CN106251861A - A kind of abnormal sound in public places detection method based on scene modeling - Google Patents

Abstract

The present invention is a kind of abnormal sound in public places detection method based on scene modeling, the method is little according to public place scene sound relative anomalies sound average amplitude, the statistical property that fluctuation range is narrower, first calculate the average amplitude of each scene acoustical signal, and set up the gauss hybrid models of different scene based on expectation-maximization algorithm；Then ask for the likelihood score of acoustical signal to be measured and model of place, carry out likelihood score coupling, then threshold condition based on most Voting principles and the continuous frame number of minimum judges that voiced frame to be measured, whether as abnormal sound, thus realizes the detection of abnormal sound.The most existing abnormal sound detection method of the present invention, scene adaptability is higher, and the error rate of detection is lower, and the real-time and the efficiency that detect the highest simultaneously.

Description

A kind of abnormal sound in public places detection method based on scene modeling

Technical field

The present invention relates to a kind of sound signal processing technology, be specifically related to abnormal sound in public places detection method.

Background technology

Prevent one of the main target threatening occurred events of public safety to be public place safety precaution.Due to public peace Total event is frequently accompanied by the various abnormal sounds such as explosive sound, shriek, shot, glass breaking sound, therefore detects and knows Abnormal sound in other acoustical signal has important function to public place safety intelligence monitoring and controlling.

The purpose of abnormal sound in public places detection is to detect to produce when occurred events of public safety occurs from acoustical signal Explosive sound, shriek, shot, the abnormal sound fragment such as glass breaking sound, be two classification problems.At present, that commonly uses is different Often sound detection method is for isolate acoustical signal fragment first with signal end detection technique from input audio signal, then Extracting the feature of this fragment, feature is classified by last Land use models sorting technique, it is achieved the detection of abnormal sound.This side The deficiency that method exists is mainly manifested in: (1) disaggregated model is that training in advance is good, the scene sample sound used during training pattern Limited amount, causes the model possibly cannot accurate description scene changes；(2) during train classification models, scene sample sound quantity is remote More than abnormal sound sample size, the imbalance of this sample can cause the inaccurate of disaggregated model decision boundary, causes abnormal sound The accuracy of sound detection reduces；(3) utilize end-point detection technology separation acoustical signal fragment before signal segment being classified, need The real-time taking more memory space and process is the highest.

Summary of the invention

In order to solve the problems referred to above, the present invention, from the angle of background modeling, gives a kind of based on scene modeling Abnormal sound in public places detection method.

The abnormal sound in public places detection method that the present invention proposes is the acoustical signal for public place, first to letter Number carry out sub-frame processing, calculate the average amplitude of every frame signal；Then according to the scene voiced frame average width of relative anomalies voiced frame It is worth little, the statistical property that fluctuation range is narrower, set up gauss hybrid models for different scenes, form the corresponding scene of each scene Model；Finally ask for the likelihood score of acoustical signal to be measured and model of place, carry out with model of place by arranging likelihood score threshold value Likelihood score mates, then threshold condition based on most Voting principles and the continuous frame number of minimum judges that whether voiced frame to be measured is as abnormal Sound, it is achieved the detection of abnormal sound.

The present invention program specifically includes scene modeling and abnormal sound two parts of detection.

(1) process of scene modeling is:

(1.1) the scene sample sound in training data and abnormal sound sample are carried out sub-frame processing, calculate every frame sound The average amplitude of tone signal.

(1.2) gauss hybrid models of sign scene is set up.First, putting down frame scene sample sound every in training data All amplitude is as the input of modeling, utilizes expectation maximization iterative algorithm to be trained, determines gauss hybrid models parameter, generates Adapt to the gauss hybrid models of this scene；Then, scene sample sound and the exception of known category information in training data are utilized The model of place set up is estimated by sample sound, based on etc. error rate principle determine the likelihood score threshold value of model of place.

In order to make abnormal sound detection that the scene of all kinds of public places is had adaptability, different scenes is divided by the present invention Do not set up independent Gaussian Mixture model of place；When carrying out abnormal sound detection, the different choice according to application scenarios is different Model of place.

