JP7172733B2 - Human Posture Estimation Device - Google Patents

Description

The present invention relates to a device for estimating the posture (orientation or movement of the trunk) of a person or an animal (hereinafter referred to as "person"), and more specifically, the apparatus measures the trunk of the person. The present invention relates to a device for estimating the posture of a person based on acceleration.

Information about a person's or animal's posture, i.e. whether a person is standing, sitting, or lying down, is itself useful in, for example, the monitoring and management of the health condition of a person and work. To provide useful information for monitoring and confirming the safety of people, etc., and to accurately detect useful information in terms of health management or safety management, such as detecting and estimating the sleeping state and drowsiness of people, etc. can be used as a reference for Therefore, various devices or systems for detecting or estimating the posture of a person have been proposed for the purpose of monitoring and managing the health and safety of the person. For example, in Patent Document 1, a three-axis acceleration sensor is arranged so that the directions of the first axis and the second axis correspond to the longitudinal direction and the vertical direction of the subject in a standing state, respectively. Detecting the stationary state of the subject from the DC signal component included in the output of the three-axis acceleration sensor attached to the upper body, and detecting the motion state of the subject from the AC signal component included in the output of the three-axis acceleration sensor. The state transition of the subject detected from the stationary state and the moving state, and the transition of the subject to the standing state and the transition to the sitting state from the displacement of the tilt angle formed by the direction of the first axis and the direction of gravity. A technique for estimating is proposed. Patent Document 2 proposes a technique for estimating the posture of a living body based on altitude data acquired by a first sensor attached to the living body. Further, in Patent Document 3, posture information of a moving body (person) is obtained using state vectors in the direction of gravity, the horizontal direction, and the due north direction calculated from information on acceleration, angular velocity, and geomagnetism measured on the moving body (person). Generating configurations are disclosed.

JP 2015-109915 JP 2016-150034 JP 2017-207456

When an acceleration sensor is attached to the trunk of a person, such as the chest, abdomen, or waist, if the posture of the person changes, the direction of the acceleration sensor changes, and the acceleration value measured there also changes. It is possible to estimate the posture of a person or the like based on (When the posture of a person or the like is detected from the acceleration of the trunk, the term "estimation" is used because the posture is not directly observed.). In particular, relatively compact portable sensors are available as acceleration sensors that can be worn on the trunk of a person. It is thought that it will be possible to measure the acceleration on the trunk of the person, etc., and estimate the posture of the person, etc. based on the acceleration value, while the person etc. is acting normally without any influence. However, when actually attaching the acceleration sensor to the trunk of a person, etc., due to differences in the physique of the person, etc., and the position and method of attaching the sensor, it is necessary to ensure that the sensor always faces in the same direction. It is not attached to the trunk, and therefore, even when the posture is the same, the measured value of acceleration by the sensor may be different. It is difficult to obtain accurate results. In relation to this point, the inventors of the present invention attached an acceleration sensor to the trunk of a person, etc., measured the acceleration in three axial directions in time series, and obtained the measurement data of the acceleration. Estimate the posture of a person using the data obtained by applying a process that compensates for the difference in the orientation of the sensor to the acceleration measurement data while walking. As a result, the inventors have found that the accuracy of the estimation result can be improved regardless of the physique of the person wearing the sensor or the position or orientation of the sensor. The present invention utilizes this finding.

Accordingly, one object of the present invention is to provide an apparatus capable of estimating the posture of a person with high accuracy based on acceleration values measured by an acceleration sensor attached to the trunk of the person.

According to the present invention, the above problem is a device for estimating the posture of a person, comprising:
an acceleration sensor that is worn on the trunk of a person or the like and measures acceleration values in three axial directions in time series;
Among the acceleration values in the three-axis directions, the acceleration values in the three-axis directions when the person or the like is in a walking state are extracted, and the extracted acceleration values in the three-axis directions are used to determine the reference vector determination means for determining a reference vector representing a representative direction of acceleration in a walking state;
converted acceleration value determination means for converting the acceleration values in the three-axis directions into converted acceleration values, which are coordinate values in a reference coordinate space in which the direction of the reference vector coincides with a predetermined reference direction;
posture estimation means for estimating the posture of the person or the like based on the transformed acceleration value.

In the above configuration, the "person" may be a person or an animal (or a walking robot), as already mentioned, and "estimating the posture of a person" , standing, sitting, lying down, walking (horizontal), climbing stairs, descending stairs, and so on. , and may include not only postures in a stationary state, but also postures during movement.) The “trunk of a person, etc.” may be any part whose direction changes depending on the posture, such as the head, neck, chest, abdomen, waist, buttocks, etc. of a person. "Acceleration values in three-axis directions" are acceleration values along three mutually different directions (three-axis directions do not necessarily have to be orthogonal). An acceleration vector acting on the trunk in the original space is determined. Note that the "acceleration sensor" may be of any type as long as it can measure acceleration values in three axial directions. A sensor is advantageously used, but is not limited to this (sensors capable of measuring uniaxial acceleration values may be mounted in three different directions and used). The “representative direction of acceleration when a person is in a walking state” means a representative direction of the direction in which the acceleration vector acting on the trunk of a person while walking is directed. The direction that may be considered to be facing most of the time, usually the direction of gravity. A "reference vector" is a vector representing a representative direction (that is, the direction of gravity) of an acceleration vector acting on the trunk of a walking person. When determining the "reference vector", the three element values representing it may be representative values of the three acceleration values in the direction measured by the acceleration sensor on the trunk of a walking person. Specifically, for example, it may be a median value, an average value, or a value obtained by principal component analysis of acceleration values in three axial directions. As described above, the "reference coordinate space" is a coordinate space in which the direction of the reference vector coincides with the arbitrarily set "predetermined reference direction". A coordinate space that coincides with the positive or negative direction of one of the x-, y-, and z-axes (i.e., the reference vector is the x-, y-, and z-axes in the reference coordinate space). matches the positive or negative direction of any one of . Then, in the above configuration, "the acceleration values in the three-axis directions are converted into converted acceleration values, which are coordinate values in a reference coordinate space in which the direction of the reference vector coincides with a predetermined reference direction." converts the element values of the acceleration vector determined by the acceleration values in the three-axis directions to the coordinate values in the "measurement coordinate space" (the coordinate space spanned by the three axes in which the acceleration sensor measures the acceleration values) (that is, , acceleration values) to coordinate values in the reference coordinate space (that is, “transformed acceleration values”).

