CN104916289A - Quick acoustic event detection method under vehicle-driving noise environment - Google Patents

Abstract

The invention provides a quick acoustic event detection method under a vehicle-driving noise environment, belongs to the field of acoustic event detection of an unmanned vehicle, and is to solve the problem that an existing unmanned vehicle cannot sense sound. The method comprises the following steps: in the training stage, serving audio signals of various target acoustic events and vehicle-driving noise signals as training data, establishing an SVM model for each target acoustic event by utilizing the audio signals, and training a BPNN noise model by utilizing the obtained energy of the vehicle-driving noise signals; and in the identification stage, carrying out filtering on the audio signals of a target acoustic event to be detected collected in real time under the vehicle-driving noise environment, carrying out noise reduction and enhancement by utilizing the BPNN noise model, then, extracting MFCC coefficients, carrying out classification identification according to the MFCC coefficients and by utilizing the SVM model to determine the category of the target acoustic event to be detected, furthermore, determining corresponding whistle language sequence codes, and carrying out retrieval on a current whistle sequence library to determine corresponding whistle language information. The acoustic event detection method is used for the unmanned vehicle.

Description

The detection method of fast acoustic event under driving noise circumstance

Technical field

The invention belongs to unmanned vehicle acoustic events detection field.

Background technology

One of core content of unmanned vehicle research is intelligent behavior decision-making, and the prerequisite of intelligent behavior decision-making is then the automatic sensing to surrounding enviroment in its driving process.The means of perception environmental information can have multiple, wherein the automatic sensing of audio visual information plays an important role in the traveling of unmanned vehicle, but a lot of interactive information between the external world and unmanned vehicle also have much based on sound, as the sound of blowing a whistle of dodging pointed out by various vehicle, the police whistle sound of police car and ambulance, the police whistle sound etc. of level crossing, around perception, these are based on the interactive information of sound, and it is most important for unmanned vehicle to make correct intelligent decision.

Summary of the invention

The object of the invention is cannot the problem of perception surrounding environment sound in order to solve existing unmanned vehicle, the invention provides the detection method of fast acoustic event under a kind of noise circumstance of driving a vehicle.

The detection method of fast acoustic event under driving noise circumstance of the present invention,

Described method comprises the steps:

Step one: in the training stage, using the sound signal of various target acoustical event and driving noise signal as training data, feature intercepting is carried out to the sound signal of each acoustic events, and marks accordingly;

Step 2: according to intercept feature, extract the MFCC coefficient of each target acoustical event, described in be characterized as the audio fragment only comprising acoustic events;

Step 3: according to the MFCC coefficient extracted, adopt the multicategory classification strategy of comparison between two, uses SVMLIB kit to be each target acoustical event establishment SVM model;

Step 4: carry out Fast Fourier Transform (FFT) to driving noise signal, obtains low frequency sub-band and the high-frequency sub-band energy of driving noise, according to each sub belt energy training BPNN noise model obtained;

Step 5: at cognitive phase, filters the sound signal of the target acoustical event to be detected of Real-time Collection under driving noise circumstance, the audio fragment that filtering and target acoustical event have nothing to do;

Step 6: utilize step 4 to obtain BPNN noise model, carries out noise reduction and enhancing to the sound signal after filtering;

Step 7: MFCC coefficient is extracted to noise reduction in step 6 and the sound signal after strengthening, according to the MFCC coefficient extracted, the SVM model adopting step 3 to obtain carries out Classification and Identification, obtains the classification of target acoustical event to be detected;

Step 8: in the acoustic events of step 7 determination classification, comprehensive analysis based on sequential and orientation is carried out to the event of blowing a whistle, obtains the sequential coding of corresponding flute language, according to the flute language sequential coding obtained, current sequence library of blowing a whistle is retrieved, determines corresponding flute language information.

In step 5, the sound signal of the target acoustical event to be detected of Real-time Collection under driving noise circumstance is filtered, the method for the audio fragment that filtering and target acoustical event have nothing to do:

The extraction of pitch period is carried out to the sound signal of Real-time Collection under driving noise circumstance, according to the pitch period extracted, judges whether it is target acoustical event, if not, then filter out, if so, then retain.

