CN101567087A - Method for detecting and tracking small and weak target of infrared sequence image under complex sky background - Google Patents

Abstract

The invention relates to a method for detecting and tracking a small and weak target of an infrared sequence image under complex sky background, which comprises the following steps of: (1) adopting a self-adaptive background eliminating algorithm which combines with the small and weak target characteristic of an image and Top-hat conversion for conducting self-adaptive background elimination and inhibiting the influence of complex background and part of noise on the small and weak target detection; (2) using an iterative threshold algorithm which is suitable for a dim-background image for dividing the possible target area in the dim-background image after background elimination; (3) constructing a dilation accumulation algorithm according to the motion law of the target in a plurality of continuous frames, and distinguishing possible false-alarm target areas; (4) using CB algorithm and logic and operation based on contour structure elements for inhibiting the false-alarm targets; (5) taking the areas in which the false-alarm targets are inhibited as possible target area which is then compared with the track prediction result in the step (7) so as to realize the updating of track; (6) conducting track judgment; and (7) conducting track prediction.

Description

The infrared sequence image under complex sky background Weak target detects and tracking

Technical field

The present invention relates to a kind of infrared sequence image under complex sky background Weak target and detect and tracking, belong to digital image processing field, relate generally to mathematical morphology, infrared image and target detection and tracking technique.All kinds of based on infrared image Target Recognition or navigational system in be widely used.

Background technology

The research of Weak target detection and tracking technique is civilian under the complex sky background, have in space flight and the military affairs important application (referring to document: Zhang Tianxu etc. determine the motion weakness target detection of width searches trend pass filtering based on three-dimensional. the pattern-recognition wall bulletin, 2007,28 (2): 246-253. (Tianxu Zhang, Meng Li, Zhengrong Zuo, Weidong Yang, Xiechang Sun.Movingdim point target detection with three-dimensional wide-to-exact search directional filtering.PatternRecognition Letters, 2007,28 (2): 246-253.)) be Flame Image Process,, Target Recognition, the important research field of gordian techniquies such as Navigation And Guidance.Because image-forming range far reaches the influence of complex sky background, the little target signal to noise ratio in the infrared image is very low, be in the complex background, movement velocity is unknown and do not have definite shape information, and this has increased the difficulty of infrared small object detection and tracking greatly.At this moment, infrared small object detects and the validity of following the tracks of under the complex background in order effectively to improve, must effectively utilize the detectability of the characteristic enhancing target of target, the difficulty that target trajectory is confirmed when simplifying target following improves the validity of target detection and tracking.

In order effectively to improve the validity of Weak target detection and tracking under the complicated infrared background, the various methods of target property of effectively utilizing are suggested.These methods comprise:

(1) based on the method for filtering (referring to document: [1] Dai Sipande etc. based on the little target detection of maximum average and maximum intermediate value. little echo signal of SPIE and data processing collection of thesis, 1999,74-83. ([1] Deshpande S.D., RondaM.H.Chan Er, V., Ph..Max-Mean and Max-Median filters for detection of small-targets.SPIE, Signal and Data Processing of Small Targets, 1999,74-83.) [2] match husband etc. the optimization of point target tracking filter. IEEE's avionics system transactions, 2000,36 (1): 15-25. ([2] Cafer C.E., Silverman J., Mooney J.M..Optimization of point target tracking filters.IEEE Transactions onAerospace and Electronic Systems, 2000,36 (1): 15-25.) [3] but Fu Lanqi etc. based on the enhancing of little target detectability under the background context that improves two-dimensional adaptive grid algorithm. IEEE's Flame Image Process transactions, 1997,6 (3): 383-397. ([3] Ffrench P.A., Zeidler J.R., Ku W.H..Enhanced detectability of smallobjects in correlated clutter using an improved 2-D adaptive lattice algorithm.IEEETransactions on Image Processing, 1997,6 (3): 383-397.))

The basic thought of these class methods is to construct corresponding wave filter according to target property, thereby by this wave filter image is carried out filtering filtering complex background, reaches the purpose that improves signal noise ratio (snr) of image, simplifies the difficulty of target detection.These class methods effect when signal noise ratio (snr) of image is low is relatively poor, or related parameter extremely difficulty choose, or algorithm is extremely complicated and be difficult to real-time application.

