JP2007187555A - Motion recognition system and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To properly recognize motion in an environment based on information from a plurality of acceleration sensors. <P>SOLUTION: Actual observation characteristic data are acquired at each prescribed time width (for example, one second) from acceleration data from sensor nodes 1-1 to 1-4, and a label of a partial operation of sample observation having high similarity is applied to the actual observation characteristic data. Every time when the label of the partial operation having high similarity is applied to the actual observation characteristic data, matching is performed between a label row applied to the actual observation characteristic data until then and a label row of the partial operation in an operation dictionary 24A, and thereby a sample operation having high likelihood in the operation dictionary 24A is recognized as a movement in the environment at that time. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a motion recognition system and a motion recognition method for recognizing motion in an environment in which the acceleration sensors are installed based on information from a plurality of acceleration sensors.

  Environmental conditions, both indoors and outdoors, change from moment to moment depending on the people present or other factors. The environmental status (whole environment or partial status change) can be grasped by distributing a plurality of sensors in the environment and collecting information from these sensors via a network. The sensor here refers to a device that detects a physical quantity or signal in the vicinity.

  If this situation change can be grasped, it becomes possible to search for a similar situation and predict a situation with a high probability of occurring next based on a comparison with a past situation. In addition, as a part of the situation, if the user can predict the action to be performed by this prediction, the user's danger situation is detected and a warning is issued to avoid the situation, or the action performed by the user Can be implemented by the system instead. In fact, automatic doors and HVAC systems (air-conditioning and air-conditioning systems) can be thought of as a simple alternative to the latter, using a combination of sensors and actuators.

  In addition, many research results on home appliance networks and sensor networks have been announced in relation to ubiquitous computing, and applications for crime prevention, medical sites, and in-vehicle devices have already begun. These achievements are largely due to improved computer performance, the emergence of new network standards, and the miniaturization of batteries, power supplies, and sensor devices.

  As a result, conventional methods for recognizing situation changes that occur in the environment are only methods that rely heavily on image recognition using cameras fixed to walls, ceilings, etc., and on installation conditions that rely on sensors with fixed observation targets. In addition, it has become possible to install a large number of sensors in the environment and collect information collected there to perform situation recognition from the bottom up (see, for example, Non-Patent Document 1).

  For example, in the “Smart Dust project” shown in Non-Patent Document 2, an ultra-small sensor node is mixed with paint and applied to a wall to realize surrounding situation recognition. In the "Smart-ITS Project", a chemical management system that issues warnings to prevent chemicals that cause severe reactions from being placed next to each other due to the proximity relationship between sensor nodes is being constructed.

Y. Nakauchi, K. Noguchi, P. Somwong, T. Matsubara, A. Namamame: Vivid Room: Human intention Detection and Activity Support Environment for Ubiquitous Autonomy, Proc. Of the 2003 IEEE / RSJ International Conference on Intelligent Robots and Systems ( IROS2003), pp.773-778, 2003. J.M.Kahn, R.H.Katz, K.S.J.Pister, Next century challenges: Mobile Networking for 'smart dust', Proc.MOBICOM, 1999,271-278. Needleman, S. B. and Wunsch, C. D. "A general method applicable to the search for similarities in the amino acid sequence of two proteins," Journal of Molecular Biology, vol.48, pp.443-453, 1970. Tsunenori Ishioka: Applied statistics, Vol.29, No.3, 141-149 (2000) on the extension of the k-means algorithm that automatically determines the number of clusters. Satoshi Hiramatsu, Masami Hattori, Atsumi Yamada, Tsuyoshi Okadome: Pervasive Association: Semantic Integration in Ubiquitous Environment (2) -Collecting Real World Relation Information Using Sensor Networks -Multimedia, Distributed, Collaboration and Mobile (DICOMO2005) Symposium Paper Pp.169-172, 2005. Furui Sadaaki, Digital Speech Processing, 8.5.1 DP Matching, pp. 162-164, Tokai University Press, 1985.

  Among conventional methods, such as automatic doors and HVAC systems, in which the operation is determined by directly specifying the threshold value for the data from the sensor, the data output by the sensor changes when the environment in which the system is installed changes. There is a possibility that the meaning changes and it may become impossible to correctly recognize the “situation where the door should be opened”. This problem has not been solved by “Smart Dust” and “Smart-ITS”, and it is desired to realize a system that can appropriately recognize the movement in the environment where the system is installed.

  The present invention has been made to solve such a problem, and the object of the present invention is to provide a motion recognition system capable of appropriately recognizing a motion in the environment based on information from a plurality of sensors. And providing a motion recognition method.

