JP5383890B1 - Object number detection device and program - Google Patents

Abstract

An object number detection apparatus and a program capable of detecting the number of detection objects even when there are a plurality of detection objects are provided.
A frequency conversion unit that converts a Doppler signal for an arbitrary reflecting object into a plurality of signals having different frequencies, and the plurality of signals having different frequencies converted by the frequency conversion unit are converted into amplitude times in respective frequency bands. An independent signal decomposing unit that decomposes into independent signals having predetermined independence using a change, an object number detecting unit that detects the number of the reflecting objects based on the independent signals decomposed by the independent signal decomposing unit, Object number detection device with [Selection] Fig. 2

Description

  The present invention relates to an object number detection device and a program.

  In recent years, devices that detect the motion of humans, cars, and other moving bodies using Doppler sensors have appeared. Since the Doppler sensor can detect minute human movements such as heartbeats and breathing, for example, a technique for health management has been developed by detecting human movements in daily life using a Doppler sensor.

  As a technique related to health management using such a Doppler sensor, in Patent Document 1, an amplitude component and a heartbeat component due to body movement are obtained from an output signal obtained from a human body surface by a Doppler sensor using independent component analysis. A technique for separating and extracting a heartbeat component is disclosed.

JP 2006-0555504 A

  However, since the technique disclosed in Patent Document 1 is based on the premise that the detection target is a single object, it is difficult to apply to a case where there are a plurality of detection targets. In addition, since the Doppler sensor generally has only a single output, there is a problem that it is difficult to detect the number of detection targets when there are a plurality of detection targets, for example.

  Therefore, the present invention has been made in view of the above problems, and an object of the present invention is a new and improved technique capable of detecting the number of detection targets even when there are a plurality of detection targets. Another object is to provide an object number detection apparatus and program.

  In order to solve the above-described problem, according to an aspect of the present invention, a frequency converter that converts a Doppler signal for an arbitrary reflecting object into a plurality of signals having different frequencies, and the plurality of the plurality of signals converted by the frequency converter Based on the independent signal decomposed by the independent signal decomposing unit, the independent signal decomposing unit that decomposes the signals of different frequencies into independent signals having predetermined independence using temporal changes in the amplitude of each frequency band An object number detection device is provided that includes an object number detection unit that detects the number of reflection objects.

  The independent signal decomposition unit may decompose the plurality of signals having different frequencies into the independent signals by independent component analysis.

  The frequency converting unit may convert the Doppler signal into the signals having different frequencies by Fourier transform.

  The independent signal decomposition unit may decompose the signal into independent signals by maximizing non-Gaussianity of the signal.

  The independent signal decomposing unit may reduce the number of dimensions of the plurality of signals having different frequencies by singular value decomposition and decompose the signal into independent signals.

  The object number detection unit may generate an integrated signal by classifying and integrating the independent signals based on the similarity of the independent signals, and detect the number of the reflective objects based on the integrated signal. .

  The object number detection unit may classify the independent signals based on similarity of the independent signals calculated by either Pearson product-moment correlation coefficient or Jensen-Shannon information amount.

  The object number detection unit may calculate a feature amount of the integrated signal, and detect the number of the reflective objects by machine learning based on the calculated feature amount.

  The object number detection unit may perform ordering of the integrated signals for calculating the feature amount based on a statistical amount of the integrated signals.

  The reflective object may be a human, animal, machine, or other periodic motion body that performs periodic motion.

  The reflective object may be a building or other periodic motion body that performs periodic vibration.

  In order to solve the above problems, according to another aspect of the present invention, a computer converts a Doppler signal for an arbitrary reflecting object into a plurality of different frequency signals, and the converted plurality of different ones. Decomposing a frequency signal into independent signals having predetermined independence using temporal changes in amplitude of each frequency band; detecting the number of reflecting objects based on the decomposed independent signals; A program for executing the above is provided.

  As described above, according to the present invention, it is possible to detect the number of detection targets even if there are a plurality of detection targets.

It is explanatory drawing for demonstrating the outline | summary of the object number detection apparatus which concerns on this embodiment. It is explanatory drawing for demonstrating the application example of the object number detection apparatus which concerns on this embodiment. It is a block diagram which shows the structure of the object number detection apparatus which concerns on this embodiment. It is a flowchart which shows operation | movement of the object number detection apparatus which concerns on this embodiment. It is explanatory drawing which shows an example of the ICA basis function obtained by the object number detection apparatus which concerns on this embodiment. It is explanatory drawing which showed the experimental environment of the object number detection apparatus which concerns on this embodiment. It is explanatory drawing which shows the integrated ICA basis function in the experiment by the object number detection apparatus which concerns on this embodiment. It is explanatory drawing which shows the integrated ICA basis function in the experiment by the object number detection apparatus which concerns on this embodiment. It is explanatory drawing which shows the integrated ICA basis function in the experiment by the object number detection apparatus which concerns on this embodiment. The experimental result by the object number detection apparatus concerning this embodiment is shown.

  Exemplary embodiments of the present invention will be described below in detail with reference to the accompanying drawings. In addition, in this specification and drawing, about the component which has the substantially same function structure, duplication description is abbreviate | omitted by attaching | subjecting the same code | symbol.

