JP2012248017A - Action model learning device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an action model learning device, method, and program for learning a highly accurate action model without requiring sensor data on a target user.SOLUTION: In a user similar model learning device 50, attribute information is calculated from physical characteristic information on each pair of source users by using learning data 30, a characteristic vector is extracted from sensor data with a label to calculate similarity, and a parameter of a user similar model 32 for each type of action is learned so as to output the similarity in response to an input of the attribute information. A physical characteristic information acquiring unit 12 of an action model learning device 10 acquires physical characteristic information on a target user, a similar user presuming unit 14 inputs attribute information calculated from physical characteristic information on a pair of the target user and source user to a user similar model 32 for each type of action to presume a similar source user having similar sensor data as that of the target user on the basis of the acquired similarity. An action model learning unit 16 learns an action model by using the sensor data on the similar source user presumed for each type of action.

Description

  The present invention relates to a behavior model learning device, method, and program, and in particular, a behavior model learning device and method for learning a behavior model used when estimating behavior based on data obtained from a sensor attached to the body. And the program.

  Currently, research on estimating daily behavior using a sensing device is actively conducted. For example, a method for recognizing 20 types of user actions such as “walking”, “watching TV”, and “brushing teeth” from a waveform measured by an acceleration sensor worn on the wrist or waist of the user has been proposed. (For example, refer nonpatent literature 1). Such studies often recognize behavior using a supervised machine learning approach. In other words, the sensor data of the user's behavior is labeled in advance, and the behavior model is learned using the labeled sensor data. Here, the label is information on the type of action and the start and end times of the action. In the method of Non-Patent Document 1, a user's behavior model is learned using sensor data with a label of the user's behavior, and high accuracy is achieved.

  However, it is difficult in terms of cost to prepare sensor data with a label for each user who performs action recognition (hereinafter referred to as “target user”). Therefore, many studies have been proposed to reduce the cost of preparing sensor data and the cost of labeling. For example, using MLLR (maximum likelihood linear regression) or MAP (maximum posterior probability) adaptation technology, learning is performed from sensor data with a label of a user (hereinafter referred to as “original user”) for acquiring sensor data for learning. The behavior model is adapted to the target user (see, for example, non-patent documents 2 to 4). In the technique of Non-Patent Document 2, the behavior model is adapted using the sensor data of the target user.

  There has also been proposed a gesture recognition system that uses a signal from the brain (sensor data of a target user) to find a recognition error and uses that information as labeled training data to relearn a gesture recognition model ( For example, refer nonpatent literature 5).

  Also, a method has been proposed that uses active learning to realize high-accuracy recognition even with a small amount of learning data (see Non-Patent Document 6, for example). In the method of Non-Patent Document 6, for a behavior that is considered difficult to recognize, the target user is requested to input a correct answer, and the model related to the target user is relearned based on the input correct answer.

  In addition, a method for learning a behavior model of a target user using sensor data of an original user with the same gender has been proposed (for example, see Non-Patent Document 7).

L. Bao and S. S. Intille, "Activity recognition from user-annotated acceleration data," Proc. PERVASIVE 2004, pp. 1--17, 2004. Naoyuki Hashida, Ren Omura, Ryota Imai: Examination for personal adaptation in daily behavior identification using accelerometer, No.3, pp.335-336, 2008. C. Leggetter and P. Woodland.Maximum likelihood linear regression for speaker adaptation of continuous density hidden Markov models.Computer Speech & Language, 9 (2): 171-185, 1995. J. Gauvain and C. Lee. Maximum a posteriori estimation for multivariate gaussian mixture observations of markov chains.IEEE Trans. On Speech and Audio Processing, 2 (2): 291--298, 2002. K. Forster, A. Biasiucci, R. Chavarriaga, J. del R. Millan, D. Roggen, and G. Troster.On the use of brain decoded signals for online user adaptive gesture recognition systems.In Pervasive 2010, pages 427- 444, 2010. M. Stikic, K. Van Laerhoven, and B. Schiele.Exploring semi-supervised and active learning for activity recognition.In Int’l Symp.on Wearable Computers, pages 81-88, 2008. G. Krassnig, D. Tantinger, C. Hofmann, T. Wittenberg, and M. Struck.User-friendly system for recognition of activities with an accelerometer. In PervasiveHealth 2010, pages 1-8, 2010.

