JP6449355B2 - Method, program, and apparatus for detecting state of moving object - Google Patents

Description

  The present invention relates to a method for detecting a state of a moving body using a sensor mounted on a mobile terminal device or the like. In particular, the present invention relates to a state detection method, a program, a portable terminal device, and the like that can detect in real time and with high accuracy whether the user moves by walking or by a specific vehicle.

  With the widespread use of multifunctional mobile terminal devices such as smartphones, attempts have been made to use data for individuals and / or many people by analyzing the data of sensors mounted on terminals and to make use of them for various services. . As sensor data, data acquired from an acceleration sensor, a geomagnetic sensor, a positioning sensor, or the like is used.

  Conventionally, a biaxial correlation value calculating means for calculating a correlation value indicating a relationship of motion between two axial directions of a plurality of axial directions based on information obtained by detecting the movement of a moving object in a plurality of axial directions And a movement status determination unit that calculates a feature value for each axial direction of a plurality of axial directions, and a movement status determination unit that determines the movement status of the movement target based on the feature value and the correlation value An apparatus has been proposed (Japanese Patent Laid-Open No. 2013-195363: Patent Document 1).

  A state in which the geomagnetic data indicating the geomagnetic value detected by the geomagnetic sensor is acquired, and the user carrying the geomagnetic sensor is moving by car based on the magnitude of the change in the geomagnetic value indicated by the acquired geomagnetic data Or a method for determining whether the vehicle is moving by train or the like (Japanese Patent Laid-Open No. 2014-66638: Patent Document 2).

  An acceleration sensor for detecting acceleration, a magnetic sensor for detecting magnetism, an acceleration detected by the acceleration sensor, and a magnitude of magnetism detected by the magnetic sensor for distinguishing and detecting a vehicle in a vehicle moving state There has been proposed an electronic device including a control unit that determines a type of movement state based on the change in the length (Japanese Patent Laid-Open No. 2015-88905: Patent Document 3).

  In addition, a series expression extraction unit that extracts a series expression in which all the positioning points have characteristics such as speed and acceleration based on the GPS trajectory segment, and a series feature quantity extraction model generation that generates a series feature quantity extraction model from the series expression And a sequence feature amount extraction unit that extracts a sequence feature amount using the sequence feature amount extraction model generated by the sequence feature amount extraction model generation unit, and an estimation model generation unit that generates an estimation model A means estimation model generation apparatus has been proposed (Japanese Patent Laid-Open No. 2006-99110: Patent Document 4).

JP 2013-195363 A JP 2014-66638 A Japanese Patent Laying-Open No. 2015-88905 JP-A-2016-99110

  In the prior art, processing is likely to be complicated, such as stepwise or alternative determination of a detectable state in order to detect a plurality of states. Further, in the stepwise determination, each step may include erroneous detection, and the accuracy is not sufficient in the end. Moreover, in implementation in portable terminal devices, such as a smart phone, it was not fully examined about performing a detection easily, without applying load to CPU etc., and was lacking in practicality.

  In view of the above problems, an object of the present invention is to provide a state detection method suitable for mounting on a portable terminal device or the like. It is intended to provide a state detection method, program, and apparatus that can be used for various purposes such as monitoring of daily activities as well as collecting life logs, etc. by enabling real-time and high-precision detection. To do. It is another object of the present invention to provide effective feature extraction and model design for improving detection accuracy.

  One embodiment according to the present invention is a method for detecting a state of a moving body, the step of calculating a plurality of feature amounts at predetermined time intervals from acquired sensor values, and the calculated plurality of features Each of calculating a plurality of state scores relating to a plurality of states that can be detected at the first time using the feature amount and at least one calculation model generated in advance, and based on the plurality of state scores, Determining a single state at a first time.

  According to the present invention, the state of the moving body is uniquely identified based on the plurality of state scores. Since a calculation model for calculating the state score is generated in advance, it is not necessary to perform complicated determination processing and the like, and detection with reduced false detection can be easily performed.

  A single state may be determined sequentially at multiple times and a single state start, a single state continuation, and / or a single state end may be determined.

