CN104881630A - Vehicle identification method based on window segmentation and fuzzy characteristics - Google Patents

Abstract

The invention discloses a vehicle identification method based on window segmentation and fuzzy characteristics. The vehicle identification method comprises extracting the edge of a front face image of a vehicle; then conducting quadrilateral fitting according to the edge so as to quickly segment a front windshield contour of the vehicle; next, extracting HSV image histogram characteristics in the range of windshield; and finally retrieving the vehicle through the contrast of region Fuzzy difference measurement of HSV histograms. According to the invention, the vehicle identification method gets rid of the unreliability of retrieval only relying on license plates, and the vehicle image retrieval results are improved.

Description

Based on the vehicle identification method of vehicle window segmentation with fuzzy feature

Technical field

The invention belongs to traffic management technology field, particularly relate to a kind of based on the vehicle identification method of vehicle window segmentation with fuzzy feature.

Background technology

The current fields such as public security of having stable political situation in urban transportation, city, the application of the supervisory systems such as electronic eyes is day by day universal.These data such as system acquisition continuous collecting vehicle image and video, extract useful information and carry out auxiliary regulatory authorities and carry out scientific management and decision-making further from the huge database collected.Such as at major urban arterial highway crossing and expressway key position, by adopting license plate recognition technology to help carry out the Intelligent treatment such as vehicle tracking, flow analysis, achieve very ten-strike.

But requiring more and more instantly high to social public security, often deliberately block because of suspected vehicles or change car plate, only identify that car plate cannot meet the Search Requirement for specific suspicion of crime vehicle, specific objective vehicle must be retrieved by the characteristic information that in image, other are difficult to change in huge database.And on the front window windshield of vehicle, generally be pasted with the marks such as annual test, compulsory insurance for traffic accident of motor-drivenvehicle and environmental protection tests, also the relatively-stationary objects in some positions such as mascot may be placed with, the color characteristic of these marks and object and relative position relation just constitute the characteristics of image of windscreen, contribute to retrieving the fast and reliable of target vehicle in a large amount of front face image data base.

Summary of the invention

The object of the invention is to the above-mentioned problem solving this field existing, a kind of target vehicle image search method is reliably provided, the method is by identifying the windshield profile of front part of vehicle, and realize retrieval according to the region Fuzzy diversity factor of image HSV histogram feature in profile, comprise the following steps:

Obtain the outline map of image to be detected; Ledgement close to level and the subvertical vertical moulding of edge figure carry out fitting a straight line, and different Straight Combination becomes different quadrilaterals; Using the highest quadrilateral of the shape similarity and the edge goodness of fit with target vehicle vehicle window sample as the elementary contour of image vehicle window to be detected; Determine the HSV histogram of image vehicle window to be detected; Carry out region Fuzzy difference angle value with the HSV histogram of target vehicle vehicle window sample to calculate; In the same way, one by one the HSV histogram of image in vehicle database and target vehicle is carried out region Fuzzy difference angle value to calculate; Result for retrieval is drawn according to region Fuzzy difference measurement.

As preferably, described outline map can be obtained by the Canny operator of region adaptivity threshold value.

As preferably, in described outline map, search for the ledgement close to level and subvertical vertical moulding, and lines are polymerized and screen; Fitting a straight line is carried out to horizontal line, filters out the horizontal line that degree of fitting is high, and calculate main level angle of inclination; Vertical line is screened and fitting a straight line; According to Straight Line Fitting Parameters, from horizontal line Resource selection two horizontal lines, respectively select a line to be combined as a quadrilateral with concentrating from left and right vertical line, different combinations forms quadrilateral collection.

Further, described horizontal line fitting a straight line, filters out the horizontal line that degree of fitting is high, and calculates in the process at main level angle of inclination, and adopt the mode of ballot to determine main horizontal tilt angle Horz_Angle, voting process is as follows:

1) use least square method, according to the point sequence fitting a straight line y=kx+b of each bar horizontal line, and digital simulation degree is that sigma is as follows, wherein x _s, y _sfor the point in horizontal line point sequence, len is horizontal line length;

y＝k·x _s+b

sigma＝∑(y-y _s) ²/len

2) according to the parameter k (straight slope) of fitting a straight line, obtain all fitting a straight lines of level angle within the scope of positive and negative 10 °, and determine angular range [Ang_Min, Ang_Max];

