JP2022148871A - Determination device - Google Patents

Abstract

To provide a determination device capable of performing processing in a short time.SOLUTION: The determination device comprises: an acquisition part acquiring a plurality of output values respectively outputted by a plurality of sensors provided on a straining body included in an object on which at least part of body weight of an object person is applied, and detecting a strain amount of the straining body; and a determination part that sets, as feature quantity, a plurality of third values as differences corresponding to a plurality of first values as the plurality of output values after first time elapsed from clock time at which at least part of the body weight is applied on the object by the object person, and a plurality of second values as the plurality of output values after second time longer than the first time elapsed from the clock time, and determines the object person by inputting the plurality of third values calculated from the output value acquired by the acquisition part to a learned model obtained by performing machine learning for determining the object person.SELECTED DRAWING: Figure 1

Description

TECHNICAL FIELD The present invention relates to an identification device, and for example, to an identification device that identifies a person sitting on a seat.

An identification device that identifies a person seated on a seat is known (for example, Patent Documents 1 to 3). A plurality of pressure sensors are installed on the seat, and a convolutional neural network is used to calculate physical characteristics such as the position of the center of gravity and the position of the maximum value of pressure when the seated person is seated. An identification device is known that identifies a seated person by calculating behavioral features such as movement of position and movement of the position of the maximum value (for example, Patent Document 1).

WO2018/179325 JP 2012-133683 A JP 2007-179474 A

In Patent Document 1, a convolutional neural network is used to extract feature amounts, and a recurrent neural network is used to extract time-series feature amounts. For this reason, a large amount of teacher data is required, and an enormous amount of time is required for arithmetic processing.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a specific device capable of processing in a short time.

The present invention provides an acquisition unit that acquires a plurality of output values respectively output by a plurality of sensors that are provided in a strain body provided in an object to which a subject applies at least part of the body weight and that detects the strain amount of the strain body. and a plurality of first values that are the plurality of output values after a first time has passed since the subject applied at least part of the body weight to the object, and a plurality of first values longer than the first time from the time A plurality of third values, which are the respective differences between the plurality of second values, which are the plurality of output values after the passage of the second time, and the plurality of third values, which are respectively corresponding differences, are used as feature amounts, and machine learning is performed to identify the subject. and a specifying unit that specifies the target person by inputting the plurality of third values calculated from the output values acquired by the acquiring unit into a learned model.

In the above configuration, the plurality of sensors may include three sensors provided at positions corresponding to vertices of a triangle when viewed from above.

In the above configuration, the object may be a seat, and the object may be seated on the object.

In the above configuration, the plurality of sensors include, when the subject is seated, a first sensor provided in front or rear with respect to the subject, a second sensor provided on the right with respect to the subject, and the It can be configured to include a third sensor provided on the left with respect to the subject.

In the above configuration, the plurality of sensors may include a fourth sensor that overlaps the subject's right thigh and a fifth sensor that overlaps the subject's left thigh when the subject is seated. .

In the above configuration, the plurality of sensors may include a sixth sensor that overlaps the back of the subject when the subject is seated.

In the above configuration, the model includes a plurality of values of at least one of the plurality of first values corresponding to the plurality of sensors and the plurality of second values corresponding to the plurality of sensors, and A model in which machine learning is performed to identify the subject using a plurality of corresponding third values as feature amounts, and the identifying unit calculates from the plurality of output values acquired by the acquiring unit. By inputting the plurality of values of the at least one and the plurality of third values respectively corresponding to the plurality of sensors obtained, the subject can be specified.

In the above configuration, the model is the sum of at least one of the plurality of values of the at least one, the plurality of third values, and the sum of the plurality of first values and the sum of the plurality of second values. is a model that performs machine learning for identifying the target person, with a value and a feature amount, and the identifying unit is a model that performs machine learning on the plurality of sensors calculated from the plurality of output values acquired by the acquiring unit. By inputting a plurality of values of the at least one, a plurality of third values, and a total value of the at least one, which correspond to each other, the target person can be specified.

In the above configuration, the model includes the plurality of first values, the plurality of second values, the plurality of third values, the sum of the plurality of first values, and the plurality of second values. and a total value are used as feature amounts, and machine learning is performed to identify the subject, and the identifying unit includes the plurality of sensors calculated from the plurality of output values obtained by the obtaining unit. the plurality of first values, the plurality of second values, the plurality of third values, the sum of the plurality of first values, and the sum of the plurality of second values, respectively corresponding to , can be configured to specify the target person.