  According to the present invention, for example, it is possible to detect in real time that the moving body has changed from a “walking movement” state to a “car movement” state. For example, if the moving body is a child, the guardian may be “on schedule” or “an unexpected situation may have occurred where it should have moved on foot, depending on the detection notification settings, etc. ”Can be grasped.

  The at least one calculation model preferably includes weights, biases, and calculation algorithms that are pre-learned to specify a magnitude relationship between state scores.

  By doing so, for example, even if there are many detection target states (3 or more, 4 or more, 5 or more, 10 or more, etc.), the state at a certain time (period) is uniquely determined without contradiction. obtain.

  The at least one calculation model is three or more models, a first model for calculating a first score from the first set of the plurality of feature quantities, and a second of the plurality of feature quantities. A second model for calculating a second score from the set, and a third model for calculating a plurality of state scores using at least the first score and the second score as feature quantities may be included.

  As described above, detection accuracy can be improved with one model using a feature amount (set) that is optimal for target detection, and false detection can be further reduced by adjustment between models.

  A single state may be determined by at least a state score exceeding a predetermined threshold. According to the present invention, accurate determination is possible by determining the state score based on a predetermined threshold.

  The sensor value includes at least a triaxial acceleration sensor value, and the plurality of feature amounts include an amplitude for each time of at least one acceleration frequency.

  The sensor value includes at least a triaxial acceleration sensor value and a triaxial geomagnetic sensor value, and the plurality of feature amounts include values related to respective changes in the horizontal component and the vertical component of the magnetic amount.

  By doing in this way, the feature-value suitable for the state determination for every time is extracted.

  The sensor value includes at least speed data, and the speed data may be referred to in determining a single state.

  In this way, false detection can be further reduced.

  Another aspect of the present invention is a program that causes a computer to execute the state detection method.

  Another aspect of the present invention is a state detection device including one or more sensors for detecting a state of a moving body, and a sensor value for acquiring a sensor value from the one or more sensors. Acquisition means; feature quantity calculation means for calculating a plurality of feature quantities at predetermined time intervals from the acquired sensor values; and database means for storing at least one calculation model and state determination rule generated in advance. A score calculating means for calculating a plurality of state scores relating to a plurality of states that can be detected at the first time using the calculated plurality of feature amounts and a calculation model; a plurality of state scores and state determination rules; And a state determination means for determining a single state at the first time.

  According to the present invention, it is possible to easily detect a moving state using a portable terminal device or the like. Since the detection is almost real time and the accuracy is sufficiently high, it can be used for various purposes.

FIG. 1 is a schematic block diagram of a state detection apparatus according to the present invention. FIG. 2 is a diagram for explaining feature amount calculation according to the present invention. FIG. 3 is a schematic block diagram of the state detection model generation apparatus according to the present invention. 4A shows the basic feature DB, FIG. 4B shows the feature DB, FIG. 4C shows the first model feature DB, and FIG. 4 shows the data structure of the second model feature DB. FIG. 5 is a diagram for explaining the correlation between the feature amount and the state. FIG. 6A shows a parameter set of the first model DB, and FIG. 6B shows a parameter set of the second model DB. FIG. 7 shows the data structure of the score DB. FIG. 8 shows a parameter set of the third model DB. FIG. 9 shows time series data of acceleration sensor values. FIG. 10 shows time-series data of geomagnetic sensor values. FIG. 11 shows time-series data of feature amounts. FIG. 12 shows time-series data of feature amounts. FIG. 13 shows time-series data of scores calculated from feature amounts. FIG. 14 shows time-series data of scores calculated from feature amounts. FIG. 15 shows time-series data of the final score and speed calculated using the score as a feature amount.

  Various features of the present invention will now be described with reference to the drawings, together with preferred embodiments not intended to limit the invention.

  FIG. 1 schematically shows a block diagram of a state detection apparatus 100 according to one embodiment of the present invention. The state detection device 100 may be a known small computer device such as a smartphone or a tablet PC.

  The state detection device 100 includes a sensor unit 10, a sensor value acquisition unit 20, a feature amount calculation unit 30, a score calculation unit 40, a state determination unit 50, a database unit 60, in order to detect the movement state of the moving object (user). And communication means 70. The state detection device 100 includes a CPU, a RAM, a ROM, an OS, an application, and the like (not shown), and each of the above units can be implemented by hardware and software (electronic device, electronic circuit, program, etc.).

