CN103208010A - Traffic state quantitative identification method based on visual features - Google Patents

Abstract

The invention belongs to the field of intelligent transportation and machine vision, and discloses a traffic state quantitative identification method based on visual features. The method comprises the following steps of: reading a video from a video acquisition card, and pre-processing each frame of image in the original video; extracting space-time related information from grayed video image frames; adding traffic state category tags for acquired space-time sequence identifiers in a mode of combining objective estimation and subjective judgment; performing dimensionality reduction on the space-time sequence identifiers added with the tags and extracting feature vectors; constructing a classifier by using the extracted feature vectors as the input of a support vector machine (SVM); and quantitatively identifying the traffic state. By adopting the method, each module is optimized, so that accumulative errors of a system are reduced, and the reliability of traffic state quantitative identification data is improved; and dimensionality reduction and feature extraction of a space-time sequence identifier image matrix are realized by adopting a method of principal component analysis (PCA) and Fisher linear discriminant analysis (Fisher LDA), and the SVM is applied to traffic state identification and classification, so that the classification is accurate and effective.

Description

A kind of traffic behavior based on visual signature quantizes recognition methods

Technical field

The invention belongs to intelligent transportation and field of machine vision, relate to a kind of method of utilizing technology such as feature extraction, support vector machine to realize the city expressway traffic behavior is quantized identification.

Background technology

In recent years, congestion in road and accident that the modernization of development of science and technology and means of transportation has caused ubiquity increase phenomenon, and the deterioration of traffic environment becomes problem demanding prompt solution.Intelligent transportation system (Intelligent Transportation System is called for short ITS) just produces and becomes research tendency and the focus in this field under such demand.In every function of ITS, the recognition result of traffic behavior provides basic Information Assurance for work such as the management in later stage and controls, and therefore validity and the degree of accuracy to total system has fundamental influence.

At present, traffic status identification both domestic and external mainly is based on the collection to traffic parameters such as speed, vehicle flowrate and time and space occupation rates, and collector commonly used has velocity radar, infrared ray coil, ultrasound wave and video detection etc.Wherein, have incomparable advantage based on the DETECTION OF TRAFFIC PARAMETERS of video, low as installation cost, easily be automated control, can extract the information of efficiently and accurately more by image processing techniques, thereby in ITS, obtain using widely.But a lot of experiments show that the extraction accuracy rate of traffic parameter sharply descends under the congestion status, thereby influence is to the discriminator of whole traffic behavior and follow-up prediction work.Recently, Chinese scholars has been placed on research emphasis and has utilized on machine vision and mode identification technology classify to traffic behavior, comprises traditional traffic parameter extraction method and directly extract the detection method that feature is carried out traffic behavior in frame of video.But in the classical model of having set up, still exist the too high information that is difficult to realize or extracts of complexity to be not enough to defectives such as effectively classification.

Summary of the invention

At the problems referred to above that exist in the traffic status identification, propose a kind of based on support vector machine (Support Vector Machine, abbreviation SVM) traffic behavior sorting technique, and realize on this basis the quantification of traffic behavior is identified, thereby finish the grasp macroscopical to the whole traffic behavior of road, for intelligent traffic administration system provides decision support effectively accurately.

The technical solution used in the present invention is: extract space-time and describe the sequence symbol and it is carried out two-layer processing from video sequence, respectively by two-way two-dimentional principal component analysis (PCA) (Principal Component Analysis, PCA) image array is carried out the dimension compression and remove redundant information, and by Fisher linear discriminant analysis (Fisher Linear Discriminant Analysis, Fisher LDA) further carries out feature extraction, the proper vector that extracts is imported as SVM, and structure multicategory classification device is divided into 4 classes with sample.In the data identification stage, obtain final traffic behavior quantized value by the distance of calculating the neighbor classified lineoid of the current proper vector distance of also weighting.

A kind of traffic behavior based on visual signature quantizes recognition methods, it is characterized in that may further comprise the steps:

Step 1 reads video from video frequency collection card, each two field picture in the original video is carried out pre-service, converts digital picture to gray level image.

Step 2 is extracted temporal and spatial correlations information in the video frame image behind gray processing, instant ceases to be busy sequence symbol S, and method is as follows:

(1) a virtual spacetime line is set in gray level image, shown in the solid line in accompanying drawing 2 the first half frame of video, the pixel value on the virtual spacetime line in every frame is designated as the vectorial L of row _t=[p _T1, p _T2..., p _Tw] ^T, wherein p is the point on the line segment.