In the above configuration, the machine learning may be machine learning using a decision tree or random forest.

The above configuration may include a machine learning unit that performs machine learning for specifying the target person using the target person and the third value as feature amounts.

The present invention comprises an acquisition unit that acquires an output value output by a sensor that is provided in a straining body provided in an object to which a subject applies at least part of the body weight and that detects the strain amount of the straining body, and the subject: a first value that is the output value at a first time after a first time has passed since the loading time when at least part of the body weight is added to the object, and a second time that is longer than the first time from the loading time A trained model that has undergone machine learning for identifying the subject using the second value that is the output value at the second time after the lapse of time and the third value that is the corresponding difference between them as a feature amount. and a specifying unit that specifies the target person by inputting the third value calculated from the output value acquired by the acquiring unit, and the first time is the same target person a plurality of times. The specific device, wherein the variation at each time of the output value when at least part of the body weight is applied is the time after overshooting.

ADVANTAGE OF THE INVENTION According to this invention, the specific apparatus which can process in a short time can be provided.

FIG. 1(a) is a schematic diagram showing a sheet on which a specific device in Example 1 is arranged, and FIGS. 1(b) and 1(c) are plan views of strain bodies. FIG. 2 is a functional block diagram of a specific device according to the first embodiment; FIG. 3 is a block diagram of a specific device in Example 1. FIG. 4(a) and 4(b) are flowcharts showing the processing of the processor in the first embodiment. FIG. 5(a) is a side view of the toilet bowl and toilet seat used in the experiment, and FIG. 5(b) is a plan view of the toilet seat showing the arrangement W of the sensors. FIGS. 6(a) to 6(d) are diagrams showing the output values with respect to time for each seated person A to D. FIG. FIG. 7 is a diagram showing the output value for each seated person's time. Figures 8(a) to 8(c) show the placement of the sensors in locations X, Y and Z, respectively. FIG. 9(a) is a perspective view of a sheet showing the arrangement of strain bodies in Modification 1 of Example 1, and FIG. 9(b) is a perspective view of the strain bodies. FIG. 10(a) is an enlarged view of FIG. 7 near the time when the user is seated, and FIG. 10(b) is a diagram showing first-order differentiation of the output value of FIG. FIG. 11(a) is a diagram showing the data of the seated person C multiple times, FIG. 11(b) is a diagram showing Max-Min output values at each time in FIG. 11(a), and FIG. 11(c) is an enlarged view of FIG. 11(b). FIG. 12(a) is a diagram showing the data of the seated person D multiple times, FIG. 12(b) is a diagram showing Max-Min output values at each time in FIG. 12(a), and FIG. 12(c) is an enlarged view of FIG. 12(b).

An embodiment will be described below with reference to the drawings.

FIG. 1(a) is a schematic diagram showing a sheet on which a specific device in Example 1 is arranged, and FIGS. 1(b) and 1(c) are plan views of strain bodies. FIG. 1(b) is a view of the strain body 20 viewed from above, and FIG. 1(c) is a view of the strain body 20 viewed from below. When the seated person 28 is seated on the seat 25, the front direction is the +X direction, the left direction is the +Y direction, the right direction is the -Y direction, and the upward direction is the Z direction.

As shown in FIG. 1( a ), the sheet 25 has a strain body 20 and a support 21 . The support 21 corresponds to, for example, the leg of the seat 25 and supports the strain body 20 . A seated person 28 sits on the strain body 20 . The support 21 is a rigid body and the strain generating body 20 is a material that flexes when the seated person 28 is seated. A cushion material may be provided between the strain generating body 20 and the seated person 28 . A sensor 10 is provided on the lower surface of the strain generating body 20 . The sensor 10 may be provided on the side surface of the strain-generating body 20 . Thus, the sensor 10 may be provided on the surface of the strain body 20. FIG. If the sensor 10 can detect the amount of displacement of the strain body 20, the sensor 10 may be provided on the surface of the strain body 20 via another member. The sensor 10 detects the amount of strain (that is, deflection) of the strain body 20 by detecting the amount of displacement of the strain body 20 . The sensor 10 may be a sensor capable of detecting the amount of displacement of the strain generating body 20, such as a strain gauge or a pressure sensor, or may be a vibration sensor. When a vibration sensor is used as the sensor 10, the amount of displacement can be calculated by integrating the detected velocity component or acceleration component once or twice. The strain-generating body 20 is resin or metal, and may be any material that is distorted (that is, bends) when pressure is applied. A specific material for the strain generating body 20 is, for example, polypropylene resin. The material used for the strain-generating body 20 is preferably hard plastic, and the elastic modulus of the strain-generating body 20 is preferably 500 MPa to 10000 MPa.