  However, the technologies of Non-Patent Documents 2 to 5 require sensor data about the target user when learning the behavior model of the target user. In particular, there is a problem that it is difficult in terms of cost for the target user to prepare sensor data with a label. Further, in the technique of Non-Patent Document 6, since the target user has to input a correct answer, this also has a problem in terms of cost. Moreover, since the technique of nonpatent literature 7 judges the learning data regarding the original user used for learning of a target user's behavior model only by sex, the accuracy of the obtained target user's behavior model will be low. There's a problem.

  The present invention has been made in view of the above problems, and provides a behavior model learning apparatus, method, and program capable of learning a behavior model with high accuracy without requiring sensor data of a target user. For the purpose.

  In order to achieve the above object, the behavior model learning device of the present invention includes an acquisition unit that acquires physical characteristic information indicating a physical characteristic of a target user, and physical characteristic information of each of a plurality of former users different from the target user. And a user similarity model constructed based on sensor data representing the behavior of each of the plurality of former users, and learning the similarity of sensor data with respect to physical feature information between the former users for each type of behavior, Based on the body feature information of the target user acquired by the acquisition means and the body feature information of the plurality of former users prepared in advance, the former user whose sensor data is similar to the target user is determined as the type of action. Using the estimation unit for each estimation and the sensor data of the original user estimated for each type of action by the estimation unit, There upon being acquired, and is configured to include a behavior model learning means for learning a behavior model for recognizing the type of action that the sensor data represents, a.

  According to the behavior model learning device of the present invention, the acquisition unit acquires the body feature information indicating the physical feature of the target user. Further, it is constructed based on the body feature information of each of the plurality of former users different from the target user and the sensor data representing the behavior of each of the plurality of former users, and the similarity of the sensor data to the body feature information between the former users is determined. A user similarity model learned for each type of action is prepared in advance. Then, the estimation means uses the target similarity model, the body characteristic information of the target user acquired by the acquisition means, the body feature information of a plurality of former users prepared in advance, and the body feature information of the target user. An original user whose sensor data is similar to the user is estimated for each type of action. Then, when the behavior model learning unit acquires the sensor data of the target user using the sensor data of the original user estimated for each type of behavior by the estimation unit, the type of behavior represented by the sensor data of the target user Learning behavioral models for recognizing

  In this way, based on the user similarity model in which the sensor data similarity to the body feature information between the original users is learned, the behavior model is learned using the sensor data of the original user estimated that the sensor data is similar to the target user. Therefore, a highly accurate behavior model can be learned without requiring sensor data of the target user.

  In addition, the user similarity model generates a plurality of pairs of original users by using learning data composed of sensor data in which the body feature information of the former user and the type of action are associated with each other. When attribute information is calculated using the body feature information of each included original user, the similarity of sensor data for each type of behavior of each original user included in the pair is calculated, and when the attribute information is input It is constructed by learning the calculated attribute information and the similarity so as to output the similarity, and the estimation means is prepared in advance with the body feature information of the target user acquired by the acquisition means. Before the attribute information is calculated based on the body feature information of the plurality of former users, and the similarity obtained by inputting the calculated attribute information to the user similarity model is equal to or greater than a predetermined value. The source user, the target user and the sensor data can be estimated as a source user similar. The former user whose similarity is equal to or greater than a predetermined value may be an original user whose predetermined value is equal to or greater than a predetermined fixed value, or the top nth value of similarity is set as a predetermined value, In other words, it may be an original user with the highest n-th similarity.

