CN104899881A - Shadow detection method for moving vehicle in video image - Google Patents

Abstract

The present invention discloses a shadow detection method for a moving vehicle in a video image, used for eliminating the shadow of a moving target in a foreground image and improving the accuracy of target detection and tracing. The shadow detection method provided by the present invention comprises steps of: building a feature image of information of fusion brightness, chrominance, edge gradient and the like; after cutting the image, taking a region with highest color difference as a starting region of vehicle body search; performing iterative search and detection on a region with highest feature value in peripheral adjacent regions until meeting a stopping criterion for iteration; if the rest sub-region set is a null set, showing that the shadow of the moving target can be neglected, otherwise, taking the rest sub-regions as shadow candidate regions; then searching shadow sub-regions from each of the shadow candidate regions; and finally combining all the searched shadow sub-regions into a complete shadow region. The method is capable of automatically determining whether the shadow exists, fusing various features more reasonably and reducing manual interventions, and is high in shadow detection rate and good in universality.

Description

Moving vehicle shadow detection method in a kind of video image

Technical field

The invention belongs to digital image processing field, particularly the shadow detection method of moving vehicle in a kind of video image.

Background technology

Shade mainly object block light source and formed dark areas, the feature according to shade can be divided into from shade and cast shadow.From shade be object itself not by the part that light irradiates, the part that should belong to target object from shade of moving target in video analysis.And cast shadow is object block light source after be incident upon shade in scene, the objects such as such as trees, vehicle, pedestrian are incident upon the shade on road surface, and this is the part of video scene but not moving target.Cast shadow in video can be divided into static shade and motion shade.Static shade is dissolved in background gradually in background modeling.Motion shade is shut out the light by the foreground target of movement in scene and is incident upon the dark areas in scene, along with target is moved together.

The motion shade being incident upon road surface due to vehicle has very similar motion feature to vehicle, and thus motion shade is usually mistaken for a part for foreground moving object.If do not eliminate this type games shade, the defect such as foreground target shape distortion, multiple moving target adhesions can be caused, the image processing operations such as follow-up vehicle detection, vehicle tracking, vehicle detection are had a huge impact, the link so detection moving vehicle shade is very important is one of hot subject of research always.

At present, the algorithm of shadow Detection can be divided into method, the type such as the method based on shading attributes feature and the method based on machine learning based on shadow model.

(1) based on the method for shadow model: mainly carry out modeling to video scene, comprise types such as source modeling and projecting direction modelings.In source modeling, one is suppose that pure white light is unique light source, shade is caused by linear attenuation due to light, and another kind is hypothesis light source is direct light (as sunshine) and the acting in conjunction that diffuses, and light is formed shade by non-linearly decaying.Such as Nadimi and Bhanu proposes a kind of double-colored model, considers two kinds of light sources to the impact of shadow region.In projecting direction modeling, Liu Zhifang etc. propose a kind of cast shadow direction model, Yuan Jiwei and Shi Zhongke further provides eight kinds of vehicle/shadow model on this basis.Method based on model needs to obtain the prior imformations such as illumination condition, moving target, scene in advance, and under numerous scenes such as background complexity, illuminance abrupt variation, the model set up exists the defect of adaptability aspect.

(2) based on the method for shading attributes feature: mainly utilize the information such as the brightness of shade, gradient, color, texture to identify shadow region.It has been generally acknowledged that the brightness step-down of shadow region (particularly penumbra region), and the change such as gradient, colourity, texture is less or ignore.The method of feature based has stronger robustness to different scenes and illumination condition, but the universality of algorithm still has much room for improvement.

(3) based on the method for machine learning: by sorter model such as structure support vector machine, neural network etc., set up the learning database of the feature such as texture, gray scale, and carry out training study to sorter model, then by the sorter trained to differentiate that whether dark areas is for shade.These class methods solve the aspect such as automatic learning, a pattern-recognition difficult problem, but the aspects such as the generalization ability of sorter, universality wait to improve.

In sum, each class methods still have certain limitation in process shadow problem, are necessary to improve shadow detection method.

Summary of the invention

Given this, the invention provides the detection method of moving vehicle shade in a kind of video image, to overcome the defect and deficiency that exist in existing method.