  In order to achieve such an object, the present invention provides a motion recognition system including a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from the acceleration sensors. A plurality of standard changes prepared in advance by cutting out data at predetermined time intervals from information from the acceleration sensor observed when the motion recognition device generates a sample motion as a motion candidate to be recognized in the environment. A feature vector whose elements are the similarity to each standard change pattern is obtained by matching with the pattern, and a vector obtained by synthesizing the feature vector at the same time of each acceleration sensor is used as the feature data of sample observation. Partial motion extraction means for extracting partial motions included in the sample motion from the sample observation feature data string, and from the sample motion A label corresponding to the partial motion is attached to the feature data of the sample observation belonging to the issued partial motion, and the label string given to the feature data of the sample observation belonging to the partial motion, the feature of the sample observation belonging to the partial motion Data representing data, motion dictionary creating means for creating a motion dictionary in which partial motions included in the sample motion and the order relation between the partial motions are written for each sample motion, and an acceleration sensor actually observed in the environment The data is extracted from the information from each of the same time intervals as the sample observation, and is matched with a plurality of standard change patterns prepared in advance, and the feature quantity is based on the similarity to each standard change pattern Find the vector and combine the feature vectors at the same time of each acceleration sensor into the actual observation feature data Then, the similarity between the feature data of the actual observation and the data representing the feature data of the sample observation belonging to the partial motion in the motion dictionary is obtained, and the label attached to the representative data having a high similarity is used as the feature data of the actual observation. Each time a label is assigned to the actual observation feature data, and between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary Motion recognition means is provided for performing matching and recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time. Note that the present invention can be realized not as a motion recognition system but also as a motion recognition method.

  According to the present invention, when a sample motion is generated as a motion candidate to be recognized in the environment, data is cut out for each predetermined time width from information from an acceleration sensor installed in the environment, and a plurality of pre-prepared A feature vector whose elements are the degree of similarity with each standard change pattern is obtained by matching with the standard change pattern, and the vector obtained by combining the feature vectors at the same time of each acceleration sensor is the feature of the sample observation Data. For example, a feature vector is obtained from the sensor data of the acceleration sensor every second, and a vector obtained by synthesizing the feature vector of each acceleration sensor is used as the feature data of the sample observation at that time. For an acceleration sensor that outputs triaxial acceleration data of X, Y, Z, etc., by cutting out the data of 1 second from the triaxial acceleration data so as to overlap each other by 0.5 seconds, Let it be sensor data.

  Then, the partial motion included in the sample motion is extracted from the feature data of each sample observation, and the label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, Label sequence attached to feature data of sample observation belonging to partial motion, data representative of feature data of sample observation belonging to that partial motion (standard motion feature vector), and partial motion included in the sample motion for each sample motion The order relation between the partial actions is described in the action dictionary.

For example, if the sample action is “sitting in a chair”, “moving” and “sitting” are extracted as partial actions, and the natural language label “moving 1” is added to the feature data of the sample observation belonging to “moving”. The natural language label “seat 1” is assigned to the feature data of the sample observation belonging to “sitting”, and the label column (“moving 1” label row, “sitting” is assigned to the feature data of this sample observation. 1 ”) is described in the action dictionary.
When the sample action is “seating from the chair”, “seating” and “moving” are extracted as partial actions, and the characteristic data of the sample observation belonging to “seating” is the natural language “seating 1”. A label is assigned, and a natural language label “Move 2” is assigned to the feature data of the sample observation belonging to “Move”, and a label row (label row “Seating 1” is assigned to the feature data of this sample observation. , “Movement 2” label sequence) is described in the action dictionary.
Then, “moving 1” and “sitting 1” are described in the motion dictionary in the order of “moving 1” and “sitting 1” as partial motions included in the sample motion corresponding to “sitting in a chair”. In addition, as the partial motions included in the sample motion corresponding to “seating from the chair”, “seating 1” and “moving 2” are described in the motion dictionary in the order of “seating 1” and “moving 2”. .

  In the present invention, a new sample operation can be defined by combining partial operations in the operation dictionary without actually generating a sample operation in the environment. For example, “movement 1”, “seating 1”, “seating 1”, and “movement 2” are partial movements, and “movement 1”, “seating 1”, “seating 1”, and “movement 2” are performed in this order. By describing it in the action dictionary, it is possible to define “use of a chair” as a sample action that combines “sitting in a chair” and “seating from a chair” that is not actually generated in the environment. .

  In this invention, at the time of actual observation, data is cut out from the information from the acceleration sensor installed in the environment at the same predetermined time width as at the time of observation of the sample operation, and a plurality of standard change patterns prepared in advance are used. The feature amount vector having the similarity with each standard change pattern as an element is obtained by performing matching between the two, and a vector obtained by synthesizing the feature amount vectors at the same time of each acceleration sensor is used as feature data of actual observation. When the actual observation feature data is obtained, the similarity between the actual observation feature data and the data representative of the sample observation feature data belonging to the partial motion in the motion dictionary (standard motion feature vector) is obtained. The label attached to the representative data having a high value is attached to the feature data of the actual observation. For example, if the data assigned with the label “Move 1” as a partial operation is similar to the actual observation feature data, the “Move 1” label is assigned to the actual observation feature data.

  When a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary. A sample operation with a high likelihood is recognized as a movement in the current environment. For example, the feature data of actual observation includes a label column of “Move 1” and a label column of “Seat 1”, and the label column of “Move 1” and the label column of “Seat 1” are “Move 1” in the motion dictionary. ”And the“ Seating 1 ”label sequence, and the order is“ Move 1 ”,“ Sitting 1 ”, then“ Sitting on a chair ”is extracted as a sample operation with high likelihood. , Recognized as movement in the environment at that time. Accordingly, motion recognition with a higher abstraction level can be performed for each predetermined time width. At this time, if sufficient sample motions are recorded in advance and a motion dictionary in which the partial motions are thoroughly examined is prepared, motion recognition with higher accuracy becomes possible.

