CN101261736A - Collaborative detection method for multi-source image motive target - Google Patents

Abstract

The invention discloses a cooperative detection method for the movable target of a multi-source image which firstly carries out image registration and unifies the coordinate system of the multi-source image for considering the multi-source image; carries out movable target detection on the sequence of each multi-source image; evaluates each detection result and corrects each detection result according to the feedback of the evaluation result until obtaining a more reliable detection output; as feedback correction can carry out fully understanding on the information of the multi-source image, a false target with smaller credit is abandoned, thus reducing the false alarm rate, increasing the confidence coefficient of the real target and improving the detection rate. By utilizing multispectral information and adopting a feedback closed ring to carry out result correction, the cooperative detection method for the movable target of the multi-source image of the invention can better solve the problems of shielding and shadow; the average detecting rate is improved from the 88.0 percent of the prior art to 92.3, thus avoiding detection miss and detection leakage.

Description

Collaborative detection method for multi-source image motive target

Technical field

The present invention relates to a kind of collaborative detection method for multi-source image motive target.

Background technology

Document " distant object based on multi-sensor information fusion detects infrared technique, 2006, Vo1.28 (12), p695-698 " discloses a kind of moving target detection method that merges based on the feature level.This method at first adopts frame-to-frame differences accumulation algorithm to obtain the moving region, carries out moving target extraction and feature level then and merges.This algorithm has carried out creditability measurement at fusing stage to the testing result of each source images, by creditability measurement the different sensors result is weighted summation, but the result is not fed back to processing unit, can't make full use of multi-sensor information, the average detected rate has only 88.0%.

Summary of the invention

In order to overcome the low deficiency of prior art accuracy of detection, the invention provides a kind of collaborative detection method for multi-source image motive target, utilize multispectral information, adopt feedback closed loop to carry out correction as a result, can be good at problems such as solution is blocked, shade, can improve accuracy of detection.

The technical solution adopted for the present invention to solve the technical problems: a kind of collaborative detection method for multi-source image motive target is characterized in that comprising the steps:

(a) at first use template matching method to finish coupling between the unique point, utilize the transformation parameter between the unique point estimated image then, and choose optimum one group as final transformation parameter, finish image registration;

(b) by continuous several two field pictures are analyzed, obtain the noise and the noise distribution density of imaging system,, adopt process of iteration, obtain cutting apart high threshold according to noise density; Then, be configured to half an of normal distribution, obtain cutting apart low threshold value; At last, cut apart for twice by high threshold and low threshold value, and further handle, obtain the visible images moving target in conjunction with mathematical morphology;

(c) at first design three adaptive thresholds that successively decrease successively, current frame image just is divided into four layers by gray-scale value, and the three first layers higher to gray-scale value, constructed one group of adaptive height threshold value respectively according to its gamma characteristic; Then the three first layers difference image of connecting is detected and carries out the neighborhood merging of layer interior testing result; Adopt mathematical morphology that testing result is further processed at last, obtain the infrared image moving target;

(d) the visible images moving target that obtains and the satisfaction of infrared image moving target result and setting are compared, determining to carry out closed loop still is the open loop operation, treat the result who enters feedback closed loop and carry out mark, when the testing result of sensor repeatedly enters feedback closed loop and still can not be when satisfied, then feedback closed loop should not be entered once more, circulation should be stopped; Have only when the number of times that is entered feedback closed loop by the result after estimating more after a little while, just carry out the correction of corresponding parameter.

The invention has the beneficial effects as follows: owing to utilize multispectral information, adopt feedback closed loop to carry out correction as a result, can be good at problems such as solution is blocked, shade, the average detected rate brings up to 92.3% by 88.0% of prior art, has avoided flase drop and omission.

Below in conjunction with drawings and Examples the present invention is elaborated.

Description of drawings

Accompanying drawing is the process flow diagram of collaborative detection method for multi-source image motive target of the present invention.

Embodiment

With reference to accompanying drawing.