(2) this row vector in t-N+1～t frame video sequence is superposeed, obtain the time ceases to be busy sequence symbol S of current time _t=[L _T-N+1, L _T-N+2..., L _t] ^TEach the time ceases to be busy sequence symbol be the matrix of a W * N, wherein W be set the time ceases to be busy sequence symbol width, N is the selected video frame number that superposes, and is the good parameter of predefined.

(3) the time ceases to be busy sequence symbol in the long enough time that comprises all traffic behaviors is gathered and extracted to the delay-line structure of definition shown in the accompanying drawing 3, thereby obtain ceases to be busy sequence symbol combination U=[S when whole ₁, S ₂..., S _t..., S _M] ^T, M be obtain the time ceases to be busy sequence symbol sample.

Step 3 takes objective estimation to differentiate the mode that combines with subjective, for the time ceases to be busy sequence symbol that has obtained adds the traffic behavior tag along sort.Specifically comprise following content:

(1) " the urban traffic control assessment indicator system " announced in 2002 according to China Ministry of Public Security is divided into unimpeded (more than the 30km/h), slight crowded (20-30km/h), crowded (10-20km/h) and serious crowded (10km/h is following) four classes according to running speed v with traffic behavior.

(2) be calculated as follows under the different speed of a motor vehicle vehicle the time ceases to be busy sequence symbol in shared " time block " ratio:

(c _l·f)/(v·N)

In the formula, f is frame per second, and N is the video frame number that superposes, c _lMean value for the wheelbase of the car that obtains according to the average living standard of the different cities people and official register type of vehicle statistics.

(3) gather and extract time ceases to be busy sequence symbol in the long enough time that comprises all traffic behaviors, in conjunction with the time ceases to be busy sequence symbol (W * dimension N) is rationally set the capacity of sample set, then according to the time block size that calculates in (2), mode by subjective judgement for the time empty descriptor add class label, will (1) in traffic behavior of setting to be labeled as successively be 1,2,3,4 classes.

Step 4 is carried out dimension-reduction treatment to the time ceases to be busy sequence symbol that adds label, extracts proper vector then.

By step 1 as can be known, the time ceases to be busy sequence symbol original matrix dimension D=W * N, known traffic behavior is divided into four class C=4 altogether in step 2.

Linear discriminant analysis, namely the purpose of Fisher LDA makes the sample data that has class label to minimize distance in the class in the maximization between class distance for trying to achieve projection matrix, this projection pattern is conducive to the Classification and Identification of later stage traffic behavior.And derive according to theory, the condition of obtaining optimum projection vector is sample total volume M〉D+C, therefore when dimension is too high, need the sample of larger amt.Simultaneously, the too high dimension of sample data can increase computation complexity.Therefore take the PCA mode that combines with LDA in two steps the sample set data to be handled.

Concrete grammar is as follows:

(1) take two-way two-dimentional principal component analytical method (LR2DPCA) to carry out the data dimensionality reduction.

1. to the time ceases to be busy sequence symbol be listed as compression

At first, definition projection properties vector y ₁For:

y ₁＝Sw ₁

In the formula, w ₁Be projection matrix.

Then, be calculated as follows sample image average Φ:

$\text{\φ=}\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{S}_i$

In the formula, M is total sample number.

This moment image divergence matrix G _tFor:

$G_t=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{(S_i-\mathrm{\φ})}^T(S_i-\mathrm{\φ})$

Choose optimum axis of projection according to minimizing mean-square error criteria, just make w ₁ ^TG _tw ₁Minimize, the solution of this problem is G _tPreceding p eigenvalue of maximum characteristic of correspondence vector.The contribution rate that adds up of p eigenwert before the quantity of p depends on.Finally can get y ₁=Sw ₁=[y ₁, y ₂..., y _p], dimension is W*p.

2. to the time ceases to be busy sequence symbol go compression

Defining the projection properties vector again is: y ₂=w ₂ ^Ty ₁, and recomputate sample image average Φ ' and the image divergence matrix G of this moment by 1. method _t', the solution of this problem is G _t' preceding q eigenvalue of maximum characteristic of correspondence vector.Thereby y ₂=w ₂ ^Ty ₁=[y ₁, y ₂..., y _q], dimension is q*p.

