CN101145280A - Independent component analysis based automobile sound identification method - Google Patents

Abstract

The present invention relates to a vehicle audio identification method based on independent component analysis, and belongs to the audio processing and mode identification technology field. The utility model has the steps that audio frequency spectrums of different vehicles are gotten by a pretreatment of different vehicle sounds, frequency spectrums are trimmed down according to characteristics of the vehicle sounds, and coefficients of frequency domain are changed to log domain, thereby enhancing the robustness of the audio identification of the vehicle. Getting the characteristics of the vehicle sounds by using the independent component analysis adapts to small sample training characteristics of vehicle identification. The samples which are ready for identifying are rebuilt in a characteristic space consisting of independent components, a Euclidean distance between the samples which are ready for identifying and the center of the vehicle is gotten, and the samples which are ready for identifying are classified according to the Euclidean distance. The present invention can identify the vehicle sounds highly efficiently and fleetly, thereby especially adapting to real time computation circumstance. The present invention can be used in a plurality of circumstances like detecting of passing vehicles in military field and intelligent traffic system in civilian field and so on.

Description

Automobile sound identification method based on independent component analysis

Technical field

The present invention relates to a kind of automobile sound identification method, belong to Audio Processing and mode identification technology based on independent component analysis.

Background technology

Independent component analysis (Independent Component Analysis, ICA) and principal component analysis (PrincipalComponent Analysis, PCA) in area of pattern recognition, method and the thinking handled have a lot of similarities, these two kinds of methods can realize the linear expression of observation data, data can be projected to more significant lower dimensional space from higher dimensional space, to reach dimensionality reduction and to reduce redundant purpose, and handle accordingly, as data structure analysis, data characteristics extraction etc.The decomposition of making by the PCA principle can only guarantee that each component that decomposes out is uncorrelated, can not guarantee these components mutually independent (unless they all are the Gaussian processes, because to the uncorrelated independence that just means of Gaussian).This just makes this decomposition lack practical significance, thereby has reduced the typicalness of the feature of extracting.In this case, adopt ICA to decompose isolated component, from each isolated component, extract features relevant again, more be of practical significance, also help further pattern-recognition.

Experiment shows that under the small sample situation, ICA is more much better than than the ability of PCA feature extraction, and this has proved that the ICA feature extraction has good rapidity, can extract effective feature from limited sample rapidly.This mainly is because PCA has only utilized the second-order statistics rule of sample based on covariance matrix, ICA then is a higher order statistical rule of utilizing sample. from the viewpoint of statistics, most of important pattern information often is included in the higher order statistical characteristic: the principal component that PCA decomposites is decorrelation (orthogonal), and ICA has not only realized the decorrelation of PCA, and the high-order statistic that obtains is separate, in statistical theory, independently be that statistical independence is comprising correlativity than uncorrelated much better than condition; PCA is based upon on the basis of Gauss's hypothesis, and ICA is based upon on the basis of non-Gauss's hypothesis, under the small sample situation, Gauss's hypothesis of sample distribution generally is invalid, at this moment the feature extraction ability of ICA is obviously good than PCA, under the large sample situation, sample distribution approaches Gaussian distribution, and at this moment the feature extraction ability of PCA will be significantly improved.

And in the identification of vehicle sounds, because environment is to the speed of the influence of extracting sound, the speed of a motor vehicle, still exist individual difference with the kind vehicle, this makes that feature database of setting up vehicle sounds is a problem of extracting validity feature from the small sample training set.Based on these characteristics of vehicle sounds identification, ICA is fit to the identification of vehicle sounds more than PCA.

Summary of the invention

The objective of the invention is to propose a kind of automobile sound identification method based on independent component analysis, it carries out discriminator by sound to vehicle, will export detected vehicle and belong to which kind of type of vehicle, and perhaps output can't be carried out sorting result.Can be used for carrying out the quantity statistics of vehicle identification classification and particular vehicle.

The present invention is achieved in that

Utilize the frequency domain information of vehicle sounds to carry out the discriminator of vehicle based on the automobile sound identification method of independent component analysis, comprise training and two processes of vehicle sounds feature identification classification of vehicle sounds feature database.Wherein:

The training process of vehicle sounds feature database is the training of supervision formula, the vehicle of each particular types is all needed to learn in advance the sound of particular vehicle.Every kind of particular vehicle sound is all carried out following feature learning process.

Step 1: utilize audio collecting device collection vehicle voice signal.The sample frequency R of audio collecting device can be as required accuracy of identification and the decision of computing speed, but (audio frequency that the mankind can hear is greatly about 20Hz ~ 20KHz) than the susceptibility height of people ear.

