CN102289819B - Method for detecting infrared motion target in real time for background adaptive estimation - Google Patents

Abstract

The invention discloses a method for detecting an infrared motion target in real time for background adaptive estimation. The method comprises the following steps of: (1) initializing a background model BG (0); (2) inputting a kth-frame original infrared image Xorg (k), wherein k is equal to 1, 2 to K; (3) restraining a space domain non-uniformity noise and a time domain random noise of the kth-frame original infrared image Xorg (k), and outputting an image Xdeal (k) which is subjected to noise restraint; (4) registering the image Xdeal (k), and outputting a registered kth-frame infrared image Xreg (k); (5) updating the background model, namely updating a background model BG (k) of the kth-frame infrared image by using an adaptive forgetting factor method; (6) performing difference operation on the kth-frame infrared image Xorg (k) and the background model BG (k), and outputting a difference operation result Xout (k); and (7) calculating the optimum segmentation threshold value Thopt of a target and a background in the Xout (k) according to a maximum inter-class variance theory to finish detection of the motion target. The method has the advantages that: by image pro-processing, the problem of over-sensitivity in noise and scene change in the conventional background difference method is effectively solved, and the detection probability of the target is improved.

Description

The infrared motion target real-time detection method that a kind of Adaptive background subtraction is estimated

Technical field

The present invention relates to a kind of infrared image object detection method, particularly a kind of moving target detecting method of suitable hardware real-time implementation.

Background technology

Detection for Moving Target is analyzed by the image sequence that sensor is obtained, and detects moving target, at aspects such as traffic monitoring, self-navigations, very important Practical significance is arranged.Typical moving target detecting method is divided into three major types: background subtraction point-score, difference image method and optical flow method.The background subtraction point-score carries out difference with present image and a known background, and passing threshold cuts apart to extract moving target.The method hypothesis background is accurately as can be known, is not subjected to the restriction of target speed, can more intactly extract moving target.The background subtraction point-score is too responsive to the scene changes that illumination and external condition cause, when there is noise in image, when particularly image and background had noise, the extraction effect of target became undesirable.The difference image method is subtracted each other by adjacent two frames in sequence image, detects target according to difference image, is the straightforward procedure that detects Moving Targets in Sequent Images.The strong adaptability of the method to environment, fast operation, weak point is can not accurately setting movement target (target location of detection is the mean place of target in two two field pictures), easily produces cavitation in the target internal zone, is difficult to detect overlapping moving target.Optical flow method does not need to know any background information, relies on time dependent light stream characteristic to calculate motion and the structural parameters of target, is the important method of moving target in the analytical sequence image.Yet, in light stream, the motion of target and structural parameters need single order or the second derivative of computed image gray scale and light stream, real image can be subject to the pollution of noise, and the computation process of derivative is the process that a noise amplifies, more the strong noise amplification is more obvious for exponent number, and the computing of optical flow method is complicated, and processing capability in real time is poor.

Summary of the invention

Goal of the invention: for the background subtraction point-score deficiency too responsive to noise and scene changes, the objective of the invention is to design a kind of respond well, calculate the infrared motion target real-time detection method that simple, as to be fit to hardware real-time implementation Adaptive background subtraction is estimated.

Technical scheme: for achieving the above object, the technical solution used in the present invention is the infrared motion target real-time detection method that a kind of Adaptive background subtraction is estimated, comprises the steps:

(1) initialization background model BG (0);

(2) the original infrared image X of input k frame _org(k), k=1,2 ..., K;

(3) to the original infrared image X of the described k frame of step (2) _org(k) spatial domain heterogeneity noise and time domain random noise suppress, and output suppresses the image X after noise _Deal(k);

(4) the k frame infrared image X after the described inhibition noise of step of registration (3) _Deal(k), output images after registration X _reg(k);

(5) background model is upgraded: by the k frame infrared image X after the described registration of adaptive forgetting factor method step of updating (4) _reg(k) background model BG (k);

(6) with the original infrared image X of k frame _org(k) make calculus of differences with background model BG (k), the X as a result of output calculus of differences _out(k);

(7) according to the described X of the theoretical calculation procedure (6) of maximum between-cluster variance _out(k) the optimal segmenting threshold Th of target and background in _opt, complete the detection of moving target.

In described step (1), utilize initial N frame original image { X _org(k) | k=1,2 ..., K} calculates background model BG (0), and the expression formula of the background value BG that coordinate (i, j) is located (i, j, 0) can be shown below:

BG(i，j，0)＝median[X _org(i，j，1)，X _org(i，j，2)，...，X _org(i，j，k)]

In formula, operational symbol median[] expression finds the solution median operation, X _orgThe original infrared image X of (i, j, k) expression k frame _org(k) pixel value located of coordinate (i, j).