<1. Overview of object number detection apparatus according to an embodiment of the present invention>
An object number detection device (object number detection device 1) according to an embodiment of the present invention is:
A. A frequency conversion unit (frequency conversion unit 11) that converts a Doppler signal for an arbitrary reflecting object (person 4, fan 5) into a plurality of signals having different frequencies;
B. Independent signal decomposing unit (independent signal decomposing unit 13) that decomposes the signals of the plurality of different frequencies converted by the frequency converting unit into independent signals having predetermined independence using temporal changes in amplitudes of the respective frequency bands. When,
C. An object number detection unit (object number detection unit 15) for detecting the number of the reflective objects based on the independent signal decomposed by the independent signal decomposition unit;
Is provided.

  First, the outline | summary of the object number detection apparatus which concerns on this embodiment is demonstrated with reference to FIGS.

  FIG. 1 is an explanatory diagram for explaining an overview of an object number detection device 1 according to the present embodiment. As shown in FIG. 1, the object number detection device 1 detects the number of objects existing in the detection range of the Doppler sensor 2 based on the output of the Doppler sensor 2.

  More specifically, the Doppler sensor 2 irradiates a transmission signal as a radio wave or a sound wave, and receives a reflected wave reflected by a moving body (reflection object) as a reception signal. Then, the object number detection device 1 detects the number of reflecting objects based on the Doppler signal obtained from the difference between the transmission signal and the reception signal output from the Doppler sensor 2. Specifically, since the radio wave or sound wave emitted by the Doppler sensor 2 is reflected by the persons 4A and 4B who perform exercise such as breathing and heartbeat, the object number detection device 1 detects that the number of objects is two. To do. Hereinafter, people 4A and 4B are collectively referred to as people 4 when it is not necessary to distinguish them.

  As shown in FIG. 1, the person 4A is standing and the person 4B is breathing while sitting on a chair. Here, the Doppler sensor 2 receives the reflected waves from the people 4A and 4B simultaneously. Therefore, the Doppler signal is a signal having a different periodicity between a signal resulting from a periodic motion such as the heartbeat or breathing of the person 4A and a signal resulting from a periodic motion such as the heartbeat or breathing of the person 4B. It is a mixed signal including at least two. The number-of-objects detection apparatus 1 detects the number of objects from such mixed signals by using independent component analysis (ICA) by utilizing that the signal sources have different periodicities.

  However, the periodicity of human breathing and heartbeat is generally not stable. Therefore, in the present specification, as shown in FIG. 2, a fan that performs a swing motion with a stable cycle is used as a reflecting object.

  FIG. 2 is an explanatory diagram for explaining an application example of the object number detection device 1 according to the present embodiment. As shown in FIG. 2, since the radio wave or sound wave irradiated by the Doppler sensor 2 is reflected by the fans 5A and 5B that perform the swing motion, the object number detection device 1 detects that the number of objects is two. To do. Note that the fans 5A and 5B perform the head swinging operation at different periods. Hereinafter, when it is not necessary to distinguish between the fans 5A and 5B, they are collectively referred to as the fan 5.

  Here, the Doppler signal obtained by the swing motion of the electric fan 5 has a widely distributed frequency and a component close to that of the person 4. Further, as shown in FIG. 2, the fact that the plurality of electric fans 5 swing each other with different periods corresponds to the person 4 breathing with different periods. As will be described in detail below, the object number detection device 1 can acquire a Doppler signal having the same component as that of the person 4, performs a periodic motion in the same manner as the person 4, and individually It is possible to identify the number of electric fans 5 that exercise at different periods. For this reason, it can be said that the present invention in which the number of fans 5 can be identified has a certain effectiveness in identifying the number of people 4.

  In addition, the reflective object which the object number detection apparatus 1 can detect the number of objects is not restricted to the person 4 or the electric fan 5. For example, the object number detection device 1 detects the number of periodic motion bodies that perform periodic motions such as heartbeats and breathing of humans 4 and animals, or mechanical and periodic motions such as a fan, an engine, and the like. Also good. In addition, the object number detection apparatus 1 may detect the number of periodic moving bodies that vibrate in a specific period or the like due to external or internal factors such as a house, a structure, or the like. In addition, the number of reflective objects that are the targets for detecting the number of objects by the object number detection device 1 may be one or two or more.

  The outline of the object number detection device 1 according to the present embodiment has been described above. Next, the configuration of the object number detection apparatus 1 according to the present embodiment will be described in detail with reference to FIG.

<2. Configuration>
FIG. 3 is a block diagram illustrating a configuration of the object number detection apparatus 1 according to the present embodiment. As shown in FIG. 3, the object number detection device 1 includes a frequency conversion unit 11, an independent signal decomposition unit 13, and an object number detection unit 15. The object number detection device 1 outputs to the output device 3 based on the Doppler signal input by the Doppler sensor 2.