3) [Ang_Min, Ang_Max] is divided into 20 bin, each bin angular span is Ang _binbecause the horizontal line filtered out is between ± 10 °, therefore Ang _binbe less than or equal to 1 °;

4) to each horizontal line, its angle A ng is obtained according to its k value, according to this angle and nearest i-th bin angle Ang _idistance, vote to i-th bin and adjacent with i two bin, ballot value vote is calculated as follows:

${\mathrm{vote}}_b=\mathrm{len}\·e^{-\frac{{(\mathrm{Ang}-{\mathrm{Ang}}_b)}^2}{{{\mathrm{Ang}}_{\mathrm{bin}}}^2}}/p\left(\mathrm{sigma}\right),b=i-1,i,i+1$

$p\left(\mathrm{sigma}\right)=\left\{\begin{array}{cc}0.5& \mathrm{sigma}<0.5\\ \mathrm{sigma}& 0.5<=\mathrm{sigma}<=2\\ 2& \mathrm{sigma}>2\end{array}\right.$

5) to each bin, after the ballot value of all horizontal lines is cumulative, determine that Horz_Angle is:

$\mathrm{Horz}\_\mathrm{Angle}=\arg \underset{{\mathrm{Ang}}_i}{\max }{\mathrm{vote}}_i$

Further, described two horizontal lines are set to as horz _tand horz _b, both degree of tilt are all less than 10 °, and wherein the angle of at least one and the difference of main level angle Horz_Angle are less than 5%; And horz _tand horz _bon the x direction of level, degree of overlapping can not be less than wherein compared with the half of strigula length, that is:

Overlap(horz _t，horz _b)>0.5*min(len _horzt,len _horzb)

Concentrate from described left and right vertical line and select a left vertical line left and right vertical line right, the two degree of overlapping in y-direction must be greater than the half of shorter vertical line length; Then calculate left, right respectively with horz _tand horz _bfour intersection point pt _i(x, y), i=1,2,3,4; Namely the solution of following equations group is asked:

y＝k _h·x+b _h,h＝horz _t,horz _b

x＝k _v·y+b _v,v＝left,right

By four intersection points from the upper left corner, called after pt counterclockwise ₁, pt ₂, pt ₃, pt ₄, form a quadrilateral Q; Repeat above step, produce the quadrilateral collection of various combination.

As preferably, calculate shape similarity and the edge goodness of fit that quadrilateral concentrates each quadrilateral and car bumper wind glass sample parameter, calculate the weight of this quadrilateral accordingly; Select the quadrilateral that weight is the highest, with the four edges of matching quadrilateral for windshield elementary contour shape similarity:

Similarity computation process is as follows:

1) dissimilar vehicle sample is selected, four some pt at artificial selection car windshield place ₁, pt ₂, pt ₃, pt ₄, calculate average following ratio t2b and standard deviation sigma accordingly _t2b; High ratio h2b with base and standard deviation sigma _h2b; Base angle average θ and standard deviation sigma _θ;

2) to each quadrilateral Q that quadrilateral is concentrated _i, calculate the parameter shown in its Fig. 2, and calculate four eigenwert f ₁, f ₂, f ₃, f ₄:

f ₁＝(len _t/len _b-t2b) ²/σ _t2b ²

f ₂＝((h ₁+h ₂)/len _b/2-h2b) ²/σ _h2b ²

f ₃＝(α-θ) ²/σ _θ ²+(β-θ) ²/σ _θ ²

f ₄＝(α-β) ²/σ _θ ²

3) according to four eigenwert f ₁, f ₂, f ₃, f ₄calculate Similarity:

$\mathrm{similarity}=e^{-(f_1+f_2+f_3+f_4)}$

Edge goodness of fit Fitness computation process is as follows:

1) to horizontal line horz _tupper all pt that satisfies condition ₁.x<=x<=pt ₄.x point set { p (x _i, y _i) | i=1,2 ... N _t, according to horz _tstraight Line Fitting Parameters k _tand b _t, calculating parameter fit _t:

${\mathrm{fit}}_t=N_t/{\mathrm{len}}_t/p\left(\underset{i=1}{\overset{N_t}{\mathrm{\&Sigma;}}}{(y_i-(k_t\&CenterDot;x_i+b_t))}^2\right)$

In formula, $p\left(\mathrm{\&theta;}\right)=\left\{\begin{array}{cc}1,& \mathrm{\&theta;}<1\\ \mathrm{\&theta;},& 1<=\mathrm{\&theta;}<3\\ 3,& \mathrm{\&theta;}>=3\end{array}\right.$