As shown in FIGS. 1(b) and 1(c), a plurality of sensors 10a to 10c are provided on the lower surface of the strain generating body 20. FIG. The sensor 10 a is provided in front of the center of gravity 23 of the seated person 28 . Sensor 10b is provided to the right of center of gravity 23 . Sensor 10c is provided to the left of center of gravity 23 . The sensors 10 a to 10 c may be provided on the upper surface of the strain body 20 or may be provided inside the strain body 20 . Only one sensor may be provided, or two or four or more sensors may be provided.

FIG. 2 is a functional block diagram of a specific device according to the first embodiment; As shown in FIG. 2 , the identification device 11 includes an acquisition unit 12 , an extraction unit 13 , a calculation unit 14 , an identification unit 16 , a machine learning unit 17 and a storage unit 18 . The output value of the sensor 10 is input to the acquisition unit 12 . A specified result is output from the specifying unit 16 . At least part of the calculation unit 14, the identification unit 16, the machine learning unit 17, and the storage unit 18 may be provided outside the identification device 11, and may be functions executed on the cloud, for example. Further, the sensor 10 may be provided inside the specific device 11 .

FIG. 3 is a block diagram of a specific device in Example 1. FIG. The identification device 11 includes, for example, a microcomputer. As shown in FIG. 3, the specific device 11 comprises a processor 30, a memory 32, an interface 34 and an internal bus . The processor 30 is a processing unit such as a CPU (Central Processing Unit) and functions as the acquisition unit 12, the extraction unit 13, the calculation unit 14, the identification unit 16, and the machine learning unit 17 in cooperation with software. The memory 32 is a volatile memory or a non-volatile memory, and stores programs executed by the processor 30, data, and the like. Also, the memory 32 functions as the storage unit 18 . The interface 34 inputs and outputs data and the like, for example, inputs data output by the sensor 10 and outputs processing results of the processor 30 . The internal bus 36 is connected between the processor 30, the memory 32 and the interface 34 and transmits data and the like between them.

4(a) and 4(b) are flowcharts showing the processing of the processor in the first embodiment. FIG. 4(a) is a flow chart for identifying a seated person, and FIG. 4(b) is a flow chart for creating a model using machine learning.

First, with reference to FIGS. 2 and 4(a), the process of identifying the seated person 28 seated on the seat 25 by the identifying device 11 will be described. The acquisition unit 12 acquires output signals from the sensors 10a to 10c (step S10). The output signal of the sensor 10 is the time-series displacement amount A(t) of the strain body 20 . From the time-series displacement amount A(t), the extraction unit 13 extracts a displacement amount A1 (first value) at a first time t1 after the first time T1 has passed since the load time t0 at which the seated person 28 sat down, and a second time T2. A displacement amount A2 (second value) at the second time t2 after the elapse is extracted (step S12). The second time T2 is longer than the first time T1. The calculation unit 14 calculates the displacement amounts A1 and A2 extracted by the extraction unit 13 (step S14). For example, the computing unit 14 computes the creep value ΔA (third value) of the strain body 20 corresponding to the difference between the displacement amounts A1 and A2=A2−A1. Further, a total value SA1 corresponding to the sum of the displacement amounts A1 and a total value SA2 corresponding to the sum of the displacement amounts A2 of the plurality of sensors 10a to 10c are calculated. Creep is a phenomenon in which the strain of the strain-generating body 20 increases over time when a constant sustained stress is applied to the strain-generating body 20 . In Example 1, after the seated person 28 is seated on the seat 25, the distribution of the stress applied to the strain-generating body 20 may change over time. The creep value ΔA includes not only the change in strain due to the creep phenomenon, but also the change in strain due to the time change of the stress distribution to the strain generating body 20 .

The specifying unit 16 uses the model stored in the storage unit 18 to specify the seated person 28 using the displacement amounts A1 and A2, the creep value ΔA, and the total values SA1 and SA2 as feature amounts (step S16). For example, the specifying unit 16 uses the model to calculate the individual identification rate score using the displacement amounts A1 and A2, the creep value ΔA, and the total values SA1 and SA2 as feature amounts. The identification unit 16 identifies the registered seated person if the calculated individual identification rate score is higher than the identification score of the registered seated person. When there are multiple registered occupants, the identification scores of the multiple occupants are compared with the calculated personal identification rate score to identify the occupant from the multiple registered occupants. Then exit.

