JP6369787B2 - Signal processing device, detection device, and program - Google Patents

Description

  The present invention generally relates to a signal processing device, a detection device, and a program, and more particularly to a signal processing device, a detection device, and a program that process a sensor signal from a sensor that receives a radio wave reflected by an object. is there.

  Conventionally, there is a human body detection device using a Doppler signal of a propagation wave reflected by a detection target.

  For example, in Patent Document 1, the Doppler signal is analyzed using an autoregressive model, the peak frequency and amplitude of the Doppler signal are calculated, and the human behavior state is determined based on the peak frequency and amplitude of the Doppler signal. A sensing device is disclosed.

JP 2010-85100 A

  For example, when detecting human breathing, the sensor signal is a low-frequency non-sinusoidal wave. However, in the prior art, it is difficult to accurately detect the target motion from the low-frequency non-sinusoidal sensor signal, and there are unreported and target motions that are not detected despite the target motion occurring. There was a possibility that a false detection would occur despite the absence.

  Conventionally, there has been a desire to improve the efficiency of target motion detection processing.

  The present invention has been made in view of the above-mentioned reasons, and its object is to detect a target motion with high accuracy from a low-frequency non-sinusoidal sensor signal and to improve the efficiency of the target motion detection process. An object is to provide a signal processing device, a detection device, and a program that can be improved.

The signal processing apparatus of the present invention applies an autoregressive model over an analysis period to a sensor signal corresponding to the movement of the target object output from a sensor that receives a radio wave reflected by the target object. Analysis processing to obtain a regression coefficient vector consisting of the autoregressive coefficient of the analysis unit, the analysis period is sequentially shifted by a shift time from the previous analysis period, and the shift time is set shorter than the time length of the analysis period, A database storing reference data for identifying the regression coefficient vector corresponding to the sensor signal when a target motion, which is a specific motion of the target object, is generated; and the analysis unit obtains the database by matching the regression coefficient vector and the reference data, the recognition unit for determining the presence or absence of the target operation, the recognition unit judging the pair is set to the target Depending on the operation, characterized in that it comprises an adjustment unit for changing the length of time and the shift time of the analysis period.

  The detection apparatus of the present invention includes a sensor that receives a radio wave reflected by an object and outputs a sensor signal corresponding to the movement of the object, and a signal processing apparatus.

  A program according to the present invention causes a computer to function as a signal processing device.

  As described above, according to the present invention, there is an effect that the target motion can be accurately detected from the low-frequency non-sinusoidal sensor signal and the efficiency of the target motion detection process can be improved. .

It is a block diagram which shows the structure of the detection apparatus in embodiment. It is a wave form diagram which shows the time-axis waveform of the sensor signal in embodiment. It is explanatory drawing showing the concept of the analysis process in embodiment. It is explanatory drawing which shows the sensor signal used for the autoregressive model in embodiment. It is explanatory drawing showing the concept of the autoregressive model in embodiment. It is explanatory drawing of the recognition process by the principal component analysis in embodiment.

  FIG. 1 shows a block configuration of the detection device 1 of the present embodiment. The detection device 1 has a function of detecting a target motion that is a specific motion of the target. In the present embodiment, the detection device 1 detects the respiration of the human body 9 with the target action as the respiration of the human body 9. Hereinafter, the detection apparatus 1 will be described.

  As shown in FIG. 1, the detection device 1 includes a sensor 11 and a signal processing device 12.

  The sensor 11 transmits a radio wave having a predetermined frequency toward the detection range, receives a radio wave reflected by the human body 9 within the detection range, and performs a Doppler frequency corresponding to a frequency difference between the transmitted radio wave and the received radio wave. The sensor signal corresponding to is output. That is, the sensor 11 is a Doppler sensor.

  Alternatively, the sensor 11 may be configured by a sensor using an FMCW (Frequency Modulated Continuous Wave) method. In this case, the sensor 11 repeats the sweep process of increasing and decreasing the frequency of the radio wave to be transmitted. By using the FMCW method for the sensor 11, the detection apparatus 1 can also measure the distance to the object.

  Hereinafter, a case where the sensor 11 is configured by a Doppler sensor will be described.