For achieving the above object, the invention provides a kind of moving vehicle shadow removing method, comprising:

(1) the image I of moving vehicle target is extracted _fwith background image I _bafter, be converted to YCbCr color space respectively, obtain image I _fwith luminance component Y, chrominance C b and the Cr component of background, then build the image I merging brightness, chromaticity information _fea, image I _feaits pixel value of middle each point is

$I_{\mathrm{fea}}(x,y)=\frac{\underset{(x,y)\∈D}{\mathrm{\Σ}}{Y}_F(x,y)}{\underset{(x,y)\∈D}{\mathrm{\Σ}}{Y}_B(x,y)}\·(1+\frac{\underset{(x,y)\∈D}{\mathrm{\Σ}}{\mathrm{Cb}}_F(x,y)-\underset{(x,y)\∈D}{\mathrm{\Σ}}{\mathrm{Cb}}_B(x,y)}{\underset{(x,y)\∈D}{\mathrm{\Σ}}[\mathrm{BW}(x,y)\·{\mathrm{Cb}}_B(x,y)]})\·(1+\frac{\underset{(x,y)\∈D}{\mathrm{\Σ}}{\mathrm{Cr}}_F(x,y)-\underset{(x,y)\∈D}{\mathrm{\Σ}}{\mathrm{Cr}}_B(x,y)}{\underset{(x,y)\∈D}{\mathrm{\Σ}}[\mathrm{BW}(x,y)\·{\mathrm{Cr}}_B(x,y)]})---\left(1\right)$

Wherein, the binaryzation foreground target image of BW for obtaining based on the moving target detecting method of background subtracting, D is the neighborhood of pixel P (x, y), then by image I _feathreshold value T is tried to achieve through maximum variance between clusters _fea, then by image I _feain be less than threshold value T _feapixel value vanishing, thus obtain image H _fea;

(2) to image I _fcarry out region segmentation, and computed image I _feach region and its background image between the absolute value summation of difference between chrominance C b and Cr, by the region D of its maximal value _seg(i) as the search initiation region of carbody, namely

$D_{\mathrm{seg}}\left(i\right)=\underset{F}{\arg \max }\left[\underset{(x,y)\∈\mathrm{Desg}}{\mathrm{\Σ}}\left(\right|{\mathrm{Cb}}_F(x,y)-{\mathrm{Cb}}_B(x,y)|+|{\mathrm{Cr}}_F(x,y)-{\mathrm{Cr}}_B(x,y)\left|\right)\right];$

(3) adopt boundary operator to image I _feacarry out rim detection, obtain edge image E, try to achieve the binary-state threshold T of edge image E according to maximum variance between clusters _edge, will threshold value T be less than _edgerim value vanishing, obtain the image I containing edge feature _edge, and by image H _feawith image I _edgesuperpose, form the characteristic image M containing edge, brightness and chrominance information, namely

M(x,y)＝I _edge+H _fea(x,y)

Wherein, $I_{\mathrm{edge}}(x,y)=\left\{\begin{array}{cc}E(x,y)& E(x,y)\&GreaterEqual;T_{\mathrm{edge}}\\ 0& E(x,y)<T_{\mathrm{edge}}\end{array}\right.;$

(4) structure suppresses the mask image of shadow region and background edge, first try to achieve the boundary B of binaryzation foreground target image BW, every bit P (the x all over border is gone through along boundary B, y), again by P (x, y) the pixel M (x, y) put in corresponding neighborhood L is set to zero, and the width calculation formula of neighborhood L is

L＝λ·arccot[α·(d _c-d _z)]

Wherein, λ is amplitude Dynamic gene, and α is rate of change Dynamic gene, d _cfor P (x, y) point is to the distance of characteristic image M centre of form C, d _zfor P (x, y) point is to characteristic image M barycenter Z distance, as L<0, L is set to zero;

(5) search initiation region D is obtained _segi the border b of (), finds out all and D along border b _segi r region that () is connected, forms set [D _seg(1), D _seg(2) ... D _seg(r)], get the region D that the pixel average of characteristic image M in this set is maximum _seg(j), then by D _seg(j) and D _segi () region merging technique forms larger region D _seg(i), namely

$D_{\mathrm{seg}}^k\left(i\right)=D_{\mathrm{seg}}^{k-1}\left(i\right)+D_{\mathrm{seg}}^{k-1}\left(j\right)---\left(6\right)$

