JP2009039466A - Action identification device, action identification system, and action identification method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an action identification device, an action identification system, and an action identification method, capable of accurately identifying a complicated action of a test subject by identifying actions which can be made in an identified posture. <P>SOLUTION: An action identification system 10 comprises a server 12, which receives acceleration data transmitted from sensor devices 16 and stores the data in a sensor DB 20. A plurality of sensor devices 16 are attached to subject's right and left wrists, ankles and thighs. Each sensor 16 is provided with an acceleration sensor, wherein acceleration data from the acceleration sensor is transmitted to the server 12. The server 12 calculates a posture feature value for detecting the posture of the subject by use of the acceleration data. When identifying the posture of the test subject according to the calculated posture feature value, the server 12 calculates an action feature value for identifying the actions which can be made in the identified posture by use of the acceleration data. Then, on the basis of the action feature value, one of the actions which can be made in the identified posture is identified. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a behavior identification device, a behavior identification system, and a behavior identification method, and more particularly, to a behavior identification device, a behavior identification system, and a behavior identification method, for example, for identifying a subject's behavior.

  An example of this type of conventional behavior identification device is disclosed in Patent Document 1. In the cellular phone device disclosed in Patent Document 1, when the owner designates a desired action (for example, “walking 1”) on the registration screen and inputs a start using the key input unit while performing the action, the mobile phone apparatus is designated in advance. The acceleration sensor unit detects the motion acceleration for the time, and the reference behavior pattern of the behavior is stored in the pattern storage unit. When the operation mode is activated after such registration, the acceleration sensor unit detects motion acceleration due to the owner's behavior, and the behavior detection unit compares it with the reference behavior pattern to identify the current behavior pattern of the subject. To do.

Another example of this type of behavior identification device is disclosed in Patent Document 2. In the motion information measurement system of Patent Document 2, motion information is collected from a motion information measurement device mounted on the user's wrist, upper arm, knee, waist, both ankles, and the like. This motion information clerk measuring device is composed of an acceleration sensor and / or a gyro sensor, and measures acceleration and / or angular velocity due to the user's motion at the wearing site, or both. The recognition of the action is performed based on the action recognition rule that the action information comparison and recognition unit has. For example, when the information from the motion information measuring device attached to the waist has a small amount of exercise and the information from the motion information measuring device attached to the wrist has a large amount of exercise, it recognizes that the motion is in a stationary state. Conversely, when the amount of exercise from the motion information measuring device attached to the waist is large, it is recognized that the exercise is a whole body exercise.
JP2003-46630 [H04M 1/247, H04Q 7/38] JP 2004-184351 [G01B 21/00, A61B 5/00, A61B 5/0488, A61B 5/11, G06F 3/00]

  However, in the technique of Patent Document 1, “walking”, “sitting”, “standing”, “boarding a vehicle” are performed according to the time transition of acceleration in the X, Y, and Z directions of the acceleration sensor unit provided in the mobile phone device. However, since only one acceleration sensor is used, it is difficult to recognize complex actions. For example, when the owner of a mobile phone device is sitting and operating a computer, even if the user can identify sitting, it cannot be identified until the computer is operated.

  In the technique of Patent Document 2, the momentum of each part is detected from data such as acceleration from the motion information measuring device attached to a plurality of parts, and the movement is a whole body movement or a movement in a stationary state. It is difficult to identify actions such as “walking”, “sitting”, and “standing”. Of course, it is difficult to identify more complex actions.

  Therefore, a main object of the present invention is to provide a novel behavior identification device, behavior identification system, and behavior identification method.

  Another object of the present invention is to provide a behavior identification device, a behavior identification system, and a behavior identification method that can accurately identify a complex behavior.

  The present invention employs the following configuration in order to solve the above problems. The reference numerals in parentheses, supplementary explanations, and the like indicate correspondence relationships with embodiments described later to help understanding of the present invention, and do not limit the present invention in any way.

  The first invention is a detection means for detecting motion detection data from a plurality of motion detection sensors attached to a subject, and a posture feature for identifying the posture of the subject based on the motion detection data detected by the detection means. Posture feature amount calculating means for calculating the amount, posture identifying means for identifying the posture of the subject based on the posture feature amount calculated by the posture feature amount calculating means, posture based on the motion detection data detected by the detecting means Action feature amount calculating means for calculating an action feature amount for identifying an action that can be taken in the posture identified by the identification means, and posture identification means based on the action feature amount calculated by the action feature amount calculating means It is an action identification apparatus provided with the action identification means which identifies a test subject's action from the some action which can be taken in the attitude | position identified by.

  In the first invention, the subject is equipped with a plurality of motion detection sensors (32), and the detection means (12) detects motion detection data from the plurality of motion detection sensors. The posture feature amount calculating means (12, S53) calculates a posture feature amount for identifying the posture of the subject (for example, standing, sitting, walking). In the embodiment, the posture feature amount is a vector calculated based on the motion detection data, and its elements are the average, standard deviation, energy, frequency domain entropy, and correlation of motion detection within a certain time window. The action identifying means (12, S57) identifies actions that can be taken in the posture identified by the posture identifying means based on the acceleration data detected by the detecting means. The behavior feature quantity calculating means (12, S59) calculates a behavior feature quantity for identifying an action that can be taken in the posture identified by the posture identification means, based on the motion detection data detected by the detection means. . The behavior identifying unit (12, S61) identifies the behavior of the subject from a plurality of behaviors that can be taken in the posture identified by the posture identifying unit, based on the behavior feature amount calculated by the behavior feature amount calculating unit.

  According to 1st invention, after detecting a test subject's action using a some motion detection sensor and identifying a test subject's attitude | position, one action is identified from the some action which can be taken in the identified attitude | position. Therefore, it is possible to accurately identify the complex behavior of the subject.

  The second invention is dependent on the first invention, wherein the plurality of motion detection sensors are attached to a plurality of different parts of the subject, and the posture feature amount calculation means is a posture identification means among all the motion detection sensors. Posture feature values are calculated based on motion detection data from all or some of the motion detection sensors for identifying actions that can be taken in the identified posture.

  In the second invention, the plurality of motion detection sensors are attached to a plurality of different parts of the subject. For example, the motion detection sensor is attached to a specific part such as the left and right wrists, left and right thighs, and left and right ankles of the subject. The posture feature amount calculation means is based on the motion detection data from all or a part of the motion detection sensors for identifying actions that can be taken in the posture identified by the posture identification means among all the motion detection sensors. To calculate the posture feature amount. For example, when identifying a certain posture (for example, standing, walking), motion detection data from motion detection sensors attached to the left and right wrists and the left and right ankles are used, but other postures (for example, When identifying the standing position and the sitting position), motion detection data from all motion detection sensors are used.

  According to the second aspect of the invention, since the posture feature amount corresponding to the posture to be identified is calculated, the posture can be accurately identified. That is, identification can be made based only on useful motion detection data.