Described step 6, utilize step 4 to obtain BPNN noise model, the method for the sound signal after filtration being carried out to noise reduction and enhancing is:

Use step 4 to obtain the high frequency noise of BPNN noise model to the sound signal after filtration to predict, use the HFS of subband spectrum-subtraction to the sound signal after filtration to carry out noise reduction and strengthen processing.

Step 7, method noise reduction in step 6 and the sound signal after strengthening being extracted to MFCC coefficient is:

The spectrum distribution of the resonance peak of the Frame of the sound signal after using the method extraction step five of LPC Power estimation to filter and energy size, use the function of the He Ne laser gain of simulation basilar membrane, be weighted Mel bank of filters;

By the Mel bank of filters after weighting to noise reduction in step 6 and the target acoustical Event Distillation MFCC coefficient after strengthening.

The function of described He Ne laser gain is:

$W_i=\underset{j=1}{\overset{k}{\Σ}}Indicator(M_i,L_i,F_j)*P_j*\exp \[-\frac{{(F_j-M_i)}^2}{{\left(\frac{L_i}{2}\right)}^{3.2}}\];$

In formula:

In the spectrum distribution obtained, subband number is B, i=1,2 ..., B, k resonance peak, j=1,2 ..., k, M _irepresent the center frequency of i-th subband, L _irepresent the width of i-th subband;

F _jrepresent the frequency at the place of a jth resonance peak, P _jrepresent the amplitude of a jth resonance peak;

Indicator (M _i, L _i, F _j) be indicative function, if resonance peak F _jwith M _icentered by, with L _ifor in the Mel subband of width, then return 1, otherwise return 0.

Beneficial effect of the present invention is, the present invention can detect that unmanned vehicle is under running environment accurately, the target acoustical event such as to blow a whistle of surrounding vehicles, and also has recall rate and lower misclassification rate more accurately in a noisy environment.Achieve the fast acoustic event detection under driving noise circumstance, have certain robustness to noise.

Accompanying drawing explanation

Fig. 1 is the principle schematic at cognitive phase in embodiment one.

Fig. 2 is the driving low frequency sub-band of noise and the principle schematic of high-frequency sub-band energy, and FL is frequency length, frequency centered by Ms, and n is frame number.

Fig. 3 is the schematic diagram of the KL divergence between Fig. 2 medium and low frequency subband and high-frequency sub-band.

Fig. 4 is the blow a whistle spectrogram of acoustic events and the frequency-region signal schematic diagram of blowing a whistle.

Fig. 5 be blow a whistle in Fig. 4 acoustic events resonance peak information weighting before and after the frequency diagram of Mel bank of filters.

Embodiment

Embodiment one: composition graphs 1 illustrates present embodiment, the detection method of fast acoustic event under the driving noise circumstance described in present embodiment,

Described method comprises the steps:

In the training stage:

(1) the car microphone array recording system in different vehicle is utilized, gather the sound signal of various target acoustical event and drive a vehicle noise signal (such as, jolt, meeting, wind made an uproar, the chatter etc. of engine), using the sound signal of various target acoustical event and driving noise signal as training data, feature intercepting is carried out to the sound signal of each acoustic events, and marks accordingly;

(2) according to intercept feature, extract the MFCC coefficient of each target acoustical event, described in be characterized as the audio fragment only comprising acoustic events;

(3) according to the MFCC coefficient extracted, adopt the multicategory classification strategy of comparison between two, use SVMLIB kit to be each target acoustical event establishment SVM model;

(4) Fast Fourier Transform (FFT) is carried out to driving noise signal, obtain low frequency sub-band and the high-frequency sub-band energy of driving noise, according to the energy training BPNN noise model obtained;

Described (4) need to set up the relation between low frequency sub-band energy and high-frequency sub-band energy; First Mel frequency domain is carried out to the division of frequency domain sub-band, dividing mode as shown in Figure 2, by calculating the KL divergence between low frequency sub-band and high-frequency sub-band, the distribution of KL divergence as shown in Figure 3, demonstrate the scheme can predicting high-frequency sub-band energy with low frequency sub-band energy, use traditional BPNN model training.