So far, ceases to be busy sequence symbol image array is converted into the time:

x＝y ₂＝w ₂ ^Ty ₁＝w ₂ ^TSw ₁

(2) the time ceases to be busy sequence symbol image array after taking Fisher LDA method to dimensionality reduction carries out further feature extraction.

Definition projection properties vector is:

y ₃＝w ₃ ^Tx

Divergence matrix (within-class scatter) in the compute classes:

$S_w=\underset{j=1}{\overset{C}{\mathrm{\Σ}}}\underset{i=1}{\overset{N_j}{\mathrm{\Σ}}}(x_i^j-u_j){(x_i^j-u_j)}^T$

In the formula,

I width of cloth image in the expression j class, and u _jAnd N _jAverage and the quantity of representing j class image.

Calculate between class scatter matrix (between-class scatter):

$S_b=\underset{j=1}{\overset{C}{\mathrm{\Σ}}}(u_j-u){(u_j-u)}^T$

In the formula, u represents the average of all images in all sample sets.

The thought of LDA is for minimizing distance in the class in the maximization between class distance.Therefore, the projection criterion of taking is the ratio of maximization the former with the latter, that is:

J(w ₃)＝w ₃ ^TS _bw ₃/w ₃ ^TS _ww ₃

This problem is converted into finds the solution S _w ^-1S _bGeneralized eigenvalue problem.Because S _w ^-1S _bOrder be C-1, final w ₃Corresponding to S _w ^-1S _bThe eigenmatrix of maximum C-1 eigenwert.

So far, with the time ceases to be busy sequence symbol image array dimension be down to the C-1=3 dimension by W * N, this 3 dimensional vector can well characterize the space time information in the traffic behavior, and meets the requirement of computation complexity, is convenient to ensuing Classification and Identification.

Step 5 will be extracted the proper vector the obtain input structural classification device as SVM in the step 3.

The sorter number is C (C-1)/2=6.At first classify in the 1st class of difference maximum and the 4th class, next structural classification device in 2 and 4,1 and 3 in the end separates all kinds of in one deck then fully.

Can't specified data whether during linear separability, suppose that it is linear inseparable situation, the solution of SVM is by a Nonlinear Mapping data to be mapped to higher dimensional space, does linear regression then in higher dimensional space.The expression formula of Nonlinear Mapping is:

y(x)＝w ^Tφ(x)+b

In the formula, x represents proper vector, and φ (x) represents mapping function, and b represents the gain in the linear function.

The building method of sorter is as follows in SVM:

Training dataset is N proper vector x ₁..., x _N, desired value is: t ₁..., t _N, t _n∈ 1,1}, and n=1,2 ..., N, and satisfy y (x _n)＞0 o'clock, t _n=1, y (x _n)＜0 o'clock, t _n=-1, t is always arranged _nY (x _n)＞0.

If the classification plane is:

The classificating thought of SVM is the maximization class interval, can push away to such an extent that classification is spaced apart 2/|| ω by geometric knowledge ||, therefore make to make satisfying under the situation of constraint condition || ω || minimize and can obtain the optimal classification face.Main solution procedure is as follows:

Introduce Lagrangian function:

In the formula, a _n〉=0, be Lagrange multiplier.

Respectively to w, b differentiate, and make that derivative is zero:

$w=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{a}_n{t}_n\mathrm{\φ}\left(x_n\right)$

$\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{a}_n{t}_n=0$

Bring the dual form that the Lagrangian function expression formula obtains this function into:

$\tilde{L}\left(a\right)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{a}_n-\frac{1}{2}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}\underset{m=1}{\overset{N}{\mathrm{\Σ}}}{a}_n{a}_m{t}_n{t}_{m}k(x_n,x_m)$

In the formula, and k (x, x')=φ (x) ^Tφ (x') is kernel function, and it can accept the input value of lower dimensional space, exports the inner product value of higher dimensional space then, thereby makes computation process needn't be concerned about concrete mapping relations φ (x), so mapping function can be expressed as:

$y\left(x\right)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{a}_n{t}_{n}k(x,x_n)+b$

According to the KKT(Karush-Kuhn-Tucker that satisfies this formula) condition can obtain the value of b:

$b=\frac{1}{N_s}\underset{n\∈S}{\mathrm{\Σ}}(t_n-\underset{m\∈S}{\mathrm{\Σ}}{a}_m{t}_{m}k(x_n,x_m))$

In the formula, N _sSummation for all support vectors.