(2) based on the method for small echo (referring to document: [1] Chen Xiaozhong, Sun Huayan. the image object based on energy feature detects. infrared and laser engineering, 2001,30 (1): 30-32.[2] Ma Laite. multi-resolution signals resolution theory: Wavelet representation for transient. IEEE's pattern analysis and machine intelligence transactions, 1989,11 (7): 674-693. ([2] Mallat, S.G..Theory formultiresolution signal decomposition:the wavelet representation.IEEE Trans.PAMI, 1989,11 (7): 674-693.) [3] Gray Si etc. based on the detection of weak target in the FLIR (Forward-Looking Infrared) image of multi-scale transform. SPIE's collection of thesis, 1994,2262:62-71. ([3] Gregoris, D.J., et al.Detection of dim targets inFLIR imagery using multiscale transforms.Proceedings of SPIE, 1994,2262:62-71.))

These class methods are mainly utilized the grain background and the interesting target district of the multiscale analysis method differentiate between images of small echo.The image information that small echo generates on different scale provides the characteristic information that target and complex background can be made a distinction, and promptly the feature of target and background on the small echo different scale has obvious difference when with small echo target image being analyzed.In view of the above, various methods based on small echo are applied to the detection and the tracking of the little target of complex background preferably.But the detection algorithm based on small echo can not be applied to the lower situation of signal noise ratio (snr) of image well.

(3) based on morphology methods (referring to document: [1] Dong Yanzhi, Shen Tongsheng, Zhou Xiaodong etc. the application in infrared image detects of wavelet transformation and mathematical morphology. laser and infrared, 2003,33 (2): 94-97.[2] Bai Xiangzhi etc. the infrared small target based on novel cap transformation strengthens, the electronic imaging magazine, 2008,17 (3): 030501. ([2] Bai Xiangzhi, Zhou Fugen, XieYongchun.New class of top-hat transformation to enhance infrared small targets.Journal ofElectronic Imaging, 2008,17 (3): 030501.))

The related main computing of these class methods is morphology Top-hat conversion.The main thought of using the little target of Top-hat change detection is with Top-hat conversion estimated background, and obtains target from original image behind the subtracting background.But these class methods effect when varying background and signal to noise ratio (S/N ratio) are low is relatively poor.

(4) based on the method for light stream (referring to document: [1] that are rich etc. based on the optical flow computation of expansion constraint. IEEE's Flame Image Process transactions, 1996,5 (5): 720-739. ([1] Bimbo A.D., Nesi P., Jorge L.C.S..Optical flowcomputation using extended constrains.IEEE Trans.IP, 1996,5 (5): 720-739.) [2] Li Xiaochun, grandson accompanies diligent, whole nation fence, Deng. a kind of fast algorithm that complex scene Middle East target is cut apart. Tsing-Hua University's journal (natural science edition), 1997,37 (8): 59-64.)

These class methods can be determined the surface structure of number, movement velocity, target range and the target of moving target by optical flow analysis.But because the complicacy and the calculated amount of aperture problem, optical flow algorithm, when light stream being applied to realistic objective detection in real time and following the tracks of, its efficiency of algorithm, algorithm evaluation, light stream Robust Estimation and light stream Fast estimation aspect still need be furtherd investigate.

(5) based on the method for multistage hypothesis (MHT) (referring to document: auspicious moral etc. a kind of multiple target tracking algorithm. IEEE's avionics system transactions, 1979,24:843-854. (Reid D.B..An Algorithm for tracking multipletargets.IEEE Trans AES, 1979,24:843-854.))

These class methods are got up numerous possible target trajectorys with the structure organization of tree, by the tree that reaches every two field picture in the sequence is carried out test of hypothesis simultaneously, to remove at any time not by the tree of check, reduce operand and memory space.But in the complex background target detection and in following the tracks of, realization is based on the required internal memory of the track algorithm of multistage hypothesis and calculated amount is too big and a certain moment situation is described (comprising target numbers and motion state) and determined, has limited the application based on the tracking of multistage hypothesis.