  The frequency conversion unit 11 converts the Doppler signal for the person 4, the fan 5, or any other reflective object into a plurality of signals having different frequencies. Specifically, the frequency conversion unit 11 performs Fourier transform and positive / negative frequency separation on the output signal output from the Doppler sensor 2. At this time, the frequency converter 11 may extract a specific frequency band that may be related to the operation of the reflecting object. The frequency conversion unit 11 outputs a signal obtained by Fourier transform, separation of positive and negative frequencies, and extraction of a specific frequency band to the independent signal decomposition unit 13.

  The independent signal decomposing unit 13 decomposes the signal converted by the frequency converting unit 11 into an independent signal having a predetermined independence by using a temporal change in the amplitude of each frequency band. Specifically, the independent signal decomposition unit 13 performs independent component analysis on the signal output from the frequency conversion unit 11. Note that the independent signal decomposition unit 13 may perform dimension reduction by singular value decomposition, which is general as preprocessing for independent component analysis. The independent signal decomposition unit 13 outputs the ICA basis function obtained as a result of the independent component analysis to the object number detection unit 15.

  The object number detection unit 15 detects the number of reflection objects based on the signal output from the independent signal decomposition unit 13. Specifically, the object number detection unit 15 classifies a time series group constituted by ICA basis functions into a partial space by clustering processing. Then, the object number detection unit 15 calculates a feature amount for each of the integrated signals, and estimates the number of objects by machine learning. The object number detection unit 15 may use SVM (Support Vector Machine) as machine learning. The object number detection unit 15 outputs the estimated number of objects to the output device 3.

  Note that the frequency conversion unit 11, the independent signal decomposition unit 13, and the object number detection unit 15 each function as an arithmetic processing unit and perform the above-described processing according to various programs. Therefore, the frequency conversion unit 11, the independent signal decomposition unit 13, the object number detection unit 15, or the object number detection device 1 is realized by, for example, a CPU (Central Processing Unit) or a microprocessor. The object number detection device 1 may include a ROM (Read Only Memory) that stores programs to be used, calculation parameters, and the like, and a RAM (Random Access Memory) that temporarily stores parameters that change as appropriate.

  The output device 3 outputs the estimation result by the object number detection unit 15 by video, image, sound, or the like. The output device 3 is realized by, for example, a CRT (Cathode Ray Tube) display device, a liquid crystal display (Liquid Crystal Display) device, a speaker, or the like.

  The Doppler sensor 2 transmits / receives a radio wave or an ultrasonic wave to / from an arbitrary reflecting object, and sends a Doppler signal, which is a signal having a frequency difference between the transmitted radio wave or ultrasonic wave and the received radio wave or ultrasonic wave, to the frequency converter 11. Output. The Doppler sensor 2 is a so-called dual channel Doppler radar having at least two transmission / reception units (not shown) for transmitting and receiving signals. In the following, with regard to the Doppler signal model, a case where one signal transmission / reception unit is provided will be described first, and then a case where a plurality of signal transmission / reception units are provided will be described.

The transmission signal V _s (t) and the reception signal V _r (t) by the Doppler sensor having one signal transmission / reception unit are expressed as follows.

Here, A _s and A _r is the amplitude, phi ₀ is signal initial phase of, f is exhibits a round trip time between the transmission frequency, T is the Doppler sensor reflecting object.

Next, the Doppler sensor 2 having a plurality of signal transmission / reception units will be described. The Doppler sensor 2 generates an in-phase signal I (t) and a quadrature signal Q (t) having a phase different by π / 2 from the in-phase signal I (t) by mixing the transmission signal V _s and the reception signal V _r. To do. The in-phase signal I (t) and the quadrature signal Q (t) are expressed as follows.

Here, the second cos (φ _r + φ _s ) and sin (φ _r + φ _s ) in the above formulas 3 and 4 can have a higher frequency compared to the first term of each formula, and therefore the low-pass filter May be removed. Then, the above formulas 3 and 4 are expressed as follows by replacing the phase difference of the first term with the distance change amount ΔR.

  Here, O represents an offset DC component, g represents noise, and λ represents a wavelength. The Doppler signal model according to Equations 5 and 6 indicates that the phase change factor is related to the change in distance, and that the in-phase signal I (t) and the quadrature signal Q (t) are not independent of each other.

<3. Operation>
The configuration of the object number detection device 1 has been described above. Below, the operation | movement which detects the number of the reflective objects by the object number detection apparatus 1 is demonstrated with reference to FIGS.

  FIG. 4 is a flowchart showing the operation of the object number detection apparatus 1 according to this embodiment. As shown in FIG. 4, first, in step S <b> 104, the frequency converter 11 performs a Fourier transform on the in-phase signal I (t) and the quadrature signal Q (t) output from the Doppler sensor 2.

(S104: Fourier transform)
As described above, the in-phase signal I (t) and the quadrature signal Q (t) that are output signals from the Doppler sensor 2 are not independent of each other. For this reason, since the output signal cannot be regarded as a mixed signal as it is, it is difficult for the object number detection device 1 to apply independent component analysis to the output signal. Therefore, the object number detection device 1 converts the output signal into a sine wave signal having a different frequency by applying a complex Fourier transform to the output signal by the frequency conversion unit 11. In the present embodiment, the frequency conversion unit 11 performs short-time Fourier transform as an example of complex Fourier transform.