In the traffic behavior sorting technique provided by the present invention, adopt Gaussian radial basis function (Gauss RadialBasis Function) in support vector machine, function expression is:

$k(x_i,x)=\exp \{-\frac{{|x-x_i|}^2}{\mathrm{\σ}}\}$

In the formula, x _iBe the kernel function center, σ is the parameter of radial extension.

Step 6 at first obtains tag along sort with vector input SVM to be identified, calculates this vector distance between neighbor classified then, at last according to this apart from quantizing the identification traffic behavior, and draw the traffic behavior quantitation curve.Method is as follows:

(1) vector input SVM to be identified is obtained tag along sort.

(2) calculate this vector distance between neighbor classified.

Each optimal classification face is the curved surface in the space, and its expression formula is: w ^Tφ (x)+b=0.

As shown in Figure 5, the intersection point of postulated point M on curved surface is M ₀, line segment M ₀M is the distance of asking.At curved surface Y an initial point Q nearer apart from the M point is set ₀, obtain the partial derivative r of this both direction on curved surface _uAnd r _wAnd Q ₀And the direction vector S between the M.Then S is decomposed U, r _uAnd r _wThree directions obtain S=dU+ar _u+ br _w, the group of solving an equation:

$\left\{\begin{array}{c}S\&CenterDot;r_u={{\mathrm{<figure-callout\; id="2"\; label="ar\; u"\; filenames="HDA00003084487200022.png"\; state="\{\{state\}\}">ar</figure-callout>}}_u}^2+{\mathrm{br}}_w\&CenterDot;r_u\\ S\&CenterDot;r_w={\mathrm{ar}}_u\&CenterDot;r_w+{{\mathrm{<figure-callout\; id="2"\; label="br\; w"\; filenames="HDA00003084487200022.png"\; state="\{\{state\}\}">br</figure-callout>}}_w}^2\end{array}\right.$

: $a=\mathrm{bk}+c,b=\frac{S\&CenterDot;r_u+{{\mathrm{<figure-callout\; id="2"\; label="cr\; u"\; filenames="HDA00003084487200022.png"\; state="\{\{state\}\}">cr</figure-callout>}}_u}^2}{{{\mathrm{kr}}_{\mathrm{<figure-callout\; id="2"\; label="u"\; filenames="HDA00003084487200022.png"\; state="\{\{state\}\}">u</figure-callout>}}}^2+r_u\&CenterDot;r_w}$

Wherein, $k=\frac{r_u\&CenterDot;r_w+{r_w}^2}{r_u\&CenterDot;r_w+{{\mathrm{<figure-callout\; id="2"\; label="r\; u"\; filenames="HDA00003084487200022.png"\; state="\{\{state\}\}">r</figure-callout>}}_u}^2},c=\frac{-S\&CenterDot;r_u\&CenterDot;r_w-S\&CenterDot;r_w}{r_u\&CenterDot;r_w+{{\mathrm{<figure-callout\; id="2"\; label="r\; u"\; filenames="HDA00003084487200022.png"\; state="\{\{state\}\}">r</figure-callout>}}_u}^2}$

Make u=u-a, w=w-b, iteration, up to | Sr _u|≤ε, | Sr _w|≤ε, wherein, the error of calculation of ε for rule of thumb providing with actual conditions.The Q of this moment _nAsk exactly near M ₀The point.

(3) identify traffic behavior according to this apart from quantification, and draw the traffic behavior quantitation curve.