(6) based on the interconnected method of data (referring to document: [1] Ba Shalong etc. based on the auxiliary interconnected tracking of multiframe data of classification. IEEE's avionics system transactions, 2005,41 (3): 868-877. ([1] Bar-Shalom Y., Kirubarajan T., Gokberk C..Tracking with classification-aided multiframe data association.IEEE Transactions on Aerospace and Electronic System, 2005,41 (3): 868-877.) [2] Pan Quan, leaf Xining, Zhang Hongcai. the broad sense Probabilistic Data Association Algorithm. electronic letters, vol, 2005,33 (3): 467-472.[3] Ba Shalong. multi-target multi-sensor following principle and method. the promise Wood, Massachusetts: ata gram publishing house, 1990. ([3] Bar-shalomYaakov.Multitarget-multisensor tracking principle and techniques.Norwood, MA:ArtechHouse, 1990.))

These class methods have obtained using preferably in multiple target tracking.But, when target numbers and effective measurement number increase, feasibility joint event number is exponential increase, cause the calculated amount of the interconnected algorithm of data to be exponential increase even the shot array phenomenon occurs, also increased the complicacy of structure feasibility joint event simultaneously, made these class methods often need to carry out respective design and just can obtain effect preferably according to concrete application.

(7) based on the corresponding method of motion (referring to document: [1] dimension etc. based on corresponding foundation of motion of expansion transient state scope. artificial intelligence, 2003,145:227-242. ([1] Veenman C.J., Reinders M.J.T., Backer E..Establishing motioncorrespondence using extended temporal scope.Artif.Intell., 2003,145:227-242.) [2] dimension etc. the constrained optimization problem of motion tracking. pattern-recognition, 2003,36:2049-2067. ([2] Veenman C.J., ReindersM.J.T., Backer E..Motion Tracking as a constrained optimization problem.Pattern Recognition, 2003,36:2049-2067.))

These class methods directly detect and follow the tracks of target according to the actual movement rule of all types of target.But these class methods are under the more situation of false-alarm targets, and operation time is longer, are difficult to satisfy the application of real-time system, make these class methods also need further research in practicability.

In addition, also have additive method to be suggested, as based on the method for curve evolution, based on the method for neuroid, based on the method for random field etc., but mostly these methods are the specific algorithm at objectives, and universality is relatively poor.

Summary of the invention

Infrared small object under the complex sky background detects must be in conjunction with the target following technology of target detection technique that utilizes single-frame images and multiple image with tracking, comes real goal under the detection and tracking complex background by the intensity profile characteristics and the movement characteristic in the multiple image of target in single-frame images.Simultaneously, in the target travel process, may there be the situation that is blocked, simply blocking under the situation and can remedy the track rejection situation of blocking generation by prediction.And under the situation of blocking for a long time, the prediction of target and block after just detect again and become the problem of comparison difficulty.

In order to address the above problem and remedy the deficiency of classic method, the present invention proposes a kind of infrared sequence image under complex sky background Weak target detects and tracking, by effectively utilizing characteristics and movement characteristic in the multiple image validity that improves complex background under target detection and tracking of target in single-frame images, and solve block the prediction of target under the situation for a long time and block after detection again, for the actual application problem in the real system provides effective solution.

In order to achieve this end, technical scheme of the present invention is as follows, at first rejects with adaptive background and iteration threshold is cut apart possible target in the detected image.Adaptive background is rejected the target under the detection of complex background effectively, also can detect once more by the target after blocking for a long time.Utilization is expanded to add up and is suppressed the false-alarm targets of complex background and noise generation with false-alarm targets then, reduces the difficulty of succeeding target tracking.Utilize the target following technology to confirm last real goal at last.In target following, the present invention has constructed new multi-frame prediction algorithm and has been used to solve the target prediction of blocking for a long time under the situation, and the target following when being blocked for a long time for target provides reference quantity.

The invention is characterized in and comprise following concrete steps:

(1) adaptive background is rejected: adopt the adaptive background rejecting algorithm of combining image Weak target feature and Top-hat conversion to carry out the adaptive background rejecting, suppress the influence that complex background and partial noise detect Weak target.It is detecting again of target that target after blocking for a long time detects again, can utilize adaptive background to reject algorithm equally and realize.

