JP2010207488A - Behavior analyzing device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a behavior analyzing device, classifying behaviors of people at optimum, and more efficiently discriminating and analyzing with high accuracy even the behaviors imperfectly associated with measurement data of a sensor in the past. <P>SOLUTION: This behavior analyzing device includes: an analyzing means for analyzing the behavior of the human body using acceleration data output from a triaxial acceleration sensor applied to the human body; and a storage means. The analyzing means includes: a basic behavior pattern defining part for separating the feature amount of posture of the human body included in change with time of the acceleration data stored as a time series data from the feature amount of motion of the human body, and defining a model composed of the motion of the basic behavior pattern and the feature amount of posture; and an expanding behavior pattern recognition part for performing the processing to separate the acceleration data newly stored as a time series data by utilizing the model of the basic behavior pattern and clustering the processed acceleration data to recognize an expanded behavior pattern. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a behavior analysis apparatus and a behavior analysis program for classifying and analyzing human behavior patterns. In particular, the present invention relates to a behavior analysis apparatus and a behavior analysis program that automatically analyze human behavior using data output by an acceleration sensor and accumulated as time series data.

  Attempts have been made to estimate a user's behavior and perform a detailed analysis using measurement data obtained from the sensor mounted on the user. For example, Patent Literature 1 discloses an action discrimination system using an acceleration sensor. The behavior discrimination system of Patent Document 1 compares output data indicating a user's posture and behavior obtained by using an acceleration sensor with posture and behavior pattern data stored in advance in a storage unit. The posture state and the motion state of the human body are discriminated. Non-Patent Document 1 discloses a method for estimating a user's behavior and posture with an information terminal equipped with a single acceleration sensor.

  In behavioral analysis using a sensor, the content of behavior may be measured and analyzed in detail, limited to a user's specific behavior. Particularly in recent years, there have been many attempts to analyze a person's walking in detail using sensors. By making it possible to know the state of walking in detail with sensors, we can monitor elderly people and infants while taking privacy into consideration, measure the amount of exercise such as steps, and maintain or improve health. It is possible to optimize the flow line, the surrounding layout, and the like in the work place.

  Patent Document 2 discloses a walking motion detection processing device that detects a human walking motion using an angular velocity sensor and an acceleration sensor. The detection processing device of Patent Document 2 detects gravitational acceleration from the values of the angular velocity vector and acceleration vector, removes the acceleration in the vertical direction from the gravitational acceleration, extracts a residual acceleration component, and obtains the obtained residual acceleration component. The principal component analysis is performed, and the walking motion is detected using the principal component vector. Patent Document 3 discloses a walking motion analyzer that analyzes walking motion by measuring body motions at a plurality of parts of a human body using sensors. The walking motion analysis device of Patent Document 3 calculates and stores a correlation between body motion and physical strength characteristics when performing a walking motion in advance, and uses the measured body motion and the calculated physical strength characteristics. The classification result of walking motion is displayed. Classification of walking motion is performed by using cluster classification based on body motion and physical strength characteristics.

  In addition, attempts have been made to estimate a plurality of behaviors such as “walking”, “running”, “sitting”, and “buying” by statistically processing the output data of the sensor worn by the user. . For example, Patent Document 4 discloses a state analysis device that attaches a plurality of sensors to a plurality of locations on the body of an observation subject and records and analyzes the behavioral state on the body. The state analysis apparatus of Patent Literature 4 calculates normalized sensor information and a moving average based on information about the movement and posture of the observation target obtained by the sensor. Based on the difference between the value of the moving average and the normalized sensor information and a predetermined threshold, it is determined whether the observation target person is in the supine position, sitting position, standing position, walking, or wheelchair running. Non-Patent Document 2 discloses a technique for identifying seven types of actions using the average, standard deviation, and energy of accelerations obtained from acceleration sensors mounted at six locations as feature quantities. The technique of Non-Patent Document 2 is considering improvement of the action identification rate by applying two speaker adaptation techniques of MAP estimation method and MLLR method used in speech recognition.

  In addition, in the Non-Patent Document 3, the inventors have examined the validity of the feature amount used for recognizing various actions from the output data of the acceleration sensor. Inventors use MFCC (Mel-Frequency Cepstial Coefficients) and feature quantities considering its primary and secondary dynamic feature quantities in Non-Patent Document 3, and optimize their frequency characteristics. The technology which improves the recognition rate of action is disclosed.