The 1st, 4 classes are defined as the unimpeded and serious crowded state of traffic conditions respectively in step 2, and generally the degree of crowding of the 2nd, 3 classes can and characterize the substantive help of traffic behavior generation to prediction, so the quantification of traffic behavior identification is primarily aimed at 2,3 classifications.Therefore, defining the 1st class traffic behavior quantized value is that 0, the 4 class traffic behavior quantized value is 1, and the quantized value of the point on the optimal classification face of the 2nd, 3 classes is respectively 1/3,2/3.According to the classification results of SVM in the step 4, if class label is 2, this proper vector to 1,2 classifying faces and 2,3 classifying faces apart from d ₁And d ₂, final quantized value v ₂For:

${\mathrm{<figure-callout\; id="2"\; label="v"\; filenames="HDA00003084487200022.png"\; state="\{\{state\}\}">v</figure-callout>}}_2=\frac{1}{3}\frac{d_2}{({\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="HDA00003084487200012.png,HDA00003084487200021.png"\; state="\{\{state\}\}">d</figure-callout>}}_1+d_2)}+\frac{2}{3}\frac{d_1}{({\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="HDA00003084487200012.png,HDA00003084487200021.png"\; state="\{\{state\}\}">d</figure-callout>}}_1+d_2)}$

If class label is 3, then calculate respectively this proper vector to 2,3 classes and 3,4 class classifying faces apart from d ₂And d ₃, final quantized value v ₃For:

${\mathrm{<figure-callout\; id="3"\; label="v"\; filenames="HDA00003084487200022.png"\; state="\{\{state\}\}">v</figure-callout>}}_3=\frac{2}{3}\frac{d_2}{({\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="HDA00003084487200012.png,HDA00003084487200021.png"\; state="\{\{state\}\}">d</figure-callout>}}_1+d_2)}+1\&CenterDot;\frac{d_3}{({\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="HDA00003084487200012.png,HDA00003084487200021.png"\; state="\{\{state\}\}">d</figure-callout>}}_1+d_2)}$

Carry out after distance calculates and quantize identification taking as above mode, add up and record each quantized value constantly continuously, can obtain the traffic behavior quantitation curve in certain bar through street a period of time, for the classification of traffic behavior prediction provides the data reference, and provide decision support effectively accurately for intelligent traffic administration system.

The present invention can obtain following beneficial effect:

(1) the designed traffic behavior of the present invention each module of quantizing identification is all chosen optimal case, be principle to reduce system's cumulative errors as much as possible, remedied the defective of existing research approach, thereby reached total optimization, can obtain reliable traffic behavior and quantize recognition data.

Ceases to be busy sequence symbol was actually and specifies the accumulation of line segment on space-time when (2) the present invention was defined, what therefore produce is virtual " background ", but not the background on the practical significance, thereby problems such as renewal that the introducing by background brings, noise have effectively been avoided.In addition, the present invention directly extracts feature from the video frame images matrix, but not traffic parameter, causes the inaccurate problem that influences the traffic behavior classification of parameter extraction thereby solved because of the traffic behavior picture material complexity increase of blocking up.

(3) for the time ceases to be busy sequence symbol image array dimensionality reduction and feature extraction, the present invention adopts the method for PCA+Fisher LDA, make both mutual supplement with each other's advantages, at first utilize the redundant composition between PCA removal data, and on projecting direction, " draw back " data pitch from, take full advantage of class label then and carry out LDA and handle, the eigenmatrix of final extraction is distributed by class as much as possible, be conducive to next step classification and quantize identification handle, make classification more accurate and effective.

(4) SVM is based on global optimum, and the Classification and Identification of nonlinear data is had incomparable advantage.The present invention is applied to the traffic status identification classification with SVM, makes classifying quality more accurate, based on the quantizing process of classification, can provide strong data support for follow-up traffic forecast, intelligent management simultaneously.

Description of drawings

Fig. 1 is the inventive method overall flow figure;

Ceases to be busy sequence symbol obtained synoptic diagram when Fig. 2 was;

Fig. 3 is the delay-line structure synoptic diagram;

Fig. 4 is sorter building method synoptic diagram;

Fig. 5 is for calculating the distance from point to curved surface synoptic diagram.

Embodiment

The hardware device that the present invention needs comprises video collector and video processor.Normal video collector (first-class as making a video recording) is set up on the overline bridge in the city expressway section, and the adjustment position makes it aim at identification to be detected track.The vision signal that collects is delivered to video processor (being generally personal computer).Algorithm in software is realized part, can use relevant interfaces such as Opencv, Directshow to realize that videos read and preprocessing part, call in Visual C++2008 that instrument such as Matlab finishes that the processing of matrix is calculated, the realization of SVM part and the integration debugging work of system.

