CN102034288B - Multiple biological characteristic identification-based intelligent door control system - Google Patents

Abstract

The invention relates to a multiple biological characteristic identification-based intelligent door control system, which comprises a voice acquisition device and an image acquisition device, wherein a door controller receives a voice signal and a human face image signal acquired by the image acquisition device respectively; the output end of the door controller is connected with a data processor and an electric control door lock respectively; the door controller transmits the received voice signal and the human face image signal into the data processor; the data processor extracts characteristics from the received voice signal and the human face image signal respectively, performs identification according to the acquired voice signal and the human face image signal and performs normalization and classification fusion treatment on the corresponding signals after the characteristics are extracted; the data processor compares the corresponding signals obtained after the classification fusion treatment with a preset sample base; and when the signals obtained after the classification fusion treatment are matched with the corresponding signals in the sample base, the door controller opens the electric control door lock. The multiple biological characteristic identification-based intelligent door control system improves the reliability of identification, has high accuracy and is convenient in identification, safe and reliable.

Description

Intelligent access control system based on multi-biological characteristic identification

Technical field

The present invention relates to a kind of gate control system, especially a kind of intelligent access control system based on multi-biological characteristic identification belongs to the technical field of gate control system.

Background technology

At present, the identity discriminating means of traditional gate control system comprise password, password, certificate or the like.Because separability, be prone to cause forgery, usurp, phenomenon such as decoding with being differentiated the people.And that the biological attribute of human body is people's a body is exclusive, and some biological attribute of people such as fingerprint, iris, sound etc. is unique, and it is applied in the generation that gate control system can be stopped phenomenons such as forging, usurp.Biometrics identification technology mainly is divided into two big types: one type is physiological characteristic identification, is respectively to utilize fingerprint, palm type, and iris, characteristics such as retina and people's face are discerned; Another kind of is behavioural characteristic identification, comprises signature and speech recognition.Adopt the biological characteristic authentication technology of single mode to receive the restriction of interference of noise and environment for use easily, and forge counterfeitly easily, be difficult to guarantee the accuracy of gate control system identification.

Summary of the invention

The objective of the invention is to overcome the deficiency that exists in the prior art, a kind of intelligent access control system based on multi-biological characteristic identification is provided, it has improved the reliability of identification, and accuracy rate is high, and convenient identification is safe and reliable.

According to technical scheme provided by the invention, said intelligent access control system based on multi-biological characteristic identification comprises sound collection equipment and image capture device; Said sound collection equipment and image capture device all link to each other with the input of access controller, and said access controller receives the voice signal of sound collection equipment collection and the facial image signal that image capture device is gathered respectively; The output terminal of access controller links to each other with data processor and electric control door lock respectively; Access controller is transferred to the voice signal and the facial image signal that receive in the data processor; Said data processor carries out feature extraction to voice signal and the facial image signal that receives respectively, voice signal and the facial image signal gathered is discerned, and corresponding signal after the feature extraction is carried out normalization and classification fusion treatment; The data processor corresponding signal of will classifying after the fusion treatment is compared with the sample storehouse of presetting; When corresponding signal mated in signal that obtains after the classification fusion treatment and the sample storehouse, access controller was opened electric control door lock.

Said data processor comprises DSP.Said data processor utilizes PCA and linear discriminant analysis method that the facial image signal is carried out feature extraction and recognition of face.

Said data processor utilizes Mei Er cepstrum coefficient and mixed Gauss model method that voice signal is carried out feature extraction and recognition of face.Voice signal and facial image signal after said data processor utilizes the z-score function to identification carry out normalization.

Voice signal after said data processor utilizes SVMs to normalization and the facial image signal fusion treatment of classifying.The output terminal of said access controller also links to each other with audible-visual annunciator; When the sample storehouse did not match in signal that obtains after the classification fusion treatment and the data processor, access controller was through audible-visual annunciator output sound and light alarm signal.