(2) iteration threshold is cut apart: a given initial value, image is divided into target and background two classes, and utilize two class gray-scale values in initial value and the image to carry out iteration, make it to converge to final threshold value;

(3) expansion adds up: according to the characteristics of motion structure expansion accumulation algorithm of target in continuous multiple frames, certain continuous a few width of cloth image is at first carried out morphology expand, add up then.Because the continuity of true infrared target motion, the real goal zone in the image after expansion adds up must be bigger zone, and other false-alarm targets zone is then less, thereby can distinguish possible false-alarm targets zone;

(4) false-alarm targets suppresses: use the possible false-alarm targets zone of expanding and adding up and produce based on the CB morphology operations and the logical and operation filtering of contour structure element, suppress false-alarm targets.

(5) track upgrades: distance is carried out relatively as the result of possible target area and step (7) trajectory predictions in the zone after false-alarm targets is suppressed, and substitutes the position, target area of predicting with possible position, target area, realizes that track upgrades;

(6) track judgement: the track that utilizes the continuous motion information judgement real goal of target.If in certain two field picture, certain track does not obtain upgrading, and this track then is described, and lose objects is once in this frame.Add up the number of times of track rejection in certain track, more than predefined number of times, then this track is the track that false-alarm targets forms, and should remove this track as if the number of times of losing; Otherwise keep this track, and the target of this track correspondence is a real goal;

(7) trajectory predictions: for the prediction of carrying out trajectory predictions and solving target under the situation of blocking for a long time, the position of calculating the next frame target with the average of the moving displacement of front multiframe, effectively utilize the movable information of multiframe in early stage, the possible position of each track reduces because the influence of prediction deviation to predicting the outcome that the fluctuation that several framing bits move causes in the prediction next frame image.

Wherein, the trajectory predictions in the described step (7) is the position of calculating the next frame target with the average of the moving displacement of front multiframe, reduces because the influence of prediction deviation to predicting the outcome that the fluctuation that several framing bits move causes, concrete grammar is:

If the number of image frames that is used to predict is nP, wherein the i frame (1≤i≤nP-1) displacement of the target of certain track correspondence is:

md _i＝||P _i-P _i1||，

Then the possible position of the target of this track correspondence of nP frame is:

$P_{\mathrm{nP}}=\underset{0\≤i\≤\mathrm{nP}-1}{\mathrm{mean}}\left(P_i\right)+\mathrm{round}(\mathrm{nP}/2)*\underset{1\≤i\≤\mathrm{nP}-1}{\mathrm{mean}}\left({\mathrm{md}}_i\right).$

Wherein, round is the round computing, and nP generally gets odd number.

Advantage of the present invention and effect are: utilize adaptive background reject to suppress the influence that complex background and noise detect Weak target, realize the detection of target and block for a long time after the detection again of target; Add up and false-alarm targets has suppressed the generation of part false-alarm targets by the expansion that effectively utilizes target behavior, reduce succeeding target and follow the tracks of the affirmation target difficulty; Utilize the continuous motion rule of real goal to construct effective trajectory predictions, renewal and decision algorithm, realize that effective detection of target reaches the prediction of blocking target location under the situation for a long time.Simultaneously, trajectory predictions is calculated the position of next frame target with the average of the moving displacement of front multiframe, reduced because the influence of prediction deviation to predicting the outcome that the fluctuation that several framing bits move causes, effectively the prediction of target under the situation is blocked in solution for a long time, simplifies the tracking of target.The present invention can effectively suppress complex sky background and noise detects and the influence of following the tracks of infrared small object, can be widely used in all kinds of Target Recognition and navigational system based on infrared image, has vast market prospect and using value.

Description of drawings

Fig. 1 is the theory diagram of infrared sequence image under complex sky background Weak target detection of the present invention with tracking.

Fig. 2 is the experimental result of the present invention to the little target image sequence of aircraft that includes 174 two field pictures, and that scheme that medium and small rectangular box marks is the result who utilizes the present invention to detect to trace into.Figure (a) (b), (c), (d), is respectively the 13rd, 21 in the sequence (e) and (f), 25,38,128 and the experimental result of the 138th frame.

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.Flow process of the present invention as shown in Figure 1, the concrete implementation detail of each several part is as follows:

1. adaptive background is rejected

It is as follows that adaptive morphology is learned background rejecting (AMCE) algorithm flow.

