CN102298781B - Motion shadow detection method based on color and gradient characteristics - Google Patents

Abstract

The invention provides a motion shadow detection method based on color and gradient characteristics. The method is characterized by: collecting several frames of initial images, and establishing a background model by using a mean value displacement method so as to obtain a background image; taking a current obtained video stream image as an input image which is used to be different with the background image, detecting a moving object and determining a moving object area; carrying out shadow detections to the moving object area by using methods based on a color characteristic and a gradient characteristic respectively; carrying out an AND operation and a morphologic process to the shadow areas obtained from the two methods so as to determine the final shadow area; real-timely updating the background model by using the current input image. By using the method of the invention, a background model accuracy is high. The color characteristic and a textural feature are combined to distinguish the moving object and the shadow. Interferences caused by noise and other factors to the shadow can be reduced so as to improve a shadow detection accuracy.

Description

Based on the Moving Shadow Detection Approach of color and gradient feature

Technical field

The technology of the present invention relates to a kind of Moving Shadow Detection Approach based on color and gradient feature.

Technical background

Such as, to being the basic task of many machine vision and video analysis application cutting apart of moving object in video sequence, video monitoring, human body detection and tracking, multimedia index and Video coding etc.Accurate moving object is cut apart also and will greatly be improved the performance of target following, identification, classification and motion analysis.But in moving object detection and the process cut apart, due to light irradiating object, can make moving target produce corresponding shade.The shade producing can move along with the motion of moving target, in the time carrying out target detection, if shade is not processed, and probably can be using shade as moving object detection out, like this, tracking, identification to the later stage have caused very large interference.

At present, shadow detection method is mainly divided into based on model or based on two kinds of features.Method based on model utilizes the geometric properties of scene, illumination and target to set up model, and as the three-dimensional structure of scene, the description of foreground object available model etc., but it is only applicable to some special scenes, and method limitation is large.The method calculation of complex in addition, is unsuitable for real-time application.On the other hand, the method based on feature is utilized color, gradient, the Texture eigenvalue identification shade of image.These class methods are generally first to detect sport foreground region, are more further divided into object and shade.But existing method is usually only considered the difference of moving target and the single feature of shade, causes detecting accuracy not high.If a kind of method is for adopting hsv color space, think that shade has reduced the brightness value of institute overlay area, but colourity and intensity value change within the specific limits, the shortcoming of the method is comparatively responsive to illumination variation, if the partial pixel of object has similar color characteristic with shade, object can be mistaken for to shade.Also have utilize the joint probability density of multiple features carry out modeling again with the method for threshold, but its data calculated amount is large, expends time in many, affects the real-time of shadow Detection, is difficult to be applied to real-time occasion.

Summary of the invention

The object of the invention is to propose a kind of Moving Shadow Detection Approach based on color and gradient feature, the method can adapt to various complex scenes, improves accuracy and the real-time of shadow Detection.

Technical solution of the present invention is as follows:

Based on a Moving Shadow Detection Approach for color and gradient feature, it is characterized in that, comprise the following steps:

Step 1: gather t frame initial pictures, set up background model, obtain background image; ([wherein the span of t is 50～300 frames, and t preferably gets 100 frames]

Step 2: using current obtained video streaming image as input picture and background image difference, detect moving target, determine motion target area;

Step 3: adopt respectively based on color characteristic and the method based on Gradient Features motion target area is carried out to shadow Detection;

Step 4: the shadow region that two kinds of methods are obtained is carried out and operated and do morphology processing, determines final shadow region;

Step 5: utilize current input image to upgrade background model, turn back to step 2.