2) to pt that satisfies condition all on left vertical line left ₁.y<=y<=pt ₂.y point set { p (x _i, y _i) | i=1,2 ... N _l, according to the Straight Line Fitting Parameters k of left _land b _l, calculating parameter fit _l:

${\mathrm{fit}}_l=N_l/{\mathrm{len}}_l/p\left(\underset{i=1}{\overset{N_l}{\mathrm{\&Sigma;}}}{(x_i-(k_l\&CenterDot;y_i+b_l))}^2\right)$

3) horz is calculated by similar approach _bwith the parameter f it of right _band fit _r;

4) according to fit _t, fit _l, fit _band fit _rcalculate Fitness:

Fitness＝(fit _t+fit _l+fit _b+fit _r)/4

Finally, Priority is calculated according to Similarity and Fitness:

Priority＝α ₁*Similarity+α ₂*Fitness

Select the highest quadrilateral of weight as image vehicle window elementary contour to be detected.

As preferably, using quadrilateral the highest for weight as image vehicle window elementary contour to be detected, the image in profile being carried out affined transformation is regular rectangular shape, is the subregion of M*N by this characteristic area even partition, and calculate the hsv color histogram of every sub regions, step is as follows:

1) by the image that represents in the quadrilateral of image vehicle window elementary contour to be detected through the regular rectangular shape of affined transformation to uniform sizes;

2) this characteristic area is divided into the even subregion R of M*N _m,n(m<=M, n<=N);

3) by image from RGB color notation conversion space to hsv color space;

4) all subregion R is calculated _m,nthe histogram Hist of middle H, S, V passage _m,n ^h, Hist _m,n ^s, Hist _m,n ^v, in histogram, all bin normalize to [0,1] interval.

As preferably, one by one face characteristics of image before the image to be checked in vehicle database and target vehicle is carried out region Fuzzy difference angle value and calculate, retrieve the most similar some images as result for retrieval according to region Fuzzy difference measurement, computing method are as follows:

1) to each subregion R in vehicle characteristics region to be checked _m,n ^t, calculate its corresponding subregion R with target vehicle characteristic area _m,n ^sfuzzy diversity factor, the channel difference angle value of each corresponding subregion is:

${F_{m,n}}^c=\underset{b=1}{\overset{b_{\max }}{\mathrm{\&Sigma;}}}\underset{k=b-}{\overset{b+1}{\mathrm{\&Sigma;}}}{{\mathrm{\&Delta;}}_{b,k}}^2\&CenterDot;(\frac{e^{-{{\mathrm{\&Delta;}}_{b,k}}^2}}{\underset{k}{\mathrm{\&Sigma;}}{e}^{-{{\mathrm{\&Delta;}}_{b,k}}^2}}\&CenterDot;(1-e^{-{(b-k)}^2}))$

Wherein c represents Hist _m,n ^h, Hist _m,n ^s, Hist _m,n ^vthree channel histogram, b _maxrepresent the maximum bin number under this histogram, △ _b,krepresent the difference of b, k two bin, formula ensure that when two histograms are when the difference of same bin or adjacent bin is less, each channel difference angle value F _m,n ^calso less;

According to each channel difference degree, the Fuzzy difference angle value calculating this corresponding subregion is:

F _m,n＝α _H*F _m,n ^H+α _s*F _m,n ^S+α _V*F _m,n ^V

Wherein α _cfor the constant factor of each passage;

2) according to the diversity factor F of corresponding subregion _m,nzoning diversity factor F;

$F=\underset{m=1}{\overset{M}{\mathrm{\&Pi;}}}\underset{n=1}{\overset{N}{\mathrm{\&Pi;}}}{F}_{m,n}*e^{-\frac{2*\min (m,M-n)}{N}}*e^{-\frac{2*\min (n,N-n)}{N}}$

Some images less with the region Fuzzy difference angle value of target vehicle in vehicle to be measured are identified as target vehicle.

As preferably, select the front image of a target vehicle, therefrom manual selection four points determine its front window home position; By the image in scope through the rectangular characteristic region of affined transformation to specified size, and the image in this characteristic area is divided into the subregion of M*N, calculates the HSV histogram of all subregion simultaneously.