  A third invention is dependent on the second invention, and the behavior feature amount calculation means includes motion detection data from all or a part of the motion detection sensors according to the behavior to be identified among all the motion detection sensors. Based on the behavior feature amount is calculated.

  In the third invention, the behavior feature amount calculating means also operates from all or a part of the motion detection sensors according to the behavior to be identified among all the motion detection sensors, similarly to the posture feature amount calculation means. A behavior feature amount is calculated based on the detection data. Therefore, for example, when identifying actions that can be taken in a standing posture (for example, standing, climbing an elevator, descending an elevator), motion detection sensors mounted on left and right wrists or left and right thighs The motion detection data from is used. Also, for example, when identifying actions that can be taken in the sitting position (for example, sitting, operating a computer while sitting, eating), the motion detection sensors attached to the left and right wrists and the left and right ankles Use motion detection data.

  According to the third aspect, the behavior feature amount corresponding to the behavior to be identified is calculated, so that the behavior can be accurately identified.

  A fourth invention is dependent on the first invention, wherein the plurality of motion detection sensors are attached to a plurality of different parts of the subject, and the behavior feature amount calculating means is configured to identify the behavior among all the motion detection sensors. Based on the motion detection data from all or part of the motion detection sensors, the behavior feature amount is calculated.

  The fourth invention is the same as the third invention except that it depends on the first invention.

  In the fourth invention as well, similar to the third invention, it is possible to accurately identify the behavior.

  5th invention is an action identification system provided with the portable terminal possessed or mounted | worn by the test subject, and the action identification apparatus, Comprising: A portable terminal detects the motion detection data from the several motion detection sensor with which the test subject was mounted | worn. Detecting means that transmits the motion detection data detected by the detection means to the action identification device. The action identification device receives the motion detection data transmitted by the transmission means. The reception means receives the motion detection data. Posture feature amount calculating means for calculating a posture feature amount for the posture of the subject based on the motion detection data, posture identification for identifying the posture of the subject based on the posture feature amount calculated by the posture feature amount calculating means Based on the motion detection data received by the means and the receiving means, actions that can be taken in the posture identified by the posture identifying means Based on the behavior feature quantity calculating means for calculating the behavior feature quantity to be separated, and a plurality of actions that can be taken in the posture identified by the posture identification means based on the behavior feature quantity computed by the behavior feature quantity computation means It is an action identification system provided with the action identification means which identifies a test subject's action.

  According to a fifth aspect of the present invention, there is provided a behavior identification system (10) comprising a portable terminal (18) possessed or attached to a subject and a behavior identification device (12) as described in the first aspect. The portable terminal includes a detection unit (50) and a transmission unit (50, 52). The detection means detects motion detection data from a plurality of motion detection sensors attached to the subject. The transmission means transmits the detected motion detection data to the action identification device. In the action identification device, the operation detection data from the portable terminal is received by the receiving means (12 communication means). Based on the received motion detection data, as described above, the posture is identified and the action is further identified.

According to the fifth invention, since the motion detection data of the motion detection sensor is transmitted to the behavior identification device from the portable terminal possessed by or attached to the subject wearing the motion detection sensor, the behavior of the subject is identified online. Can do.
A sixth invention is a behavior identification method of a computer for identifying the posture and behavior of a subject, (a) detecting motion detection data from a plurality of motion detection sensors attached to the subject, and (b) step ( Based on the motion detection data detected by a), calculate the posture feature amount for the subject's posture, (c) identify the subject's posture based on the posture feature amount calculated by step (b), (d) Based on the motion detection data detected in step (a), calculate an action feature amount for identifying possible actions in the posture identified in step (c), and (e) step This is a behavior identification method for identifying a subject's behavior from a plurality of behaviors that can be taken in the posture identified in step (c) based on the behavior feature amount calculated in (d).

  In the sixth invention, similarly to the first invention, it is possible to accurately identify the complex behavior of the subject.

  According to the present invention, since the behavior of the subject is detected using the plurality of motion detection sensors and the posture of the subject is identified, one behavior is identified from the plurality of behaviors that can be taken in the identified posture. Can accurately identify complex behaviors.

  The above object, other objects, features and advantages of the present invention will become more apparent from the following detailed description of embodiments with reference to the drawings.

  Referring to FIG. 1, a behavior identification system 10 according to an embodiment of the present invention includes a server 12 that also functions as a behavior identification device, and is applied to a certain environment and is a person (subject) existing in the environment. Record your actions and identify them. A plurality of databases (DB) are connected to the server 12. Specifically, the server 12 is connected to a sensor DB 20, a posture identification sensor set DB 22, and a behavior identification sensor set DB 24.

  The sensor DB 20 is a database for recording the behavior of the subject. Specifically, the sensor DB 20 mainly stores acceleration data from the sensor device 16 to be described later in time series. The posture identification sensor set DB 22 stores sensor sets (sensor combinations) for identifying postures. In this embodiment, as described later, postures are classified into three types of “standing position”, “sitting position”, and “walking”, and a plurality of acceleration data necessary for identifying these postures are output. The identification information of the sensor device 16 is stored in the posture identification sensor set DB 22 as a sensor set. The action identification sensor set DB 24 stores a sensor set (a combination of sensors) for identifying actions. In this example, actions can be taken in a “standing” position, “standing”, “elevating the elevator” and “elevating the elevator”, and “sitting” in a “sitting” position, There are ten types of “computer work”, “reading” and “meal” and “walking”, “up stairs” and “down stairs” which can be taken in the posture of “walking”. Identification information of a plurality of sensor devices 16 that output acceleration data necessary for identifying an action in each posture class is stored in the action identification sensor set DB 24 as a sensor set.

  Note that the posture identification sensor set DB 22 and the action identification sensor set DB 24 are constructed by learning using acceleration data collected in advance when the subject behaves, as will be described later (FIGS. 19 and 20). And FIG. 21).

  The server 12 is connected to a repeater 18 worn or possessed by the subject via a wired or wireless communication line (network) 14. In this embodiment, since the subject may move, the network 14 is provided with a plurality of access points (not shown), and the repeater 18 is connected to the network 14 via any of the access points. The repeater 18 is communicably connected to a plurality of sensor devices 16 attached to the subject.

  FIG. 2A is a block diagram showing a specific configuration of the sensor device 16, and the sensor device 16 includes a CPU 30. The CPU 30 is connected to an acceleration sensor 32, a RAM 34, and a Bluetooth module 36. An antenna 38 is connected to the Bluetooth module 36. For example, the acceleration sensor 32 is a multi-axis (3-axis) acceleration sensor, has a sampling frequency of 200 Hz, and can measure acceleration up to ± 3 G (G is gravity). In the sensor device 16, acceleration data (acceleration data) detected by the acceleration sensor 32 is supplied to the CPU 30. The CPU 30 stores (temporarily stores) the acceleration data from the acceleration sensor 32 in the RAM 34, and the acceleration data for the predetermined time is transmitted via the Bluetooth module 36, the antenna 38, and the network 14 every predetermined time (for example, 10 seconds). To the server 12.