The present invention is based on the video of gathering on the city expressway, from video, extract the time ceases to be busy sequence symbol to characterize time and space information, then it is carried out data compression and feature extraction, and construct the multicategory classification device as the input vector of SVM, realize quantification identification to traffic behavior by computational data to the distance on neighbor classified plane at last.Below in conjunction with accompanying drawing, elaborate the specific embodiment of the present invention.

A kind of traffic behavior based on visual signature quantizes recognition methods, and process flow diagram may further comprise the steps as shown in Figure 1:

Step 1 reads video from video frequency collection card, each two field picture in the original video is carried out pre-service: obtain gray level image after every red, green, blue component in the digital picture is averaged.Take following gray processing formula among the present invention:

f(x,y)＝1/3(R(x,y)+G(x,y)+B(x,y))

In the formula, f (x y) is the gray-scale value of image after the pre-service,, R (x, y), G (x, y), (x y) is respectively red, green, blue color component in the original video frame to B.

Step 2 is extracted temporal and spatial correlations information in the video frame image behind gray processing, instant ceases to be busy sequence symbol S.The time ceases to be busy sequence symbol obtain synoptic diagram as shown in Figure 2.

Step 3 takes objective estimation to differentiate the mode that combines with subjective, for the time ceases to be busy sequence symbol that has obtained adds the traffic behavior tag along sort.

Step 4 is carried out dimension-reduction treatment to the time ceases to be busy sequence symbol that adds label, extracts proper vector then.

At first take two-way two-dimentional principal component analytical method (LR2DPCA) that image array is carried out dimension-reduction treatment, respectively image is compressed in the row direction with on the column direction, twice PCA contribution rate all is set at 90% in the computation process.Time ceases to be busy sequence symbol image array after taking Fisher LDA method to dimensionality reduction then carries out further feature extraction.The image array dimension dimensionality reduction 3 of ceases to be busy sequence symbol when carrying out afterwards in above two steps.

Step 5 will be extracted the proper vector the obtain input structural classification device as SVM in the step 3, the building method of sorter as shown in Figure 4.

Step 6 at first obtains tag along sort with vector input SVM to be identified, calculates this vector distance between neighbor classified then, at last according to this apart from quantizing the identification traffic behavior, and draw the traffic behavior quantitation curve.

Claims (4)

1. the traffic behavior based on visual signature quantizes recognition methods, it is characterized in that may further comprise the steps:

Step 1 reads video from video frequency collection card, each two field picture in the original video is carried out pre-service, converts digital picture to gray level image;

Step 2 is extracted temporal and spatial correlations information in the video frame image behind gray processing, instant ceases to be busy sequence symbol S;

Step 3 takes objective estimation to differentiate the mode that combines with subjective, for the time ceases to be busy sequence symbol that has obtained adds the traffic behavior tag along sort, specifically comprises following content:

(1) " the urban traffic control assessment indicator system " announced in 2002 according to China Ministry of Public Security is divided into unimpeded (more than the 30km/h), slight crowded (20-30km/h), crowded (10-20km/h) and serious crowded (10km/h is following) four classes according to running speed v with traffic behavior;

(2) be calculated as follows under the different speed of a motor vehicle vehicle the time ceases to be busy sequence symbol in shared " time block " ratio:

(c _l·f)/(v·N)

In the formula, f is frame per second, and N is the video frame number that superposes, c _lMean value for the wheelbase of the car that obtains according to the average living standard of the different cities people and official register type of vehicle statistics;

(3) gather and extract time ceases to be busy sequence symbol in the long enough time that comprises all traffic behaviors, in conjunction with the time ceases to be busy sequence symbol (W * dimension N) is rationally set the capacity of sample set, then according to the time block size that obtains in (2), mode by subjective judgement for the time empty descriptor add class label, will (1) in traffic behavior of setting to be labeled as successively be 1,2,3,4 classes;

Step 4 is carried out dimension-reduction treatment to the time ceases to be busy sequence symbol that adds label, extracts proper vector then;

Step 5 will be extracted the proper vector the obtain input structural classification device as SVM in the step 3;

The sorter number is C (C-1)/2=6; At first classify in the 1st class of difference maximum and the 4th class, next structural classification device in 2 and 4,1 and 3 in the end separates all kinds of in one deck then fully;

