JP2845850B2 - Automatic target detection method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for automatically detecting a target having a predetermined sound source, and more particularly to a method for automatically detecting a target having a moving engine.

[0002]

2. Description of the Related Art Conventionally, in this type of automatic target detection system, an input audio signal is subjected to spectrum analysis, and a target is detected by paying attention to frequency, amplitude, Doppler frequency information and the like obtained as a result.

FIG. 11 shows an example of a basic configuration of a conventional automatic target detection system for this type of sound source. In the figure, a transducer group 1 converts a sound pressure level of an acoustic signal generated from a sound source into a voltage. The amplifier 2 and the band-pass filter 3 amplify the acoustic signal and attenuate signals outside the required frequency band. Baseband conversion 4 and low-pass filter 5
Is frequency-mixed with the local oscillator output for ease of processing and simplification of the H / W configuration, and is converted to an intermediate frequency that spreads around the local oscillator frequency. The A / D converter 6 converts an analog sound signal into a digital signal, and the directivity synthesis 7 uses the phase of each transducer to improve a signal-to-noise ratio (hereinafter, referred to as S / N) and azimuth resolution. To create a directional beam. The Fourier transform 8 performs a Fourier transform on the signal resulting from the directivity synthesis from a broadband signal including all frequency components in the frequency band in which the signal is engaged to a narrowband frequency-specific signal within a predetermined time, Output as a Fourier component. The phase calculation 9 calculates the azimuth using the phase difference of the Fourier components between the beams, and the amplitude calculation 10 calculates the power spectrum from the Fourier components. The target determination 10 compares the frequency transition (Doppler information 13) of the power spectrum with the passage of time with a reference having a specific frequency 12 or the like as a database, or uses a neural network in which target characteristics have been learned in advance. And outputs it as an automatically detected target together with the azimuth.

Conventionally, this type of automatic target detection system basically has the above-described configuration, and various systems have been developed by a target determination after frequency analysis for obtaining a power spectrum and a technique associated therewith. Yes (10, 11, 1 in the figure)
2). For example, JP-A-2-22799 discloses a method of extracting a characteristic shape characteristic of a frequency change with time and comparing it with a reference characteristic stored in advance.
Japanese Patent Application Laid-Open No. 4-147078 discloses a method of determining a target through a neural network in which frequency information and amplitude information of a power spectrum are learned in advance by a teacher signal.

[0005]

In these conventional automatic target detection methods, a unique frequency component of the target, that is, a power spectrum in which the frequency characteristic appears remarkably, is obtained from the sound signal generated by the target. Although the data processing of the amplitude and frequency shift that are continuous in time is performed, the acoustic signal emitted by the target is generated as a result of changing the state of the surrounding sound or transient sound that appears suddenly in addition to this inherent frequency component. It has a sound. Therefore, in the conventional method, automatic detection is performed using only a part of the characteristic of the sound signal of the target, and it is difficult to perform automatic detection because other useful acoustic characteristics are not used. There is a problem such as detection. In addition, there is a problem in that since a temporally continuous acoustic signal is processed, it cannot respond to a transient signal such as a hatch opening / closing sound or a hull striking sound.

[0006]

In order to solve the above-mentioned problems, an automatic target detection method according to the present invention is characterized in that an acoustic signal is inputted and a sound such as an opening / closing sound of a hatch of a ship or a striking sound of a hull is taken. A transient sound detector that detects a short-time generated sound whose amplitude and frequency changes in the same manner, a transient sound detector that uses this detection result as feature information data, and an acoustic signal is input to detect a continuously generated sound generated by an engine or the like. A continuous sound detector that uses the detection result as feature information data, and the number of piston movements (cylinder rate) and the number of explosions (firing rate) within a unit time of the engine cylinder specific to the target by inputting an acoustic signal Spectral pattern that detects the fundamental frequency component and its harmonic components generated based on the periodic motions, etc., classifies the time variation pattern, and outputs this as feature information data. When the target is in contact with the surrounding medium like a propeller and disturbs the surrounding medium, the ambient noise and the radiated noise of the target are amplitude-modulated in accordance with the fluctuation of the medium. It has a feature detection unit that detects a plurality of different features of the modulation frequency detection unit that detects and estimates the rotation speed of the propeller and the like and uses this as feature information data,
A data integration unit that integrates and outputs each feature information data detected by different feature detection units, and a target certainty factor indicating how likely the detection target is as a target, classifies and outputs the targets It has a target determination unit.