  FIG. 2B is a block diagram showing a specific configuration of the repeater 18, and the repeater 18 includes a CPU 50. An interface 52, a RAM 54, and a Bluetooth module 56 are connected to the CPU 50. In addition, an antenna 58 is connected to the Bluetooth module 56. In the repeater 18, acceleration data transmitted from the sensor device 16 is given to the CPU 50 via the antenna 58 and the Bluetooth module 56. The CPU 50 stores (temporarily stores) acceleration data from the sensor device 16 in the RAM 54, and at a timing when connected to the network 14, the CPU 50 converts the acceleration data stored in the RAM 54 into an interface 52 such as a LAN (wireless LAN) adapter. The data is transmitted to the server 12 via the network 14. The server 12 receives acceleration data from the network 14 via an interface (not shown) such as a LAN adapter, and stores (registers) the received acceleration data in the sensor DB 20.

  Although detailed explanation is omitted, it is determined in advance which part of the subject each sensor device 16 is to be worn, and the acceleration data transmitted from each sensor device 16 includes the sensor device 16. Identification information (ID) for identification is attached as a label. For example, the MAC address of the Bluetooth module 36 can be used as the ID. However, a ROM that can be accessed by the CPU 30 may be provided inside the sensor device 16, and an ID may be stored in the ROM. On the other hand, in the server 12, in a memory (not shown) such as a ROM, a RAM, and a hard disk provided therein, a part to which the sensor device 16 is to be attached is described corresponding to the ID of the sensor device 16. Table is stored. Therefore, the server 12 can easily identify the acceleration data from the sensor device 16 attached to which part by referring to the label (ID) and the table attached to the acceleration data.

  As shown in FIG. 3, the plurality of sensor devices 16 (16a, 16b, 16c, 16d, 16e, 16f) and the repeater 18 are attached to the subject. Specifically, the sensor device 16a is attached to the right wrist, the sensor device 16b is attached to the left wrist, the sensor device 16c is attached to the right thigh, the sensor device 16d is attached to the left thigh, and the sensor device 16e is attached to the right foot. The sensor device 16f is attached to the left ankle. The repeater 18 is attached to the waist of the subject.

  Although illustration is omitted, each sensor device 16 and the relay 18 are attached to the subject by a fixing tool such as a rubber band. However, the repeater 18 may be carried by the subject or may be put in a bag, a waist pouch or a rucksack used by the subject.

  In such a behavior identification system 10, as described above, the behavior of the subject is recorded and identified. Conventionally, many studies have been made on identification of daily activities using an acceleration sensor. In behavior identification using an acceleration sensor, it is important to evaluate conditions such as the mounting position of the acceleration sensor and the sampling frequency of data at the time of data collection.

  As a study that evaluated the relationship between the sensor and the recognition rate for multiple positions (parts) of the subject's body, “L.Bao and SS Insille ,: Activity recognition from user-annotated acceleration data , in Proc. of Pervasive 2004, vol. LNCS 3001, A. Ferscha and F. Mattern (Eds.), Springer-Verlag, Berlin Heidelberg, pp. 1-17, 2004. (Reference 1) and “N. Kern, B. Schiele, and A. Schmidt, “Multi-sensor activity context detection for wearable computing,” in European Symposium on Ambient Intelligence (EUSAI), Eindhoven, The Netherlands, Nov. 2003. (reference 2) It is done.

  In the former, the mounting position is evaluated using acceleration sensors that are mounted asymmetrically on five positions of the body, and about 80% of identification results are obtained with respect to identification of 20 kinds of daily activities in only two positions of the right wrist and left thigh. Have gained. However, the arrangement of the acceleration sensors is asymmetrical. For example, in the case of a wrist, it is necessary to evaluate the difference between left and right. In the latter, the optimum mounting position of the acceleration sensor is evaluated by changing the mounting position of the acceleration sensor used for identification for each action to be identified and dividing the upper body and the lower body. However, the evaluation regarding the identification performance regarding all the combinations of the wearing positions in which the acceleration sensors are arranged symmetrically and the difference between the dominant hand and the dominant foot of the subject is considered has not been made.

  Moreover, the value of the sampling frequency when collecting acceleration data varies for each research. In many studies, data is collected at a sampling frequency of 50 Hz or higher (references 1 and 3). However, Reference 3 is “N. Ravi, N. Dandekar, P. Mysore, and ML Littman,: Activity recognition from accelerometer data, American association for artificial intelligence (www.aaai.org), July 2005.” .

  Also, “CV Bouten, KT Koekkoek, M. Verduin, R. Kodde, and JD Janssen,“ A triaxial Accelerometer and portable data processing unit for the assessment of daily physical activity, ”IEEE Trans. On Bio-Medical Eng., Vol. 44, no. 3, pp 136-147, 1997. (Ref. 4), it is necessary to have a sampling frequency of 20Hz or more to measure human behavior such as “walking” and “running”. There are evaluation results. On the other hand, “Akihiro Kawahara, Hiroyuki Morikawa, Yuki Aoyama,“ Context Estimation Using Small Wireless Sensors and its Applications, ”NPO Wearable Computer Research and Development Organization, Wearable Computing Research Group, Vol. 1, "No. 3, pp. 2-6, Dec. 2005." (Reference 5), a single biaxial acceleration sensor that samples at 10Hz or less is used, and a discrimination rate of nearly 100% is obtained. . However, the actions targeted by Reference 5 are four types of actions such as “standing”, “sitting”, “walking”, and “running” that are relatively easy to identify, and are necessary for more complex action identification in daily life. There has been no assessment of sampling frequency requirements.

  Therefore, first, as conditions for data collection in everyday action identification using the acceleration sensor 32, (1) discrimination performance and optimum sensor arrangement evaluation using a maximum of six acceleration sensors 32 arranged symmetrically in each part of the body, (2) An experiment was conducted on the discrimination performance evaluation based on the sampling frequency for each action.

  For example, in extracting feature values from acceleration, time-series data (acceleration data) is divided into sliding windows of a fixed time, and feature values are obtained for each window. In addition, since the method of calculating | requiring this feature-value is disclosed by said reference 1 and reference 3, please refer to these.

  Further, in this embodiment, the parameters of the sampling frequency of acceleration data and the number of samples in the window are set in the same manner as in Reference Document 2 above, and 50 Hz data is divided into 256 sample windows that overlap by 128 samples for each action. To do. Therefore, one window length is 5.120 seconds. FIG. 4 shows a method of dividing into sliding windows. Each window is indicated by a dotted frame. However, in FIG. 4, for easy understanding, adjacent windows are displayed with a slight shift in the vertical and horizontal directions.