In described step (3), can use finite support territory high frequency constant statistics Nonuniformity Correction method to the original infrared image X of k frame _org(k) spatial domain heterogeneity noise suppresses.The method that suppresses can be:

Utilize gauss low frequency filter g with the original infrared image X of k frame _org(k) be divided into high fdrequency component

And low frequency component

Two parts:

$X_{\mathrm{org}}\left(k\right)=X_{\mathrm{org}}^{\mathrm{high}}\left(k\right)+X_{\mathrm{org}}^{\mathrm{low}}\left(k\right)$

$X_{\mathrm{org}}^{\mathrm{low}}\left(k\right)=X_{\mathrm{org}}\left(k\right)\⊗g$

In formula, operational symbol

The expression convolution operation;

Gauss low frequency filter g (i, the j) expression formula that coordinate (i, j) is located is shown below;

$g(i,j)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}{e}^{-\left(\frac{i^2+j^2}{2{\mathrm{\σ}}^2}\right)}$

In formula, σ is the variance of gauss low frequency filter g;

Calculate the original infrared image X of k frame _org(k) biasing coefficient O ^High(k) and gain coefficient G ^High(k), the biasing coefficient O that locates of k frame coordinate (i, j) ^High(i, j, k) and gain coefficient G ^High(i, j, k) expression formula is shown below respectively:

$O_{\mathrm{tem}}^{\mathrm{high}}(i,j,k)=\frac{1}{\mathrm{Num}}\×X_{\mathrm{org}}^{\mathrm{high}}(i,j,k)+(1-\frac{1}{N})\×O^{\mathrm{high}}(i,j,k-1)$

$G^{\mathrm{high}}(i,j,k)=\frac{1}{\mathrm{Num}}\×|X_{\mathrm{org}}^{\mathrm{high}}(i,j,k)-O^{\mathrm{high}}(i,j,k-1)|+(1-\frac{1}{N})\×G^{\mathrm{high}}(i,j,k-1)$

O ^high(i，j，0)＝0

$O_{\mathrm{tem}}^{\mathrm{high}}(i,j,0)=0$

G ^high(i，j，0)＝1

In formula,

Candidate's coefficient of setovering,

It is the original infrared image of k frame

X _org(k) the pixel value X that locates of coordinate (i, j) _orgThe high fdrequency component of (i, j, k), Num is the accumulation frame number, ε is error constant,

It is the finite support territory of finite support territory high frequency constant statistics Nonuniformity Correction method;

Therefore, X _org(i, j, k) output after finite support territory high frequency constant statistics Nonuniformity Correction method processing Expression formula is shown below:

$X_{\mathrm{deal}}^{\mathrm{nu}}(i,j,k)=(X_{\mathrm{org}}(i,j,k)-O^{\mathrm{high}}(i,j,k))/G^{\mathrm{high}}(i,j,k)$

Wherein,

It is the image after the heterogeneity squelch.

To the image of k frame after the heterogeneity squelch

Can carry out the time domain random noise and suppress, the image X of output k frame after the time domain random noise suppresses _Deal(k):.

$X_{\mathrm{deal}}^H(i,j,k)=\mathrm{median}\left[X_{\mathrm{deal}}^{\mathrm{nu}}(i,j+q^{\′},k)\right|q^{\′}=-2,-\mathrm{1,0,1,2}]$

$X_{\mathrm{deal}}(i,j,k)=\mathrm{median}\left[X_{\mathrm{deal}}^H(i+p^{\′},j,k)\right|p^{\′}=-2,-\mathrm{1,0,1,2}]$

Wherein, operational symbol median[] expression finds the solution median operation,

Right

The Output rusults of medium filtering in the horizontal direction, X _Deal(i, j, k) is right

The Output rusults of medium filtering, namely right in the vertical direction

Output image after the time domain random noise suppresses.

In described step (4), can adopt the k frame infrared image X after template registration method is exported registration _reg(k):

At first, the image X after the k frame suppresses noise _Deal(k) choose size centered by coordinate (i, j) and be the standard form of P * Q;

Secondly, adopt square error function MSE as criterion, the image X after the k-1 frame suppresses noise _Deal(k-1) in centered by coordinate (i, j) size mate for the image-region of M * N, P＜M, Q＜N, calculating least mean-square error MSE _min, least mean-square error MSE _minCorresponding coordinate (i+ Δ i, j+ Δ j) is registration position;

Square error function MSE expression formula is as follows:

$\mathrm{MSE}(i+\mathrm{\Δi},j+\mathrm{\Δj},k)=\frac{1}{P\×Q}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{q=1}{\overset{Q}{\mathrm{\Σ}}}{(X_{\mathrm{deal}}(i+\mathrm{\Δi}+p-\frac{p}{2},j+\mathrm{\Δj}+q-\frac{Q}{2},k)-X_{\mathrm{deal}}(i+\mathrm{\Δi}+p-\frac{P}{2},j+\mathrm{\Δj}+q-\frac{Q}{2},k-1))}^2$