Advantage of the present invention: access controller receives voice signal, the facial image signal of sound collection equipment and image capture device simultaneously; Data processor utilizes Mei Er cepstrum coefficient and mixed Gauss model method to carry out feature extraction and voice recognition to voice signal; Data processor utilizes PCA and linear discriminant analysis method to carry out feature extraction and sound device to the facial image signal; Data processor is to the voice signal after extracting and discerning and the facial image signal carries out Z-score normalization and the SVMs classification is merged; Reduce the wrong rate such as grade of identification, improved the safe reliability of gate control system, safe and reliable.

Description of drawings

Fig. 1 is a structured flowchart of the present invention.

Fig. 2 is a workflow diagram of the present invention.

Embodiment

Below in conjunction with concrete accompanying drawing and embodiment the present invention is described further.

As shown in Figure 1: as to the present invention includes image capture device, sound collection equipment, electric control door lock, access controller, data processor and audible-visual annunciator.

The multi-biological characteristic fusion is meant and utilizes the biological characteristic different characteristic, carries out the fusion of certain aspect, its objective is the limitation that overcomes or evade single characteristic; The present invention is through two biological characteristics to facial image and voice signal multidigit gate inhibition.Merge the influence that authentication has reduced unfavorable factor through combining different biological features information, for the deficiency of the living things feature recognition that solves single mode has been brought effective solution.Facial image has relative uniqueness and stability with acoustic information, and gathers conveniently, has untouchable and non-infringement.Being suitable for gate control system uses as diagnostic characteristics.

As shown in Figure 1: said image capture device links to each other with the input end of access controller with the output terminal of sound collection equipment, and the color collecting device of said image adopts camera, as the collection apparatus instrument, gathers visitor's facial image; Sound collection equipment adopts microphone, as the collection apparatus instrument, gathers visitor's voice signal.The output terminal of access controller links to each other with data processor; Said data processor comprises DSP (digital signal processor); Said data processor can carry out recognition of face and voice recognition, stores relevant people's face information and acoustic information in the said data processor in advance.Access controller receives the facial image signal of image capture device collection and the voice signal that sound collection equipment is gathered, and above-mentioned signal is transferred in the data processor.Said data processor utilizes Mei Er spectrum scramble coefficient (MFCC) and mixed Gauss model (GMM) to carry out feature extraction and voice recognition to the voice signal that receives; Data processor utilizes PCA (PCA) and linear discriminant analysis method (LDA) to carry out feature extraction and recognition of face to the facial image signal that receives.In order to reduce classification wrong rate such as grade relatively; Data processor through the Z-score function to voice signal and facial image signal after the identification being carried out the normalization processing; Elimination utilizes the result difference of different biological features identification, reduces owing to the different errors that cause of classification; Data processor utilizes SVMs (SVM) method that above-mentioned normalization result is classified after carrying out the normalization processing again, reaches decision-making level's fusion treatment.Compare between the sample storehouse that prestores in voice signal and the facial image signal and the data processor of data processor after above-mentioned fusion treatment; When corresponding signal is complementary in the voice signal of gathering and facial image signal and the sample storehouse; Data processor sends the instruction of opening the door to access controller, and access controller is opened electric control door lock; When the voice signal of gathering and facial image signal and sample storehouse did not match, access controller sent the sound and light alarm signal through audible-visual annunciator, guarantees the safety of gate control system.

Data processor utilizes PCA (PCA) and linear discriminant analysis method (LDA) to carry out feature extraction and recognition of face to the facial image signal; Wherein, Principal component analysis (PCA) (PCA) is a kind of statistical method; It is by means of an orthogonal transformation, and the former random vector that component is relevant changes into the incoherent new random vector of component, and the covariance matrix that on algebraically, shows as former random vector is transformed into the diagonal form battle array; On how much, show as the orthogonal coordinate system with former coordinate system transformation Cheng Xin, sample point scatters and opens most.Linear discriminant analysis (LDA) has great influence at area of pattern recognition, and the linear discriminant analysis method can be applicable to the image dimensionality reduction, and it is a kind of method of carrying out dimensionality reduction based on the classification of sample.Its projection matrix is through distributing between the maximization class, and interior distribution of type of minimizing obtains simultaneously.