  The sensor means 10 may include a known triaxial acceleration sensor, triaxial geomagnetic sensor, and positioning sensor. The triaxial acceleration sensor can measure acceleration (AXt, AYt, AZt) (G) in the X-axis, Y-axis, and Z-axis directions at time (t). The triaxial geomagnetic sensor can measure magnetic quantities (MXt, MYt, MZt) (μT) in the X-axis, Y-axis, and Z-axis directions at time (t). The positioning sensor can acquire the latitude and longitude at the time (t), the instantaneous speed (m / s) of the currently facing device, and the like based on GPS positioning, Wi-Fi positioning, base station positioning, and the like. it can. The speed data is obtained directly from the positioning sensor value, or is calculated based on position information (latitude, longitude) of the positioning sensor.

  The sensor value acquisition unit 20 can acquire a sensor value from the sensor unit 10. The feature amount calculation unit 30 can calculate the feature amount from the sensor value. The DB unit 60 stores a score calculation model 61 and a state determination rule 62 generated in advance. The score calculation means 40 can calculate the state score with reference to the score calculation model 61 with the calculated feature quantity as an input. The state determination unit 50 receives the calculated state score as an input, refers to the state determination rule 62, determines the state of the moving body at a certain time, and can output it to an output unit (not shown) such as a display. . The determination result 63 is stored in the DB means 60. The communication means 70 may include a mailer or the like, and can appropriately notify the determination result 63 to a set communication destination or the like. For example, a detection of a specific state (for example, “get on the train”, “get off the train”, etc.) can be sent by e-mail for a guardian who wants to watch the child go to school.

  FIG. 2 shows an outline of feature amount calculation according to one embodiment of the feature amount calculation unit 30. In the figure, S1, S2,..., S1 ′, S2 ′,... Represent sampling times according to a predetermined sampling rate. In the illustrated example, S1, S2,... Correspond to a sampling rate of 200 ms, and 5 samples can be acquired per second (5 Hz). S1 ′, S2 ′,... Correspond to a sampling rate of 2000 ms, and one sample can be acquired every 2 seconds from the start of sampling. The sampling rate is not limited to these, but the sampling rate is preferably 10 Hz, 5 Hz, or 1 Hz or less in order to reduce the load on the CPU, increase the processing speed, reduce power consumption, and the like. According to the method of the present invention, the state of the moving body can be sufficiently detected with a small number of samples.

  The sampling may be started when the user inputs an application activation instruction. Alternatively, it may be in a constantly activated state.

  As shown in the figure, the feature amount calculation unit 30 calculates the combined acceleration from the triaxial acceleration sensor value at each sampling time, calculates the amplitude of each frequency at a certain time using a plurality of past data, and further calculates the frequency of each frequency. The moving average of the amplitude can be calculated as the feature amount (A, B, C, D, E).

Specifically, with the acceleration sensor values at the sampling time t as AXt, AYt, and AZt, the acceleration composite value (without gravity value) ACCt is calculated as follows.

The amplitude at each frequency is calculated by performing a Fourier transform based on the past 1.8 seconds (each 9 data) of ACCt.

Of the obtained (facc (0) t, ..., facc (8) t), (|| facc (0) t ||, ..., || facc (4) t ||) It becomes the amplitude of each frequency spectrum at the sampling time t. Further, a moving average of (|| facc (0) t ||,..., || facc (4) t ||) for the past 10 seconds (50 data) is calculated.

  The calculated five values are defined as a feature amount A, a feature amount B, a feature amount C, a feature amount D, and a feature amount E, respectively.

  In the example of FIG. 2, sampling is started in S1, each feature amount (A, B, C, D, E) is calculated in S58, and thereafter, each feature is sequentially performed in steps of S59, S60,. A quantity is calculated.

  Further, the feature amount calculation unit 30 calculates the pitch angle and roll angle of the terminal from the average value of acceleration, calculates the horizontal component and the vertical component from the triaxial geomagnetic sensor value, and calculates the value calculated at the previous time. Each difference is taken, and a moving average value and a standard deviation of a plurality of differences can be calculated as feature amounts.