The present invention has following beneficial effect: this vehicle image search method utilizes the region Fuzzy difference angle value of image HSV histogram feature within the scope of vehicle windscreen to identify similar image, break away from the unreliability of single dependence license number search, improve the effect of vehicle image retrieval.

Accompanying drawing explanation

Fig. 1 is based on the process flow diagram of vehicle window segmentation with the vehicle identification method of fuzzy feature;

The parameter schematic diagram of Fig. 2 quadrilateral Q;

Fig. 3 calculates vehicle window region HSV histogram.

Embodiment

Below in conjunction with specific embodiment, and by reference to the accompanying drawings, technical scheme of the present invention is further described:

Embodiment 1: in step S01, when using the Canny operator of region adaptivity threshold value to ask for outline map, Canny operator needs setting height two threshold values.Conventional method adopts artificial setting or overall self-adaptation to determine, this method adopts region adaptivity method definite threshold and runs Canny operator, and step is as follows:

1) image is divided into 4*4 totally 16 pieces of subregions;

2) at each subregion, the accumulation grey level histogram in this region is calculated;

$H\left(i\right)=\underset{g=0}{\overset{i}{\mathrm{\&Sigma;}}}I\left(g\right)$

Wherein i is i-th gray level, and scope 0 ~ 255, I (g) is the number of pixels of g for gray scale in image, chooses high threshold Th according to accumulation grey level histogram _hwith Low threshold Th _l.

${\mathrm{Th}}_h=\arg \underset{i}{\min }H\left(i\right)>0.7*H\left(255\right)$

Th _l＝0.4*Th _h

3) Canny operator is run according to region threshold.When the process of Canny operator is to certain subregion, adopt the high-low threshold value in this region to process, in trans-regional boundary neighborhood, adopt the threshold average process in adjacent subarea territory.

Embodiment 2: each quadrilateral Q concentrated quadrilateral, calculates shape similarity Similarity and the edge goodness of fit Fitness of quadrilateral and car bumper wind glass sample parameter, calculates the weight Priority of this quadrilateral accordingly;

Shape similarity Similarity reflects the similarity degree of windshield shape in found quadrilateral Q and sample.Concrete computation process is as follows:

1) select 100 and comprise the dissimilar vehicle sample such as car, minibus, lorry, SUV, four some pt at artificial selection car windshield place ₁, pt _2,pt ₃, pt ₄, calculate average following ratio t2b=0.7716 accordingly, standard deviation sigma _t2b=0.0245; The high ratio h2b=0.3570 with base, standard deviation sigma _h2b=0.0346; Average θ=72.12, base angle, standard deviation sigma _θ=2.5373.

2) to each quadrilateral Q that quadrilateral is concentrated _i, calculate its parameter as shown in Figure 2, and calculate four eigenwert f ₁, f ₂, f ₃, f ₄:

f ₁＝(len _t/len _b-t2b) ²/σ _t2b ²

f ₂＝((h ₁+h ₂)/len _b/2-h2b) ²/σ _h2b ²

f ₃＝(α-θ) ²/σ _θ ²+(β-θ) ²/σ _θ ²

f ₄＝(α-β) ²/σ _θ ²

3) according to four eigenwert f ₁, f ₂, f ₃, f ₄calculate Similarity:

$\mathrm{similarity}=e^{-(f_1+f_2+f_3+f_4)}$

Edge goodness of fit Fitness reflects the four edges of Q and four the line horz participating in matching Q _t, horz _b, the laminating degree of left and right on image space positions, computation process is as follows:

1) to horizontal line horz _tupper all pt that satisfies condition ₁.x<=x<=pt ₄.x point set { p (x _i, y _i) | i=1,2 ... N _t, according to horz _tstraight Line Fitting Parameters k _tand b _t, calculating parameter fit _t:

${\mathrm{fit}}_t=N_t/{\mathrm{len}}_t/p\left(\underset{i=1}{\overset{N_t}{\mathrm{\&Sigma;}}}{(y_i-(k_t\&CenterDot;x_i+b_t))}^2\right)$

In formula, $p\left(\mathrm{\&theta;}\right)=\left\{\begin{array}{cc}1,& \mathrm{\&theta;}<1\\ \mathrm{\&theta;},& 1<=\mathrm{\&theta;}<3\\ 3,& \mathrm{\&theta;}>=3\end{array}\right.$

2): to pt that satisfies condition all on left vertical line left ₁.y<=y<=pt ₂.y point set { p (x _i, y _i) | i=1,2 ... N _l, according to the Straight Line Fitting Parameters k of left _land b _l, calculating parameter fit _l:

${\mathrm{fit}}_l=N_l/{\mathrm{len}}_l/p\left(\underset{i=1}{\overset{N_l}{\mathrm{\&Sigma;}}}{(x_i-(k_l\&CenterDot;y_i+b_l))}^2\right)$

3) horz is calculated by similar approach _bwith the parameter f it of right _band fit _r;

4) according to fit _t, fit _l, fit _band fit _rcalculate Fitness:

Fitness＝(fit _t+fit _l+fit _b+fit _r)/4

Priority is calculated according to Similarity and Fitness:

Priority＝α ₁*Similarity+α ₂*Fitness

Wherein α ₁get 0.7, α ₂get 0.3, because of according to statistics, the contribution of shape similarity to window edge matching is larger.

Select quadrilateral Q that weights Priority is the highest as windshield elementary contour, by the part of this front part of vehicle image in elementary contour through the characteristic area of affined transformation to uniform sizes, and be divided into the subregion of M*N (12*8).Then image is transformed into HSV space by rgb space, 36 bin are divided into according to H passage, channel S is divided into 10 bin, V passage is divided into 16 bin, and all bin are normalized to [0,1] interval mode calculates the hsv color histogram of every sub regions, as face characteristics of image before this vehicle.As shown in Figure 3.

Above-mentioned embodiment is used for explaining and the present invention is described, instead of limits the invention, and in the protection domain of spirit of the present invention and claim, any amendment make the present invention and change, all fall into protection scope of the present invention.

Claims (9)

1., based on vehicle window segmentation and the vehicle identification method of fuzzy feature, it is characterized in that, comprise the following steps:

S01 obtains the outline map of image to be detected;

Ledgement close to level and the subvertical vertical moulding of S02 edge figure carry out fitting a straight line, and different Straight Combination becomes different quadrilaterals;

S03 is using the highest quadrilateral of the shape similarity and the edge goodness of fit with target vehicle vehicle window sample as the elementary contour of image vehicle window to be detected;

S04 determines the HSV histogram of image vehicle window to be detected;

The HSV histogram of S05 and target vehicle vehicle window sample carries out region Fuzzy difference angle value and calculates;

The HSV histogram of image in vehicle database and target vehicle in the same way, is carried out region Fuzzy difference angle value and calculates by S06 one by one;

S07 draws result for retrieval according to region Fuzzy difference measurement.

2. split the vehicle identification method with fuzzy feature according to according to claim 1 based on vehicle window, it is characterized in that: the outline map described in step S01 uses the Canny operator of region adaptivity threshold value to obtain.

3. split the vehicle identification method with fuzzy feature according to according to claim 1 based on vehicle window, it is characterized in that, step S02 specifically comprises: in outline map, search for the ledgement close to level and subvertical vertical moulding, and be polymerized lines and screen; Fitting a straight line is carried out to horizontal line, filters out the horizontal line that degree of fitting is high, and calculate main level angle of inclination; Vertical line is screened and fitting a straight line; According to Straight Line Fitting Parameters, from horizontal line Resource selection two horizontal lines, respectively select a line to be combined as a quadrilateral with concentrating from left and right vertical line, different combinations forms quadrilateral collection.

4. split the vehicle identification method with fuzzy feature according to according to claim 3 based on vehicle window, it is characterized in that: to horizontal line fitting a straight line, filter out the horizontal line that degree of fitting is high, and calculate in the process at main level angle of inclination, adopt the mode of ballot to determine main horizontal tilt angle Horz_Angle, voting process is as follows:

1) use least square method, according to the point sequence fitting a straight line y=kx+b of each bar horizontal line, and digital simulation degree is that sigma is as follows, wherein x _s, y _sfor the point in horizontal line point sequence, len is horizontal line length;

y＝k·x _s+b

sigma＝∑(y-y _s) ²/len

2) according to the parameter k (straight slope) of fitting a straight line, obtain all fitting a straight lines of level angle within the scope of positive and negative 10 °, and determine angular range [Ang_Min, Ang_Max];

3) [Ang_Min, Ang_Max] is divided into 20 bin, each bin angular span is Ang _binbecause the horizontal line filtered out is between ± 10 °, therefore Ang _binbe less than or equal to 1 °;