  The behavior model learning device of the present invention can further include a user similarity model learning means for learning the user similarity model.

  The behavior model learning method of the present invention is a behavior model learning method in a behavior model learning device including an acquisition unit, an estimation unit, and a behavior model learning unit, wherein the acquisition unit includes physical characteristics of a target user. And the estimation means is constructed based on the body feature information of each of a plurality of former users different from the target user, and sensor data representing the behavior of each of the plurality of former users, and The similarity of sensor data with respect to body feature information between the former users, a user similarity model determined for each type of action, the body feature information of the target user acquired by the acquisition means, and the plurality of prepared in advance Based on the body feature information of the former user, the former user whose sensor data is similar to the target user is estimated for each kind of behavior, and the behavior model The learning means recognizes the type of action represented by the sensor data when the sensor data of the target user is acquired using the sensor data of the original user estimated for each kind of action by the estimating means. It is a method of learning an action model for the purpose.

  Further, the behavior model learning device of the present invention can be a behavior model learning method in the behavior model learning device configured to further include a user similarity model learning unit, and the user similarity model learning unit includes the user similarity model. You can learn the model.

  The behavior model learning program of the present invention is a program for causing a computer to function as each means constituting the behavior model learning device.

  As described above, according to the behavior model learning device, method, and program of the present invention, the sensor data is obtained from the target user based on the user similarity model in which the similarity of the sensor data to the body feature information between the original users is learned. Since the behavioral model is learned using the sensor data of the former user estimated to be similar to the sensor data, it is possible to learn the behavioral model with high accuracy without requiring the sensor data of the target user. .

It is a block diagram which shows the functional structure of the action model learning apparatus and user similar model learning apparatus of this Embodiment. It is a figure which shows an example of the attribute information calculated from the body feature information and sensor data of a pair of former users, and a similarity. It is a flowchart which shows the content of the user similarity model learning process routine in the user similarity model learning apparatus of this Embodiment. It is a flowchart which shows the content of the action model learning process routine in the action model learning apparatus of this Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  The behavior model learning apparatus 10 according to the present embodiment includes a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory) that stores a program for executing a behavior model learning processing routine described later. ). As shown in FIG. 1, this computer can be functionally represented by a configuration including a body feature information acquisition unit 12, a similar user estimation unit 14, a behavior model learning unit 16, and a storage unit 20. The storage unit 20 stores learning data 30, a user similarity model 32, and a behavior model 34.

  First, the construction of the user similarity model 32 used in the similar user estimation unit 14 will be described here. The user similarity model 32 is constructed by the user similarity model learning device 50.

  The user similar model learning device 50 is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a user similar model learning processing routine described later. As shown in FIG. 1, this computer functionally includes a user pair generation unit 52, an attribute information calculation unit 54, a similarity calculation unit 56, a user similarity model learning unit 58, and a storage unit 20. Can be represented. The storage unit 20 stores learning data 30 and a user similarity model 32.

  The learning data 30 stored in the storage unit 20 includes sensor data representing behavior acquired from an acceleration sensor worn by the former user and body feature information indicating the physical characteristics of the former user. The storage unit 20 stores learning data regarding a plurality of former users. A label is assigned to the sensor data in advance. Here, the “label” means a label indicating the type of action represented by the sensor data (for example, “walking action” or “running action”), and the corresponding action type and the start and end of the action Contains time information.

  The user pair generation unit 52 generates a plurality of pairs of original users by selecting two original users from the set of original users included in the learning data 30.

  The attribute information calculation unit 54 calculates attribute information for each pair of original users generated by the user pair generation unit 52 from the body feature information of the original users in the pair.

  The body feature information includes basic information such as height, weight, and gender, and information related to an action to be recognized. For example, in the case of an action of brushing teeth, information on the dominant hand of the action can be included. Values that can be taken by physical feature information include numerical information and class information. For example, height and weight are numerical information, and sex information is class information (male or female class).