  First, the complex signal f (t) = I (t) + Q (t). The frequency converter 11 performs a complex Fourier transform on the complex signal f (t) to obtain a complex Fourier coefficient F (ω) represented by the following equation.

  Here, ω represents a frequency. The frequency conversion unit 11 performs short-time Fourier transform on the complex Fourier coefficient F (ω) using the window function w (t).

  Here, F (t, ω) is a multidimensional complex signal corresponding to each frequency ω. Here, the absolute value of the multidimensional complex signal F (t, ω) is expressed as the multidimensional amplitude spectrum signal A (t, ω) by the following equation.

  As described above, in step S <b> 104, the frequency conversion unit 11 performs a short-time Fourier transform on the output signal from the Doppler sensor 2. Subsequently, in step S108, the frequency converter 11 separates positive and negative frequencies.

(S108: Separation of positive and negative frequencies)
The frequency conversion unit 11 separates positive and negative frequencies for the multidimensional amplitude spectrum signal A (t, ω). The frequency ω of the multidimensional amplitude spectrum signal A (t, ω) has a positive frequency and a negative frequency, and each has a different meaning. That is, the positive frequency is related to the speed at which the reflecting object approaches the Doppler sensor 2, and the negative frequency is related to the speed at which the reflecting object moves away from the Doppler sensor 2. For this reason, the frequency converter 11 separates them so that the amplitude spectrum signal A _{+ ω} (t, ω) corresponding to the positive frequency represented by the following expression and the amplitude spectrum signal A _−ω corresponding to the negative frequency are expressed. (T, ω) is obtained.

Next, the frequency converter 11 may extract a specific frequency band that may be related to the operation of the reflecting object. For example, the frequency conversion unit 11 corresponds to a frequency in the range of 0.2 to 10.0 Hz for the amplitude spectrum signal A _{+ ω} (t, ω) as a frequency band that can be related to the swinging operation of the electric fan 5. The amplitude spectrum signal A ′ _{+ ω} (t, ω) represented by the equation may be extracted.

On the other hand, for the amplitude spectrum signal A _−ω (t, ω), the frequency converter 11 corresponds to a frequency in the range of −10.0 to −0.2 Hz, and the amplitude spectrum signal A ′ _−ω represented by the following equation. (T, ω) may be extracted.

  As described above, in step S108, the frequency converter 11 separates positive and negative frequencies. Subsequently, in step S <b> 112, the independent signal decomposition unit 13 performs independent component analysis on the signal output by the frequency conversion unit 11.

(S112: Independent component analysis)
As described above, the multidimensional amplitude spectrum signal A (t, ω) is divided into two groups of positive and negative frequencies. Here, the independent signal decomposition unit 13 regards the time-varying signal for each frequency band as a mixed signal and applies independent component analysis thereto. Note that the time-varying signal for each frequency band is regarded as a mixed signal because the time-varying signal is considered to be composed of mixed components of various movements of the signal source in the frequency domain.

The independent signal decomposition unit 13 performs independent component analysis on each of the amplitude spectrum signals A ′ _{+ ω} (t, ω) and A ′ _−ω (t, ω) output by the frequency conversion unit 11. At this time, the independent signal decomposing unit 13 may perform dimension reduction by singular value decomposition SVD (•), which is general as preprocessing for independent component analysis.

The singular value is represented by the standard deviation of the main component used as a measure of importance. The independent signal decomposition unit 13 calculates the cumulative contribution rate of singular values and performs threshold processing, so that the neck of the electric fan 5 is _derived from the amplitude spectrum signals A ′ _{+ ω} (t, ω) and A′− _ω (t, ω). Unnecessary information about the pretending action may be deleted, leaving a useful component simply.

In the present embodiment, as an example, the independent signal decomposition unit 13 employs a main component having a cumulative contribution rate of up to 80%. Here, the amplitude spectrum signal _{A'+ omega} by independent signal decomposition unit 13 (t, ω), A' -ω (t, ω) n obtained by singular value decomposition for each of the m-dimensional principal component signal, S _p (t) and S _q (t) are represented by the following equations, respectively.

  Here, p = 1, 2,..., N, q = 1, 2,. Since the number of dimensions changes depending on the cumulative contribution rate, n and m differ depending on the data.

Subsequently, the independent signal decomposition unit 13 applies independent component analysis to S _p (t) and S _q (t). Independent component analysis is a widely used technique for mixed signals, and is characterized in that the mixed signal can be decomposed so that each becomes a statistically independent component. A basic independent component analysis model is given below.

  Here, s is a vector composed of signal sources, x is an observation vector including a mixed signal, and A is a mixing matrix. Further, the signal source s is unknown, and the information about the mixing matrix A is also unknown. Therefore, in order to estimate the independent component from only the observation signal x, the signal source s is estimated by the following separation matrix W based on the assumption that the signal source vector s is statistically independent.

Here, the separation matrix W is an approximation of A- ¹ . The separation matrix W is estimated by maximizing the non-Gaussian nature of the signal source using a measure such as kurtosis or negentropy.