Step 2: the pre-service before the frequency spectrumization.At first with the making zero of average amplitude of sound waveform (so-called making zero of average amplitude equals the sampled value addition of all sampled points to 0 exactly, can by all sampled values and a suitable constant are realized the making zero of average amplitude of waveform in the Calais mutually); The voice data that obtains is divided into N frame (F _n; N=1,2 ..., N), every frame comprises M sampled voice point (S _Ni; N=1,2 .., N; I=1,2 ..., M), and for guaranteeing that smoothly can making between every frame and the consecutive frame of result has overlapping of L sampled point.Utilize of the influence of pre-service hamming window smoothing filter filtering weakening Gibbs' effect again to the Fourier transform of back.Hamming window smoothing filter is specific as follows:

$w_i=0.54-0.46\cos \left(2\mathrm{\π}\frac{i}{M-1}\right),0,1,...,M-1---\left(1\right)$

x _ni＝x _niw _i,n＝1，2，...，N;i＝1，2，...,M (2)

Step 3: set up frequency spectrum.At first utilize fast fourier transform that frequency domain is arrived in the vehicle sounds signal transformation of every frame, this method is only handled the power spectrum of signal, for the size that reduces calculated amount and feature database can be selected preceding T power spectrum coefficient (Φ _n"={ φ " _N1, φ " " _N2..., φ " _NT; N=1,2 .., N).For the convenience of further handling preceding T power spectrum coefficient carried out normalized according to formula (3) at last, even $\underset{i=0}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\φ}}_{\mathrm{ni}}^{\′}=1.$

${\mathrm{\Φ}}_n^{\′}=\{{\mathrm{\φ}}_{n1}^{\′},{\mathrm{\φ}}_{n2}^{\′}\text{,...,}{\mathrm{\φ}}_{\mathrm{nt}}^{\′}\}=\{\frac{{\mathrm{\φ}}_{n1}^{\′\′}}{T},\frac{{\mathrm{\φ}}_{n2}^{\′\′}}{T},...,\frac{{\mathrm{\φ}}_{\mathrm{nt}}^{\′\′}}{T}\};n=1,2,...,i=1,2,...,i=1,2,...,T$ $\begin{array}{ccc}\underset{i=0}{\mathrm{\Σ}}{\mathrm{\φ}}_{\mathrm{ni}}^{\′\′}& \underset{i=0}{\mathrm{\Σ}}{\mathrm{\φ}}_{\mathrm{ni}}^{\″\′}& \underset{i=0}{\mathrm{\Σ}}{\mathrm{\φ}}_{\mathrm{ni}}^{\″\′}\end{array}$ (3)

Step 4: frequency spectrum adjustment.Frequency spectrum is carried out following logarithmetics to be handled:

φ _ni＝F ₂1g(F ₁φ′ _ni+1);n＝1，2，...,N,i＝1，2，...,T (4)

Wherein, the F in the formula (4) ₁And F ₂The following rule of definite use: in the following step 6 during the PCA dimension-reduction treatment along with the increase of sound characteristic training set, the eigenwert of the covariance matrix of sound characteristic training set will change, F ₁And F ₂The eigenwert of choosing the covariance matrix that will make the sound characteristic training set change less.

The second, narration utilizes independent component analysis to extract the vehicle sounds feature.Through Φ after the above pre-service _n={ φ _N1, φ n ₂..., φ _NT; N=1,2 ..., N.Following vehicle sounds feature extraction and identifying are all based on this power spectrum.

Step 5: with Φ _n; N=1,2 ..., N is connected to the row vector { φ of a N * T dimension ₁₁, φ ₁₂..., φ _1T, φ ₂₁, φ ₂₂..., φ _2T... .. φ _N1, φ _N2..., φ _NT, for the training sample of K similar vehicle sounds, form K * N*T matrix X, sample average $\stackrel{\_}{X}=E\left(X\right).$

Step 6: can take more training sample in order to improve discrimination, like this can be with PCA to its dimensionality reduction, get its preceding d principal component, X behind the dimensionality reduction is d * N*T matrix. regard X as be made up of d N*T dimension observation signal vector matrix, establishing this group observation signal is to be formed by d isolated component linear hybrid.

Step 7: X is carried out independent component analysis, isolate d isolated component, by these isolated components u1, u2 ... one group of base in ud constitutive characteristic space, the subspace that this d base vector is opened has just formed the feature space of describing a class vehicle sounds.

