CN1251144C - Infra red detecting and tracing method for weak target under complex background condition - Google Patents

Abstract

The present invention relates to a method for detecting and tracing infrared weak targets under the condition of complex backgrounds. An image background is obtained through morphological filtering, the image background is subtracted from an original image, and a des-mean value image with the targets and noise is obtained. The image preprocessing technique based on the volume of an elliptic paraboloid is adopted, the signal-to-noise ratio of the processing result is increased, a detecting threshold value is selected by using a criterion of a constant false alarm rate, and a single frame test result is obtained. The false target is further removed by using the continuity of the motion trace of the target. The method of the nearest neighbor association is adopted to track the target, and kalman filtering based on a model with uniform motion in a straight line is adopted to update the target state. If the target is lost, the target is tenaciously tracked by using the trace for predicting a possibly position. The present invention simplifies the calculation under the condition of guaranteeing the algorithm accuracy to increase the real time of the algorithm and the performance of detecting and tracking the infrared target. The present invention with extensive market perspective and application value can be widely applied to various military and civil systems.

Description

The detection of infrared small object and tracking under the complex background condition

Technical field:

The present invention relates to the detection and the tracking of infrared small object under a kind of complex background condition, be a core technology of Infra-Red Search ﹠ Track System, precise guidance system, infraed early warning system, big visual field targeted surveillance system, satellite remote sensing system, safety check system etc., in all kinds of military, civilian systems, all can be widely used.

Background technology:

Infrared imagery technique is a kind of contactless measuring technology, and it can detect invisible heat radiation that target sends and the technology that converts visual picture to easily.Information is obtained the related research that key area is infrared detection technique and method, and its critical role becomes increasingly conspicuous.Owing to advantages such as infrared imagery technique has good concealment, investigative range is wide, bearing accuracy is high, identification camouflage ability is strong, operating distance is far away and lightweight is small and exquisite, low consumption is reliable enjoy favor, can be widely used in fields such as security monitoring, national defense and military and industrial automation detection.

Along with the development of infrared eye technology, thermal imaging system adds one dimension or two-dimentional opto-mechanical scanner from past employing unit or polynary discrete detector, has developed into the gazing type imaging device without optical mechaical scanning.Based on the infrared thermal imaging detection system of staring focal plane arrays (FPA), no matter on temperature control and spatial resolution, still on frame frequency and spectral response, all be greatly improved.Since the focal plane stare thermal imaging system exclusive premium properties, become the new and high technology that research and develop energetically countries in the world.As one of key link of Intelligentized Information, infrared target detection, imaging tracking and recognition technology are the bottleneck problem of puzzlement and restriction infrared imaging detection Practical Performance and technological difficulties and anxious to be solved always, caused at present domestic and international expert's great attention, and carried out deep, extensive studies around this problem.

In infrared target detection and tracing process, need intercept and capture as soon as possible and the locking tracking target, the technical barrier that detection of low signal-to-noise ratio infrared small object and tracking are faced under the complex background mainly contains:

1. target is little, does not have information such as size, shape and texture, and traditional image processing method can't be used;

2. background complexity, noisiness the unknown;

3. signal to noise ratio (S/N ratio) is low, and target is submerged among the ground unrest;

4. data volume is big, is difficult to real-time processing.

The researchist has proposed certain methods at target detection in the low signal-to-noise ratio infrared image and tracking both at home and abroad, as three-dimensional matched filtering, block sequential likelihood ratio detection method, dynamic programming, high-order correlation method etc., but calculated amount is very big, is difficult to handle in real time sequence of video images.Tracing it to its cause, mainly is not have a kind of effective image pre-processing method to improve the signal to noise ratio (S/N ratio) of handling the back image, causes target detection and association, track algorithm difficulty to increase; Do not utilize the weak characteristics of distant object maneuverability, reduced data processing procedure etc.

Summary of the invention:

The objective of the invention is to above-mentioned deficiency at prior art, the detection and the tracking of infrared small object under a kind of complex background condition are provided, carry out Digital Signal Analysis and Processing with imaging hardware system is supporting, the performance of the detection of raising system, imaging and tracking satisfies the performance requirement of domestic real system.