  The sensor signal output from the sensor 11 is an analog time axis signal corresponding to the movement of the human body 9. FIG. 2 shows a sensor signal (solid line) output from the sensor 11. The sensor 11 converts the difference between the transmission and reception frequencies into a voltage, and the signal strength of the sensor signal increases or decreases in accordance with the respiration of the human body 9 existing 3 m ahead of the sensor 11. When the sensor signal is positive, the human body 9 is exhaling, and when the sensor signal is negative, the human body 9 is inhaling. In the example of FIG. 2, the respiration cycle is about 0.5 to 0.8. It is assumed to be seconds. Moreover, the broken line of FIG. 2 shows background noise.

  The sensor 11 includes a transmission control unit 11a, a transmission unit 11b, a transmission antenna 11c, a reception antenna 11d, and a reception unit 11e.

  The transmission unit 11b transmits the radio wave toward the detection range via the transmission antenna 11c. The transmission control unit 11a controls the frequency, transmission timing, and the like of the radio wave transmitted by the transmission unit 11b. The radio wave transmitted by the transmitter 11b can be, for example, a millimeter wave having a frequency of 24.15 GHz. The radio wave transmitted by the transmission unit 11b is not limited to a millimeter wave, and may be a microwave. Moreover, the value of the frequency of the radio wave transmitted by the transmitter 11b is not particularly limited.

  The receiving unit 11e receives the radio wave reflected by the human body 9 within the detection range via the receiving antenna 11d, and outputs a sensor signal having a frequency corresponding to the frequency difference between the transmitted radio wave and the received radio wave.

  Further, when the FMCW method is used, the sensor 11 generates a beat signal having a beat frequency corresponding to the difference between the frequency of the transmission radio wave (transmission frequency) and the frequency of the reception radio wave (reception frequency). Output as a signal. The FMCW method is originally used to measure the distance to an object, but a sensor signal similar to a Doppler signal can be obtained by analyzing the phase components of the transmitted radio wave and the received radio wave. In the present embodiment, in order to focus on signal processing after the sensor signal is generated, detailed description of the sensor signal generated by the FMCW sensor 11 is omitted.

  The signal processing device 12 has a function of performing signal processing on the sensor signal output from the sensor 11. The signal processing device 12 includes an amplification unit 12a, an A / D conversion unit 12b, a smoothing unit 12c, an analysis unit 12d, a database 12e, a recognition unit 12f, and an output unit 12g. Furthermore, the signal processing device 12 preferably includes a learning unit 12h, a setting unit 12i, and an adjustment unit 12j.

  The amplifying unit 12a is constituted by an amplifier using an operational amplifier, for example, and amplifies the sensor signal.

  The A / D converter 12b converts the sensor signal amplified by the amplifier 12a into a digital sensor signal and outputs it. As an example, the sampling rate of the A / D converter 12b is set to 1000 sps.

  In the present embodiment, the smoothing unit 12c receives the digital sensor signal output from the A / D conversion unit 12b, derives a moving average value of 100 samples, and further calculates a median value (median for every 50 samples) of the moving average value. ). That is, the smoothing unit 12c performs a moving average process and a median process on the digital sensor signal output from the A / D conversion unit 12b, thereby outputting a smoothed digital sensor signal as shown in FIG. .

  The analysis unit 12d stores the sensor signal smoothed by the smoothing unit 12c, applies an autoregressive model to the stored sensor signal, and obtains a regression coefficient vector composed of autoregressive coefficients of the autoregressive model. I do. The analysis unit 12d uses the autoregressive model shown in Equation 1. Equation 1 is an autoregressive model of the sensor signal input to the analysis unit 12d. By adding the error term [e] to the value obtained by multiplying the explanatory variable matrix [X] by the regression coefficient vector [a], An objective variable vector [Y] is obtained. [] Represents a vector.

  The objective variable vector [Y], the explanatory variable matrix [X], the regression coefficient vector [a], and the error vector [e] are expressed by Equation 2 in the case of a cubic autoregressive model. Note that Equation 2 shows a third-order autoregressive model for the sake of explanation, but the order of the autoregressive model used by the analysis unit 12d is not limited to the third order.