[0007]

Next, the present invention will be described with reference to specific examples.

FIG. 1 is a block diagram showing one embodiment of the present invention.

In FIG. 1, an acoustic signal emitted from a sound source is input to a transducer group M1, converted from a sound pressure level to a voltage level, and subjected to a pre-processing M2 to facilitate subsequent processing. Here, pre-processing M2 is amplification,
Applying a band-pass filter, baseband conversion, and low-pass filter. Thereafter, the A / D conversion M3 converts the analog amount into a digital amount, and performs directivity synthesis M4 by adjusting the phase calculated from the sound wave arrival time between the transducers. Up to this point, it is equivalent to the conventional method. The feature detection unit S1 has a transient sound detection processing unit M5, a continuous sound detection processing unit M6, a spectrum pattern detection processing unit M7, and a modulation frequency detection processing unit M8 as its internal configuration. Is input to the detection processing unit.

The transient sound detection processing section M5 performs a cross-correlation process on a transient signal such as a hatch opening / closing sound and a hull hitting sound with a reference signal in a database.
Those whose results exceed a certain threshold are detected as transient sounds in which the target has occurred. The cross-correlation processing may perform pattern matching between the spectrum appearing in the frequency domain as well as the spectrum of the reference signal in addition to the time domain.

[0011] The continuous sound detection processing unit M6 calculates the ambient noise level of the acoustic signal of the wide band frequency and performs threshold processing, and a signal that exceeds the threshold and is continuous in time is regarded as a continuously generated sound of an engine or the like. Output after removing self-generated noise.

The spectrum pattern detection processing section M7 performs frequency analysis processing such as Fourier transform on the acoustic signal,
Detects fundamental frequency components such as the cylinder rate and firing rate of the engine specific to the target target and its harmonic components, tracks the movement of each detected frequency component over time (hereinafter referred to as line tracking), and its fluctuation pattern Is output.

The modulation frequency detection processing unit M8 has a mechanism in which the target moves periodically like a propeller. When the target moves in contact with the surrounding medium, the medium is disturbed and the surrounding noise and the radiation of the target are emitted. The noise is amplitude-modulated in accordance with the fluctuation of the medium. The modulated noise is detected, an envelope component is obtained, a frequency analysis process is performed, and the frequency interval between the harmonics (hereinafter referred to as pitch) of the propeller is calculated. The number of revolutions, the number of revolutions of the engine shaft, etc. are estimated and output. Instead of directly examining the frequency difference between each harmonic to determine the pitch between the harmonics, a logarithmic transformation is performed on the power spectrum, and the resulting spectrum is Fourier-transformed. The method used may be used.

Each feature detection processing unit M of the feature detection unit S1
Each of the feature detection results output from 5-M8 is compared by the data integration processing unit M9 of the data integration unit S2 with the coincidence of the time and orientation of the output results of each detection processing unit. The results of each feature detection are integrated as a signal from, and if they are out of the range, they are output as separate targets.

The integrated data is determined by the target determination processing unit M10 of the target determination unit S3 based on a combination of weight values for each detection information (level, variation pattern, frequency, signal type, etc.) of each feature detection processing unit. The target certainty factor indicating the target confidence is calculated, and those exceeding the threshold are listed as automatic detection targets, and the frequency characteristics and fluctuation patterns of the detection information are coded and combined, compared with the reference in the database, and the type of the target is determined. Classify and output.