4) to each horizontal line, its angle A ng is obtained according to its k value, according to this angle and nearest i-th bin angle Ang _idistance, vote to i-th bin and adjacent with i two bin, ballot value vote is calculated as follows:

${\mathrm{vote}}_b=\mathrm{len}\&CenterDot;e^{-\frac{{(\mathrm{Ang}-{\mathrm{Ang}}_b)}^2}{{{\mathrm{Ang}}_{\mathrm{bin}}}^2}}/p\left(\mathrm{sigma}\right),b=i-1,i,i+1$

$p\left(\mathrm{sigma}\right)=\left\{\begin{array}{cc}0.5& \mathrm{sigma}<0.5\\ \mathrm{sigma}& 0.5<=\mathrm{sigma}<=2\\ 2& \mathrm{sigma}>2\end{array}\right.$

5) to each bin, after the ballot value of all horizontal lines is cumulative, determine that Horz_Angle is:

$\mathrm{Horz}\_\mathrm{Angle}=\arg \underset{{\mathrm{Ang}}_i}{\max }{\mathrm{vote}}_i$

5. split the vehicle identification method with fuzzy feature according to according to claim 3 based on vehicle window, it is characterized in that: described two horizontal lines are horz _tand horz _b,degree of tilt is all less than 10 °, and wherein the angle of at least one and the difference of main level angle Horz_Angle are less than 5%; And horz _tand horz _bon the x direction of level, degree of overlapping can not be less than wherein compared with the half of strigula length, that is:

Overlap(horz _t，horz _b)>0.5*min(len _horzt,len _horzb)

Concentrate from described left and right vertical line and select a left vertical line left and right vertical line right, the two degree of overlapping in y-direction must be greater than the half of shorter vertical line length; Then calculate left, right respectively with horz _tand horz _bfour intersection point pt _i(x, y), i=1,2,3,4; Namely the solution of following equations group is asked:

y＝k _h·x+b _h,h＝horz _t,horz _b

x＝k _v·y+b _v,v＝left,right

By four intersection points from the upper left corner, called after pt counterclockwise ₁, pt _2,pt ₃, pt ₄, form a quadrilateral Q; Repeat above step, produce the quadrilateral collection of various combination.

6. split the vehicle identification method with fuzzy feature according to according to claim 1 based on vehicle window, it is characterized in that, step S03 specifically comprises:

Shape similarity Similarity computation process is as follows:

1) dissimilar vehicle sample is selected, four some pt at artificial selection car windshield place ₁, pt ₂, pt ₃, pt ₄, calculate average following ratio t2b and standard deviation sigma accordingly _t2b; High ratio h2b with base and standard deviation sigma _h2b; Base angle average θ and standard deviation sigma _θ;

2) to each quadrilateral Q that quadrilateral is concentrated _i, calculate the parameter shown in its Fig. 2, and calculate four eigenwert f ₁, f ₂, f ₃, f ₄:

f ₁＝(len _t/len _b-t2b) ²/σ _t2b ²

f ₂＝((h ₁+h ₂)/len _b/2-h2b) ²/σ _h2b ²

f ₃＝(α-θ) ²/σ _θ ²+(β-θ) ²/σ _θ ²

f ₄＝(α-β) ²/σ _θ ²

3) according to four eigenwert f ₁, f ₂, f ₃, f ₄calculate Similarity:

$\mathrm{similarity}=e^{-(f_1+f_2+f_3+f_4)}$

Edge goodness of fit Fitness computation process is as follows:

1) to horizontal line horz _tupper all pt that satisfies condition ₁.x<=x<=pt ₄.x point set { p (x _i, y _i) | i=1,2 ... N _t, according to horz _tstraight Line Fitting Parameters k _tand b _t, calculating parameter fit _t:

${\mathrm{fit}}_t=N_t/{\mathrm{len}}_t/p\left(\underset{i=1}{\overset{N_t}{\mathrm{\&Sigma;}}}{(y_i-(k_t\&CenterDot;x_i+b_t))}^2\right)$

In formula, $p\left(\mathrm{\&theta;}\right)=\left\{\begin{array}{cc}1,& \mathrm{\&theta;}<1\\ \mathrm{\&theta;},& 1<=\mathrm{\&theta;}<3\\ 3,& \mathrm{\&theta;}>=3\end{array}\right.$

2) to pt that satisfies condition all on left vertical line left ₁.y<=y<=pt ₂.y point set { p (x _i, y _i) | i=1,2 ... N _l, according to the Straight Line Fitting Parameters k of left _land b _l, calculating parameter fit _l:

${\mathrm{fit}}_l=N_l/{\mathrm{len}}_l/p\left(\underset{i=1}{\overset{N_l}{\mathrm{\&Sigma;}}}{(x_i-(k_l\&CenterDot;y_i+b_l))}^2\right)$

3) horz is calculated by similar approach _bwith the parameter f it of right _band fit _r;

4) according to fit _t, fit _l, fit _band fit _rcalculate Fitness:

Fitness＝(fit _t+fit _l+fit _b+fit _r)/4

Finally, Priority is calculated according to Similarity and Fitness:

Priority＝α ₁*Similarity+α ₂*Fitness

Select the highest quadrilateral of weight as image vehicle window elementary contour to be detected.

7. split the vehicle identification method with fuzzy feature according to according to claim 1 based on vehicle window, it is characterized in that, step S04 specifically comprises: using quadrilateral the highest for weight as image vehicle window elementary contour to be detected, image in profile being carried out affined transformation is regular rectangular shape, be the subregion of M*N by this characteristic area even partition, and calculate the hsv color histogram of every sub regions, step is as follows:

1) by the image that represents in the quadrilateral of image vehicle window elementary contour to be detected through the regular rectangular shape of affined transformation to uniform sizes;

2) this characteristic area is divided into the even subregion R of M*N _m,n(m<=M, n<=N);

3) by image from RGB color notation conversion space to hsv color space;

4) all subregion R is calculated _m,nthe histogram Hist of middle H, S, V passage _m,n ^h, Hist _m,n ^s, Hist _m,n ^v, in histogram, all bin normalize to [0,1] interval.

8. split the vehicle identification method with fuzzy feature according to according to claim 1 based on vehicle window, it is characterized in that: face characteristics of image before the image to be checked in vehicle database and target vehicle is carried out region Fuzzy difference angle value and calculates by described method one by one, retrieve the most similar some images as result for retrieval according to region Fuzzy difference measurement, computing method are as follows:

1) to each subregion R in vehicle characteristics region to be checked _m,n ^t, calculate its corresponding subregion R with target vehicle characteristic area _m,n ^sfuzzy diversity factor, the channel difference angle value of each corresponding subregion is:

${F_{m,n}}^c=\underset{b=1}{\overset{b_{\max }}{\mathrm{\&Sigma;}}}\underset{k=b-1}{\overset{b+1}{\mathrm{\&Sigma;}}}{{\mathrm{\&Delta;}}_{b,k}}^2\&CenterDot;(\frac{e^{{{-\mathrm{\&Delta;}}_{b,k}}^2}}{\underset{k}{\mathrm{\&Sigma;}}{e}^{{{-\mathrm{\&Delta;}}_{b,k}}^2}}\&CenterDot;(1-e^{{-(b-k)}^2}))$

Wherein c represents Hist _m,n ^h, Hist _m,n ^s, Hist _m,n ^vthree channel histogram, b _maxrepresent the maximum bin number under this histogram, Δ _b,krepresent the difference of b, k two bin, formula ensure that when two histograms are when the difference of same bin or adjacent bin is less, each channel difference angle value F _m,n ^calso less;

According to each channel difference degree, the Fuzzy difference angle value calculating this corresponding subregion is:

F _m,n＝α _H*F _m,n ^H+α _s*F _m,n ^S+α _V*F _m,n ^V

Wherein α _cfor the constant factor of each passage;

2) according to the diversity factor F of corresponding subregion _m,nzoning diversity factor F;

$F=\underset{m=1}{\overset{M}{\mathrm{\&Pi;}}}\underset{n=1}{\overset{N}{\mathrm{\&Pi;}}}{F}_{m,n}*e^{-\frac{2*\min (m,M-n)}{M}}*e^{-\frac{2*\min (n,N-n)}{N}}$

Some images less with the region Fuzzy difference angle value of target vehicle in vehicle to be measured are identified as target vehicle.

9. split the vehicle identification method with fuzzy feature according to according to claim 1 based on vehicle window, it is characterized in that, the HSV histogram of target vehicle described in step S06 is determined by the following method: the front image selecting a target vehicle, and therefrom manual selection four points determine its front window home position; By the image in scope through the rectangular characteristic region of affined transformation to specified size, and the image in this characteristic area is divided into the subregion of M*N, calculates the HSV histogram of all subregion simultaneously.