In the present invention, matching is performed between the label sequence given to the actual observation feature data and the label sequence of the partial motion in the motion dictionary, and the sample operation with a high likelihood in the motion dictionary is changed to the environment at that time. However, the sample operation with the highest likelihood may be recognized as the motion in the environment at that time, or the top N sample operations may be recognized as the motion in the environment at that time. .
For example, there is a label column of “movement 1”, “seating 1”, “seating 1”, “movement 2” in the feature data of actual observation, and this “movement 1”, “seating 1”, “seating 1”. , “Move 2” label sequence matches “Move 1”, “Seat 1”, “Leave 1”, “Move 2” in the motion dictionary, and the order is “Move 1”, “Seat If “1”, “seating 1”, and “moving 2” are in order, “usage of chair”, “sitting in chair”, “seating from chair” are given as the descending order of likelihood. Recognize as movement.

  In the present invention, matching between the cut-out data and a plurality of standard change patterns prepared in advance may be performed by DP matching as shown in Non-Patent Document 6. Also, the matching between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary is symbolic DP matching (by dynamic programming) as shown in Non-Patent Document 3. (Character string similarity calculation). In the present invention, the extraction of the partial motion included in the sample motion may be performed by clustering as shown in Non-Patent Document 4.

  Also, it is obtained by extracting data at predetermined time intervals from information from an acceleration sensor observed when a sample motion is generated as a motion candidate to be recognized in the environment, and performing fast Fourier transform on the extracted data A feature vector whose elements are frequency components is obtained, and a vector obtained by synthesizing the feature vector at the same time of each acceleration sensor is used as sample observation feature data, and the partial operation included in the sample operation from this sample observation feature data column A partial motion extracting means for extracting the sample motion, and a label corresponding to the partial motion is attached to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the feature data of the sample observation belonging to the partial motion is attached. Label string, data representative of sample observation feature data belonging to the partial motion, and support A motion dictionary creating means for creating a motion dictionary in which partial motions included in the sample motion and an order relation between the partial motions are written for each pull motion, and sample observation from information from an acceleration sensor actually observed in the environment A vector obtained by cutting out data at the same predetermined time width, obtaining a feature vector whose elements are frequency components obtained by performing fast Fourier transform of the cut-out data, and synthesizing feature vectors at the same time of each acceleration sensor Is the actual observation feature data, and the similarity between the actual observation feature data and the data representative of the sample observation feature data belonging to the partial motion in the motion dictionary is obtained, and the label is given to the representative data with high similarity Each time a label is attached to the actual observation feature data, Matching is performed between the label sequence given to the actual observation feature data and the partial sequence label in the motion dictionary, and the sample motion with high likelihood in the motion dictionary is used as the motion in the environment at that time. You may make it provide the motion recognition means to recognize.

  Also, it is obtained by extracting data every predetermined time width from information from an acceleration sensor observed when a sample motion is generated as a motion candidate to be recognized in the environment, and performing weblet conversion of the extracted data A feature vector whose elements are weblet coefficients is obtained, and a vector obtained by combining feature vectors at the same time of each acceleration sensor is used as feature data for sample observation. A partial motion extraction means for extracting a motion, and a label corresponding to the partial motion is attached to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and is attached to the feature data of the sample observation belonging to the partial motion. Data representing the characteristic data of sample observations belonging to the And a motion dictionary creation means for creating a motion dictionary in which the partial motions included in the sample motion and the order relation between the partial motions are written for each sample motion, and sample observation from information from an acceleration sensor actually observed in the environment Cut out the data at the same predetermined time width as, and obtain a feature vector whose elements are weblet coefficients obtained by performing weblet conversion of the cut out data, and synthesize the feature vector at the same time of each acceleration sensor The obtained vector is used as the actual observation feature data, and the similarity between the actual observation feature data and the data representative of the sample observation feature data belonging to the partial motion in the motion dictionary is obtained. Labeling means for assigning existing labels to actual observation feature data, and labels to actual observation feature data Each time, a matching is performed between the label sequence given to the actual observation feature data and the label sequence of the partial motion in the motion dictionary, and the sample operation with high likelihood in the motion dictionary is You may make it provide the movement recognition means recognized as a movement in an environment.

  In addition, data is extracted every predetermined time width from information from the acceleration sensor observed when a sample motion is generated as a motion candidate to be recognized in the environment, and an autoregressive model of the extracted data is estimated. A feature vector that has an autoregressive coefficient as an element is obtained, and a vector obtained by synthesizing feature vectors at the same time of each acceleration sensor is used as sample observation feature data, and is included in the sample operation from this sample observation feature data column A partial motion extracting means for extracting a partial motion, and a label corresponding to the partial motion is attached to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and is attached to the feature data of the sample observation belonging to the partial motion. Labeled data, data representative of sample observation feature data belonging to the partial motion, and For each sample motion, a motion dictionary creating means for creating a motion dictionary in which partial motions included in the sample motion and the order relation between the partial motions are written, and sample observation from information from an acceleration sensor actually observed in the environment Data is cut out at the same predetermined time width, and a feature vector whose elements are autoregressive coefficients obtained by estimating the autoregressive model of the cut out data is obtained, and the feature vector at the same time of each acceleration sensor is synthesized. The obtained vector is used as the actual observation feature data, and the similarity between the actual observation feature data and the data representative of the sample observation feature data belonging to the partial motion in the motion dictionary is obtained. Labeling means for assigning a label to the actual observation feature data, and each time a label is attached to the actual observation feature data , Matching is performed between the label sequence given to the actual observation feature data and the partial sequence label sequence in the motion dictionary, and the sample motion with high likelihood in the motion dictionary is moved in the environment at that time. Motion recognition means for recognizing as follows may be provided.