Next, a process of creating a learned model by the identifying device 11 will be described with reference to FIGS. 2 and 4(b). The displacement amounts A1 and A2, the creep value ΔA, and the total values SA1 and SA2 are regarded as one set, and a plurality of sets are prepared for each of a plurality of specific seated persons. The machine learning unit 17 acquires a plurality of sets for a plurality of specific seated persons (step S15). A plurality of sets may be acquired in the same manner as steps S10 to S14 in FIG. Multiple sets may be obtained from memory 32 or an external memory. The machine learning unit 17 performs machine learning using a plurality of sets as teacher data, and creates a machine-learned model (step S18). The machine learning unit 17 stores the created model in the storage unit 18 (step S20). Then exit.

[experiment]
An experiment was conducted to identify a person sitting on a toilet seat. FIG. 5(a) is a side view of the toilet bowl and toilet seat used in the experiment, and FIG. 5(b) is a plan view of the toilet seat showing the arrangement W of the sensors. As shown in FIG. 5(a), a toilet seat 22 is provided on a toilet bowl 24 as a strain-generating body. A seated person 28 sits on the toilet seat 22 . The material of the toilet seat 22 is polypropylene resin. As shown in FIG. 5B, three sensors 10a to 10c are provided on the lower surface of the toilet seat 22 (surface on the -Z side). The sensors 10a-10c are strain gauges that detect the strain of the toilet seat 22. FIG. Sensor 10a is located at the front (+X) of toilet seat 22, and sensors 10b and 10c are located at the right (-Y) and left (+Y) of toilet seat 22, respectively. Layout W is the layout of the sensors 10a to 10c in FIG.

The output values of the sensors 10a to 10c were measured for five seated persons. FIGS. 6(a) to 6(d) are diagrams showing the output values with respect to time for each seated person A to D. FIG. The vertical axis represents the output values of the sensors 10a to 10c, and the unit is arbitrary coordinates [a. u. ]. The magnitude of the output values of the sensors 10a to 10c is the strain amount of the toilet seat 22 detected by the strain gauges, and correlates with the displacement amount of the toilet seat 22. FIG. The horizontal axis is time, and the seated person sat down at 10 seconds indicated by the arrow. Measurement data of four seated persons A to D are shown. The output value versus time for the fifth occupant has a trend between occupants A and B, but is not shown.

As shown in FIGS. 6(a) to 6(d), the output values of the sensors 10c of the seated persons A and B are larger than those of the sensors 10a and 10b. Seated persons C and D have similar output values from sensors 10a to 10c. From this, it can be seen that the centers of gravity of the seated persons A and B are located on the left side. In this way, the position of the center of gravity when the user is seated can be known from the output values of the sensors 10a to 10c. Further, by summing the output values of the sensors 10a to 10c, the amount of displacement corresponding to the weight of the seated persons A to D can be obtained.

For the seated persons A, B, and D, the increase in the output value of the sensor 10a after being seated is small, and the increase in the output values of the sensors 10b and 10c is large. From this, it can be seen that the seated persons A, B, and D sit slouching forward, and the center of gravity moves backward over time after being positioned forward. Seated person C has a small increase in the output values of the sensors 10b and 10c over time. From this, it can be seen that the seated person C does not tend to sit in a slouched position. In this way, the tendency of the output values of the sensors 10a to 10c (that is, the amount of displacement of the toilet seat 22) with respect to time indicates the characteristics of the sitting manner of the seated person.

FIG. 7 is a diagram showing the output value for each seated person's time. The vertical axis is the output value of the sensor 10b. As shown in FIG. 7, the output value at time t1 after a certain period of time T1 has passed since the load time t0 when the user was seated is defined as a displacement amount A1, and the output value at time t2 after a certain period of time T2 has passed is defined as a displacement amount A2. ΔA=A2−A1 corresponds to the toilet seat 22 creep value. A study was made to include the creep value ΔA in the feature quantity for machine learning.