In the apparatus of the present invention described above, basically, the acceleration value measured by the acceleration sensor is determined by changing the direction of the acceleration sensor attached to the trunk of the person as the posture of the person changes. changes, the posture of a person or the like is estimated based on the acceleration value. However, in the case of such a configuration, as already described, the physique and body shape of the person wearing the acceleration sensor, the mounting position of the acceleration sensor on the trunk, the orientation of the acceleration sensor at the mounting position, and the like are different. Therefore, it is not always possible to attach the acceleration sensor in the same direction to the trunk, and it is difficult to directly grasp the orientation of the acceleration sensor attached to the trunk of a person. The direction of the measurement axis of the acceleration sensor (the direction in which the acceleration value is obtained) changes every time the acceleration sensor is attached to the trunk, and even when the posture is the same, the acceleration value measured by the acceleration sensor changes from time to time. Thus, when the posture of a person or the like is estimated by directly referring to the acceleration value, the accuracy of the estimation result may be lowered. That is, even when the person is in the same posture, for example, even when the person is standing, depending on the direction of the acceleration sensor attached to the trunk, the direction of the acceleration sensor in the measurement coordinate space of the acceleration sensor Since the direction of gravity changes, the acceleration values measured by the acceleration sensor will differ, and the result of estimating the posture of a person or the like may change each time the acceleration sensor is worn.

Therefore, in the configuration of the apparatus of the present invention, an acceleration sensor is attached to the trunk of a person, etc., and after measuring acceleration values in three axial directions on the trunk, the acceleration values are calculated as described above. Arithmetic processing is executed to compensate for the influence of the difference in orientation of the acceleration sensor with respect to the trunk, and the posture of a person or the like is estimated using the compensated acceleration value. Improvement is planned.

Specifically, in the arithmetic processing for compensating for the influence of the difference in orientation of the acceleration sensor with respect to the trunk in the acceleration values measured by the acceleration sensor, The direction of the acceleration vector when a person or the like is standing in the measurement coordinate space of the sensor, that is, the direction of gravity is detected. Therefore, in the configuration of the present invention, the acceleration values in the three-axis directions measured in time series by the acceleration sensor are the acceleration values in the three-axis directions when a person or the like is walking substantially horizontally. Values are extracted, and the extracted time-series acceleration values are used to determine a reference vector representing a representative direction of acceleration when a person or the like is in a walking state. When a person walks in a substantially horizontal direction, the posture of the person is usually in a standing position. Although it vibrates along with it, its direction is considered to be generally directed in the direction of gravity. Therefore, the direction of the vector represented by the representative value of the acceleration values in the three-axis directions measured by the acceleration sensor in the walking state of a person is It can be specified as the direction of gravity. In addition, in detecting the direction of gravity in the measurement coordinate space of the acceleration sensor when the person is standing, the acceleration value when the person is walking is used instead of when the person is standing still. , when a person is walking, a vibrational change is observed in the acceleration value due to walking. It can be distinguished, but when a person stands still, there is no vibrational change in the acceleration value, and when the direction of the acceleration sensor is unknown, the acceleration value in the standing posture is referred to it. This is because it is difficult to distinguish between the sitting position and the lying position just by doing so.

Any method can be used to extract the acceleration values in the three-axis directions when a person or the like is walking in a substantially horizontal direction from the acceleration values in the three-axis directions measured in time series by the acceleration sensor. be. As mentioned above, the acceleration value of a person while walking is characterized by vibrational changes that accompany the walking motion. By detecting or identifying the presence or absence of characteristics of the acceleration value during walking, the acceleration value in the walking state of a person or the like can be extracted as time-series data. For example, in one aspect, a discriminator configured using machine learning can be used to extract acceleration value data while a person or the like is walking. In this case, specifically, time-series data of acceleration values when a person or the like is in various states including walking, and corresponding time-series data of the posture of the person or the like (correct data) are collected. A feature quantity selected as reflecting the characteristics of the posture of a person, especially the walking state, is calculated from time-series data of acceleration values prepared as training data, and the characteristics of the time-series data of acceleration values are calculated. The parameters of the discriminator are determined by machine learning so that the amount is used as input data and the orientation of correct data is output as the discrimination result. Then, when identifying the walking section of an actual person, etc., the feature amount of the acceleration values in the three-axis directions measured in time series by the acceleration sensor is transferred to the classifier constructed using the parameters determined as described above. By inputting, the data of the section in which the posture of the person is in the walking state is identified and extracted from the time-series data of the acceleration value. In the configuration of such a discriminator, for example, a method using machine learning capable of configuring an arbitrary discriminator such as a random forest method, a support vector machine method, a k-NN method, and a neural network is adopted. good. It should be noted that the extraction of acceleration value data while a person is walking can be done by any method that does not rely on machine learning. It may be achieved by determining that it is an acceleration value during walking, and such a case also belongs to the scope of the present invention.

As described above, when acceleration values are extracted while a person is walking, the reference vector representing the representative direction of the acceleration when the person is in a walking state is obtained by referring to the extracted acceleration value data. The element values are determined using acceleration values, that is, coordinate values in each axial direction in the measurement coordinate space of the acceleration sensor. Specifically, as described above, the element values of the reference vector may be, for example, median values, average values, and values obtained by principal component analysis of acceleration values in three axial directions. Thus, the reference vector obtained here is an acceleration vector directed in the direction of gravity, and its element values are coordinate values in the measurement coordinate space of the acceleration sensor when the posture of a person, etc., is standing. It is what is represented.

When the reference vector represented by the coordinate values in the measurement coordinate space of the acceleration sensor is determined, the coordinate space in which the direction of the reference vector is the predetermined reference direction, i.e., the reference coordinate space, can be determined. Become. The predetermined reference direction is one direction that may be arbitrarily selected as described above, and is typically one axis in three-dimensional space, for example, the negative direction of the x-axis (x, y, z )=(−1,0,0), etc. may be chosen. Then, if the element values of the acceleration vector determined by the acceleration values measured by the acceleration sensor are represented by the coordinate values in the reference coordinate space, regardless of the direction of the acceleration sensor attached to the trunk, Since the direction of gravity in a standing position is represented by coordinate values in a coordinate space in the predetermined reference direction, the influence of the difference in orientation of the acceleration sensor attached to the trunk is compensated. becomes. Thus, as described above, the process of converting the acceleration values in the three-axis directions into the converted acceleration values, which are the coordinate values in the reference coordinate space, is executed.