  According to the present invention, actual observation feature data is obtained for each predetermined time width from information from a plurality of acceleration sensors installed in the environment, and sample observation partial operations having high similarity to the actual observation feature data Each time a label is assigned, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, and the likelihood in the motion dictionary is high. The sample operation is recognized as the movement in the environment at that time, and the movement in the environment can be appropriately recognized based on information from the plurality of acceleration sensors.

Hereinafter, the present invention will be described in detail with reference to the drawings.
FIG. 1 is a diagram showing an outline of an embodiment of a motion recognition system according to the present invention. In the figure, reference numerals 1 (1-1 to 1-4) denote sensor nodes installed indoors. The sensor node 1-1 is on the door D, the sensor node 1-2 is on the chair A, and the sensor node 1-3. Is attached to the desk B, and the sensor node 1-4 is attached to the chair C.

  In this example, the sensor node in Non-Patent Document 5 is used as the sensor node 1. The sensor node 1 handled here outputs three-axis (XYZ direction) acceleration data, temperature, humidity, illuminance, presence / absence of human proximity, and the like.

  2A, 2B, 2C, and 3D, triaxial acceleration data S1, S2, S3 output from the sensor nodes 1-1, 1-2, 1-3, and 1-4. S4 is shown. In the present embodiment, the 3-axis acceleration data S1, S2, S3, and S4 output from the sensor nodes 1-1, 1-2, 1-3, and 1-4 are transferred to the motion recognition device 2 via the network 3. Based on the acceleration data S1, S2, S3, S4 from the sensor nodes 1-1, 1-2, 1-3, 1-4. In this case, the sensor nodes 1-1, 1-2, 1-3, and 1-4 correspond to the acceleration sensor referred to in the present invention.

  The motion recognition device 2 is realized by hardware including a processor and a storage device, and a program that realizes various functions in cooperation with these hardware, and has a motion recognition function as a function unique to the present embodiment. ing. Hereinafter, the motion recognition function of the motion recognition apparatus 2 will be described with reference to the functional block diagram shown in FIG.

[Create a motion dictionary]
First, a sample operation is generated in the room as a motion candidate to be recognized. For example, as a sample operation, “sitting in a chair” and “seating from a chair” are generated. Here, as a simple example, a case where “seating on chair A” and “seating from chair A” are individually generated will be described as an example.

  When “sitting in chair A” and “seating from chair A” are generated in the room, acceleration data S2 from sensor node 1-2 changes. FIG. 4A shows a change in the acceleration data S2 from the sensor node 1-2 when “sitting in the chair A” is generated. FIG. 4B shows a change in the acceleration data S2 from the sensor node 1-2 when “seating from the chair A” is generated.

  The change α of the acceleration data S2 when seated on the chair A is the change α1 when the chair A is moved to secure a space for the person to sit, and when the person sits down on the seat Change α2. The change β of the acceleration data S2 when the person leaves the chair A includes a change β1 when the person stands up from the seat and a change β2 when the person moves the chair A to return to the original position. It consists of.

  The acceleration data S1 to S4 (information from a plurality of acceleration sensors) from the sensor nodes 1-1 to 1-4 when the sample operation is generated are given to the information input unit 2A of the motion recognition device 2. The information input unit 2A includes acceleration feature quantity detection modules 21-1 to 21-4. The acceleration feature quantity detection modules 21-1 to 21-4 receive acceleration data (from sensor nodes 1-1 to 1-4). Triaxial acceleration data) S1 to S4 are given.

  The acceleration feature quantity detection module 21 (21-1 to 21-4) has a width of 1 second from the triaxial acceleration data S (S1 to S4) sent from the sensor node 1 (1-1 to 1-4). Is cut out so as to overlap each other by 0.5 seconds (see FIG. 5), and used as sensor data from the sensor node 1 (1-1 to 1-4). Then, data is cut out from the sensor data every second, and matching is performed with 11 standard change patterns (patterns P1 to P11 shown in FIG. 6) prepared in advance, and similar to each standard change pattern. The degree is calculated, a feature quantity vector having the similarity to each standard change pattern as an element is obtained and sent to the motion recognition unit 2B.

  In this example, the time width of the data extracted from the triaxial acceleration data is 1 second, and the overlap width is 0.5 seconds. This time width and overlap width are the sensor node (acceleration sensor) performance, sensor node It shall be adjusted according to the characteristics of the object to be mounted and the characteristics of the environment where the sensor network is installed. In this embodiment, matching with the standard change pattern is performed by DP matching as shown in Non-Patent Document 6.

  The motion recognition unit 2B includes a feature vector collection module 22, a sample motion database (sample motion DB) 23, a motion dictionary creation unit 24, a maximum likelihood partial motion determination module 25, and a symbolic DP matching processing module 26. The feature quantity vectors from the acceleration feature quantity detection modules 21-1 to 21-4 are given to the feature quantity vector collection module 22.