  In this embodiment, as an example of independent component analysis, a technique called a fast ICA method is used (A. Hyvarinen and E. Oja, “A Fast Fixed-Point Algorithm for Independent Component Analysis,” National. No. 7, pp. 1483-1492, October, 1997). This method is known for its very fast convergence called a fixed point algorithm, and Negentropy or the like is used as a measure of non-Gaussianity. The signal estimated by the fastICA method is called an ICA basis function. An important property of this ICA basis function is that it retains sign and rank indefiniteness.

When the independent component analysis processing is set to ICA (·), the independent signal decomposition unit 13 maximizes or substantially maximizes the independence (non-Gaussian property) expressed by the following equation as a result of the independent component analysis. X _p (t) and X _q (t) are obtained.

  Note that the number of ICA basis functions extracted by the independent signal decomposition unit 13 to be independent from each other is equal to the number of input signals. Here, an example of the ICA basis function obtained by the independent signal decomposition unit 13 for the Doppler signal obtained from the single electric fan 5 will be described with reference to FIG.

  FIG. 5 is an explanatory diagram illustrating an example of the ICA basis function obtained by the independent signal decomposition unit 13 according to the present embodiment. As shown in FIG. 5, it can be said that each ICA basis function has a partly common feature.

  Therefore, the number-of-object detection unit 15 is a method of independent subspace analysis (A. Hyvarenen and P. Hoyer, “Independent subspace analysis emergence of phase and flow infure features. Network, vol. 2, pp. 1059-1064, July, 1999.), based on the idea that ICA basis functions form independent subspaces, clustering of ICA basis functions is performed in step S116 described later. I do. However, like the method of independent subspace analysis, it is considered that the independence of the subspaces does not have to be guaranteed, but there is a restriction that the subspaces must be independent.

  An application of applying the method of independent subspace analysis to a single mixed signal using a spectrogram has been proposed by Casey and Westner (M. A. Casey and A. Westner, “Separation of mixed audio sources”. by independent subspace analysis, "Proc. Int. Comp. Music Conf., Berlin, Germany, pp. 154-161, May, 1997.). Casey and Westner proposed a framework for applying independent subspace analysis techniques to audio signal data, and made it possible to reconstruct valuable information from specific subspaces. On the other hand, in this embodiment, the object number detection unit 15 classifies a time series group constituted by ICA basis functions into a partial space by clustering processing.

  In order to estimate the subspace, accuracy is required for the measurement of the similarity between ICA basis functions and the application of the clustering method. In the present embodiment, the object number detection apparatus 1 accurately measures the similarity between ICA basis functions, in particular, in step S116 described later.

As described above, in step S112, the independent signal decomposing unit 13 performs independent component analysis on the signal output from the frequency converting unit 11. Subsequently, in step S116, the object number detection unit 15 performs clustering on the multidimensional signals X _p (t) and X _q (t).

(S116: Clustering)
The object number detection unit 15 performs clustering on the assumption that a dependency relationship exists in each of the multidimensional signals X _p (t) and X _q (t). In the present embodiment, the object number detection unit 15 calculates the similarity between signals and performs clustering by the group average method as an example of clustering processing. The number-of-objects detection unit 15 uses either Pearson's product-moment correlation coefficient Cor (x, y) or Jensen Shannon's information amount D _JS (q || p) as an index for calculating the similarity. May be used. The Pearson product moment correlation coefficient Cor (x, y) is expressed by the following equation.

The Jensen-Shannon information amount D _JS (q || p) is expressed by the following equation.

Here, D _KL (p || q) represents a Cullback library information amount expressed by the following equation.

  This index represents the stochastic similarity between the stochastic distributions of each ICA basis function. The distribution is estimated using a Gaussian kernel which is a kernel function. However, since this index remains undefined, the object number detection unit 15 takes an absolute value for the ICA basis function before calculating the distribution. This index is obtained by symmetrizing the amount of information of the Cullback library, and is also used in the proposal by Casey and Westner described above.

  The object number detection unit 15 performs clustering based on the similarity calculated by the above-described index. In the present embodiment, the object number detection unit 15 performs clustering by a group average method, which is one of hierarchical clustering methods, as an example of clustering. The group average method can evaluate the quality of the cluster based on all the similarities, and the single connection method and the complete connection method that identify the similarity of the clusters with the similarity between single components. Can avoid problems. Note that the number k of clusters in the group average method is set to be equal to or greater than the number of reflecting objects.

  Subsequently, the object number detection unit 15 integrates signals belonging to each cluster formed by clustering into a single signal (integrated signal) for each cluster. In the present embodiment, as an example, the object number detection unit 15 integrates a single signal by taking an average between signals belonging to each cluster. Further, the integrated signal can be regarded as a signal having the characteristics of each partial space.

  Here, there is a problem about indefiniteness of order called a permutation problem. As one method for solving this problem, in this embodiment, focusing on the fact that the object number detection unit 15 can be divided into a large fluctuation and a small fluctuation of the integrated signal, the order of each statistical quantity is changed. decide.

If a series of processing from clustering to integration is referred to as Clustering _c (•), the object number detection unit 15 integrates k integrated signals Y _c ⁺ (t) and Y _c ⁻ (t) expressed by the following equations. Get.