  The objective variable vector [Y] is a row vector whose objective variables Y (n) to Y (n-4) are sensor signal values (output of the smoothing unit 12c) that are continuous in time series. The explanatory variable matrix [X] is a 5 × 3 matrix in which the values of sensor signals that are continuous in time series (output of the smoothing unit 12c) are explanatory variables Y (n−1) to Y (n−7). . The component of the regression coefficient vector [a] is the autoregressive coefficient a, and in Equation 2, which is a third-order autoregressive model, the autoregressive coefficients a1, a2, and a3 are used as the components of the regression coefficient vector [a]. The error vector [e] is a row vector composed of five components e (n) to e (n−4). Y (n) to Y (n-7) are values of sensor signals that are continuous in time series, as shown in FIG. Note that n is a positive integer, and (n), (n-1), (n-2), (n-3). . . Indicates a sample identification code.

  Then, the analysis unit 12d determines the regression coefficient vector [a] using the equation shown in Equation 3, and obtains the autoregressive coefficient a (for example, a1, a2, a3). In this case, the regression coefficient vector [a] is determined so that the error vector [e] is minimized (least square solution). For example, when a second-order autoregressive model is used, the autoregressive model is represented as a vector diagram in FIG. [X] [a] is the sum of [X1] [a1] and [X2] [a2], and the difference between [X] [a] and [Y] is the error term [e]. Then, the regression coefficient vector [a] is determined so that the error vector [e] is minimized.

The analysis unit 12d can also determine the regression coefficient vector [a] using the equation shown in Equation 4. In this case, the regression coefficient vector [a] is obtained as a least mean square error solution. In Equation 4, σ ² n is a variance value of background noise, σ ² s is a variance value of a signal (in this embodiment, a respiratory signal of the human body 9), and I is a unit matrix.

  The analysis unit 12d can also determine the regression coefficient vector [a] using QR decomposition.

  The regression coefficient vector [a] determined by the analysis unit 12d reflects the characteristics of the time axis waveform (output of the sensor 11) of the sensor signal. The time-axis waveform of this sensor signal is determined not only by the frequency and amplitude but also by various factors such as the amount of change in frequency per unit time and the amount of change in amplitude per unit time. That is, it can be said that the regression coefficient vector [a] reflects each element constituting the time axis waveform of the sensor signal, and the regression coefficient vector [a] is different for each time axis waveform of the sensor signal.

  And the analysis part 12d has memorize | stored the sensor signal smooth | blunted by the smoothing part 12c, and as shown in FIG. 3, with respect to the sensor signal (output of the smoothing part 12c) in the analysis period T1, it is a regression coefficient vector. The above analysis processing for determining [a] is performed. In the present embodiment, the analysis period T1 is set to 0.8 (seconds).

  Further, as shown in FIG. 3, the analysis unit 12d repeatedly performs the analysis process by shifting the start point of the next analysis period T1 by the shift time T2 (seconds) from the start point of the previous analysis period T1. That is, the next analysis period T1 is set with a delay of the shift time T2 from the previous analysis period T1. In the present embodiment, the shift time T2 is set to 0.2 seconds. That is, the time required for one analysis process is less than the shift time T2, and the analysis unit 12d performs the analysis process for the next analysis period T1 delayed by the shift time T2 after the analysis process being executed is completed. Thus, an analysis process is performed every shift time T2. Therefore, the analysis unit 12d can repeat the analysis process at a cycle shorter than the analysis period T1, so that the number of analysis processes that can be executed per fixed time can be increased, and the analysis process can be highly efficient.

  The recognizing unit 12f determines the presence or absence of the target motion by comparing the regression coefficient vector [a] determined by the analyzing unit 12d with the reference data in the database 12e. Here, the recognition unit 12f determines the presence or absence of breathing of the human body 9 as the target action. The reference data used for the process of determining whether the human body 9 is breathing is data for identifying the regression coefficient vector [a] corresponding to the sensor signal when the human body 9 is breathing.

  The recognizing unit 12f preferably determines whether or not the human body 9 is breathing by performing pattern recognition processing based on principal component analysis, for example.

  The learning unit 12h acquires the regression coefficient vector [a] determined by the analysis unit 12d as learning data when there is a human body 9 breathing within the detection range of the sensor 11. Further, the learning unit 12h also acquires the regression coefficient vector [a] determined by the analysis unit 12d as learning data when there is no human body 9 breathing within the detection range of the sensor 11. And the learning part 12h stores the reference data obtained by performing a principal component analysis with respect to learning data in the database 12e. Here, the reference data stored in the database 12e is data used for pattern recognition, and is category data in which the motion of the object is associated with the projection vector and the discrimination boundary value. The learning unit 12h acquires the learning data a plurality of times in each case where the human body 9 breathing within the detection range of the sensor 11 is present / not present.