Can't specified data whether during linear separability, suppose that it is linear inseparable situation, the solution of SVM is by a Nonlinear Mapping data to be mapped to higher dimensional space, does linear regression then in higher dimensional space; The Nonlinear Mapping expression formula is:

y(x)＝w ^Tφ(x)+b

In the formula, x represents proper vector, and φ (x) represents mapping function, and b represents the gain in the linear function;

The building method of sorter is as follows in SVM:

Training dataset is N proper vector x ₁..., x _N, desired value is: t ₁..., t _N, t _n∈ 1,1}, and n=1,2 ..., N, and satisfy y (x _n)＞0 o'clock, t _n=1, y (x _n)＜0 o'clock, t _n=-1, t is always arranged _nY (x _n)＞0;

If the classification plane is:

The classificating thought of SVM is the maximization class interval, can push away to such an extent that classification is spaced apart 2/|| ω by geometric knowledge ||, therefore make to make satisfying under the situation of constraint condition || ω || minimize the optimal classification face that can obtain; Main solution procedure is as follows:

Introduce Lagrangian function:

In the formula, a _n〉=0, be Lagrange multiplier;

Respectively to w, b differentiate, and make that derivative is zero:

$w=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{a}_n{t}_n\mathrm{\&phi;}\left(x_n\right)$

$\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{a}_n{t}_n=0$

Bring the dual form that the Lagrangian function expression formula obtains this function into:

$\tilde{L}\left(a\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{a}_n-\frac{1}{2}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}{a}_n{a}_m{t}_n{t}_{m}k(x_n,x_m)$

In the formula, and k (x, x')=φ (x) ^Tφ (x') is kernel function, and it can accept the input value of lower dimensional space, exports the inner product value of higher dimensional space then, thereby makes computation process needn't be concerned about concrete mapping relations φ (x), so mapping function can be expressed as:

$y\left(x\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{a}_n{t}_{n}k(x,x_n)+b$

According to the KKT(Karush-Kuhn-Tucker that satisfies this formula) condition can obtain the value of b:

$b=\frac{1}{N_s}\underset{n\&Element;S}{\mathrm{\&Sigma;}}(t_n-\underset{m\&Element;S}{\mathrm{\&Sigma;}}{a}_m{t}_{m}k(x_n,x_m))$

In the formula, N _sSummation for all support vectors;

In the traffic behavior sorting technique, adopt Gaussian radial basis function (Gauss Radial BasisFunction) in support vector machine, function expression is:

$k(x_i,x)=\exp \{-\frac{{|x-x_i|}^2}{\mathrm{\&sigma;}}\}$

In the formula, x _iBe the kernel function center, σ is the parameter of radial extension;

Step 6 at first obtains tag along sort with vector input SVM to be identified, calculates this vector distance between neighbor classified then, at last according to this apart from quantizing the identification traffic behavior, and draw the traffic behavior quantitation curve.

2. a kind of traffic behavior based on visual signature according to claim 1 quantizes recognition methods, it is characterized in that, the method for ceases to be busy sequence symbol S was as follows when step 2 was extracted:

(1) a virtual spacetime line is set in gray level image, the pixel value on the virtual spacetime line in every frame is designated as the vectorial L of row _t=[p _T1, p _T2..., p _Tw] ^T, wherein p is the point on the line segment;

(2) this row vector in t-N+1～t frame video sequence is superposeed, obtain the time ceases to be busy sequence symbol S of current time _t=[L _T-N+1, L _T-N+2..., L _t] ^TEach the time ceases to be busy sequence symbol be the matrix of a W * N, wherein W be set the time ceases to be busy sequence symbol width, N is the selected video frame number that superposes;

(3) the time ceases to be busy sequence symbol in the long enough time that comprises all traffic behaviors is gathered and extracted to definition delay-line structure, obtains ceases to be busy sequence symbol combination U=[S when whole ₁, S ₂..., S _t..., S _M] ^T, M be obtain the time ceases to be busy sequence symbol sample.