Here, c = 1, 2,..., K. Further, it is assumed that the integrated signals Y _c ⁺ (t) and Y _c ⁻ (t) are arranged in descending order of variance as an example of the statistics.

As described above, in step S116, the object number detection unit 15 performs clustering on the multidimensional signals X _p (t) and X _q (t). Subsequently, in step S120, the object number detection unit 15 calculates the feature amounts of the integrated signals Y _c ⁺ (t) and Y _c ⁻ (t) integrated as a result of clustering.

(S120: feature amount calculation)
First, the feature amount is calculated for each of the object number detection unit 15 and the integrated signals Y _c ⁺ (t) and Y _c ⁻ (t). In the present embodiment, as an example of the feature amount, the object number detection unit 15 includes three types of feature amounts in the frequency domain focusing on signal periodicity and two types of non-Gaussian feature amounts focusing on signal distribution. Is calculated. Here, the feature quantity in the frequency domain is intended that the power spectrum of the integrated signal will have various frequency bands and powers, and the non-Gaussian feature quantity has a statistically different distribution of the integrated signal. Intended to be.

-Feature quantity in frequency domain The number-of-objects detection unit 15 calculates a frequency domain entropy H expressed by the following equation as a feature quantity.

Here, F _i represents a power spectrum obtained by Fourier transform. The frequency domain entropy H is a measure for measuring the complexity of the entire frequency domain. In addition to the frequency domain entropy H, the number-of-objects detection unit 15 calculates a total of three types, that is, the peak value of the amplitude of the power spectrum and the value of the frequency at the peak as the feature amount.

Non-Gaussian Feature Quantity The object number detection unit 15 calculates the skewness skewness and kurtosis kurtosis expressed below as feature quantities.

  Here, the skewness skewness indicates the asymmetry of the distribution, and the kurtosis kurtosis indicates the degree of deviation from the Gaussian distribution of the input distribution.

  From the above, the dimension of the feature amount is 2 × 5 × k as a product of 2 indicating the type of frequency (positive / negative), 5 indicating the type of feature amount, and the number of clusters k.

As described above, in step S120, the object number detection unit 15 calculates the feature amounts of the integrated signals Y _c ⁺ (t) and Y _c ⁻ (t). Subsequently, in step S124, the object number detection unit 15 estimates the number of objects by machine learning.

(S124: Estimating the number of objects by machine learning)
The object number detection unit 15 estimates the number of objects by machine learning based on the feature amount calculated in step S120. In the present embodiment, as an example of machine learning, the object number detection unit 15 estimates the number of objects by SVM.

<4. Experiment>
The inventor conducted an experiment to demonstrate the effectiveness of the present invention. Hereinafter, the contents and results of the experiment by the present inventor will be described with reference to FIGS.

[Experiment environment]
FIG. 6 is an explanatory diagram showing an experimental environment of the object number detection apparatus 1 according to the present embodiment. As shown in FIG. 6, the present inventor conducted experiments in two experimental environments of pattern α and pattern β in which the distance between the Doppler sensor 2 and the fans 5A and 5B is different. Specifically, the pattern α has a distance of 1.0 meter between the Doppler sensor 2 and the electric fan 5, more specifically, the electric fan 5B, and the pattern β has a distance of 2.0 meters between the Doppler sensor 2 and the electric fan 5. The Doppler sensor 2 is a 24-GHz continuous wave microwave Doppler radar. The electric fan 5A has an oscillation period of about 10 seconds, and the electric fan 5B has an oscillation period of about 18 seconds. Further, the heights of the fans 5A and 5B and the Doppler sensor 2 from the ground were equally about 0.8 meters.

[Data acquisition and analysis]
The present inventor uses three patterns of label α and pattern β as three types of label data when one of the fans 5A or 5B is operating and when both the fans 5A and 5B are operating. Obtained in the experimental environment. That is, the present inventor has obtained a total of six types of label data. The inventor sampled each data for 9 minutes at a sampling frequency of 500 Hz. In addition, the inventor used a Hamming window as the window function w (t) of Expression 8, the window width was set to 2048 samples, and the window was shifted every 20 samples. In addition, the present inventor set the threshold of the cumulative contribution rate in singular value decomposition to 0.8 and sets the number of clusters k in the group average method to 2.

Since the sampling frequency is 500 Hz, the resolution of the frequency ω is about 0.24 Hz. Further, the frequencies _{ω of the} amplitude spectrum signals A ′ _{+ ω} (t, ω) and A ′ _−ω (t, ω) have about 40 dimensions, respectively. Further, the dimension of the feature amount calculated by the object number detection unit 15 is 2 × 5 × k as described above, and the cluster number k is set to 2, so that it is 20 dimensions.

  And this inventor made SVM identify the data acquired by the above setting. The present inventor used the soft margin SVM because the acquired data are well mixed with each other. The inventor has set the parameter C of the soft margin SVM to 1 and maximizes the margin using a Gaussian kernel. The present inventor set the size of the determination window for identification to 40 seconds and performed identification by SVM with a 1 second interval shift.