As described above, the present invention includes the feature detecting section S1 having a plurality of feature detecting sections M5 to M8 therein, and integrates each piece of feature information data detected by each feature detecting section. A data processing unit S2 for outputting, and a target determining unit S3 for classifying and outputting the type of the target in order to determine the likelihood of the target
Consists of

Next, the processing units M5 to M10 will be described in detail below. The present invention focuses on integrating and using a plurality of feature detection information data. Therefore, in the feature detection processes M5 to M8, any method can be used as long as a target feature can be detected. .

FIG. 2 is a diagram for explaining an example of the transient sound detection processing M5, and FIG. 2A shows the processing flow. The acoustic signal is subjected to detection processing M501, and an envelope component of the signal is obtained. Next, cross-correlation processing M502 is performed on a plurality of pre-registered reference signals in the database, then threshold processing M503 is performed, and a signal whose calculated cross-correlation coefficient exceeds a threshold is output as a detection target.
(B) shows an example of a signal after the detection processing M501 is output, and (C) shows an example of processing data after the cross-correlation processing M502.

FIG. 3 shows a processing flow of an example of the continuous sound detection processing section M6. The input acoustic signal is subjected to detection processing M601.
After that, the directionally synthesized beam directed in each direction is time-integrated for each beam to reduce noise variation, and the output of the beam having the minimum level among the results is multiplied by K as a reference level. Then, ambient noise calculation processing M602 for determining a threshold level is performed. The signal subjected to the detection processing is subjected to time integration processing M603 separately from the ambient noise calculation processing M602, and a signal exceeding a threshold is obtained in a threshold calculation processing M604. The next continuity determination processing M605 is performed to check whether or not the target exists in the azimuth gate centered on the predicted azimuth obtained using the azimuth data for the past N times. It is determined whether or not there is a continuous presence, and if it is continuous for L seconds or more, it is output as a target. After performing a self-noise removal process M606 for removing a direction in which self-radiation noise is clearly superimposed from the output results detected as targets, the output is output to the subsequent stage. FIG.
4A shows an output example of the threshold processing M604, and FIG. 4B shows an output result example of the self-noise M606. FIG. 4A shows a state where a threshold value is calculated from the obtained ambient noise level, and a value exceeding the threshold value is detected. FIG. 4
(B) shows an example of data detected, not detected, or removed by the continuity determination process and the self-noise removal process.

FIG. 5 shows a spectrum pattern detection processing unit M7.
2 shows a processing flow of an example. The input acoustic signal is subjected to an FFT operation in a spectrum operation processing M701, and then detected to calculate a power spectrum of the input signal. In the maximum value detection processing M702, the power spectrum output from M701 is stored in N samples in the time direction, and the maximum value in N samples is obtained for each frequency cell (peak hold). In the binarization processing M703, a frequency average level and a standard deviation are obtained for each beam, and the average + K *
The threshold level is determined by the standard deviation, "0" is assigned to frequency cells having a spectrum less than the threshold, and "1" is assigned to frequency cells having a spectrum equal to or greater than the threshold, and the signal is assigned to "1" and "0".
Value. In most cases, the binarized signal spectrum has a spread in frequency. Therefore, in the case where a frequency gate is provided and a plurality of frequencies existing in the gate are separated by a predetermined cell interval or more by the thinning processing M704, Once the frequency cells are allotted "1" to all of the gates, the width is expanded to the continuous frequency band, then "1" is allocated only to the center cell, and "0" is allocated to the other cells to make them thinner. . In the encoding process M705, the spectrum and the frequency one timing before are compared, and the change of the frequency cell is encoded according to the difference between the frequency cell number one timing before and the current frequency cell number. For example, a sign “+” is assigned if the cell has moved to the higher frequency cell number by 5 cells, and a sign “I” or the like is assigned if the change is within ± 2 cells because the change is small. These codes are line-tracked in accordance with the proximity of the frequency and the continuity of the appearance time in the list processing M706, converted from the codes into a line list, and classified for each line list. For example, group A; "+ / II
II ... * ", Group B;" IIIIII ...
* ". Thereafter, the collation processing M7
07 to see if they match the variation pattern of the line in the database in the group classified by 07
If there is a match, it is output as a detection target. FIG.
6A shows an output example of the spectrum of the signal after the spectrum calculation processing M701, FIG. 6B shows an output example of the binarized and thinned signal after the thinning processing M704, and FIG.
(C) shows an output example of a code in the line list after the list processing M706.

