CN101556739A - Vehicle detecting algorithm based on intrinsic image decomposition - Google Patents

Abstract

The invention relates to a vehicle detecting algorithm in an intelligent traffic system, in particular to a vehicle detecting algorithm based on intrinsic image decomposition, which is universal at day and night. The specific practice thereof comprises the steps of: at first, carrying out derivation and filter to an initial input image on a log-domain to convert the initial input image to a gradient domain; then calculating the difference between the gradient map of a current image and the gradient map of a background image on the gradient domain to obtain the gradient map of a moving foreground image; using the method of intrinsic image decomposition to process the gradient map of the moving foreground image to obtain the gradient map of a photogram and the gradient map of a target image; extracting the pixels with the gradient amplitude larger than a threshold T in the gradient map of the target image and using the pixels as moving target points; and aiming at a certain region, considering that the vehicle passes through the region if the number of the moving target points exceeds a certain proportion of the total number of the pixels in the region. The vehicle detecting algorithm based on intrinsic image decomposition has the beneficial effects of removing the effects of shadows and illumination and being capable of accurately carrying out detection to the vehicles at day and night in real time.

Description

Vehicle detecting algorithm based on the intrinsic image decomposition

Technical field

The present invention relates to a kind of intelligent traffic vehicle detection algorithm, particularly a kind of general round the clock intelligent traffic vehicle detection technique based on the intrinsic image decomposition.

Background technology

Along with the development of intelligent transportation system (ITS), become an important research direction in the computer vision as the vehicle detecting algorithm of its core technology.Can the accurate profile that obtain vehicle be followed the tracks of and waited processing very important for target classification afterwards, also has influence on the degree of accuracy of total system.

Under hypographous condition and night environment vehicle detection be always research difficult point and hot issue.In the vehicle detection by day, shade is main interference.And vehicle at night detects the interference that not only is subjected to shade, also faces the influence of car light.At first, the headlight that light is very strong is enough to cause the change of whole image illumination condition, thereby the vehicle at night detection is difficult to realize by the method for subduction background behind the background modeling.Secondly, under the irradiation of headlight, the preceding road surface of car can be by photograph very bright, the zone of this piece light is the same with shade all moves with vehicle, if directly carry out background difference or inter-frame difference, headlight and shade will be detected together with moving vehicle, cause very big difficulty to accurately being partitioned into moving vehicle.When adopting the method for virtual coil, video detection system judges entering or when leaving state, just more need to remove the influence of shade and strong illumination at the coil position vehicle so that accurately be partitioned into moving vehicle.

Existing motion shadow detection method can be divided into following two classes: the first, and set up the geometric model of shade from the angle in space, but it needs prioris such as scene, object and illumination, and be not suitable for complicated situation; The second, based on the method for shadows pixels attribute; For example: based on the shadow detection method of hsv color space or normalization RGB color space, the distinct disadvantage of these methods is to suppose that shade does not change the color of overlay area; The method of those colors based on shade, texture or gradient characteristic for example again, these methods have similar characteristic in some part of running into vehicle and shade character and are, and tend to occur vehicle body and be taken as shadow removal, and the situation that real shade is retained.

The detection algorithm at night that existing vehicle at night detection technique then compares less, common can be divided into based on the method for vehicle body privileged site (comprising headlight, taillight and chassis etc.) location with based on the method for marginal information.These algorithms all mainly are at certain some specific occasions or a certain specific application, otherwise are exactly poor effect.

On the whole, generally all include respectively in the existing video detection system and overlap algorithms, also do not have a kind of general round the clock effective vehicle checking method now at two of day and night.