If the length and width of image f is respectively Wide and Height:

The first step: calculate R=CBC _B(f),

With

Being respectively morphology expands and erosion operation;

Second step: establish i=0;

The 3rd step: calculate JV (i)=maxw _i-minw _i, maxw wherein _iAnd minw _iBe that size is L _w* L _hAnd the rectangular window w of center on pixel i _iThe minimum and maximum value of middle pixel grey scale;

The 4th step: press the image background that following formula is estimated pixel i correspondence:

$\mathrm{BG}\left(i\right)=\left\{\begin{array}{c}f\left(i\right),f\left(i\right)-R\left(i\right)<\mathrm{JV}\left(i\right)\\ R\left(i\right),\mathrm{else}\end{array}\right.;$

The 5th step: if i＜Wide * Height, i=i+1 changed for the 3rd step over to;

Otherwise, changed for the 6th step over to;

The 6th step: reject the image background of estimating with following formula:

I＝f-BG。

Above-mentioned AMCE algorithm flow is rejected the image background BG that estimates from f, keep target when suppressing image background, thereby realizes the enhancing of target.W among the AMCE _iSize should can obtain by the priori or the experiment of target less than the size in territory, probable target area.

2. iteration threshold is cut apart

In the ideal case, adaptive background is rejected with original image and is deducted the target image that will be left only to comprise impact point and picture noise after the image background of estimation, but in actual conditions,, the background estimating algorithm also includes the residual image background in the target image because to the inaccuracy of background estimating, making.The residual image background is low in a large number gray-value pixel, makes target image become a width of cloth dark background image.The histogram of such image does not have tangible peak value, but the low-pixel value number of pixels is maximum.

Characteristics at low gray-scale value target image can be divided into image target and background two classes, and a given initial value, utilize two class gray-scale values in this initial value and the image to carry out iteration, make it to converge to final threshold value.This process is adaptive to the dark background image under the various situations.

The adaptive iteration threshold step is as follows:

(1) takes from the initial value of the average of the entire image after the adaptation background is rejected as Threshold

$\mathrm{Threshold}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}I\left(r_i\right),$

Wherein, N is a number of pixels, r _i=(x _i, y _i) be the pixel coordinate of image, I (r _i) be r _iThe pixel value of corresponding pixel points.

(2) with Threshold image is divided into two classes, pixel value is a target greater than the class of Threshold, and its average is designated as mean _Up, pixel value is a background less than the class of Threshold, its average is designated as mean _Down, the computing formula of new threshold value Threshold is as follows:

Threshold＝(1-1/f(x))×mean _down+(1/f(x))×mean _up，

Wherein,

x＝mean _up-mean _down，

${\mathrm{mean}}_{\mathrm{up}}=\frac{1}{N_{\mathrm{up}}}\underset{i=1}{\overset{N_{\mathrm{up}}}{\mathrm{\&Sigma;}}}I\left(r_i\right),$ N _UpBe the number of pixels of target,

${\mathrm{mean}}_{\mathrm{down}}=\frac{1}{N_{\mathrm{down}}}\underset{i=1}{\overset{N_{\mathrm{down}}}{\mathrm{\&Sigma;}}}I\left(r_i\right),$ N _DownNumber of pixels for background.

(3) end iteration if the value of Threshold no longer changes, otherwise return (2).

In above iterative process, in the computing formula of threshold value Threshold, the system of selection of f (x) is as follows:

1. the span of f (x) should (1, ∞)

This makes the value of Threshold can not surpass tonal range.

2. f (x) is the nondecreasing function of x

This will make f (x) that Threshold is had autoregulation, the reasons are as follows:

When Threshold is in high gray-scale value district, x=mean _Up-mean _DownBigger, (1-1/f (x)) is bigger at this moment, and (1/f (x)) is less, and then the calculating of Threshold value can increase mean in (2) step _DownProportion Threshold is reduced; On the contrary, when Threshold is in low gray-scale value district, x=mean _Up-mean _DownLess, (1-1/f (x)) is less at this moment, and (1/f (x)) is bigger, and the calculating of Threshold value can reduce mean in (2) step _DownProportion Threshold is increased; And when Threshold reaches reasonable threshold value, mean _UpAnd mean _DownValue can not change, f (x) can not change yet, Threshold will be stabilized in reasonable threshold value.Like this, along with the algorithm that carries out of iterative process can automatic adjusting Threshold value threshold value to the end.