Composition graphs 1, at cognitive phase:

(1) utilize microphone array recording system Real-time Collection in unmanned vehicle to all voice signals, the voice signal first carried out gathering carries out pre-service, and pre-service comprises pre-emphasis, framing, windowing and FFT conversion;

(2) pretreated signal is carried out to the extraction of pitch period, according to the pitch period extracted, judge whether it is target acoustical event, if not, then filter out, if so, then retain.This step is the audio fragment that filtering and target acoustical event have nothing to do;

(3) by the extraction of pitch period and after detecting, the spectrum distribution of the resonance peak of the Frame of the sound signal after using the method extraction step five of LPC Power estimation to filter and energy size, use the function of the He Ne laser gain of simulation basilar membrane, Mel bank of filters is weighted; Mel bank of filters around resonance peak is weighted, highlights the frequency around resonance peak.Thus the He Ne laser imitating people's ear basilar memebrane maps.This meets the hearing mechanism of people's ear, as wanted very to know that others speaks in numerous noisy ground unrest, must improve the decibel of sound, covering other people sound;

(4) use step 4 to obtain the high frequency noise of BPNN noise model to the sound signal after filtration to predict, use the HFS of subband spectrum-subtraction to the sound signal after filtration to carry out noise reduction and strengthen processing;

(5) extract MFCC coefficient to noise reduction and the sound signal after strengthening, according to the MFCC coefficient extracted, the SVM model adopting step 3 to obtain carries out Classification and Identification, obtains the classification of target acoustical event to be detected;

Extract MFCC coefficient: by the Mel bank of filters after weighting to noise reduction in step 6 and the target acoustical Event Distillation MFCC coefficient after strengthening;

The function of described He Ne laser gain is:

$W_i=\underset{j=1}{\overset{k}{\Σ}}Indicator(M_i,L_i,F_j)*P_j*\exp \[-\frac{{(F_j-M_i)}^2}{{\left(\frac{L_i}{2}\right)}^{3.2}}\];$

In formula:

In the spectrum distribution obtained, subband number is B, i=1,2 ..., B, k resonance peak, j=1,2 ..., k, M _irepresent the center frequency of i-th subband, L _irepresent the width of i-th subband;

F _jrepresent the frequency at the place of a jth resonance peak, P _jrepresent the amplitude of a jth resonance peak;

Indicator (M _i, L _i, F _j) be indicative function, if resonance peak F _jwith M _icentered by, with L _ifor in the Mel subband of width, then return 1, otherwise return 0.

(6) in the acoustic events determining classification, comprehensive analysis based on sequential and orientation is carried out to the event of blowing a whistle, obtain the sequential coding of corresponding flute language, obtain the sequential coding of corresponding flute language, according to the flute language sequential coding obtained, current sequence library of blowing a whistle is retrieved, determines corresponding flute language information;

Table 1 utilizes present embodiment to the test result of target acoustical event detection, test data is unmanned vehicle various acoustic events of blowing a whistle in the process of moving, comprise yowl flute (each yowl flute duration is greater than 1.5 seconds), comprise 100, shortly to blow a whistle (each short duration of blowing a whistle is less than 1 second), comprising is 100, flute language 25, (being made up of length sequence of blowing a whistle).Result indicates present embodiment can the target acoustical event such as detection and Identification flute language fast in a noisy environment; Fig. 4 is the blow a whistle spectrogram of acoustic events and the frequency-region signal schematic diagram of blowing a whistle.Fig. 5 be according to the resonance peak information weighting of acoustic events of blowing a whistle in Fig. 4 before and after the frequency diagram of Mel bank of filters.