(2) process that acoustical signal to be measured carries out abnormal sound detection comprises the following steps:

(2.1) acoustical signal to be measured of input is carried out sub-frame processing, calculate the average amplitude of every frame acoustical signal.

(2.2) according to the model of place that the different choice of application scenarios is different, every frame acoustical signal and model of place are calculated Likelihood score.

The computational methods of likelihood score are: on the basis of built vertical model of place, by the average amplitude of every frame acoustical signal As the input of its probability density function, calculate this frame acoustical signal and the likelihood score of each single Gauss model in model of place, And further according to its likelihood score with whole model of place of the weight computing of each single Gauss model.

(2.3) every frame acoustical signal is carried out likelihood score coupling, it is achieved abnormal sound detects.

The method of likelihood score coupling: first, by the comparison of current frame signal likelihood score Yu model of place likelihood score threshold value, Preliminary judgement present frame acoustical signal is the probability of abnormal sound signal, the most tentatively sentences when likelihood score is less than likelihood score threshold value Determining current frame signal is abnormal sound signal, is otherwise scene acoustical signal；Then, in conjunction with former frame signal, current frame signal With the preliminary judgement result of a later frame signal, most Voting principle is used to determine the type of current frame signal.Finally, by different The often statistical analysis of sound clip length, arranges minimum frame number threshold condition continuously, filters part based on this threshold value and is similar to different The shorter scene sound clip of Chang Shengyin, detects the abnormal sound signal in acoustical signal to be measured.

It is an advantage of the current invention that: different scenes is set up different models of place by (1), overcome existing method scene The shortcoming of bad adaptability；(2) only scene is modeled, can avoid owing to scene sample sound is uneven with abnormal sound sample The inaccurate problem of detection weighed and cause；(3) need not to utilize end-point detection to obtain sound event, the real-time of detection and effect Rate is higher, takies memory space less.

Accompanying drawing explanation

The entire block diagram of Fig. 1 the present invention program；

Scene modeling figure in Fig. 2 the present invention program；

Abnormal sound detection procedure chart in Fig. 3 the present invention program.

Detailed description of the invention

Below in conjunction with the accompanying drawings, the detailed description of the invention of the present invention is further elaborated.

Fig. 1 is the entire block diagram of the present invention program, is specifically related to a kind of abnormal sound in public places based on scene modeling Detection method.The method, for the acoustical signal of public place, first carries out pre-place to scene tone signal and acoustical signal to be measured Reason, i.e. carries out sub-frame processing, and calculates the average amplitude of every frame signal signal；Then according to scene voiced frame relative anomalies sound Sound frame average amplitude is little, the statistical property that fluctuation range is narrower, by EM Algorithm for Training scene tone signal, sets up and is adapted to this scene Gauss hybrid models, form the model of place of this scene；Finally calculate the likelihood score of acoustical signal to be measured and model of place, and Carry out likelihood score coupling and draw testing result, the most first the likelihood score threshold value of gained likelihood score with model of place is compared, then Threshold conditions based on most Voting principles and the continuous frame number of minimum judge that voiced frame to be measured is whether as abnormal sound, it is achieved exception The detection of sound.

The present invention program use acoustical signal be sample frequency be 16kHz, sampling resolution is the public place sound of 16bit Sound, wherein comprises multiple scene sound of public place and multiple accidental abnormal sound the most that may be present.Public place Scene sound have a very wide distribution due to it, randomness is big, and its acoustical signal Normal Distribution, feature is that energy is less, energy Amount fluctuation range is narrower.And the energy of abnormal sound is bigger than scene sound, and distribution compares concentration, by public place Scene sound sets up model, is mated with model of place by acoustical signal to be measured, it is possible to according to abnormal sound and scene sound Difference test go out abnormal sound accidental in public place.

The present invention program mainly includes scene modeling and abnormal sound two parts of detection.

Fig. 2 is the scene modeling figure in the present invention program, concretely comprises the following steps:

(1) the scene sample sound in training data and abnormal sound sample are carried out sub-frame processing, calculate every frame sound The average amplitude of signal.