The transformation from the acceleration values in the three axial directions to the transformed acceleration values may be performed in various manners. In one aspect, converting the coordinate values (acceleration values) of a vector in the measurement coordinate space to the coordinate values in the reference coordinate space is performed in a predetermined reference direction (e.g., , the negative direction of the x-axis) coincides with the direction of the reference vector. An angular deviation with direction may be calculated. When the reference vector is rotated with this angular deviation as the rotation angle, the coordinate values of the reference vector will match the coordinate values of the predetermined reference direction. When the acceleration value vector determined by the acceleration values in the 3-axis directions measured in time series by the acceleration sensor is rotated, the coordinate values of the acceleration value vector are changed to the coordinate values in the reference coordinate space. will be converted to That is, the acceleration values in the three-axis directions measured by the acceleration sensor are subjected to coordinate transformation corresponding to rotating the reference vector so as to match the predetermined reference direction, so that the acceleration values in the three-axis directions are converted to the reference coordinates. A process of transforming to a post-transformation acceleration value, which is a coordinate value in space, is accomplished. Specifically, for example, as will be described later in detail, such arithmetic processing is performed by converting the coordinate values of the reference vector in the measurement coordinate space into the coordinate values in the reference coordinate space. may be determined using a rotation angle that makes the direction of the direction coincide with a predetermined reference direction, and may be achieved by using the rotation conversion matrix to convert the acceleration values in the three-axis directions measured by the acceleration sensor.

As described above, once the transformed acceleration values are obtained, the posture of a person or the like is estimated using those values. As described above, the post-conversion acceleration value is the acceleration vector measured by the acceleration sensor on the trunk expressed as coordinate values in the reference coordinate space. The element values of the acceleration value vectors after conversion are approximately the same, and when the person is in a different posture, the element values of the acceleration value vectors after conversion also have different values. It is expected that it will be possible to identify the posture of a person with high accuracy.

Any method can be used to identify the posture of a person or the like based on the converted acceleration value. As one aspect, a classifier constructed using machine learning can also be used for the recognition of the posture of a person or the like here. In that case, specifically, first, time-series data of acceleration values in three axial directions measured by an acceleration sensor when a person or the like is in various states, and the corresponding time-series of the posture of the person or the like data (correct answer data) are prepared as teacher data. As described above, the time-series data of the acceleration values measured by the acceleration sensor is subjected to the conversion process for matching the direction of the reference vector with the predetermined reference direction, thereby converting the acceleration values in the three-axis directions into the converted acceleration values. , thus calculating the time-series data of the post-conversion acceleration values. Then, from the time-series data of the acceleration values after conversion, the feature amount selected as reflecting the feature of the posture of the person is calculated, and the feature amount is used as input data, and the posture of the correct data is output as the identification result. The parameters of the discriminator are determined by machine learning. Then, when estimating the posture of a person, the converted acceleration obtained from the acceleration values in the three axial directions measured in time series by the acceleration sensor is sent to the discriminator configured using the parameters determined as described above. By inputting the feature amount of the value, it becomes possible to identify and extract the posture of a person or the like. In the configuration of such a discriminator, for example, a method using machine learning capable of configuring an arbitrary discriminator such as a random forest method, a support vector machine method, a k-NN method, and a neural network is adopted. good. The posture of a person may be identified by an arbitrary method that does not rely on machine learning, for example, the posture of a person may be determined according to the threshold range to which an arbitrary feature amount calculated from the converted acceleration value belongs. .

In the configuration of the apparatus of the present invention described above, an air pressure sensor is used as a sensor attached to the trunk of a person, etc. to measure the air pressure value in the trunk. It may be used for extraction and estimation of the posture of a person or the like when the person or the like is walking in a substantially horizontal direction. Based on the air pressure value, it is expected that it will be possible to more accurately identify whether a person is walking horizontally or going up and down stairs. In this case, the air pressure value may also be used as one of the input data in constructing the discriminator based on machine learning.

In the configuration of the apparatus of the present invention described above, the acceleration sensor (and the atmospheric pressure sensor) may be accommodated in a housing that can be worn on the trunk of a person or the like. A device for performing calculation and identification processing for estimating the posture of a person from measurement data (acceleration value, air pressure value) may be housed in the housing or may be realized by an external computer. good.

Thus, according to the above apparatus of the present invention, in the configuration for estimating the posture of a person based on the acceleration values measured by the acceleration sensor attached to the trunk of the person, Since the difference in the orientation of the acceleration sensor is compensated for, there is robustness against the orientation of the acceleration sensor, and an improvement in the accuracy of the result of estimating the posture of a person is expected. In addition, since the device of the present invention can be achieved with a relatively compact portable acceleration sensor and a calculation device for the measured values, it is possible to reduce the effects on the orientation and movement of the body of the person as much as possible. This is advantageous in that it is possible to relatively easily and inexpensively estimate the posture of a person, etc., while they are acting normally.

Other objects and advantages of the present invention will become apparent from the following description of preferred embodiments of the invention.

FIG. 1(A) schematically shows a housing attached to the trunk of a subject including sensors for measuring acceleration values and air pressure values in an embodiment of a posture estimation device according to the present invention. It is a diagram. FIG. 1B is a block diagram showing the internal configuration of an embodiment of a posture estimation device according to the present invention. FIG. 1(C) is a diagram showing, in the form of a block diagram, the configuration of a device for machine learning of discriminator parameters used for posture estimation in an embodiment of a posture estimation device according to the present invention. FIGS. 2A, 2B, and 2C are schematic diagrams of a person standing, sitting, and lying down. FIG. 2(D) shows reference vectors in the measurement coordinate space of the acceleration sensor, and FIG. 2(E) shows reference vectors in the reference coordinate space. FIGS. 3A and 3B are examples of acceleration values measured by a triaxial acceleration sensor attached to the trunk of the subject. (A) is when the subject is walking, and (B) is when the subject is standing (still). FIG. 4A is a diagram showing, in the form of a flow chart, processing for estimating the posture of a subject in the embodiment of the posture estimation device according to the present invention. FIG. 4B is a diagram showing, in the form of a flowchart, the process of determining classifier parameters used for posture estimation in this embodiment. FIG. 4C is a diagram for explaining the timing of calculation of feature amounts in this embodiment. FIG. 5 shows distributions of X-axis acceleration values and Y-axis acceleration values measured by a 3-axis acceleration sensor attached to the trunk of the subject when the posture of the subject is standing, sitting, and lying. It is a diagram. (A) is a plot of the measurement values of the acceleration sensor, and (B) is the acceleration value corrected to compensate for the orientation deviation when the acceleration sensor is worn (after conversion) according to the teaching of the present embodiment. acceleration value). FIG. 6(A) shows the position of the housing containing the sensor worn by the subject when the verification experiment of the effect of the posture estimation apparatus according to the present embodiment was conducted. (C) shows the behavioral protocol imposed on the subjects during the experiment. FIG. 7 shows the F value in the estimation result in the verification experiment of the effect of the attitude estimation device in this embodiment.