One, image registration.

The approximate range of at first manual selected characteristic point then with the unique point of the angle point in this scope as image, uses the Harris factor to extract.

Secondly, carry out characteristic matching.Feature Points Matching was divided into for two steps, and its detailed process is as follows:

1, unique point is slightly mated.

Thick coupling is F ₁, F ₂In to set up one in each unique point be (2k+1) * (2k+1) template at center with this unique point.Select F then ₁In a unique point, with its template and F ₂In the template at each reference mark compare, calculate the correlativity of two templates.

F ₁Middle unique point (x _i, y _i) and F ₂Middle unique point (x ' _j, y ' _j) degree of correlation be calculated as follows:

${\mathrm{cor}}_{\mathrm{ij}}=\frac{\underset{u=-k}{\overset{k}{\mathrm{\Σ}}}\underset{v=-k}{\overset{k}{\mathrm{\Σ}}}[f_1(x_i+u,y_i+v)-{\stackrel{\‾}{f}}_1(x_i,y_i)][f_2(x_j^{\′}+u,y_j^{\′}+v)-{\stackrel{\‾}{f}}_2(x_j^{\′},y_j^{\′})]}{{\left\{\underset{u=-k}{\overset{k}{\mathrm{\Σ}}}\underset{u=-k}{\overset{k}{\mathrm{\Σ}}}{[f_1(x_i+u,y_i+v)-{\stackrel{\‾}{f}}_1(x_i,y_i)]}^2\underset{u=-k}{\overset{k}{\mathrm{\Σ}}}\underset{v=-k}{\overset{k}{\mathrm{\Σ}}}{[f_2(x_j^{\′}+u,y_j^{\′}+v)-{\stackrel{\‾}{f}}_2(x_j^{\′},y_j^{\′})]}^2\right\}}^{1/2}}---\left(1\right)$

Wherein: f ₁Expression I ₁The gray scale of unique point, f ₂Expression I ₂The gray scale of unique point, the mean value of all pixel grey scales in the f representation template.By above formula as can be known ,-1≤cor _Ij≤ 1.

Use formula (1), and threshold value T is set ₁, choose greater than T ₁Unique point right, the feature point set M that is slightly mated thus ₁={ (x _i, y _j) | 0＜i＜=n} and M ₂=(x ' _i, y ' _i) | 0＜i＜=n}, wherein M ₁In (x _i, y _i) and M ₂In (x ' _i, y ' _i) coupling, n represents the unique point logarithm that mates.

2, the smart coupling of unique point.

This paper is example with the affine model, introduces the work of the smart compatible portion of unique point.

The formula of affined transformation model is as follows:

$x=\frac{m_1{x}^{\′}+m_2{y}^{\′}+m_3}{m_7{x}^{\′}+m_8{y}^{\′}+1}---\left(2\right)$

$y=\frac{m_4{x}^{\′}+m_5{y}^{\′}+m_6}{m_7{x}^{\′}+m_8{y}^{\′}+1}---\left(3\right)$

From M ₁, M ₂In optional four pairs of matching characteristic points, (x ₁, y ₁), (x ₂, y ₂), (x ₃, y ₃), (x ₄, y ₄) and (x ' ₁, y ' ₁), (x ' ₂, y ' ₂), (x ' ₃, y ' ₃), (x ' ₄, y ' ₄).According to one group of transformation model parameter of affined transformation Model Calculation P _i=(m ₁, m ₂, m ₃, m ₄, m ₅, m ₆, m ₇, m ₈).

Calculate M ₂In other each unique point, as (x ' ₁, y ' ₁) (corresponding point in M1 are (x ₁, y ₁)), at parameter P _iFollowing corresponding point coordinate (x ' ₁, y ' ₁).At I ₁In extract with (x ' ₁, y ' ₁) be the template at center, calculate then (x ' ₁, y ' ₁) and (x ₁, y ₁) the degree of correlation.If its degree of correlation is greater than a certain threshold value T ₂, think that then this meets this group parameter to point.Statistics meets parameter P _iUnique point logarithm sum _i

Select the unique point logarithm of formula (4) sum as final coupling between image, and with that group parameter P of its correspondence parameter as the inter-image transformations model.At this moment not only reject the incorrect point of coupling in the thick coupling, also calculated each parameter of inter-image transformations model.