In this embodiment, five types of feature quantities of average, standard deviation, energy, frequency domain entropy, and correlation are calculated for each window based on acceleration. These are the feature quantities used in the studies of the above Reference 1 and Reference 3. The average and standard deviation are the average and standard deviation of the acceleration of each axis, respectively. The energy was obtained as the sum of the square of the amplitude by performing FFT (Fast Fourier Transform) on the acceleration of each axis. However, the DC component and the folded part are excluded. If the FFT amplitude component of the acceleration of one axis is F ₁ , F ₂ , F ₃ ,. Here, w is the number of samples in the window (256 in this embodiment).

[Equation 1]

Further, frequency domain entropy (hereinafter sometimes simply referred to as “entropy”) is obtained based on the probability distribution p. As shown in Equation 2, the probability distribution p is obtained by normalizing the FFT component F _i excluding the DC component and the folded portion with the sum of all components.

[Equation 2]

  Thus, entropy is obtained according to Equation 3.

[Equation 3]

  The correlation coefficient is a correlation coefficient related to acceleration between two axes. At this time, the target two axes include not only combinations within one acceleration sensor 32 but also combinations straddling arbitrary two acceleration sensors 32. When (x, y) is a vector having an acceleration of a certain axis, cov (x, y) is a covariance of x and y, and σx and σy are standard deviations of x and y, respectively, between xy The correlation coefficient corr (x, y) is defined by Equation 4.

[Equation 4]

  In this embodiment, in order to identify the daily behavior of the subject, supervised learning using the behavior to be identified as a label is applied to the feature amount extracted by the above-described method. As the class classifier, three types of well-known SVM (Support Vector Machine), NN (Nearest Neighbor), and C4.5 are adopted.

  In addition, as described above, daily activities to be identified include “standing”, “sitting”, “walking”, “up stairs”, “down stairs”, “computer work”, “reading”, “food” , “Elevator up” and “Elevator down”. These ten types of actions can be classified into any of the above-mentioned three postures of “standing position”, “sitting position”, and “walking”.

  As described above, in this embodiment, the sensor device 16 is mounted at six positions: the left wrist, the right wrist, the left thigh, the right thigh, the left ankle, and the right ankle of the subject. The reason for choosing the wrist and thigh as the wearing position is that past studies have confirmed its effectiveness. For example, the above reference 1 and “N. Kern, B. Schiele, and A. Schmidt,: Multi-sensor activity context detection for wearable computing, in European Symposium on Ambient Intelligence (EUSAI), Eindhoven, the Netherlands, Nov. 2003. "(Reference 6) introduces a study that evaluated the mounting position of an acceleration sensor. Furthermore, in the study introduced in References 1 and 3, it can be seen that the acceleration pattern at the ankle is more effective than the acceleration pattern at the thigh for periodic actions and actions involving up and down movements. Acceleration sensors are also attached to the left and right ankles. In addition, the acceleration sensor 32 (sensor device 16) is arranged (mounted) symmetrically in order to evaluate the mounting position in the whole body, including the comparison between the dominant hand / dominant foot and the non-dominant hand / non-dominant foot. Because.

  For example, data collection is performed separately for a set of acceleration data (training data) for use in learning of the classifier and a set of acceleration data (test data) for evaluating the identification results. The training data and the test data are acceleration data obtained when the subject behaves according to the scenario including the above-described ten kinds of behaviors set in advance by the experimenter. However, the training data is composed of acceleration within a total time of about 100 minutes per subject, and the test data is composed of acceleration within a total time of about 30 minutes per subject. This training data and test data were collected for three subjects. Three subjects are men in their twenties and thirties, and each has a right hand and a right foot.

  The action labeling was performed by recording the start time and end time of each action in seconds when the experimenter accompanies the subject and the subject executes the action according to the scenario. After the data collection was completed, only the labeled part was cut out from the collected acceleration data, and the above-described feature amount for each window was calculated. However, the five seconds before and after the start time and end time of the action are considered to be during the action change, and the data during that time is excluded from the calculation of the feature amount.

  Data was collected from all acceleration sensors 32 according to the technique described above. Next, using the training data of all subjects, learning is performed for each class classifier of SVM, NN, and C4.5, and then all test data of all subjects is given to determine the discrimination performance for each of the ten types of actions. It was. At this time, as combinations of sensors used for learning and evaluation, all combinations (63 ways) in which the number of sensors is 1 to 6 were used as comparison conditions.

  FIG. 5 shows a combination of mounting positions of the acceleration sensor 32 that obtains the highest identification result in each classifier, and the identification result. Here, the identification rate is a ratio (True Positive Rate) of the number of samples correctly recognized in all samples. When SVM is used as the class classifier, the highest identification result (89.8) is obtained when the acceleration sensors 32 of the four sensor devices 16a, 16b, 16e and 16f of “both wrists and both ankles” are used. %). Further, when NN is used as the class classifier, the highest discrimination result (87.6) is obtained when the acceleration sensors 32 of the three sensor devices 16a, 16e and 16f of “right wrist and both ankles” are used. %). Further, when C4.5 is used as the class classifier, the highest discrimination result (83) is obtained when the acceleration sensors 32 of the three sensor devices 16a, 16e and 16f of “right wrist and both ankles” are used. 0.0%).

  FIG. 6 shows a line graph obtained by obtaining a combination of acceleration sensors 32 that obtains the highest identification result for each classifier and plotting the identification results. As shown in FIG. 6, the superiority of using SVM as a classifier can be confirmed. Therefore, hereinafter, only the case where SVM is used as a classifier will be described.

  FIG. 7 is a bar graph showing the influence of the dominant hand (right hand in this embodiment) and the dominant hand (right foot in this embodiment) on the identification result when SVM is used as the class classifier. As shown in FIG. 7, when the right wrist (16a) is removed from the four positions (16a, 16b, 16e, 16f) that are the optimal mounting positions of the acceleration sensor 32, the identification performance is greatly deteriorated, and the right wrist (16a) is worn. It can be seen that the contribution of the acceleration sensor 32 of the sensor device 16a is high. FIG. 7 also shows that the identification result is equally right and left when one of the ankles (16e or 16f) is removed from the four positions (16a, 16b, 16e, and 16f) that are the optimal mounting positions of the acceleration sensor 32. It was not possible to confirm the difference in the contribution of the ankle.

  FIG. 8 shows the identification results for each of the ten types of actions with respect to the SVM results shown in FIG. As shown in FIG. 8, when the number of sensor devices 16 (acceleration sensor 32) used for identification is reduced from four to four (16a, 16b, 16e, 16f) when worn on both wrists and both ankles. In addition, the improvement of the identification results was observed in all actions except for “up of the elevator”. A confusion matrix of identification results for each action is shown in FIGS. As can be seen with reference to FIGS. 9 and 10, the number of misidentified samples is reduced by reducing the number of sensors 16 (acceleration sensor 32) used for identification to four, that is, both wrists and both ankles. ing.