Specifically, in order to calculate the horizontal component and the vertical component of the magnetic quantity, the inclination (pitch angle PITt, roll angle ROLt) of the sensor-equipped terminal is calculated from the triaxial acceleration sensor value at the sampling time t. From the pitch angle PITt, the roll angle ROLt, and the triaxial geomagnetic sensor values (MX _T , MY _T , MZ _T ) at time _T , the sensor values (MGX _T , MGY _T , MGZ _T ) when the vertical direction is the Z axis. Is calculated. Time t represents a time with a sampling rate of 200 ms, time T represents a time with a sampling rate of 2000 ms, and time t and time T at the time of calculation indicate the same time.

Based on the sensor values (MGX _T , MGY _T , MGZ _T ) when the vertical direction is the Z axis, the horizontal component MH _T and the vertical component MV _T of the magnetic sensor value are calculated.

Further, the absolute value (MHΔ _T , MVΔ _T ) of the difference between (MH _T , MV _T ) and (MH _T−1 , MV _T−1 ) is calculated.

According to the time series, the feature value 1, feature value 2, feature value 3, and feature value 4 are calculated by calculating the mean value and standard deviation of the past 120 seconds (each 60 data) of (MHΔ _T , MVΔ _T ). Is done.
(Feature 1)
(Feature 2)
(Feature 3)
(Feature 4)

  In the example of FIG. 2, each feature amount (1, 2, 3, 4) is calculated at the sampling time S61 ′ (S610), and thereafter, each feature amount is S62 ′, S63 ′,. Can be obtained.

  Furthermore, the feature amount calculation unit 30 can calculate the moving average and standard deviation of the difference of the combined acceleration as the feature amount. The combined value and difference (absolute value), average (feature value 5), and standard deviation (feature value 6) of the triaxial acceleration sensor values at time T are calculated as follows.

(Feature 5)
(Feature 6)

  In the example of FIG. 2, the average and standard deviation of the past 120 seconds (60 data) are calculated at the sampling time S61 ′, and thereafter, the feature amount 5 and the feature amount 6 are calculated in order S62 ′, S63 ′,. Is done.

  Note that how many data the moving average is in the past may be set as appropriate, and may be smaller or larger than the above example.

  The score calculation unit 40 (FIG. 1) uses the calculated feature quantities A, B, C, D, E and the feature quantities 1, 2, 3, 4, 5, 6 to determine the probability of the state (reliability level). ) To calculate the state score associated with The calculation model 61 is referred to for the calculation.

  The calculation model 61 is generated in advance and stored in the DB means 60. A block diagram of a state detection model generation device 200 of one embodiment for generating the calculation model 61 is schematically shown in FIG.

  The state detection model generation device 200 is configured by a computer, and includes a labeled basic feature DB 201, a feature extraction unit 202, a labeled feature DB 203, a model DB generator 204, a first model feature DB 205, and a first model. The generation unit 206, the first model DB 207, the second model feature DB 208, the second model generation unit 209, the second model DB 210, the score DB 211, the third model generation unit 212, the third model DB 213, the determination rule DB 214, and the input An output unit 215 is provided. The state detection model generation apparatus 200 includes hardware such as a CPU, RAM, ROM, and HDD (not shown), and software such as an OS and applications, and each DB and each unit are implemented by cooperation of hardware and software. The vehicle detection model generation apparatus 200 may be configured by a single computer or may be configured by a plurality of computers.

  FIG. 4A shows the data structure of the labeled basic feature DB 201. As shown in the figure, the labeled basic feature DB 201 stores the time t, the triaxial geomagnetic sensor value, the triaxial acceleration sensor value, the speed, and the basic label. The basic label may arbitrarily represent the actual state of the moving object at time t, and may be, for example, a subway, an elevated railway, a bullet train, a monorail, a bus, a passenger car, a walk, a run, a bicycle, or a stop. .

  The feature amount extraction unit 202 can extract various feature amounts for generating a model. The feature quantity extraction unit 202 at least in the same manner as the feature quantity calculation means 30, the feature quantities A, B, C, D, E and the feature quantities 1, 2, 3, E, from the basic feature quantity (FIG. 4A). 4, 5, 6 can be calculated.