[Experimental result: ICA basis function]
FIG. 7 is an explanatory diagram showing an integrated ICA basis function in the experiment by the object number detection apparatus 1 according to the present embodiment. The ICA basis function shown in FIG. 7 is a result of acquiring data for about 40 seconds in a state where the two fans 5A and 5B are operating. 7A and 7B show the result of clustering into a subspace for positive frequencies, and FIGS. 7C and 7D show the subspace for negative frequencies. The result of clustering is shown. Further, (a) and (c) show the results of integration into the subspace 1 from the determination of the order, respectively, and similarly (b) and (d) show the results of integration into the subspace 2, respectively.

  Next, FIG. 8 shows an ICA basis function in a state where only the electric fan 5A is operated, and FIG. 9 shows an ICA basis function in a state where only the electric fan 5B is operated. The meanings of (a), (b), (c), and (d) in FIGS. 8 and 9 are the same as the meanings of (a), (b), (c), and (d) described above with reference to FIG.

  8 and 9 are explanatory diagrams showing integrated ICA basis functions in the experiment by the object number detection apparatus 1 according to the present embodiment. As shown in FIG. 8 and FIG. 9, when comparing the partial space 1 and the partial space 2, the waveform vibrates smoothly in the partial space 1, whereas the waveform vibrates violently in the partial space 2. . Further, when comparing the positive frequency and the negative frequency, the waveforms are similar but the temporal peak positions (positions on the horizontal axis) are different. The inventor has inferred that the difference in peak position is caused by the swinging motion of the electric fan 5.

[Experimental result: Recognition by SVM]
The object number detection device 1 detects the number of the fans 5 during the swinging operation by SVM. That is, the object number detection device 1 detects that there is one electric fan 5 when, for example, only the electric fan 5A performs the swinging motion. Moreover, this inventor does not consider the case where the electric fan 5 does not exist. The reason is that the object number detection device 1 can easily determine from the amplitude of the output signal of the Doppler sensor 2 that there is no reflecting object such as the electric fan 5. Hereinafter, with reference to FIG. 10, an experiment result of the object number detection by the SVM of the object number detection apparatus 1 will be described.

  FIG. 10 shows an experimental result by the object number detection apparatus 1 according to the present embodiment. 10 (1) and 10 (2), A (α), B (α), and Mix (α) are respectively a state in which only the fan 5A is present and a state in which only the fan 5B is present in the experimental environment pattern α. The state in which the electric fans 5A and 5B exist together is shown. Similarly, A (β), B (β), and Mix (β) in FIGS. 10 (1) and 10 (2) are in a state where only the fan 5A exists in the experimental environment pattern β, and only the fan 5B exists. State in which the electric fans 5A and 5B exist together.

The present inventor adopted A (β), B (β), and Mix (β), which are the three states in the experimental environment pattern β, as training data for identification by SVM. The inventor then combined A (β) and B (β) and labeled one fan as “a single fan” and labeled Mix (β) as two fans as “two fans”. . Note that the test data includes all data.

  FIG. 10 (1) shows the result when the correlation coefficient shown in the above equation 18 is used for the similarity in the clustering of the ICA basis function, and FIG. 10 (2) shows the above equation 19 in the similarity. The results when using the Jensen-Shannon information amount shown in Fig. 5 are shown. 10 (1) and 10 (2), the leftmost column indicates training data, and the uppermost column indicates test data.

  10 (1) and 10 (2), the recognition result of the number of objects by the object number detection device 1, that is, the ratio of whether or not the number of objects has been correctly recognized is expressed as a percentage. More specifically, the horizontal axis is “a single fan”, the vertical axis is A (α), B (α), A (β), B (β) cells, and the horizontal axis is “two fans”. The cells whose vertical axis is Mix (α) and Mix (β) indicate the rate at which the object number detection apparatus 1 correctly recognizes the number of objects. Also, the horizontal axis is “two fan”, the vertical axis is A (α), B (α), A (β), B (β) cells, and the horizontal axis is “a single fan”. , Mix (α) and Mix (β) cells indicate the percentages in which the object number detection apparatus 1 cannot correctly recognize the object number.

  Here, as described above, the output signal of the Doppler sensor 2 is proportional to the speed and the distance. Since the inventor uses the experimental environment pattern β in which the distance between the Doppler sensor 2 and the electric fan 5 is 2.0 meters as training data, the number of objects in the experimental environment pattern α is different from the training data. It is inferred that it was difficult to identify this.

  FIG. 10 (3) shows the difference in the correct answer rate according to the similarity calculation index in the clustering of the ICA basis functions. The correct answer rate is the rate at which the object number detection apparatus 1 can correctly recognize the number of objects. As shown in FIG. 10 (3), using the Jensen-Shannon information amount as a similarity calculation index has better object number recognition accuracy than using the correlation coefficient. Further, the recognition accuracy when the Jensen-Shannon information amount is used as the similarity calculation index is 95.7% at the minimum as shown in FIG.

<5. Effect>
As described above, the object number detection apparatus 1 can estimate the number of reflective objects that perform periodic motion. In the present embodiment, the object number detection device 1 detects the number of objects for a plurality of fans 5, but the present invention can also be applied to a person 4. This is because the Doppler signal obtained by the swing motion of the fan 5 has a wide frequency distribution and a component close to that of the person 4, and the plurality of fans 5 perform the swing motion at different periods. This is because the person 4 corresponds to breathing with different periods. As described above, the present invention can detect the number of people 4 and therefore can be used as crime prevention or monitoring technology in ordinary homes or public spaces.