JP 2003-61935 A JP 2005-114537 A JP 2008-173250 A JP 2006-116325 A

Kurasawa Hiroshi, Kawahara Akihiro, Morikawa Hiroyuki, Aoyama Yuki “Attitude Estimation Method Using 3-Axis Accelerometer Considering Sensor Placement”, Information Processing Society of Japan Research Report 2006-UBI-11- (3), Vol. 2006. No. 54, pp. 15-22, 2006 Naoyuki Hashida, Ren Omura, Ryota Imai “Contribution to identification accuracy of personal adaptation technology in daily behavior identification using acceleration sensor”, Information Processing Society of Japan Research Report 2008-UBI-19- (13), Vol. 2008. No. 66, pp. 69-74, 2008 Shinichi Takeuchi, Shinya Ito, Tetsugo Tamura, Satoru Hayami “Recognition of Daily Behavior Using Acceleration Information”, IEICE Technical Report, vol. 108, no. 453, CAS 2008-142, pp. 229-234, 2009

  In the conventional behavior analysis technology, the behavior pattern to be analyzed is limited to one to several types, and the user's movement and posture are measured by a sensor, and whether the obtained data is applicable to the behavior pattern to be analyzed. The purpose is to identify exactly. Alternatively, an object is to provide a method or means for detecting more detailed individual differences or characteristics of specific actions such as “walking” and “running”.

  However, it is difficult to classify human daily behavior into a small number of patterns, and there are many behaviors that have not yet been classified. The behavior that has not been classified is mostly the case where the identification pattern or identification model that can identify the behavior is not defined from the measured value by the sensor, but this behavior was tried to be analyzed by the conventional analysis technology. In some cases, a lot of work is required.

  In order to analyze an action by a conventional analysis device, first, a category including a series of contents of the action to be analyzed is set. Next, a large amount of measurement data is prepared by measuring the behavior to be analyzed at a certain time interval. Then, it is necessary to define the identification pattern and the identification model by processing the information included in the measurement data, extracting the feature amount representing the behavior, and causing the analysis device to learn the feature amount. In this way, a great deal of effort is required to define and analyze new behaviors using conventional behavior analysis techniques. For this reason, the technique which can identify and classify | categorize a new action more efficiently was calculated | required.

  Conventionally, in clustering of data output from sensors, a method has been widely used in which an evaluation function is defined for a feature space to obtain an optimal number of categories and an optimal classification. However, there is a limitation that the information obtained from the acceleration sensor is a mixture of posture information and motion information, which may be an obstacle to clustering. For this reason, there has been a demand for a technique for more accurately identifying and analyzing actions using time-series information in which posture and movement information are mixed.

  The present invention has been made in view of the above problems, and classifies human behavior optimally, and even if it is behavior that has not been sufficiently correlated with measurement data by a sensor so far, it is more efficient and highly accurate. It is made in order to provide the behavior analysis apparatus and behavior analysis program which identify and analyze.

  The behavior analysis apparatus according to the present invention includes analysis means for analyzing the behavior of the human body using acceleration data output from triaxial acceleration attached to the human body, and storage means. The analysis means includes a basic action pattern definition unit and an extended action pattern recognition unit. The basic behavior pattern definition unit separates the feature quantity of the human body motion and the feature quantity of the human body posture included in the temporal change of the acceleration data continuously output and accumulated as time series data, and is determined in advance. The basic action is associated with each feature quantity, a model composed of the basic action pattern movement and the posture feature quantity is defined, and the model is stored in the storage means. The extended action pattern definition unit uses the basic action pattern model to perform processing to separate newly accumulated acceleration data as time-series data, and clusters the processed acceleration data to recognize the extended action pattern It is characterized by that.

  The basic behavior pattern definition unit of the present invention is characterized in that the feature quantity of the human body motion and the feature quantity of the posture are separated based on the distribution of the feature quantity included in the acceleration data accumulated as time series data.

  The basic behavior pattern definition unit of the present invention obtains each feature amount after separating acceleration data in a high frequency region as human body motion information and low frequency region acceleration data as human body posture information. And

  The extended action pattern recognition unit of the present invention separates the feature quantity of the human body motion and the feature quantity of the human body posture included in the temporal change of the acceleration data, and divides each feature quantity for each unit time. A basic element model is created by decomposing acceleration data.