Summary of the invention

The present invention is directed to the deficiency of prior art, proposed a kind of general round the clock vehicle detecting algorithm that decomposes based on intrinsic image, this algorithm can be real-time carries out vehicle detection.The steps include: at first on log-domain, original input picture to be carried out differentiate filtering, it is transformed into gradient field; On gradient field, calculate gradient map poor of the gradient map of present image and background image, obtain the gradient map of foreground image; With the method that intrinsic image decomposes the gradient map of foreground image is handled, obtained the gradient map of shadow image and the gradient map of target image; At last, the gradient map of target image is carried out moving Object Segmentation, extract gradient amplitude in the gradient map of target image greater than the pixel of threshold value T as the moving target point; At a certain zone, if the certain proportion that outnumbers this zone interior pixel point total number of moving target point, then thinking has vehicle to pass through in this zone.

Wherein the concrete acquisition methods of the gradient map of background image is: the image that is used to extract background that at first will preserve in advance (n frame) all is transformed into log-domain; Again every width of cloth image of log-domain is carried out the differentiate filtering of two dimension; The value of same pixel in the filtered every width of cloth image of differentiate put together constitutes an one-dimensional vector, and each such one-dimensional vector is carried out medium filtering, obtains the drawing for estimate of the background image of gradient field.Promptly on log-domain, original image can be decomposed into:

i(x，y，t)＝m(x，y，t)+b(x，y) (1)

I (x wherein, y, t) represent the gradient map of t original image constantly, b (x, the y) gradient map of presentation video background (shadow that comprises the road surface fixed object), m (x, y, t) represent the gradient map of t sport foreground image constantly, it has comprised a large amount of sport foreground information, but also is not information of vehicles accurately.

The method that intrinsic image decomposes is: set up the joint classification device based on spatial color information and time domain colouring information, each pixel among the sport foreground image gradient figure is classified, obtain the gradient map of target image.

Joint classification device based on spatial color information and time domain colouring information is defined as:

1) in normalization RGB color space,

Be the red component of normalization color space, Horizontal direction gradient and vertical gradient be expressed as respectively With Obtain

In point (x, the gradient of y) locating

$d\tilde{R}(x,y)=\left\{\begin{array}{c}{\tilde{R}}_x(x,y),\mathrm{if}\left|{\tilde{R}}_x(x,y)\right|>\left|{\tilde{R}}_y(x,y)\right|\\ {\tilde{R}}_y(x,y),\mathrm{otherwise}\end{array}\right.$

2) for the green component after the normalization

Adopting uses the same method obtains

In point (x, the gradient of y) locating

$d\tilde{G}(x,y)=\left\{\begin{array}{c}{\tilde{G}}_x(x,y),\mathrm{if}\left|{\tilde{G}}_x(x,y)\right|>\left|{\tilde{G}}_y(x,y)\right|\\ {\tilde{G}}_y(x,y),\mathrm{otherwise}\end{array}\right.$

3) with point (x, the color characteristics value V that y) locates (x y) is defined as:

$V(x,y)=\max \left\{\right|d\tilde{R}(x,y)|,|d\tilde{G}(x,y)\left|\right\}$

Then, the sorter based on spatial color information is defined as:

Sorter based on the time domain colouring information is defined as:

(x, the normalization color-set of y) locating becomes to be respectively wherein adjacent two two field pictures at same pixel

With

4) according to aforementioned two sorters, combined detector is defined as:

T ₁And T ₂Be constant.

The beneficial effect of the vehicle detecting algorithm that decomposes based on intrinsic image of the present invention is: removed the influence of shade and illumination, can be real-time and accurately the vehicle at daytime and night have been detected.

Description of drawings

Fig. 1 is the general flow chart of a kind of vehicle detecting algorithm embodiment of the present invention;

Fig. 2 is for extracting the process flow diagram of sport foreground module;

Fig. 3 is the process flow diagram of intrinsic image decomposing module.

Embodiment:

Below, the present invention is further illustrated with embodiment in conjunction with the accompanying drawings.