The linear discriminant analysis face identification method is described below:

Suppose total N image in the original image storehouse, suppose that each image has d pixel, then primitive man's face image vector can be expressed as X ₁, X ₂, Λ X _N(vectorial dimension is made as d), wherein N ₁Individual facial image belongs to 1 type, N ₂Individual facial image belongs to 2 types, N _CIndividual facial image belongs to C class (N wherein ₁+ N ₂+ ... + N _C=N).Then the average of all kinds of facial images is:

${\mathrm{\μ}}_i=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{X}_i,(i=\mathrm{1,2},L,C)---\left(1\right)$

The average of total facial image is:

$\mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{X}_i---\left(2\right)$

Dispersion matrix S between the sample class _BWith within class scatter matrix S _wBe defined as:

$S_B=\underset{i=1}{\overset{C}{\mathrm{\Σ}}}{N}_i({\mathrm{\μ}}_i-\mathrm{\μ}){({\mathrm{\μ}}_i-\mathrm{\μ})}^T---\left(3\right)$

$S_W=\underset{i=1}{\overset{C}{\mathrm{\Σ}}}\underset{X_k\∈X_i}{\mathrm{\Σ}}(X_k-{\mathrm{\μ}}_i){(X_k-{\mathrm{\μ}}_i)}^T---\left(4\right)$

If S _w(within class scatter matrix) is nonsingular, then will type of acquisition between dispersion with type in the W of the maximum projecting direction of the ratio of dispersion _OptSatisfy following formula:

$W_{\mathrm{opt}}=\arg \max \frac{\left|W^T{S}_{B}W\right|}{\left|W^T{S}_{W}W\right|}=[w_1,w_2,K,w_m]---\left(5\right)$

{ w wherein _i| i=1,2, L, m} are the S that satisfies following formula _BAnd S _wM corresponding eigenvalue of maximum { λ _i| i=1,2, L, the pairing proper vector of m}: S _BW _i=λ _iS _WW _i(i=1,2, L, m).Notice that this matrix has only C-1 nonzero eigenvalue at most, C is the classification number.

Utilize the linear projection operator, each width of cloth facial image can shine upon the proper vector that obtains a low dimension.This vectorial element is done inner product operation respectively through each column vector of image vector and projection operator and is obtained.

The PCA method still is that the LDA method all needs a large amount of experiment samples, and also can not satisfy the requirement of sample size in the real life fully, and in order to solve the problem that small sample possibly exist, data processor adopts the method for PCA+LDA to carry out recognition of face.

If to p ⁱ _j(j=1,2 ..., S; I=1,2 ..., the vectorial characteristic projection of j people's face of i class people face that K) obtains for PCA, S is every type a sample number, K is the training sample sum.

At first calculate the average μ of all kinds of samples _iWith total sample average μ, from the corresponding class average of the figure image subtraction of each sample, promptly all kinds of training sample centralizations then, deduct total sample average and obtain μ from all kinds of averages _i, form data matrix to the training sample image of all centralizations, and seek orthogonal basis for this data matrix through the PCA method.If the orthogonal basis of obtaining is U, with the image projection of all centralizations to orthogonal basis.

Wherein,

identifies sample characteristics; Project to the average of all centralizations on the orthogonal basis, accomplished the process of PCA:

At last in the following formula with the parameter substitution LDA that tries to achieve:

Collecting and distributing degree matrix S between type of finding the solution _BWith within class scatter matrix S _WWherein:

$S_W=\underset{I=1}{\overset{C}{\mathrm{\Σ}}}{S}_I---\left(9\right)$

Calculate generalized eigenvalue Λ and characteristic of correspondence vector V:

S _BV＝λS _WV (11)

According to the descending series arrangement proper vector of character pair value, K-1 proper vector before only keeping, Here it is Fisher base vector.Project to the original image that rotates through on the Fisher base vector, project to original image on the orthogonal basis U earlier in other words, continue the projected image that obtains to project on the Fisher base vector W again.Carry out Classification and Identification work then, calculate the discrimination of recognition of face.

Mei Er cepstrum coefficient (Mel-Frequency Cepstrum Coefficient; MFCC) be a simulation to people's ear filter function; Being linear growth in low frequency part, being the growth of index at HFS, is the characteristic of effective a kind of Speaker Identification.At present, the Application on Voiceprint Recognition for text-independent mainly adopts the MFCC characteristic.Mei Er cepstrum coefficient characteristic parameter extraction process is that pre-service divides frame, FFT (Fast Fourier Transform (FFT)), Mei Er wave filter, takes the logarithm and DCT (discrete cosine transform) conversion process.