Wherein, n is D _segm the number of pixel in () region, k is current iteration counting, iterates until the region D that obtains _segstop when () meets formula (7) i, i.e. region D _segi in the periphery subregion of (), no longer domain of the existence pixel average is greater than the subregion of threshold xi, threshold xi gets the region D that segment that in boundary B, neighborhood L width is greater than zero is touched _segthe characteristic image M pixel average of (j);

$\max \left\{\frac{1}{n}\underset{(x,y)\&Element;\mathrm{Desg}\left(m\right)}{\mathrm{\&Sigma;}}M(x,y)\right\}<\mathrm{\&xi;},m=r+1,r+2...r+k---\left(7\right)$

Region D _segi () other subregion untouched forms 2 set Q ₁=[D _seg(r+1), D _seg(r+2) ... D _seg] and Q (r+k) ₂=[D _seg(r+k+1), D _seg(r+k+2) ... D _seg(N)], Q ₁d _segthe adjacent subarea territory set of (i), Q ₂for not with D _segi regional ensemble that () is adjacent, if Q ₁and Q ₂being empty set, there is not shade in account for motion target, namely completes the testing process of whole shade candidate region; Otherwise, judge set Q ₂whether the area pixel mean value in middle region is greater than threshold xi, if be not more than, gathers Q ₁and Q ₂all elements composition region be the candidate regions of shade, if be greater than, illustrated that two moving targets there occurs adhesion, then will Q have been gathered ₂by formula (6), shadowing analysis is carried out to another target, until Q will be gathered when meeting (7) formula ₁and Q ₂all elements composition region be the candidate regions of shade;

(6) with shade candidate regions for starting point, adopt region-growing method to search the shadow region of shade candidate regions and periphery thereof, the criterion of region growing search is

S(j)＝{P(x,y):M(x,y)＜T _edge&|I _fea(x,y)-I _fea(x+d,y+d)|＜ζσ&BW(x,y)＝1}

Go through all over set Q ₁and Q ₂subregion representated by each element, in search subregion and neighboring area, obtains shade subregion S (r+1), S (r+2) ... S (N), all shade subregions obtain whole shadow region S of moving vehicle by logical "or" computing, namely

S＝D _seg(r+1)|D _seg(r+2)|…D _seg(N)|S(r+1)|S(r+2)|…S(N)。

Moving vehicle shadow detection method in described a kind of video image, in described step (1), as background image I _bluminance component Y, chrominance C b and Cr component in any channel components is less than normal when causing denominator to be zero, adopt the mean value of this channel components to substitute

The present invention considers the multicharacteristic informations such as brightness, colourity and edge gradient simultaneously, shadow region brightness is utilized to be the ratio decay of its background and the colourity change feature such as small, adopt the attenuation ratio value between the brightness of movement destination image and its background image, chromatic component, construct shadow character image.This characteristic image has been decayed the texture of shadow region effectively, and strengthens the texture of nonshaded area further.After image region segmentation, with the maximum region of colour difference for start point search vehicle body region, and untouched remaining area is considered as shade candidate regions.Again from each its neighboring area of shade candidate regions search, all shade subregions of final integration, obtain the whole shadow region of vehicle.Compared with existing shadow detection method, experimental result shows: the method proposed independently can differentiate whether there is shade, there is good shadow Detection performance and universality, reduce the impact of shade on subsequent video analysis, be applicable to the application such as target following, vehicle, vehicle Flow Detection that traffic video is analyzed.

Accompanying drawing explanation

Fig. 1 is a two field picture of the present invention's video sequence in an embodiment;

Fig. 2 is the background image I of the present invention's video scene in an embodiment _b;

Fig. 3 is that the present invention is in an embodiment containing the moving vehicle image I of shadow region _f;

Fig. 4 is the moving vehicle image I that brightness and chrominance information are merged in the present invention in an embodiment _fea;

Fig. 5 is the present invention's region, search starting point of moving vehicle car body and shade candidate regions in an embodiment;

Fig. 6 is the moving vehicle characteristic image M that gray scale, edge and colourity are merged in the present invention in an embodiment;

Fig. 7 is the mask image that the present invention suppresses shadow region marginal information in an embodiment;

Fig. 8 is the characteristic image M that the present invention eliminates shadow region marginal information in an embodiment;

Fig. 9 is the testing result of the present invention's moving vehicle shade in an embodiment.