(2) gauss hybrid models of sign scene is set up.

In step (1), the frame length to acoustical signal framing is 8ms, and it is 8ms that frame moves, the most overlapping between adjacent two frames.

For the i-th frame signal x in acoustical signal_iM the computing formula of () average amplitude is:

Wherein x_iRepresenting the average amplitude of the i-th frame signal, N is the data point number in a frame acoustical signal.

In step (2), for certain scene of public place, scene sound is utilized to set up the Gaussian Mixture mould characterizing scene The process of type is: using the average amplitude of frame scene sample sound every in training data as the input of modeling, utilize expectation maximum Change iterative algorithm to be trained, determine gauss hybrid models parameter, thus generate the gauss hybrid models adapting to this scene.

In the model, if the number of single Gauss model is M, the average of each single Gauss model, variance matrix are respectively μ_j,C_j, j=1,2 ... M, input data are the average amplitude x of a frame scene sound_i, then this gauss hybrid models is:

$p\left(x_i\right)=\underset{j=1}{\overset{M}{\Σ}}{\mathrm{\α}}_j{N}_j(x_i;{\mathrm{\μ}}_j,C_j)$

Wherein, α_jFor the weight coefficient shared by each Gaussian component, meet:

$\underset{j=1}{\overset{M}{\Σ}}{\mathrm{\α}}_j=1$

N_j(x_i；μ_j,C_j) it is the probability density function of each Gaussian component, represent that a frame signal is with each single Gauss seemingly So degree, is defined as:

$N_j(x_i;{\mathrm{\μ}}_j,C_j)=\frac{1}{\sqrt{{\left(2\mathrm{\π}\right)}^n\left|C_j\right|}}\exp \[-\frac{1}{2}{(x_i-{\mathrm{\μ}}_j)}^T{C_j}^{-1}(x_i-{\mathrm{\μ}}_j)\]$

In order to determine this gauss hybrid models, i.e. to determine the weight coefficient of each single Gauss, average, variance matrix, this Invention uses expectation maximization training algorithm, by calculating the maximum of the log-likelihood function of gauss hybrid models, and can be really Determine above-mentioned parameter.

The log-likelihood function of gauss hybrid models is:

$p\left(x_i\right)=\underset{i=1}{\overset{n}{\Σ}}log\left\{\underset{j=1}{\overset{M}{\Σ}}{\mathrm{\α}}_j{N}_j(x_i;{\mathrm{\μ}}_j,C_j)\right\}$

The flow process of expectation maximization (EM) algorithm is:

1) initialize

By k means clustering algorithm, sample is clustered, initializes one group of parameter, utilize each Gaussian component average and Variance matrix is as μ_j0, C_j0, α_j0For the weight coefficient shared by each Gaussian component；

2) estimating step

In gauss hybrid models, estimate the probability that input sample data is generated by each single Gauss model, i.e. input sample Posterior probability be:

${\mathrm{\β}}_{ij}=\frac{{\mathrm{\α}}_j{N}_j(x_i;\mathrm{\Φ})}{\underset{k=1}{\overset{M}{\Σ}}{\mathrm{\α}}_j{N}_k(x_i;\mathrm{\Φ})},1\≤i\≤n,1\≤j\≤M$

Wherein, β_ijBelong to the posterior probability of jth Gaussian component for input sample, n is number of samples, and M divides for single Gauss Amount number, N is the probability density function of each Gaussian component, and α is the weight coefficient shared by each Gaussian component, and Φ represents Gauss Parameters in model.

3) maximization steps

The log-likelihood function maximizing gauss hybrid models updates weights, average and variance matrix, and formula is as follows:

Renewal weights:

${\mathrm{\α}}_j=\frac{\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\β}}_{ij}}{N}$

Wherein, β_ijBelong to the posterior probability of jth Gaussian component for input sample, N is that the probability of each Gaussian component is close Degree function.

Renewal average:

${\mathrm{\μ}}_j=\frac{\underset{i=1}{\overset{N}{\Σ}}{x}_i{\mathrm{\β}}_{ij}}{\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\β}}_{ij}}$

Wherein, x_iFor input sample, β_ijThe posterior probability of jth Gaussian component is belonged to for input sample.