  The feature quantity vector collection module 22 synthesizes the feature quantity vectors from the acceleration feature quantity detection modules 21-1 to 21-4 every second, and uses the synthesized vector as sample feature data at each time sample operation DB 23. To store.

  The motion dictionary creation unit 24 extracts partial motions included in the sample motion from the sample observation feature data sequence stored in the sample motion DB 23, and extracts the partial motion included in the sample motion from the sample motion. A label corresponding to the partial motion is given, and a label string given to the sample observation feature data belonging to the partial motion is described in the motion dictionary 24A. Further, data representative of sample observation feature data (standard motion feature vector) belonging to the partial motion is extracted and described in the motion dictionary 24A. In addition, a label corresponding to each sample operation is given, and the partial operations included in the sample operation and the order relationship between the partial operations are described in the operation dictionary 24A.

  In the present embodiment, the extraction of the partial motion included in the sample motion is performed by clustering as shown in Non-Patent Document 4. In this case, the “xmeans” algorithm is applied to the sample observation feature data sequence stored in the sample operation DB 23, clustering is performed according to the vector distance between the feature data, and this corresponds to the classified partial operations. A label is assigned to the cluster, and the representative vector of the cluster is set as a standard motion feature vector corresponding to the partial motion.

  Note that the “xmeans” algorithm is an algorithm classified by non-hierarchical clustering. By using the sequential repetition of the “k-means” algorithm and the division stop criterion by BIC, the number of clusters considered to be optimal in terms of information theory can be obtained. Can be determined automatically.

  For example, if the sample motion is “sitting in chair A”, “movement” corresponding to the change α1 of acceleration data S2 and “seat” corresponding to α2 are extracted as partial motions included in the sample motion. The In this case, the motion dictionary creation unit 24 assigns a natural language label of “movement 1” to the sample observation data belonging to “movement” extracted from “sitting in chair A”, and the label string is assigned to the action dictionary 24A. Describe in. Further, a natural language label of “seat 1” is assigned to the sample observation data belonging to “seat” extracted from “seat on chair A”, and the label string is described in the motion dictionary 24A. In addition, a natural language label of “sitting in chair A” is assigned to the sample action corresponding to “sitting in chair A”, and “moving 1” and “sitting 1” are included in “sitting in chair A”. As the partial actions to be performed, “movement 1” and “sitting 1” are described in the action dictionary 24A in this order.

  For example, when the sample operation is “seating from chair A”, “partition” corresponding to change β1 of acceleration data S2 and “movement” corresponding to β2 are included as partial operations included in the sample operation. Extracted. In this case, the motion dictionary creation unit 24 assigns a natural language label “seating 1” to the sample observation data belonging to “seating” extracted from “seating from chair A”, and the label string is given. Described in the action dictionary 24A. Further, a natural language label of “move 2” is assigned to the sample observation data belonging to “move” extracted from “seating from chair A”, and the label string is described in the action dictionary 24A. In addition, a natural language label “seating from chair A” is assigned to the sample action corresponding to “seating from chair A”, and “seating 1” and “movement 2” are separated from “chair A”. The partial motions included in the “seat” are described in the motion dictionary 24A in the order of “seating 1” and “moving 2”.

  In the present embodiment, “use of chair A” that combines “sitting in chair A” and “seating from chair A”, which is not actually generated in the environment, as necessary. It is defined as a sample operation (see FIG. 7). That is, “movement 1”, “seating 1”, “seating 1”, and “movement 2” are partial motions, and “movement 1”, “seating 1”, “seating 1”, and “movement 2” are performed in this order. By describing in the motion dictionary 24A, “use of chair A” that combines “sitting in chair A” and “seating from chair A” that is not actually generated in the room is defined as a sample motion. To do.

  As described above, in this embodiment, a new sample operation can be defined by combining partial operations in the operation dictionary 24A without actually generating a sample operation in the room. Moreover, the label provided here may be freely generated as necessary, or the vocabulary that can be used based on a thesaurus (synonyms) may be limited.

[Motion recognition during actual observation]
In this motion recognition system, the motion recognition device 2 recognizes the motion in the room using the motion dictionary 24A described above as follows. During actual observation, acceleration data (triaxial acceleration data) S1 to S4 from the sensor nodes 1-1 to 1-4 are given to the acceleration feature amount detection modules 21-1 to 21-4 of the motion recognition device 2.

  The acceleration feature quantity detection module 21 (21-1 to 21-4), as in the case of sample observation, transmits triaxial acceleration data S (S1 to S4) sent from the sensor node 1 (1-1 to 1-4). ), The data having a width of 1 second are cut out so as to overlap each other by 0.5 seconds, and used as sensor data from the sensor node 1 (1-1 to 1-4). Then, data is cut out from the sensor data every second, and matching is performed with 11 standard change patterns (patterns P1 to P11 shown in FIG. 4) prepared in advance, and similar to each standard change pattern. The degree is calculated, and a feature quantity vector having the similarity to each standard change pattern as an element is obtained.

  The feature quantity vectors detected by the acceleration feature quantity detection modules 21-1 to 21-4 are given to the feature quantity vector collection module 22. The feature quantity vector collection module 22 synthesizes the feature quantity vectors from the acceleration feature quantity detection modules 21-1 to 21-4 every second, and uses the synthesized vector as actual observation data to obtain the maximum likelihood partial motion determination module 25. Send to.