DESCRIPTION OF SYMBOLS 1... Housing 10... Posture estimation device 12... Machine learning device P... Examinee

Configuration of Apparatus Referring to FIG. 1(A), one embodiment of the apparatus for estimating posture of a person or the like according to the present invention is configured to measure the torso, for example, the head, neck, and chest of a subject P who is a person or the like. , the abdomen, the waist, and the buttocks (wearable) housing 1, and components used for posture estimation are housed in the housing 1. The housing 1 may house an acceleration sensor that measures an acceleration value referred to for posture estimation, and further house an air pressure sensor that measures an air pressure value. The acceleration sensor is a sensor that can measure acceleration values in three-axis directions that determine acceleration vectors that change with changes in the posture of the subject (changes in the direction of the trunk). may be used, but is not limited to this (for example, three sensors that measure acceleration in one direction may be prepared, and may measure acceleration values in different directions). In addition, the acceleration sensor can measure the acceleration on the trunk of the subject while the subject is acting normally without affecting the orientation and movement of the subject's body as much as possible. A relatively compact sensor that can be housed in a portable housing 1 is advantageously used. On the other hand, the atmospheric pressure sensor can measure changes in the atmospheric pressure that occur when the subject walks up and down stairs, or climbs or descends a slope. A relatively compact sensor that can be housed in a portable housing 1 is advantageous because it is possible to measure air pressure while the examinee is acting normally without affecting orientation and movement as much as possible. used for It should be understood that the form of housing 1 containing the sensor may vary depending on the form of sensor employed.

Further, in one aspect of the present embodiment, the housing 1 further includes a processing unit and a processing unit that receive the output of the above sensor and identify the posture of the subject in a manner described later. A display that displays the output value of the unit and/or the operation status of the device, a communication unit that transmits the output value of the processing unit to an external device or facility, and an operation that receives instructions and operations on the device by the subject or user A panel or the like may be provided. In this case, the processing unit may be a conventional small computer device including a microcomputer, memory or flash memory or the like. Alternatively, in another aspect, the processing unit and the like may be configured in a computer device (not shown) separate from the housing 1. In that case, the output of the sensor is It may be transmitted to the computing device via wired or wireless communication. Furthermore, part of the process of identifying the subject may be executed within the housing, and part of the process may be executed by an external computer device.

In the device (posture estimation device) 10 for estimating the posture of the subject according to the present embodiment, an acceleration sensor attached to the trunk of the subject as shown in FIG. A first acceleration value feature quantity extraction unit that receives the measured acceleration value from time to time and extracts or calculates an acceleration feature quantity from the acceleration value, and receives the acceleration feature quantity from the first acceleration value feature quantity extraction unit. Then, referring to them, a classifier configured using parameters (classifier parameters) stored in the memory identifies the walking section in which the subject is walking in the time-series data of acceleration values. a walking zone extracting unit for extracting a walking zone, a reference vector determining unit for determining a reference vector by referring to the acceleration value of the extracted walking zone, and determining a rotation transformation matrix for coordinate transformation of the acceleration value, and a rotation transformation matrix and a second acceleration value feature quantity extractor for extracting or calculating a feature quantity from the coordinate-converted acceleration value (post-conversion acceleration value). and a discriminator that receives the feature quantity of the converted acceleration value from the second acceleration value feature quantity extraction unit, refers to them, and uses parameters (discriminator parameters) stored in the memory. and a posture identification unit that identifies the posture of the subject by. In addition, as shown in the figure, an air pressure value feature quantity extraction unit is provided which receives the air pressure value measured by the air pressure sensor attached to the trunk of the subject every moment and extracts or calculates the pressure feature quantity from the air pressure value. The atmospheric pressure feature quantity may be referred to when the walking segment extraction unit extracts the walking segment and the posture identifying unit identifies the posture of the subject. Then, the posture identification result, that is, the estimation result is displayed on the display or stored in an arbitrary storage device. Whether the configuration for estimating the subject's posture is contained in the housing 1 or configured in a separate computer device, the operation of each unit in the device is controlled by the program stored in the memory. It should be understood that what is achieved by the actuation accordingly.

The parameters of the discriminator used in the process of extracting the walking segment of the subject and the parameters of the discriminator used in the process of estimating the posture of the subject may be determined by machine learning, stored in memory. Such machine learning processing may be executed in a processing unit within the housing 1, and the execution results may be stored in the memory within the housing 1, or may be stored in a separate unit from the housing 1 It may be executed in a computer device, and execution results may be stored in memory.

Referring to FIG. 1C, in a device (machine learning device) 12 that executes machine learning, as in posture estimation device 10, an acceleration value measured by an acceleration value sensor is received and Acceleration feature quantity extraction unit for extracting or calculating acceleration feature quantity; ) as supervised learning data, machine learning processing is performed using such learning data, and the subject identifies and extracts the walking section during walking in the acceleration value time series data. A machine learning (walking identification) unit that determines parameters, and a recognition result (correct value) of the posture of the subject at the time of measurement of the acceleration value measured by the acceleration value sensor and the acceleration value. A walking segment extraction unit that extracts the walking segment in which the subject is walking in the acceleration value time-series data by referring to the posture identification result (correct value), and extracting the acceleration value of the extracted walking segment A reference vector determination unit that determines a reference vector by referring to the reference vector and determines a rotation transformation matrix for coordinate transformation of the acceleration value, and acceleration value coordinates that perform coordinate transformation of the acceleration value measured by the acceleration sensor using the rotation transformation matrix a transformation unit, a second acceleration value feature amount extraction unit for extracting or calculating a feature amount from the coordinate-transformed acceleration value (transformed acceleration value), and obtaining the transformed acceleration feature amount and the transformed acceleration feature amount. When it is determined, that is, the identification result (correct value) of the posture of the subject at the time of measurement of the acceleration value is used as supervised learning data, and machine learning processing is performed using this learning data. A machine learning (posture identification) unit that determines parameters for a classifier that identifies the posture of the body, and a memory that stores classifier parameters for walking recognition and classifier parameters for posture recognition. you can

In addition, an air pressure sensor is prepared for the subject together with the acceleration sensor, and the air pressure value measured by the air pressure sensor attached to the trunk of the subject is received every moment, and the air pressure feature value is extracted or calculated from the air pressure value. A pressure value feature value extraction unit is provided, and the pressure feature value is referred to as input data in learning data in machine learning in the machine learning (walking identification) unit and the machine learning (posture identification) unit. It's okay to be like this.