Least mean-square error MSE _minExpression formula is as follows:

MSE _min(i+Δi，j+Δj，k)＝min{MSE(i+Δi，j+Δj，k)}

Wherein, (Δ i, Δ j) expression is with respect to the displacement of coordinate (i, j) ,-(M-P)/2≤Δ i≤(M-P)/2 ,-(N-Q)/2≤Δ j≤(N-Q)/2;

At last, to the image X after k frame inhibition noise _Deal(k) carry out registration:

X _reg(i，j，k)＝X _deal(i+Δi，j+Δj，k)。

In described step (5), the original infrared image X of k frame _org(k) background model BG (i, j, the k) expression formula located of coordinate (i, j) can be shown below:

BG(i，j，k)＝(1-α(i，j，k))×X _reg(i，j，k)+α(i，j，k)×BG(i，j，k-1)

In formula, X _reg(i, j, k) is the original infrared image X of k frame _org(k) the registration output located of coordinate (i, j), α (i, j, k) is the original infrared image X of k frame _org(k) forgetting factor located of coordinate (i, j), expression formula is shown below:

$\mathrm{\α}(i,j,k)=\frac{|H(i,j,k)-H(i,j,k-1)|}{H(i,j,k-1)}$

$H(i,j,k)=-\underset{m=1}{\overset{M_1}{\mathrm{\Σ}}}\underset{n=1}{\overset{N_1}{\mathrm{\Σ}}}(\left(\frac{X_{\mathrm{reg}}(i,j,k)}{\underset{m=1}{\overset{M_1}{\mathrm{\Σ}}}\underset{n=1}{\overset{N_1}{\mathrm{\Σ}}}{X}_{\mathrm{reg}}(i,j,k)}\right)\·\lg \left(\frac{X_{\mathrm{reg}}(i,j,k)}{\underset{m=1}{\overset{M_1}{\mathrm{\Σ}}}\underset{n=1}{\overset{N_1}{\mathrm{\Σ}}}{X}_{\mathrm{reg}}(i,j,k)}\right))$

In formula, H (i, j, k) is the original infrared image X of k frame _org(k) M centered by coordinate (i, j) ₁* N ₁The neighborhood entropy of the local neighborhood of size, operational symbol lg[] operation of expression logarithm.

In described step (6), the original infrared image X of k frame _org(k) the output X that locates of coordinate (i, j) _out(i, j, k) expression formula can be:

X _out(i，j，k)＝X _org(i，j，k)-BG(i，j，k)

In formula, X _org(i, j, k) is the original infrared image X of k frame _org(k) pixel value located of coordinate (i, j), BG (i, j, k) is the pixel value that background model BG (k) coordinate (i, j) is located.

In described step (7), Th _optExpression formula can be as follows:

${\mathrm{Th}}_{\mathrm{opt}}=\arg \underset{0\≤\mathrm{Th}\≤255}{\max }({\mathrm{\ω}}_0{({\mathrm{\μ}}_0-\mathrm{\μ})}^2+{\mathrm{\ω}}_1{({\mathrm{\μ}}_1-\mathrm{\μ})}^2)$

Wherein, operational symbol arg[] be to return to optimal value operation, operational symbol max[] be to find the solution maxima operation, Th is segmentation threshold, ω ₀And ω ₁Respectively that moving target is at the probability of background area and target area appearance, μ ₀The mean value of background area, μ ₁Be the mean value of target area, μ is the mean value of background and target area, and their expression formula is as follows respectively:

${\mathrm{\ω}}_0=\underset{v=0}{\overset{\mathrm{Th}}{\mathrm{\Σ}}}p\left(v\right),$ ${\mathrm{\μ}}_0=\frac{1}{{\mathrm{\ω}}_0}\underset{v=0}{\overset{\mathrm{Th}}{\mathrm{\Σ}}}v\·p\left(v\right)$

${\mathrm{\ω}}_1=\underset{v=\mathrm{Th}+1}{\overset{T}{\mathrm{\Σ}}}p\left(v\right),$ ${\mathrm{\μ}}_1=\frac{1}{{\mathrm{\ω}}_1}\underset{v=\mathrm{Th}+1}{\overset{T}{\mathrm{\Σ}}}v\·p\left(v\right)$

μ＝ω ₀μ ₀+ω ₁μ ₁

In formula, p (v) is that the gradation of image value is the probability of the pixel appearance of v, and T is gray level, and Th is segmentation threshold;

The expression formula of p (v) is: $p\left(v\right)=\frac{1}{M_2\×N_2}\underset{X_{\mathrm{out}}(i,j,k)=v}{\mathrm{\Σ}1}$

Wherein,

Expression is used for computed image gray-scale value X _out(i, j, k) equals the number of pixels of v, M ₂And N ₂Columns and the line number of difference presentation video.