  The calculated feature amount is stored in the labeled feature amount DB 203. FIG. 4B shows the data structure of the labeled feature DB 203. As shown in the figure, each feature amount is stored in association with time. When the sampling rate of the feature amount is different (time t, time T), data may be stored for each sampling rate.

  The feature extraction of one embodiment is designed to reduce false positives. For example, the inventor according to the present invention has found that there is a difference in the change in the vertical component and the horizontal component of the magnetic quantity, and these changes well reflect “train movement” and “car movement”.

  FIG. 5 shows a diagram in which corresponding states are plotted with the horizontal axis representing the standard deviation of the change (difference) in the horizontal component of the magnetic quantity and the vertical axis representing the standard deviation of the change (difference) in the vertical component of the magnetic quantity. In the figure, open circles represent “car (including bus and passenger car) movement”, light gray × represents “train (including subway and elevated railway) movement”, and dark gray dots Indicates “walking or stopping”. As illustrated, “car movement” and “train movement” are well separated. Therefore, the standard deviation of the change (difference) of the horizontal component of the magnetic quantity and the standard deviation of the change (difference) of the vertical component can be feature quantities for identifying “car” and “train”.

  As shown in FIG. 5, “walking and stopping, etc.” (dark gray points) overlap with “car movement” and “train movement”, and there is a risk of erroneous detection in the analysis based on the feature amount. For example, even if it is determined that the train is a “train” or a “car”, it may actually be “walking” or “stop”.

  The detection model according to the present invention generates a model for each specific state when a specific feature amount (set) is advantageous for detection of a specific state, and detects one model by other models. By reducing it, it is generated so as to improve the detection accuracy as a whole.

  In one embodiment, the 1st model and the 2nd model are generated, and the 3rd model is generated based on these. In one embodiment, the first model uses features A, B, C, D, and E to detect “walking” and the like, and the second model uses features 1, 2, 3, 4, 5 , 6 is used to detect “car”, “train”, and the like.

  The model DB generation unit 204 can select a feature amount suitable for generation of each model, and can appropriately assign a label.

  FIG. 4C shows the data structure of the first model feature DB 205. In one embodiment, the selected feature quantities A, B, C, D, E and the label 1 are stored for each time in order to generate a model for detecting “walking”. Label 1 can be determined by the basic label and the subject of score calculation (eg, “walking”). For example, when the basic label at time t means “walking”, the label 1 information becomes “walking”, and when the basic label at time t does not mean “walking” (train, car, stop, etc.), label 1 information Can be “other”.

  The first model generation unit 206 generates a first model by learning the feature amount read from the first model feature amount DB 205 and the label 1 information. For the generation of the first model, logistic regression, neural network (NN), support vector machine (SVM), known libraries such as deep learning (for example, LIBSVM, LIBLINER, CAMFE, TENSORFLOW, CHAINER, etc.) can be used. . As a result of learning, for example, when a certain time has a “walking” label, a parameter set (weight) for the “walking score” is set so that the score of the time is higher than the score of the time having the “other” label. , Bias, calculation algorithm).

FIG. 6A shows the determined parameter set stored in the first model DB 207. In one embodiment, two sets of parameters for calculating a “walking score” representing the reliability of the walking state and a “running score” representing the reliability of the running state are stored. The weights of the feature amounts A, B, C, D, and E for calculating the “walking score” are W _WA , W _WB , W _WC , W _WD , W _WE , the bias is b _W , and the score calculation algorithm is shown in the f _W. Further, the weights of the feature amounts A, B, C, D, and E for calculating the “running score” are W _RA , W _RB , W _RC , W _RD , W _RE , the bias is b _R , and the algorithm is f _Indicated by _R. Instead of or in addition to the above, a set of parameters for calculating a “walking score” that integrates “walking” and “running” may be stored. In the following description, a “walking score” is used as appropriate.

  FIG. 4D shows the data structure of the second model feature DB 208. In one embodiment, for detection of “car” or “train”, the selected feature quantities 1, 2, 3, 4, 5, 6 and the label 2 are stored every time (T). Label 2 can also be determined by the base label and the subject of score calculation (eg, “car”). For example, when the basic label at time T (t) means “car” to be detected, the label 2 information is “car”, and the basic label at time T (t) does not mean “car” to be detected ( In the case of train, walk, stop, etc.), the label 2 information can be “other”.