Renewal variance matrix:

$c_j=\frac{\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\β}}_{ij}(x_i-{\mathrm{\μ}}_j^T){(x_i-{\mathrm{\μ}}_j^T)}^T}{\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\β}}_{ij}}$

Wherein, β_ijThe posterior probability of jth Gaussian component, x is belonged to for input sample_iFor input sample, μ_jFor each high The average of this component.

4) condition of convergence is determined

Constantly iteration E and M step, repeats to update three values above, until meeting condition:

| P (X | Φ)-P (X | Φ) ' | ＜ ε

The results change that before and after i.e., twice iteration obtains is less than the most then terminating iteration, and wherein P exists for input sample Likelihood score in gauss hybrid models, Φ represents the parameters in Gauss model, usual ε ＜ 10^-5。

Determined the parameter of each Gaussian component of gauss hybrid models by above-mentioned expectation maximization iterative algorithm, thus complete The gauss hybrid models of this scene is set up.

Then model of place arranges likelihood score threshold value, and it is different that this threshold value is used for the type of this frame acoustical signal of preliminary judgement Chang Shengyin or scene sound.When meeting threshold value, this frame acoustical signal of preliminary judgement is abnormal sound signal, represents with 1, otherwise For scene acoustical signal, represent with 0.Likelihood score Threshold is: utilize the scene of known category information in training data The model of place set up is estimated by sample sound and abnormal sound sample, based on etc. error rate principle determine model of place Likelihood score threshold value.

What said process completed model of place corresponding to some scene in public place sets up process, different in order to make Often sound detection has adaptability to the scene of all kinds of public places, and the different scenes of public place are set up solely by the present invention respectively Vertical Gaussian Mixture model of place；When carrying out abnormal sound detection, according to the scene mould that the different choice of application scenarios is different Type.

Fig. 3 is the abnormal sound detection procedure chart in the present invention program, mainly includes following step:

(1) acoustical signal to be measured is carried out framing pretreatment, calculate the average amplitude of every frame acoustical signal；

(2) likelihood score of every frame signal and model of place is calculated；

(3) every frame acoustical signal is carried out likelihood score coupling, it is achieved abnormal sound detects.

In step (1), it is 8ms that the present invention arranges the frame length of a frame acoustical signal to be measured, and frame moves as 8ms, adjacent two frames Between the most overlapping, average amplitude computational methods are identical with the method during Fig. 2 scene modeling.

In step (2), for kth frame acoustical signal x_kM (), if its average amplitude is x_k, calculate one by below equation In frame acoustical signal and model of place each single Gauss model likelihood score and with the likelihood score of this model of place:

${\mathrm{Pix}}_j(x_k;{\mathrm{\μ}}_j,C_j)=\frac{1}{\sqrt{{\left(2\mathrm{\π}\right)}^n\left|C_j\right|}}\exp \[-\frac{1}{2}{(x_k-{\mathrm{\μ}}_j)}^T{C_j}^{-1}(x_k-{\mathrm{\μ}}_j)\]$

$P=\underset{j=1}{\overset{M}{\Σ}}{\mathrm{\α}}_j{\mathrm{Pix}}_j(x_k;{\mathrm{\μ}}_j,C_j)$

Wherein, α_j,μ_j,C_jFor weights, average and the covariance matrix of single Gauss model each in model of place, M is scene The number of single Gauss model, Pix in model_jFor this frame acoustical signal and the likelihood score of each single Gauss model, P in model of place Likelihood score for this frame acoustical signal Yu this model of place.

In step (3), by the comparison of current frame signal likelihood score Yu model of place likelihood score threshold value, preliminary judgement is worked as Front frame acoustical signal is the probability of abnormal sound signal, and when likelihood score is less than likelihood score threshold value, then preliminary judgement present frame is believed Number it is abnormal sound signal, represents with 1, be otherwise scene acoustical signal, represent with 0.