Embodiment

Below in conjunction with the drawings and specific embodiments, the method for moving vehicle shadow Detection in video image of the present invention is set forth further.

In the present invention's video image in an embodiment, moving vehicle shadow detection method comprises the following steps.

Steps A. set up background model for traffic monitoring scene as shown in Figure 1.Background real-time update is obtained to the background of moving vehicle, as shown in Figure 2.The PBAS method in moving target detecting method is adopted to obtain moving vehicle target, as shown in Figure 3.

Step B. is by movement destination image I _fwith background image I _bbe converted to YCbCr color space from rgb color space, obtain the value of brightness Y, chrominance C b and Cr tri-components respectively, conversion formula is as follows

Y＝16+65.481*R+128.553*G+24.966*B

Cb＝128-37.797*R-74.203*G+112*B

Cr＝128+112*R-93.785*G-18.214*B

Step C. design of graphics is as I _fea, reach and cut down shadow region texture and the object strengthening car body texture, as shown in Figure 4.Image I _feain each calculated for pixel values formula such as formula shown in (1)

$I_{\mathrm{fea}}(x,y)=\frac{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{Y}_F(x,y)}{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{Y}_B(x,y)}\&CenterDot;(1+\frac{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cb}}_F(x,y)-\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cb}}_B(x,y)}{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}[\mathrm{BW}(x,y)\&CenterDot;{\mathrm{Cb}}_B(x,y)]})\&CenterDot;(1+\frac{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cr}}_F(x,y)-\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cr}}_B(x,y)}{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}[\mathrm{BW}(x,y)\&CenterDot;{\mathrm{Cr}}_B(x,y)]})---\left(1\right)$

Wherein, BW is the binary image of foreground target, and D is the neighborhood of pixel P (x, y), and in the present embodiment, Size of Neighborhood D value gets 3, i.e. 8 neighborhoods.When denominator is 0, adopt background image I respectively _bbrightness Y, chrominance C b and Cr mean value replace corresponding denominator.

Image I is tried to achieve according to maximum variance between clusters _feabinary-state threshold T _fea, in view of being less than threshold value T _feapixel normally shade or black car body cause, for reducing the impact that this kind of pixel is searched for car body, therefore, will threshold value T be less than _feavalue become 0, obtain image H _fea, namely

$H_{\mathrm{fea}}(x,y)=\left\{\begin{array}{cc}{I}_{\mathrm{fea}}(x,y)& I_{\mathrm{fea}}(x,y)\&GreaterEqual;T_{\mathrm{fea}}\\ 0& I_{\mathrm{fea}}(x,y)<T_{\mathrm{fea}}\end{array}\right.$

Step D. adopts Meanshift algorithm to carry out image I _fsplit some subregions and form set D _seg(N), image I is got according to formula (2) _fchrominance C b, Cr component differ maximum region D with background image _segi (), as the region, search starting point of carbody, as shown in Figure 5, the region that wherein color is the darkest is region, search starting point, namely

$D_{\mathrm{seg}}\left(i\right)=\underset{F}{\arg \max }\left[\underset{(x,y)\&Element;\mathrm{Desg}}{\mathrm{\&Sigma;}}\left(\right|{\mathrm{Cb}}_F(x,y)-{\mathrm{Cb}}_B(x,y)|+|{\mathrm{Cr}}_F(x,y)-{\mathrm{Cr}}_B(x,y)\left|\right)\right];$

Step e. adopt boundary operator to image I _feacarry out rim detection, adopt Sobel boundary operator to obtain edge image E in the present embodiment, its computing formula is

$E(x,y)=\left(\right|G_x|+|G_y\left|\right)(1+\frac{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cb}}_F(x,y)-\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cb}}_B(x,y)}{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}[\mathrm{BW}(x,y)\&CenterDot;{\mathrm{Cb}}_B(x,y)]})\&CenterDot;(1+\frac{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cr}}_F(x,y)-\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cr}}_B(x,y)}{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}[\mathrm{BW}(x,y)\&CenterDot;{\mathrm{Cr}}_B(x,y)]})$

Wherein, G _x=I _fea(x-1, y-1)+2I _fea(x, y-1)+I _fea(x+1, y-1)-I _fea(x-1, y+1)-2I _fea(x, y+1)-I _fea(x+1, y+1),

