JP6648435B2 - Discrimination device, discrimination method, program, model generation device, and model generation method - Google Patents

Description

  The present invention relates to a determination device, a determination method, a program, a model generation device, and a model generation method.

  2. Description of the Related Art In recent years, a technique has been developed for determining the movement of a moving object based on a signal derived from the movement of a moving object such as a person detected by a non-contact sensor. For example, Patent Literature 1 below discloses an invention relating to a control circuit including at least one sensitivity correction unit for correcting the sensitivity of a sensor in processing a signal detected by the sensor. Note that the sensitivity correction means adjusts the amplification factor of the amplifier circuit in the control circuit, suppresses a relatively fast-moving frequency component using a first-order lag filter, or uses a second-order lag filter, for example. And a function of suppressing frequency components of relatively slow movement. In such a configuration, by switching the sensitivity correction means, it is possible to adjust the sensitivity for determining the presence or absence of a person.

JP 2010-230492 A

  According to the technology disclosed in Patent Document 1, by adjusting the sensitivity in accordance with the movement of the moving object and the environment of the moving object, it is possible to improve the accuracy of determining whether or not a person is present. However, in an environment where the sensor is installed, the sensor can detect not only complicated movement of a person but also various disturbances (for example, electric noise and mechanical movement). The signal detected by the sensor may include a plurality of disturbances. Therefore, there is a limit in discriminating the motion of a moving object in an environment that may include various disturbances by simply switching the sensitivity correction unit disclosed in Patent Document 1 and adjusting the sensitivity.

  Therefore, the present invention has been made in view of the above-described problem, and an object of the present invention is to provide a new and highly accurate moving object that can determine the motion of a moving object from signals including various disturbances. An object of the present invention is to provide an improved discriminating apparatus.

In order to solve the above problem, according to an aspect of the present invention, based on a sensor signal indicating a movement of a moving body detected by a sensor, a feature amount calculating unit that calculates a feature amount of the movement of the moving body, Using a regression learning model generated based on a plurality of data pairs of a feature amount of the motion and an evaluation value corresponding to the predetermined motion, the motion of the moving object calculated by the feature amount calculation unit An evaluation value calculation unit that calculates an evaluation value corresponding to the movement of the moving object, and a threshold value that is compared with the evaluation value corresponding to the movement of the moving object to determine the movement of the moving object according to the feature amount of And an adjustment unit for adjusting , wherein the evaluation value corresponding to the predetermined movement is a plurality of pairs of a characteristic amount of the predetermined movement and an evaluation value set in advance corresponding to the predetermined movement. Multiple regressions generated based on data Depending on the error between a plurality of determination results corresponding to the predetermined motion is determined using a model is determined, the determination device is provided.

  The regression learning model is a regression learning model generated by support vector regression, and the evaluation value calculation unit includes a support vector that configures the regression learning model generated by the support vector regression, and a motion vector of the moving object. An evaluation value corresponding to the movement of the moving object may be calculated using the feature amount.

  The feature amount calculating unit may calculate the feature amount based on a waveform of the sensor signal.

  The feature amount calculation unit may calculate the magnitude of the amplitude of the sensor signal as the feature amount.

  The feature amount calculation unit may calculate a frequency of the waveform of the sensor signal in a predetermined section as the feature amount.

  The adjustment unit may adjust the threshold value stepwise.

  The predetermined movement may include a movement of a living body, and the determination of the movement of the moving body may include determining whether the movement of the moving body is the movement of the living body.

  The sensor may be a Doppler sensor.

According to another embodiment of the present invention, there is provided a method for solving the above-mentioned problem, comprising: calculating a feature amount of the movement of a moving object based on a sensor signal indicating the movement of the moving object detected by a sensor; Using a regression learning model generated based on a plurality of data pairs of the feature amount of the motion and the evaluation value corresponding to the predetermined motion, the moving object is determined in accordance with the feature amount of the motion of the moving object. in calculating the evaluation value that corresponds to the motion, see contains the steps of: adjusting the threshold to be compared with the evaluation value corresponding to the movement of the moving object for determination of movement of the moving object, the predetermined The evaluation value corresponding to the motion is a plurality of regression learning models generated based on a plurality of data sets each of which is a combination of the feature amount of the predetermined motion and an evaluation value set in advance corresponding to the predetermined motion. The location determined using Is determined according to the error between a plurality of determination results corresponding to the motion determination method is provided.

According to another embodiment of the present invention, there is provided a computer configured to calculate a feature amount of the motion of a moving object based on a sensor signal indicating the motion of the moving object detected by a sensor. An amount calculator, and a regression learning model generated based on a plurality of data sets each of which includes a feature amount of the predetermined motion and an evaluation value corresponding to the predetermined motion, and is calculated by the feature amount calculator. An evaluation value calculation unit that calculates an evaluation value corresponding to the motion of the moving object according to the feature amount of the motion of the moving object, and the evaluation value corresponding to the motion of the moving object for determining the motion of the moving object And an adjustment unit that adjusts a threshold value to be compared with, the evaluation value corresponding to the predetermined movement is a feature amount of the predetermined movement, and an evaluation value set in advance corresponding to the predetermined movement. And multiple data Is determined according to the error between a plurality of determination results corresponding to the predetermined motion is determined using a plurality of regression learning model generated Zui, the program is provided.

According to another embodiment of the present invention, there is provided a feature value for calculating a feature value of a predetermined motion based on a sensor signal indicating a predetermined motion of a moving object detected by a sensor. A calculation unit, using a plurality of regression learning models generated based on a plurality of data pairs of a feature amount of the predetermined motion and a reference evaluation value set in advance corresponding to the predetermined motion. A regression learning model evaluation value determination unit that determines an evaluation value corresponding to the predetermined motion in accordance with an error between a plurality of determination results corresponding to the determined predetermined motion; and a feature amount of the predetermined motion. And a model generation unit configured to generate a regression learning model based on a plurality of data sets each of which includes an evaluation value corresponding to the predetermined motion.

To solve the above problem, according to another aspect of the present invention, a step of calculating a feature amount of the predetermined movement based on a sensor signal indicating a predetermined movement by a moving object detected by a sensor; It is determined using a plurality of regression learning models generated based on a plurality of data pairs of the feature amount of the predetermined motion and a reference evaluation value set in advance corresponding to the predetermined motion. Determining an evaluation value corresponding to the predetermined motion according to an error between a plurality of determination results corresponding to the predetermined motion; a characteristic amount of the predetermined motion; and an evaluation value corresponding to the predetermined motion. Generating a regression learning model based on a plurality of data pairs of values.

  As described above, according to the present invention, it is possible to determine the motion of a moving object with higher accuracy from signals including various disturbances.

It is a figure showing the outline of the discrimination system concerning one embodiment of the present invention. FIG. 1 is a block diagram illustrating a configuration example of a model generation device according to a first embodiment of the present invention. FIG. 3 is a diagram illustrating an example of a data list stored in a signal DB according to the embodiment. FIG. 4 is a diagram for describing the number of zero crossings of a signal data waveform. FIG. 4 is a diagram illustrating an example of an evaluation value allocation processing flow by an evaluation value generation unit for each data set stored in the data list illustrated in FIG. 3. It is a figure showing an example of an evaluation list stored in an evaluation value DB concerning the embodiment. 4 is a flowchart illustrating an operation example of the model generation device according to the same embodiment. It is a block diagram showing an example of composition of a discriminating device concerning the embodiment. 4 is a flowchart illustrating an operation example of the determination device according to the same embodiment. It is a block diagram showing an example of composition of a model generation device concerning a 2nd embodiment of the present invention. FIG. 3 is a block diagram illustrating a configuration example of an evaluation value generation unit and a database storing various data used in the evaluation value generation unit according to the embodiment. It is a figure showing an example of data list 1 (no disturbance) stored in feature DB concerning the embodiment. FIG. 4 is a diagram illustrating an example of a data list 2 (with disturbance) stored in a feature amount DB according to the embodiment. It is a figure showing an example of reference evaluation value list 1 (no disturbance) stored in reference evaluation value DB concerning the embodiment. It is a figure showing an example of reference evaluation value list 2 (with disturbance) stored in reference evaluation value DB concerning the embodiment. It is a figure for explaining evaluation value decision processing by an evaluation value decision part concerning the embodiment. It is a graph which shows the distribution of the unmanned rate calculated by each reference regression learning model. 5 is a flowchart illustrating an operation example of an evaluation value generation unit according to the same embodiment. It is a flow chart which shows a processing flow of a reference model generation part concerning the embodiment. It is a flow chart which shows a processing flow of a discrimination verification part concerning the embodiment. It is a flow chart which shows a processing flow of an evaluation value deciding part concerning the embodiment. FIG. 2 is a block diagram illustrating an example of a hardware configuration of the information processing apparatus according to the embodiment of the present invention.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the specification and the drawings, components having substantially the same functional configuration are denoted by the same reference numerals, and redundant description is omitted.

<1. Overview of discrimination system>
A discriminating system 1 according to an embodiment of the present invention applies a regression learning model generated based on a predetermined motion by a model generating device 10 to a discriminating device 20 to discriminate a motion of a moving object detected by a sensor 2. This is a system that is determined by the device 20. Hereinafter, an outline of the determination system 1 will be described.

