to claim 3,
The part state determination model generation unit generates the part state determination model information for a plurality of persons,
The behavior classification system further holds attribute information of the plurality of persons,
The part state estimation unit includes the acceleration measured by the acceleration sensor and the part state determination model information generated from one or more persons having an attribute similar to the attribute of the person wearing the acceleration sensor. An action classification system for estimating a state of the one or more parts of the person based on the person. The behavior classification system according to claim 3,
The part state determination model information includes a rule for determining a state of a part at a certain point of time based on a state of a part at a point before that. The behavior classification system according to claim 2,
The action content classification unit
The person's action is divided into a plurality of sections, and the person's actions in each section are classified at a time point where a change amount of the value indicating the state of the one or more parts is larger than a predetermined threshold. A behavior classification system. The behavior classification system according to claim 6,
The behavior content classification unit classifies the behavior of the person in each section by clustering values indicating the states of the one or more parts of the plurality of sections. The behavior classification system according to claim 6,
The model information further includes action content determination model information including a rule for determining the content of the action based on the state of the one or more parts.
The behavior classification unit classifies the behavior of the person in each section based on one or more states of the estimated person and the behavior content determination model information. system. The behavior classification system according to claim 6,
The behavior classification system characterized by further having a display part which outputs the screen data for displaying the result of having classified the action of the person for every section. The behavior classification system according to claim 9,
The display unit further outputs screen data for displaying a state of the one or more parts corresponding to the behavior classified for each section,
When information for identifying an action that has been executed in any of the sections is input, the input information is used as a candidate for information for identifying an action in another section classified as the same type of action as the section. An action classification system characterized by output. The behavior classification system according to claim 9,
Further holding information indicating changes in values indicating the health status of the person,
The said display part outputs the screen data for displaying the action with high correlation with the value which shows the said health state among the said classified action, The action classification system characterized by the above-mentioned. The behavior classification system according to claim 2,
The predetermined part to which the acceleration sensor is attached is a forearm of a person,
The acceleration sensor measures triaxial acceleration,
The physical feature amount is an angle indicating the amplitude of the triaxial acceleration, the frequency of the triaxial acceleration, and the amount of rotation of the forearm about the longitudinal direction of the forearm, calculated from the triaxial acceleration of the predetermined part. And at least one of the longitudinal directions of the forearm,
The state of the one or more parts includes at least one of an upper arm orientation, an elbow angle, a thigh orientation, and an entire body orientation. The behavior classification system according to claim 2,
A sensor device, a computer, and a network that relays data communicated between the sensor device and the computer,
The sensor device includes the acceleration sensor and a communication unit that transmits acceleration data measured by the acceleration sensor to the computer via the network.
The computer includes a communication unit that receives the acceleration data, a processor, and a storage device.
The part state determination model information is stored in the storage device,
The part state estimation unit and the action content classification unit are the processor that executes a program stored in the storage device. A behavior classification device having a processor and a storage device connected to the processor,
The storage device
Model information including a rule for determining the behavior of the person based on a physical feature amount calculated from the acceleration of a predetermined part of the person;
Holding acceleration information measured by an acceleration sensor attached to the predetermined part of the person,
The processor classifies the person's action based on the physical feature amount calculated from the acceleration information and the model information. A behavior classification method by a computer having a processor and a storage device connected to the processor,
The storage device
Model information including a rule for determining the behavior of the person based on a physical feature amount calculated from the acceleration of a predetermined part of the person;
Holding acceleration information measured by an acceleration sensor attached to the predetermined part of the person,
The behavior classification method includes a procedure in which the processor classifies the behavior of the person based on a physical feature amount calculated from the acceleration information and the model information.