Obviously, when $f\left(x\right)=\frac{1}{2}$ The time, this algorithm is just changed in quality and is iteration threshold algorithm commonly used.

As seen from the above analysis, the value difference of f in different iterative process (x), so mean in the adjustment process _UpAnd mean _DownAdjustment factor can regulate automatically according to Threshold, and these adjustment factors are different corresponding to different Threshold, can regard multiple dimensioned adjusting as, this just is different from iteration threshold algorithm commonly used at all, and this multiple dimensioned adjustment process will make Threshold converge on more reasonably threshold value.

3. the curvature of f (x) should be less

This makes that the regulating action of f (x) is unlikely excessive, to reduce iterations.

4. the convergence of iterative process

Suppose Threshold for desiring the reasonable threshold value of split image, the gradation of image average of the target and background that this moment is corresponding is respectively mean _UpAnd mean _Down, their difference is x=mean _Up-mean _Down, because this moment, Threshold was reasonable threshold value, mean _Up, mean _DownAnd x will no longer change.

If a certain corresponding initial threshold of step is Threshold1 in the iterative process, and Threshold1＜Threshold, the gradation of image average and the difference thereof of the target and background that this moment is corresponding are mean _Up1, mean _Down1 and x1, then know x1＜x this moment by the characteristics of dark background image, i.e. Δ x1=x-x1＞0 is known by (2) step and to be carried out will strengthening mean after an iteration _UpIt is Threshold2 that 1 proportion increases Threshold1, and the gradation of image average and the difference thereof of the target and background that this moment is corresponding are mean _Up2, mean _Down2 and x2, by the characteristics of dark background image, mean _Up2＞mean _Up1, mean _Down2＞mean _Down1, and mean _Up2-mean _Up1＞mean _Down2-mean _Down1, i.e. x2＞x1, so Δ x2=x-x2＜Δ x1, this just means the mean that carries out along with iterative process _Up-mean _DownDifference DELTA x between the x corresponding with reasonable threshold value will diminish gradually, and when Δ x was zero, corresponding threshold was reasonable threshold value, and this moment, iterative process was ended.

Equally, if a certain corresponding initial threshold of step is Threshold1 in the iterative process, and Threshold1＞Thershold, along with the carrying out of iterative process, Threshold1 also will converge to reasonable threshold value.

Experiment showed, that iterative process generally carries out just can terminating in reasonable threshold value below 10 times.

According to above condition, it is as follows to construct f (x):

f(x)＝log ₁₀(10+αx)，

The α value is a constant, and is different at different image α values.The average of image is more little, and the α value is more little, generally between (0,10).Utilize the result images after Threshold binaryzation adaptive background is rejected just can obtain possible target image IPTar.

3. expand and add up

The true characteristics of motion of infrared target in continuous multiple frames is continuous, and there is randomness the position that false-alarm targets occurs in continuous multiple frames.Therefore, expansion accumulation algorithm can be used for the possible false-alarm targets of mark.The algorithm detailed step is as follows:

If present frame is i frame two-value target image IPTar _i, number of image frames to be added up is nS.

The first step: get j=i-nS frame bianry image, all pixels are 0 among the initialization accumulation result image A ddImage;

Second step: with j two field picture structural element B ₁Expand B ₁Shape commonly used circular, B ₁Size often get 1-3;

The 3rd step: direct and AddImage addition with the result images in second step;

The 4th step: if j＜i, j=i+1 changed for second step over to,

Otherwise algorithm stops;

Because the continuity of true infrared target motion, the real goal zone in the image after expansion adds up must be bigger zone, and other false-alarm targets zone is then less.Therefore, this attribute can be used to suppress false-alarm targets.

4. false-alarm targets suppresses

The result who adds up by expansion as can be known, the size in real goal zone is greater than other zones.Therefore, available morphology operations removes the less zone of size and rebuilds current frame image with logic and operation, and then the false-alarm targets in the current frame image just can be suppressed preferably.The detailed process of algorithm is as follows:

The first step: use C _BCalculation process AddImage removes the less zone of size; C _B(f)=and min{f, CBC _B(f) };

Second step: with the original bianry image IPTar of present frame _iWith C _BAddImage image after the computing carries out the logical and operation, rebuilds present frame two-value target image.