In step (3), on the basis of likelihood score threshold value preliminary judgement every frame signal type, more continuous frame number is used to throw Ticket principle determines that current frame signal is scene sound or abnormal sound further.The decision method of frame number Voting principle continuously For: set former frame signal, present frame acoustical signal, a later frame signal starting type be respectively(It is 0 or 1), Last type T of present frame is judged according to below equation_k:

$T_k=\left\{\begin{array}{cc}1,& \begin{array}{cc}if& sum({\stackrel{\‾}{T}}_{k-1}+{\stackrel{\‾}{T}}_k+{\stackrel{\‾}{T}}_{k+1})\end{array}\≥2\\ 0,& else\end{array}\right.$

Work as T_kWhen being 1, it is determined that this frame signal is abnormal sound, it is when 0, to be judged to scene sound.

In step (3), after determining the type of each frame signal, by the statistical analysis to abnormal sound fragment length, Minimum successive frame number sieve is set and selects threshold value, filter part based on this threshold value and be similar to the shorter scene sound clip of abnormal sound, Detect abnormal sound complete in acoustical signal to be measured.

It is an advantage of the current invention that: different scenes is set up different models of place by (1), overcome existing method scene The shortcoming of bad adaptability；(2) only scene is modeled, can avoid owing to scene sample sound is uneven with abnormal sound sample The inaccurate problem of detection weighed and cause；(3) need not to utilize end-point detection to obtain sound event, the real-time of detection and effect Rate is higher, takies memory space less.

In order to verify the performance of abnormal sound detection method of the present invention, in family, office, ATM, bank, shop Tested respectively Deng under 5 scenes.First, when choosing under each scene, the scene voice data of a length of a hour is trained also Set up the gauss hybrid models of this scene；Then, the time scene of a length of a hour in addition to training data is chosen under each scene Sound is as test data, and in test data, 200 shots of random superposition, explosive sound, glass breaking sound and shriek etc. are different Chang Shengyin, forms testing data.By this testing data is tested, calculate abnormal sound detection method of the present invention False drop rate and loss.Table 1 show the experimental result of the inventive method.

Table 1 the present invention program experimental result

Scene type	Loss	False drop rate
			Family	0.50%	0
Office	0.50%	0
			ATM	8.00%	9.50%
Bank	5.00%	5.00%
			Shop	6.50%	6.50%

Test result indicate that: loss and false drop rate that abnormal sound is detected under different scenes by the inventive method are the lowest In 10%, wherein the scene such as family, office has lower loss and false drop rate relative to the scene such as bank, shop.With Existing abnormal sound detection method is compared, and abnormal sound in public places detection is had by abnormal sound detection method of the present invention Lower error rate and more preferable scene adaptability, it is possible to adapt to the application of different scene.

Claims (7)

1. an abnormal sound in public places detection method based on scene modeling, it is characterised in that: first to public place Acoustical signal carries out sub-frame processing, calculates the average amplitude of every frame signal；Then according to scene voiced frame relative anomalies voiced frame Average amplitude is little, the statistical property that fluctuation range is narrower, by EM Algorithm for Training scene tone signal, sets up the height being adapted to this scene This mixed model, forms the model of place of this scene；Finally calculate the likelihood score of acoustical signal to be measured and model of place, and carry out Likelihood score coupling draws testing result, the most first the likelihood score threshold value of gained likelihood score with model of place is compared, then based on The threshold condition of most Voting principles and the continuous frame number of minimum judges that voiced frame to be measured is whether as abnormal sound, it is achieved abnormal sound Detection.

2. abnormal sound in public places detection method based on scene modeling as claimed in claim 1, it is characterised in that described side Method specifically include scene modeling and abnormal sound detection two parts:

(1) process of scene modeling is:

(1.1) the scene sample sound in training data and abnormal sound sample are carried out sub-frame processing, calculate every frame sound letter Number average amplitude；

(1.2) using the average amplitude of every frame acoustical signal as input data, the Gauss characterizing scene is set up for different scenes Mixed model, and utilize scene sample sound and the abnormal sound sample scene to setting up of known category information in training data Model is estimated, based on etc. error rate principle determine the likelihood score threshold value of model of place；

(2) step of abnormal sound detection is:

(2.1) acoustical signal to be measured is carried out framing pretreatment, calculate the average amplitude of every frame acoustical signal；

(2.2) according to the model of place that the different choice of application scenarios is different, the likelihood score of every frame signal and model of place is calculated；

(2.3) every frame acoustical signal is carried out likelihood score coupling, then based on most Voting principles and the threshold value of the continuous frame number of minimum Whether condition criterion voiced frame to be measured is abnormal sound, it is achieved abnormal sound detects.