As shown in Figure 1, decompose for extracting sport foreground and intrinsic image the core of this algorithm, below in conjunction with accompanying drawing these two parts is described:

1. extraction sport foreground

This step carries out the first time to original image sequence and decomposes, with obtain Background b (x, y) and the sport foreground part m in each two field picture (t), concrete process flow diagram is seen Fig. 2 for x, y.

At first, with input image sequence i (t-n * Δ t), n=0,1,2 ... k} is transformed into log-domain (k=9 in the algorithm, interval of delta t can oneself be adjusted), then through f _x=[0,1 ,-1] ^TAnd f _y=[0,1 ,-1] is carried out differentiate filtering, is obtained:

i _x/y(x，y，t)＝f _x/y*i(x，y，t)

i _x/y(x，y，t)＝f _x/y*[b(x，y)+m(x，y，t)]

i _x/y(x，y，t)＝b _x/y(x，y)+m _x/y(x，y，t)

F wherein _X/y={ f _x, f _y, i _X/y(x, y, t)={ i _x(x, y, t), i _y(x, y, t) }, * represents convolution algorithm.The background component b of image (x, y) satisfy:

b _x/y(x，y)＝f _x/y*b(x，y)

Output m after differentiate filtering _X/y(x, y t) are sparse matrix.

Respectively the x of output and the gradient image sequence of y both direction are carried out medium filtering, can obtain the estimated value of background image gradient

With

${\hat{b}}_{x/y}(x,y)={\mathrm{median}}_t{i}_{x/y}(x,y,t)$

${\hat{m}}_{x/y}(x,y,t)=i_{x/y}(x,y,t)-{\hat{b}}_{x/y}(x,y)$

After having obtained the gradient map of background image, only need from the gradient map of original image, deduct the gradient map m that this part just can obtain prospect part (comprising moving object, motion shade, illumination variation) _x(x, y, t) and m _y(x, y, t):

m _x(x，y，t)＝I _x(x，y，t)-b _x(x，y)

m _y(x，y，t)＝I _y(x，y，t)-b _y(x，y)

2. intrinsic image decomposes

Obtain after the gradient of foreground image, this step is that sorter of design is separated movement destination image from foreground image, obtains the target image gradient map, and concrete process flow diagram is seen Fig. 3.

Foreground image can regard as irradiation image and target image and.In gradient field, their relation satisfies:

m _x/y(x，y，t)＝o _x/y(x，y，t)+s _x/y(x，y，t)

Wherein, o _X/y(x, y t) are the gradient of moving target; s _X/y(x, y t) are the gradient of shadow.This algorithm will utilize the color characteristic of image sequence in time domain and spatial domain to combine to reach this purpose.

A. based on the sorter of spatial color information

Only consider current frame image,, design a two-value sorter foreground image is classified in the gradient of each location of pixels according to spatial domain change color information.

At the RGB color space, each pixel in the image can (R, G B) represent with vector.Represent mode according to the colouring information that top document is introduced, rgb space can be expressed as to normalization:

$\tilde{R}=\frac{R}{R+G+B}$

$\tilde{G}=\frac{G}{R+G+B}$

$\tilde{B}=\frac{B}{R+G+B}$

Be respectively normalized redness, green and blue component, because

Can pass through $\tilde{B}=1-\tilde{R}-\tilde{G}$ Obtain, so use

Represent that the color group achievement is passable.

Horizontal direction gradient and vertical gradient be expressed as respectively With With

Representative

Point (x, the gradient of y) locating, and the value that defines it is:

$d\tilde{R}(x,y)=\left\{\begin{array}{c}{\tilde{R}}_x(x,y),\mathrm{if}\left|{\tilde{R}}_x(x,y)\right|>\left|{\tilde{R}}_y(x,y)\right|\\ {\tilde{R}}_y(x,y),\mathrm{otherwise}\end{array}\right.$

Gradient

Can use the same method and obtain:

$d\tilde{G}(x,y)=\left\{\begin{array}{c}{\tilde{G}}_x(x,y),\mathrm{if}\left|{\tilde{G}}_x(x,y)\right|>\left|{\tilde{G}}_y(x,y)\right|\\ {\tilde{G}}_y(x,y),\mathrm{otherwise}\end{array}\right.$

And then with point (x, the color characteristics value V that y) locates (x y) is defined as:

$V(x,y)=\max \left\{\right|d\tilde{R}(x,y)|,|d\tilde{G}(x,y)\left|\right\}$

(x y) compares with a threshold value, constitutes a simple two-value sorter with color characteristics value V.Concrete sorter is defined as:

$\left\{\begin{array}{cc}{\hat{o}}_{x/y}(x,y)={\hat{m}}_{x/y}(x,y),{\hat{s}}_{x/y}(x,y)=0,& V(x,y)\≥T_1\\ {\hat{s}}_{x/y}(x,y)={\hat{m}}_{x/y}(x,y),{\hat{o}}_{x/y}(x,y)=0,& \mathrm{else}\end{array}\right.$

If (x y) greater than this threshold value, then is divided into the gradient that belongs to target image with foreground image in the gradient of this location of pixels to V; Otherwise, then foreground image is divided into the gradient that belongs to the shadow image in the gradient of this location of pixels.

B. based on the sorter of time domain colouring information

This step is set up a two-value sorter, will be classified as the gradient that belongs to target image by the time domain gradient that reflectance varies causes, all the other gradients then are left the gradient of shadow image.

Suppose that (x, the color-set of y) locating becomes to be respectively C to adjacent two two field pictures at same pixel ₁(x, y) and C ₂(x, y).Here C ₁(x, y) and C ₂(x y) is the RGB vector.Normalization C ₁(x, y) and C ₂(x y) obtains With

Use dot product

Be illustrated in that (x y) locates color and forms the degree that time domain changes, and is caused by illumination if the pixel time domain changes, then $[{\hat{C}}_1(x,y)\·{\hat{C}}_2(x,y)]=1,$ Cos (0) just.Consider The noise, if

Value be lower than one and be slightly less than 1 threshold value, then the pixel of this point is changed being considered as causing by reflectance varies.Concrete sorter is defined as:

$\left\{\begin{array}{cc}{\hat{o}}_{x/y}(x,y)={\hat{m}}_{x/y}(x,y),{\hat{s}}_{x/y}(x,y)=0,& {\hat{C}}_1(x,y)\&CenterDot;{\hat{C}}_2(x,y)<T_2\\ {\hat{s}}_{x/y}(x,y)={\hat{m}}_{x/y}(x,y),{\hat{o}}_{x/y}(x,y)=0,& \mathrm{else}\end{array}\right.$

C. joint classification device

Two sorters in front are combined, stipulate that final classification criterion is:

$\left\{\begin{array}{c}{\hat{o}}_{x/y}(x,y)={\hat{m}}_{x/y}(x,y),{\hat{s}}_{x/y}(x,y)=0,\\ \mathrm{ifV}(x,y)\&GreaterEqual;{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="A20091009879100131.png"\; state="\{\{state\}\}">T</figure-callout>}}_1\mathrm{and}{\hat{C}}_1(x,y)\&CenterDot;{\hat{C}}_2(x,y)<T_2\\ {\hat{s}}_{x/y}(x,y)={\hat{m}}_{x/y}(x,y),{\hat{o}}_{x/y}(x,y)=0,\\ \mathrm{otherwise}\end{array}\right.$

The result who obtains by this joint classification device combines the target image gradient map that obtains after two conditions.

The specific practice of moving Object Segmentation part is:

In the goal gradient image, the Grad than other location of pixels is big usually to be positioned at the Grad of object edge location of pixels.Therefore, Grad and threshold value that needs only each location of pixels of goal gradient image compares the edge that just can detect moving target.