Described step 1 is:

The front t two field picture reference picture of Image estimation as a setting that utilizes current input image, utilizes this t two field picture to set up background model;

If x ₁, x ₂..., x _tfor the pixel set of certain same position in t two field picture, them, image is at a group observations sample of this point as a setting, and the estimated value method that adopts mean shift method to obtain background image this point is:

$m_h\left(x\right)=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}G\left(\frac{x_i-x}{h}\right)w\left(x_i\right)x_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}G\left(\frac{x_i-x}{h}\right)w\left(x_i\right)}$

The initial value of given x is the arbitrary value in sample set, kernel function

allowable error ε, adopts mean shift method cyclically to carry out three steps below, until termination condition meets, wherein h is the size [span 10～50 of h, value is 20 conventionally] of mean shift method search window, and the number that n is sample point is t, w (x _i) be one and be assigned to sampled point x _iweight, span is nonnegative number, (conventionally value be 1)

The first step: calculate m _h(x);

Second step: if || m _h(x)-x|| < ε, end loop;

The 3rd step: m _h(x) be assigned to x, and continue to carry out the first step; Obtain m _h(x) be the solution of background model, background image is at the best estimate of this point;

Described step 2 is: in front input picture and background, in the RGB component difference of corresponding pixel points, select a maximal value, be labeled as DV (x, y), in the time that DV (x, y) is greater than threshold value TL, judgement (x, y) pixel is motion pixel, is labeled as g (x, y)=1, otherwise mark g (x, y)=0; For the image of M*N size, the computing formula of TL is as follows

$\mathrm{TL}=\min (\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{I}_R(x_i,y_j),\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{I}_G(x_i,y_j),\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{I}_B(x_i,y_j))---\left(2\right)$

Wherein I _r(x _i, y _j) be pixel (x _i, y _j) the value of R passage, I _g(x _i, y _j) be pixel (x _i, y _j) the value of G passage, I _b(x _i, y _j) be pixel (x _i, y _j) the value of B passage;

Then use morphological method closed operation to carry out filtering processing to g (x, y) bianry image, eliminate the isolated point being caused by noise.

Described step 3 is:

Step a: utilize Sobel gradient operator to carry out rim detection to the RGB triple channel of original image moving region, and three-channel RGB Grad is had to the point of consistent change direction as the point of candidate shadow region, remember that now the set of point is S _c, obviously S _cin hypographous point not only, also may have the point of the lower object of brightness;

Step b: image conversion is arrived to C ₁c ₂c ₃color space, then the imagery exploitation Sobel gradient operator after conversion is carried out to rim detection; Due to C ₁c ₂c ₃color space is to illumination-insensitive, the point S set therefore now obtaining _min, only comprise the point of object;

C ₁c ₂c ₃color space definition is

$C_1=\arctan \left(\frac{R}{\max (G,B)}\right)$

$C_2=\arctan \left(\frac{G}{\max (R,B)}\right)$

$C_3=\arctan \left(\frac{B}{\max (R,G)}\right)$

C ₁c ₂c ₃be a kind of color character invariant, what it was described is the function of the color scheme of pixel, and the feature invariant amount of pixel is not subject to the impact of the factors such as visual angle, shade, surface direction and illumination condition.

Step c: comprehensive above step a and b, the point S set of shadow region _s=S _c-S _mthereby, the point S set of acquisition shadow region _sin the profile forming S set a little _ms1;

The described method based on Gradient Features is that the gradient vector figure after structure normalization also compares the gradient vector figure of prospect and background, thereby distinguishes the shadow region of moving target and its generation.Concrete algorithm is as follows:

Steps A: calculate the gradient vector of background image, along image level direction, carry out convolution with Soble operator and the background image of 0 °, 45 °, 90 °, 135 ° 4 directions successively, utilize formula 6 to calculate 4 yardstick g of each pixel (x, y) _{f, 1}(x, y), g _{f, 2}(x, y), g _{f, 3}(x, y) and g _{f, 4}(x, y), obtains scaling vector g _f(x, y)={ g _{f, 1}, g _{f, 2}, g _{f, 3}, g _{f, 4}| _{x, y}, g _f(x, y) reflected the trend of background image in this pixel graded, is also the form of expression of background gradient:

g _f，i(x，y)＝f(x，y)*H _i(x，y) i＝1，2，3，4

Wherein f (x, y) is the pixel value on coordinate (x, y), H _i(x, y) is 0 °, 45 °, 90 °, the 135 ° Soble operators in 4 directions, is respectively:

$\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right],$ $\left[\begin{array}{ccc}0& 1& 2\\ -1& 0& 1\\ -2& -1& 0\end{array}\right],$ $\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right],$ $\left[\begin{array}{ccc}2& 1& 0\\ 1& 0& -1\\ 0& -1& -2\end{array}\right];$

The normalization of step B background image gradient vector:

According to formula

i=1,2,3,4 are normalized the gradient vector of each pixel of background, obtain the gradient vector after normalization

and normalized gradient vector is stored, storehouse (the gradient vector comparison of context vault and present image hereinafter) as a setting, according to gradient direction principle of invariance, the gradient vector after normalization is a build-in attribute of background, under level of illumination visible, does not change with illumination.