FIG. 7 shows a processing flow of an example of the modulation frequency detection processing section M8. In the low-pass filter of M801, frequency components higher than the frequency band in which it is engaged are removed.
A detection process M802 is performed on the output of the low-pass filter to obtain a fluctuation envelope component, and a spectrum calculation process M80 is performed.
In No. 3, FFT processing M803A is performed and detection processing M803B is performed.
To obtain a power spectrum, and perform time integration processing M803.
At C, the value of each frequency cell is time-averaged over the past N samples to suppress noise fluctuations and obtain a spectrum including the modulation frequency due to the target and its harmonics. After that, the signal is band-limited by a band-pass filter M803D so that only the frequency band required in the subsequent stage is output. In the threshold processing M804, a small positive value is given to a spectrum having a value less than the threshold, which is a threshold using the average spectrum level of the entire frequency band as a reference value. The reason for giving a small positive value instead of 0 is
This is because no error occurs during the G operation. In the cepstrum calculation processing M805, the output spectrum of the threshold processing M804 is subjected to LOG conversion processing M805A, which is further subjected to FFT.
As a result of the processing M805B and the detection processing M805C, a cepstrum is obtained. Thereafter, the time axis integral M80 is set for each quefrency cell (a cell corresponding to a frequency in the spectrum).
Perform 5D to suppress the fluctuation of the noise component. This cepstrum represents the appearance interval (pitch) of harmonics existing in the power spectrum, and the dimension is time. After this,
In the threshold processing M806, the moving average processing is performed on the quefrency to obtain an average noise level near the corresponding quefrency with respect to the fluctuation of the cepstrum on the quefrency axis, and based on this, a threshold is obtained to remove the noise cepstrum. Lastly, the cepstrum detected by exceeding the threshold value by the data collation processing M807 is compared with a specific target cylinder rate or a quefrency representing a firing rate registered in advance in the database, and if they match, it is output as a target. FIG. 8A shows an output example of a modulated signal after the low-pass filter M801, and FIG. 8B shows an output example of a harmonic spectrum appearing at the frequency interval f1 after the spectrum calculation processing M803. C) Cepstrum operation processing M
1 / f1 representing the pitch of harmonics of the spectrum after 805
3 shows an example of output of a cepstrum signal appearing at the position of.

FIG. 9A shows a processing flow of an example of the data integration processing section M9. The coincidence of the time of the detection target from each feature detection system is checked by the time gate process M901.
The detection times are compared between a plurality of feature detection systems where the target exists. If the detection time is within a certain time gate, the signals are regarded as signals from the same target as if the time matches, and the output data of each feature detection process To integrate. Next, direction gate processing M9
In 02, the coincidence of the detection directions between the feature detection systems is checked in the integrated data after the time gate processing. One detection system is compared with another detection system, and if it is within a certain gate, it is output as it is as if the azimuth is matched, but if it is out of the azimuth gate, it is separated from the integration as a separate target and separated Output to FIG. 9B shows an example of a case where integration is performed and a case where integration is not performed.