  The maximum likelihood partial motion determination module 25 obtains the similarity between the actual observation data from the feature vector collection module 22 and the standard motion feature vector representing the sample observation feature data belonging to the partial motion in the motion dictionary 24, A label attached to a standard motion feature vector having a high degree of similarity is attached to the actual observation data. For example, if the actual observation data is similar to the standard motion feature vector to which the label “movement 1” is assigned as a partial action, the label “movement 1” is assigned to the actual observation data.

  When a label is assigned to the actual observation data by the maximum likelihood partial action determination module 25, the symbolic DP matching processing unit 26 and the partial sequence in the action dictionary 24A and the label sequence assigned to the actual observation data so far. Symbolic DP matching is performed with the label sequence of the two, and a sample operation with a high likelihood in the motion dictionary 24A is recognized as a motion in the room at that time. Since symbolic DP matching is described in Non-Patent Document 3, description thereof is omitted here.

  In this example, symbolic DP matching is performed with the partial action label sequence in the action dictionary 24A, and the label of the sample action corresponding to the series of partial actions whose likelihood exceeds the threshold (included partial actions and Labels of sample operations having the same order) are output in descending order of likelihood as indoor movements observed at that time.

  For example, in the actual observation data, there are a label column of “Movement 1” and a label column of “Seating 1”, and the label column of “Movement 1” and the label column of “Seating 1” are “Movement 1” in the motion dictionary 24A. ”And the“ seat 1 ”label sequence, and the order is“ moving 1 ”and“ seat 1 ”, the label“ sitting on chair A ”is the sample operation with high likelihood. Is output.

  Further, the actual observation data includes a label column of “seating 1” and a label column of “moving 2”. The label column of “seating 1” and the label column of “moving 2” are “ If the label sequence of “seating 1” matches the label sequence of “moving 2”, and the order is “seating 1” and “moving 2”, then the sample operation from “chair A” has a high likelihood. The label “Away” is output.

  In addition, the actual observation data includes label columns of “movement 1”, “seating 1”, “seating 1”, and “movement 2”, and this “movement 1”, “seating 1”, “seating 1”, The label column of “Movement 2” matches the label sequence of “Movement 1”, “Seating 1”, “Leave 1”, “Movement 2” in the motion dictionary 24A, and the order is “Movement 1”, “Seating” If “1”, “seating 1”, and “moving 2” are in order, “usage of chair A”, “sitting in chair A”, “seating from chair A” are given in descending order of likelihood. A label is output.

  In the above-described embodiment, the sample operation is “sitting in chair A” or “seating from chair A” to make the explanation easy to understand, but chair A, desk B, chair C, door Needless to say, D is combined to generate various sample operations.

  In the embodiment described above, in the acceleration feature quantity detection module 21 (21-1 to 21-4), data is cut out from the sensor data from the sensor node 1 (1-1 to 1-4) every second. , Matching between a plurality of standard change patterns prepared in advance to calculate the similarity to each standard change pattern, and to obtain a feature vector having the similarity to each standard change pattern as an element However, in the acceleration feature quantity detection module 21 (21-1 to 21-4), data is cut out from the sensor data from the sensor node 1 (1-1 to 1-4) every second, and the cut out data It is also possible to obtain a feature quantity vector having frequency components obtained by performing the fast Fourier transform as described above.

  Also, instead of obtaining a feature quantity vector whose element is a frequency component obtained by performing fast Fourier transform of the cut-out data, a feature having a weblet coefficient obtained by performing a weblet transform of the cut-out data as an element A quantity vector may be obtained, or a feature quantity vector having an autoregressive coefficient obtained by estimating an autoregressive model of the cut out data as an element may be obtained.

It is a figure which shows the outline of one Embodiment of the motion recognition system which concerns on this invention. It is a figure which illustrates the acceleration data from a sensor node. It is a functional block diagram of the motion recognition apparatus in the motion recognition system which concerns on this invention. It is a figure which shows the change of the acceleration data of 3 axes | shafts from a sensor node at the time of generating "seating from chair A" and generating "seating from chair A" as sample operation | movement. It is a figure explaining the extraction according to the time width and overlap width from triaxial acceleration data. It is a figure which illustrates the standard change pattern used with the acceleration feature-value detection module of a motion recognition apparatus. It is a figure explaining the example which defines "use of a chair" which combined "sitting in a chair" and "seating from a chair" as a sample operation | movement.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 (1-1 to 1-4) ... Sensor node, 2 ... Motion recognition apparatus, 2A ... Information input part, 2B ... Motion recognition part, 3 ... Network, 21 (21-1-24) ... Acceleration feature-value detection module , 22 ... feature vector collection module, 23 ... sample motion DB, 24 ... motion dictionary creation unit, 24A ... motion dictionary, 25 ... maximum likelihood partial motion determination module, 26 ... symbolic DP matching processing module, S (S1-S4) ) ... acceleration data, P1-P11 ... standard change pattern, A, C ... chair, B ... desk, D ... door.