  The second model generation unit 209 generates a second model by learning the feature amount read from the second model feature amount DB 208 and the label 2 information. As described above, a known machine learning library or the like can be used for generating the model. For example, in the time-series data of “car score”, the parameter set is learned so that the time score having the “car” label is higher than the score of the time having the “other” label. In one embodiment, a parameter set for three scores of “car”, “train” and “walk, stop, etc.” is learned.

  FIG. 6B shows a learned parameter set stored in the second model DB 210. The data structure has a weight, a bias, and a calculation algorithm for each of the feature amounts 1, 2, 3, 4, 5, and 6 for each score to be calculated. As a result, time-sequential data of each score representing a predetermined state likelihood is obtained from the feature quantities 1, 2, 3, 4, 5, 6 for each time.

  FIG. 7 shows the data structure of the score DB 211. For the generation of the third model, the obtained score is stored every time (t = T). In one embodiment, a “walking score” is calculated for time t, and a “car score”, “train score”, and “walking / stopping score” are calculated for time T. For the purpose of detecting a state corresponding to time, time t and time T are associated with each other. For example, in one embodiment, if time t is 200 ms and time T is 2000 ms, a moving average of 10 samples of walking scores is taken and each score is associated with a common time (t = T). .

  The state label can be determined by the basic label and the state to be detected. In one embodiment, the detection target states are “walking (including walking and running)”, “car (including bus and passenger car)”, and “train (including subway and elevated railway)”. Therefore, for example, when the basic label at a certain time (t = T) means “walking”, the state label becomes a “walking” label, and the basic label at a certain time (t = T) does not mean “walking” ( In the case of trains, cars, stops, etc., it can be an “other” label.

  When taking the moving average or the like, when the same time (period) includes a plurality of basic labels, a label for generating a model may be determined according to any of the basic labels. For example, a label for generating a model may be determined based on the most basic labels, the basic label at the start time, and the basic label at the end time.

  The third model generation unit 212 can learn each score (feature amount) and state label information read from the score DB 211 using a known machine learning library or the like to generate a third model.

FIG. 8 shows a parameter set of the third model of one embodiment stored in the third model DB 213. In one embodiment, the parameter set includes the original four scores as feature amounts, weights for each feature amount (W ₁ , W ₂ , W ₃ ,...), And bias (b ₁ , b ₂ ,. ..) And a calculation algorithm (f ₁ , f ₂ ,...) Are included for each score to be calculated.

  By doing in this way, determination of the several state detected by a different feature-value is matched. The state score output by the third model can clearly indicate the magnitude relationship between the scores so as to appropriately reflect the state of time. Any state is not confusing, and even if there are many states to be detected (for example, three or more, four or more, six or more, or ten or more), the state can be determined as a single unit. A large number of test data (not shown) tested the state score based on the third model, and it was confirmed that at least “car”, “train”, and “walk” were sufficiently separated.

  Note that feature quantity extraction and model generation are not limited to the above embodiment. As long as it matches the purpose of accurately detecting the state associated with the time, the feature amount, the combination thereof, the number of models, and the like may be arbitrary.

The determination rule DB 214 can define a state determination rule based on the score according to the third model. For example, the rules may include:
L1. When the walking score at a certain time is equal to or greater than a predetermined threshold, the state at that time is determined to be walking movement.
L2. When the train score at a certain time is greater than the car score and equal to or greater than a predetermined threshold, the state at that time is determined as the train movement state.
L3. When the car score at a certain time is greater than the train score and is equal to or greater than a predetermined threshold, the state at that time is determined as the vehicle movement state.
L4. If any score at a certain time falls below each threshold value, it is determined that it is not in any detection target state (stop, etc.).
Priorities may be defined above, for example, rule L1 may take precedence over rules L2 and L3.