3. a kind of traffic behavior based on visual signature according to claim 1 quantizes recognition methods, it is characterized in that, step 4 is carried out dimension-reduction treatment to the time ceases to be busy sequence symbol that adds label and extract the method for proper vector as follows:

(1) take two-way two-dimentional principal component analytical method (LR2DPCA) to carry out the data dimensionality reduction;

1. to the time ceases to be busy sequence symbol be listed as compression

At first, definition projection properties vector y ₁For:

y ₁＝Sw ₁

In the formula, w ₁Be projection matrix;

Then, be calculated as follows sample image average Φ:

$\mathrm{\&Phi;}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{S}_i$

In the formula, M is total sample number;

At this moment, image divergence matrix G _tFor:

$G_t=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{(S_i-\mathrm{\&Phi;})}^T(S_i-\mathrm{\&Phi;})$

Choose optimum axis of projection according to minimizing mean-square error criteria, just make w ₁ ^TG _tw ₁Minimize, the solution of this problem is G _tPreceding p eigenvalue of maximum characteristic of correspondence vector; The contribution rate that adds up of p eigenwert before the quantity of p depends on; Finally can get y ₁=Sw ₁=[y ₁, y ₂..., y _p], dimension is W*p;

2. to the time ceases to be busy sequence symbol go compression

Defining the projection properties vector again is: y ₂=w ₂ ^Ty ₁, and recomputate sample image average Φ ' and the image divergence matrix G of this moment by 1. method _t', the solution of this problem is G _t' preceding q eigenvalue of maximum characteristic of correspondence vector; Thereby y ₂=w ₂ ^Ty ₁=[y ₁, y ₂..., y _q], dimension is q*p;

So far, ceases to be busy sequence symbol image array is converted into the time:

x＝y ₂＝w ₂ ^Ty ₁＝w ₂ ^TSw ₁

(2) the time ceases to be busy sequence symbol image array after taking Fisher LDA method to dimensionality reduction carries out further feature extraction;

Definition projection properties vector is:

y ₃＝w ₃ ^Tx

Divergence matrix (within-class scatter) in the compute classes:

$S_w=\underset{j=1}{\overset{C}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{N_j}{\mathrm{\&Sigma;}}}(x_i^j-u_j){(x_i^j-u_j)}^T$

In the formula, I width of cloth image in the expression j class, and u _jAnd N _jAverage and the quantity of representing j class image;

Calculate between class scatter matrix (between-class scatter):

$S_b=\underset{j=1}{\overset{C}{\mathrm{\&Sigma;}}}(u_j-u){(u_j-u)}^T$

In the formula, u represents the average of all images in all sample sets;

The thought of LDA is for minimizing distance in the class in the maximization between class distance, the projection criterion of taking is the ratio of maximization the former with the latter, that is:

J(w ₃)＝w ₃ ^TS _bw ₃/w ₃ ^TS _ww ₃

This problem is converted into finds the solution S _w ^-1S _bGeneralized eigenvalue problem; Because S _w ^-1S _bOrder be C-1, final w ₃Corresponding to S _w ^-1S _bThe eigenmatrix of maximum C-1 eigenwert.

4. a kind of traffic behavior based on visual signature according to claim 1 quantizes recognition methods, it is characterized in that the method that the step 6 traffic behavior quantizes identification is as follows:

(1) vector input SVM to be identified is obtained tag along sort;

(2) calculate this vector distance between neighbor classified;

(3) identify traffic behavior according to this apart from quantification, and draw the traffic behavior quantitation curve;

The 1st, 4 classes are defined as the unimpeded and serious crowded state of traffic conditions respectively in step 2, and generally the degree of crowding of the 2nd, 3 classes can and characterize the substantive help of traffic behavior generation to prediction, so the quantification of traffic behavior identification is primarily aimed at 2,3 classifications; Therefore, defining the 1st class traffic behavior quantized value is that 0, the 4 class traffic behavior quantized value is 1, and the quantized value of the point on the optimal classification face of the 2nd, 3 classes is respectively 1/3,2/3; According to the classification results of SVM in the step 4, if class label is 2, this proper vector to 1,2 classifying faces and 2,3 classifying faces apart from d ₁And d ₂, final quantized value v ₂For:

$v_2=\frac{1}{3}\frac{d_2}{(d_1+d_2)}+\frac{2}{3}\frac{d_1}{(d_1+d_2)}$

If class label is 3, then calculate respectively this proper vector to 2,3 classes and 3,4 class classifying faces apart from d ₂And d ₃, final quantized value v ₃For:

$v_3=\frac{2}{3}\frac{d_2}{(d_1+d_2)}+1\&CenterDot;\frac{d_3}{(d_1+d_2)}$

Add up and record each quantized value constantly continuously, can obtain the traffic behavior quantitation curve in certain bar through street a period of time.

Images