  This improvement in discrimination performance will be confirmed in the canonical feature space. FIG. 11 shows the first and first in the canonical feature space when the acceleration sensors 32 of all six sensor devices 16a, 16b, 16c, 16d, 16e and 16f attached to both wrists, both thighs and both ankles are used. It is a scatter diagram on two variables. FIG. 12 shows the first and second variables in the canonical feature space when the acceleration sensors 32 of the four sensor devices 16a, 16b, 16e and 16f attached to both wrists and ankles are used. It is a scatter diagram.

  Here, the cumulative contributions up to the second variable in FIGS. 11 and 12 are 84.5% and 78.7%, respectively. In FIG. 11 and FIG. 12, it can be confirmed that 10 types of behaviors are distributed in three types: behaviors that take a standing posture, behaviors that take a sitting posture, and behaviors that involve walking. Also, in FIG. 11, three types of actions that take a standing posture and four types of behaviors that take a sitting posture are clearly distributed, whereas in FIG. 12, three types of actions that take a standing posture. And the four types of actions that take a sitting posture are not clear.

  Further, in the case shown in FIG. 11, four types of actions (“sitting”, “computer work”, “reading”, “meal”) taking a sitting posture are concentrated in the lower right area. This suggests that it is difficult to correctly identify the four types of actions that take a sitting posture. Specifically, regarding the four types of actions that take a sitting posture, when the intra-class / inter-class variance ratio is obtained using the first and second variables in FIGS. 11 and 12, the case shown in FIG. Is 4.88, and in the case shown in FIG. 12 (when the acceleration sensors 32 of the four sensor devices 16a, 16b, 16e and 16f are used) is 20.04. This means that the four types of action classes that take a sitting posture are separated in the canonical feature space when the acceleration sensors 32 of the thigh sensor devices 16c and 16d are removed. For example, FIG. 13A, FIG. 13B, and FIG. 14 show the correctness corresponding to each posture when the acceleration data of the acceleration sensors 32 of the four sensor devices 16a, 16b, 16e, and 16f is used. The distribution for each action in the quasi-feature space is shown. As can be seen from FIG. 13A, FIG. 13B, and FIG. 14, in any case, actions in the same posture are separated and can be easily identified.

  Further, FIG. 15 and FIG. 16 show posture identification results and action identification results that can be taken in each posture class (category) when another sample is used. FIG. 15 shows the results when the posture and all actions are identified based on the acceleration data from the acceleration sensors 32 of all the sensor devices 16. On the other hand, FIG. 16 shows the results when the posture and all actions are identified based on the acceleration data from the acceleration sensors 32 of the four sensor devices 16 excluding the thigh sensor device 16. However, in FIG. 15 and FIG. 16, the behavior is classified and shown in the categories of standing, sitting and walking.

  Referring to FIG. 15 and FIG. 16, when the posture identification results are seen, the thigh sensor devices 16 c and 16 d are more clearly identified when the identification is based on the acceleration data from the acceleration sensors 32 of all the sensor devices 16. It can be seen that the identification rate is higher than that in the case of identification based on the acceleration data from the acceleration sensors 32 of the four sensor devices 16 except for the above. Further, when looking at the identification results of the actions included in the standing and walking categories, the case of identification based on the acceleration data from the acceleration sensors 32 of all the sensor devices 16 and the thigh sensor devices 16c and 16d are excluded. It can be seen that the discrimination rate is almost the same in the case of discrimination based on the acceleration data from the acceleration sensors 32 of the four sensor devices 16. However, when the identification result of the sitting position is seen, the acceleration sensors of the four sensor devices 16 excluding the thigh sensor devices 16c and 16d are more than in the case of identifying based on the acceleration data from the acceleration sensors 32 of all the sensor devices 16. It can be seen that the identification rate is considerably higher when the identification is based on the acceleration data from 32.

  As described above, the sensor devices 16c and 16d attached to both thighs are effective for identifying a difference in posture, but are effective for identifying a behavior between similar postures, in particular, a behavior between sitting postures. I understand that it is not.

  Here, in the above example, the sensor device 16 (acceleration sensor 32) is attached to both wrists, both thighs, and both ankles to identify 10 types of actions. The attachment position varies depending on the application (use) environment or the like of the behavior identification system 10. Further, the posture and behavior to be identified vary depending on the usage environment of the behavior identification system 10 and the like. Therefore, the server 12 extracts acceleration data (sensor device 16) used for posture identification from all acquired acceleration data (samples), and further identifies actions that can be taken in the identified posture. The sensor device 16 is extracted for each posture. Then, referring to the set of sensor devices 16 (posture identification sensor set DB22) used for identifying the extracted posture and the set of sensor devices 16 for identifying behavior (behavior identification sensor set DB24), The posture and behavior are identified.

  Further, when other conditions are the same, by suppressing the sampling frequency low, it is possible to save power in the acceleration sensor in both the collection and transmission of acceleration data, and it is possible to extend the battery driving time. In order to realize a daily behavior identification system that can withstand long-term use, the relationship between sampling frequency and identification rate was evaluated.

  In the sampling frequency evaluation experiment, data obtained by sampling the acceleration data collected as described above at 50 Hz is used. From this acceleration data, acceleration data of 25 Hz, 12.5 Hz, 6.25 Hz, and 3.125 Hz were generated by downsampling. For these five different sampling frequencies, the identification of each case was sought. At this time, based on the combination of the optimal mounting positions of the acceleration sensor 32 obtained in the above-described experiment, the sensor data used for learning and evaluation is only the data of “both wrists and both ankles”, and the class classifier is also SVM. Only was used. When performing feature calculation, the number of samples included in the window varies depending on the sampling frequency, but the window size length is kept constant and the number of samples is variable.

  FIG. 17 shows a graph showing the identification rate for each action when the sampling frequency is changed. Referring to the graph shown in FIG. 17, it can be seen that when the sampling frequency is lowered as a whole, the identification accuracy tends to be lowered. On the other hand, “computer work” can be cited as an action showing the opposite tendency. Comparing the result of sampling at 50 Hz shown in FIG. 10 with the result of sampling at 6.25 Hz as shown in FIG. 18, it can be seen that the reduction in the identification rate is small even when the sampling frequency is lowered. . However, in the “sitting” posture, the number of samples misidentifying “computer work” as “meal” decreases due to a decrease in sampling frequency. On the contrary, “meal” increases the number of samples misidentified as “reading” due to a decrease in frequency, and “reading” increases the number of samples misidentified as “computer work” due to a decrease in frequency. This means that when the sampling frequency is 6.25 Hz, three types of actions (“computer work”, “meal”, and “reading”) regarding the “sitting” posture cannot be correctly identified. Referring to FIG. 17, it is considered that a sampling frequency of 12.5 Hz or more is necessary for identification of actions in the “sitting” posture. In this embodiment, since the main purpose is not to save power in the acceleration sensor 32 (sensor device 16), the sampling frequency is fixed to 50 Hz.