  FIG. 6 shows an image of a two-dimensional scatter diagram, a projection axis L1 and a recognition boundary L2 when two autoregressive coefficients a1 and a2 are orthogonal to each other when a secondary autoregressive model is used. For illustration purposes, it is illustrated in two dimensions. The learning unit 12h corresponding to the principal component analysis includes a group G0 corresponding to learning data when the human body 9 does not exist in the detection range of the sensor 11, and a group G1 corresponding to learning data when the human body 9 exists in the detection range. And set. Then, the learning unit 12h corresponding to the principal component analysis calculates the average value of the distribution of data (schematically shown by the broken line M0 and the solid line M1) obtained by projecting the scattered points in the groups G0 and G1 onto the projection axis. The projection axis L1 is determined under the condition that the interval is maximized and the variance is maximized. Furthermore, the learning unit 12h corresponding to the principal component analysis determines a recognition boundary L2 between a region where the human body 9 exists in the detection range and a region where the human body 9 does not exist in the detection range.

  Then, the recognizing unit 12f that performs principal component analysis determines the presence or absence of breathing of the human body 9 by comparing the regression coefficient vector [a] determined by the analyzing unit 12d with reference data (see FIG. 6) of the database 12e. To do.

  Or it is preferable that the recognition part 12f determines the presence or absence of the respiration of the human body 9 by performing the recognition algorithm using multiple regression analysis. Also in this case, the reference data stored in the database 12e is created by the learning unit 12h.

  Alternatively, the recognition unit 12f preferably determines whether the human body 9 is breathing by performing a recognition algorithm using a discriminant analysis method. Also in this case, the reference data stored in the database 12e is created by the learning unit 12h. The discriminant analysis method includes a method using a dimensionless norm, a method using a Fisher correct rate, a method using a Mahalanobis distance, and the like. Of these three discriminant analysis methods, the former is simpler than the latter.

  And the signal processing apparatus 12 is provided with the output part 12g which outputs the detection result by the recognition part 12f. The output unit 12g outputs an output signal indicating that the respiration of the human body 9 has been detected when the recognizing unit 12f detects the respiration of the human body 9. When the recognizing unit 12f does not detect the respiration of the human body 9, the output unit 12g outputs an output signal indicating non-detection.

  As described above, the signal processing device 12 includes the analysis unit 12d, the database 12e, and the recognition unit 12f. The analysis unit 12d applies the autoregressive model over the analysis period T1 to the sensor signal corresponding to the movement of the human body 9 output from the sensor 11 that receives the radio wave reflected by the human body 9 (target object). The analysis unit 12d sequentially performs an analysis process for obtaining the regression coefficient vector [a] including the autoregressive coefficient of the autoregressive model while shifting the analysis period T1 by the shift time T2 from the previous analysis period T1. The analysis unit 12d sets the shift time T2 to be shorter than the time length of the analysis period T1. The database 12e stores reference data for identifying a regression coefficient vector [a] corresponding to a sensor signal when a target motion (for example, breathing) that is a specific motion of the human body 9 is occurring. The recognition unit 12f determines the presence or absence of the target motion of the human body 9 by comparing the regression coefficient vector [a] obtained by the analysis unit 12d with reference data.

  Here, the sensor signal that detects the respiration of the human body 9 has a low-frequency non-sinusoidal waveform. Conventional means for detecting the presence of a low-frequency non-sinusoidal periodic signal include means using frequency analysis such as FFT, and means for comparing the time-axis waveform of the sensor signal itself with the sample signal of the time-axis waveform. is there. However, in the method using frequency analysis, a resolution of 0.1 Hz or more is required for the FFT processing, and a sensor signal of at least 10 seconds or more is required for one analysis processing, which increases hardware resources. Even in a method using collation of the time axis waveform itself, a long sensor signal is required for one analysis process.