Beneficial effect: the present invention compared with prior art has following remarkable advantage: (1) by the image pre-service, efficiently solves traditional background subtraction point-score to noise and scene changes sensitive issue too, has improved target detection probability; (2) finite support territory high frequency constant statistics Nonuniformity Correction method (SC-HFCS) is proposed, utilize the heterogeneity of SC-HFCS method adaptively correcting image, solved traditional scene correcting algorithm speed of convergence when improving signal noise ratio (snr) of image slow, there is the technical barrier of " ghost ", reduced false alarm rate; (3) for the context update problem of background subtraction point-score, designed the forgetting factor adaptive updates model based on the neighborhood entropy, effectively improved context update speed, solved in essence the interference problem of prospect to background; (4) there are not labyrinth and high exponent arithmetic(al) in detection method, are fit to the hardware real-time implementation.

Description of drawings

Fig. 1 is the process flow diagram of the infrared motion target real-time detection method of Adaptive background subtraction estimation of the present invention.

Embodiment

Below in conjunction with the drawings and specific embodiments, further illustrate the present invention, should understand these embodiment only is used for explanation the present invention and is not used in and limits the scope of the invention, after having read the present invention, those skilled in the art all fall within the application's claims limited range to the modification of the various equivalent form of values of the present invention.

In conjunction with Fig. 1, the below illustrates with example the infrared motion target real-time detection method that Adaptive background subtraction of the present invention is estimated.Use non-refrigeration gazing type infrared eye, the focal plane arrays (FPA) size is 320 * 240, and working frame frequency is 25HZ, and data bits is 8 bits.Infrared image (hereinafter to be referred as image) is by the special image disposable plates of optical fiber transmission to the DSP+FPGA framework, and the detection of moving target realizes in dsp processor, satisfies real-time processing requirement, and concrete implementation step is as follows:

(1) initialization background model BG (0) gathers ten initial frame original image { X _org(k) | k=1,2 ..., it is as follows that 10}, coordinate (i, j) locate BG (i, j, 0) expression formula:

BG(i，j，0)＝median[X _org(i，j，1)，X _org(i，j，2)，...，X _org(i，j，10)]

Wherein, operational symbol median[] expression finds the solution median operation, X _org(i, j, k) expression k frame original image X _org(k) pixel value located of coordinate (i, j), 1≤i≤240,1≤j≤320.

(2) input k frame original image X _org(k), the image size is 320 * 240;

(3) to k frame original image X _org(k) spatial domain heterogeneity noise suppresses, and proofreaies and correct output to be It is the image after the heterogeneity noise processed.The template size of setting gauss low frequency filter g is 5 * 5, and variance is 1, and spatial domain 5 * 5 Gauss's templates are as shown in table 1:

0.002	0.013	0.220	0.013	0.002
					0.013	0.06	0.098	0.06	0.013
0.220	0.098	0.162	0.098	0.220
					0.013	0.06	0.098	0.06	0.013
0.002	0.013	0.220	0.013	0.002

Table 1

Calculate X _org(k) high fdrequency component

Expression formula is as follows:

$X_{\mathrm{org}}^{\mathrm{low}}\left(k\right)=X_{\mathrm{org}}\left(k\right)-X_{\mathrm{org}}\left(k\right)\⊗g$

Wherein, operational symbol

It is convolution operation.

By

Calculate X _org(k) correction coefficient of heterogeneity noise, coefficient O namely setovers ^High(k) and gain coefficient G ^High(k).The biasing coefficient O that k frame coordinate (i, j) is located ^High(i, j, k) and gain coefficient G ^High(i, j, k) expression formula is as follows respectively:

$O_{\mathrm{tem}}^{\mathrm{high}}(i,j,k)=\frac{1}{\mathrm{Num}}\×X_{\mathrm{org}}^{\mathrm{high}}(i,j,k)+(1-\frac{1}{N})\×O^{\mathrm{high}}(i,j,k-1)$

$G^{\mathrm{high}}(i,j,k)=\frac{1}{\mathrm{Num}}\×|X_{\mathrm{org}}^{\mathrm{high}}(i,j,k)-O^{\mathrm{high}}(i,j,k-1)|+(1-\frac{1}{N})\×G^{\mathrm{high}}(i,j,k-1)$

O ^high(i，j，0)＝0

$O_{\mathrm{tem}}^{\mathrm{high}}(i,j,0)=0$

G ^high(i，j，0)＝1

Wherein,

Candidate's coefficient of setovering,

X _orgThe high fdrequency component of (i, j, k), Num is the accumulation frame number, ε is error constant,

The finite support territory of SC-HFCS algorithm, 1≤i≤240,1≤j≤320.