For realizing such purpose, in the technical scheme of the present invention, utilize the mathematical morphology filter technology to obtain the background of image earlier, original image is deducted the background image of acquisition, what obtain containing target and noise removes the average image.Because the imaging of point target in actual detector is not a point, but one " convex closure ".The present invention proposes thus infrared image is carried out image preconditioning technique based on the elliptic paraboloid volume, that adopts quadric surface to approach to contain noise removes average image local pixel value, obtain quadric coefficient value, and calculate quadric volume, with quadric volume as eigenwert.Utilize the character of quadric surface invariant further to simplify volume calculation.Utilize the constant false alarm rate criterion to select detection threshold, obtain the single frames testing result.According to the regularity of target travel, utilize the continuity of target trajectory further to remove false target.Simple relatively arest neighbors correlating method is adopted in target following, according to the characteristics a little less than the maneuverability of distant object, adopts based on the Kalman filtering of uniform rectilinear motion model and upgrades dbjective state.When track rejection, the possible position of the filter value outside forecast target in the moment keeps stable memory tracking power before adopting.

Method of the present invention comprises following concrete steps:

1. image goes average to handle (Demeaning): adopt 5 * 5 complete zero flat-top shape structural elements that infrared image is carried out gray scale unlatching and then closed mathematical morphology filter operation, compare less all kinds of bright, dark noise of size and remote little target with structural element to remove, obtain infrared image background, with original image subtracted image background, what obtain containing target and noise removes the average image.

2. based on the image pre-service of elliptic paraboloid volume: according to the image-forming principle of infrared imaging system, the imaging of point target in actual detector is not a point, but one " convex closure ", be similar to the downward opening elliptic paraboloid in the quafric curve, that adopts the secondary elliptical parabolic surface to approach to contain noise removes average image local pixel value, obtain quadric coefficient value, and calculate the volume that this local quadric surface is comprised, with quadric volume as eigenwert.Utilize the character of quadric surface invariant further to simplify volume calculation, the matrix inversion operation of having avoided quadrature similarity transformation method to introduce.Utilize the constant false alarm rate criterion to select detection threshold,, obtain the single frames testing result if eigenwert, is then thought the possibility target greater than detection threshold.

3. the multiframe of target is confirmed and followed the tracks of: continuous 3 frames of possible target that detect when single frames appear in 3 * 3 neighborhoods, then track initial.Employing is upgraded target trajectory based on the Kalman filtering of uniform rectilinear motion model, and generates oval tracking gate thus.If when falling into the measuring value of oval tracking gate and having only one, directly carry out target trajectory and upgrade; If the measuring value that falls into oval tracking gate during more than one, then adopts the arest neighbors data association to carry out target trajectory and upgrades, promptly target trajectory upgrades by the nearest measured value of distance one-step prediction value.

4. trajectory predictions behind the track rejection: adopting Kalman filtering to carry out track when upgrading, if track rejection, the possible position of a few frame targets behind constantly the state filtering value outside forecast keeps stable memory tracking power before then utilizing.

The present invention adopts mathematical morphology filter to obtain image background, and its algorithm can pass through the hardware Parallel Implementation, has improved processing speed greatly.In image pre-service based on the elliptic paraboloid volume, utilize the character of quadric surface invariant further to simplify volume calculation, avoided the matrix inversion operation of general real coefficient quadratic equation by introducing in the quadrature similarity transformation method standardisation process, the real-time and the robustness of algorithm have been improved, significantly improve the signal to noise ratio (S/N ratio) of output image, also simplified follow-up tracking and data association algorithm simultaneously greatly.According to the characteristics a little less than the distant object maneuverability, adopt based on the Kalman filtering of uniform rectilinear motion model and upgrade dbjective state, under the situation that guarantees tracking accuracy, simplified filtering calculating.The present invention improved the real-time of algorithm guaranteeing to have simplified calculating under the condition of arithmetic accuracy, improved the performance of infrared target detection, tracking greatly, can be widely used in all kinds of military, civilian systems, has vast market prospect and using value.

Description of drawings:

Fig. 1 is a disposal route The general frame of the present invention.