FIG. 10 shows a processing flow of the target determination processing section M10. The integrated data information outputs a certainty factor for each feature detection unit in a target certainty factor calculation process M1001, and the target certainty factor is calculated using these. Continuous sound certainty calculation M10
In step 04, the S / N value is extracted from the detection information output from the continuous sound detection processing unit M6, and the corresponding certainty factor is output by referring to a previously prepared S / N lookup table. In the transient sound certainty factor calculation M1005, the transient sound type is extracted from the detection information output from the transient sound detection processing unit M5, and the corresponding certainty factor is output by referring to a type reference table prepared in advance. In the spectral pattern certainty factor calculation M1006, the variation pattern code is extracted from the detection information output from the spectrum pattern detection processing unit M7, and the corresponding certainty factor is output by referring to a previously prepared pattern code reference table. In the modulation frequency certainty calculation M1007, the modulation frequency is extracted from the detection information output from the modulation frequency detection processing unit M8, and the corresponding certainty is output by comparing it with a modulation frequency reference table prepared in advance. In the target certainty factor calculation M1008, the target certainty factors are calculated by combining the respective certainty factors output from the four feature certainty factor calculations from M1004 to M1007. Here, the calculation is performed using the MYCIN-OR method [WINSTON, 1979] used for the medical expert system.
This is a method of adding two confidence factors (positive values less than 1) and then subtracting the multiplied value. If there are three or more elements, this procedure is repeated for the calculated result. Its output does not exceed 1. Target determination processing M1002
Thus, a target certainty degree or more with a target certainty factor is recognized as a target. Furthermore, the combination pattern of each piece of detection information of the data information integrated by the target classification process M1003 is compared with the combination pattern of the combination pattern reference table previously provided as a reference. Classify and output.

[0024]

As described above, the present invention has a plurality of different characteristic detecting units, namely, a transient sound detecting unit, a continuous sound detecting unit, a spectrum pattern detecting unit, and a modulation frequency detecting unit. After integration, the degree of certainty of the target is calculated, and the target is classified and output according to the combination of the detection information. Therefore, since the target is detected not only from one feature but also from a plurality of features and the integrated determination is made, not only the target can be easily detected, but also erroneous detection can be reduced. Further, it is possible to detect not only a temporally continuous signal but also a suddenly generated transient signal.

[Brief description of the drawings]

FIG. 1 is a diagram showing an embodiment of a target automatic detection method according to the present invention.

2A and 2B are diagrams showing a transient sound detection operation according to an embodiment of the present invention, wherein FIG. 2A shows a processing process, FIG. 2B shows an example of an output signal of a detection process M501, and FIG. FIG. 14 is a diagram illustrating a processing result after the correlation processing M502.

FIG. 3 is a diagram showing a process of a continuous sound detection process according to the embodiment of the present invention.

4A and 4B are diagrams showing an operation in the process of FIG. 3, wherein FIG. 4A shows an example of an output result of a threshold process M604, and FIG. 4B shows a process result after a continuity determination process M605.

FIG. 5 is a diagram showing a process of a spectrum pattern detection process according to the embodiment of the present invention.

6A and 6B are diagrams showing an operation in the process of FIG. 5; FIG. 6A shows an example of a processing output of a spectrum calculation process M701;
FIG. 7C is a diagram illustrating a processing output example of a thinning process M704, and FIG. 7C is a diagram illustrating a processing output example of a list process M706.

FIG. 7 is a diagram showing a process of a modulation frequency detection process according to the embodiment of the present invention.

FIGS. 8A and 8B are diagrams showing the process of FIG. 7, wherein FIG. 8A shows an example of an output signal of a low-pass filter M801, FIG. 8B shows an example of an output of a spectrum operation process M803, and FIG. 8C shows a cepstrum operation process M805; Here is an output example.

FIG. 9 is a diagram showing a data integration process according to the embodiment of the present invention;
FIG. 7A is a diagram illustrating an integration process, and FIG.

FIG. 10 is a diagram showing a process of a target determination process according to the embodiment of the present invention.