  FIG. 1 is a diagram showing an outline of a discrimination system 1 according to an embodiment of the present invention. The configuration of the determination system 1 differs depending on the processing to be performed. For example, referring to FIG. 1, when performing a model generation process for generating a regression learning model, the discrimination system 1 has a configuration including a sensor 2 and a model generation device 10. On the other hand, when performing the process of determining the motion of the moving object, the determination system 1 has a configuration including the sensor 2 and the determination device 20. Note that the sensor 2 and the model generating device 10 may be integrated to form a model generating device, or the sensor 2 and the determining device 20 may be integrally configured to form a determining device.

(Sensor)
The sensor 2 is a non-contact type sensor installed on a ceiling of a room or the like, as shown in FIG. The sensor 2 emits a radiation wave such as light, an electromagnetic wave, or a sound wave, for example, toward the inside of a room, which is a detection area, and uses a moving body (for example, a person P, an air conditioner AC, or a curtain CU in FIG. 1). A radio wave sensor that receives the reflected wave may be used. The radio wave sensor may be, for example, a Doppler sensor. At this time, the frequency of the reflected wave changes from the frequency of the radiated wave due to the Doppler effect caused by the motion of the object such as vibration. The sensor 2 generates a signal having a difference frequency between the frequency of the radiation wave and the frequency of the reflected wave. The sensor 2 may be equipped with, for example, a quadrature detection method. In this case, the sensor 2 may generate two kinds of signals of a cosine wave component (I component) and a sine wave component (Q component). The sensor 2 outputs the generated signal to either the model generation device 10 or the discrimination device 20.

  The sensor 2 can be installed at an arbitrary position as long as the movement of the moving body can be detected. In the example shown in FIG. 1, the sensor 2 is configured such that the transmitting unit for the radiation wave and the receiving unit for the reflected wave are integrated, but the sensor 2 has a configuration in which the transmitting unit and the receiving unit are separated. May be realized by: Further, the radiation wave emitted by the sensor 2 may be a wave of an arbitrary frequency band as long as the Doppler effect can be generated by the vibration of the object. For example, the radiation wave is preferably a wave in a frequency band in which a motion of a person can be detected. Further, the environment in which the sensor 2 is installed in the model generation processing and the determination processing may be different. For example, in the model generation processing, since it is necessary to detect a signal of a single movement of a moving object, the sensor 2 may be provided in a room such as a soundproof room that can block disturbance. On the other hand, in the determination process, the installation location of the sensor 2 is not particularly limited as long as it is required to detect the movement of a moving body such as a person. For example, the sensor 2 may be installed not only in a room but also in a section partitioned by a partition, a fence, or the like. The sensor 2 is installed inside a building or a vehicle body such as an exercise facility, a medical facility, a nursing facility, a nursing home, an attraction facility, a criminal facility, a house, a hotel, a condominium, a building, a car, a railcar, and a ship. Is also good. Further, the type of the moving object is not limited to the person, furniture, home appliance, or the like as described above.

  In the present embodiment, the sensor 2 is preferably a radio wave sensor (Doppler sensor). For example, when the sensor 2 is an infrared sensor, the movement of a moving object can be detected. However, when the moving object is stationary, it is difficult to detect the moving object. On the other hand, when the sensor 2 is a radio wave sensor, even when the moving body is stationary, vibration of the trunk due to respiration of the moving body can be detected. Therefore, it is easy to detect a state in which the moving object is stationary.

(Model generation device)
The model generation device 10 is a device used for a model generation process of a predetermined moving body motion. The model generation device 10 may be realized by, for example, one or a plurality of information processing devices on a network. More specifically, the model generation device 10 may be realized by a server, a PC (Personal Computer), or the like. The model generation device 10 acquires a signal derived from the movement of a predetermined moving object from the sensor 2 and generates an evaluation value corresponding to the movement of the predetermined moving object. Then, the model generation device 10 generates a regression learning model based on the signal of the movement of each predetermined moving object and the evaluation value corresponding to the movement of each moving object. The generated regression learning model is used in the discrimination device 20.

  Here, in this specification, the movement of a moving object includes not only the movement of a person, but also movement of a person other than a person, such as movement of a home appliance, vibration of the ground, or vibration of furniture due to wind. Further, the movement of the person may be further subdivided. For example, a motion while sleeping or a motion while sitting may be included in the motion of the person. The evaluation value is a value used for determining the motion of a moving object, and is associated with the motion of each moving object.

(Determination device)
The discriminating device 20 is a device used for discriminating the motion of a specific moving object. The determination device 20 may be realized by, for example, one or a plurality of information processing devices on a network. More specifically, the determination device 20 may be realized by a server, a PC, or the like. The determination device 20 determines the motion of the moving object by using a signal derived from the motion of the moving object detected by the sensor 2 and the regression learning model generated by the model generation device 10. For example, the determination device 20 may determine the motion of the moving object by comparing an evaluation value obtained by using a regression learning model with respect to the motion of the moving object detected by the sensor 2 with an adjustable threshold. Good. For example, the determination device 20 can transmit a determination result to an external device or the like via a wired or wireless communication via a communication unit (not shown). Thus, for example, the external device can perform a process using the determination result. By the processing using the determination result, for example, it is possible to grasp the presence or absence of the person P. It should be noted that the processing using the determination result can be performed in the main body of the determination device 20 instead of the external device as described above.

  Heretofore, technology for determining the presence or absence of a person has been developed in the past. For example, according to the invention disclosed in Japanese Patent Application Laid-Open No. 2010-230492, the sensitivity for determining the presence or absence of a person can be adjusted. The sensitivity could only be adjusted for signals containing the above factors. Therefore, it has been difficult to adjust the sensitivity for improving the discrimination accuracy for complex signals including various disturbances.

  In the discrimination system 1 of the present invention, the model generation device 10 generates a regression learning model based on a plurality of predetermined scenarios, and the discrimination device 20 determines the motion of a moving object using the regression learning model. In such a configuration, the similarity (evaluation value) between the motion of the moving object and a predetermined scenario which is a component of the regression learning model is calculated by applying the motion of the moving object to the regression learning model. That is, by applying a regression learning model to a signal including a plurality of disturbances, it is possible to represent a similar motion as a continuous evaluation value. Therefore, for example, by setting a threshold for determination as a sensitivity adjustment element within a range of possible evaluation values, it is possible to determine whether or not the movement of a moving object is a predetermined movement by comparing with a threshold. it can. If the evaluation value is a continuous value, the threshold value can also be set continuously, so that the sensitivity for determining the predetermined movement can be finely adjusted. Therefore, the sensitivity can be adjusted even for a signal in which a plurality of disturbances coexist, in which the sensitivity cannot be adjusted conventionally. Therefore, the accuracy of determining the motion of the moving object can be further improved.

<2. First Embodiment>
Hereinafter, the model generation device 10 and the discrimination device 20 according to the first embodiment of the present invention will be described.

[2.1. Model generation device]
(1) Configuration FIG. 2 is a block diagram illustrating a configuration example of the model generation device 10 according to the first embodiment of the present invention. Referring to FIG. 2, the model generation device 10 includes a signal acquisition unit 110, a signal DB (database) 111, a filter unit 120, a feature amount calculation unit 130, an evaluation value generation unit 140, an evaluation value DB 141, a model generation unit 150, A regression learning model DB 151 is provided. The model generation device 10 includes a plurality of sets each including a feature space including a feature vector calculated from a signal derived from a known predetermined motion detected by the sensor 2 and an evaluation value corresponding to the predetermined motion. Generate a regression learning model from the data. The generated regression learning model is used in the discrimination device 20. Hereinafter, the model generation device 10 according to the present embodiment will be described.

(Signal acquisition unit)
The signal acquisition unit 110 acquires signal data output from the sensor 2 via a communication unit or an input unit (not shown). In the present embodiment, the sensor 2 is a Doppler sensor. The signal data output from the sensor 2 is data composed of a time-series digital signal, and has two components, I (n) as an I component and Q (n) as a Q component. The signal data is collected by the sensor 2 for a specified time. The prescribed time may be, for example, 10 minutes. The sampling frequency of the signal data is not particularly limited, but is preferably 100 Hz or more in order to determine the motion of the person. The signal acquisition unit 110 acquires a signal that has detected a predetermined movement as signal data, and outputs the acquired signal data to the signal DB 111.

(Signal DB)
The signal DB 111 is a database that stores the signal data acquired by the signal acquisition unit 110. The database has a plurality of data lists, and the data list is constituted by a plurality of data sets. FIG. 3 is a diagram illustrating an example of a data list stored in the signal DB 111. Referring to FIG. 3, the data list 1000 includes data files (A1.dat, B2.dat,..., C6.dat) of signal data derived from a predetermined motion acquired by the signal acquisition unit 110, , Attributes, and scenarios as a data set. The ID is an index assigned to each acquired signal data. The attribute is an item indicating the type of movement of the moving object, and a plurality of attributes can be set in one data list. In the data list 1000 shown in FIG. 3, for example, three types of attributes, "unmanned (no disturbance)", "unmanned (disturbed)", and "manned" are set. Here, “unmanned (no disturbance)” means a state in which no motion by a moving object including a person is detected. “Unattended (with disturbance)” means that there is no person, but the motion of another moving object can be detected. Further, “maned” means a state where only movement by a person can be detected.