Each of the five seated persons was seated 50 times, and the output values of the sensors 10a to 10c were obtained. The features used were the times T1 and T2 of 10 seconds and 40 seconds, respectively, the displacement amounts A1 and A2 and the creep value ΔA of each of the three sensors 10a to 10c, and the sum of A1 of the three sensors 10a to 10c. It is the sum SA2 of the values SA1 and A2. Let 3 A1, 3 A2, 3 ΔA, SA1 and SA2 be one set. 50 sets of data were prepared for each seated person. Of the 5×50 sets of data, 5×40 sets were used as teacher data, machine learning was performed, and a model was created. The remaining 5×10 sets of data were used as verification data, and the correct answer rate was defined as the rate at which the seated person answered correctly when the seated person was identified. As cross-validation, combinations of training data and verification data from 50 sets of data were performed under 20 conditions. Decision trees and random forests were used as machine learning methods. The depth of the tree was set to 5.

Table 1 is a table showing the average percentage of correct answers when cross-validation is performed under 20 conditions at the placement W of the sensors 10a to 10c.

"With ΔA" is the correct answer rate when machine learning is performed using three A1, three A2, three ΔA, SA1 and SA2 as feature amounts. “No ΔA” is the correct answer rate when machine learning is performed using three A1s, three A2s, SA1s and SA2s as feature amounts. "Only ΔA" is the percentage of correct answers when machine learning is performed using only three ΔAs as feature amounts. It shows the correct answer rate when using decision trees and random forests as machine learning methods.

As shown in Table 1, the correct answer rate is higher when random forest is used than when decision tree is used. A correct answer rate of 60% or more can be ensured even if only ΔA is used as a feature amount. When ΔA is added to the feature amount without ΔA, the correct answer rate is improved to about 80%. Thus, by using ΔA as a feature amount, the seated person can be specified. By adding ΔA to A1, A2, SA1 and SA2 and using it as a feature quantity, the correct answer rate is further improved.

The correct answer rate was examined by changing the arrangement of the sensors. Figures 8(a) to 8(c) show the placement of the sensors in locations X, Y and Z, respectively. As shown in FIG. 8(a), in arrangement X, five sensors 10a to 10c, 10f and 10g are arranged. The arrangement of the sensors 10a to 10c is the same as arrangement W in FIG. 5(b). Sensors 10f and 10g were placed on the right rear and left rear, respectively. As shown in FIG. 8B, in arrangement Y, five sensors 10a to 10e are arranged. The arrangement of the sensors 10a to 10c is the same as arrangement W in FIG. 5(b). The sensors 10d and 10e were arranged in front (+X direction) of the sensors 10f and 10g in the arrangement X of FIG. 8(a). As shown in FIG. 8(c), in arrangement Z, seven sensors 10a to 10g are arranged. The arrangement of the sensors 10a to 10g is a combination of the arrangement X of FIG. 8(a) and the arrangement Y of FIG. 8(b).

The percentage of correct answers was calculated for the placements W to Z. Table 2 is a table showing the percentage of correct answers for placements W to Z.

The correct answer rate in the upper row is the average correct answer rate of the 20 cross-validated conditions, and the correct answer rate in parentheses in the lower row is the lowest correct answer rate among the 20 cross-validated conditions. When there are 5 sensors, the feature amounts are 5 A1, 5 A2, 5 ΔA, SA1 and SA2, and when there are 7 sensors, the feature amounts are 7 A1 and 7 A2. , seven ΔA, SA1 and SA2.

As shown in Table 2, the correct answer rate improves as the number of sensors increases. In the random forest of configuration Z with 7 sensors, the average correct answer rate is 90%, and the lowest correct answer rate is 80%. Among five placements X and Y, the rate of correct answers for placement Y is high. As shown in FIGS. 8(a) and 8(b), the weight of the seated person 28 is less likely to be applied to the sensors 10g and 10f than to the sensors 10d and 10e. Therefore, placement X is considered to have a lower percentage of correct answers than placement Y.

According to the first embodiment, the identification unit 16 uses the creep value ΔA as a feature amount, and the learned model that has performed machine learning for identifying the seated person 28 is calculated from the displacement amount acquired by the acquisition unit 12. The seated person 28 is specified by inputting the creep value ΔA. As a result, the seated person can be identified as shown by ΔA in Table 1 only. Since a huge amount of teaching data as in Patent Document 1 is not required, processing can be performed in a short time.

Although one sensor may be provided, it is preferable to provide a plurality of sensors. By using a plurality of creep values ΔA corresponding to the plurality of sensors 10a to 10g as feature amounts to identify the seated person 28, the identification accuracy can be further improved. The number of sensors is preferably 3 or more, more preferably 5 or more. A sum of creep values of a plurality of sensors 10a to 10g may be used as the feature quantity.