Adopt C in the algorithm first step _BComputing mainly is because C _BComputing has the characteristics of protection image detail, so C _BComputing can not damage the real goal zone when removing false-alarm targets.C _BThe size of the structural element of using in the computing usually should be less than the real goal zone, and concrete big I is obtained by priori.Because false-alarm targets does not have fixed shape, so the shape of structural element can be used rhombus or rectangle commonly used.It is 3 rhombus that the present invention advises adopting radius.

Because the inhibition of false-alarm targets among the AddImage, algorithm can reconstruct the real goal zone in the present frame and removing the part false-alarm targets with operation in second step reaches the purpose that false-alarm targets suppresses.Result images after this step is designated as RT.

5. track upgrades

Because adaptive background is rejected the inhibition of image complex background and expansion adds up and false-alarm targets suppresses inhibition to false-alarm targets, the false-alarm targets quantity among the result images RT significantly reduces, but still has a small amount of false-alarm targets to obtain keeping.At this moment, can effectively utilize the target following technology of target travel information to be used to confirm real goal.Herein, the purpose that track upgrades is that the position with the possible target of the position of the possible target in the existing predicted current frame and detected present frame compares, to upgrade the target location of each track in present frame.

Detailed step of updating when there be NR zone in certain track Trace in present frame RT is as follows:

The first step: the predicted position P that gets present frame among the Trace _Cur, make i=0,0≤i≤NR-1;

Second step: get i regional P among the RT _i

The 3rd step: calculate P _iWith P _CuBetween apart from d _i

The 4th step: if i≤NR-1, i=i+1 changed for second step over to;

Otherwise, changed for the 5th step over to;

The 5th step: ask all d _iIn minimum value d _j

The 6th step: if d _j≤ mindT then uses P _iUpgrade the P among the Trace _Cu

Otherwise Trace does not upgrade;

MindT in the above-mentioned algorithm flow can obtain according to priori usually.Above-mentioned algorithm has guaranteed that track brings in constant renewal in according to the target detection situation of present frame, guarantees that the information of target continuous motion can be used effectively.

6. track judgement

The purpose of track judgement is thereby whether utilize the continuous motion information of target to adjudicate certain track be the purpose that the movement locus of real goal reaches recognition objective.Because the step of front has effectively suppressed false-alarm targets, the utilization of therefore simple movable information just can effectively be adjudicated the real goal track.

If in certain two field picture, track Trace does not obtain upgrading, and this track then is described, and lose objects is once in this frame.If the number of times of Trace lose objects surpasses certain pre-set threshold nTL, then Trace is not the track of real goal; Otherwise Trace is the track of real goal.NTL is a predetermined threshold, and picture noise is serious more, and nTL is big more.

This decision method is based on following brass tacks: if target is a real goal, then because the continuity of target travel, the track of this target correspondence must be continuous and not have than the situation that has track rejection in the multiframe; And false-alarm targets the situation of track rejection occurs owing to its randomness that occurs makes its corresponding track in than multiframe in the continuous multiple frames image.

Simultaneously, because effective inhibition of false-alarm targets, the false-alarm targets that can form longer track is few.Therefore, by finding the track of real goal after the track judgement.The position of the point of the present frame of this track correspondence is the position of real goal at present frame, thereby reaches the purpose of identification real goal.

7. trajectory predictions

One of purpose of the present invention is in order to solve the multiframe occlusion issue.Target can not be detected when target is blocked by object, for possible position that still can recognition objective must be predicted the position of target according to the characteristics of motion of target.Simultaneously, even target is not blocked, the prediction of target next frame position helped also the renewal for present frame target location in the track provides reference quantity in the target update.

When target trajectory is predicted, can handle, but simple linear prediction just can be approximately this basic assumption of straight line according to target travel rule in a short time with linear prediction.Therefore, under the situation that target is blocked for a long time, simple linear prediction just not too has been fit to trajectory predictions.The reason that produces this result is: the motion of next frame is predicted in simple linear prediction usually with the motion of former frames (being generally the 2-3 frame).And the motion of former frames may exist than great fluctuation process, and the deviation that produces under the situation of the less frame of prediction is also less.But this fluctuation may be introduced than mistake under the situation of prediction multiframe.Therefore, must effectively utilize the movable information of multiframe in early stage to predict the result of following multiframe.For this reason, the present invention proposes prediction algorithm based on the multiframe displacement of targets.