According to the third model of one embodiment, erroneous detection of “walking movement” as “car movement” or “train movement” is reduced, but “stop” is referred to as “car movement” or “train movement”. It cannot be said that there is no risk of false detection of “movement”. For example, it is difficult to detect a state that moves slowly in a line as “walking movement”. Therefore, by detecting only “walking movement”, the “almost stopped” state can be erroneously detected as “car movement” or “train movement”. . Thus, the decision rule may include the following to utilize speed data.
L5. When it is determined as “car movement” or “train movement” based on the score, if the speed data is below a predetermined threshold, it is not determined as “car movement” or “train movement”, but “stop etc.” decide.

Furthermore, the judgment rule is actually “car movement” or “train movement”, so that it is not erroneously detected as “stop etc.” only by speed data, or the state of “train” etc. Since it is assumed that it will persist, using past determination results may be included.
L6. Only when the state determined immediately before is “stop, etc.”, the L5 rule is adopted with reference to “speed data”.
L7. When the state determined immediately before is a “detection target state” such as a car, a train, or a walk, in the subsequent determination, the threshold value that is the “detection target state” is a threshold value 2 that is reduced from the initial threshold value 1. Change to

  The determination rules (threshold value, etc.) stored in the determination rule DB 214 may be appropriately set by an administrator or the like based on the interpretation of the output result of the calculation model. The administrator can change, add, delete, etc. the determination rule via the input / output unit 215 of the state detection model generation device 200.

  The DB means 60 (FIG. 1) of the state detection apparatus 100 includes a calculation model 61 (for example, the first model, the second model, and the third model) determined in advance as described above, and a determination rule 62 (for example, L1 to L1). L7) is stored.

  The score calculation means 40 refers to the calculation model 61 and calculates each score from the calculated feature amount.

  The state determination unit 50 refers to the determination rule 62 and determines the state from the calculated score. The determination result 63 is sequentially stored in the DB means 60. The determination result 63 may be output for each determination (real time) to output means (display or the like) (not shown). An output time interval or the like may be arbitrarily set by user input.

An embodiment of the state detection method according to the present invention using a series of test data is shown below. A summary of the test is shown in Table 1 below.

  FIG. 9 shows time-series data of the acquired triaxial acceleration sensor values. For the sake of explanation, the correct state (value on the vertical axis—1: walking, −4: train, −8: car) is shown in the lower bar graph (the same applies to the following figures).

  FIG. 10 shows time-series data of the triaxial geomagnetic sensor values acquired.

  FIG. 11 shows time-series data of the calculated feature amounts A, B, C, D, and E (moving average of the amplitudes of the five frequencies).

  FIG. 12 shows the calculated feature quantity 1 (average value of magnetic quantity horizontal component difference), feature quantity 2 (average value of magnetic quantity vertical component difference), feature quantity 3 (standard deviation of magnetic quantity horizontal component difference), feature Time series data of a quantity 4 (standard deviation of the magnetic quantity vertical component difference), a feature quantity 5 (average value of the synthesized acceleration difference), and a feature quantity 6 (standard deviation of the synthesized acceleration difference) are shown. When the actual state is “train”, it is observed that the feature amount 2 (average value of the magnetic amount vertical component difference) and the feature amount 4 (standard deviation of the magnetic amount vertical component difference) are remarkably increased. The Further, when the actual state is “walking”, it is observed that the feature amount 5 (average value of the combined acceleration difference) and the feature amount 6 (standard deviation of the combined acceleration difference) are projected.

  FIG. 13 shows a walking score (solid line) and a running score (broken line) calculated by the first model of the example. Even if the correct answer was "walking", the "running" score tended to be slightly higher. Separately, a “walking score” (not shown) in which “walking” and “running” are integrated obtained a score that well reflects the correct answer (walking = walking).

  FIG. 14 shows the vehicle score (fine dotted line), train score (dashed line), and walking / stop score (solid line) calculated by the second model of the example. The magnitude relationship between the scores generally represents the actual state. However, there were times when walking and stopping scores were not well separated.

  FIG. 15 shows time series data of each score and speed calculated by the third model of the example. The determination result using the determination rule of the embodiment (value on the vertical axis 1: walking determination, 4: train determination, 8: car determination) is shown in dark gray and contrasted with the correct answer (light gray). As shown, 100% of the gait was detected. Trains and cars have an accuracy of almost 100% for the period when they are detected as “cars” and “is a train”, and over-detection (false positive) was suppressed. Although the correct answer was “train” or “car”, there was a false negative that was not detected, but the time was extremely short, from a few seconds to a few minutes, and it was sufficiently practical.