3. abnormal sound in public places detection method based on scene modeling as claimed in claim 1 or 2, it is characterised in that seemingly So the computational methods of degree are: on the basis of built vertical model of place, using the average amplitude of every frame acoustical signal as its probability The input of density function, calculates this frame acoustical signal and the likelihood score of each single Gauss model in model of place, and root further According to its likelihood score with whole model of place of weight computing of each single Gauss model, computing formula is as follows:

${\mathrm{Pix}}_j(x_k;{\mathrm{\μ}}_j,C_j)=\frac{1}{\sqrt{{\left(2\mathrm{\π}\right)}^n\left|C_j\right|}}\exp \[-\frac{1}{2}{(x_k-{\mathrm{\μ}}_j)}^T{C_j}^{-1}(x_k-{\mathrm{\μ}}_j)\]$

$P=\underset{j=1}{\overset{M}{\Σ}}{\mathrm{\α}}_j{\mathrm{Pix}}_j(x_k;{\mathrm{\μ}}_j,C_j)$

Wherein, α_j,μ_j,C_jFor weights, average and the covariance matrix of single Gauss model each in model of place, M is model of place The number of middle single Gauss model, Pix_jFor this frame acoustical signal and the likelihood score of each single Gauss model in model of place, P is for being somebody's turn to do Frame acoustical signal and the likelihood score of this model of place.x_kKth frame acoustical signal x_kThe average amplitude of (m).

4. abnormal sound in public places detection method based on scene modeling as claimed in claim 1 or 2, it is characterised in that enter The process of row likelihood score coupling is: first, by the comparison of current frame signal likelihood score Yu model of place likelihood score threshold value, tentatively Judging the present frame acoustical signal probability as abnormal sound signal, when likelihood score is less than likelihood score threshold value, then preliminary judgement is worked as Front frame signal is abnormal sound signal, is otherwise scene acoustical signal；Then, in conjunction with former frame signal, current frame signal and after The type preliminary judgement result of one frame signal, uses most Voting principle to determine the type of current frame signal；Finally, by different The often statistical analysis of sound clip length, arranges minimum frame number threshold condition continuously, filters part based on this threshold value and is similar to different The shorter scene sound clip of Chang Shengyin, detects the abnormal sound signal in acoustical signal to be measured.

5. abnormal sound in public places detection method based on scene modeling as claimed in claim 2, it is characterised in that described company The decision method of continuous frame number Voting principle is: set former frame signal, present frame acoustical signal, the starting type of a later frame signal divide It is notWhereinIt is 0 or 1, judges last type T of present frame according to below equation_k:

$T_k=\left\{\begin{array}{c}1,ifsum({\stackrel{\‾}{T}}_{k-1}+{\stackrel{\‾}{T}}_k+{\stackrel{\‾}{T}}_{k+1})\≥2\\ 0,else\end{array}\right.$

Work as T_kWhen being 1, it is determined that this frame signal is abnormal sound, it is when 0, to be judged to scene sound.

6. abnormal sound in public places detection method based on scene modeling as claimed in claim 1 or 2, it is characterised in that right In certain scene of public place, the process utilizing scene sound to set up the gauss hybrid models characterizing scene is: will train number According to, the average amplitude of every frame scene sample sound is as the input of modeling, utilizes expectation maximization iterative algorithm to be trained, Determine gauss hybrid models parameter, thus generate the gauss hybrid models adapting to this scene.

7. abnormal sound in public places detection method based on scene modeling as claimed in claim 1 or 2, it is characterised in that institute Multiple scene sound that the acoustical signal stating public field comprises public place and multiple accidental abnormal sound the most that may be present Sound.