G _y＝I _fea(x-1,y-1)+2I _fea(x-1,y)+I _fea(x-1,y+1)-I _fea(x+1,y-1)-2I _fea(x+1,y)-I _fea(x+1,y+1)

Maximum variance between clusters is adopted to try to achieve the binary-state threshold T of edge image E _edge, will threshold value T be less than _edgerim value become 0, to reduce the interference that slight edge causes, namely

$I_{\mathrm{edge}}(x,y)=\left\{\begin{array}{cc}E(x,y)& E(x,y)\&GreaterEqual;T_{\mathrm{edge}}\\ 0& E(x,y)<T_{\mathrm{edge}}\end{array}\right.---\left(3\right)$

Then, by image H _feawith edge image I _edgesuperpose, obtain the characteristic image M having merged edge gradient, brightness and colourity, the value of each pixel is

M(x,y)＝I _edge+H _fea(x,y) (4)

The characteristic image M obtained in the present embodiment as shown in Figure 6.

Step F. obtain binaryzation foreground target image BW, calculate the centre of form C (x of binaryzation BW _c, y _c), centroid calculation formula is

$x_c=\frac{1}{n}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}i\&CenterDot;\mathrm{BW}(i,j),y_c=\frac{1}{n}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}j\&CenterDot;\mathrm{BW}(i,j)$

Wherein $n=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\mathrm{BW}(i,j)$

Due to the marginal information skewness of characteristic image M, show as that car body area marginal information is enriched and the marginal information in shadow region is very rare, therefore characteristic image M barycenter is closer to car body, its barycenter Z (x _z, y _z) computing formula be

$x_z=\frac{1}{n}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}i\&CenterDot;I_{\mathrm{edge}}(i,j)\&CenterDot;\mathrm{BW}(i,j),y_z=\frac{1}{n}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}j\&CenterDot;I_{\mathrm{edge}}(i,j)\&CenterDot;\mathrm{BW}(i,j)$

Wherein $n=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\mathrm{BW}(i,j)$

In order to cut down the edge in shadow region and background area, build the mask image suppressing shadow region marginal information, as shown in Figure 7.In order to retain car body marginal information, the width of neighborhood L is variable, and time in shadow region, L value is comparatively large, and L value is less when nonshaded area.If the distance that any point in the boundary B of binaryzation foreground image BW is P (x, y) to centre of form C is d _c, P (x, y) point is d to barycenter Z distance _z, then the width of neighborhood L is

L＝λ·arccot[α·(d _c-d _z)] (5)

In the present embodiment, as L<0, putting L is 0; λ is amplitude Dynamic gene, and in the present embodiment, λ value is 1; α is rate of change Dynamic gene, and in the present embodiment, α value is 1.The effect of this mask image is that the pixel M (x, y) in P (x, y) vertex neighborhood L is set to 0.To characteristic image M after the operation of mask image, restrained effectively the marginal information in shadow region, as shown in Figure 8.

Step G. obtains region, search starting point D _segi the border b of (), finds out all and D along border b _segi region that () is connected, supposes that region quantity is r, forms set [D _seg(1), D _seg(1) ... D _seg(r)], calculate the pixel average in each region of characteristic image M, average maximum D _seg(j) region and original D _segi () merges, namely

$D_{\mathrm{seg}}^k\left(i\right)=D_{\mathrm{seg}}^{k-1}\left(i\right)+D_{\mathrm{seg}}^{k-1}\left(j\right)---\left(6\right)$

Wherein, n is D _segm the number of pixel in () region, k is current iteration counting.

Then loop iteration tries to achieve region D _seg(i), until termination of iterations when meeting formula (7), namely in characteristic image M with D _segin i other subregion that () is adjacent, no longer domain of the existence pixel characteristic mean value is all less than the subregion of threshold xi, threshold xi be approximately equal to zero decimal, the value of threshold xi of adjusting according to the difference of video scene.