  Further, the extended action pattern recognition unit of the present invention is characterized in that data processing is performed to merge feature values of acceleration data that has been processed to be decomposed.

  Furthermore, the extended action pattern recognition unit of the present invention collates a time-dependent change of newly accumulated acceleration data with a basic action pattern model and a basic element model to obtain an action pattern that is not defined in the basic action pattern. A feature amount of a motion to be represented and a feature amount of a posture are detected, a new extended action pattern model is defined, and related quantitative measurement is performed to recognize the new extended action pattern.

  The present invention further provides a behavior analysis program. The behavior analysis program of the present invention is included in a procedure for accumulating acceleration data representing arbitrary behavior continuously output from a three-axis acceleration sensor attached to a human body as time-series data, and a change with time of the acceleration data. Define the model consisting of the basic behavior pattern features by associating the procedure that separates the features of human movement and the features of the posture of the human body with the basic behaviors that have been determined in advance. A procedure for storing the model in the storage means, a procedure for performing processing for separating acceleration data newly accumulated as time-series data using the model of the basic behavior pattern, and processed acceleration data And a procedure for recognizing an extended action pattern by clustering the computer.

  The behavior analysis apparatus according to the present invention can identify and analyze a behavior more accurately by separating and analyzing the feature quantity of the human body motion and the feature quantity of the posture from the acceleration data output from the acceleration sensor. Is possible.

  The extended behavior pattern recognition unit of the behavior analysis device of the present invention can automatically define a behavior model for which no model is defined and perform quantitative analysis of the behavior. As a result, the amount of work involved in the analysis is reduced, and behavior analysis can be performed more quickly and efficiently. This also facilitates the recognition and quantitative analysis of more diverse behavior patterns.

FIG. 1 is a diagram schematically showing a state in which a subject wearing a triaxial acceleration sensor is lying on the side. FIG. 2 shows data obtained by accumulating acceleration detected by the “walking” action output from the acceleration sensor in time series. FIG. 3 shows data obtained by accumulating accelerations detected by the action of “sitting on a chair” output from the acceleration sensor in time series. FIG. 4 shows data obtained by accumulating acceleration detected by the action “standing from a chair” output from the acceleration sensor in time series. FIG. 5 shows data obtained by accumulating the acceleration detected by the “sleeping in bed” action output from the acceleration sensor in time series. FIG. 6 shows data obtained by accumulating acceleration detected by the action “getting out of bed” output from the acceleration sensor in time series. FIG. 7 is a flowchart of a procedure performed by the basic action pattern definition unit shown in the embodiment. FIG. 8 is a diagram showing the results of frequency analysis of acceleration data of “walking” behavior for each x, y, and z component. FIG. 9 is a diagram illustrating examples of posture feature values and motion feature values applied to the “walking” behavior stored in advance by the basic behavior pattern definition unit. FIG. 10 is an example of data defining a basic behavior pattern model. FIG. 11 is a flowchart of a procedure until the extended action pattern recognition unit learns an extended action pattern. FIG. 12 is a flowchart of a procedure until the extended action pattern recognition unit recognizes an extended action pattern. FIG. 13 is a diagram showing acceleration data and a model of an action newly recognized by the process of the extended action pattern recognition unit. FIG. 14 is a diagram illustrating an example of a screen that outputs the behavior recognized by the behavior analysis device according to the present embodiment. FIG. 15 is a flowchart of a procedure until the extended action pattern recognition unit performs quantitative measurement of actions related to the extended action pattern. FIG. 16 is a block diagram schematically illustrating the configuration of the behavior analysis apparatus.

  In the following, as an example of a preferred embodiment for carrying out the present invention, six types of behaviors of “walking”, “sitting in a chair”, “standing from a chair”, “sleeping on a bed”, “getting up from a bed”, and “stopping” are basic behaviors The basic behavior pattern model is defined using the data of the triaxial acceleration sensor worn by the subject, and the behavior pattern that is not classified as the basic behavior is newly defined from the newly input triaxial acceleration sensor data. An analysis apparatus that recognizes and defines a model will be described in detail with reference to the drawings.