  The scenario is an item indicating a specific moving object set in advance in each attribute and the movement of the moving object. The predetermined movement set as the scenario is a movement selected in advance by a user of the model generation device 10 or the like. In the data list 1000 shown in FIG. 3, “air conditioner” and “curtain” are set as scenarios for the attribute “unmanned (with disturbance)”. This means that the signal data acquired as the attribute of "unmanned (with disturbance)" is signal data derived from the vibration of the air conditioner and the curtain. Similarly, for the attribute “maned”, “sit”, “sleep”, and “walk” are set as scenarios. This means that the signal data acquired as the “maned” scenario is signal data derived from a sitting motion, a sleeping motion, and a walking motion of a person.

  As described above, the signal DB 111 stores a data list including a plurality of data sets including IDs, attributes, scenarios, and signal data corresponding to the scenarios. Note that these data sets and data lists may be appropriately added, modified, or deleted by a user of the model generation device 10 via an input unit or a communication unit (not shown). The generated data list is output to the filter unit 120.

(Filter section)
The filter unit 120 performs a filtering process on the signal data included in the data list acquired from the signal DB 111. The signal data included in the data list contains a lot of high frequency component noise unrelated to the motion of the moving object. Therefore, the filter unit 120 removes or reduces these noises by performing LPF (Low Pass Filter) processing.

First, the filter unit 120 acquires the signal data I _m (n) and Q _m (n) of ID = m stored in the data list (n = 1 to N). Here, N indicates the number of samplings included in the signal data.

_{Here, I} m (n) and _Q m (n) is second-order IIR (Infinite Impulse Response: Infinite Impulse Response) LPF processing by the filter may be performed. The cutoff frequency in the LPF processing may be, for example, 50 Hz. By such LPF processing, high-frequency noise components included in signal data can be removed or reduced. Here, the signal data after the LPF processing by the IIR filter is performed is referred to as I _m ′ (n) and Q _m ′ (n), respectively.

The LPF processing by the IIR filter is performed by repeatedly performing the following equations (1) and (2) in order from n = 1 to N. Here, x (n) is an input value and y (n) is an output value. For example, in the following formula (1), the x (n) _I m (n) and _Q m (n) is substituted. In the following equation (2), I _m ′ (n) and Q _m ′ (n) are output as y (n). Further, k, a ₀ , a ₁ , a ₂ , b ₁ , and b ₂ are coefficients set in the design stage according to the frequency band.

  Note that A (0) = A (-1) = 0.

After I _m '(n) and _{Q m'} (n) is output, the A (n-2) is assigned the value of A (n-1), the A (n-1) A ( n ) Is substituted. The LPF processing is completed by performing this substitution processing from n = 1 to N. The data list including the signal data on which the LPF processing has been completed is output to the feature amount calculation unit 130.

  Although the filter unit 120 according to the present embodiment uses an IIR filter as the LPF processing, the present invention is not limited to this example. For example, the filter unit 120 performs LPF processing by transforming the signal into the frequency domain by Fourier transform, removing the frequency in a frequency band outside a predetermined range, and then transforming the signal again into the time domain by inverse Fourier transform. Is also good. Further, the filter unit 120 may perform the LPF processing by using a known technique other than the above-described technique.

(Feature calculation unit)
The feature amount calculation unit 130 calculates a feature amount of the signal data after the LPF processing stored in the data list acquired from the filter unit 120. Here, the characteristic amount is a characteristic amount calculated based on the waveform of the signal data. For example, the feature amount includes a feature amount calculated based on the amplitude of the waveform of the signal data, the frequency component of the waveform, and the like. In the present embodiment, the feature value calculation unit 130 calculates the amplitude feature value and the frequency feature value of the waveform of the signal data as the feature value.

In the present embodiment, the amplitude feature amount is an amplitude value at time n of the signal data after the LPF processing. The amplitude feature amount V _m (n) is calculated by the following equation (3). Further, the amplitude feature amount V _m (n) may be a statistic such as an average value or a maximum value of the amplitude values of the respective sampling data in the unit period starting from the time n.

The frequency feature quantity may be, for example, the number of zero crossings of the waveform of the signal data in the unit period. FIG. 4 is a diagram for explaining the number of zero crossings of the waveform of the signal data. As indicated by a circle in FIG. 4 crossing, zero crossing number, the waveform of the signal data I _m '(n) or Q _m' (n) is the time axis in the unit period T _a and (amplitude = zero value) It is the number of times to do. Specifically, the zero crossing number, among the L sampling data in the unit period T _a for the time n and the top is a number satisfying the following formulas (4) and (5). Unit interval T _a may be, for example, one second, in which case, L is 100 (i.e., sampling frequency = 100 Hz of the time-series digital signal data outputted from the sensor 2) may be used.

In the above formula unit period _{T a} for the time n and the top (4) and Formula (5) the number of i satisfying each _f 1 m When (n) and _f 2m (n), the zero crossing number _f m (n) Is f _1m (n) + f _2m (n).

The feature quantity calculator 130, the above in addition to the amplitude characteristic amount _V m (n) and zero crossing number _f m (n), further signal data _{I m} '(n) or _{Q m'} waveform (n) May be calculated based on. For example, the feature amount calculation unit 130 calculates the number of local maximum points or local minimum points of the waveform in a unit period, statistics such as the variation or average value of the local maximum points or minimum points in a plurality of unit periods, or the power spectrum in the frequency domain. , May be calculated as a feature amount. By increasing the types of the feature amounts, the features of the waveform of the signal data can be more finely extracted.

Feature amount calculation unit 130, first, the L-number of sampling data included from the signal data to the unit interval T _a and J pairs extracted, respectively calculates the feature quantity for each sampling data unit sections T _a. In other words, C × J features are extracted from one signal data (C is the number of types of features). Next, the feature amount calculation unit 130 calculates a feature amount for all the signal data included in the acquired data list. Then, the feature amount calculation unit 130 generates a feature vector having the calculated feature amount as a component. The feature vector forms a vector of F _m (j) = [V _m (j), f _m (j)]. Here, j indicates the oldest sampling time among the L sampling data as described above. The feature amount calculation unit 130 performs the above-described processing on all the signal data stored in the data list, and generates a feature vector F _m (j) for each signal data. Then, the feature amount calculation unit 130 outputs the generated feature vector F _m (j) to the model generation unit 150.

(Evaluation value determination unit)
The evaluation value generation unit 140 generates an evaluation value corresponding to each scenario in the data list stored in the signal DB 111. Here, the evaluation value is a parameter used to determine a motion by a specific moving object. For example, when the evaluation value is set in the range of 0 to 1, for a scenario that is a motion by a specific moving object (here, a motion of a person), 1 is set as an evaluation value, and for a scenario that is not a motion of a person, , A value other than 1 may be set as the evaluation value. Here, for example, an evaluation value of 0 may be set for a scenario that is clearly not a motion of a person, but an intermediate value between 0 and 1 may be set as an evaluation value for a scenario that is confusing to the motion of a person. . Also, different evaluation values may be set for each scenario, even if the attributes are the same. In this way, by setting the evaluation value in association with the scenario, it is easy to make an adjustment for determining a motion that is confusing with a motion by a specific moving object such as a person in a subsequent determination process.

  The evaluation value generation unit 140 may, for example, assign a prescribed evaluation value according to the attribute of the data list stored in the signal DB 111. FIG. 5 is a diagram illustrating an example of an evaluation value allocation processing flow by the evaluation value generation unit 140 for each data set stored in the data list 1000 illustrated in FIG. Referring to FIG. 5, first, the evaluation value generation unit 140 acquires a data list stored in the signal DB 111 (S101). The evaluation value generation unit 140 reads attributes of the data list for all scenarios, and sets evaluation values according to the read attributes.

The evaluation value generation unit 140 determines whether the read attribute is “unmanned (no disturbance)” (S103). When the read attribute is “unmanned (no disturbance)” (S103 / YES), the evaluation value generation unit 140 sets the evaluation value of the data set to 0.0 (S105). Next, the evaluation value generation unit 140 determines whether the read attribute is “unattended (with disturbance)” (S107). When the read attribute is “unattended (with disturbance)” (S107 / YES), the evaluation value generation unit 140 sets the evaluation value of the data set to 0.5 (S109). If the read attribute is neither “unmanned (no disturbance)” nor “unmanned (disturbed)” (“maned” in the example shown in FIG. 3) (S107 / NO), the evaluation value generating unit 140 The evaluation value for the set is set to 1.0 (S101). Based on the above processing flow, the evaluation value generation unit 140 can allocate an evaluation value for each data set. When there are M scenarios, the value set in step S105, S109, or S111 is substituted for the evaluation value Y _m (j) corresponding to the m-th scenario.

  In the present embodiment, the evaluation value is assigned to each scenario by the evaluation value generation unit 140, but the present invention is not limited to this example. For example, the evaluation value generation unit 140 may allocate an evaluation value input from an input unit (not shown) to each scenario. Thus, the user of the model generation device 10 can freely set the evaluation value for each scenario. Further, the range of the evaluation value is not limited to the range (0.0 to 1.0) shown above, and can be set freely.