$\max \left\{\frac{1}{n}\underset{(x,y)\&Element;\mathrm{Desg}\left(m\right)}{\mathrm{\&Sigma;}}M(x,y)\right\}<\mathrm{\&xi;},m=r+1,r+2...r+k---\left(7\right)$

Region D _segi () other subregion untouched forms 2 set Q ₁=[D _seg(r+1), D _seg(r+2) ... D _seg] and Q (r+k) ₂=[D _seg(r+k+1), D _seg(r+k+2) ... D _seg(N)], Q ₁d _segthe periphery subregion set of (i), Q ₂be not with D _segi regional ensemble that () is adjacent.

Q in the present embodiment ₁and Q ₂set is non-NULL, and account for motion target exists shade.Again due to set Q ₂middle area pixel mean value is all less than threshold xi, and the shade only having a moving target is described.Then, Q will be gathered ₁and Q ₂all elements be considered as the candidate regions of shade.The mainly shadow region, the region involved by segment that in boundary B, neighborhood L width value is greater than zero, as shown in Figure 6 and Figure 7, therefore, threshold xi gets the characteristic image M pixel average that neighborhood L width value is greater than the region involved by segment of zero, and in the present embodiment, threshold xi is 0.1.D after iteration ends _segi () as the gray area in Fig. 5, i.e. carbody part, the white space be not colored is then shade candidate region.

Step H. is from Q ₁and Q ₂each element [D of set _seg(r+1), D _seg(r+2) ... D _seg(N) subregion] sets out, and adopts region-growing method at characteristic image I _feasearch shadow region S (j) in (x, y), the criterion of region growing calculates according to formula (8).

S(j)＝{P(x,y):M(x,y)＜T _edge&|I _fea(x,y)-I _fea(x+d,y+d)|＜ζσ&BW(x,y)＝1} (8)

Obtain shade subregion S (r+1), S (r+2) ... after S (N), all subregions are obtained whole shade subregion S of moving vehicle by logical "or" computing, namely

S＝D _seg(r+1)|D _seg(r+2)|…D _seg(N)|S(r+1)|S(r+2)|…S(N)

In the present embodiment, as shown in Figure 9, its grey area is the shade detected to whole shade subregion S, and white portion is the car body of vehicle.As can be seen from Figure 9, the present invention detects the effective of shade from moving target.

Claims (2)

1. a moving vehicle shadow detection method in video image, is characterized in that, comprise following steps:

(1) the image I of moving vehicle target is extracted _fwith background image I _bafter, be converted to YCbCr color space respectively, obtain image I _fwith luminance component Y, chrominance C b and the Cr component of background, then build the image I merging brightness, chromaticity information _fea, image I _feaits pixel value of middle each point is

$I_{\mathrm{fea}}(x,y)=\frac{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{Y}_F(x,y)}{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{Y}_B(x,y)}\&CenterDot;(1+\frac{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cb}}_F(x,y)-\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cb}}_B(x,y)}{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}[\mathrm{BW}(x,y)\&CenterDot;{\mathrm{Cb}}_B(x,y)]})\&CenterDot;(1+\frac{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cr}}_F(x,y)-\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}{\mathrm{Cr}}_B(x,y)}{\underset{(x,y)\&Element;D}{\mathrm{\&Sigma;}}[\mathrm{BW}(x,y)\&CenterDot;{\mathrm{Cr}}_B(x,y)]})---\left(1\right)$

Wherein, the binaryzation foreground target image of BW for obtaining based on the moving target detecting method of background subtracting, D is the neighborhood of pixel P (x, y), then by image I _feathreshold value T is tried to achieve through maximum variance between clusters _fea, then by image I _feain be less than threshold value T _feapixel value vanishing, thus obtain image H _fea;

(2) to image I _fcarry out region segmentation, and computed image I _feach region and its background image between the absolute value summation of difference between chrominance C b and Cr, by the region D of its maximal value _seg(i) as the search initiation region of carbody, namely

$D_{\mathrm{seg}}\left(i\right)=\underset{F}{\arg \max }\left[\underset{(x,y)\&Element;\mathrm{Dseg}}{\mathrm{\&Sigma;}}\left(\right|{\mathrm{Cb}}_F(x,y)-{\mathrm{Cb}}_B(x,y)|+|{\mathrm{Cr}}_F(x,y)-{\mathrm{Cr}}_B(x,y)\left|\right)\right];$

(3) adopt boundary operator to image I _feacarry out rim detection, obtain edge image E, try to achieve the binary-state threshold T of edge image E according to maximum variance between clusters _edge, will threshold value T be less than _edgerim value vanishing, obtain the image I containing edge feature _edge, and by image H _feawith image I _edgesuperpose, form the characteristic image M containing edge, brightness and chrominance information, namely