$\mathrm{sum}=\underset{i}{\max }\left({\mathrm{sum}}_i\right),(0<i\&le;C_n^4)---\left(4\right)$

Repeat said process, up to having selected M ₁, M ₂In the combination (C altogether that had a few _n ⁴Kind) till, just finished the work of characteristic matching part.

Two, visible images moving object detection.

In record present frame and preceding first frame, preceding the 3rd frame, the point set that gray-scale value does not change obtains point set 1.

In the record point set 1, in preceding second frame, the point set that gray-scale value changes obtains point set 2.

Calculate in every frame, because the density p of the noise that image-forming mechanism causes.

An initial arbitrarily high threshold is set, carries out iteration, when binaryzation, the density p of the little noise that obtains thinks that the threshold value that obtains this moment is exactly high threshold th during very near ρ ^*Promptly | ρ-ρ ' |≤ξ, wherein ξ is a very little number of prior appointment.

According to normal distribution half sets up noise model, tries to achieve low threshold value tl ^*

Present frame and former frame are carried out difference, difference image is carried out cutting apart for twice according to high threshold and low threshold value.

According to the neighborhood judgement, carry out neighborhood and merge, carry out morphology again and handle, obtain final testing result.

Three, infrared image moving object detection.

Segmentation threshold θ between step 1, three adaptation layers that successively decrease of calculating ₁, θ ₂, θ ₃, segmentation threshold adopts the OSTU method to calculate.And calculate and cut apart height threshold value Th in corresponding three groups of adaptation layers ₁, Tl ₁, Th ₂, Tl ₂, Th ₃, Tl ₃

The high threshold Th of high brightness layer ₁:

Th ₁＝θ ₁-θ ₂ (5)

The low threshold value Tl of high brightness layer ₁:

Tl ₁＝η ₁ (6)

The high threshold Th of intermediate light layer ₂:

Th ₂＝θ ₂-θ ₃ (7)

The low threshold value Tl of intermediate light layer ₂:

Tl ₂＝η ₂ (8)

The high threshold Th of low-light level layer ₃:

Th ₂＝θ ₃ (9)

The low threshold value Tl of low-light level layer ₃:

Tl ₃＝η ₃ (10)

η ₁, η ₂, η ₃Segmentation threshold for difference result in high brightness layer, intermediate light layer, the low-light level layer.Difference processing result in each layer is the difference of each brightness layer and its gray average, and it is significantly bimodal that its histogram is, and two peak-to-peak separation threshold values can think to be exactly low threshold value to be asked.

Step 2, to the high brightness layer:

(a) with current frame image f _t(x, y) middle gray-scale value is greater than θ ₁Pixel respectively with f _T-k(x, y), f _T-k-1(x, y), f _T-k-2(x, y) pixel of middle correspondence position is carried out the difference image computing, obtains f ₁(x, y), f ₂(x, y) and f ₃(x, y);

(b) to f ₁(x, y), f ₂(x, y) and f ₃(x y) uses Th respectively ₁And Tl ₁Carry out binaryzation, obtain f ' ₁(x, y), f ' ₂(x, y), f ' ₃(x, y) and f " ₁(x, y), f " ₂(x, y), f " ₃(x, y), promptly

$f_i^{\&prime;}(x,y)=\left\{\begin{array}{ccc}255& \mathrm{if}& f_i(x,y)\&GreaterEqual;{\mathrm{Th}}_1\\ 0& \mathrm{else}\end{array}\right.,f_i^{\&prime;\&prime;}(x,y)=\left\{\begin{array}{ccc}255& \mathrm{if}& f_i(x,y)\&GreaterEqual;{\mathrm{Tl}}_1\\ 0& \mathrm{else}\end{array}\right.---\left(11\right)$