Like arrangements W-Y, sensors 10a-10g preferably include three sensors provided at positions corresponding to the vertices of a triangle when viewed from above. Thereby, the seated person can be specified based on the characteristics of the movement of the center of gravity when the seated person sits down.

When the seated person 28 is seated, the sensors 10a to 10g are sensors 10a, 10g, or 10f (first sensor) provided in front or rear of the seat with respect to the seated person 28, and sensors 10b or 10d (first sensor) provided to the right of the seat. 2 sensor), preferably including sensor 10c or 10e (third sensor) provided to the left. Thereby, the seated person 28 can be identified based on the characteristics of the seating behavior of the seated person 28, such as whether the seated person 28 is bending forward or not, and whether the center of gravity is on either side. Comparing Arrangements X and Y, even if the sensors 10f and 10g are arranged behind the seated person 28 as in Arrangement X, the response rate is low. It is considered that this is because the load of the seated person 28 is not so much applied to the positions of the sensors 10f and 10g of the arrangement X. Thus, it is preferable to arrange the sensors 10a to 10g at positions where the load of the seated person 28 is large.

The sensors 10a to 10g include a sensor 10d (second sensor) provided on the right side of the seated person 28, a sensor 10e (third sensor) provided on the left side, and a sensor 10b (third sensor) that overlaps the right thigh of the seated person 28. 4 sensor) and sensor 10c (fifth sensor) overlying the left thigh. Thereby, the seated person 28 can be identified based on the characteristics of the weight change applied to the strain body 20 from the thighs.

A seated person 28 is specified using a plurality of displacement amounts A1 and A2 and a creep value ΔA corresponding to a plurality of sensors 10a to 10g as feature amounts. Thereby, the seated person 28 can be identified based on characteristics such as the position of the center of gravity when the seated person 28 is seated. Therefore, the accuracy of specifying the seated person 28 can be further improved. In order to reflect features such as the position of the center of gravity in machine learning, at least one of the displacement amounts A1 and A2 may be used as a feature amount.

Furthermore, a plurality of creep values ΔA corresponding to a plurality of sensors 10a to 10g and total values SA1 and SA2 are used as feature amounts to identify the seated person . Thereby, the seated person 28 can be identified based on characteristics such as the weight of the seated person 28 . Therefore, the accuracy of specifying the seated person 28 can be further improved. In order to reflect features such as the weight of the seated person 28 in machine learning, at least one of the total values SA1 and SA2 may be used as a feature amount.

In order to reflect both the position of the center of gravity of the seated person 28 and the weight of the seated person 28 in machine learning, a plurality of displacement amounts A1 and A2 corresponding to a plurality of sensors 10a to 10g, a plurality of creep values ΔA, a total value It is preferable to identify the seated person 28 using SA1 and SA2 as feature amounts.

[Modification 1 of Embodiment 1]
FIG. 9(a) is a perspective view of a sheet showing the arrangement of strain bodies in Modification 1 of Example 1, and FIG. 9(b) is a perspective view of the strain bodies. As shown in FIG. 9A, a pedestal 42, a seat cushion 44, and a seat back 46 are provided as a car seat 41 in a four-wheeled vehicle. The pedestal 42 is, for example, a metal member. The seat cushion 44 and the seat back 46 are cushion materials. Three strain-generating bodies 20 a are provided in the seat cushion 44 and two strain-generating bodies 20 b are provided in the seat back 46 . The strain bodies 20a are provided on the front, right and left sides of the center of gravity of the seated person. A load is applied to the strain body 20a mainly from the buttocks of the seated person. The strain-generating bodies 20b are provided on the lower left and right sides. A load is mainly applied to the strain body 20b from the back of the seated person. As shown in FIG. 9(b), the sensors 10 for detecting the strain of the strain-generating bodies 20a and 20b are provided on the strain-generating bodies 20a and 20b.

As in Modification 1 of Embodiment 1, the seat on which the strain-generating bodies 20a and 20b are provided may be a car seat 41. FIG. The strain-generating body 20 may be provided on a seat, such as a sofa or an office chair, on which a person sits. By providing the strain-generating body 20b at a location where the back of the seated person contacts, the accuracy of specifying the seated person is further improved. In this way, the sensor 10 (sixth sensor) may be arranged to overlap the back of the seated person when the seated person is seated.