Step 8: the feature database of setting up environmental noise class and various vehicle sounds.Environmental noise as a classification, can be set up different environmental noise classes for improving discrimination: overcast and rainy roadside noise-like, overcast and rainy open-air noise-like, sunny day roadside noise-like, high wind environmental noise class or the like.Also should under various typical environment, gather sound and extract the feature database that feature is formed a class vehicle the feature extraction of various vehicles.

The 3rd, narration utilizes independent component analysis identification vehicle sounds.

Step 9:, in the feature space of all sound class, rebuild its model according to following three formula for a vehicle sounds x to be identified.

$\[a_1,a_2,...,a_d\]=(x-\stackrel{\‾}{X})\×{\left(p^{+}\right)}^{-1},---\left(5\right)$

$\tilde{x}=\underset{i=1}{\overset{d}{\mathrm{\Σ}}}{a}_i{u}_i+\stackrel{\‾}{X},---\left(6\right)$

$E\left(x\right)={\left|\right|x-\stackrel{\‾}{x}\left|\right|}^2,---\left(7\right)$

Formula (5) is for asking projection coefficient after a certain feature space goes average, and formula (6) is illustrated in the reconstruction model of this feature space to vehicle sounds, and formula (7) is illustrated in the reconstruction model of this feature space and the reconstruction error between the vehicle sounds to be identified.Wherein, (p ⁺) ^-1Be the pseudoinverse of isolated component composition matrix, x is at the reconstruction model of feature space to vehicle sounds x to be identified, and E (x) is the error of reconstruction model and actual former input.

For vehicle sounds to be identified, if with respect to the reconstruction error minimum of certain feature space, show that this vehicle sounds meets the description in individual features space most, then corresponding class is recognition result.

For vehicle sounds to be identified, given sample x to be identified is if r satisfies

$E^r\left(x\right)=\underset{i}{\min }\left\{E^i\left(x\right)\right\}---\left(8\right)$

Vehicle sounds sample then to be identified belongs to r class vehicle sounds pattern.

But method step brief summary of the present invention is: obtain the frequency spectrum of vehicle sounds by appropriate pre-service, and frequency spectrum reduced according to the characteristics of vehicle sounds, and with the transformation of coefficient of frequency domain to log-domain, improved the robustness of vehicle sounds identification; Utilize independent component analysis to extract the feature of vehicle sounds, can adapt to the small sample training characteristics of vehicle identification; Sample to be identified is rebuild on the feature space that isolated component constitutes, and obtained sample to be identified and vehicle class centre distance, and sort out to the sample identified according to this distance.

The invention has the beneficial effects as follows: the automobile sound identification method based on independent component analysis can be used for fields such as intelligent transportation or military surveillance according to vehicle calculating vehicle number.

Description of drawings

Fig. 1 is the theory diagram that the present invention is based on the automobile sound identification method of independent component analysis.

Embodiment:

The present invention is described in further detail below in conjunction with accompanying drawing.

The target of demo system is that Light-duty Vehicle and heavy goods vehicles are distinguished in roadside vehicle classification identification.

Automobile sound identification method based on independent component analysis utilizes the frequency domain information of vehicle sounds to carry out the discriminator of vehicle, training and two processes of vehicle sounds feature identification classification of comprising the vehicle sounds feature database, the spectral vectors that obtains through adjusting is these two common processing procedures, and hereinafter that this is a part of processing procedure unification is called preprocessing process.

1, preprocessing process

Step 1: utilize audio collecting device collection vehicle voice signal.Because the concentration of energy of vehicle sounds 90% should be not less than 8000HZ according to the sampling thheorem sample frequency below 4000HZ, the sample frequency of audio collecting device is elected 22050HZ as like this.

Step 2: the pre-service before the frequency spectrumization.At first with the making zero of average amplitude of sound waveform; The voice data that obtains is divided into N frame (F _n; N=1,2 ..., N), every frame comprises 4096 sampled voice point (S _Ni; N=1,2 ..., N; I=1,2 ..., 4096), and for guaranteeing that smoothly can making between every frame and the consecutive frame of result has overlapping of 512 sampled points.Utilize of the influence of pre-service hamming window smoothing filter filtering weakening Gibbs' effect again to the Fourier transform of back.Hamming window smoothing filter is specific as follows:

$w_i=0.54-0.46\cos \left(2\mathrm{\π}\frac{i}{4095}\right),i=\mathrm{0,1},...,4095---\left(9\right)$

x _ni＝x _niw _i,n＝1，2，...,N;i＝1,2，...,4096 (10)