Step C: calculate the gradient vector of current frame image, and normalization: along current frame image horizontal direction, carry out convolution with Soble operator and the sequence image of 4 directions successively, calculate 4 yardstick g of each pixel (x, y) _{c, 1}(x, y), g _{c, 2}(x, y), g _{c, 3}(x, y) and g _{c, 4}(x, y), obtains scaling vector g _c(x, y)={ g _{c, 1}, g _{c, 2}, g _{c, 3}, g _{c, 4}| _{x, y}, g _c(x, y) reflected the trend of current frame image in this pixel graded, utilizes formula:

$g_i^{*}=\frac{g_i}{\underset{j=1}{\overset{4}{\mathrm{\&Sigma;}}}\left|g_i\right|},$ i＝1，2，3，4

Be normalized, obtained vector $g_c^{*}(x,y)=\{g_{c,1}^{*},g_{c,2}^{*},g_{c,3}^{*},g_{c,4}^{*}{|}_{x,y}\};$

Step D: the difference d that calculates the normalized gradient vector of pixel corresponding in current frame image pixel and context vault:

$d=\left|\right|g_f^{*}(x,y)-g_c^{*}(x,y)\left|\right|=\underset{i=1}{\overset{4}{\mathrm{\&Sigma;}}}|g_{c,i}^{*}-g_{f,i}^{*}|$

Template is cut apart in employing

threshold value Δ ε=0.3 splits motion shade from motion target area, cutting apart in template, be denoted as 1 be the point of motion shadow region;

Step e: scanning entire image, repeating step D, until the normalized gradient vector of entire image and background image normalized gradient vector is more complete, realizes motion shade cutting apart from motion target area, and motion shadow region point is gathered and is labeled as S _ms2.

In described step 4, morphology processing procedure is:

For further improving moving shadow detection result, to the S obtaining _ms1region and S _ms2aND-operation and OR operation are carried out in region, obtain "AND" figure and "or" figure.Then, utilize connected component analysis to obtain each connected region, be called an agglomerate.In "AND" figure, obtain the center of mass point of each agglomerate, then judge which agglomerate in "or" figure contains this center of mass point, if contained, retain this agglomerate in "or" figure. finally the agglomerate of the center of mass point not containing in "AND" figure and contain in "or" figure is deleted, finally the result of "or" figure is carried out to closing operation of mathematical morphology again one time, obtain final moving target.

The described Moving Shadow Detection Approach based on color and gradient feature, it is characterized in that, the described wherein Moving Shadow Detection Approach based on color and gradient feature, only the pixel of the moving region on present image carries out shadow Detection, to improve arithmetic speed.

Beneficial effect:

Compared with prior art, beneficial effect of the present invention is as follows:

First, proposed the background model based on mean-shift algorithm, can extract background image to having in scene in the interference of moving target and the situation of light variation, background model accuracy is high.

Secondly, adopt and respectively shadow Detection is carried out in moving region based on color invariance and the method based on gradient, again the shade detecting is carried out to morphology processing, reduced noise and the interference of other factors to shade, thereby improved the correctness of shadow Detection.

Again, in last handling process, two shadow regions are carried out respectively the operation of "AND" and "or" and determined shade connected region after the OR operation shadow region as final detection by the barycenter of shade connected region after AND-operation, further reduced false drop rate.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is background modeling process flow diagram;

Fig. 3 is the shadow region testing process based on color characteristic;

Fig. 4 is the shadow region testing process based on tonsure feature.

Embodiment

Below in conjunction with accompanying drawing and specific implementation process, the present invention is described in further details, but protection scope of the present invention is not limited to following embodiment.