M(x,y)＝I _edge+H _fea(x,y)

Wherein, $I_{\mathrm{edge}}(x,y)=\left\{\begin{array}{cc}E(x,y)& E(x,y)\&GreaterEqual;T_{\mathrm{edge}}\\ 0& E(x,y)<T_{\mathrm{edge}}\end{array}\right.;$

(4) structure suppresses the mask image of shadow region and background edge, first try to achieve the boundary B of binaryzation foreground target image BW, every bit P (the x all over border is gone through along boundary B, y), again by P (x, y) the pixel M (x, y) put in corresponding neighborhood L is set to zero, and the width calculation formula of neighborhood L is

L＝λ·arccot[α·(d _c-d _z)]

Wherein, λ is amplitude Dynamic gene, and α is rate of change Dynamic gene, d _cfor P (x, y) point is to the distance of characteristic image M centre of form C, d _zfor P (x, y) point is to characteristic image M barycenter Z distance, as L<0, L is set to zero;

(5) search initiation region D is obtained _segi the border b of (), finds out all and D along border b _segi r region that () is connected, forms set [D _seg(1), D _seg(2) ... D _seg(r)], get the region D that the pixel average of characteristic image M in this set is maximum _seg(j), then by D _seg(j) and D _segi () region merging technique forms larger region D _seg(i), namely

$D_{\mathrm{seg}}^k\left(i\right)=D_{\mathrm{seg}}^{k-1}\left(i\right)+D_{\mathrm{seg}}^{k-1}\left(j\right)---\left(6\right)$

Wherein, n is D _segm the number of pixel in () region, k is current iteration counting, iterates until the region D that obtains _segstop when () meets formula (7) i, i.e. region D _segi in the periphery subregion of (), no longer domain of the existence pixel average is greater than the subregion of threshold xi, threshold xi gets the region D that segment that in boundary B, neighborhood L width is greater than zero is touched _segthe characteristic image M pixel average of (j);

$\max \left\{\frac{1}{n}\underset{(x,y)\&Element;\mathrm{Dseg}\left(m\right)}{\mathrm{\&Sigma;}}M(x,y)\right\}<\mathrm{\&xi;},m=r+1,r+2...r+k---\left(7\right)$

Region D _segi () other subregion untouched forms 2 set Q ₁=[D _seg(r+1), D _seg(r+2) ... D _seg] and Q (r+k) ₂=[D _seg(r+k+1), D _seg(r+k+2) ... D _seg(N)], Q ₁d _segthe adjacent subarea territory set of (i), Q ₂for not with D _segi regional ensemble that () is adjacent, if Q ₁and Q ₂being empty set, there is not shade in account for motion target, namely completes the testing process of whole shade candidate region; Otherwise, judge set Q ₂whether the area pixel mean value in middle region is greater than threshold xi, if be not more than, gathers Q ₁and Q ₂all elements composition region be the candidate regions of shade, if be greater than, illustrated that two moving targets there occurs adhesion, then will Q have been gathered ₂by formula (6), shadowing analysis is carried out to another target, until Q will be gathered when meeting (7) formula ₁and Q ₂all elements composition region be the candidate regions of shade;

(6) with shade candidate regions for starting point, adopt region-growing method to search the shadow region of shade candidate regions and periphery thereof, the criterion of region growing search is

S(j)＝{P(x,y):M(x,y)＜T _edge&|I _fea(x,y)-I _fea(x+d,y+d)|＜ζσ&BW(x,y)＝1}

Go through all over set Q ₁and Q ₂subregion representated by each element, in search subregion and neighboring area, obtains shade subregion S (r+1), S (r+2) ... S (N), all shade subregions obtain whole shadow region S of moving vehicle by logical "or" computing, namely

S＝D _seg(r+1)|D _seg(r+2)|…D _seg(N)|S(r+1)|S(r+2)|…S(N)。

2. moving vehicle shadow detection method in a kind of video image according to claim 1, is characterized in that, in described step (1), as background image I _bluminance component Y, chrominance C b and Cr component in any channel components is less than normal when causing denominator to be zero, adopt the mean value of this channel components to substitute.