In the configuration of FIG. 1(C), the acceleration sensor, the atmospheric pressure sensor, the first and second acceleration feature extraction units, the atmospheric pressure feature extraction unit, the reference vector determination unit, and the acceleration value coordinate conversion unit It may be the same as that in the estimating device 10 (may be shared). Whether the machine learning device 12 for the subject is housed in the housing 1 or configured as a separate computer device, the operation of each part in the device is controlled by the program stored in the memory. It should be understood that what is achieved by the actuation accordingly.

Operation of the Device (1) Outline of Improvement in Posture Estimation As explained in the section "Summary of the Invention", in the posture estimation device according to the present invention, Figs. ), when the subject P is equipped with an acceleration sensor housed in the housing 1, the direction of gravity g with respect to the measurement axis (solid arrow) of the acceleration value of the acceleration sensor is Since the acceleration value measured by the acceleration sensor changes due to the change in the posture of the examinee P, the posture of the examinee P changes between the standing position (A) and the sitting position (B) based on the change in the acceleration value. , recumbent position (C), and so on. Regarding posture estimation based on the acceleration value, the acceleration value measured by the acceleration sensor changes depending on the orientation of the acceleration sensor with respect to the trunk of the subject P. Since the orientation may change each time the acceleration sensor is worn, it is difficult to identify the posture of the subject P only by referring to the acceleration values obtained by the acceleration sensor. For example, if the accelerometer is mounted so that its measurement axis faces upward when the posture is standing, and the posture is lying down, the accelerometer is placed so that its measurement axis faces sideways when the posture is standing. It is indistinguishable from a standing posture, and even if the posture is the same, if the direction of the measurement axis of the accelerometer changes, the acceleration value will change. Therefore, it may be determined that they are in different postures. That is, when the direction of the measurement axis of the acceleration sensor with respect to the trunk of the subject P is not always the same, or when it cannot be grasped, it becomes difficult to accurately estimate the posture based on the acceleration value. However, the direction of the measurement axis of the acceleration sensor is not necessarily the same because the physique of the subject and the position and mounting method of the sensor are different. is not easy either.

By the way, in estimating the posture based on the above acceleration values, the acceleration acting on the trunk undergoes temporary fluctuations due to body movements, but the vector (acceleration vector) is generally always of the same magnitude. , faces the direction of gravity (g) (see FIG. 2(D)). This is because the coordinate space), that is, the coordinate axis of the coordinate space when the acceleration vector is represented by the acceleration value (the measurement axis of the acceleration sensor) changes each time the acceleration sensor is attached. Therefore, when estimating the posture, instead of using the acceleration values measured by the acceleration sensor as they are, an acceleration vector determined by the acceleration values measured by the acceleration sensor is set in a specific direction with respect to the trunk. A value obtained by coordinate transformation of the acceleration value measured by the acceleration sensor so that it is represented by the coordinate value in the coordinate space (the coordinate space fixed with respect to the trunk) (post-transformation acceleration value) is used, the post-transformation acceleration value is the same value when the posture is the same, and becomes a different value when the posture is different. Therefore, it is expected that the accuracy of posture estimation is improved.

In this regard, as already mentioned, the acceleration vector acting on the trunk basically points in the direction of gravity. , the direction of gravity always faces the same direction with respect to the trunk when the posture is in a standing position. , see FIG. 2(E)) can be used as a coordinate space fixed with respect to the trunk. Also, when the posture is standing, sitting, or lying still, the acceleration value measured by the acceleration sensor only stops at a certain value, as illustrated in FIG. When the orientation of the acceleration sensor is unknown, the posture cannot be distinguished from the acceleration value. As exemplified in A), the body motion associated with walking causes a characteristic vibration in the acceleration value measured by the acceleration sensor. This can be detected, and thus the acceleration value when the posture is standing can be extracted as time-series data. Then, since it can be assumed that the direction of most of the acceleration vectors except for the characteristic vibration represented by the extracted acceleration value during walking is the direction of gravity, the direction of gravity is detected, and the direction of gravity is detected. The direction can be set as a reference coordinate space defined as one reference direction. Actually, since the median value of the acceleration value during walking in FIG. The direction of the acceleration vector can be detected.

Thus, in estimating the posture, the apparatus of the present embodiment first detects the walking section of the subject from the acceleration values measured by the acceleration sensor attached to the subject. An acceleration vector directed in the direction of gravity is determined as a reference vector from the acceleration values in the walking section. Then, the acceleration value measured by the acceleration sensor attached to the subject is converted into the coordinate value in the reference coordinate space in which the reference vector coincides with the predetermined reference direction, that is, the converted acceleration value. The acceleration vector determined by the acceleration value measured by the acceleration sensor is represented by the coordinate value in the coordinate space fixed to the trunk of the subject, and the converted acceleration value is used to move the subject pose estimation is performed. As will be described in detail below, in the identification of the walking section of the subject using the acceleration value measured by the acceleration sensor and the identification of the subject's posture using the converted acceleration value, is a supervised learning of the acceleration values measured by the acceleration sensor when the subject takes various postures and the identification result (correct value) of the subject's posture at the time of measurement of the acceleration value. A discriminator configured by machine learning using data may be used. In addition to the acceleration value measured by the acceleration sensor, the air pressure value measured by the pressure sensor is referred to as an input variable in the identification of the walking section of the subject and the identification of the posture of the subject. may be Since the air pressure value reflects the displacement of the trunk in the direction of gravity, it is possible to more accurately identify whether the subject's movement is horizontal or includes vertical movement (such as climbing stairs). becomes.

(2) Estimation of Posture In estimating the posture of a person, etc., in the device of this embodiment, as shown in FIG. Section identification (steps 5 and 6), reference vector determination (step 8), coordinate transformation of acceleration values (step 9), and orientation identification (steps 10 and 11) are executed in sequence, and the estimation results are output. you can Each process will be described below.

(i) Acquisition of acceleration values and atmospheric pressure values (steps 1 to 3)
The acceleration value and the air pressure value measured for posture estimation are measured and stored in time series by the acceleration sensor and the air pressure sensor, respectively, as described above. In the subsequent series of processing, from the time-series data of these measured values, the feature amount calculated for each unit called epoch as schematically shown in FIG. , the following processing is executed each time the measurement of the acceleration value and the atmospheric pressure value is completed one epoch. Note that one epoch may overlap with the preceding and following epochs (the ratio of overlap with the preceding and succeeding epochs may be changed arbitrarily), or may not overlap. As shown in the figure, for example, an epoch with a width of 4 seconds may be overlapped with the epoch before and after it by 2 seconds, and at the end of each epoch Ct1, Ct2, . . . A feature amount is extracted or calculated using the measurement data. Therefore, in this case, as shown in FIG. 4A, the measurement (step 1) and storage (step 2) of the acceleration value and the atmospheric pressure value are repeatedly executed until each epoch is completed. The next process is executed (step 4) each time the process is completed (step 3).