Embodiment 1:

Moving Shadow Detection Approach based on color and gradient feature as shown in Figure 1, comprises the steps:

1. set up background model

Obtain the color characteristic information of the pixel in the front N two field picture of current input image, start to detect target from the N+1 frame image of input: present frame image pixel point is as estimation point, according to mean_shift algorithm, obtain the probable value that estimation point belongs to background model, and current frame picture point is upgraded to background model as new sampled point, as shown in Figure 2.Specifically be divided into again following steps:

1. by the pixel value of each pixel (x, y) of front N two field picture as a sample space, and be divided into the class for q

C＝{{c _i，w _i}}，i＝1，…，q (1)

Wherein c _ithe eigenwert of pixel, w _iit is all kinds of weights.

2. calculate the weight of each classification

$w_i=\frac{l_i}{m},i=1,\&CenterDot;\&CenterDot;\&CenterDot;,q---\left(2\right)$

Wherein l _ifor the number of Different categories of samples point, the sum that m is sample point

3. utilizing mean-shift (mean shift) algorithm to obtain reliable background model estimates

$\hat{C}=c_{i^{*}},$ i ^*＝arg max{w _i} (3)

Wherein

for the estimated value of the most reliable background model pixel

Mean-shift (mean shift method) described above obtains reliable background model method of estimation and is

$m_h\left(x\right)=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}G\left(\frac{x_i-x}{h}\right)w\left(x_i\right)x_i}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}G\left(\frac{x_i-x}{h}\right)w\left(x_i\right)}---\left(4\right)$

A given initial point x, kernel function G (X), allowable error ε, execution three steps below of Mean Shift algorithm circulation, until termination condition is satisfied,

The first step. calculate m _h(x);

If second step. || m _h(x)-x|| < ε, end loop;

The 3rd step. m _h(x) be assigned to x, and continue to carry out the first step.

Obtain m _h(x) be the best estimate of background model pixel value.

2, background difference

In calculating current input image and background, in the RGB component difference of corresponding pixel points, select a maximal value, be labeled as DV (x, y) suc as formula (5), in the time that DV (x, y) is greater than threshold value TL, judgement (x, y) pixel is motion pixel, is labeled as g (x, y)=1, otherwise mark g (x, y)=0; Then use morphological method to carry out filtering processing to g (x, y) bianry image, remove the isolated area that area is less, non-connected region simultaneously.

DV(x，y)＝max(|I(x，y).c-B(x，y).c|).c＝R，G，B (5)

For the image of M*N size, TL is defined as follows

$\mathrm{TL}=\min (\frac{1}{\mathrm{MN}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}I(x_i,y_j).c),$ .c＝R，G，B (6)

3, the shadow Detection based on color and gradient feature

Be not subject to the impact of the factors such as visual angle, shade, surface direction and illumination condition according to color character invariant.Color character invariant can obtain from the RGB triple channel value transform of pixel, and conventional color character invariant has normalized rgb color space and C ₁c ₂c ₃color space.Wherein C ₁c ₂c ₃color space is:

$C_1=\arctan \left(\frac{R}{\max (G,B)}\right)$

$C_2=\mathrm{atc}\tan \left(\frac{G}{\max (R,B)}\right)---\left(7\right)$

$C_3=\arctan \left(\frac{B}{\max (R,G)}\right)$

Concrete grammar flow process as shown in Figure 3.

Step 1, according to the low feature in brightness ratio peripheral region, shadow region, is designated as candidate shadow region region mark one low brightness.Exactly rim detection is carried out in original image moving region specifically, the three-channel gradient of RGB has the point of consistent descent direction, just thinks the point of candidate shadow region.Remember that now the set of point is S _c, obviously S _cin hypographous point not only, also may have the point of the lower object of brightness.

Step 2, first arrives C by image conversion ₁c ₂c ₃color space, then the image after conversion is carried out to rim detection.The similar step 1 of operating process.Due to C ₁c ₂c ₃color space is to illumination-insensitive, the point S set therefore now obtaining _min, only comprise the point of object.