Step 3: set up frequency spectrum.At first utilize fast fourier transform that frequency domain is arrived in the vehicle sounds signal transformation of every frame.Because phase information is unimportant for voice recognition, this method is only handled the power spectrum of signal, the concentration of energy that experiment shows most of vehicles about 80% is lower than the scope of 2000Hz in frequency, and about 90% concentration of energy is lower than the scope of 4000Hz in frequency.For the size that reduces calculated amount and feature database can be selected preceding 1200 power spectrum coefficient (Φ _n＂={ φ ＂ _N1, φ ＂ _N2..., φ ＂ _N1200; N=1,2 ..., N) cover the frequency range of 5.4HZ-6460Hz.For the convenience of further handling preceding 1200 power spectrum coefficients are carried out normalized according to formula (11) at last.

${\mathrm{\Φ}\′}_n=\{{\mathrm{\φ}\′}_{n1},{\mathrm{\φ}\′}_{n2},...{,\mathrm{\φ}\′}_{n1200}\}=\{\frac{{\mathrm{\φ}\″}_{n1}}{\underset{i=0}{\overset{1200}{\mathrm{\Σ}}}{\mathrm{\φ}}_{\mathrm{ni}}^{\′\′}},\frac{{\mathrm{\φ}\″}_{n2}}{\underset{i=0}{\overset{1200}{\mathrm{\Σ}}}{\mathrm{\φ}}_{\mathrm{ni}}^{\′\′}},...,\frac{{\mathrm{\φ}\″}_{n1200}}{\underset{i=0}{\overset{1200}{\mathrm{\Σ}}}{\mathrm{\φ}}_{\mathrm{ni}}^{\′\′}}\};n=1,2,...,N,i=1,2,...,1200---\left(11\right)$

Step 4: frequency spectrum adjustment.For the identification that makes vehicle sounds robust more, avoid the result of specific frequency spectrum details left and right sides vehicle sounds identification, vehicle sounds feature extraction and identifying need consider that entire spectrum distributes.In order to reach this purpose frequency spectrum is carried out following logarithmetics processing:

φ _ni＝F ₂lg(F ₁φ′ _ni+1);n＝1，2，...,N，i＝1，2，...，1200 (12)

Wherein, F ₁=10000, F ₂=100.

2, the foundation of vehicle sounds feature database

Φ _n={ φ _N1, φ _N2..., φ _N1200; N=1,2 ..., N, a frame (4096 sampled points) that is a vehicle sounds sample is through pretreated power spectrum coefficient, and the foundation of following vehicle sounds feature database is based on this power spectrum.

Step 1: sound and the noise of gathering different automobile types.Gather environment: overcast and rainy roadside, sunny day roadside, high wind environment roadside; The speed of collection vehicle: 20km/h, 80km/h, 120km/h, every class vehicle is gathered three cars of this kind.Each class will have 9 training samples like this.The noise class is gathered environment: overcast and rainy roadside, and sunny day roadside, will there be 3 samples in high wind environment roadside.The foundation of feature database is established as example with the feature database of Light-duty Vehicle.

Step 2: with Φ _n; N=1,2 ..., N is connected to the row vector { φ of a N * 1200 dimensions ₁₁, φ ₁₂..., φ _1,1200, φ ₂₁, φ ₂₂..., φ _2,1200... .. φ _N1, φ _N2..., φ _{N, 1200}, for the training sample of 9 similar vehicle sounds, form 9 * N*1200 matrix X, sample average X=E (X)

Step 3: can take more training sample in order to improve discrimination, like this can be with PCA to its dimensionality reduction, get its preceding 100 principal components, X behind the dimensionality reduction is 100 * N*1200 matrix. regard X as be made up of 100 N*T dimension observation signal vectors matrix, establishing this group observation signal is to be formed by 100 isolated component linear hybrid.

Step 4: adopt the FastICA algorithm to carry out independent component analysis to X, isolate 100 isolated components, by these isolated components u ₁, u ₂... u ₁₀₀One group of base in constitutive characteristic space, the subspace that these 100 base vectors are opened has just formed the feature space of describing a class vehicle sounds.

Step 5: with X and u ₁, u ₂... u ₁₀₀Preservation is as the feature database of a class vehicle.

3, the identification vehicle sounds

Step 1:, respectively at the Light-duty Vehicle sound class, rebuild its model in other feature space of heavy goods vehicles sound class and noise-like according to following three formula for a vehicle sounds x to be identified.