Probability density function in the statistical model is that a speaker's characteristic distribution probability density function just can become this speaker's template to the complete description of speaker characteristic in this distribution of feature space.

GMM (mixed Gauss model) is a probability statistics model, distinguishes the speaker through the descriptive statistics that the target speaker characteristic is distributed, and its statistical parameter can be represented words person's characteristic information effectively, therefore can utilize GMM to do the mapping of feature space.It has and text-independent, advantage such as processing speed is fast, recognition effect is good, successfully apply to the Speaker Identification of text-independent and affirmation in.

GMM is the weighted sum of M member's Gaussian probability density, is expressed as:

$p\left(x\right|\mathrm{\λ})=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{a}_i{b}_i\left(x\right)---\left(12\right)$

Wherein, x is a D dimension random vector; b _i(x) (i=1,2 ..., M) be each member's Gaussian probability-density function; a _i(i=1,2 ..., M) be mixed weight-value.

Complete GMM can be expressed as:

λ _i＝{a _i，μ _i，∑ _i}，(i＝1，2，L，M) (13)

Each member's density function is the gauss of distribution function of a D dimension variable, is expressed as:

$b_i\left(x\right)=\frac{1}{{\left(2\mathrm{\π}\right)}^{D/2}{\left|\mathrm{\Σ}\right|}^{1/2}}\exp [-\frac{1}{2}{(x-{\mathrm{\μ}}_i)}^T{\mathrm{\Σ}}^{-1}(x-{\mathrm{\μ}}_i)]---\left(14\right)$

For the length tested speech time series X=(x that is T ₁, x ₂..., x _T), its GMM likelihood probability can be expressed as:

$P\left(X\right|{\mathrm{\λ}}_i)=\underset{t=1}{\overset{T}{\mathrm{\Π}}}p\left(x_t\right|{\mathrm{\λ}}_i)---\left(15\right)$

$L\left(X\right|{\mathrm{\λ}}_i)=\lg \left[P\left(X\right|{\mathrm{\λ}}_i)\right]$ (16)

$=\underset{t=1}{\overset{T}{\mathrm{\Σ}}}\lg \left[p\left(x_t\right|{\mathrm{\λ}}_i)\right]$

Utilization Bayes' theorem during identification in N unknown words person's model, obtains the corresponding words person of the maximum model of likelihood probability and is recognition result:

$t^{*}=\underset{1\≤i\≤N}{\arg }\left\{\max \right[L\left(X\right|{\mathrm{\λ}}_i)\left]\right\}---\left(17\right)$

In practical application, the scale of each GMM model is taken as 30-50 Gaussian distribution usually.

In recognition of face and vocal print identifying, each sub-recognition system can provide the similarity that a matching result s (score) comes characterization test sample and matching template.Because the sample data of handling is variant, the decision-making of face identification system and Application on Voiceprint Recognition system can there are differences in form.In order to eliminate of the influence of numerical value form difference to last classification results; At first utilize normalized function being converted into the s ' that same codomain interval can be compared mutually, utilize the stronger svm classifier device of classification capacity that normalized s ' is classified then from the interval s of the different codomains of different sorters with measure.In this article, adopt normalized function to adopt the Z-score function.

The form of the Z-score function that specifically is applied to test:

$s^{\′}=\frac{s-\mathrm{mean}\left(\right\{s\left\}\right)}{\mathrm{std}\left(\right\{s\left\}\right)}---\left(18\right)$

Here s ' is the matching result after the normalization, and mean () and std () represent respectively and average and ask the mean square deviation computing, and { s} is the coupling marking set of a sorter to all test sample books.This method for normalizing is fairly simple, only needs estimation average and variance just can carry out normalization to s.