(ii) Identification of Walking Segments (Steps 5-6)
Each time an epoch of acceleration/air pressure measurement/storage is completed, if a reference vector, which will be described later, has not been determined (step 4), the feature values of the acceleration/air pressure are referred to in order to determine the reference vector. Then, the section in which the subject is walking is specified and extracted from the time-series data of the acceleration value and the air pressure value. Specifically, first, feature amounts for each epoch of the acceleration value and the atmospheric pressure value are calculated. Such a feature amount may be appropriately selected from the time-series data within each epoch of acceleration values and atmospheric pressure values. Typically, a statistic for each predetermined time of time-series data may be employed as the feature amount. Specifically, the feature values include, for example, the average value, median value, standard deviation, root mean square (RMS) of acceleration values of each axis, energy in a specific frequency region, peak position/height of autocorrelation function , coefficients for linear regression of atmospheric pressure waveforms, etc. can be used. Also, multiple feature amounts (for example, seven feature amounts) may be referred to in identifying the section during walking.

Thus, when the feature values of the acceleration value and the air pressure value are calculated, they are referenced to determine whether or not the subject was walking in the measurement section of the acceleration value and the air pressure value from which those feature values were obtained. is identified. Such identification may be performed using a classifier configured using parameters (classifier parameters) determined by a supervised machine learning technique using learning data as described later. The discriminator may be constructed according to any supervised machine learning approach such as random forest method, support vector machine method, gradient boosting method, k-NN method. The discriminator parameters are prepared in advance, stored in memory, and recalled for use, as described below. A specific process for identifying whether or not the subject is walking from the feature amount may be executed by any algorithm, and is typically executed using functions and modules prepared in a programming language. may be

(iii) Determination of Reference Vector (Step 8)
When the walking section is detected by identifying whether the subject was walking or not, it is determined whether or not the reference vector can be determined (step 7). As will be explained later, the reference vectors are selected in the representative directions of the acceleration vectors representing them with reference to the acceleration values in the three-axis directions in the walking section. is too short, the reference vector cannot be selected with high accuracy. Thus, a walk segment long enough to select a reference vector may be detected before determination of the reference vector is performed (i.e. Steps 1-6 may be directly repeated until .

Thus, when time-series data of acceleration values for a walking section of sufficient length are accumulated, representative values of acceleration values in three axial directions are selected as element values (bx, by, bz) of the reference vector. (see FIG. 2(D))). Specifically, such a representative value may be a median value, an average value, or a value obtained by principal component analysis of acceleration values in three axial directions (see FIG. 3A). As described above, the direction of the reference vector represented by the representative value of each acceleration value in the three axial directions is considered to be in the direction of gravity. The element values (bx, by, bz) of the vector Vr represent the vector in the direction of gravity g with coordinate values in the measurement coordinate space [xyz] of the acceleration sensor.

The reference vector Vr of the gravitational direction g is the measurement coordinate of the acceleration sensor regardless of the direction of the coordinate axis of the measurement coordinate space of the acceleration sensor (that is, regardless of the direction in which the acceleration sensor is mounted). In space, the same direction can be selected virtually always with respect to the trunk in a standing position. Therefore, as shown in FIG. 2(E), when the subject is walking and the posture of the trunk is standing, the direction of gravity is a predetermined reference direction (for example, the negative direction of the x-axis). direction) can be used as a reference coordinate space fixed with respect to the trunk as described above, and the acceleration values measured by the acceleration sensor (the coordinate values in the measurement coordinate space of the acceleration vector ) into coordinate values in the reference coordinate space, the acceleration value after the conversion is a value that varies with changes in the direction of the trunk regardless of the direction in which the acceleration sensor is worn.

により与えられ、基準ベクトルＶｒをｙ軸周りに回転する回転行列Ｒｙは、

により与えられ、かくして、基準ベクトルＶｒがｘ－ｙ平面上に延在するよう回転して得られたベクトルＶｒ’の座標値（ｂｘ’，ｂｙ’，ｂｚ’）は、

により与えられる。次いで、ベクトルＶｒ’がｚ軸周りにて回転角γだけ回転され、これにより、基準ベクトルＶｒがｘ軸の負方向に一致させられることとなる。ここで、回転角γは、

により与えられ、基準ベクトルＶｒをｚ軸周りに回転する回転行列Ｒｚは、

により与えられる。従って、基準ベクトルＶｒの座標値（ｂｘ，ｂｙ，ｂｚ）は、回転変換行列Ｒｚ・Ｒｙを乗ずることによって、ｘ軸の負方向に一致することになるので（この操作は、加速度センサの計測座標空間のｘ軸の負方向が基準ベクトルの方向に一致するように計測座標空間の座標軸を回転することに相当する。）、加速度センサの計測した３軸方向の加速度値Ａｍに対して、回転変換行列Ｒｚ・Ｒｙを乗ずることによって、３軸方向の加速度値Ａｍに対応する加速度ベクトルが基準座標空間に於ける座標値Ａｓにて表されることとなる。かくして、ステップ９に於いては、基準ベクトルＶｒの要素値（ｂｘ，ｂｙ，ｂｚ）を用いて、Ｒｙ、Ｒｚを決定し、更に、加速度センサの計測した３軸方向の加速度値Ａｍに対して、
Ａｓ＝Ｒｚ・Ｒｙ・Ａｍ …（６）
の変換処理を実行し、変換後加速度値（Ａｓ）が算出される。 (iv) coordinate transformation of acceleration values (step 9);
Once the reference vector is determined, it becomes possible to convert the acceleration value measured by the acceleration sensor into a coordinate value in a reference coordinate space in which the reference vector is set in a predetermined reference direction. In this respect, the conversion process corresponds to a rotational conversion operation that rotates the direction of the reference vector represented by the coordinate values (bx, by, bz) in the measurement coordinate space of the acceleration sensor in a predetermined reference direction. For example, assuming that the predetermined reference direction is set to the negative direction of the x-axis, that is, the (-1, 0, 0) direction, referring to FIGS. to (-1, 0, 0) (that is, to align the negative direction of the x-axis of the accelerometer with the direction of the reference vector), the reference vector Vr is set around the y-axis and around the z-axis as It should be rotated to Specifically, for example, first, the reference vector Vr is rotated about the y-axis by a rotation angle β and arranged on the xy plane. Here, the rotation angle β is