[a ₁,a ₂,…,a ₁₀₀]＝(x-X)×(p ⁺) ^-1 (13)

$\tilde{x}=\underset{i=1}{\overset{100}{\mathrm{\Σ}}}{a}_i{u}_i\text{+}\stackrel{\_}{X},---\left(14\right)$

$E\left(x\right)={\left|\right|x-\tilde{x}\left|\right|}^2,---\left(15\right)$

Formula (13) is for asking projection coefficient after a certain feature space goes average, and formula (14) is illustrated in the reconstruction model of this feature space to vehicle sounds, and formula (15) is illustrated in the reconstruction model of this feature space and the reconstruction error between the vehicle sounds to be identified.Wherein, (p ⁺) ^-1Be the pseudoinverse of isolated component composition matrix,

For at the reconstruction model of feature space to vehicle sounds x to be identified, E (x) is the error of reconstruction model and actual former input.

For vehicle sounds to be identified, if with respect to the reconstruction error minimum of certain feature space, show that this vehicle sounds meets the description in individual features space most, then corresponding class is recognition result.

Step 2: for vehicle sounds to be identified, given sample x to be identified is if r satisfies

$E^r\left(x\right)=\underset{i}{\min }\left\{E^i\left(x\right)\right\}---\left(16\right)$

Vehicle sounds sample then to be identified belongs to r class vehicle sounds pattern.

Sum up above explanation as can be known, the present invention includes following steps:

1) various types of vehicles sound is carried out pre-service, obtain the frequency spectrum of various types of vehicles sound;

2) utilize independent component analysis method to extract the feature of various types of vehicles sound respectively;

3) sample to be identified is rebuild its model respectively on the feature space of the various types of vehicles that constitutes with isolated component, ask for the Euclidean distance of sample to be identified and various types of vehicles respectively, with the class of vehicle of the pairing model correspondence of its middle distance reckling as recognition result.

Above embodiment is the unrestricted technical scheme involved in the present invention in order to explanation only, although the present invention is had been described in detail with reference to above preferred embodiment, those of ordinary skill in the art should be appreciated that technical scheme of the present invention can make amendment, is out of shape or is equal to replacement; And do not break away from the spirit and scope of technical solution of the present invention, all should be encompassed among the claim scope of the present invention.

Claims (4)

1. automobile sound identification method based on independent component analysis is characterized in that may further comprise the steps:

1) various types of vehicles sound is carried out pre-service, obtain the frequency spectrum of various types of vehicles sound;

2) utilize independent component analysis method to extract the feature of various types of vehicles sound respectively;

3) sample to be identified is rebuild its model respectively on the feature space of the various types of vehicles that constitutes with isolated component, ask for the Euclidean distance of sample to be identified and various types of vehicles respectively, with the class of vehicle of the pairing model correspondence of its middle distance reckling as recognition result.

2. the automobile sound identification method based on independent component analysis according to claim 1 is characterized in that:

Described pre-treatment step 1) further comprise following substep:

1.1) utilize audio collecting device to gather the various types of vehicles voice signal;

1.2) at first with making zero of average amplitude; The voice data that obtains is divided into the N frame and makes between every frame and the consecutive frame overlapping of L sampled point arranged; Again each frame is used the hamming window filtering;

1.3) frequency domain is arrived in the vehicle sounds signal transformation of every frame, and draw its power spectrum.

3. the automobile sound identification method based on independent component analysis according to claim 2 is characterized in that:

In step 1.3) in, in order to reduce the size of calculated amount and feature database, select preceding T power spectrum coefficient, and the power spectrum coefficient is carried out normalized;

Continue frequency spectrum set-up procedure 1.4 afterwards):

1.4) for the identification that makes vehicle sounds robust more, with the frequency spectrum logarithmetics.

4. according to the described automobile sound identification method of one of claim 1-3, it is characterized in that based on independent component analysis:

The characterization step 2 of described extraction various types of vehicles sound) further comprise following substep:

2.1) every class vehicle sounds is connected to the row vector through the power spectrum coefficient of each frame of obtaining after the pre-service, the sample of similar vehicle sounds is formed matrix X, and write down its sample average

2.2) utilize principal component method to matrix X dimensionality reduction;

2.3) matrix behind the dimensionality reduction is carried out independent component analysis, by one group of base in these isolated component constitutive characteristic spaces, the subspace that these base vectors are opened has just formed the feature space of describing a class vehicle sounds;

2.4) with sample average

Save as the feature database of a class vehicle with isolated component.

Images