Therefore, k frame coordinate (i, j) is located pixel X _org(i, j, k) output after the SC-HFCS algorithm process

Expression formula is as follows:

$X_{\mathrm{deal}}^{\mathrm{nu}}(i,j,k)=(X_{\mathrm{org}}(i,j,k)-O^{\mathrm{high}}(i,j,k))/G^{\mathrm{high}}(i,j,k)$

Wherein,

The image of k frame after the heterogeneity noise processed.

(4) to the k two field picture

Carry out the time domain random noise and suppress, the image X of output k frame after the time domain random noise suppresses _Deal(k), expression formula is as follows:

$X_{\mathrm{deal}}^H(i,j,k)=\mathrm{median}\left[X_{\mathrm{deal}}^{\mathrm{nu}}(i,j+q^{\′},k)\right|q^{\′}=-2,-\mathrm{1,0,1,2}]$

$X_{\mathrm{deal}}(i,j,k)=\mathrm{median}\left[X_{\mathrm{deal}}^H(i+p^{\′},j,k)\right|p^{\′}=-2,-\mathrm{1,0,1,2}]$

Wherein, operational symbol median[] expression finds the solution median operation,

Right

The Output rusults of medium filtering in the horizontal direction, X _Deal(i, j, k) is right

The Output rusults of medium filtering, namely right in the vertical direction

Output image after the time domain random noise suppresses, 1≤i≤240,1≤j≤320.

(5) utilize template registration method to X _Deal(k) carry out registration, obtain the k two field picture X after registration _reg(k).

At first, at the image X of k frame after the time domain random noise suppresses _Deal(k) choose size centered by coordinate (i, j) and be the standard form (P=Q=30) of P * Q;

Secondly, adopt square error function (MSE) as criterion, at k-1 two field picture X _Deal(k-1) big or small centered by coordinate (i, j) in is the image-region work coupling (M=N=60) of M * N, calculates least mean-square error MSE _min, least mean-square error MSE _minCorresponding coordinate (i+ Δ i, j+ Δ j) is optimum matching (registration) position.

Square error function (MSE) expression formula is as follows:

$\mathrm{MSE}(i+\mathrm{\Δi},j+\mathrm{\Δj},k)=\frac{1}{P\×Q}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{q=1}{\overset{Q}{\mathrm{\Σ}}}{(X_{\mathrm{deal}}(i+\mathrm{\Δi}+p-\frac{p}{2},j+\mathrm{\Δj}+q-\frac{Q}{2},k)-X_{\mathrm{deal}}(i+\mathrm{\Δi}+p-\frac{P}{2},j+\mathrm{\Δj}+q-\frac{Q}{2},k-1))}^2$

Least mean-square error MSE _minExpression formula is as follows:

MSE _min(i+Δi，j+Δj，k)＝min{MSE(i+Δi，j+Δj，k)}

Wherein, (Δ i, Δ j) expression is with respect to the displacement of coordinate (i, j) ,-(M-P)/2≤Δ i≤(M-P)/2 ,-(N-Q)/2≤Δ j≤(N-Q)/2.

At last, to the X of k frame after the time domain random noise suppresses _Deal(k) carry out registration:

X _reg(i，j，k)＝X _deal(i+Δi，j+Δj，k)。

(6) upgrade X _reg(k) background model BG (k), the background model BG that coordinate (i, j) is located (i, j, k) expression formula is as follows:

BG(i，j，k)＝(1-α(i，j，k))×X _reg(i，j，k)+α(i，j，k)×BG(i，j，k-1)

Wherein, X _reg(i, j, k) is k frame original image X _org(k) the registration output located of coordinate (i, j), α (i, j, k) is k frame original image X _org(k) forgetting factor of the registration output located of coordinate (i, j), expression formula is as follows:

$\mathrm{\α}(i,j,k)=\frac{|H(i,j,k)-H(i,j,k-1)|}{H(i,j,k-1)}$

$H(i,j,k)=-\underset{m=1}{\overset{M_1}{\mathrm{\Σ}}}\underset{n=1}{\overset{N_1}{\mathrm{\Σ}}}(\left(\frac{X_{\mathrm{reg}}(i,j,k)}{\underset{m=1}{\overset{M_1}{\mathrm{\Σ}}}\underset{n=1}{\overset{N_1}{\mathrm{\Σ}}}{X}_{\mathrm{reg}}(i,j,k)}\right)\·\lg \left(\frac{X_{\mathrm{reg}}(i,j,k)}{\underset{m=1}{\overset{M_1}{\mathrm{\Σ}}}\underset{n=1}{\overset{N_1}{\mathrm{\Σ}}}{X}_{\mathrm{reg}}(i,j,k)}\right))$

Wherein, H (i, j, k) is k frame original image X _org(k) M centered by (i, j) ₁* N ₁The neighborhood entropy of the local neighborhood of size, X _reg(i, j, k) is k frame original image X _org(k) the registration output located of coordinate (i, j), operational symbol lg[] be the logarithm operation, 1≤i≤240,1≤j≤320.