  After allocating the evaluation value to each scenario, the evaluation value generation unit 140 associates the ID of each scenario with the evaluation value, and outputs the scenario to the evaluation value DB 141 as an evaluation value list.

(Evaluation value DB)
The evaluation value DB 141 is a database that stores the evaluation value list acquired from the evaluation value generation unit 140. FIG. 6 is a diagram illustrating an example of the evaluation list stored in the evaluation value DB 141. Referring to FIG. 6, in the evaluation list 2000 stored in the evaluation value DB 141, the evaluation values are stored as data sets together with IDs, attributes, and scenarios. Since the ID, the attribute, and the meaning of the scenario are the same as those of each element included in the data list stored in the signal DB 111, the description is omitted. The evaluation value list stored in the evaluation value DB 141 is output to the model generation unit 150. The value of the evaluation value stored in each data set may be set, added, modified, or deleted by a user of the model generation device 10 via an input unit, a communication unit, or the like (not shown). In that case, the model generation device 10 does not necessarily need to include the evaluation value generation unit 140.

(Model generation unit)
The model generation unit 150 generates a regression learning model based on a plurality of sets of data including the feature vector output from the feature amount calculation unit 130 and the evaluation values included in the evaluation value list acquired from the evaluation value DB 141. I do. Specifically, first, the model generation unit 150 generates a feature space including J extracted data and feature vectors F ₁ (j) to F _M (j) corresponding to M scenarios, and generates M pieces of feature space. A regression learning model is generated based on data that is a set of evaluation values Y ₁ (j) to Y _M (j) corresponding to the scenario. Since the evaluation value is assigned by the evaluation value generation unit 140 for each corresponding scenario, the value of the evaluation value Y _m (j) corresponding to the same scenario (that is, signal data) is the same regardless of the value of j. is there.

  Here, in the present embodiment, the regression learning model can be generated using, for example, SVR (Support Vector Regression). The SVR can be realized based on, for example, the contents disclosed in SMOLA, Alex J. and Bernhard SCHOELKOPF, “A Tutorial on Support Vector Regression”, 1998. The regression learning model is not limited to the above example, and may be generated by a known regression analysis method such as a kernel method or a multiple regression analysis.

  The regression learning model generated by the SVR is composed of a support vector matrix s composed of K elements, their weight vectors c, and thresholds b. Specifically, the support vector matrix s is defined by a matrix represented by the following equation (6), and the weight vector c is defined by a vector represented by the following equation (7). The threshold b is a value uniquely determined. Further, the number K of elements of the support vector matrix s is uniquely determined in the process of calculation by SVR.

  In addition, it is also possible to generate a regression learning model by mapping data constituted by the feature space and the evaluation value to a non-linear space by using a kernel trick. In the present embodiment, a Gaussian kernel function represented by the following equation (8) may be used as a kernel trick. In this case, the coefficient σ defining the spread of the Gaussian function is an element constituting the regression learning model.

  The model generation unit 150 outputs to the regression learning model DB 151 a regression learning model generated from the data constituted by the feature spaces and the evaluation values corresponding to the M scenarios using a regression analysis such as SVR.

(Regression learning model DB)
The regression learning model DB 151 is a database that stores the regression learning model output from the model generation unit 150. The regression learning model DB 151 stores, for example, a support vector matrix s, a weight vector c, a threshold value b, and a coefficient σ of a Gaussian kernel function of the regression learning model generated by the above equations (6) to (8). Is done. One or a plurality of regression learning models may be stored in the regression learning model DB 151. For example, a regression learning model corresponding to the type of the moving object or the type of the motion of the moving object determined by the determination device 20 described later may be stored in the regression learning model DB 151.

  The data related to the regression learning model stored in the regression learning model DB 151 is copied and stored in the regression learning model DB 231 of the discriminating apparatus 20. Accordingly, the determination device 20 can determine the motion of the moving object using the regression learning model generated by the model generation device 10.

(2) Operation FIG. 7 is a flowchart illustrating an operation example of the model generation device 10 according to the first embodiment of the present invention. First, the signal acquisition unit 110 acquires a signal derived from a predetermined movement detected by the sensor 2 as signal data (S201). The signal acquisition unit 110 outputs a plurality of pieces of signal data derived from a plurality of predetermined movements to the signal DB 111. The signal DB 111 stores a plurality of signal data as a data list together with IDs, attributes, and parameters, and outputs the data to the filter unit 120 and the evaluation value generation unit 140 (S203).

  Next, the filter unit 120 performs an LPF process on each signal data included in the data list (S205). Each signal data on which the LPF processing has been performed is output to the feature value calculation unit 130. The characteristic amount calculation unit 130 calculates a characteristic amount for each signal data after the LPF processing (S207). Next, the feature amount calculation unit 130 generates a feature vector including one or more calculated feature amounts (S209). Data including the generated feature vector is output to the model generation unit 150.

  Next, the model generation unit 150 generates a feature space including feature vectors corresponding to M scenarios (S211). On the other hand, the evaluation value generation unit 140 determines an evaluation value for each signal data based on the attributes included in the evaluation value list (S213). When the evaluation value is determined in advance for each signal data, step S211 may be omitted. Data including the determined evaluation value is output to the model generation unit 150.

  Then, the model generation unit 150 generates a regression learning model based on the data that pairs the generated feature space and the evaluation value (S215). Thereafter, the model generation unit 150 outputs the generated regression learning model to the regression learning model DB 151 (S217).

  The model generation device 10 according to the present embodiment has been described above. Subsequently, the determination device 20 according to the present embodiment will be described.

[2.2. Identification device]
(1) Configuration FIG. 8 is a block diagram illustrating a configuration example of the discrimination device 20 according to the first embodiment of the present invention. Referring to FIG. 8, the determination device 20 includes a signal acquisition unit 210, a signal DB 211, a filter unit 220, a feature value calculation unit 230, an evaluation value calculation unit 240, a regression learning model DB 241, a determination unit 250, and an adjustment unit 260. .

(Signal acquisition unit)
The signal acquisition unit 210 acquires signal data output from the sensor 2 via a communication unit or an input unit (not shown). When performing the discrimination processing on the acquired signal in real time, the signal acquisition unit 210 outputs the acquired signal data to the filter unit 220. In addition, when performing the discrimination processing on the acquired signal data afterwards, the signal acquisition unit 210 may output the acquired signal data to the signal DB 211.

(Signal DB)
The signal DB 211 is a database that sequentially stores the signal data acquired by the signal acquiring unit 210. When performing the determination process on the stored signal data, the signal data stored in the signal DB 211 is output to the filter unit 220.

(Filter section)
The filter unit 220 performs a filtering process on the signal data acquired from the signal acquisition unit 210 or the signal DB 211. The filter process performed by the filter unit 220 may be an LPF process similarly to the filter unit 120 of the model generation device 10. In this case, the LPF processing may be based on an IIR filter. Since the IIR filter is the same as the above-described IIR filter, the description is omitted. The filter unit 220 outputs the LPF-processed signal data to the feature amount calculation unit 230.

(Feature calculation unit)
The feature amount calculation unit 230 calculates a feature amount of the signal data acquired from the filter unit 220. The feature value calculated by the feature value calculation unit 230 is the same as the feature value calculated by the feature value calculation unit 130 of the model generation device 10. For example, the feature value calculation unit 230 calculates the amplitude feature value and the frequency feature value of the waveform of the signal data as the feature value. The specific calculation method of these feature amounts is the same as the calculation method by the feature amount calculation unit 130 of the model generation device 10, and thus the description is omitted.

  The feature amount calculation unit 230 generates a feature vector having the calculated feature amount as a component. Specifically, the feature amount calculation unit 230 calculates a two-dimensional feature vector F (k) = [V (k), f (k)] from the signal data I (k) or Q (k) at time k. Generate Note that V (k) is an amplitude feature amount, and f (k) is the number of zero crossings in a unit period. The unit period described above is the same as the unit period used when calculating the number of zero crossings by the feature amount calculation unit 130 of the model generation device 10. The feature amount calculation unit 230 outputs the generated feature vector F (k) to the evaluation value calculation unit 240.

(Evaluation value calculation unit)
The evaluation value calculation unit 240 uses the feature vector F (k) acquired from the feature amount calculation unit 230 and one regression learning model stored in the regression learning model DB 241 to detect the motion of the moving object detected by the sensor 2. Is calculated. The regression learning model stored in the regression learning model DB 241 is a regression learning model generated by the model generation device 10. The evaluation value calculation unit 240 calculates, for example, a value of an output value z (k) obtained by substituting the feature vector F (k) into a discriminant function represented by the following equation (9) as an evaluation value. The function g (xi _{, k} , σ) shown in the following equation (9) is a kernel function shown in the above equation (8).

Here, the value of x _{i, k} is given by the following equation (10). Note that V _i and f _i are the values of the i component of the support vector matrix s represented by the above equation (6).

  The output value z (k) calculated by the evaluation value calculation unit 240 is output to the determination unit 250 as an evaluation value.