  The configuration of the behavior analysis apparatus in the present embodiment is schematically shown in FIG. The behavior analysis apparatus is composed of one computer 3, and the CPU 4 and the storage means 5 are built in the computer 3. The basic action pattern definition unit and the extended action pattern recognition unit of the analysis unit are stored as programs in the storage unit of the computer 3, and necessary processing is executed by the CPU 4 of the computer 3.

  In the present embodiment, a triaxial acceleration sensor having a wireless communication function is used as the triaxial acceleration sensor 1 that continuously outputs the acceleration detected by being mounted on the body of the subject 2. The acceleration sensor 1 outputs the detected acceleration as a component value of a three-dimensional (x, y, z) orthogonal coordinate system defined in the sensor, and transmits the component value to the computer of the behavior analysis device by wireless communication. The sampling frequency of the acceleration sensor 1 is a maximum of 100 Hz. As shown in FIG. 1, the acceleration sensor 1 was mounted on the waist of the subject 2 to detect acceleration. Since the acceleration sensor 1 is small, it does not obstruct the behavior of the subject 2.

  The basic action pattern definition unit performs the following processing on the acceleration data of each action output by the acceleration sensor 1 to define a multi-dimensional feature value in order to define each basic action pattern model. Extract and define a model composed of these features. FIG. 7 shows a flowchart of a procedure until the basic action pattern definition unit defines a model of the basic action pattern.

  In step 1 of FIG. 7, the basic behavior pattern definition unit divides the acceleration data detected by the acceleration sensor 1 and accumulated in time series into basic behaviors based on information input by the operator when detecting acceleration. Label the. 2 to 6 show acceleration data divided for each action and labeled with a basic action name as a label. The acceleration data is output as components in three orthogonal directions of x, y, and z. FIG. 2 shows data obtained by accumulating accelerations detected by “walking” behavior in time series. FIG. 3 shows data obtained by accumulating accelerations detected by the action of “sitting on a chair” in time series. FIG. 4 shows data obtained by accumulating the acceleration detected by the action “standing from the chair” in time series. FIG. 5 shows data obtained by accumulating acceleration detected in the “sleeping on bed” behavior in time series. FIG. 6 shows data obtained by accumulating acceleration detected in the action “getting up from bed” in time series. In addition, although a “stop” action is defined as the basic action, the data output from the acceleration sensor 1 when it is completely stopped is only the gravitational acceleration. Modeling behavior patterns is omitted.

The basic action pattern definition unit performs frequency analysis such as Fourier transform using acceleration data segmented for each action. Alternatively, arithmetic processing for each unit time can be performed on the segmented acceleration data. In this case, data is obtained at the sampling period of acceleration before processing. The feature amount obtained from the acceleration data by this processing can be defined as follows. Here, t represents the frame number of the segmented acceleration data (time when processing per unit time is performed). The feature quantity is treated as vector data of K = 3J dimension in which the J dimension for each axis is arranged in the directions x, y, z of the acceleration axis. A K-dimensional feature vector _ct is obtained for each frame t. This whole is represented as C.

Next, the basic action pattern definition unit defines a model used for learning a model representing the basic action pattern. The basic action pattern definition unit of the present embodiment defines the model as follows. In the following, m is a model number for representing a basic action pattern. In this embodiment, since there are six types of basic actions, the number of models M is six at the maximum. s is an action pattern having a certain duration, and the number of states is S. A parameter B ^m of the occurrence probability of the feature amount in the state s of the model number m is expressed by the following equation.

At this time, let P ^m be the probability that the observed feature quantity C occurred from the model m. A model number that maximizes P ^m is obtained, and this model number is set as the category of the recognized action pattern. In other words, an expression representing the category of the behavior pattern is shown below.

  When a hidden Markov model is applied to define a basic behavior pattern model, the transition probability between states is also considered. Instead of the occurrence probability, a model that uses a distance from the average value of feature values arranged in the time direction, or a model that considers time expansion and contraction by DP (dynamic programming) matching can be defined. A forward-backward algorithm can be used as a method for estimating the time-series data of feature quantities and the probability model for each state while obtaining a hidden Markov model. In other cases, a temporal alignment between the model and the observed data is obtained. Then, a model can be defined by obtaining a statistic such as an average value from a set of feature values corresponding to a certain time interval.