As shown in Figure 1, actual infrared image adopts the mathematical morphology filter technology to obtain to remove the average image earlier, carries out the image pre-service based on volume again, comprises quadric surface match and curved surface volume calculation, utilize the constant false alarm rate criterion to select detection threshold, obtain the single frames testing result.Employing is carried out target following based on the Kalman filtering and the arest neighbors correlating method of uniform rectilinear motion model.

Fig. 2 is the image pre-service result of the present invention's employing based on the elliptic paraboloid volume.

Wherein, Fig. 2 (a) is about 2 emulating image for signal to noise ratio (S/N ratio), and Fig. 2 (b) is a result, and signal to noise ratio (S/N ratio) brings up to 6.57.Fig. 2 (c) is that true infrared image adopts the image after morphologic filtering goes average, and Fig. 2 (d) is a result, and signal to noise ratio (S/N ratio) also is greatly improved.

Fig. 3 is target detection and tracing figure.

Wherein Fig. 3 the picture left above is true infrared image, and Fig. 3 top right plot is that morphologic filtering removes the average image, and Fig. 3 lower-left figure is the image pre-service result based on the elliptic paraboloid volume, the target of Fig. 3 bottom-right graph for confirming.

Embodiment:

In order to understand technical scheme of the present invention better, embodiments of the present invention are further described below in conjunction with accompanying drawing.

Fig. 1 is the detection of infrared small object and the disposal route The general frame of tracking under a kind of complex background condition of the present invention's proposition.The concrete implementation detail of each several part is as follows:

1. morphologic filtering goes average to handle

Adopt mathematical morphology filter to obtain infrared image background.With structural element b input picture f is carried out the gray scale expansion and corrodes being designated as respectively:

(f b)(x，y)＝max{f(x-s，y-t)+b(s，t)|(x-s)，(y-t)∈D _f，(s，t)∈D _b}

(1)

(f b)(x，y)＝min{f(x+s，y+t)-b(s，t)|(x+s)，(y+t)∈D _f，(x，t)∈D _b}

(2) D in the formula _fAnd D _bIt is respectively the field of definition of f and b.

With structural element b input picture f is carried out that gray scale is opened and closure is designated as respectively:

fοb＝(fb)b (3)

f·b＝(fb)b (4)

Open operation commonly used is eliminated and is compared the less bright noise of size with structural element in the reality, eliminate with closed procedure and compare the less dark noise of size with structural element, and keep the integral image gray-scale value and big bright, dark areas is unaffected substantially.Because distant object shows as little bright spot and speck in infrared image, therefore, infrared image is opened and then closed procedure, to remove all kinds of bright, dark noises and little target, acquisition image background with 5 * 5 complete zero flat-top shape structural elements.With original image subtracted image background, what obtain containing target and noise removes the average image.

2. based on the image pre-service of elliptic paraboloid volume

Quadric general equation is in the note space

F(x ₁，x ₂，x ₃)＝a ₁₁x ₁ ²+a ₂₂x ₂ ²+a ₃₃x ₃ ²+2a ₁₂x ₁x ₂+2a ₁₃x ₁x ₃+2a ₂₃x ₂x ₃

+2a ₁₄x ₁+2a ₂₄x ₂+2a ₃₄x ₃+a ₄₄＝0 (5)

Wherein, a ₁₁, a ₂₂, a ₃₃, a ₁₂, a ₁₃, a ₂₃Be not zero entirely.After orthogonal matrix diagonalization coordinate transform, surface equation becomes

F ' (x ₁', x ₂', x ₃')=λ ₁x ₁' ²+ λ ₂x ₂' ²+ λ ₃x ₃' ²+ 2a ₁₄' x ₁'+2a ₂₄' x ₂'+2a ₃₄' x ₃'+a ₄₄=0 (6) wherein works as λ ₁, λ ₂, λ ₃In have only one to be zero, might as well establish λ ₃=0, and λ ₁λ ₂＞0 o'clock, quadric surface was an elliptic paraboloid.