(7) with k frame original image X _org(k) make calculus of differences with background model BG (k), the X as a result of output calculus of differences _out(k).The output X that coordinate (i, j) is located _out(i, j, k) expression formula is as follows:

X _out(i，j，k)＝X _org(i，j，k)-BG(i，j，k)

Wherein, X _org(i, j, k) is k frame original image X _org(k) pixel value located of coordinate (i, j), BG (i, j, k) is the pixel value that background model BG (k) coordinate (i, j) is located, 1≤i≤240,1≤j≤320.

(8) calculate X theoretical according to maximum between-cluster variance _out(k) the optimal segmenting threshold Th of target and background in _opt, complete the detection of moving target.Th _optExpression formula is as follows:

${\mathrm{Th}}_{\mathrm{opt}}=\arg \underset{0\≤\mathrm{Th}\≤255}{\max }({\mathrm{\ω}}_0{({\mathrm{\μ}}_0-\mathrm{\μ})}^2+{\mathrm{\ω}}_1{({\mathrm{\μ}}_1-\mathrm{\μ})}^2)$

Wherein, operational symbol arg[] be to return to optimal value operation, operational symbol max[] be to find the solution maxima operation, Th is segmentation threshold, ω ₀And ω ₁Respectively that moving target is at the probability of background area and target area appearance, μ ₀The mean value of background area, μ ₁Be the mean value of target area, μ is the mean value of background and target area, and their expression formula is as follows respectively:

${\mathrm{\ω}}_0=\underset{v=0}{\overset{\mathrm{Th}}{\mathrm{\Σ}}}p\left(v\right),$ ${\mathrm{\μ}}_0=\frac{1}{{\mathrm{\ω}}_0}\underset{v=0}{\overset{\mathrm{Th}}{\mathrm{\Σ}}}v\·p\left(v\right)$

${\mathrm{\ω}}_1=\underset{v=\mathrm{Th}+1}{\overset{T}{\mathrm{\Σ}}}p\left(v\right),$ ${\mathrm{\μ}}_1=\frac{1}{{\mathrm{\ω}}_1}\underset{v=\mathrm{Th}+1}{\overset{T}{\mathrm{\Σ}}}v\·p\left(v\right)$

μ＝ω ₀μ ₀+ω ₁μ ₁

Wherein, p (v) is that the gradation of image value is the probability of the pixel appearance of v, and T is gray level, and Th is segmentation threshold.P (v) expression formula is as follows:

$p\left(v\right)=\frac{1}{240\×320}\underset{X_{\mathrm{out}}(i,j,k)=v}{\mathrm{\Σ}1}$

Wherein,

Expression is used for computed image gray-scale value X _out(i, j, k) equals the number of pixels of v, v=1, and 2 ..., 255,1≤i≤240,1≤j≤320.

Claims (6)

1. the infrared motion target real-time detection method that Adaptive background subtraction is estimated, is characterized in that, comprises the steps:

(1) initialization background model BG (0);

(2) the original infrared image X of input k frame _org(k), k=1,2 ..., K;

(3) to the original infrared image X of the described k frame of step (2) _org(k) spatial domain heterogeneity noise and time domain random noise suppress, and output k frame suppresses the image X after noise _Deal(k);

(4) the k frame infrared image X after the described inhibition noise of step of registration (3) _Deal(k), the k frame infrared image X after the output registration _reg(k);

(5) background model is upgraded: by the k frame infrared image X after the described registration of adaptive forgetting factor method step of updating (4) _reg(k) background model BG (k);

(6) with the original infrared image X of k frame _org(k) make calculus of differences with background model BG (k), the X as a result of output calculus of differences _out(k);

(7) according to the described X of the theoretical calculation procedure (6) of maximum between-cluster variance _out(k) the optimal segmenting threshold Th of target and background in _opt, complete the detection of moving target;

In described step (3), use finite support territory high frequency constant statistics Nonuniformity Correction method to the original infrared image X of k frame _org(k) spatial domain heterogeneity noise suppresses, and the method for described inhibition is:

Utilize gauss low frequency filter g with the original infrared image X of k frame _org(k) be divided into high fdrequency component

And low frequency component

Two parts:

$X_{\mathrm{org}}\left(k\right)=X_{\mathrm{org}}^{\mathrm{high}}\left(k\right)+X_{\mathrm{org}}^{\mathrm{low}}\left(k\right)$