If the number of image frames that is used to predict is nP, wherein the i frame (1≤i≤nP-1) displacement of the target of certain track correspondence is:

md _i＝||P _i-P _i-1||，

Then the possible position of the target of this track correspondence of nP frame is:

$P_{\mathrm{nP}}=\underset{0\&le;i\&le;\mathrm{nP}-1}{\mathrm{mean}}\left(P_i\right)+\mathrm{round}(\mathrm{nP}/2)*\underset{1\&le;i\&le;\mathrm{nP}-1}{\mathrm{mean}}\left({\mathrm{md}}_i\right).$

Wherein, round is the round computing, and nP generally gets odd number.

As can be seen from the above equation, has reduced because the influence of prediction deviation to predicting the outcome that the fluctuation that several framing bits move causes the position that this prediction algorithm calculates the next frame target with the average of the moving displacement of front multiframe.

After the trajectory predictions, its result is admitted to track and upgrades the track that is used to upgrade real goal.

In order to show effect of the present invention, adopt the little target image sequence of aircraft that includes 174 two field pictures to verify validity of the present invention.Image background complexity in this image sequence, picture noise are serious, target is blocked for a long time by building body background.Experimental result as shown in Figure 2, image shown in the figure is the 13rd, 21 in the image sequence, the 25,38,128 and the 138th frame.As can be seen, although the image background complexity, adaptive background estimates to have suppressed preferably the influence of complex background to effect of the present invention from the result.Simultaneously, the serious noise of image may produce a large amount of false-alarm points, and these false-alarm points also are suppressed, and make the present invention can obtain effect preferably.What is more important, the target in the image sequence was once blocked for some time for a long time, but the present invention has still predicted the possible position of target preferably, has embodied validity of the present invention.

Claims (2)

1, a kind of infrared sequence image under complex sky background Weak target detects and tracking, and it is characterized in that: this method comprises following concrete steps:

(1) adopts the adaptive background rejecting algorithm of combining image Weak target feature and Top-hat conversion to carry out the adaptive background rejecting, suppress the influence that complex background and partial noise detect Weak target;

(2) the iteration threshold algorithm that utilize to be fit to the dark background image is cut apart the territory, probable target area in the dark background image after background is rejected;

(3) according to the characteristics of motion structure expansion accumulation algorithm of target in continuous multiple frames, distinguish possible false-alarm targets zone;

(4) use C based on the contour structure element _BComputing and logical and operation suppress false-alarm targets;

(5) result of the zone after the false-alarm targets inhibition as possible target area and step (7) trajectory predictions compared, realize that track upgrades;

(6) carry out the track judgement; If in certain two field picture, certain track does not obtain upgrading, and this track then is described, and lose objects is once in this frame; Add up the number of times of track rejection in certain track, if the number of times of track rejection is more than predefined number of times, then this track is the track that false-alarm targets forms, and should remove this track; Otherwise keep this track, and the target of this track correspondence is a real goal;

(7) carry out trajectory predictions; Effectively utilize the movable information of multiframe in early stage, adopt possible position based on target in each track in the prediction algorithm prediction next frame image of multiframe displacement of targets.

2, the trajectory predictions in the step according to claim 1 (7), it is characterized in that: the position of calculating the next frame target with the average of the moving displacement of front multiframe, the influence of prediction deviation that the fluctuation that minimizing moves owing to several framing bits causes to predicting the outcome, its concrete grammar is:

If the number of image frames that is used to predict is nP, wherein the i frame (1≤i≤nP-1) displacement of the target of certain track correspondence is:

md _i＝||P _i-P _i-1||，

Then the possible position of the target of this track correspondence of nP frame is:

$P_{\mathrm{nP}}=\underset{0\&le;i\&le;\mathrm{nP}-1}{\mathrm{mean}}\left(P_i\right)+\mathrm{round}(\mathrm{nP}/2)*\underset{1\&le;i\&le;\mathrm{nP}-1}{\mathrm{mean}}\left({\mathrm{md}}_i\right);$

Wherein, round is the round computing, and nP generally gets odd number.

Images