The rotation matrix Ry, which rotates the reference vector Vr around the y-axis, is given by

and thus the coordinate values (bx', by', bz') of the vector Vr' obtained by rotating the reference vector Vr so that it extends on the xy plane are

given by The vector Vr' is then rotated about the z-axis by a rotation angle γ, which aligns the reference vector Vr in the negative direction of the x-axis. Here, the rotation angle γ is

The rotation matrix Rz that rotates the reference vector Vr around the z-axis is given by

given by Therefore, the coordinate values (bx, by, bz) of the reference vector Vr are aligned in the negative direction of the x-axis by multiplying the rotation transformation matrix Rz·Ry (this operation is equivalent to the measurement coordinates of the acceleration sensor (corresponding to rotating the coordinate axes of the measurement coordinate space so that the negative direction of the x-axis of the space coincides with the direction of the reference vector), and the rotation transformation By multiplying the matrix Rz·Ry, the acceleration vector corresponding to the acceleration value Am in the three axial directions is represented by the coordinate value As in the reference coordinate space. Thus, in step 9, Ry and Rz are determined using the element values (bx, by, bz) of the reference vector Vr. ,
As=Rz.Ry.Am (6)
, and the converted acceleration value (As) is calculated.

As described above, in the converted acceleration values obtained from the acceleration values in the three axial directions measured by the acceleration sensor, the distribution range of the values can be separated more clearly depending on the posture of the subject. have been discovered. With reference to FIGS. 5A and 5B, the distribution range of each posture of the acceleration values in the three axial directions measured by the acceleration sensor ( Comparing A) with the distribution range (B) of the post-conversion acceleration value for each orientation, it is understood that the latter has a clearer boundary of the distribution range of the acceleration value depending on the orientation. Thus, as described above, it is expected that the accuracy of the estimation result will be improved when the posture is identified using the converted acceleration value.

(v) pose identification (steps 10-11)
When the converted acceleration value is obtained through the series of processes described above, the converted acceleration value is used to calculate the feature amount of the converted acceleration value and atmospheric pressure value for each epoch in the same manner as in step 5. (Step 10). Such a feature amount may be appropriately selected from the time-series data within each epoch of the converted acceleration value and atmospheric pressure value. Typically, a statistic for each predetermined time of time-series data may be employed as the feature amount. The features include, for example, the average value, median value, standard deviation, root mean square (RMS) of the converted acceleration values for each axis, energy in a specific frequency region, peak position and height of the autocorrelation function, atmospheric pressure waveform can be used, such as the coefficient when linear regression is performed. Also, multiple feature amounts (for example, seven feature amounts) may be referred to in posture identification.

Thus, when the feature values of the acceleration value and the air pressure value are calculated, they are referenced to identify the posture of the subject in the epoch when those feature values were obtained. Specifically, the postures identified may be standing, sitting, lying down, walking (horizontal), climbing stairs, descending stairs, and the like. Such identification, as in the case of walking section identification, uses a classifier configured using parameters (classifier parameters) determined by a supervised machine learning method using learning data as described later. may be implemented using The discriminator may be constructed according to any supervised machine learning approach such as random forest method, support vector machine method, gradient boosting method, k-NN method. The discriminator parameters are prepared in advance, stored in memory, and recalled for use, as described below. A specific process of identifying the posture of the subject from the feature quantity may be executed by any algorithm, typically using functions and modules prepared in a program language. Here, the obtained orientation identification result is the final orientation estimation result of the apparatus.

The series of processes described above describes the case of performing posture identification in real time. An estimation of pose over time may be performed. In that case, first, the section in which the subject is walking is identified from the obtained measurement data, the reference vector is determined using the data of the walking section, and the coordinate values of the reference vector are determined. may be used to perform coordinate transformation of the acceleration values in the obtained measurement data, and thus posture identification may be performed using the transformed acceleration values (the determination processing in step 7 is unnecessary. .).

(3) Machine learning for estimating posture In estimating the posture of the apparatus of the present embodiment, the parameters of the discriminator for discriminating the walking section of the subject and the posture of the subject Parameters of the discriminator for discrimination are determined in advance by a machine learning technique and stored in memory. As already mentioned, any discriminator may be constructed according to any supervised machine learning method such as random forest method, support vector machine method, gradient boosting method, k-NN method.

Referring to FIG. 4B, in the machine learning process, first, in preparation of learning data (step 21), acceleration values are measured by the acceleration sensor while the subject assumes various postures. and measure the air pressure value with the air pressure sensor, record the posture of the subject, use the measurement data of the acceleration value and the air pressure value as input data, and associate the posture when each measurement data is obtained as correct data. supervised learning data are prepared. The subject's postures included in such learning data may include standing, sitting, lying down, climbing stairs, and descending stairs in addition to horizontal walking.

Then, in determining the parameters of the discriminator for discriminating the walking section of the subject, from the measurement data of the acceleration value and the air pressure value in the above learning data, in the case of step 5 in FIG. Similarly, the acceleration feature amount and the pressure feature amount are calculated (step 22). Therefore, when the feature amount in the walking section is input, the parameters of the discriminator may be appropriately determined so that the walking section is identified (step 23). As already mentioned, the specific processing may be executed using any algorithm or any function or module prepared in any programming language, and the discriminator parameter is the algorithm to be used, the function of the programming language or It will be understood by those skilled in the art that the specific values of the discriminator parameters are appropriately set according to the specifications of the module since they are determined by the specifications of the module.

On the other hand, in determining the parameters of the discriminator for discriminating the posture of the subject, first, the correct value data in the learning data is referred to, and the section during walking of the subject is the acceleration value. are identified and extracted from the measurement data of (step 24). Then, in each of the extracted time-series data of the acceleration values in the three axial directions in the walking section of the subject, the representative value of the acceleration value is , are selected as the element values of the reference vector and the reference vector is determined (step 25). After that, the acceleration values in the learning data are subjected to the same coordinate conversion as in step 9, the converted acceleration values are calculated (step 26), and the converted acceleration values and the atmospheric pressure values are calculated in the same manner as in step 10. is calculated. Thus, when the feature values of the converted acceleration value and the atmospheric pressure value are obtained, using these feature values and the corresponding correct value data (posture), depending on the machine learning method adopted for posture identification, When the feature amount corresponding to each posture is input, determination of parameters of the discriminator may be appropriately executed so that the posture is output as a discrimination result (step 28). As already mentioned, the specific processing may be executed using any algorithm or any function or module prepared in any programming language, and the discriminator parameter is the algorithm to be used, the function of the programming language or It will be understood by those skilled in the art that the specific values of the discriminator parameters are appropriately set according to the specifications of the module since they are determined by the specifications of the module.