  On the other hand, the signal processing device 12 of the present embodiment can set the analysis period T1 of about one cycle of respiration and detect the presence or absence of respiration using the sensor signal of this analysis period T1. That is, the signal processing device 12 can shorten the length of the sensor signal necessary for analyzing the low-speed non-sinusoidal periodic signal. Further, since the signal processing device 12 performs analysis using the autoregressive model and performs recognition processing using the regression coefficient vector [a], the amount of reference data can be reduced. Further, since the signal processing device 12 can repeat the analysis process at a cycle shorter than the analysis period T1, the number of analysis processes that can be executed per fixed time can be increased, and the efficiency of the analysis process can be increased. .

  Therefore, the signal processing device 12 can accurately detect the target motion from the low-frequency non-sinusoidal sensor signal, and can improve the efficiency of the target motion detection process. The target motion is not limited to the movement of the human body 9 such as breathing. For example, the motion of a small animal such as a dog or a cat, the shaking of a structure such as a window or a door may be the target motion. The content is not limited.

  The analysis unit 12d applies an autoregressive model to the moving average value of the sensor signal, and sets the time length of the analysis period T1 to a time length corresponding to the breathing (target motion) cycle of the human body 9. It is preferable to do.

  Therefore, the signal processing device 12 can accurately detect the target action from the low-frequency non-sinusoidal sensor signal.

  The recognition unit 12f preferably uses discriminant analysis, principal component analysis, or multiple regression analysis.

  Therefore, the signal processing device 12 can accurately detect the target action from the low-frequency non-sinusoidal sensor signal.

  The signal processing device 12 preferably includes a learning unit 12h. The learning unit 12h generates reference data based on the sensor signal, and stores the generated reference data in the database 12e.

  Therefore, since the signal processing device 12 generates the reference data from the actual sensor signal, the detection accuracy of the target motion is further improved.

  Furthermore, it is preferable that the signal processing device 12 sets a plurality of target operations as detection targets. In this case, the signal processing device 12 targets movements of human limbs, movement of people, movement of small animals such as dogs and cats, shaking of structures such as windows and doors, etc. in addition to breathing of the human body 9. be able to. The database 12e stores a plurality of reference data. Each of the plurality of reference data is data for identifying the regression coefficient vector [a] corresponding to the sensor signal when each of the plurality of target motions occurs.

  Then, the setting unit 12i selects any one of a plurality of predetermined target motions as the target motion determined by the recognition unit 12f. The setting unit 12i includes a touch panel that can be operated by a user, a switch, and the like. Alternatively, the setting unit 12i may select the target action by remote operation using wired communication or wireless communication.

  Here, the waveform of the sensor signal varies depending on the target operation to be detected. Therefore, the optimal values of the analysis period T1 and the shift time T2 of the analysis unit 12d also differ depending on the target operation to be detected. For example, it is preferable that the time length of the analysis period T1 is set to a value equal to (or substantially equal to) the period of the target motion. The shift time T2 is preferably set to a value that is longer than the time required for one analysis process of the analysis unit 12d and that matches (or substantially matches) the period in which the feature point of the waveform of the target motion appears.

  Therefore, the adjustment unit 12j stores in advance data of an analysis period T1 and a shift time T2 corresponding to each of a plurality of target actions. Then, the adjustment unit 12j sets the analysis period T1 and the shift time T2 of the analysis unit 12d according to the target operation set by the setting unit 12i.

  Then, the analysis unit 12d performs an analysis process for determining the regression coefficient vector [a] using the analysis period T1 and the shift time T2 set by the adjustment unit 12j. The recognizing unit 12f determines the presence or absence of the target motion using the reference data corresponding to the target motion set by the setting unit 12i.

  The setting unit 12i can also sequentially switch a target action selected from a plurality of target actions at regular time intervals. In this case, the analysis unit 12d performs an analysis process for determining the regression coefficient vector [a] while sequentially switching the analysis period T1 and the shift time T2. The recognizing unit 12f sequentially switches the reference data to determine the presence / absence of each of the plurality of target actions.

  That is, the database 12e preferably stores a plurality of reference data for identifying the regression coefficient vector [a] corresponding to the sensor signal when each of the plurality of target motions occurs. In this case, the recognizing unit 12f determines the presence or absence of the target motion by collating the regression coefficient vector [a] obtained by the analyzing unit 12d with one or more reference data.

  Therefore, the signal processing device 12 can detect the presence / absence of each of a plurality of target actions, and versatility is improved.