  The basic action pattern definition unit performs learning for defining a model for each basic action pattern, as shown in Step 2 of FIG. The learning performed in this step is supervised learning. The data used for learning is acceleration data accumulated as time-series data that is divided into basic actions in step 1 and labeled with actions. The basic behavior pattern definition unit learns a behavior pattern model using the feature amount obtained from the value of this label and time-series acceleration data. When the hidden Markov model is applied, its parameters are learned by an EM (Expectation Maximization) algorithm.

  In step S3, the basic behavior pattern definition unit separates posture information and motion information included in the learned acceleration data accumulated in time series. In step S4, a basic behavior pattern model including a posture model and a motion model is defined. The basic action pattern definition unit has multiple procedures for defining each model by separating posture information and motion information, and examines the contents of learned acceleration data and executes one of the procedures. .

  An example of a procedure for separating posture information and motion information is to perform frequency analysis of acceleration data and separate low frequency components as posture information and high frequency components as motion information. . FIG. 8 shows the result of frequency analysis of the acceleration data of the “walking” basic action for each of the x, y, and z components. In FIG. 8, the horizontal axis represents frequency (Hz), and the vertical axis represents the power spectrum of each frequency. The basic behavior pattern definition unit in the present embodiment is a “walking” behavior model, and defines a waveform appearing in a frequency region of 0 Hz or more and 1 Hz or less as a feature amount derived from posture, and exceeds 1 Hz and is 25 Hz or less. A waveform appearing in the frequency domain is defined as a feature amount derived from motion, and posture data and motion data are separated.

  In addition, the basic behavior pattern definition unit can perform a procedure for separating posture information and motion information based on the distribution of feature values included in the learned acceleration data. FIG. 9 shows an example of the posture feature amount and the motion feature amount related to the “walking” behavior that the basic behavior pattern definition unit of the present embodiment stores in advance as the value of the temporal change in the acceleration data.

  Further, the basic action pattern definition unit includes the following procedure for separating posture data and motion data, and changes the processing content according to the content of the learned acceleration data to change the motion model in step S3. And we define the model of posture.

  For example, the basic behavior pattern definition unit performs a frequency analysis of the learned acceleration data, and then performs a filtering operation in the time domain that removes components in the high frequency domain. Then, only a low frequency component is extracted and defined as a posture model. As an example of a specific process of the filtering operation, there is a process of obtaining a moving average of acceleration data and removing a motion component. Furthermore, it is possible to formulate the learned acceleration data as a mixture of a plurality of information, estimate the mixing ratio, and separate posture data and motion data. When a mixed Gaussian distribution (mixed Gaussian model) is used as the output probability distribution of the hidden Markov model, the posture model and the motion model can be defined by this processing.

  Furthermore, when the content related to the movement of the learning data provided to the basic action pattern definition unit is selected as follows, the posture data and the movement data can be more easily separated. That is, if the basic behavior pattern definition unit can present multiple patterns with the same posture and change pattern but different movement speed and size as learning data, the basic behavior pattern definition unit learned Processing that obtains invariant components in a plurality of patterns and separates them as components of the model of the posture can be performed. This process has the effect of reducing the learning burden of the basic action pattern definition unit and increasing the accuracy of the defined model.

  In step S5, the basic behavior pattern definition unit stores the defined basic behavior pattern model in the storage means of the computer. An example of model data based on a mixed Gaussian distribution of basic behavior patterns is shown in FIG. In the data shown in FIG. 10, the name label of the basic action is first described. Further, the data includes items of <number of states>, <states>, <number of mixtures>, <mixture>, <average>, and <variance> for defining a model of a basic action pattern. In the item <number of states>, information on the number of states stored is described. In the <state> item, information on the number of the stored state is described. In the item <number of mixtures>, information on the number of stored Gaussian distributions is described. In the <mixing> item, information of the number of the stored Gaussian distribution is described. In the items <average> and <variance>, information on the stored characteristic quantities of the Gaussian distribution is described.

  The extended behavior pattern recognition unit uses the basic behavior pattern model created by the basic behavior pattern definition unit in the previous steps to process the newly accumulated acceleration data, and the posture and motion derived from the basic behavior pattern And a model for recognizing the extended action pattern, which is a new action pattern, is created. FIG. 11 shows a flowchart of a procedure performed by the extended action pattern recognition unit. A detailed description will be given below according to the flowchart.