Thus, as described above, the discriminator parameters determined by machine learning are each stored in the memory and used when estimating the subject's posture by referring to the acceleration value and the air pressure value.

Verification Experiment In order to verify the effectiveness of the present invention described above, the following experiment was conducted. It should be understood that the following examples illustrate the effectiveness of the present invention and do not limit the scope of the invention.

An experiment was conducted to estimate the posture of a subject using the acceleration value and the pressure value measured by the acceleration sensor and the pressure sensor attached to the subject by the apparatus of the present embodiment. In the experiment, as shown in FIG. 6(A), the subject was equipped with housings containing an acceleration sensor and an air pressure sensor on each of the abdomen and the left and right flanks. ) and (C), the subjects who acted according to the protocol shown in (C) measured the acceleration values and air pressure values in three axial directions by each sensor, and recorded the posture at that time. The measurement was performed twice for each of the four subjects (24 sets (4×3×2 sets) of data groups were obtained as learning data). The following feature values extracted from the measurement data were used as the feature values. : Mean value of acceleration value for each axis, root mean square value, coefficient of regression line of atmospheric pressure value (7 features in total). A classifier for walking section identification and posture identification was configured using the random forest method. The discriminator classifies the state of the subject into walking or not in the horizontal direction by using the above-mentioned feature values calculated from the acceleration value and the air pressure value in identifying the walking section. In identifying the posture, the posture of the subject is classified into walking (horizontal), climbing stairs, descending stairs, and standing using the above feature values calculated from the converted acceleration value and atmospheric pressure value. It was configured to be categorized as one of position (static), sitting (static) and lying (static). The evaluation of the posture estimation is performed by estimating the posture of each subject using the measurement values of each sensor of each subject in the procedure of LOSOCV (Leave One Subject Out Cross Validation) below, and calculating the F value of the estimation result. (harmonic mean of precision and recall) was calculated. In the procedure of LOSOCV, the data of one of the three sensors attached to one subject is used as test data, and the discriminator parameters are determined by machine learning using the data of the remaining subjects, Under the determined discriminator parameter setting, the process of estimating the posture and calculating the F value of the estimation result using the test data is performed 12 times so that the data of all the sensors of all the subjects become the test data. repeated.

Referring to FIG. 7, in the results of posture estimation executed as described above, the acceleration values measured by the acceleration sensor are used as the input variables of the discriminator (without coordinate transformation), and the above Comparing the F value between the case of using the transformed acceleration value obtained according to the teaching of this embodiment (with coordinate transformation) and the latter, the accuracy of the estimation results for all postures except for climbing stairs is improved. did. In particular, in the supine position, the method according to the present invention gave an F value of 1, which indicates that the estimation was perfect without a single error. In addition, it is understood that the improvement of the performance in the sitting position and the standing position is remarkably improved, and the method of the present invention achieves improvement in the accuracy of posture estimation. (For stair climbing, although the performance of the method of the present invention is lower, the decrease is slight, and the accuracy of walking and climbing stairs is improved, so in estimating the posture of the subject while walking It is judged that the accuracy is generally improved even if the

Although the above description has been made with respect to the embodiments of the present invention, many modifications and changes can be easily made by those skilled in the art, and the present invention is limited only to the above-exemplified embodiments. It will be clear that the invention is non-limiting and can be applied to a variety of devices without departing from the concept of the invention.

For example, the identification of the walking section and/or the identification of the posture is not based on a classifier configured by machine learning, but the acceleration feature value, the converted acceleration feature value, or the pressure feature value is experimentally or theoretically set as appropriate. by determining whether it falls within a specified threshold range. For the coordinate transformation of the acceleration values, if the coordinate values in the measurement coordinate space of the acceleration sensor can be transformed into the coordinate values in the reference coordinate space in which the reference vector coincides with the predetermined reference direction, the rotation transformation matrix exemplified above can be used. In addition to the method used, a method using a quaternion or the like may be used.

Claims (1)

A device for estimating the posture of a human, an animal, or a walking robot (hereinafter referred to as "human, etc."),
an acceleration sensor that is attached to the trunk of a person and measures acceleration values in three axial directions in chronological order and an air pressure sensor that measures the air pressure value in the trunk;
Among the acceleration values and pressure values in the three-axis directions, the acceleration values and pressure values in the three-axis directions when the person or the like is in a substantially horizontal walking state are extracted; reference vector determination means for determining a reference vector representing a representative direction of acceleration when the person or the like is in a substantially horizontal walking state, using the acceleration value and the air pressure value;
converted acceleration value determination means for converting the acceleration values in the three-axis directions into converted acceleration values, which are coordinate values in a reference coordinate space in which the direction of the reference vector coincides with a predetermined reference direction;
posture estimation means for estimating the posture of the person or the like based on the converted acceleration value and the atmospheric pressure value;
The reference vector determination means includes a first discriminator configured using machine learning,
The first discriminator comprises time-series data of acceleration values and air pressure values when the person or the like is in various states including the substantially horizontal walking state, and time-series data of the acceleration values and air pressure values. Using teacher data consisting of time-series correct data of the posture of the person, etc. corresponding to the data, selected from the time-series data of the acceleration value and the air pressure value that reflects the characteristics of the posture of the person, etc. Determining a parameter for outputting the posture of the correct data as an identification result when the feature amount is given as input data,
The first discriminator constructed using the parameters is provided with feature amounts of acceleration values in three axial directions measured in time series by the acceleration sensor and air pressure values measured in time series by the air pressure sensor. By inputting, it is configured to identify and extract the data of the section in which the posture of the person is in the substantially horizontal walking state in the time-series data of the acceleration value and the air pressure value,
the posture estimation means includes a second discriminator configured using machine learning;
The second discriminator comprises time-series data of acceleration values in three axial directions measured by the acceleration sensor and air pressure values measured by the air pressure sensor when the person or the like is in various states. time-series correct data of the posture of a person corresponding to the time-series data of the acceleration value and the air pressure value of , and calculating a feature amount selected as reflecting the characteristics of the posture of the person, etc. from the time-series data of the converted acceleration value and the atmospheric pressure value, Determining a parameter so as to output the posture of the correct data as an identification result using the feature values as input data,
Converted acceleration values obtained from acceleration values in three axial directions measured in time series by the acceleration sensor to the second discriminator configured using the parameters and pressure values measured by the pressure sensor It is configured to identify and estimate the posture of the person by inputting the feature amount of
Device.

Images