[Modification 2 of Embodiment 1]
Modification 2 of Example 1 is an example of a method of determining the time t0 at which the extraction unit 13 is seated. FIG. 10(a) is an enlarged view of FIG. 7 near the time when the user is seated, and FIG. 10(b) is a diagram showing first-order differentiation of the output value of FIG. As shown in FIG. 10(a), for seated persons A and C, the output value overshoots after the time when the output value starts to rise (the time indicated by the arrow and the time when the horizontal axis is 10 seconds). Seated persons B and D do not overshoot. Also, the rise of the output value differs depending on the seated persons AD. As shown in FIG. 10(b), the first derivative of the output value overshoots after the time (arrow) when the output value begins to rise. The overshoot peak differs depending on the seated persons AD.

The time when the seated person 28 is seated on the seat includes the moment when a part of the seated person 28 touches the seat, the time when a predetermined percentage of the seated person's 28 weight, which is added to the seat when the seated person 28 is stabilized after sitting, is added to the seat, and the like. Various definitions are possible. Therefore, in FIG. 10A, the extraction unit 13 determines the time when the output value reaches the predetermined threshold value Th as the time t0 when the seated person 28 is seated. In FIG. 10A, the time when the output value of the seated person C reaches the threshold value Th is t0. If the seated person is light, such as a child, the output value may not reach the threshold Th. Therefore, as shown in FIG. 10(b), the extraction unit 13 may determine the time when the first derivative of the output value reaches a predetermined threshold value Th as the time t0 when the seated person 28 is seated. In FIG. 10(b), the time at which the first derivative of the output value of the seated person C reaches the threshold value Th is t0. Further, the extraction unit 13 may determine the time when the first derivative of the output value peaks as the time t0 when the seated person 28 is seated. The extracting unit 13 determines the time t0 each time it acquires output value data.

[Modification 3 of Embodiment 1]
Modification 3 of Example 1 is an example of a method of determining time t1 after the first time T1 has elapsed from time t0 at which the seated person sat down. We investigated the variability of the sitting data for multiple times. FIGS. 11(a) and 12(a) are diagrams showing multiple data of seated persons C and D, respectively, FIGS. 11(b) and 12(b) are respectively FIGS. FIGS. 11(c) and 12(c), which show the Max-Min output values at each time in (a), are enlarged views of FIGS. 11(b) and 12(b), respectively. Max-Min is the maximum output value minus the minimum output value at each time. Dots in FIGS. 11(c) and 12(c) indicate points for which Max-Min is calculated, and straight lines are lines connecting the dots.

As shown in FIGS. 11(a) and 12(a), the output value data varies. In the seated person C, the output value overshoots in almost all the data, while in the seated person D, the output value does not overshoot in almost all the data. As shown in FIGS. 11(b) and 12(b), Max-Min for both seated persons C and D becomes a maximum value M1 with the lapse of time, and then becomes a minimum value M2. Thus, Max-Min is overshooting. Seated person C's overshoot is larger than seated person's D overshoot.

If a time with a large Max-Min value is used as the first time t1, the accuracy of identifying the seated person will drop. Therefore, the first time t1 is set after the output value is stabilized. Specifically, the first time t1 is set to the time after the Max-Min (variation) of the output value at each time when the same seated person sits on the seat a plurality of times overshoots. For example, the first time t1 is determined so that the first time t1 is after the output value variation overshoots for all of the seated persons, based on the output values obtained when each of the plurality of seated persons is seated a plurality of times. . The determination of the first time t1 may be performed by the extraction unit 13, or the first time t1 may be set in advance. After determining the first time t1 and the second time t2, the same first time t1 and second time t2 are used for a plurality of seated persons.

The time after the overshoot may be the time after the Max-Min peak. For example, when K is 0 or more and less than 1, Max-Min becomes M1-K×(M1-M2). The time t1 may be set as the time when the time is reached. For example, K is preferably 0.5 or more, more preferably 0.8 or more. Time t1 may be obtained for a plurality of seated persons, and the latest time t1 may be set as the first time t1. Standard deviation or the like may be used as the variation at each time in addition to Max-Min. When the seated person sits on the seat, the first time T1 is preferably 4 seconds or longer, more preferably 5 seconds or longer, in order to stabilize the output value. If the first time T1 is too long, the output value will be saturated. Therefore, the first time T1 is preferably 20 seconds or less. If the difference between the second time T2 and the first time T1 (that is, t2-t1) is too small, the output value will not change much. Therefore, the difference between the second time T2 and the first time T1 is preferably 10 seconds or more. If the second time T2 is too long, the output value will be saturated. Therefore, the difference between the second time T2 and the first time T1 is preferably 100 seconds or less.