(c) eliminate the influence of noise spot to testing result, if current frame image and front k, coming to the same thing after k+1 and the k+2 frame difference image binaryzation, then this pixel is target or background, otherwise thinks noise spot, that is:

$d_1^{\&prime;}(x,y)=\left\{\begin{array}{ccc}{f}_1^{\&prime;}(x,y)& \mathrm{if}& f_1^{\&prime;}(x,y)=f_2^{\&prime;}(x,y)=f_3^{\&prime;}(x,y)\\ 0& \mathrm{else}& \end{array}\right.---\left(12\right)$

$d_1^{\&prime;\&prime;}(x,y)=\left\{\begin{array}{ccc}{f}_1^{\&prime;\&prime;}(x,y)& \mathrm{if}& f_1^{\&prime;\&prime;}(x,y)=f_2^{\&prime;\&prime;}(x,y)=f_3^{\&prime;\&prime;}(x,y)\\ 0& \mathrm{else}& \end{array}\right.---\left(13\right)$

(d) at high threshold Th ₁The neighborhood of the testing result that obtains is with low threshold value Tl ₁The result who detects goes to replace, and finishes in the high brightness layer and detects, and obtains " seed region " for the first time.

Step 3, the pixel of brightness layer carries out the operation of step 2 at the neighborhood of for the first time " seed region " and in belonging to, and obtains the second time " seed region ".

Step 4, carry out the operation of step 3, obtain for the third time " seed region ", promptly final actual detection result at the neighborhood of for the second time " seed region " and the pixel that belongs to the low-light level layer.

Final actual detection result is carried out morphology handle, the fragment problems in detecting is handled.

The visible images moving target that obtains and the satisfaction of infrared image moving target result and setting are compared, and determining to carry out closed loop still is the open loop operation; Treat the result who enters feedback closed loop afterwards and carry out mark,, should stop circulation when the testing result of this sensor repeatedly enters feedback closed loop and still can not then should not enter feedback closed loop once more when satisfied; Have only when the number of times that is entered feedback closed loop by the result after estimating more after a little while, just carry out the correction of corresponding parameter.

After testing, use the average detected rate of method of the present invention to reach 92.3%.

Claims (1)

1, a kind of collaborative detection method for multi-source image motive target is characterized in that comprising the steps:

(a) at first use template matching method to finish coupling between the unique point, utilize the transformation parameter between the unique point estimated image then, and choose optimum one group as final transformation parameter, finish image registration;

(b) by continuous several two field pictures are analyzed, obtain the noise and the noise distribution density of imaging system,, adopt process of iteration, obtain cutting apart high threshold according to noise density; Then, be configured to half an of normal distribution, obtain cutting apart low threshold value; At last, cut apart for twice by high threshold and low threshold value, and further handle, obtain the visible images moving target in conjunction with mathematical morphology;

(c) at first design three adaptive thresholds that successively decrease successively, current frame image just is divided into four layers by gray-scale value, and the three first layers higher to gray-scale value, constructed one group of adaptive height threshold value respectively according to its gamma characteristic; Then the three first layers difference image of connecting is detected and carries out the neighborhood merging of layer interior testing result; Adopt mathematical morphology that testing result is further processed at last, obtain the infrared image moving target;

(d) the visible images moving target that obtains and the satisfaction of infrared image moving target result and setting are compared, determining to carry out closed loop still is the open loop operation, treat the result who enters feedback closed loop and carry out mark, when the testing result of sensor repeatedly enters feedback closed loop and still can not be when satisfied, then feedback closed loop should not be entered once more, circulation should be stopped; Have only when the number of times that is entered feedback closed loop by the result after estimating more after a little while, just carry out the correction of corresponding parameter.

Images