(Regression learning model DB)
The regression learning model DB 241 stores at least one regression learning model calculated by the model generation device 10. The regression learning model DB 241 stores part or all of the regression learning model stored in the regression learning model DB 151 of the model generation device 10. The regression learning model stored in the regression learning model DB 241 is output to the evaluation value calculation section 240. The type of the regression learning model output to the evaluation value calculation unit 240 is freely selected by the user of the discriminating apparatus 20 or the like. For example, a regression learning model suitable for the environment in which the determination device 20 is used, the structure of the building, the material, and the like may be selected. Note that the regression learning model stored in the regression learning model DB 241 can be appropriately added, changed, or deleted via an input unit or a communication unit (not shown).

(Discriminator)
The determination unit 250 determines the motion of the moving object detected by the sensor 2 based on the evaluation value (output value) z (k) calculated from the evaluation value calculation unit 240. In the present embodiment, the determination unit 250 determines whether the motion of the moving object detected by the sensor 2 is a motion of a person based on the evaluation value z (k). More specifically, the determination unit 250 compares the evaluation value z (k) with the sensitivity threshold Th, and determines whether or not the motion of the moving object is a motion of a person based on the magnitude relationship. Here, assuming that a determination result value indicating whether or not the movement is a person motion is d, the value of the determination result value d is determined based on the following comparison formula. First, when z (k) ≧ Th, d = 1. When d = 1, the motion of the moving object is determined to be the motion of the person. On the other hand, when z (k) <Th, d = 0. When d = 0, it is determined that the motion of the moving object is not the motion of the person.

  The determination unit 250 outputs the value of the determination result value d to an external device, a storage unit (not shown) included in the determination device 20, or the like. Further, the determination unit 250 may output the evaluation value z (k) instead of or together with the determination result value d. For example, when the appropriate value of the sensitivity threshold Th adjusted by the adjustment unit 260 described later is calibrated in the calibration of the determination device 20, in order to verify the actual value of the evaluation value z (k) and the movement of the moving object. , Z (k) can be output.

(Adjustment unit)
The adjustment unit 260 adjusts the sensitivity threshold Th that is compared with the evaluation value corresponding to the movement of the moving object for the determination of the movement of the moving object by the determination unit 250. The adjustment unit 260 adjusts the sensitivity threshold Th via an input unit or a communication unit (not shown). The evaluation value z (k) as the target of the sensitivity threshold Th is a continuous value obtained by the discriminant function represented by the above equation (9). The discriminant function is a continuous function including a support vector matrix s, a weight vector c, and a threshold b included in the regression learning model. That is, the regression learning model forms a continuous function generated based on each scenario. Therefore, for example, even if the signal detected by the sensor 2 is a signal in which a plurality of disturbances are mixed, the one-dimensional evaluation value reflecting the characteristics of the signal corresponding to each scenario is obtained by the discriminant function. Is output. Therefore, the adjustment unit 260 can adjust the sensitivity to the determination of the movement of the moving object in an environment where a plurality of disturbances exist only by setting the sensitivity threshold Th to an arbitrary value.

  The adjustment unit 260 sets an arbitrary value as the sensitivity threshold Th within a range of possible values of the evaluation value z (k) output by the above discriminant function. The value of the sensitivity threshold Th may change continuously or may change discretely. For example, the adjustment unit 260 may adjust the value of the sensitivity threshold Th in a stepwise manner. Thus, the user using the determination device 20 can easily adjust the value of the sensitivity threshold Th. A predetermined value may be set in advance as an initial value of the sensitivity threshold Th, or a value according to an environment in which the determination device 20 is used may be set. Further, the adjustment of the value of the sensitivity threshold Th may be manually set by a user using the discriminating apparatus 20, but by using a statistical technique such as logistic regression analysis or linear discriminant analysis, the optimal sensitivity threshold Th may be adjusted. A value may be set. The set value of the sensitivity threshold Th is output to the determination unit 250.

(2) Operation FIG. 9 is a flowchart showing an operation example of the discrimination device 20 according to the first embodiment of the present invention. First, the signal acquisition unit 210 acquires a signal derived from the motion of the moving object detected by the sensor 2 as signal data (S301). The signal acquisition unit 210 outputs the acquired signal data to the signal DB 211 or the filter unit 220.

  Next, the filter unit 220 performs LPF processing on the acquired signal data (S303). The signal data on which the LPF processing has been performed is output to the feature amount calculation unit 230. The feature amount calculation unit 230 calculates the feature amount of the signal data after the LPF processing (S305). Next, the feature amount calculation unit 230 generates a feature vector composed of the calculated one or more feature amounts (S307). The data including the generated feature vector is output to the evaluation value calculation unit 240.

  The evaluation value calculation unit 240 calculates an evaluation value corresponding to the movement of the moving object using the acquired feature vector and one regression learning model stored in the regression learning model DB 241 (S309). Next, the determination unit 250 determines the motion of the moving object by comparing the evaluation value with the sensitivity threshold adjusted by the adjustment unit 260 (S311). Then, the determination unit 250 outputs the determination result to an external device, a storage unit, or the like (S313).

[2.3. effect]
So far, the first embodiment of the present invention has been described with reference to FIGS. According to the first embodiment of the present invention, the model generation device 10 uses SVR or the like based on a plurality of data pairs of a plurality of predetermined motion feature amounts and evaluation values corresponding to the predetermined motions. Generate a regression learning model. Then, the discrimination device 20 uses the regression learning model generated by the model generation device 10 to determine the motion of the moving object detected by the sensor 2. The regression learning model forms a discriminant function that is a continuous function, and the discriminator 20 uses the discriminant function to calculate a one-dimensional evaluation value corresponding to the motion of the moving object. Then, the determination device 20 determines the movement of the moving object by comparing the evaluation value with a sensitivity threshold value that can be adjusted to an arbitrary value. In such a configuration, by using a regression learning model reflecting a plurality of predetermined movements, even if various disturbances accompanying the movement of the moving body detected by the sensor 2 are included in the signal, the discrimination device 20 can be used. Can set the sensitivity threshold freely without depending on the type of the disturbance. Thereby, the sensitivity relating to the determination of the movement of the moving object can be more finely adjusted. Therefore, for example, it is possible to perform sensitivity adjustment for discrimination even for a signal that includes a disturbance that could not be discriminated conventionally, such as a disturbance that is not a target of the sensitivity correction unit or a mixture of a plurality of disturbances. . Therefore, the accuracy of determining the motion of the moving object can be further improved.

<3. Second Embodiment>
Subsequently, a second embodiment of the present invention will be described. In the present embodiment, the evaluation value generation unit 140 included in the model generation device 10 has a function of appropriately determining an evaluation value corresponding to each scenario. In the first embodiment of the present invention, a predetermined value is assigned in advance for each scenario, but the predetermined value is a value determined based on the subjective or experience of the user of the model generation device 10 or the like. . Therefore, when generating a regression learning model, these evaluation values do not always generate an optimal model. Therefore, the evaluation value generation unit 140 according to the present embodiment prepares a plurality of data lists including signal data derived from a plurality of predetermined movements, generates a plurality of reference regression learning models for each data list, and Using the reference regression learning model, discrimination processing is performed for all data lists used for generating the reference regression learning model. Then, the evaluation value generation unit 140 compares the discrimination results obtained using each reference regression learning model, and determines an appropriate evaluation value based on the comparison result. As a result, the evaluation value for the predetermined movement can be determined mechanically without subjectivity of the user or the like. Therefore, the quality of the regression learning model by the model generation device 10 is improved, and the accuracy of the determination result by the determination device 20 can be further improved. Hereinafter, the model generation device 10 and the evaluation value generation unit 140 according to the second embodiment will be described.

[3.1. Model generation device]
(1) Configuration FIG. 10 is a block diagram illustrating a configuration example of a model generation device 10 according to the second embodiment of the present invention. Referring to FIG. 10, the model generation device 10 includes a signal acquisition unit 110, a signal DB 111, a filter unit 120, a feature amount calculation unit 130, a feature amount DB 131, an evaluation value generation unit 140, an evaluation value DB 141, a model generation unit 150, A regression learning model DB 151 is provided. The signal acquisition unit 110, the signal DB 111, the filter unit 120, the feature value calculation unit 130, the evaluation value DB 141, the model generation unit 150, and the regression learning model DB 151 according to the present embodiment include the components according to the first embodiment. Since they have the same function, the description is omitted. The feature amount DB 131 is a database that stores data including the feature amount of the signal data calculated by the feature amount calculation unit 130 in association with a data set stored in the signal DB 111. The feature amount DB 131 will be described later in detail. Hereinafter, the configuration and the like of the evaluation value generation unit 140 will be described.

[3.2. Evaluation value generator]
(1) Configuration FIG. 11 is a block diagram illustrating a configuration example of the evaluation value generation unit 140 and a database that stores various data used in the evaluation value generation unit 140. Referring to FIG. 11, evaluation value generation section 140 includes reference model generation section 410, discrimination verification section 420, and evaluation value determination section 430.