Step 3, comprehensive above two steps, the point S set of shadow region _s=S _c-S _m.

Shadow detection method based on Gradient Features is, the gradient vector figure after structure normalization also compares the gradient vector figure of prospect and background, thereby the shadow region of distinguishing moving target and its generation as shown in Figure 4.Concrete algorithm is as follows:

Step 1 is calculated the gradient vector of background image, along image level direction, carries out convolution successively with Soble operator and the background image of 4 directions, utilizes formula (8) to calculate 4 yardstick g of each pixel (x, y) _{f, 1}(x, y), g _{f, 2}(x, y), g _{f, 3}(x, y) and g _{f, 4}(x, y) obtains scaling vector g _f(x, y)={ g _{f, 1}, g _{f, 2}, g _{f, 3}, g _{f, 4}| _{x, y}, g _f(x, y) reflected the trend of background image in this pixel graded, is also the form of expression of road surface background texture.

g _i(x，y)＝f(x，y)*H _i(x，y) i＝1，2，3，4 (8)

Wherein H _i(x, y) is 4 Soble operators in direction.

The normalization of step 2 background image gradient vector utilizes formula

$g_i^{*}=\frac{g_i}{\underset{j=1}{\overset{4}{\mathrm{\&Sigma;}}}\left|g_i\right|}\text{,}$ i＝1，2，3，4 (9)

The gradient vector of each pixel of background is normalized, obtains the gradient vector after normalization

and normalized gradient vector is stored as to background dictionary according to gradient direction principle of invariance, the gradient vector after normalization is a build-in attribute of background, under level of illumination visible, does not change with illumination.

Step 3 is calculated the gradient vector of current frame image, and normalization. and in like manner, the process of repeating step 1, along image level direction, carries out convolution with Soble operator and the sequence image of 4 directions successively, calculates 4 yardstick g of each pixel (x, y) _{c, 1}(x, y), g _{c, 2}(x, y), g _{c, 3}(x, y) and g _{c, 4}(x, y), obtains scaling vector g _c(x, y)={ g _{c, 1}, g _{c, 2}, g _{c, 3}, g _{c, 4}| _{x, y}.G _c(x, y) reflected the trend of current frame image in this pixel graded. be normalized, obtained vector

Step 4, according to formula (8), compares the normalized gradient vector of the pixel in current frame image with the normalized gradient vector of corresponding pixel in background dictionary storehouse:

$d=\left|\right|g_f^{*}(x,y)-g_c^{*}(x,y)\left|\right|=\underset{i=1}{\overset{4}{\mathrm{\&Sigma;}}}|g_{c,i}^{*}-g_{f,i}^{*}|---\left(10\right)$

Making the template of cutting apart of moving target is F (x, y), owing to being subject to external environment and electromagnetic interference (EMI), often

therefore,, in order to reduce the impact of this interference, need setting threshold Δ ε to adjudicate background and moving target

$F(x,y)=\{\frac{\begin{array}{cc}0,& d>\mathrm{\&Delta;\&epsiv;}\end{array}}{\begin{array}{cc}1,& d\&le;\mathrm{\&Delta;\&epsiv;}\end{array}}---\left(11\right)$

Step 5 scans entire image. repeating step 4, until complete the comparison of entire image and background image normalized gradient vector. cutting apart in template, being denoted as 1 region is background shadow region, road surface, and being denoted as 0 region is motion target area. so just motion shade is cut apart out from motion target area.

Illustrate: the described Moving Shadow Detection Approach based on color and gradient feature, only the pixel of the moving region on foreground image carries out shadow Detection, to improve arithmetic speed.

4, the morphology processing of shadow region

Carry out AND-operation and OR operation to what obtain based on color characteristic and the shadow Detection result based on tonsure respectively, obtain "AND" figure and "or" figure.Then, utilize connected component analysis to obtain each connected region, be called an agglomerate. in "AND" figure, obtain the center of mass point of each agglomerate, then which agglomerate in judgement or figure contains this center of mass point, if contained, retain this agglomerate in "or" figure. finally by the not agglomerate deletion containing center of mass point in "AND" figure in "or" figure, the result of "or" figure is carried out to a morphology again and process and connected component analysis, determine final moving target.