  In step S11, the extended behavior pattern recognition unit learns newly added acceleration data. In step S12, the basic behavior pattern model is learned again. The extended behavior pattern recognition unit creates a subset of learning data by associating the newly added learning data with the behaviors of the six basic behavior patterns at the time of relearning in step S12. Next, in step S13, the extended action pattern recognition unit divides the re-learned model of the basic action pattern for each unit time, thereby decomposing the model of the basic action pattern and creating an element model. The extended action pattern recognition unit in the present embodiment decomposes the basic action pattern model by dividing it into units of one second.

  Further, in step S14, the extended behavior pattern recognition unit further subdivides the element model subdivided for each unit time according to the distribution of the feature amount. In step S15, a basic element model for describing the extended action pattern is created by integrating (clustering) the subdivided models. The procedure for integrating the basic element model using the bottom-up integration method is described below.

  In step S16, a basic element model describing an extended action pattern is obtained by obtaining an optimal combination of subsets from a set of element models obtained so far. That is, the element model obtained in the stage up to step S15 is obtained by performing temporal subdivision and subdivision by distribution of feature amounts on the pattern included in the added learning data. Among these combinations, an optimum combination representing the added learning data is obtained.

  Finding all combinations of element models to represent learning data requires a large amount of computation, so a bottom-up integration method generally used in clustering methods is applied. Let N be the number of element models at the start of clustering. First, of the element models, two models that maximize the evaluation criteria by being integrated are selected and integrated. At this time, the number of element models is N-1. Thereafter, this process is repeated until the number of element models reaches a predetermined number L. Considering the set of integrated original element models as a subset, it means that a method for dividing N element models into L exclusive subsets at the start of clustering is obtained.

  In the procedure of step S14 decomposition and step S15 integration performed by the extended action pattern recognition unit, a criterion for evaluating the validity of the process for the model or a subset of the model is required. When a probabilistic model is used to define the behavior pattern, the likelihood can be used as a measure representing the degree to which the learning data describes the model. In addition, when the entire model set is divided into a combination of subsets of independent actions, the goodness of separation of the model set is determined between the class that is the model set and the class that is the set of other models. It can be represented by the ratio of the variance and the variance within the class. Furthermore, as a model for identification, it is possible to use the identification performance when a combination of subsets of independent classes is used as an evaluation criterion. The model whose validity has been evaluated in this way is defined and temporarily stored as a basic element model in step S16.

  An extended action pattern definition unit that creates and defines a basic element model for recognizing an extended action pattern performs a procedure for recognizing an extended action pattern that is a new action pattern. The procedure for recognizing the extended action pattern is shown in FIG. The extended behavior pattern recognition unit acquires acceleration data newly accumulated in time series in step S21. In step S22, the acceleration data is compared with the basic behavior pattern model and the created basic element model. In step S23, the extended action pattern recognition unit detects an action pattern that is not defined in the basic action pattern and corresponds to the basic element model, assigns a class name as a new action pattern, and sets the start time and Specify end time and likelihood. The event described in step S23 of FIG. 12 is a behavior pattern corresponding to a new classification of behavior patterns or an automatically formed behavior class. In step S24, the extended action pattern recognition unit recognizes the identified new action pattern as an extended action pattern, and defines the model.

  An example of an action pattern newly recognized by the extended action pattern recognition unit by the procedure of FIG. 12 is given in FIG. The behavior pattern shown in FIG. 13 is an example in which the subject 2 who has stopped in a standing state has recognized the behavior of an intermediate process until lying down and standing still. From the series of acceleration measurement values shown in FIG. 13, a series of extended action pattern categories such as “sit down”, “change direction”, “shift hips”, and “depress state” were recognized.

  In FIG. 14, an example of the screen which the display apparatus 6 outputs the action which the extended action pattern recognition part of the action analysis apparatus of this Embodiment recognized is shown. On the right side of the image, the elapsed time and the action category recognized by the extended action pattern recognition unit are displayed. On the left side of the screen, an image of the behavior of the subject 2 who is currently wearing the acceleration sensor 1 can be displayed as necessary.