(Reference model generator)
The reference model generation unit 410 corresponds to each of the feature amounts (feature vectors) included in the plurality of data lists stored in the feature amount DB 131 and each of the plurality of data lists stored in the reference evaluation value DB 151. A reference regression learning model is generated using the reference evaluation values included in the reference evaluation value list. The reference regression learning model is a regression learning model generated using a plurality of data sets based on a set of a feature amount included in each of the plurality of data lists and a reference evaluation value. In the present embodiment, two types of data lists (data list 1 (without disturbance) and data list 2 (with disturbance)) are prepared. Therefore, two types of models, a model corresponding to the data list 1 and a model corresponding to the data list 2, are generated for the reference regression learning model. In the present embodiment, two types of reference regression learning models are generated. However, the present invention is not limited to such an example, and three or more types of data lists are used, and reference regression learning models corresponding to the data lists are used. A learning model may be generated.

  Each of the plurality of data lists has a different configuration. FIGS. 12 and 13 are diagrams illustrating an example of the data list 1 (without disturbance) 3000 and the data list 2 (with disturbance) 4000 stored in the feature amount DB 131. Referring to FIG. 12 and FIG. 13, each data list is configured by a data set including an ID, an attribute, a scenario, a signal data file including signal data, and a feature data file including feature data. The feature amount data included in the feature amount data file includes a feature amount of a predetermined motion calculated by the feature amount calculation unit 130 in advance. Here, the data list 1 includes data sets whose attributes are “unmanned (no disturbance)” and “manned”. On the other hand, the data list 2 includes data sets whose attributes are “unmanned (no disturbance)”, “unmanned (disturbed)”, and “manned”. Although the data lists 1 and 2 include data sets having the same attribute and scenario, the signal data included in each data set does not necessarily have to be the same. Further, a plurality of signal data may be included for a scenario having the same content.

  FIGS. 14 and 15 are diagrams showing the configuration of the reference evaluation value list 1 (without disturbance) 5000 and the reference evaluation value list 2 (with disturbance) 6000 stored in the reference evaluation value DB 151. Referring to FIGS. 14 and 15, each data list is configured by a data set including an ID, an attribute, a scenario, and a reference evaluation value. Similarly to the data list 1 and the data list 2 described above, the reference evaluation value list 1 does not include a data set whose attribute is “unmanned (with disturbance)”, while the reference evaluation value list 2 has the attribute “ Unattended (with disturbance) "data set. In these reference evaluation value lists, the reference evaluation value of the data set whose attribute is “unmanned” is set to 0, and the reference evaluation value of the data set whose attribute is “manned” is set to 1. Data list 1 corresponds to reference evaluation value list 1, and data list 2 corresponds to reference evaluation value list 2.

  Data list 1 (and reference evaluation value list 1) does not include a data set having the attribute “unmanned (with disturbance)”. Therefore, when the reference regression learning model (referred to as reference regression learning model 1) generated using the data list 1 and the reference evaluation value list 1 is applied to the discriminating process, the data list 2 and the reference evaluation value list 2 are combined. It is considered that the probability of discriminating a motion other than a person as a person's motion is higher than a case where a reference regression learning model generated using the above (referred to as reference regression learning model 2) is applied to the discrimination processing. On the other hand, in the reference evaluation value list 2, the reference evaluation value of the attribute “unmanned (with disturbance)” is set to 0. Therefore, when the reference regression learning model 2 is applied to the discrimination process, the probability that the motion of the person is discriminated as a motion other than the person is higher than when the reference regression learning model 1 is applied to the discrimination process. Can be Therefore, the evaluation value generation unit 140 according to the present embodiment determines evaluation values corresponding to all scenarios according to the difference in the determination result between the reference regression learning models. This makes it possible to generate a regression learning model that can improve the discrimination accuracy.

  The reference model generation unit 410 generates a reference regression learning model 1 and a reference regression learning model 2, and stores these reference regression learning models in the reference model DB 161.

(Discrimination verification section)
The discriminating verification unit 420 determines the reference regression learning model stored in the reference model DB 161 for the feature amounts included in all data lists used for generating the reference regression learning model, which are stored in the feature amount DB 131. Of each reference regression learning model is verified, and an evaluation value corresponding to each scenario is determined based on the verification result. Specifically, the discriminating verification unit 420 first obtains a feature amount from all the feature amount data included in the data list 1 and the data list 2. Then, the discriminating verification unit 420 determines whether or not a specific moving object is moving (here, whether or not there is a person) using the reference regression learning model 1 and the reference regression learning model 2 for each feature amount. . The discrimination verification unit 420 compares the difference between the reference regression learning models with respect to the discrimination result (unmanned rate) of the presence or absence of a person. Then, the discrimination verification unit 420 determines an evaluation value corresponding to each scenario based on the difference. Hereinafter, the processing in the determination verification unit 420 will be described in order.

  First, the discriminating verification unit 420 performs a discriminating process on each of the feature amounts included in the data list used for generating the reference regression learning model, using each of the reference regression learning models. The discriminating process by the discriminating verification unit 420 is the same process as the process by the evaluation value calculating unit 240 and the discriminating unit 250 of the discriminating apparatus 20 according to the first embodiment. For example, it is assumed that one reference regression learning model generated by the reference model generation unit 410 includes the following support vector matrix s', weight vector c ', and threshold value b'. Specifically, the support vector matrix s 'is defined by a matrix represented by the following equation (11), and the weight vector c' is defined by a vector represented by the following equation (12). The threshold value b 'is a value uniquely determined.

  Further, similarly to the first embodiment, it is also possible to generate a regression learning model by mapping data constituted by a feature space and an evaluation value onto a non-linear space using a kernel trick. In the present embodiment, a Gaussian kernel represented by the following equation (13) may be used as a kernel trick.

Here, the number of scenarios included in all data lists is M, and the feature vector corresponding to the m-th scenario is F _m (k). The discriminant verification unit 420 calculates, as an evaluation value, an output value z _m (k) obtained by substituting the feature vector F _m (k) into a discriminant function represented by the following equation (14).

Here, the value of x _{i, k} is given by the following equation (15). Here, V _i and f _i are the values of the i component of the support vector matrix s represented by the above equation (11).

Next, the determination verification unit 420 determines whether or not the m-th scenario is a motion of a human body, based on the evaluation value z _m (k) obtained by Expression (14). More specifically, the determination verification unit 420 determines whether the m-th scenario is a human body motion based on whether the evaluation value z _m (k) is equal to or greater than 0. Here, if the determination result value indicating whether a human body motion and d _m, based on the following comparison expression, the discrimination result value the value of d _m are determined. First, when z _m (k) ≧ 0, d _m = 1. If d _m = 1, the m-th scenario is determined to be a person's movement. On the other hand, when z _m (k) <0, d _m = 0. If d _m = 0, it is determined that the m-th scenario is not a motion of a person.

Determination verification section 420 calculates a d _m for all scenarios included in all data lists. Then, the discrimination verification unit 420 calculates the unmanned rate eval [%] according to the following equation (16).

  The discrimination verification unit 420 performs the discrimination process and the unmanned rate (eval) calculation process for each of the plurality of reference regression learning models generated by the reference model generation unit 410. In the present embodiment, the discriminating verification unit 420 performs the discriminating process for all scenarios using each of the reference regression learning model 1 and the reference regression learning model 2, and performs the unmanned rate calculation process based on the discrimination result. I do. Then, discrimination verification section 420 outputs the calculation result of the unmanned rate to evaluation value determination section 430.

(Evaluation value determination unit)
The evaluation value determination unit 430 determines an evaluation value used for the process of generating a regression learning model based on the determination result obtained from the determination verification unit 420. More specifically, the evaluation value determination unit 430 determines the evaluation value based on the difference between the unmanned rates obtained from the discrimination verification unit 420 for each reference regression learning model. Hereinafter, the evaluation value determination processing by the evaluation value determination unit 430 will be described with reference to FIG.

  FIG. 16 is a diagram for explaining the evaluation value determination processing by the evaluation value determination unit 430. Referring to FIG. 16, the data list 7000 includes an ID corresponding to a scenario, a reference evaluation value, an unmanned rate in the reference regression learning model 1, an unmanned rate in the reference regression learning model 2, a difference between the unmanned rates, and an evaluation value. It is included. The reference evaluation value is a temporary evaluation value stored in advance in the reference evaluation value DB 151, and is set to 0 for a scenario with an attribute of “unmanned” and set to 1 for a scenario with an attribute of “attended”. The unmanned rate in the reference regression learning model 1 and the unmanned rate in the reference regression learning model 2 are unmanned rates calculated using the respective reference regression learning models. For example, in the case of ID 1001 (attribute “unmanned (no disturbance)”, scenario “none”), both show an unmanned rate of 100%. In this case, the evaluation value determination unit 430 determines the value (1) given in advance as the reference evaluation value as the evaluation value. On the other hand, in the ID 1004 (attribute “maned”, scenario “sitting”), the unmanned rate in the reference regression learning model 1 indicates 50%, but the unmanned rate in the reference regression learning model 2 indicates 90%. I have. As described above, the calculated unmanned rate differs depending on the reference regression learning model depending on the scenario. Therefore, it is necessary to determine an evaluation value so that a scenario that is objectively difficult to determine is not output as a result different from an expected result of the determination.