  The server 12 executes the operation as described above according to the flowcharts shown in FIGS. Specifically, the server 12 executes the posture / behavior identification feature amount learning process according to the flowcharts shown in FIGS. 19 to 21, and executes the posture / behavior identification process according to the flowchart shown in FIG.

  As illustrated in FIG. 19, when the server 12 starts the posture / action identification feature amount learning process, the server 12 reads a posture identification learning data set from the sensor DB 20 in step S1. Specifically, the server 12 reads acceleration data detected by each sensor device 12 attached to the subject as a behavior record of a subject from the sensor DB 20. Here, in this posture identification learning data set, posture classes are labeled corresponding to the acceleration data. That is, identification information (ID) indicating any of standing, sitting and walking is added. However, in this embodiment, the action identification learning data set to be described later is the same as the attitude learning data set. Therefore, not only the attitude class identification information but also the action identification information is included in the acceleration data. It has been added. Similar to the training data described above, these identification information can be labeled when the acceleration data is acquired. However, if the behavior of the subject is recorded with a video camera or the like, it can be labeled afterwards according to the recording.

  In the subsequent step S3, a set of all sensors is read as a sensor set S. In the next step S5, posture identification feature extraction processing is executed. Here, the server 12 calculates a feature amount (posture identification feature amount) for identifying a posture, that is, a feature vector fp. This feature vector fp is expressed by Equation 5. As described above, the average, standard deviation, energy, entropy, and correlation coefficient are calculated for each window from the accelerations detected by all the acceleration sensors 32, and these are the elements of the feature vector fp.

[Equation 5]
fp = F (S) = {fp1, fp2,..., fpNp}
Subsequently, in step S7, the feature vector fp is subjected to principal component analysis, and a feature (feature vector fpr) having a contribution rate larger than the threshold value ηp is selected. The selected feature vector fpr is expressed by Equation 6.

[Equation 6]
fpr = {fpr1, fpr2,... fprk} (rk ≦ Np)
As can be seen from Equation 6, the feature vector fpr is represented by the selected rk feature quantities.

Here, as described above, there are five types of feature amounts, but the feature amount is obtained for each axis of acceleration, and the correlation coefficient is obtained for all combinations between two different axes. Thus, the feature vector fp is obtained. Therefore, for example, when using acceleration detected by the four acceleration sensors 32, 12 (3 axes × 4) pieces of “average”, “standard deviation”, “energy” and “entropy” are used. A feature amount is obtained, and 66 ( ₁₂ C ₂ = 12 × 11/2) feature amounts are obtained as correlation coefficients. That is, the feature vector fp in this case has 114 dimensions. At this time, the variable Np = 114. However, the correlation coefficient obtains the correlation between the acceleration of each axis of the four acceleration sensors 32 (the acceleration of 12 axes) and the acceleration of the other axes.

  When the principal component analysis is performed on the feature vector fp, eigenvalues λi (i = 1, 2,..., Np) from the first principal component to the Np principal component are calculated in descending order of eigenvalues. At this time, the contribution rate is defined as a value (percentage) obtained by dividing each eigenvalue by the total value of the eigenvalues. That is, the contribution ratio of each eigenvalue λi is expressed by Equation 7.

[Equation 7]

  It is determined whether or not the value of the contribution rate exceeds the threshold value ηp, and the feature vector fpr is obtained. For example, the threshold ηp is determined to be 1 (%). The greater the eigenvalue λi, the greater the variance of the principal component scores, and the greater the eigenvector (weighted composite feature vector of the original feature axis) corresponding to (belonging to) each eigenvalue λi is to explain the entire feature space (the amount of information is large). ). However, it can be said that the feature axis having an extremely small eigenvalue λi is not necessary because the force explaining the entire feature space is small. Therefore, by deleting a feature axis (feature amount) having a small eigenvalue λi from the feature vector fp, the feature vector is compressed to a feature vector fpr having a small dimension.

  In this embodiment, the feature vector fpr is obtained using the contribution rate, but the cumulative contribution rate may be used. Here, the cumulative contribution rate means the contribution rate accumulated in order from the first principal component. In this case, for example, the threshold value ηp is determined to be 95 (%), and a feature vector fpr employing the first principal component to the smallest k-th principal component (k ≦ Np) exceeding the threshold value ηp is generated.

  Returning to FIG. 19, in the next step S9, a sensor set Sm not used for the feature vector fpr is obtained. That is, the sensor device 16 (acceleration sensor 32) in which the feature quantity (element or component) exists in the feature vector fp but the feature quantity does not exist in the feature vector fpr is detected. In step S11, it is determined whether the sensor set Sm is an empty set φ.

If “NO” in the step S11, that is, if the sensor set Sm is not the empty set φ, the sensor set S is updated in a step S13 (S = S−Sm). That is, the number of sensor devices 16 (acceleration sensors 32) used for posture identification is reduced. Thereafter, the process returns to step S5. On the other hand, if “YES” in the step S11, that is, if the sensor set Sm is the empty set φ, the posture identifying sensor set S is written in the posture identifying sensor set database in a step S15. That is, the server 12 stores the sensor set S in the posture identification sensor set DB 22 as the posture identification sensor set Sp ^* .

Subsequently, as shown in FIG. 20, in step S <b> 17, a behavior identification learning data set is read from the sensor DB 20. Here, the same data as the data read in step S1 is read. In the next step S19, the posture ωpiε {p1, p2,..., PNp} is selected. In this posture ωpi, labels (identification information) of all posture classes are described. The posture ωpi is determined in advance by an administrator or a user of the behavior identification system 10 and is stored in a DB such as the sensor DB 20 or an internal memory of the server 12 although illustration is omitted. In step S21, it is determined whether the variable i exceeds the maximum value Np (i> Np). That is, it is determined whether or not a behavior identification sensor set Sa ^* to be described later is obtained for all posture classes included in the posture ωpi.

  If “YES” in the step S21, that is, if the variable i exceeds the maximum value Np, the posture / behavior identification feature amount learning process is ended as it is. On the other hand, if “NO” in the step S21, that is, if the variable i is equal to or less than the maximum value Np, the action A (ωpi) as shown in the equation 8 is read in a step S23. Here, the action A (ωpi) is a label of action (identification information) that can be taken in the posture ωpi, that is, a label of an action that can be taken in the posture class indicated by the posture ωpi.

[Equation 8]
A (ωpi) = {a1, a2,..., ANpi}
In subsequent step S25, as in step S3, all sensor sets are read out as set S. As shown in FIG. 21, in the next step S27, sensor data (acceleration data) Z (S, A (ωpi)) corresponding to the action label is read. Subsequently, in step S29, feature amounts (average, standard deviation, energy, entropy, correlation coefficient) are extracted from the sensor data Z. That is, as shown in Equation 9, an m-dimensional feature vector X is obtained.