  Specifically, in the calculation of the attribute information, when the original user in a certain original user pair is the user A and the user B, first, the user A is defined as the base user and the user B is defined as the object user. From the numerical body feature information (for example, age), the base user value, the object user value, the difference between the object user value and the base user value, and the ratio between the object user value and the base user value Are calculated. FIG. 2 shows an example of the calculated attribute. The first line is an example of calculating the attributes of users A and B. In this example, the ages of users A and B are 29 and 32, respectively. Therefore, the difference is 3, and the ratio is 1.1. From the class physical feature information (e.g., gender, etc.), three attributes are calculated: the base user value, the object user value, and the difference between the object user value and the base user value. Here, the difference is 1 when the values of the classes are different, and 0 when the values are the same. In the example of the first line in FIG. 2, the sexes of the users A and B are both men. Therefore, the difference is zero. In this way, the attribute information calculation unit 54 calculates attribute information from all body feature information for all pairs of original users generated by the user pair generation unit 52.

  The similarity calculation unit 56 calculates the similarity of the behavior of the two former users in each pair for each type of behavior. The similarity between the behaviors of the two former users is calculated from sensor data with a label related to the former users included in the learning data 30.

  Specifically, first, feature extraction is performed from time-series sensor data. Feature extraction is basically performed using each time window (window). For example, a sliding window is set every second, and features are extracted from the sensor data in the window for each window. The sensor data used here is data measured by an acceleration sensor and has characteristics in the magnitude and periodic change of the value. To capture the magnitude of the value, the average of the values in the sensor data window can be used as a feature as it is. The feature indicating the periodic change of the sensor data can be extracted by performing a Fourier analysis on the data series in the window. For example, fast Fourier transform can be applied to a data series, and each Fourier coefficient can be used as a feature as it is. For example, as in the technique of Non-Patent Document 1, energy and entropy can be calculated from Fourier coefficients and used as features. Such feature extraction processing is performed for each axis of the acceleration sensor attached to the body.

  The similarity is calculated using the feature set extracted as described above (can be treated as a feature vector expressed by a vector). A feature vector for a time window consists of features extracted from the data in that window for all axes of all sensors. A series of feature vectors can be obtained by performing this process by sliding the time window.

  Next, a method of calculating the similarity for each type of action from the feature vector group calculated from sensor data with labels of certain users A and B will be described. At this time, user A is a base user and B is an object user. A mixed Gaussian model (GMM) can be used to calculate the similarity. GMM is learned using a feature vector group extracted from sensor data corresponding to a certain type of action of user A. And the likelihood with respect to GMM of the feature vector extracted from the sensor data corresponding to the kind of the action of the user B is calculated | required. Let the average of the likelihood of all the feature vectors corresponding to the action of the user B be the similarity of the user B to the user A for the action. Non-Patent Document 8 (A. Dempster, N. Laird, D. Rubin, et al. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society. Series B (Methodological), 39 ( 1): 1-38, 1977.) can be used. Note that feature extraction and similarity calculation are not limited to the above method, and a conventionally known method can be used.

  Thereby, for example, as shown in FIG. 2, attribute information and similarity are obtained for each pair of original users generated by the user pair generation unit 52. In FIG. 2, “base” means “base user”, and “obj” means “object user”.

  The user similarity model learning unit 58 learns the user similarity model 32 using the attribute information and the degree of similarity obtained for each pair of original users. The user similarity model 32 learns for each type of action. That is, when the attribute information calculated for the original user pair is input for each type of action (for each label of the sensor data with label), the similarity calculated for the original user pair is output. Thus, the parameters of the user similarity model 32 are learned. For the learning, a model capable of outputting a numerical value such as a regression model or a neural network can be used. The constructed user similarity model 32 is stored in the storage unit 20.

  Next, each part of the behavior model learning device 10 will be described in detail.

  The body feature information acquisition unit 12 acquires the body feature information of the target user input by an input device (not shown).