$I_2=\left|\begin{array}{cc}{a}_{11}& a_{12}\\ a_{12}& a_{22}\end{array}\right|+\left|\begin{array}{cc}{a}_{11}& a_{13}\\ a_{13}& a_{33}\end{array}\right|+\left|\begin{array}{cc}{a}_{22}& a_{23}\\ a_{23}& a_{33}\end{array}\right|={\mathrm{\λ}}_1{\mathrm{\λ}}_2+{\mathrm{\λ}}_2{\mathrm{\λ}}_3+{\mathrm{\λ}}_3{\mathrm{\λ}}_1$

It is a quadric little variable.

Go in the average image at morphologic filtering, establish with current pixel (x ', y ') and can be expressed as for the pixel in (2k+1) * (2k+1) neighborhood at center,

I[x '+i, y '+j] || i|≤k, | j|≤k} approaches the pixel value that contains noise with quadric surface

z＝ax ²+bxy+cy ²+dx+ey+f (7)

Can obtain parameter in the equation by the least-squares algorithm of classics, by minimizing target function

${\left|\right|\mathrm{Ax}-I\left|\right|}^2={\mathrm{\Σ}}_{i=-k}^k{\mathrm{\Σ}}_{j=-k}^k{({\mathrm{ai}}^2+\mathrm{bij}+{\mathrm{cj}}^2+\mathrm{di}+\mathrm{ej}+f-I(x^{\′}+i,y^{\′}+j))}^2$

Obtain least square solution

x＝(A ^TA) ^-1A ^TI

Wherein

x＝[a，b，c，d，e，f] ^T，

$A=\left[\begin{array}{cccccc}{k}^2& k^2& k^2& -k& -k& 1\\ & \·\·\·& & & \·\·\·& \\ i^2& \mathrm{ij}& j^2& i& j& 1\\ & \·\·\·& & & \·\·\·& \\ k^2& k^2& k^2& k& k& 1\end{array}\right]\begin{array}{c}(-k,-k)\\ \\ (i,j)\\ \\ (k,k)\end{array}$

I＝(I(x′-k，y′-k)，…，I(x′-i，y′-j)，…，I(x′+k，y′+k)) ^T

After the elliptic paraboloid curve fitting, can obtain its extreme point by extremum conditions

$\frac{\mathrm{\δz}}{\mathrm{\δx}}(x_0,y_0)=\frac{\mathrm{\δz}}{\mathrm{\δy}}(x_0,y_0)=0$

Draw

$x_0=\frac{2\mathrm{cd}-\mathrm{be}}{b^2-4\mathrm{ac}},y_0=\frac{2\mathrm{ae}-\mathrm{bd}}{b^2-4\mathrm{ac}}$

(7) formula is carried out the translational coordination conversion

x′＝x-x ₀，y′＝y-y ₀

Obtain

z＝ax′ ²+bx′y′+cy′ ²+f′ (8)

Wherein $f^{\′}=\frac{{\mathrm{cd}}^2-\mathrm{bde}+{\mathrm{ae}}^2}{b^2-4\mathrm{ac}}+f$

(8) formula is turned to the quadratic form standard form with the orthogonal transformation method with it, then have

z＝λ ₁x ²+λ ₂y ²+f′ (9)

If the opening of wanting elliptic paraboloid down, and extreme point is positioned at the top on xoy plane, need satisfy condition

λ ₁＜0，λ ₂＜0，f′＞0 (10)

Adopt energy that little target comprises as characteristic quantity, it can pass through

V＝λ ₁λ ₂f′＝I ₂×f′ (11)

Characterize.Because at the general real coefficient quadratic equation of above-mentioned quadrature similarity transformation methodization is the inversion operation that will introduce matrix in the standard form calculating process, so the algorithm computation amount is big and matrix inversion can cause the algorithm instability.Utilize the character of quadric surface invariant to come further abbreviation computing below, avoided matrix inversion.