$X_{\mathrm{org}}^{\mathrm{low}}\left(k\right)=X_{\mathrm{org}}\left(k\right)\⊗g$

In formula, operational symbol

The expression convolution operation;

Gauss low frequency filter g (i, the j) expression formula that coordinate (i, j) is located is shown below;

$(i,j)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}{e}^{-\left(\frac{i^2+j^2}{2{\mathrm{\σ}}^2}\right)}$

In formula, σ is the variance of gauss low frequency filter g;

Calculate the original infrared image X of k frame _org(k) biasing coefficient O ^High(k) and gain coefficient G ^High(k), the biasing coefficient O that locates of k frame coordinate (i, j) ^High(i, j, k) and gain coefficient G ^High(i, j, k) expression formula is shown below respectively:

$O_{\mathrm{tem}}^{\mathrm{high}}(i,j,k)=\frac{1}{\mathrm{Num}}\×X_{\mathrm{org}}^{\mathrm{high}}(i,j,k)+(1-\frac{1}{\mathrm{Num}})\×O^{\mathrm{high}}(i,j,k-1)$

$G^{\mathrm{high}}(i,j,k)=\frac{1}{\mathrm{Num}}\×|X_{\mathrm{org}}^{\mathrm{high}}(i,j,k)-O^{\mathrm{high}}(i,j,k-1)|+(1-\frac{1}{\mathrm{Num}})\×G^{\mathrm{high}}(i,j,k-1)$

O ^high(i,j,0)＝0

$O_{\mathrm{tem}}^{\mathrm{high}}(i,j,k)=0$

G ^high(i,j,0)＝1

In formula,

Candidate's coefficient of setovering,

The original infrared image X of k frame _org(k) the pixel value X that locates of coordinate (i, j) _orgThe high fdrequency component of (i, j, k), Num is the accumulation frame number, ε is error constant, $|O_{\mathrm{tem}}^{\mathrm{high}}(i,j,k)-O^{\mathrm{high}}(i,j,k-1)|\≤\mathrm{\ϵ}$ It is the finite support territory of finite support territory high frequency constant statistics Nonuniformity Correction method;

Therefore, X _org(i, j, k) output after finite support territory high frequency constant statistics Nonuniformity Correction method processing

Expression formula is shown below:

$X_{\mathrm{deal}}^{\mathrm{nu}}(i,j,k)=(X_{\mathrm{org}}(i,j,k)-O^{\mathrm{high}}(i,j,k))/G^{\mathrm{high}}(i,j,k)$

Wherein,

It is the image after the heterogeneity squelch;

In described step (5), the original infrared image X of k frame _org(k) background model BG (i, j, the k) expression formula located of coordinate (i, j) is shown below:

BG(i,j,k)＝(1-α(i,j,k))×X _reg(i,j,k)+α(i,j,k)×BG(i,j,k-1)

In formula, X _reg(i, j, k) is the original infrared image X of k frame _org(k) the registration output located of coordinate (i, j), α (i, j, k) is the original infrared image X of k frame _org(k) forgetting factor located of coordinate (i, j), expression formula is shown below:

$\mathrm{\α}(i,j,k)=\frac{|H(i,j,k)-H(i,j,k-1)|}{H(i,j,k-1)}$

In formula, H (i, j, k) is the original infrared image X of k frame _org(k) M centered by coordinate (i, j) ₁* N ₁The neighborhood entropy of the local neighborhood of size.

2. the infrared motion target real-time detection method estimated of Adaptive background subtraction according to claim 1, is characterized in that, in described step (1), utilizes initial N frame original image { X _org(k) | k=1,2 ..., N} calculates background model BG (0), and the expression formula of the background value BG that coordinate (i, j) is located (i, j, 0) is shown below:

BG(i,j,0)＝median[X _org(i,j,1),X _org(i,j,2),...,X _org(i,j,k)，...,X _org(i,j,N)]

In formula, operational symbol median[] expression finds the solution median operation, X _orgThe original infrared image X of (i, j, k) expression k frame _org(k) pixel value located of coordinate (i, j).

3. the infrared motion target real-time detection method estimated of Adaptive background subtraction according to claim 1, is characterized in that, to the image of k frame after the heterogeneity squelch

Carry out the time domain random noise and suppress, the image X of output k frame after the time domain random noise suppresses _Deal(k):

$X_{\mathrm{deal}}^H(i,j,k)=\mathrm{median}\left[X_{\mathrm{deal}}^{\mathrm{nu}}(i,j+q^{\′},k)\right|q^{\′}=-2,-1,\mathrm{0,1,2}]$

$X_{\mathrm{deal}}(i,j,k)=\mathrm{median}\left[X_{\mathrm{deal}}^H(i+p^{\′},j,k)\right|p^{\′}=-2,-\mathrm{1,0,1,2}]$

Wherein, operational symbol median[] expression finds the solution median operation,

Right

The Output rusults of medium filtering in the horizontal direction, X _Deal(i, j, k) is right

The Output rusults of medium filtering, namely right in the vertical direction

Output image after the time domain random noise suppresses.