  The similar user estimation unit 14 uses the user similarity model 32 learned by the user similarity model learning device 50 and uses the physical information feature of the target user acquired by the physical feature information acquisition unit 12 to perform sensor for each type of action. Estimate the original user whose data is estimated to be similar.

  Specifically, a plurality of pairs of the target user and the original user are generated from the target user and each original user included in the learning data 30. And about each pair, attribute information is calculated by the same process as the said attribute information calculation part 54 using the body feature information of the target user and former user in each pair. Then, the calculated attribute information for each pair is input to the user similarity model 32, and the similarity is acquired. At this time, among the similarities acquired for each pair, the original user in the pair that is the upper predetermined number (n) is extracted as a similar original user similar to the target user. In addition, you may extract the former user in the pair whose similarity becomes more than a predetermined value as a similar former user. Further, since the user similarity model 32 is created for each type of action, a similar source user is extracted for each type of action.

  The behavior model learning unit 16 acquires sensor data with labels of n similar source users for each type of behavior estimated by the similar user estimation unit 14 from the learning data 30 stored in the storage unit 20, and features from the sensor data. A vector series is extracted, and the behavior model 34 is learned for each type of action using the extracted feature vector series. The learning of the behavior model 34 uses a well-known learning method of the behavior model using sensor data with a label. For example, a hidden Markov model used for modeling time-series data can be used (for example, Non-Patent Document 9 “L. Rabiner. A tutorial on hiddenMarkov models and selected applications in speech recognition. Proceedings of the IEEE, 77 ( 2): 257-286, 1989 ").

  Next, a user similarity model learning process routine in the user similarity model learning device 50 will be described with reference to FIG. This routine starts in a state in which learning data 30 relating to a plurality of former users is stored in the storage unit 20.

  In step 100, by selecting two former users from the set of former users included in the learning data 30, a plurality of pairs of former users are generated.

  Next, in step 102, for each pair of original users generated in step 100, for example, attribute information as shown in FIG. 2 is calculated from the body feature information of the original users in the pair.

  Next, in step 104, a feature vector is extracted from the labeled sensor data regarding the original user included in the learning data 30 of the two original users in each pair generated in step 102, and the extracted feature vector is used. The similarity of the behavior of two former users is calculated for each type of behavior.

  Next, in step 106, the attribute calculated for the pair of the original user for each type of action using the attribute information for each pair calculated in step 102 and the similarity for each pair calculated in step 104. When the information is input, the parameters of the user similarity model 32 are learned so that the similarity calculated for the original user pair can be output. The learned user similarity model 32 is stored in the storage unit 20, and the process is terminated.

  Next, a behavior model learning process routine in the behavior model learning device 10 will be described with reference to FIG. This routine starts in a state where the storage unit 20 stores learning data 30 regarding a plurality of former users and a user similarity model 32 learned by the user similarity model learning device 50.

  In step 150, the inputted body feature information of the target user is acquired. Next, in step 152, a plurality of pairs of the target user and the original user are generated from the target user and each of the original users included in the learning data 30.

  Next, in step 154, for each pair generated in step 152, the attributes similar to those in step 102 of the user similarity model learning process are used, using the physical feature information of the target user and the original user in each pair. Calculate information.

  Next, in step 156, the user similarity model 32 stored in the storage unit 20 is called, and the attribute information for each pair calculated in step 154 is input to the user similarity model 32, and the similarity is acquired. Then, among the similarities acquired for each pair, the original user in the top n pairs is extracted as a similar original user similar to the target user. This process is performed on the user similarity model 32 generated for each type of action, and a similar source user is estimated for each type of action.

  Next, in step 158, sensor data with labels of n similar-source users for each type of action estimated in step 156 is acquired from the learning data 30 stored in the storage unit 20, and the feature vector series is obtained from the sensor data. And the behavior model 34 is learned for each type of behavior using the extracted feature vector series. The learned behavior model 34 is stored in the storage unit 20, and the process is terminated.