[Equation 9]
X = {x1, x2,..., Xm}
In subsequent step S31, the feature vector X is subjected to principal component analysis, and a feature having a contribution rate larger than the threshold value τ is selected. The selected feature vector Xr is expressed by Equation 10. These are the same as generating the feature vector fpr from the above-described feature vector fp.

[Equation 10]
Xr = {xr1, xr2,... Xrl} (rl <K)
As can be seen from Equation 10, the feature vector Xr is represented by the selected rl feature quantities.

  In the next step S33, a sensor set Sn that is not used for the feature vector Xr is obtained. That is, the sensor device 16 (acceleration sensor 32) that has an element in the feature vector X but does not have an element in the feature vector Xr is detected. In step S35, it is determined whether the sensor set Sn is an empty set φ.

If “NO” in the step S35, that is, if the sensor set Sn is not the empty set φ, the sensor set S is updated in a step S37 (S = S−Sn). That is, the number of sensor devices 16 (acceleration sensors 32) used for action identification is reduced. Thereafter, the process returns to step S27. On the other hand, if “YES” in the step S35, that is, if the sensor set Sn is an empty set φ, the action identifying sensor set S is written in the action identifying sensor set database in a step S39 and shown in FIG. Return to step 19. That is, in step S39, the server 12 stores the sensor set S as the posture identification sensor set Sa ^* in the posture identification sensor set DB 24.

As described above, FIG. 22 is a flowchart showing the posture / behavior identification process. As shown in FIG. 22, when starting the posture / behavior identification process, the server 12 reads the posture identification sensor set Sp ^* and the behavior identification sensor set Sa ^* from each database (22, 24) in step S51. In the next step S53, feature extraction is performed using the data (acceleration data) of the sensor set Sp ^* . That is, a feature vector Fp as shown in Equation 11 is obtained. Each element (feature amount) of the feature vector Fp is the same as the feature vector fp described above.

[Equation 11]
Fp = {fp1, fp2,..., FpN}
In step S55, posture identification processing is executed (ωp = classify_posture (Fp)). Here, a posture identification process (a method for identifying a posture class) will be described. As described above, the posture classes are the three classes of standing, sitting, and walking, and supervised learning is applied as the feature amount of each sliding window. The scatter diagram in the feature space of each class forms a cluster as shown in FIG. 11 or FIG. As a classifier used for learning, a k-nearest neighbor (k-Nearest Neighbor), a linear discriminant function, a naive Bayes method, a decision tree learning, a support vector machine (SVM), or the like can be applied.

  However, data on clusters according to the scatter diagrams as shown in FIG. 11 or FIG. 12, FIG. 13 (A), FIG. 13 (B) and FIG. 14 is stored in advance in the internal memory of the server 12 as teacher data. . Here, “data about a cluster” means, for example, information that can identify a class (posture or action) corresponding to each cluster, and coordinate data indicating the cluster in the scatter diagram.

  For example, when the k-nearest neighbor method is applied as a discriminator, k samples from the order in which the Euclidean distance from the cluster defined by the teacher data is close when mapping on the feature space (FIGS. 11 and 12). And the class (posture class) with the largest number among the k samples can be used as the identification result.

  When constructing a multi-class support vector machine, it is performed by training a pair-wise classifier that discriminates each class from other individual classes for the number of classes. In the embodiment, an SMO (Sequential Minimal Optimization) method is used, and a one-dimensional polynomial kernel is used as a kernel function. However, regarding the SMO method, “SS Keerthi, S. K, Shevade, C. Bhattacharyya, KR K, Murthy: Improvements to Platt's SMO Algorithm for SVM Classifier Design, Neural Computation, 13 (3), pp. 637-649, 2001 "(reference 7).

  The identification results (identification rates) shown in FIG. 9, FIG. 10, FIG. 15 and FIG. 16 are results when an SVM based on the SMO method is used as a discriminator.

When the posture class is identified in step S55, in step S57, a sensor set Sa ^* (ωp) used for action identification is determined based on the identified posture ωp. That is, a sensor set Sa ^* (ωp) for identifying actions that can be taken in the posture class (ωp) identified in step S55 is determined. In subsequent step S59, feature extraction is performed using the data of the determined sensor set Sa (ωp). That is, a feature vector Fap as shown in Equation 12 is obtained.

[Equation 12]
Fap = {fa1, fa2,..., FaM}
Next, in step S61, action identification processing is executed (ωa = classify_action (Fap)). Since this action identification process is the same as the above-described posture identification process in step S55, a duplicate description is omitted. Accordingly, in step S61, one action is identified (specified) from the plurality of actions included in the posture class ωp. In subsequent step S63, the action identification result ωa is output. For example, the action identification result ωa is displayed as text on a monitor (not shown) connected to the server 12, or is output to a speaker (not shown) connected to the server 12 by voice, or both of them. Or output by means. However, the action identification result ωa may be transmitted (output) by e-mail to another terminal (computer or telephone). In addition to simply outputting the action identification result ωa, the identification result (posture and action) may be labeled on the acceleration data.

  In step S65, it is determined whether or not the posture / behavior identification process is to be terminated. Here, it is determined whether or not an end instruction from the user is input or whether the processing for all acceleration data is ended. If “NO” in the step S65, that is, if the posture / action identifying process is not ended, the process returns to the step S53 as it is, and the posture / action identifying process is executed for the next sample. On the other hand, if “YES” in the step S65, that is, if the posture / behavior identification processing is ended, the posture / behavior identification processing is ended as it is.

  According to this embodiment, since one action that can be taken in the posture class is identified after the posture is identified, the behavior of the subject can be accurately identified. In addition, since a proper sensor set is used for posture and action identification, posture and action can be more accurately identified.

  In this embodiment, acceleration data is collected by the mobile terminal and transmitted to the server so that the server identifies the action. However, the present invention is not limited to this. By storing posture pattern and action pattern data in the internal memory of the mobile terminal, the mobile terminal can also identify the action. That is, it is possible to identify postures and actions regardless of online or offline.

  In this embodiment, transmission data including acceleration data is transmitted from the mobile terminal to the server at regular intervals. However, after collecting all acceleration data, the transmission data may be transmitted to the server. Good. Here, all the acceleration data is acceleration data for the behavior of the subject in one day, acceleration data that can be stored in the memory of the mobile terminal, or acceleration that can be recorded by the remaining amount of the battery that drives the mobile terminal. Means data.

  Furthermore, in this embodiment, the transmission data is transmitted from the portable terminal to the server through the repeater. However, the transmission data may be transmitted directly from the portable terminal to the server without passing through the repeater.

  Furthermore, in this embodiment, 10 types of actions are classified into 3 types of postures to identify postures and behaviors. However, if there are 2 or more postures, 4 or more types may be used. As long as there are two or more types of actions, there may be eleven or more types. That is, it should not be limited to the posture and action of the embodiment.