The test result of table 1 target acoustical event detection

Claims (5)

1. the detection method of fast acoustic event under noise circumstance of driving a vehicle, it is characterized in that, described method comprises the steps:

Step one: in the training stage, using the sound signal of various target acoustical event and driving noise signal as training data, feature intercepting is carried out to the sound signal of each acoustic events, and marks accordingly;

Step 2: according to intercept feature, extract the MFCC coefficient of each target acoustical event, described in be characterized as the audio fragment only comprising acoustic events;

Step 3: according to the MFCC coefficient extracted, adopt the multicategory classification strategy of comparison between two, uses SVMLIB kit to be each target acoustical event establishment SVM model;

Step 4: carry out Fast Fourier Transform (FFT) to driving noise signal, obtains low frequency sub-band and the high-frequency sub-band energy of driving noise, according to each sub belt energy training BPNN noise model obtained;

Step 5: at cognitive phase, filters the sound signal of the target acoustical event to be detected of Real-time Collection under driving noise circumstance, the audio fragment that filtering and target acoustical event have nothing to do;

Step 6: utilize step 4 to obtain BPNN noise model, carries out noise reduction and enhancing to the sound signal after filtering;

Step 7: MFCC coefficient is extracted to noise reduction in step 6 and the sound signal after strengthening, according to the MFCC coefficient extracted, the SVM model adopting step 3 to obtain carries out Classification and Identification, obtains the classification of target acoustical event to be detected;

Step 8: in the acoustic events of step 7 determination classification, comprehensive analysis based on sequential and orientation is carried out to the event of blowing a whistle, obtains the sequential coding of corresponding flute language, according to the flute language sequential coding obtained, current sequence library of blowing a whistle is retrieved, determines corresponding flute language information.

2. the detection method of fast acoustic event under driving noise circumstance according to claim 1, it is characterized in that, in step 5, the sound signal of the target acoustical event to be detected of Real-time Collection under driving noise circumstance is filtered, the method for the audio fragment that filtering and target acoustical event have nothing to do:

The extraction of pitch period is carried out to the sound signal of Real-time Collection under driving noise circumstance, according to the pitch period extracted, judges whether it is target acoustical event, if not, then filter out, if so, then retain.

3. the detection method of fast acoustic event under driving noise circumstance according to claim 1, is characterized in that,

Described step 6, utilize step 4 to obtain BPNN noise model, the method for the sound signal after filtration being carried out to noise reduction and enhancing is:

Use step 4 to obtain the high frequency noise of BPNN noise model to the sound signal after filtration to predict, use the HFS of subband spectrum-subtraction to the sound signal after filtration to carry out noise reduction and strengthen processing.

4. the detection method of fast acoustic event under driving noise circumstance according to claim 1, is characterized in that, step 7, and method noise reduction in step 6 and the sound signal after strengthening being extracted to MFCC coefficient is:

The spectrum distribution of the resonance peak of the Frame of the sound signal after using the method extraction step five of LPC Power estimation to filter and energy size, use the function of the He Ne laser gain of simulation basilar membrane, be weighted Mel bank of filters;

By the Mel bank of filters after weighting to noise reduction in step 6 and the target acoustical Event Distillation MFCC coefficient after strengthening.

5. the detection method of fast acoustic event under driving noise circumstance according to claim 4, it is characterized in that, the function of described He Ne laser gain is:

$W_i=\underset{j=1}{\overset{k}{\Σ}}Indicator(M_i,L_i,F_j)*P_j*\exp \[-\frac{{(F_j-M_i)}^2}{{\left(\frac{L_i}{2}\right)}^{3.2}}\];$

In formula:

In the spectrum distribution obtained, subband number is B, i=1,2 ..., B, k resonance peak, j=1,2 ..., k, M _irepresent the center frequency of i-th subband, L _irepresent the width of i-th subband;

F _jrepresent the frequency at the place of a jth resonance peak, P _jrepresent the amplitude of a jth resonance peak;

Indicator (M _i, L _i, F _j) be indicative function, if resonance peak F _jwith M _icentered by, with L _ifor in the Mel subband of width, then return 1, otherwise return 0.