  As described above, according to the present invention, it is possible to detect a state with high accuracy in real time. For example, the first detection of the “train” state can be output as “(now) on the train”, and when another state is detected following the “train” state, the “(now) get off the train” Can be output. The form of output is not limited, and various information related to the state associated with the time can be presented.

  The state detection method according to the present invention can be installed in a known mobile terminal device such as a smartphone in the form of an application program. The load relating to the calculation and determination processing is small, high-speed and high-precision determination is possible, and the utility is excellent.

  Those skilled in the art will appreciate that many different modifications are possible without departing from the spirit and aspects of the invention. Accordingly, it goes without saying that the embodiments of the present invention are merely examples, and do not limit the scope of the present invention.

DESCRIPTION OF SYMBOLS 100 State detection apparatus 10 Sensor means 20 Sensor value acquisition means 30 Feature-value calculation means 40 Score calculation means 50 State determination means 60 Database means 61 Score calculation model 62 State determination rule 63 Determination result 70 Communication means

Claims (9)

A method for detecting the state of a moving object in real time ,
Calculating a plurality of feature amounts at predetermined time intervals, respectively, from the acquired sensor values;
Calculating a plurality of state scores relating to a plurality of states that can be detected at the first time using the plurality of calculated feature quantities and a plurality of calculation models generated in advance;
Based on said plurality of status score, it viewed including the steps of: determining a single state in the first time,
The plurality of calculation models include a specific state score calculation model for calculating a specific state score related to a specific state using a feature amount selected from the plurality of feature amounts, and the plurality of calculation models using the calculated specific state score. A state score calculation model for calculating the state score of
A state detection method , wherein a state score related to the single state among the plurality of state scores at the first time is higher than other state scores in a time series including the first time .   Each single state at a plurality of times is determined sequentially, thereby determining the start of a single state, the continuation of a single state, and / or the end of a single state The state detection method according to claim 1. The plurality of calculation models are three or more models;
A first model for calculating a first score from a first set of the plurality of feature quantities;
A second model for calculating a second score from a second set of the plurality of feature quantities;
The state detection method according to claim 1, further comprising: a third model for calculating the plurality of state scores using at least the first score and the second score as feature amounts.   The state detection method according to claim 1, wherein a state score of the determined single state exceeds at least a predetermined threshold value. The sensor value includes at least a triaxial acceleration sensor value;
Wherein the plurality of feature amounts, state detection method according to claim 1, characterized in that it comprises a moving average of the amplitude of at least one of the acceleration frequency at a certain time. The sensor value includes at least a triaxial acceleration sensor value and a triaxial geomagnetic sensor value;
2. The state detection method according to claim 1, wherein the plurality of feature amounts include values related to changes in a horizontal component and a vertical component of a magnetic amount. The sensor value includes at least velocity data;
The state detection method according to claim 1, wherein speed data is referred to in the step of determining the single state. The program which makes a computer perform the state detection method in any one of Claim 1 thru | or 7 . A state detection device including one or more sensors for detecting a state of a moving object in real time ,
Sensor value acquisition means for acquiring sensor values from the one or more sensors;
Feature quantity calculating means for calculating a plurality of feature quantities at predetermined time intervals from the acquired sensor values;
Database means for storing a plurality of pre-generated calculation models and state determination rules;
Score calculating means for respectively calculating a plurality of state scores relating to a plurality of states that can be detected at a first time using the plurality of calculated feature quantities and the calculation model;
A state determination unit that determines a single state at the first time based on the plurality of state scores and the state determination rule ;
The plurality of calculation models include a plurality of specific state score calculation models that respectively calculate specific state scores relating to a specific state from feature amounts selected from the plurality of feature amounts, and the plurality of calculation states based on the calculated specific state scores. A state score calculation model for calculating the state score of
A state detection device , wherein a state score related to the single state among the plurality of state scores at the first time is higher than other state scores in a time series including the first time .