  Independent component analysis outputs a number of ICA basis functions equal to the number of input signals. For this reason, the object number detection apparatus 1 is required to generate a large number of input signals for independent component analysis in order to finely express the characteristics of the Doppler signal. In response to such a request, the object number detection device 1 represents a Doppler signal in a spectrogram by Fourier transform, and further separates it with positive and negative frequencies, so that a large number of signals for one input from the Doppler sensor 2 can be obtained. Is generated. Therefore, the object number detection device 1 can express the characteristics of the Doppler signal in detail.

  In general, for an overcomplete problem that the number of signal sources is larger than the number of sensors, obtaining information on a plurality of signal sources based on a single output from the Doppler sensor 2 This is difficult unless the motion frequencies corresponding to the sources are far apart. In response to such a problem, the object number detection apparatus 1 performs independent component analysis on the assumption that the temporal change in amplitude in frequency units obtained by short-time Fourier transform is a mixed signal in which the amplitude information of each signal source is mixed. By doing so, it is possible to separate features.

  Further, the object number detection apparatus 1 performs clustering processing on the result of independent component analysis. This processing is based on the independence mitigation measure in which the estimated independent components form an independent subspace, which is shown by the above-described method of independent subspace analysis. Actually, the distribution in the frequency domain of the Doppler signal obtained by observing the reflecting object has a certain skirt centered on the frequency at which the amplitude is peaked. Since the shapes of the time-varying signals of the amplitudes at the base frequencies are similar to each other, it can be expected that similarity appears in a part between the independent components estimated by the independent component analysis. Therefore, it can be said that the object number detection apparatus 1 performing clustering on the result of the independent component analysis is effective as a method for capturing the characteristics of the Doppler signal.

<6. Summary>
The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to such examples. It is obvious that a person having ordinary knowledge in the technical field to which the present invention pertains can come up with various changes or modifications within the scope of the technical idea described in the claims. Of course, it is understood that these also belong to the technical scope of the present invention.

  For example, in the above-described embodiment, the electric fan 5 that performs a single operation called the neck swing operation is the object of object number detection, but the present invention is not limited to such an example. For example, an object composed of a plurality of different movement parts may be handled as a plurality of objects. That is, the object number detection device 1 may detect the number of parts that perform different movements in one object.

1 Object number detection device 1
11 Frequency converter 11
13 Independent signal decomposition unit 13
15 Object number detection unit 15
2 Doppler sensor 2
3 Output device 3
4 people 4
5 Fan 5

Claims (12)

A frequency converter that converts a Doppler signal for an arbitrary reflecting object into a plurality of signals having different frequencies;
An independent signal decomposing unit that decomposes the signals of the plurality of different frequencies converted by the frequency converting unit into independent signals having predetermined independence using temporal change in amplitude of each frequency band;
An object number detection unit that detects the number of the reflective objects based on the independent signal decomposed by the independent signal decomposition unit;
An object number detection device comprising:   The object number detection device according to claim 1, wherein the independent signal decomposing unit decomposes the signals having different frequencies into the independent signals by independent component analysis.   3. The object number detection device according to claim 1, wherein the frequency conversion unit converts the Doppler signal into the plurality of signals having different frequencies by Fourier transform. 4.   The said independent signal decomposition | disassembly part is an object number detection apparatus as described in any one of Claims 1-3 decompose | disassembled into the said independent signal by maximizing the non-Gaussian property of a signal.   5. The object number detection device according to claim 1, wherein the independent signal decomposition unit reduces the number of dimensions of the signals having different frequencies by singular value decomposition and then decomposes the independent signal into the independent signal. .   The object number detection unit generates an integrated signal by classifying and integrating the independent signals based on the similarity of the independent signals, and detects the number of the reflective objects based on the integrated signal. The object number detection apparatus as described in any one of 1-5.   The said object number detection part classifies the said independent signal based on the similarity of the said independent signal calculated by either Pearson's product-moment correlation coefficient or Jensen Shannon information content. Object number detection device.   The object number detection device according to claim 6 or 7, wherein the object number detection unit calculates a feature amount of the integrated signal, and detects the number of the reflection objects by machine learning based on the calculated feature amount.   The object number detection device according to claim 8, wherein the object number detection unit performs ordering of the integrated signals for calculating the feature amount based on a statistic of the integrated signals.   10. The object number detection device according to claim 1, wherein the reflective object is a periodic motion body that performs periodic motion such as a person, an animal, a machine, or the like.   The said reflection object is an object number detection apparatus as described in any one of Claims 1-9 which is a periodic exercise body which performs a periodic vibration other than a building. On the computer,
Converting the Doppler signal for any reflective object into signals of different frequencies;
Decomposing the transformed signals of different frequencies into independent signals having predetermined independence using temporal changes in amplitude of each frequency band;
Detecting the number of reflective objects based on the resolved independent signal;
A program for running

Images