4. the infrared motion target real-time detection method estimated of Adaptive background subtraction according to claim 1, is characterized in that, in described step (4), adopts the k frame infrared image X after template registration method output registration _reg(k):

At first, the image X after the k-1 frame suppresses noise _Deal(k-1) choose size centered by coordinate (i, j) and be the standard form of P * Q;

Secondly, adopt square error function MSE as criterion, the image X after the k frame suppresses noise _Deal(k) in centered by coordinate (i, j) size mate for the image-region of M * N, P＜M, Q＜N, calculating least mean-square error MSE _min, least mean-square error MSE _minCorresponding coordinate (i+ Δ i, j+ Δ j) is registration position;

Square error function MSE expression formula is as follows:

$\mathrm{MSE}(i+\mathrm{\Δi},j+\mathrm{\Δj},k)=\frac{1}{P\×Q}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}\underset{q=1}{\overset{Q}{\mathrm{\Σ}}}{(X_{\mathrm{deal}}(i+\mathrm{\Δi}+p-\frac{P}{2},j+\mathrm{\Δj}+q-\frac{Q}{2},k)-X_{\mathrm{deal}}(i+p-\frac{P}{2},j+q-\frac{Q}{2},k-1))}^2$

Least mean-square error MSE _minExpression formula is as follows:

MSE _min(i+Δi,j+Δj,k)＝min{MSE(i+Δi,j+Δj,k)}

Wherein, (Δ i, Δ j) expression is with respect to the displacement of coordinate (i, j) ,-(M-P)/2≤Δ i≤(M-P)/2 ,-(N-Q)/2≤Δ j≤(N-Q)/2;

At last, to the image X after k frame inhibition noise _Deal(k) carry out registration:

X _reg(i,j,k)＝X _deal(i+Δi,j+Δj,k)。

5. the infrared motion target real-time detection method estimated of Adaptive background subtraction according to claim 1, is characterized in that, in described step (6), and the original infrared image X of k frame _org(k) the output X that locates of coordinate (i, j) _out(i, j, k) expression formula is:

X _out(i,j,k)＝X _org(i,j,k)-BG(i,j,k)

In formula, X _org(i, j, k) is the original infrared image X of k frame _org(k) pixel value located of coordinate (i, j), BG (i, j, k) is the pixel value that background model BG (k) coordinate (i, j) is located.

6. the infrared motion target real-time detection method estimated of Adaptive background subtraction according to claim 1, is characterized in that, in described step (7), and Th _optExpression formula is as follows:

${\mathrm{Th}}_{\mathrm{opt}}=\arg \underset{0\≤\mathrm{Th}\≤255}{\max }({\mathrm{\ω}}_0{({\mathrm{\μ}}_0-\mathrm{\μ})}^2+{\mathrm{\ω}}_1{({\mathrm{\μ}}_1-\mathrm{\μ})}^2)$

Wherein, operational symbol arg[] be to return to optimal value operation, operational symbol max[] be to find the solution maxima operation, Th is segmentation threshold, ω ₀And ω ₁Respectively that moving target is at the probability of background area and target area appearance, μ ₀The mean value of background area, μ ₁Be the mean value of target area, μ is the mean value of background and target area, and their expression formula is as follows respectively:

${\mathrm{\ω}}_0=\underset{v=0}{\overset{\mathrm{Th}}{\mathrm{\Σ}}}p\left(v\right),$ ${\mathrm{\μ}}_0=\frac{1}{{\mathrm{\ω}}_0}\underset{v=0}{\overset{\mathrm{Th}}{\mathrm{\Σ}}}v\·p\left(v\right)$

${\mathrm{\ω}}_1=\underset{v=\mathrm{Th}+1}{\overset{T}{\mathrm{\Σ}}}p\left(v\right),$ ${\mathrm{\μ}}_1=\frac{1}{{\mathrm{\ω}}_1}\underset{v=\mathrm{Th}+1}{\overset{T}{\mathrm{\Σ}}}v\·p\left(v\right)$

μ＝ω ₀μ ₀+ω ₁μ ₁

In formula, p (v) is that the gradation of image value is the probability of the pixel appearance of v, and T is gray level;

The expression formula of p (v) is: $p\left(v\right)=\frac{1}{M_2\×N_2}\underset{X_{\mathrm{out}}(i,j,k)=v}{\mathrm{\Σ}1}$

Wherein,

Expression is used for computed image gray-scale value X _out(i, j, k) equals the number of pixels of v, M ₂And N ₂Columns and the line number of difference presentation video.

Images