Claims (8)

A motion recognition system comprising a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from these acceleration sensors,
The motion recognition device includes:
Data is extracted every predetermined time interval from information from the acceleration sensor observed when a sample motion is generated as a motion candidate to be recognized in the environment, and between a plurality of standard change patterns prepared in advance. Is used to obtain a feature vector whose elements are the similarity to each standard change pattern, and the vector obtained by combining the feature vectors at the same time of each acceleration sensor is used as the feature data of the sample observation. Partial motion extraction means for extracting a partial motion included in the sample motion from a sequence of data;
A label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the label string attached to the feature data of the sample observation belonging to the partial motion belongs to the partial motion. Action dictionary creating means for creating data representing the characteristic data of sample observation, and an action dictionary describing the partial actions included in the sample action and the order relation between the partial actions for each sample action;
Data is cut out from the information from the acceleration sensor actually observed in the environment at the same predetermined time width as the sample observation, and each standard is matched with a plurality of standard change patterns prepared in advance. A feature vector whose element is a similarity to the change pattern is obtained, and a vector obtained by synthesizing feature vectors at the same time of each acceleration sensor is used as actual observation feature data, and the actual observation feature data and a part in the motion dictionary A label attaching means for obtaining a similarity with data representing the feature data of sample observation belonging to the operation, and attaching a label attached to the representative data with high similarity to the feature data of the actual observation;
Each time a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, A motion recognition system comprising: motion recognition means for recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time. A motion recognition system comprising a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from these acceleration sensors,
The motion recognition device includes:
It is obtained by extracting data every predetermined time width from information from the acceleration sensor observed when a sample motion is generated as a motion candidate recognized in the environment, and performing fast Fourier transform of the extracted data A feature vector whose elements are frequency components is obtained, and a vector obtained by synthesizing the feature vector at the same time of each acceleration sensor is used as sample observation feature data, and the partial operation included in the sample operation from this sample observation feature data column Partial motion extraction means for extracting
A label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the label string attached to the feature data of the sample observation belonging to the partial motion belongs to the partial motion. Action dictionary creating means for creating data representing the characteristic data of sample observation, and an action dictionary describing the partial actions included in the sample action and the order relation between the partial actions for each sample action;
The data is cut out from the information from the acceleration sensor actually observed in the environment at the same predetermined time width as the sample observation, and the frequency component obtained by performing the fast Fourier transform of the cut out data is an element. The vector obtained by obtaining the quantity vector and combining the feature quantity vectors at the same time of each acceleration sensor is used as the actual observation feature data, and the actual observation feature data and the sample observation feature data belonging to the partial motion in the motion dictionary are represented. A label attaching means for obtaining a similarity to the data and assigning a label attached to the representative data having a high similarity to the feature data of the actual observation;
Each time a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, A motion recognition system comprising: motion recognition means for recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time. A motion recognition system comprising a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from these acceleration sensors,
The motion recognition device includes:
It is obtained by cutting out data at predetermined time intervals from information from the acceleration sensor observed when a sample motion is generated as a motion candidate recognized in the environment, and performing weblet conversion of the cut out data A feature vector whose elements are weblet coefficients is obtained, and a vector obtained by combining feature vectors at the same time of each acceleration sensor is used as feature data for sample observation. Partial motion extraction means for extracting motion;
A label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the label string attached to the feature data of the sample observation belonging to the partial motion belongs to the partial motion. Action dictionary creating means for creating data representing the characteristic data of sample observation, and an action dictionary describing the partial actions included in the sample action and the order relation between the partial actions for each sample action;
Data is extracted from the information from the acceleration sensor actually observed in the environment at the same predetermined time width as the sample observation, and a weblet coefficient obtained by performing weblet conversion of the extracted data is used as an element. The feature vector is obtained, and the vector obtained by combining the feature vectors at the same time of each acceleration sensor is used as the feature data of the actual observation, and the feature data of the actual observation and the feature data of the sample observation belonging to the partial motion in the motion dictionary are represented. A label attaching means for obtaining a similarity to the data to be obtained, and attaching a label attached to representative data having a high similarity to the feature data of the actual observation;
Each time a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, A motion recognition system comprising: motion recognition means for recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time. A motion recognition system comprising a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from these acceleration sensors,
The motion recognition device includes:
It is obtained by cutting out data every predetermined time width from information from the acceleration sensor observed when a sample motion is generated as a motion candidate recognized in the environment, and estimating an autoregressive model of the cut out data. A feature vector that has an autoregressive coefficient as an element is obtained, and a vector obtained by synthesizing feature vectors at the same time of each acceleration sensor is used as sample observation feature data, and is included in the sample operation from this sample observation feature data column Partial motion extraction means for extracting the partial motion;
A label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the label string attached to the feature data of the sample observation belonging to the partial motion belongs to the partial motion. Action dictionary creating means for creating data representing the characteristic data of sample observation, and an action dictionary describing the partial actions included in the sample action and the order relation between the partial actions for each sample action;
Data is cut out from the information from the acceleration sensor actually observed in the environment at the same predetermined time width as the sample observation, and an autoregressive coefficient obtained by estimating an autoregressive model of the cut out data is used as an element. A vector obtained by synthesizing the feature vector at the same time of each acceleration sensor as the feature data of the actual observation, and the feature data of the actual observation and the feature data of the sample observation belonging to the partial motion in the motion dictionary A label assigning means for obtaining a similarity with representative data and assigning a label attached to representative data with a high similarity to the feature data of the actual observation;