(x, y) result of presentation class is right with E

Carry out operational analysis, as fruit dot (x y) belongs to the edge of moving target, then E (x, y)=1, otherwise E (x, y)=0, criterion is as follows:

$E_x(x,y)={\hat{o}}_x(x,y,r)+{\hat{o}}_x(x,y,g)+{\hat{o}}_x(x,y,b)$

$E_y(x,y)={\hat{o}}_y(x,y,r)+{\hat{o}}_y(x,y,g)+{\hat{o}}_y(x,y,b)$

$E(x,y)=\left\{\begin{array}{cc}1,& \mathrm{if}{E}_x(x,y)>T_3\mathrm{or}{E}_y(x,y)>T_3\\ 0,& \mathrm{else}\end{array}\right..$

Wherein, With

(x y) locates the Grad of each color component of RGB color space x direction to be respectively point;

With

(x y) locates the Grad of each color component of RGB color space y direction to be respectively point.

At some zones, (ratio that x, the number of pixel y)=1 account for total pixel number in this zone surpasses threshold value S, and judging in this zone has moving vehicle to pass through if satisfy E in this zone.

Claims (4)

1, a kind of vehicle detecting algorithm that decomposes based on intrinsic image is characterized in that comprising following steps: at first on log-domain, original input picture is carried out differentiate filtering, it is transformed into gradient field; On gradient field, calculate gradient map poor of the gradient map of present image and background image, obtain the gradient map of sport foreground image; With the method that intrinsic image decomposes the gradient map of foreground image is handled, obtained the gradient map of shadow image and the gradient map of target image; Extract in the gradient map of target image gradient amplitude greater than the pixel of threshold value T, as the moving target point; At a certain zone, if the certain proportion that outnumbers this zone interior pixel point total number of moving target point, then thinking has vehicle to pass through in this zone.

2, the vehicle detecting algorithm that decomposes based on intrinsic image as claimed in claim 1, it is characterized in that the concrete acquisition methods that extracts the gradient map of background image in the background image part is: the image that is used to extract background that at first will preserve in advance (n frame) all is transformed into log-domain; Again every width of cloth image of log-domain is carried out the differentiate filtering of two dimension; The value of same pixel in the filtered every width of cloth image of differentiate put together constitutes an one-dimensional vector, and each such one-dimensional vector is carried out medium filtering, obtains the drawing for estimate of the background image of gradient field.

3, the vehicle detecting algorithm that decomposes based on intrinsic image as claimed in claim 1 or 2, it is characterized in that, the intrinsic image decomposition method is: set up the joint classification device based on spatial color information and time domain colouring information, each pixel in the sport foreground image is classified, thereby obtain final moving target.

4, the vehicle detecting algorithm that decomposes based on intrinsic image as claimed in claim 3 is characterized in that, is defined as based on the joint classification device of spatial color information and time domain colouring information:

1) in normalization RGB color space,

Be the red component of normalization color space,

Horizontal direction gradient and vertical gradient be expressed as respectively

With

Obtain

In point (x, the gradient of y) locating

$d\tilde{R}(x,y)=\left\{\begin{array}{cc}{\tilde{R}}_x(x,y),& \mathrm{if}\left|{\tilde{R}}_x(x,y)\right|>\left|{\tilde{R}}_y(x,y)\right|\\ {\tilde{R}}_y(x,y),& \mathrm{otherwise}\end{array}\right.$

2) for the green component after the normalization Adopt with

Same method obtains In point (x, the gradient of y) locating

3) with point (x, the color characteristics value V that y) locates (x y) is defined as:

$V(x,y)=\max \left\{\right|d\tilde{R}(x,y)|,|d\tilde{G}(x,y)\left|\right\}$

Then, the sorter based on spatial color information is defined as:

Sorter based on the time domain colouring information is defined as:

(x, the normalization color-set of y) locating becomes to be respectively wherein adjacent two two field pictures at same pixel

With

4) according to aforementioned two sorters, combined detector is defined as:

T ₁And T ₂Be constant.

Images