  FIG. 17 is a graph showing the distribution of the unmanned rate calculated by each reference regression learning model. In the graph of FIG. 17, ○ indicates the unmanned rate of the scenario including the attribute “unmanned” included in the data list 1, □ indicates the unmanned rate of the scenario including the attribute “maned” included in the data list 1, The unmanned rate of the scenario including the attribute “unmanned” included, and △ is the unmanned rate of the scenario including the attribute “maned” included in the data list 2. As shown in FIG. 17, ○ and Δ indicate that the difference in the unmanned rate by the reference regression learning model is small, but there are cases where the difference in the unmanned rate is increased by the reference regression learning model, such as □ and ×. Therefore, when the difference between the unmanned rates increases, the evaluation value determining unit 430 determines the value of the evaluation value as 0.5 as an intermediate value in order to suppress the influence of the difference in the reference regression learning model. On the other hand, when the difference in the unmanned rate is small, it is considered that the influence of the difference in the reference regression learning model is small, so the value given as the reference evaluation value is determined as it is as the evaluation value. As described above, by flexibly changing the evaluation value according to the difference in the unmanned rate, a regression learning model without subjectivity can be generated.

Referring to FIG. 16 again, the evaluation value determination unit 430 determines the values of the evaluation values for all the scenarios based on the difference between the unmanned rates. For example, the evaluation value determination unit 430 determines the evaluation value to be 0.5 when the difference in the unmanned rate for the same scenario between the reference regression learning model 1 and the reference regression learning model 2 is equal to or greater than _a predetermined threshold Tha. I do. In the present embodiment, the predetermined threshold value Th _a is 20%. When the difference between the unmanned rates is equal to or greater than 20%, the evaluation value determination unit 430 determines the evaluation value as 0.5 for the above scenario. On the other hand, when the difference in the unmanned rate is less than 20%, the evaluation value determination unit 430 determines the value given as the reference evaluation value as the evaluation value.

The value of the predetermined threshold Th _a can be freely changed for precision adjustment of the regression learning model. Further, in the present embodiment, the evaluation value of the three values by a predetermined threshold value Th _a is set, the number of the predetermined threshold may be plural. Further, when the difference between the unmanned rates is large, 0.5 is set as the intermediate value of the evaluation value, but the value as the intermediate value is not particularly limited. For example, when there are a plurality of predetermined thresholds as described above, various values may be set as intermediate values according to the number of steps set by the thresholds. Further, for example, the value of the evaluation value determined by the evaluation value determination unit 430 may be determined based on the calculated value of the unmanned rate eval, a statistic, or the like. For example, the value of the evaluation value corresponding to one scenario may be a value corresponding to the average value of the unmanned rate calculated using a plurality of reference regression learning models.

  The evaluation value generator 140 outputs an evaluation value list including the evaluation value determined by the evaluation value determiner 430 to the evaluation value DB 141.

(2) Operation FIG. 18 is a flowchart illustrating a processing flow of the evaluation value generation unit 140 according to the second embodiment of the present invention. Referring to FIG. 18, the evaluation value generation unit 140 first generates a reference regression learning model corresponding to each data list by the reference model generation unit 410 (S400). Next, the evaluation value generation unit 140 performs a discrimination process for all scenarios using each reference regression learning model by the discrimination verification unit 420, and calculates an unmanned rate (S500). Then, the evaluation value generation unit 140 compares the unmanned rate between the reference regression learning models by the evaluation value determination unit 430, and determines the evaluation value of each scenario according to the comparison result (S600). Hereinafter, the above three processes will be described in order.

  FIG. 19 is a flowchart illustrating a processing flow of the reference model generation unit 410 according to the second embodiment of the present invention. Referring to FIG. 19, the reference model generation unit 410 first obtains one data list from the feature DB 131 (S401). Next, the reference model generation unit 410 extracts feature data including feature vectors corresponding to all scenarios included in the data list (S403). The reference model generation unit 410 acquires a reference evaluation value list corresponding to the data list from the reference evaluation value DB 151 (S405). Then, the reference model generation unit 410 generates a reference regression learning model based on data in which the feature space including the plurality of feature vectors and the reference evaluation value included in the reference evaluation value list are paired (S407). The reference model generation unit 410 performs the above-described processing on a plurality of data lists (and reference evaluation value lists), and generates a plurality of reference regression learning models.

  FIG. 20 is a flowchart illustrating a processing flow of the discrimination verification unit 420 according to the second embodiment of the present invention. Referring to FIG. 20, the discriminating verification unit 420 first extracts a feature vector of each scenario from all data lists stored in the feature DB 131 (S501). Next, the discriminant verification unit 420 calculates an evaluation value by substituting the extracted feature vector into a discriminant function constituted by one reference regression learning model, and outputs a discrimination result (S503). Then, the discrimination verification unit 420 calculates an unmanned rate using the discrimination results corresponding to all scenarios (S505). Using the reference regression learning model 1 and the reference regression learning model 2, the discrimination verification unit 420 calculates the unmanned rate in each model for all scenarios.

FIG. 21 is a flowchart illustrating a processing flow of the evaluation value determination unit 430 according to the second embodiment of the present invention. Referring to FIG. 21, the evaluation value determination unit 430 first obtains a reference evaluation value corresponding to one scenario from the reference evaluation value DB 151 (S601). Next, the evaluation value determination unit 430 acquires the unmanned rate in each reference regression learning model calculated by the discrimination verification unit 420 (S603). Evaluation value decision unit 430 calculates the difference between the unmanned rates obtained respectively (diff) (S605), compares the magnitude of the difference diff and the predetermined threshold value Th _a of the unmanned rate (S607). If the difference diff unattended rate is the threshold value Th _a higher (S607 / YES), the evaluation value determining unit 430, an evaluation value corresponding to the scenario to determine the 0.5 (S609). On the other hand, if the difference diff unmanned rate is below the threshold Th _a, evaluation value determination unit 430 determines the value of the reference evaluation value corresponding to the scenario as the evaluation value (S611). The evaluation value determination unit 430 performs the above-described evaluation value determination processing for all scenarios.

[3.3. effect]
So far, the second embodiment of the present invention has been described with reference to FIGS. According to the second embodiment of the present invention, the evaluation value generation unit 140 generates a plurality of reference regression learning models each including a plurality of different data lists and reference evaluation value lists, and generates the reference regression learning models. The presence / absence of a person is determined for the scenarios included in the plurality of data lists. Then, the evaluation value generation unit 140 determines an evaluation value corresponding to each scenario based on the difference in the discrimination result (unmanned rate) between the reference regression learning models. In such a configuration, the evaluation value used for generating the regression learning model is determined using the regression analysis technique, so that the evaluation value corresponding to each scenario can be determined mechanically. Thus, for example, the evaluation value corresponding to a scenario in which it is difficult to determine whether or not a person is moving, such as the attribute “unmanned (disturbed)” or the scenario “sit”, excludes the subjectivity of the user or the like. Can be determined. Therefore, the model generation device 10 according to the present embodiment can generate a regression learning model using an evaluation value list including more appropriate evaluation values. Therefore, it is possible to further improve the accuracy of determining whether the movement of the moving object is a specific movement.

<4. Example of hardware configuration>
As above, the discrimination system 1 according to each embodiment of the present invention has been described. Next, a hardware configuration of an information processing device according to an embodiment of the present disclosure will be described with reference to FIG. FIG. 22 is a block diagram illustrating a hardware configuration example of the information processing device 900 according to the embodiment of the present disclosure. The illustrated information processing device 900 can realize, for example, the model generation device 10 and the determination device 20 in the above embodiment.

  Referring to FIG. 22, the information processing apparatus 900 includes a CPU (Central Processing Unit) 901, a ROM (Read Only Memory) 902, a RAM (Random Access Memory) 903, and a host bus 904. Further, the information processing device 900 includes a bridge 905, an external bus 906, an interface 907, an input device 908, an output device 909, a storage device 910, a drive 911, and a network interface 912.

  The CPU 901 functions as an arithmetic processing device and a control device, and controls overall operations in the information processing device 900 according to various programs. Further, the CPU 901 may be a microprocessor. Note that the CPU 901 controls the entire operation of a function including each functional unit in the information processing device 900 or a part of the operation. For example, in the CPU 901, the ROM 902 stores programs used by the CPU 901, calculation parameters, and the like. The RAM 903 temporarily stores a program used in the execution of the CPU 901 and parameters that change as appropriate in the execution. These are mutually connected by a host bus 904 including a CPU bus and the like.

  The host bus 904 is connected via a bridge 905 to an external bus 906 such as a PCI (Peripheral Component Interconnect / Interface) bus. Note that the host bus 904, the bridge 905, and the external bus 906 do not necessarily need to be separately configured, and these functions may be implemented on one bus.

  An input device 908 includes an input unit such as a mouse, a keyboard, a touch panel, a button, a microphone, a switch, and a lever for a user to input information, and an input control circuit that generates an input signal based on an input by the user and outputs the input signal to the CPU 901. It is composed of By operating the input device 908, the user of the information processing device 900 can input various data to the information processing device 900 and instruct a processing operation. Note that the input device 908 implements a function of an input unit (not shown).

  The output device 909 includes a display device such as a CRT display device, a liquid crystal display (LCD) device, an OLED device, and a lamp. Further, the output device 909 includes a sound output device such as a speaker and headphones. The output device 909 outputs the reproduced content, for example. Specifically, the display device displays various information such as reproduced video data as text or images. On the other hand, the audio output device converts reproduced audio data, text data displayed on a display device, and the like into audio and outputs the audio. Note that the output device 909 implements the function of a display unit (not shown).