Each time a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, A motion recognition system comprising: motion recognition means for recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time. In a system comprising a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from these acceleration sensors,
Data is extracted every predetermined time interval from information from the acceleration sensor observed when a sample motion is generated as a motion candidate to be recognized in the environment, and between a plurality of standard change patterns prepared in advance. Is used to obtain a feature vector whose elements are the similarity to each standard change pattern, and the vector obtained by combining the feature vectors at the same time of each acceleration sensor is used as the feature data of the sample observation. Extracting a partial motion included in the sample motion from a sequence of data;
A label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the label string attached to the feature data of the sample observation belonging to the partial motion belongs to the partial motion. Action dictionary creating means for creating data representing the characteristic data of sample observation, and an action dictionary describing the partial actions included in the sample action and the order relation between the partial actions for each sample action;
Data is cut out from the information from the acceleration sensor actually observed in the environment at the same predetermined time width as the sample observation, and each standard is matched with a plurality of standard change patterns prepared in advance. A feature vector whose element is a similarity to the change pattern is obtained, and a vector obtained by synthesizing feature vectors at the same time of each acceleration sensor is used as actual observation feature data, and the actual observation feature data and a part in the motion dictionary Obtaining a similarity to the data representing the feature data of the sample observation belonging to the operation, and assigning a label attached to the representative data having a high similarity to the feature data of the actual observation;
Each time a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, And recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time. In a system comprising a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from these acceleration sensors,
It is obtained by extracting data every predetermined time width from information from the acceleration sensor observed when a sample motion is generated as a motion candidate recognized in the environment, and performing fast Fourier transform of the extracted data A feature vector whose elements are frequency components is obtained, and a vector obtained by synthesizing the feature vector at the same time of each acceleration sensor is used as sample observation feature data, and the partial operation included in the sample operation from this sample observation feature data column Extracting the
A label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the label string attached to the feature data of the sample observation belonging to the partial motion belongs to the partial motion. Creating data representing the characteristic data of sample observation, and a motion dictionary describing the partial motions included in the sample motion and the order relationship between the partial motions for each sample motion;
The data is cut out from the information from the acceleration sensor actually observed in the environment at the same predetermined time width as the sample observation, and the frequency component obtained by performing the fast Fourier transform of the cut out data is an element. The vector obtained by obtaining the quantity vector and combining the feature quantity vectors at the same time of each acceleration sensor is used as the actual observation feature data, and the actual observation feature data and the sample observation feature data belonging to the partial motion in the motion dictionary are represented. Obtaining a similarity to the data, and attaching a label attached to representative data having a high similarity to the feature data of the actual observation;
Each time a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, And recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time. In a system comprising a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from these acceleration sensors,
It is obtained by cutting out data at predetermined time intervals from information from the acceleration sensor observed when a sample motion is generated as a motion candidate recognized in the environment, and performing weblet conversion of the cut out data A feature vector whose elements are weblet coefficients is obtained, and a vector obtained by combining feature vectors at the same time of each acceleration sensor is used as feature data for sample observation. Extracting the motion;
A label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the label string attached to the feature data of the sample observation belonging to the partial motion belongs to the partial motion. Creating data representing the characteristic data of sample observation, and a motion dictionary describing the partial motions included in the sample motion and the order relationship between the partial motions for each sample motion;
Data is extracted from the information from the acceleration sensor actually observed in the environment at the same predetermined time width as the sample observation, and a weblet coefficient obtained by performing weblet conversion of the extracted data is used as an element. The feature vector is obtained, and the vector obtained by combining the feature vectors at the same time of each acceleration sensor is used as the feature data of the actual observation, and the feature data of the actual observation and the feature data of the sample observation belonging to the partial motion in the motion dictionary are represented. A labeling step for determining the similarity to the data to be performed, and giving the label given to the representative data with high similarity to the feature data of the actual observation,
Each time a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, And recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time. In a system comprising a plurality of acceleration sensors installed in an environment and a motion recognition device that recognizes movement in the environment based on information from these acceleration sensors,
It is obtained by cutting out data every predetermined time width from information from the acceleration sensor observed when a sample motion is generated as a motion candidate recognized in the environment, and estimating an autoregressive model of the cut out data. A feature vector that has an autoregressive coefficient as an element is obtained, and a vector obtained by synthesizing feature vectors at the same time of each acceleration sensor is used as sample observation feature data, and is included in the sample operation from this sample observation feature data column Extracting partial motions;
A label corresponding to the partial motion is assigned to the feature data of the sample observation belonging to the partial motion extracted from the sample motion, and the label string attached to the feature data of the sample observation belonging to the partial motion belongs to the partial motion. Creating data representing the characteristic data of sample observation, and a motion dictionary describing the partial motions included in the sample motion and the order relationship between the partial motions for each sample motion;
Data is cut out from the information from the acceleration sensor actually observed in the environment at the same predetermined time width as the sample observation, and an autoregressive coefficient obtained by estimating an autoregressive model of the cut out data is used as an element. A vector obtained by synthesizing the feature vector at the same time of each acceleration sensor as the feature data of the actual observation, and the feature data of the actual observation and the feature data of the sample observation belonging to the partial motion in the motion dictionary Obtaining a similarity to the representative data, and attaching a label attached to the representative data having a high similarity to the feature data of the actual observation;
Each time a label is assigned to the actual observation feature data, matching is performed between the label sequence assigned to the actual observation feature data and the partial motion label sequence in the motion dictionary, And recognizing a sample operation with a high likelihood in the motion dictionary as a motion in the environment at that time.

Images