Adopt SVM to classify to the result after the normalization, made full use of the information of s ' after the normalization; Compare with direct employing svm classifier device, carry out normalization and can the result be adjusted to the same numerical value interval that can compare, make classification results more accurate.Adopt the system performance of normalization SVM amalgamation mode to improve a lot than direct phase add mode, and when the more raisings of number of categories obvious more.This is owing to when number of categories is big more, the performance of each subsystem meeting variation, and the relation between each subsystem also complicates.If in the time of only to the direct addition of decision-making mark of each subsystem, system performance can be along with increasing of number of categories worse and worse.The SVM Fusion Model can be described the nonlinear relationship between each subsystem comparatively fully on the contrary, so performance can improve to some extent.Thereby when number of categories was big more, the advantage that SVM merges was just obvious more.

As shown in Figure 2: during work, input acoustic image checkout equipment continues to detect voice signal or picture signal; When the visitor, sound collection equipment and image capture device start, and gather visitor's voice signal and facial image signal respectively.Access controller outputs to voice signal and facial image signal in the data processor, and data processor utilizes Mei Er cepstrum coefficient and mixed Gauss model to carry out feature extraction and voice recognition to voice signal; Utilize PCA and LDA that the facial image signal is carried out feature extraction and recognition of face simultaneously.After voice signal and the facial image identification of data processor after, carry out normalization through the Z-score function and handle, and normalized data are carried out fusion treatment, thereby form decision-making level's fusion through SVM to identification.Data after decision-making level merges and the sample storehouse in the data processor compare, and when the data in fused data and sample storehouse were complementary, data processor sent matched signal to access controller, and access controller is opened electric control door lock, accomplish the action of opening the door; When the data in fused data and the sample storehouse did not match, data processor enabling counting device compared comparing number of times; When relatively reaching set point number, access controller is through audible-visual annunciator output sound and light alarm signal; When number of comparisons during less than set point number, access controller receives the acquired signal of sound collection equipment and image capture device again.

The present invention is through gathering facial image signal and voice signal, and adopts existing disposal route that facial image and voice signal are carried out feature extraction and identification, can be applied to airport, bank, public security organ, attendance checking system or other aspects.

Access controller of the present invention receives voice signal, the facial image signal of sound collection equipment and image capture device simultaneously; Data processor utilizes Mei Er cepstrum coefficient and mixed Gauss model method to carry out feature extraction and voice recognition to voice signal; Data processor utilizes PCA and linear discriminant analysis method to carry out feature extraction and sound device to the facial image signal; Data processor is to the voice signal after extracting and discerning and the facial image signal carries out Z-score normalization and the SVMs classification is merged; Reduce the wrong rate such as grade of identification, improved the safe reliability of gate control system, safe and reliable.

Claims (1)

1. intelligent access control system based on multi-biological characteristic identification; Comprise sound collection equipment and image capture device; It is characterized in that: said sound collection equipment and image capture device all with the linking to each other of the input of access controller, said access controller receives the facial image signal that voice signal that sound collection equipment gathers and image capture device are gathered respectively; The output terminal of access controller links to each other with data processor and electric control door lock respectively; Access controller is transferred to the voice signal and the facial image signal that receive in the data processor; Said data processor carries out feature extraction to voice signal and the facial image signal that receives respectively, voice signal and the facial image signal gathered is discerned, and corresponding signal after the feature extraction is carried out normalization and classification fusion treatment; The data processor corresponding signal of will classifying after the fusion treatment is compared with the sample storehouse of presetting; When corresponding signal mated in signal that obtains after the classification fusion treatment and the sample storehouse, access controller was opened electric control door lock;

Said data processor comprises DSP;

Said data processor utilizes PCA and linear discriminant analysis method that the facial image signal is carried out feature extraction and recognition of face;

Said data processor utilizes Mei Er cepstrum coefficient and mixed Gauss model method that voice signal is carried out feature extraction and recognition of face;

Voice signal and facial image signal after said data processor utilizes the z-score function to identification carry out normalization;

Voice signal after said data processor utilizes SVMs to normalization and the facial image signal fusion treatment of classifying;

The output terminal of said access controller also links to each other with audible-visual annunciator; When the sample storehouse did not match in signal that obtains after the classification fusion treatment and the data processor, access controller was through audible-visual annunciator output sound and light alarm signal.

Images