  Moreover, it is preferable to provide the adjustment part 12j which changes the time length of the analysis period T1 and the shift time T2 according to the object operation | movement set as the determination object of the recognition part 12f.

  Therefore, the analysis period T1 and the shift time T2 suitable for the target operation to be detected are set.

  Furthermore, the above-described detection device 1 includes a sensor 11 that receives a radio wave reflected by the human body 9 (object) and outputs a sensor signal corresponding to the movement of the human body 9, and a signal processing device 12. And

  Therefore, the detection apparatus 1 can accurately detect the target motion from the low-frequency non-sinusoidal sensor signal and can improve the efficiency of the target motion detection process.

  The sensor 11 is preferably a Doppler sensor or a sensor using an FMCW (Frequency Modulated Continuous Wave) method.

  Therefore, the detection apparatus 1 can detect minute movements such as respiration.

  Further, the signal processing device 12 is equipped with a computer constituted by a microcomputer or the like, and each function of the signal processing device 12 is realized by the computer executing a program. The computer constituting the signal processing device 12 includes a processor and an interface that operate according to a program as main hardware configurations. This type of processor includes a DSP (Digital Signal Processor), a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), etc., and implements the following functions of the signal processing device 12 by executing a program. If it is possible, the kind is not ask | required.

  As a program providing form, a computer-readable ROM (Read Only Memory), a form stored in advance in a recording medium such as an optical disc, or the like is supplied to the recording medium via a wide area communication network including the Internet. There are forms.

  In other words, the program causes the computer to function as the signal processing device 12.

  Therefore, the computer that executes this program can accurately detect the target motion from the low-frequency non-sinusoidal sensor signal, and can improve the efficiency of the target motion detection process.

  The above-described embodiment is an example of the present invention. For this reason, the present invention is not limited to the above-described embodiment, and various modifications can be made depending on the design and the like as long as the technical idea according to the present invention is not deviated from this embodiment. Of course, it is possible to change.

DESCRIPTION OF SYMBOLS 1 Detection apparatus 11 Sensor 12 Signal processing apparatus 12a Amplification part 12b A / D conversion part 12c Smoothing part 12d Analysis part 12e Database 12f Recognition part 12g Output part 12h Learning part 12i Setting part 12j Adjustment part 9 Human body (object)

Claims (8)

Applying an autoregressive model over the analysis period to the sensor signal corresponding to the movement of the object output from the sensor that receives the radio wave reflected by the object, and a regression coefficient comprising the autoregressive coefficient of the autoregressive model Analyzing processing for obtaining a vector, sequentially performing the analysis period shifted by a shift time from the previous analysis period, and setting the shift time shorter than the time length of the analysis period;
A database storing reference data for identifying the regression coefficient vector corresponding to the sensor signal when a target motion that is a specific motion of the target is occurring;
A recognizing unit that determines the presence or absence of the target motion by comparing the regression coefficient vector obtained by the analyzing unit with the reference data ;
A signal processing apparatus comprising: an adjustment unit configured to change a time length of the analysis period and the shift time according to the target operation set as a determination target of the recognition unit .   The analysis unit applies an autoregressive model to the moving average value of the sensor signal, and sets the time length of the analysis period to a time length corresponding to the period of the target motion. The signal processing apparatus according to claim 1.   The signal processing apparatus according to claim 1, wherein the recognition unit uses discriminant analysis, principal component analysis, or multiple regression analysis. The database stores a plurality of the reference data for identifying the regression coefficient vector corresponding to the sensor signal when each of the plurality of target actions is occurring,
The said recognition part determines the presence or absence of the said target operation | movement by collating the said regression coefficient vector which the said analysis part calculated | required with one or more said reference data. Signal processing equipment. The signal processing apparatus according to claim 1, further comprising: a learning unit that generates the reference data based on the sensor signal and stores the generated reference data in the database. A detection apparatus comprising: a sensor that receives radio waves reflected by an object and outputs a sensor signal corresponding to the movement of the object; and the signal processing device according to claim 1. The detection device according to claim 6, wherein the sensor is a Doppler sensor or a sensor using an FMCW (Frequency Modulated Continuous Wave) method. A program for causing a computer to function as the signal processing apparatus according to claim 1.

Images