  As described above, according to the behavior model learning device of the present embodiment, based on the idea that users with similar physical characteristics will also have similar sensor data. For the input of attribute information calculated from the body feature information of the pair of users, a user similarity model that outputs the similarity of the behavior of the original user in the pair is used. Based on the similarity obtained by inputting attribute information calculated from the body feature information of the pair of the target user and the original user to the user similarity model, the similarity source in which the target user and the action (sensor data) are similar A user is estimated, and a behavior model for each type of behavior is learned using the estimated sensor data with label of the similar source user. For this reason, the target user can learn a highly accurate behavior model only by inputting body feature information and without requiring sensor data of the target user.

  In the above embodiment, the case where the behavior model learning device 10 and the user similarity model learning device 50 are configured as separate computers with the storage unit 20 as a common configuration has been described. You may make it do.

  The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  Moreover, although the above-described behavior model learning device and user-similar model learning device have a computer system inside, if the “computer system” uses a WWW system, a homepage providing environment (or Display environment).

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 Behavior model learning apparatus 12 Body feature information acquisition part 14 Similar user estimation part 16 Behavior model learning part 20 Storage part 30 Learning data 32 User similarity model 34 Action model 50 User similarity model learning apparatus 52 User pair production | generation part 54 Attribute information calculation part 56 Similarity Calculation Unit 58 User Similarity Model Learning Unit

Claims (6)

Acquisition means for acquiring physical feature information indicating physical characteristics of the target user;
The degree of similarity of sensor data with respect to body feature information of each of a plurality of former users different from the target user, and sensor data representing the behavior of each of the plurality of former users. Based on the user similarity model learned for each type of action, the body feature information of the target user acquired by the acquisition means, and the body feature information of the plurality of former users prepared in advance. Estimating means for estimating the former user whose sensor data is similar to that of the user for each type of action,
An action model for recognizing the kind of action represented by the sensor data when the sensor data of the target user is acquired using the sensor data of the original user estimated for each kind of action by the estimating means An action model learning means for learning
A behavior model learning apparatus including: The user similarity model generates a plurality of pairs of original users using learning data composed of sensor data in which the body feature information of the former user and the type of action are associated, and is included in the pair Attribute information is calculated using body feature information of each original user, the similarity of sensor data for each type of behavior of each original user included in the pair is calculated, and the similarity is obtained when the attribute information is input. Constructed by learning the calculated attribute information and the similarity so as to output the degree,
The estimation unit calculates the attribute information based on the body feature information of the target user acquired by the acquisition unit and the body feature information of the plurality of former users prepared in advance, and calculates the calculated attribute information. The behavior model learning device according to claim 1, wherein the former user whose similarity obtained by inputting to the user similarity model is equal to or greater than a predetermined value is estimated as an former user whose sensor data is similar to the target user.   The behavior model learning device according to claim 1, further comprising user similarity model learning means for learning the user similarity model. An action model learning method in an action model learning device including an acquisition means, an estimation means, and an action model learning means,
The acquisition means acquires physical feature information indicating the physical features of the target user,
The estimation means is constructed based on body feature information of each of a plurality of former users different from the target user, and sensor data representing each of the plurality of former users, and for body feature information between the former users The similarity degree of sensor data is learned for each type of action, the user similarity model, the body feature information of the target user acquired by the acquisition means, and the body feature information of the plurality of former users prepared in advance On the basis of the original user whose sensor data is similar to the target user for each type of action,
The behavior model learning unit uses the sensor data of the original user estimated for each type of behavior by the estimation unit, and when the sensor data of the target user is acquired, the type of behavior represented by the sensor data A behavior model learning method for learning a behavior model for recognizing an object. The behavior model learning device further includes a user similarity model learning means,
The behavior model learning method according to claim 4, wherein the user similarity model learning unit learns the user similarity model.   The behavior model learning program for functioning a computer as each means which comprises the behavior model learning apparatus of any one of Claims 1-3.

Images