  As shown in FIG. 15, the extended action pattern recognition unit not only recognizes a new action pattern, but also detects a feature quantity of a motion and a feature quantity of a posture included in the new action pattern, and creates a new extension action It is possible to perform a process for describing the new behavior in more detail by quantitatively measuring the posture and movement related to the pattern. Here, the detection and separation of the motion feature value and the posture feature value can be performed by the same procedure as the procedure previously performed in step 3 by the basic action pattern definition unit. By repeating such recognition of the extended behavior pattern, a set of models representing independent classes for describing a new behavior pattern can be obtained. The number of classes is larger than the number of categories of the six basic action patterns defined at the beginning, and many new actions can be automatically recognized by repeating the processing procedure by the extended action pattern recognition unit. it can.

  As described above, the behavior analysis apparatus according to the present embodiment separates the posture model and the motion model in the basic behavior pattern model. For this reason, even when a new extended action pattern is recognized, posture recognition and motion recognition can be performed separately. Another aspect of behavior quantification based on the extended behavior pattern is that information regarding a position on the time axis obtained as a result of recognition can be used. That is, the duration, frequency, time interval between actions and actions, etc. can be obtained as quantitative information.

  With the behavior analysis device and behavior analysis program of the present invention, it becomes possible to automatically classify and analyze human behavior patterns. As a result, a system that monitors the actions of people who need support, context-aware services (situation recognition services) using mobile devices, and programs for improving work environments by recording and analyzing actions and actions performed in offices and factories It can be used for application design of operations at production sites and workflow analysis in service fields.

1 Acceleration sensor 2 Subject 3 Computer 4 CPU
5 Storage means 6 Display device

Claims (7)

Analysis means for analyzing the behavior of the human body using acceleration data output from a three-axis acceleration sensor attached to the human body, and storage means,
The analysis means is
Separating the feature quantity of the human body posture and the feature quantity of the movement of the human body included in the temporal change of the acceleration data continuously output and accumulated as time series data, and a predetermined basic action A basic action pattern definition unit that associates with each of the feature quantities, defines a model composed of the feature quantities of the basic action pattern, and stores the model in the storage unit;
Extended behavior pattern recognition unit that performs processing for separating acceleration data newly accumulated as time series data using the model of the basic behavior pattern, and recognizes the extended behavior pattern by clustering the processed acceleration data And a behavior analysis device.   The basic behavior pattern definition unit separates the feature amount of the human body posture and the feature amount of the movement of the human body based on a distribution of feature amounts included in the acceleration data accumulated as time series data. The behavior analysis apparatus according to claim 1.   The basic action pattern definition unit obtains respective feature amounts after separating acceleration data in a high frequency region as information on movement of the human body and acceleration data in a low frequency region as information on posture of the human body. The behavior analysis apparatus according to claim 1.   The extended behavior pattern recognition unit separates the feature quantity of the motion and the feature quantity of the posture included in the time-dependent change of the acceleration data, and sets the feature quantity of the motion and the feature quantity of the posture for each unit time. The behavior analysis apparatus according to claim 1, wherein the basic element model is created by decomposing the acceleration data by dividing the acceleration data.   The behavior analysis apparatus according to claim 4, wherein the extended behavior pattern recognition unit performs data processing that merges feature amounts of the acceleration data that has been processed to be divided.   A change in the acceleration data newly accumulated over time is collated with the model of the basic action pattern and the basic element model, and a feature quantity of motion and a feature quantity of posture representing an action pattern not defined in the basic action pattern The behavior analysis apparatus according to claim 4, wherein a new extended behavior pattern model is defined, and the new extended behavior pattern is recognized by performing a related quantitative measurement. A procedure for accumulating acceleration data representing any action continuously output from a three-axis acceleration sensor worn on a human body as time-series data;
A step of separating the feature quantity of the human body motion and the feature quantity of the human body posture included in the time-dependent change of the acceleration data;
A procedure for associating a predetermined basic action with each of the feature quantities and defining a model composed of the feature quantities of the basic action pattern;
A procedure for storing the model in the storage means;
Using the model of the basic behavior pattern, a procedure for performing processing for separating the acceleration data newly accumulated as time series data,
A behavior analysis program for causing a computer to execute a procedure for recognizing an extended behavior pattern by clustering the processed acceleration data.