FIG. 11 shows a conventional example.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Transducer group 2 Amplification part 3 Band pass filter 4 Base band conversion 5 Low pass filter 6 A / D conversion 7 Directivity synthesis 8 Fourier transformation 9 Phase calculation 10 Amplitude calculation 11 Target judgment 12 Specific frequency 13 Doppler information S1 Feature detection part S2 Data integration section S3 Target determination section M1 Transducer group M2 Preprocessing section M3 A / D conversion section M4 Directivity synthesis section M5 Transient sound detection processing section M6 Continuous sound detection processing section M7 Spectrum pattern detection processing section M8 Modulation frequency detection processing section M9 Data integration processing unit M10 Target determination processing unit M501 Detection processing M502 Cross-correlation processing M503 Threshold processing M601 Detection processing M602 Ambient noise level calculation processing M603 Threshold processing M604 Continuity determination processing M605 Self-noise removal processing M701 Vector calculation processing M702 Maximum value detection processing M703 Binarization processing M704 Thinning processing M705 Encoding processing M706 List processing M707 Collation processing M801 Low-pass filter M802 Detection processing M803 Spectrum calculation processing M804 Threshold processing M805 Cepstrum calculation processing M806 Threshold processing Matching processing M803A FFT calculation M803B Detection processing M803C Time axis integration M803D Bandpass filter M804A Frequency axis integration M805A LOG calculation processing M805B FFT processing M805C Detection processing M805D Time axis integration M901 Time target processing M902 M1 azimuth M1 100 degree gate processing Judgment process M1003 Target classification process M1004 Continuous sound target certainty factor calculation M1005 Tiger Stringent sound confidence factor computing M1006 spectral patterns confidence factor computing M1007 modulation frequency confidence factor computing M1008 target confidence calculation

──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Takahiro Kato 3-12-2-6402 Nagase, Yokosuka City, Kanagawa Prefecture (56) References JP-A 8-271620 (JP, A) JP-A 8-220210 (JP) JP-A-8-211151 (JP, A) JP-A-7-44517 (JP, A) JP-A-6-75038 (JP, A) JP-A-8-15402 (JP, A) 4-359799 (JP, A) JP-A-4-208887 (JP, A) (58) Fields investigated (Int. Cl. ⁶ , DB name) G01S 3/80-3/86 G01S 5/18-5 / 30 G01S 7/52-7/64 G01S 15/00-15/96 G01H 3/00 G01V 1/00

Claims (4)

(57) [Claims] 1. An audio signal emitted from a detection target is detected,
In the automatic target detection method for recognizing the arrival direction and the type of the target, the sound signal is input, and a short-time generated sound (hereinafter, referred to as a transient sound) whose amplitude or frequency changes within a short time is detected. A transient sound detection unit that uses the detection result as feature information data, and inputs the acoustic signal to detect a continuously generated sound, obtains a target detailed azimuth based on the detection result, and generates a continuous sound as feature information data. A detection unit, a spectrum pattern for inputting the acoustic signal, detecting a fundamental frequency component and a harmonic component of the sound generated based on the periodic motion, classifying the time-varying pattern, and outputting this as characteristic information data. A detection unit, which receives the acoustic signal, detects noise in which ambient noise or target radiated noise is amplitude-modulated in accordance with the fluctuation of the medium, and detects the noise to estimate the rotational speed. A target automatic detection method, comprising: a modulation frequency detection unit for setting information data. 2. A data synthesizing section for synthesizing and outputting each feature information data detected by the plurality of detecting sections, and classifying the target by obtaining a target certainty factor indicating how likely the detection target is as the target. 2. The automatic target detection method according to claim 1, further comprising: a target determination unit that outputs the result. 3. The data integration unit and the target determination unit further include a time gate processing unit that integrates the feature information data detected by the plurality of feature detection units for each detection time, and a result integrated by the detection time. Integrate by the azimuth gate processing unit that performs integrated processing for each detected azimuth, weighted and combined the integrated feature information data, and found the target certainty that indicates how likely the detected target is as the expected target, 3. The automatic target detection method according to claim 2, wherein the target is classified and output according to a combination of the characteristic information. 4. The sound generated in a short period of time is a sound of opening and closing a hatch of a ship and a sound of striking a hull, and a signal generated based on the periodic motion is a sound due to a piston motion or an explosion of an engine. 2. The target automatic detection method according to claim 1, wherein the sound is a sound obtained by rotating the propeller.