  The storage device 910 is a device for storing data in the information processing device 900 according to the embodiment of the present invention. The storage device 910 may include a storage medium, a recording device that records data on the storage medium, a reading device that reads data from the storage medium, a deletion device that deletes data stored on the storage medium, and the like. The storage device is composed of, for example, a hard disk drive (HDD) or a solid state drive (SSD). The storage device 910 stores programs executed by the CPU 901 and various data. The storage device 910 stores each DB provided in each information processing device, such as the signal DB 111. Note that the storage device 910 implements the function of a storage unit (not shown).

  The drive 911 is a reader / writer for a storage medium, and is built in or external to the information processing apparatus 900. The drive 911 reads information recorded on a removable storage medium 96 such as a mounted magnetic disk, optical disk, magneto-optical disk, or semiconductor memory, and outputs the information to the RAM 903. The drive 911 can also write information to the removable storage medium 96.

  The network interface 912 is a communication interface including, for example, a communication device for connecting to another device. Further, the network interface 912 may be a communication device compatible with a wireless LAN (Local Area Network), a communication device compatible with LTE (Long Term Evolution), or a Bluetooth communication device. Further, the network interface 912 may be a wire communication device that performs wired communication. Note that the network interface 912 implements the function of a communication unit (not shown).

  The example of the hardware configuration of the information processing apparatus 900 has been described above. Each of the above components may be configured using a general-purpose member, or may be configured by hardware specialized for the function of each component. Such a configuration can be appropriately changed according to the technical level at the time of implementation.

<5. Summary>
As described above, the preferred embodiments of the present invention have been described in detail with reference to the accompanying drawings, but the present invention is not limited to such examples. It is obvious that those skilled in the art to which the present invention pertains can conceive various changes or modifications within the scope of the technical idea described in the claims. It is understood that these also belong to the technical scope of the present invention.

  For example, in the above embodiment, it is assumed that a person is present in an area where the sensor 2 can detect, but the present invention is not limited to this example. For example, a sensor may be provided so that the sensor can detect a person only when the person is in a position such as a chair or a bed. This makes it possible to limit the area to be determined, thereby improving the accuracy of determining the motion of the moving object. Further, the discrimination device may be used in combination with the result of the presence or absence of a person with another information processing device (for example, a device that measures the amount of movement and posture of the person). In such a configuration, processing by another information processing device can be performed only when a person is present in a specific area. Therefore, the influence of disturbance that can occur is reduced, and the measurement accuracy of the amount of movement and posture of the person can be improved.

  Further, the determination device may include each function of the model generation device. In such a configuration, the discriminating apparatus not only performs the discriminating process using the regression learning model stored in the regression learning model DB, but also generates a regression learning model using signal data or the like obtained from outside. it can. This makes it possible to generate not only a general regression learning model but also a regression learning model specialized for a place or a scene to which the discriminating device is applied. Therefore, the accuracy of determining the motion of the moving object to be determined can be further improved.

  Further, in the above-described embodiment, an example has been described in which a motion of a person is set as a determination target, but the present invention is not limited to such an example. For example, the present invention is applicable to various moving body movements such as movements of other living bodies, machines and cranks. In this case, for example, when the sensor is a radio wave type sensor, by appropriately setting the wavelength of the radiation wave in accordance with the displacement of the signal waveform to be determined, it is possible to extract a signal representing any motion of the moving object. It is possible.

  In addition, each step in the processing of each device and each functional unit in this specification does not necessarily need to be processed in a time series in the order described in the flowchart. For example, each step in the processing of each device and each functional unit may be processed in an order different from the order described in the flowchart, or may be processed in parallel.

  Further, a computer program for causing hardware such as a CPU, a ROM, and a RAM built in the model generation device 10 and the discrimination device 20 to perform functions equivalent to those of the components of the model generation device 10 and the discrimination device 20 is also provided. Can be created. A storage medium storing the computer program is also provided.

Reference Signs List 1 discrimination system 2 sensor 10 model generation device 20 discrimination device 110, 210 signal acquisition unit 120, 220 filter unit 130, 230 feature amount calculation unit 140 evaluation value generation unit 150 model generation unit 240 evaluation value calculation unit 250 determination unit 260 adjustment unit 410 Reference model generation unit 420 Discrimination verification unit 430 Evaluation value determination unit

Claims (12)

A feature amount calculation unit that calculates a feature amount of the movement of the moving object based on a sensor signal indicating the movement of the moving object detected by the sensor;
Using a regression learning model generated based on a plurality of data that is a set of a feature amount of a predetermined motion and an evaluation value corresponding to the predetermined motion, the moving object calculated by the feature amount calculation unit. An evaluation value calculation unit that calculates an evaluation value corresponding to the motion of the moving object according to the feature amount of the motion,
An adjustment unit that adjusts a threshold value to be compared with the evaluation value corresponding to the movement of the moving object for determination of the movement of the moving object,
Equipped with a,
The evaluation value corresponding to the predetermined motion is a plurality of pieces of data generated based on a plurality of sets of a feature amount of the predetermined motion and an evaluation value set in advance corresponding to the predetermined motion. A discriminating apparatus , which is determined according to an error between a plurality of determination results corresponding to the predetermined motion determined using a regression learning model . The regression learning model is a regression learning model generated by support vector regression,
The evaluation value calculation unit calculates an evaluation value corresponding to the motion of the moving object by using a support vector included in the regression learning model generated by the support vector regression and a feature amount of the motion of the moving object. The discriminating apparatus according to claim 1.   The discriminating apparatus according to claim 1, wherein the feature amount calculation unit calculates the feature amount based on a waveform of the sensor signal.   The discrimination device according to claim 3, wherein the feature amount calculation unit calculates the magnitude of the amplitude of the sensor signal as the feature amount.   The discrimination device according to claim 3, wherein the feature amount calculation unit calculates a frequency of the waveform of the sensor signal in a predetermined section as the feature amount. The adjustment section adjusts the threshold value stepwise, discriminating device according to any one of claims 1-5. The predetermined movement includes movement of a living body,
The determination device according to any one of claims 1 to 6 , wherein the determination of the movement of the moving object includes determination of whether the movement of the moving object is the movement of the living body. Wherein the sensor is a Doppler sensor, discriminating apparatus according to any one of claims 1-7. Calculating a feature amount of the motion of the moving object based on a sensor signal indicating the motion of the moving object detected by the sensor;
Using a regression learning model generated based on a plurality of data pairs of a predetermined motion feature amount and an evaluation value corresponding to the predetermined motion, according to the motion feature amount of the moving object, Calculating an evaluation value corresponding to the movement of the moving object;
Adjusting a threshold value to be compared with the evaluation value corresponding to the movement of the moving object for determination of the movement of the moving object;
Only including,
The evaluation value corresponding to the predetermined motion is a plurality of pieces of data generated based on a plurality of sets of a feature amount of the predetermined motion and an evaluation value set in advance corresponding to the predetermined motion. A discrimination method , wherein the discrimination method is determined according to an error between a plurality of determination results corresponding to the predetermined motion determined using a regression learning model . Computer
A feature amount calculation unit that calculates a feature amount of the movement of the moving object based on a sensor signal indicating the movement of the moving object detected by the sensor;
Using a regression learning model generated based on a plurality of data that is a set of a feature amount of a predetermined motion and an evaluation value corresponding to the predetermined motion, the moving object calculated by the feature amount calculation unit. An evaluation value calculation unit that calculates an evaluation value corresponding to the motion of the moving object according to the feature amount of the motion,
An adjustment unit that adjusts a threshold value to be compared with the evaluation value corresponding to the movement of the moving object for determination of the movement of the moving object,
To function as,
The evaluation value corresponding to the predetermined motion is a plurality of pieces of data generated based on a plurality of sets of a feature amount of the predetermined motion and an evaluation value set in advance corresponding to the predetermined motion. A program that is determined according to an error between a plurality of determination results corresponding to the predetermined movement determined using a regression learning model . A feature amount calculating unit that calculates a feature amount of the predetermined movement based on a sensor signal indicating a predetermined movement by the moving body detected by the sensor;
The feature amount of the predetermined motion and the plurality of regression learning models generated based on a plurality of data pairs of a reference evaluation value set in advance corresponding to the predetermined motion are determined. A regression learning model evaluation value determination unit that determines an evaluation value corresponding to the predetermined motion in accordance with an error between a plurality of determination results corresponding to the predetermined motion,
A model generation unit that generates a regression learning model based on a plurality of data sets each of which is a feature amount of the predetermined motion and an evaluation value corresponding to the predetermined motion;
A model generation device comprising: Calculating a characteristic amount of the predetermined movement based on a sensor signal indicating a predetermined movement by the moving body detected by the sensor;
The feature amount of the predetermined motion and the plurality of regression learning models generated based on a plurality of data pairs of a reference evaluation value set in advance corresponding to the predetermined motion are determined. Determining an evaluation value corresponding to the predetermined motion according to an error between a plurality of determination results corresponding to the predetermined motion;
A step of generating a regression learning model based on a plurality of data pairs of the feature amount of the predetermined motion and an evaluation value corresponding to the predetermined motion;
And a model generation method.

Images