  Further, in this embodiment, a case has been described in which an action for one subject is identified. However, there may be a plurality of subjects. In such a case, a posture identification sensor set and a behavior identification sensor set may be created for each subject, and the posture and behavior may be identified using the sensor set corresponding to the subject. Further, by taking the average of the acceleration data of each subject, it is possible to create a posture identifying sensor set and a behavior identifying sensor set that can be used for all subjects. In such a case, the storage capacity of the database can be saved.

  Further, in this embodiment, only the case where the acceleration sensor is used as the sensor for detecting the behavior (motion) of the subject has been described, but a gyro sensor (three axes) can be used instead of the acceleration sensor. In this case, on the basis of the angular velocity detected by the gyro sensor, five types of feature amounts of average, standard deviation, energy, frequency domain entropy, and correlation are calculated for each sliding window.

FIG. 1 is an illustrative view showing one example of an action identification system of the present invention. FIG. 2 is an illustrative view showing an electrical configuration of the sensor device and the repeater of FIG. FIG. 3 is an illustrative view showing a state in which the subject wears the sensor device and the repeater of FIG. FIG. 4 is an illustrative view showing an example in which right wrist acceleration data is divided by a sliding window of a fixed time. FIG. 5 is a bar graph showing a combination of acceleration sensor mounting positions for obtaining the highest discrimination result in each classifier and the discrimination result. FIG. 6 is a line graph obtained by obtaining a combination of acceleration sensors for obtaining the highest discrimination result for each classifier and plotting the discrimination results. FIG. 7 is a bar graph showing the effect of dominant hand or dominant foot on the identification result when SVM is used as a classifier. FIG. 8 is a line graph showing identification results for each of the ten types of actions with respect to the SVM results shown in FIG. FIG. 9 is an illustrative view showing a confusion matrix of identification results for each action when all acceleration sensors are used. FIG. 10 is an illustrative view showing a confusion matrix of identification results for each action when using acceleration sensors attached to both wrists and both ankles. FIG. 11 is a scatter diagram on the first and second variables in the canonical feature space when all the acceleration sensors are used. FIG. 12 is a scatter diagram on the first and second variables in the canonical feature space when using acceleration sensors attached to both wrists and both ankles. FIG. 13 is an illustrative view showing a distribution for each action in the canonical feature space according to the posture of walking and standing when using acceleration sensors mounted on both wrists and both ankles. FIG. 14 is an illustrative view showing a distribution for each action in the canonical feature space corresponding to the posture of the sitting position when using acceleration sensors attached to both wrists and both ankles. FIG. 15 is an illustrative view showing identification results of postures and identification results of actions that can be taken in each posture when all sensor devices are used. FIG. 16 is an illustrative view showing posture identification results and action identification results that can be taken in each posture when using acceleration sensors mounted on both wrists and both ankles. FIG. 17 is a line graph showing changes in the identification rate for each sampling frequency when using acceleration sensors mounted on both wrists and both ankles. FIG. 18 is an illustrative view showing a discrimination result at a sampling frequency of 6.25 Hz when using acceleration sensors attached to both wrists and both ankles. FIG. 19 is a flowchart showing a part of the server attitude / behavior identification feature amount learning process shown in FIG. 20 is another part of the server attitude / behavior identification feature amount learning process shown in FIG. 1, and is a flowchart subsequent to FIG. FIG. 21 is another part of the server attitude / behavior identification feature amount learning process shown in FIG. 1, and is a flowchart subsequent to FIG. FIG. 22 is a flowchart showing the server attitude / behavior identification processing shown in FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Action identification system 12 ... Server 14 ... Network 16 (16a, 16b, 16c, 16d, 16e, 16f) ... Sensor apparatus 18 ... Repeater 20, 22, 24 ... Database 30, 50 ... CPU
32 ... Acceleration sensor 34, 54 ... RAM
36, 56 ... Bluetooth module 52 ... Interface

Claims (6)

Detection means for detecting motion detection data from a plurality of motion detection sensors attached to the subject;
Posture feature amount calculating means for calculating posture feature amount for identifying the posture of the subject based on the motion detection data detected by the detection portion;
Posture identification means for identifying the posture of the subject based on the posture feature value calculated by the posture feature value calculation means;
An action feature amount calculating means for calculating an action feature amount for identifying an action that can be taken in the posture identified by the posture identifying means based on the motion detection data detected by the detecting means; and the action feature A behavior identification device comprising behavior identification means for identifying the behavior of the subject from a plurality of behaviors that can be taken in the posture identified by the posture identification means based on the behavior feature amount calculated by the amount calculation means. The plurality of motion detection sensors are attached to a plurality of different parts of the subject,
The posture feature amount calculating means is configured to detect from all or a part of the motion detection sensors for identifying actions that can be taken in the posture identified by the posture identification means among all the motion detection sensors. The behavior identification device according to claim 1, wherein the posture feature amount is calculated based on motion detection data.   The behavior feature amount calculating means calculates the behavior feature amount based on motion detection data from all or a part of the motion detection sensors according to the behavior to be identified among all the motion detection sensors. The action identification device according to claim 2. The plurality of motion detection sensors are attached to a plurality of different parts of the subject,
The behavior feature amount calculating means calculates the behavior feature amount based on motion detection data from all or a part of the motion detection sensors according to a behavior to be identified among all the motion detection sensors. The action identification device according to claim 1. A behavior identification system comprising a portable terminal possessed or worn by a subject and a behavior identification device,
The portable terminal is
Detection means for detecting motion detection data from a plurality of motion detection sensors attached to the subject, and transmission means for transmitting the motion detection data detected by the detection means to the behavior identification device,
The behavior identification device includes:
Receiving means for receiving motion detection data transmitted by the transmitting means;
Posture feature value calculating means for calculating posture feature value for identifying the posture of the subject based on the motion detection data received by the receiving means;
Posture identification means for identifying the posture of the subject based on the posture feature value calculated by the posture feature value calculation means;
An action feature amount calculating means for calculating an action feature amount for identifying an action that can be taken in the posture identified by the posture identifying means based on the motion detection data received by the receiving means; and the action feature A behavior identification system comprising behavior identification means for identifying the behavior of the subject from a plurality of behaviors that can be taken in the posture identified by the posture identification means based on the behavior feature amount calculated by the quantity calculation means. A computer behavior identification method for identifying the posture and behavior of a subject,
(a) detecting motion detection data from a plurality of motion detection sensors attached to the subject;
(b) Based on the motion detection data detected by the step (a), calculate a posture feature amount for identifying the posture of the subject,
(c) based on the posture feature amount calculated in step (b), identifying the posture of the subject,
(d) based on the motion detection data detected in the step (a), calculating an action feature amount for identifying an action that can be taken in the posture identified in the step (c); and
(e) A behavior identification method for identifying the behavior of the subject from a plurality of behaviors that can be taken in the posture identified in the step (c) based on the behavior feature amount calculated in the step (d).