Adopt method of completing the square to turn to the quadratic form canonical form to (8) formula, do conversion

y＝y′ (12)

$\begin{array}{c}x=x^{\′}+\frac{b}{2a}{y}^{\′}\\ y=y^{\′}\end{array}---\left(12\right)$

Obtain

$z={\mathrm{ax}}^2+(c-\frac{b^2}{4a})y^2+f^{\′}---\left(13\right)$

According to the Reversible Linear Transformation quadratic form is the inertial theorem of standard form, and the positive and negative number in the coefficient of standard form is constant as can be known.Therefore condition (10) is equivalent to:

$a<0,c-\frac{b^2}{4a}<0,f^{\&prime;}>0----\left(14\right)$

For (8) formula, each coefficient is in its corresponding quadric surface general equation (5)

a ₁₁＝a，a ₂₂＝c，a ₃₃＝0， $a_{12}=\frac{b}{2},$ a ₁₃＝0，a ₂₃＝0

The quadric surface invariant

$I_2=\left|\begin{array}{cc}{a}_{11}& a_{12}\\ a_{12}& a_{22}\end{array}\right|+\left|\begin{array}{cc}{a}_{11}& a_{13}\\ a_{13}& a_{33}\end{array}\right|+\left|\begin{array}{cc}{a}_{22}& a_{23}\\ a_{23}& a_{33}\end{array}\right|=\mathrm{ac}-\frac{b^2}{4}$

Utilize quadric surface invariant character as can be known, little target energy eigenwert (11) formula is equivalent to

$V=I_2\&times;f^{\&prime;}=(\mathrm{ac}-\frac{b^2}{4})\&times;f^{\&prime;}---\left(15\right)$

As seen, (15) formula has been avoided matrix inversion operation, and algorithm performance and (11) formula are identical.

For with the concentration of energy of target center, (15) formula is taken advantage of a scale-up factor to target

$V_c=\frac{V}{1+{\left(x_0\right)}^2+{\left(y_0\right)}^2}----\left(16\right)$

Figure 2 shows that the simulation result based on the image pre-processing method of elliptic paraboloid volume, wherein (a) is about in 2 the Gaussian noise for target is submerged in signal to noise ratio (S/N ratio), (b) is result, and signal to noise ratio (S/N ratio) is brought up to about 6.5; (c) remove the average image for actual infrared image after by mathematical morphology filter, (d) be result, as can be seen, the signal to noise ratio (S/N ratio) of exporting the result has obtained very big enhancing.

3. the multiframe of target is confirmed and is followed the tracks of

To image pre-service result, can determine that according to given false alarm rate the threshold value that detects obtains the single frames testing result based on volume.Because the target range sensor is far away, a little less than the maneuverability of target, and the frame speed of image is very fast, so target can be approximate with uniform rectilinear motion model.

The tracker model: writ attitude vector is $X\left(K\right)={[x\left(k\right),\stackrel{\&CenterDot;}{x}\left(k\right),y\left(k\right),\stackrel{\&CenterDot;}{y}\left(k\right)]}^T,$ Measuring vector is Z (k)=[x (k), y (k)] ^T, system equation can be written as

X(k+1)＝ΦX(k)+Gw(k)

Z(k)＝HX(k)+v(k) (17)

Wherein

$\mathrm{\&Phi;}=\left[\begin{array}{cccc}1& T& 0& 0\\ 0& 1& 0& 0\\ 0& 0& 1& T\\ 0& 0& 0& 1\end{array}\right],G=\left[\begin{array}{cc}{T}^2/2& 0\\ T& 0\\ 0& T^2/2\\ 0& T\end{array}\right],H=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 0& 1& 0\end{array}\right],$

W (k), v (k) are the zero-mean white Gaussian noise of Q (k) and R (k) for variance.

Track initial: if continuous 3 frames of possible target that detect appear in 3 * 3 neighborhoods, track initial then.

Tracking gate and data association: suppose being estimated as of k-1 state vector constantly

The one-step prediction of state vector is

$\hat{X}\left(k\right|k-1)=\mathrm{\&Phi;}\hat{X}(k-1|k-1)$

Then the new breath of system is

$d\left(k\right)=Z\left(k\right)-H\hat{X}\left(k\right|k-1)----\left(18\right)$

New breath variance battle array is

S(k)＝HP(k|k-1)H ^T+R(k)

Wherein, P (k|k-1) is the one-step prediction variance.