In the first embodiment and its modification, an example was described in which the identifying device 11 identifies a person sitting on the seat 25. good too. The object is the seat 25 in FIG. 1(a), and the toilet seat 22 in FIG. 5(a). The object may be, for example, a bed on which the subject lies, or a weight scale on which the subject stands.

Although the embodiments of the present invention have been described in detail above, the present invention is not limited to such specific embodiments, and various modifications and variations can be made within the scope of the gist of the present invention described in the scope of claims. Change is possible.

10, 10a to 10g sensor 11 identification device 12 acquisition unit 16 identification unit 17 machine learning unit 20, 20a, 20b strain body 22 toilet seat

Claims (12)

an acquisition unit that acquires a plurality of output values respectively output by a plurality of sensors that are provided in a strain body included in an object to which a subject applies at least part of the body weight and that detects the strain amount of the strain body;
a plurality of first values, which are the plurality of output values after a first time has passed since the subject added at least part of the body weight to the object; and a second value longer than the first time from the time. A plurality of second values, which are the plurality of output values after the lapse of time, and a plurality of third values, which are respectively corresponding differences, are used as feature amounts, and machine learning is performed to identify the subject. a specifying unit that specifies the target person by inputting the plurality of third values calculated from the output values acquired by the acquiring unit into the model of
A specific device with The identification device according to claim 1, wherein the plurality of sensors includes three sensors provided at positions corresponding to vertices of a triangle when viewed from above. 3. The identification device according to claim 1, wherein the object is a seat, and the object is seated on the object. When the subject is seated, the plurality of sensors includes a first sensor provided in front or rear of the subject, a second sensor provided on the right of the subject, and the subject. 4. The identification device of claim 3, including a third sensor located to the left of the . 5. The identification device according to claim 4, wherein the plurality of sensors includes a fourth sensor that overlaps the subject's right thigh and a fifth sensor that overlaps the subject's left thigh when the subject is seated. The identification device according to any one of claims 3 to 5, wherein the plurality of sensors includes a sixth sensor that overlaps the back of the subject when the subject is seated. The model includes a plurality of values of at least one of a plurality of first values corresponding to the plurality of sensors and a plurality of second values corresponding to the plurality of sensors, and a plurality of values corresponding to the plurality of sensors. A model that performs machine learning for identifying the subject using the third value as a feature amount,
The specifying unit inputs at least one of the plurality of values and the plurality of third values respectively corresponding to the plurality of sensors calculated from the plurality of output values acquired by the acquisition unit, The specifying device according to any one of claims 1 to 6, which specifies the target person. The model includes a plurality of values of the at least one, a plurality of third values, and a sum of at least one of a sum of the plurality of first values and a sum of the plurality of second values. A model that has been subjected to machine learning for identifying the subject as a feature amount,
The specifying unit is configured to provide at least one of the plurality of values, the plurality of third values, and at least one of the plurality of values corresponding to the plurality of sensors calculated from the plurality of output values acquired by the acquisition unit. The identification device according to claim 7, wherein the target person is identified by inputting a total value. The model includes the plurality of first values, the plurality of second values, the plurality of third values, the sum of the plurality of first values, the sum of the plurality of second values, and is a feature quantity, and a machine learning model for identifying the subject,
The identifying unit is configured to configure the plurality of first values, the plurality of second values, and the plurality of third values corresponding to the plurality of sensors calculated from the plurality of output values acquired by the acquiring unit. The identification device according to claim 8, wherein the target person is identified by inputting a value, a total value of the plurality of first values, and a total value of the plurality of second values. The specific device according to any one of claims 1 to 9, wherein the machine learning is machine learning using a decision tree or random forest. The specifying device according to any one of claims 1 to 10, further comprising a machine learning unit that performs machine learning for specifying the target person using the target person and the third value as feature amounts. an acquisition unit that acquires an output value output by a sensor that is provided in a strain body included in an object to which a subject applies at least part of the body weight and that detects the strain amount of the strain body;
a first value that is the output value at a first time after a first time has elapsed since the time when the subject applied at least part of the body weight to the object; A third value, which is the difference between the second value, which is the output value at the second time after the second time has elapsed, and the third value, which is the corresponding difference, is used as a feature amount, and machine learning is performed to identify the subject. a specifying unit that specifies the target person by inputting the third value calculated from the output value acquired by the acquiring unit into the model of
with
The specific device, wherein the first time is a time after the variation in the output value at each time overshoots when the same subject applies at least part of the weight a plurality of times.

Images