Claims (3)

1. the Moving Shadow Detection Approach based on color and gradient feature, is characterized in that, comprises the following steps:

Step 1: gather t frame initial pictures, set up background model, obtain background image;

Step 2: using current obtained video streaming image as input picture and background image difference, detect moving target, determine motion target area;

Step 3: adopt respectively based on color characteristic and the method based on Gradient Features motion target area is carried out to shadow Detection;

Step 4: the shadow region that two kinds of methods are obtained carry out with operation and or operate and do morphology processing, determine final shadow region;

Step 5: judge whether to finish, if not, utilize current input image to upgrade background image, turn back to step 2; If so, stop;

Described step 1 is:

The front t two field picture reference picture of Image estimation as a setting that utilizes current input image, utilizes this t two field picture to set up background model;

If x ₁, x ₂..., x _tfor the pixel set of certain same position in t two field picture, them, image is at a group observations sample of this point as a setting, and the estimated value method that adopts mean shift method to obtain background image this point is:

$m_h\left(x\right)=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}G\left(\frac{x_i-x}{h}\right)w\left(x_i\right)x_i}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}G\left(\frac{x_i-x}{h}\right)w\left(x_i\right)},$

The initial value of given x is the arbitrary value in sample set, kernel function

allowable error ε, adopts mean shift method cyclically to carry out three steps below, until termination condition meets, wherein h is the size of mean shift method search window, and the number that n is sample point is t, w (x _i) be one and be assigned to sampled point x _iweight, span is nonnegative number;

The first step: calculate m _h(x);

Second step: if || m _h(x)-x||< ε, end loop;

The 3rd step: m _h(x) be assigned to x, and continue to carry out the first step; Obtain m _h(x) be the solution of background model, background image is at the best estimate of this point;

Detection method based on color characteristic in described step 3 is:

Step a: utilize Sobel gradient operator to carry out rim detection to the RGB triple channel of input picture moving region, and three-channel RGB Grad is had to the point of consistent change direction as the point of candidate shadow region, remember that now the set of point is S _c, obviously S _cin hypographous point not only, also may have the point of the lower object of brightness;

Step b: image conversion is arrived to C ₁c ₂c ₃color space, then the imagery exploitation Sobel gradient operator after conversion is carried out to rim detection; Due to C ₁c ₂c ₃color space is to illumination-insensitive, the point S set therefore now obtaining _min, only comprise the point of object;

C ₁c ₂c ₃color space definition is

$C_1=\arctan \left(\frac{R}{\max (G,B)}\right)$

$C_2=\arctan \left(\frac{G}{\max (R,B)}\right)$

$C_3=\arctan \left(\frac{B}{\max (R,G)}\right)$

Step c: comprehensive above step a and b, the point S set of shadow region _s=S _c-S _mthereby, the point S set of acquisition shadow region _sin the profile forming S set a little _ms1;

The described method based on Gradient Features is that the gradient vector figure after structure normalization also compares the gradient vector figure of prospect and background, thereby distinguishes the shadow region of moving target and its generation; Concrete algorithm is as follows:

Steps A: calculate the gradient vector of background image, along image level direction, carry out convolution with Sobel operator and the background image of 0 °, 45 °, 90 °, 135 ° 4 directions successively, utilize following formula to calculate 4 yardstick g of each pixel (x, y) _{f, 1}(x, y), g _{f, 2}(x, y), g _{f, 3}(x, y) and g _{f, 4}(x, y), obtains scaling vector g _f(x, y)={ g _{f, 1}, g _{f, 2}, g _{f, 3}, g _{f, 4}| _x,y, g _f(x, y) reflected the trend of background image in this pixel graded, is also the form of expression of background gradient:

g _f,i(x,y)=f(x,y)*H _i(x,y) i=1,2,3,4，

Wherein f (x, y) is the pixel value on coordinate (x, y), H _i(x, y) is 0 °, 45 °, 90 °, the 135 ° Sobel operators in 4 directions, is respectively:

$\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right],\left[\begin{array}{ccc}0& 1& 2\\ -1& 0& 1\\ -2& -1& 0\end{array}\right],\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right],\left[\begin{array}{ccc}2& 1& 0\\ 1& 0& -1\\ 0& -1& -2\end{array}\right];$

The normalization of step B background image gradient vector:

According to formula

i=1,2,3,4 are normalized the gradient vector of each pixel of background, obtain the gradient vector after normalization

and normalized gradient vector is stored, storehouse as a setting, according to gradient direction principle of invariance, the gradient vector after normalization is a build-in attribute of background, under level of illumination visible, does not change with illumination;

Step C: the gradient vector that calculates current frame image, and normalization: along current frame image horizontal direction, carry out convolution with Sobel operator and the sequence image of 0 °, 45 °, 90 °, 135 ° 4 directions successively, calculate 4 yardstick g of each pixel (x, y) _{c, 1}(x, y), g _{c, 2}(x, y), g _{c, 3}(x, y) and g _{c, 4}(x, y), obtains scaling vector g _c(x, y)={ g _{c, 1}, g _{c, 2}, g _{c, 3}, g _{c, 4}| _x,y, g _c(x, y) reflected the trend of current frame image in this pixel graded, utilizes formula:

$g_i^{*}=\frac{g_i}{\underset{j=1}{\overset{4}{\mathrm{\&Sigma;}}}\left|g_i\right|},i=\mathrm{1,2,3,4}$

Be normalized, obtained vector $g_c^{*}(x,y)=\{g_{c,1}^{*},g_{c,2}^{*},g_{c,3}^{*},g_{c,4}^{*}{|}_{x,y}\};$

Step D: the difference d that calculates the normalized gradient vector of pixel corresponding in current frame image pixel and context vault:

$d=\left|\right|g_f^{*}(x,y)-g_c^{*}(x,y)\left|\right|=\underset{i=1}{\overset{4}{\mathrm{\&Sigma;}}}|g_{c,i}^{*}-g_{f,i}^{*}|$

Template is cut apart in employing threshold value △ ε=0.3 splits motion shade from motion target area, cutting apart in template, be denoted as 1 be the point of motion shadow region;

Step e: scanning current frame image, repeating step D, until the normalized gradient vector of the entire image of current frame image and background image normalized gradient vector is more complete, realizes motion shade cutting apart from motion target area, and the set of motion shadow region point is labeled as to S _ms2.

2. the Moving Shadow Detection Approach based on color and gradient feature according to claim 1, is characterized in that, described step 2 is: in current input image and background, in the RGB component difference of corresponding pixel points, select a maximal value, be labeled as DV(x, y), work as DV(x, y) while being greater than threshold value TL, judgement (x, y) pixel is motion pixel, is labeled as g(x, y)=1, otherwise mark g(x, y)=0; For the image of M*N size, the computing formula of TL is as follows

$\mathrm{TL}=\min (\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{I}_R(x_i,y_j),\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{I}_G(x_i,y_j),\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{I}_B(x_i,y_j))---\left(2\right)$

Wherein I _r(x _i, y _j) be pixel (x _i, y _j) the value of R passage, I _g(x _i, y _j) be pixel (x _i, y _j) the value of G passage, I _b(x _i, y _j) be pixel (x _i, y _j) the value of B passage;

Then using morphological method closed operation to g(x, y) bianry image carries out filtering processing, eliminates the isolated point being caused by noise.

3. the Moving Shadow Detection Approach based on color and gradient feature according to claim 1, is characterized in that, in described step 4, morphology processing procedure is:

For further improving moving shadow detection result, to the S obtaining _ms1region and S _ms2aND-operation and OR operation are carried out in region, obtain "AND" figure and "or" figure, then, utilize connected component analysis to obtain each connected region, be called an agglomerate, in "AND" figure, obtain the center of mass point of each agglomerate, then judge which agglomerate in "or" figure contains this center of mass point, if contained, retain this agglomerate in "or" figure; Finally the agglomerate of the center of mass point not containing in "AND" figure and contain in "or" figure is deleted, finally the result of "or" figure is carried out to closing operation of mathematical morphology again one time, obtain final moving target.

Images