The norm of the new vectorial d of breath of order (k) is

g(k)＝d ^T(k)S ^-1(k)d(k) (19)

Wherein, g (k) obeys χ _M ²Distribute, M is for measuring dimension.

A definition ellipsoid (being called tracking gate again) in measurement space makes and measures with certain probability distribution in tracking gate

${\tilde{V}}_k\left(\mathrm{\&gamma;}\right)=[Z:g\left(k\right)\&le;\mathrm{\&gamma;}]----\left(20\right)$

Wherein, γ can pass through χ ²Distribution table checks in.

If after certain frame is handled, in tracking gate, have only an echo, then target trajectory directly upgrades; If have in the tracking gate more than an echo, then target trajectory upgrades by the nearest measured value of distance one-step prediction value.

The track of target upgrades by the standard Kalman filtering algorithm

X(k|k-1)＝ΦX(k-1|k-1)

P(k|k-1)＝ΦP(k-1|k-1)Φ ^T+GQ(k-1)G ^T

K(k)＝P(k|k-1)H ^T[HP(k|k-1)H ^T+R] ^-1

X(k|k)＝X(k|k-1)+K(k)[Z(k)-H(k)X(k|k-1)]

P(k|k)＝[I-K(k)H]P(k|k-1) (21)

Target detection and tracking are as shown in Figure 3, the picture left above is original infrared image, top right plot be mathematical morphology filter obtain remove the average image, lower-left figure is the image pre-service result based on the elliptic paraboloid volume, the target of bottom-right graph for confirming marked by highlighted cross.

4. trajectory predictions behind the track rejection

In target detected and follow the tracks of after, if target is blocked by of short duration, even or by after the aforesaid image pre-service, when still having some frames not detect target, can be according to target positional information and motion state before this, dope next step possible position of target, when target occurs once more, but tenacious tracking and be unlikely to lose objects still.When distance is far away, a little less than the maneuverability of infrared target, can be by the filter value of the aforementioned dbjective state vector that draws based on the target following of uniform rectilinear motion model, outside forecast draws the possible position of target.

Suppose being estimated as of k state vector constantly Ensuing a few frame target is not detected, and then the predicted value in the target of k+n frame is:

$\hat{X}(k+n|k)={\mathrm{\&Phi;}}^n\hat{X}\left(k\right|k)----\left(22\right)$

In actual applications, along with the postponement of time and possible target maneuver, the tracking data in past is more and more uncorrelated with following quiet condition, and along with the increase of n, precision of prediction can descend.Generally choose n＜8.

Claims (1)

1, the detection and the tracking of infrared small object under a kind of complex background condition is characterized in that comprising following concrete steps:

1) image goes average to handle: adopt 5 * 5 complete zero flat-top shape structural elements that infrared image is carried out gray scale unlatching and then closed mathematical morphology filter operation, less all kinds of bright, dark noise of size and remote Weak target are compared in removal with structural element, obtain infrared image background, with original image subtracted image background, what obtain containing target and noise removes the average image;

2) based on the image pre-service of elliptic paraboloid volume: that adopts the secondary elliptical parabolic surface to approach to contain noise removes average image local pixel value, obtain quadric coefficient value, and calculate the volume that this local quadric surface is comprised, with quadric volume as eigenwert, utilize the constant false alarm rate criterion to select detection threshold, if eigenwert, is then thought the possibility target greater than detection threshold, obtain the single frames testing result;

3) multiframe of target is confirmed and followed the tracks of: continuous 3 frames of possible target that detect when single frames appear in 3 * 3 neighborhoods, then track initial; Employing is upgraded target trajectory based on the Kalman filtering of uniform rectilinear motion model, and generate oval tracking gate thus, when the measuring value that falls into oval tracking gate has only one, directly carrying out target trajectory upgrades, the measuring value that falls into oval tracking gate is during more than one, then adopt the arest neighbors data association to carry out target trajectory and upgrade, promptly target trajectory upgrades by the nearest measured value of distance one-step prediction value;

4) trajectory predictions behind the track rejection: when the employing Kalman filtering was carried out the track renewal, if track rejection, the possible position of a few frame targets behind the state filtering value outside forecast in the moment kept stable memory tracking power before then utilizing.

Images