CN101676744B - Method for tracking small target with high precision under complex background and low signal-to-noise ratio - Google Patents

Abstract

The invention provides a method for tracking a small target with high precision under a complex background and a low signal-to-noise ratio (SNR). The method comprises pre-treating an image under the complex background and low SNR, portioning and extracting the target based on a target adaptive threshold of the binomial distribution judgment rule, adopting a curve fitting algorithm based on Kalman filter thinking improvement to carry out motion prediction on the target, using data fusion of infrared and visible light transducers to improve detection probability of the target and reduce false-alarm probability, and when the shape of the target alters, using a shape of which edge characteristics are normalized to identify and seek characteristic invariables so as to achieve precise tracking of the target.

Description

A kind of small target with high precision under complex background and low signal-to-noise ratio tracking

(1) technical field:

The present invention relates to a kind of small target with high precision under complex background and low signal-to-noise ratio tracking; Especially refer to the automatic detection of target and recognition technology, Flame Image Process and data fusion method under a kind of pattern-recognition and Based Intelligent Control, the complex background; Through fusion to visible light and infrared information; Realized detection and tracking, belonged to technical field of information processing Weak target under the complex background.

(2) background technology:

In modern times in the war; A large amount of uses of various precision guided weapons make the air defense operation more and more difficult that becomes; The application of cruise missile, stealth aircraft, armed helicopter, antiradiation missile, scout-attack unmanned plane makes antiaircraft situation become more and more severeer, how attacks target and implement effectively to it that attack is the Modern Aerial Defense urgent problem preserving to detect under the own situation.And at present the air defense of China mainly resists conventional airplane and sets up, so be more weak to the antagonism of above-mentioned target, its main cause is that the detection to above-mentioned target has difficulties.Particularly the threat of antiradiation missile makes the application of air defense radar be very restricted.It is reported that before the war in Iraq that took place in 2003, the air defense of Yi Fang more than 80% is intact; But after war was fired, the antiaircraft weapon of Yi Fang almost had no to contribute, and its main cause is made electromagnetism power exactly on top of in U.S. army's hand; The air defense radar of Iraqi military is as long as a start; Just suffer the attack of U.S. army's antiradiation missile in a few minutes, the will to fight that U.S. army has more destroyed Yi Jun when destroying the Yi Jun air defense is though so the most of antiaircraft weapon of Yi Jun is excellent; But but dare not use, become useless a pile scrap iron.The main cause that the Yi Jun air defense operation fails effectively to play a role is to lack the detection means that effectively is directed against U.S. army's air-supported threat.

Photodetection system resists above-mentioned target threat and has unique advantage, and Photodetection system and radar are used, and have very strong complementarity.There is the not available advantage of following radar in photoelectricity track production system:

1) passive type working method, non-radiating electromagnetic wave, good concealment;

2) be operated in the light wave scope, do not receive electronic interferences;

3) target is difficult for stealthy;

4) during low-angle tracking, do not receive the influence of ground clutter, the no lower space of silence;

5) tracking accuracy and distance accuracy are high;

6) the target image visual and clear is easy to Target Recognition.

The key issue that the photodetection tracker need solve is that development has intelligentized video object recognition and tracking device and high-precision servo tracking turntable.The design of Photodetection system should be as much as possible found and recognition objective when remote because aerial target is very small and weak usually when remote, so under the complex background Weak target to detect identification and follow the tracks of be to need the key technical problem that solves.

(3) summary of the invention:

The object of the present invention is to provide a kind of small target with high precision under complex background and low signal-to-noise ratio tracking; To be implemented under the utmost point low signal-to-noise ratio; In time detect, discern Weak target; After detecting target, stable tracking target, not lose objects under complex background (bait, target shake, imaging noise, multiple goal, intersection etc.) situation.

The present invention is a kind of small target with high precision under complex background and low signal-to-noise ratio tracking; The multimode data fusion can improve the detection probability and the accurate tracking precision of target under its complex background that adopts; Take into full account the requirement that systematization and through engineering approaches are used simultaneously, in design, consider multiple general requirment, multiple information interface; Integrated powerful soft, hardware resource need to change software and can realize the processing to different target such as marine, aerial not changing following of hardware case.

Technical scheme of the present invention is:

Small target with high precision under complex background and low signal-to-noise ratio tracking of the present invention is applied on the multimode multi-target accurate tracking apparatus of independent development, with the performance index of verification system.This multimode multi-target accurate tracking apparatus is made up of following three parts: digital servo platform, integrated information processing platform, compression and transmission equipment; Small target with high precision under complex background and low signal-to-noise ratio tracking of the present invention mainly is achieved in integrated information processing platform.In this device:

1) digital servo platform:

This digital servo platform is by CCD (Charge Coupled Device; Being the CCD imageing sensor) video camera, infrared sensor, high accuracy number servo turntable, handle and monitor form, and also can select two ccd video cameras or two infrared sensors as required for use.Ccd video camera in this digital servo platform can be simulating signal input or digital signal input, and its resolution of the infrared sensor of employing is 768 * 576.

This digital servo platform is the support platform of image acquiring device, and ccd video camera wherein and infrared sensor are installed in high accuracy number servo turntable two ends respectively, can move with the high accuracy number servo turntable.This high accuracy number servo turntable can rotate according to the control command that receives simultaneously, and target is carried out accurate tracking, makes target remain on the center, visual field of image acquiring device.

Ccd video camera is used in combination with infrared sensor, obtains visible light and infrared target image information simultaneously, the target signature in comprehensive two kinds of information, thereby the detection probability and the accurate tracking precision of raising target.

2) integrated information processing platform:

This integrated information processing platform is made up of information interface, high speed digital signal processor, servo control processor.Its high speed digital signal processor adopts the signal processing system based on DSP (digital signal processor).High speed digital signal processor receives the image information of importing into from ccd video camera and infrared sensor, and completion is to the realization of the target's feature-extraction in visible light and the infrared image under the complex background low signal-to-noise ratio condition, characteristic matching, target travel prediction and estimation, precise tracking method.Servo control processor is according to the result of target prediction and tracking; Confirm the direction of motion of high accuracy number servo turntable; And send control command to the high accuracy number servo turntable, the high accuracy number servo turntable is followed the tracks of target with the result who follows the tracks of according to prediction.

Here adopt two separate signal processors; Be high speed digital signal processor and servo control processor; Be high speed digital signal processor and servo control processor, respectively the control information of image information and servo platform handled, in image pre-service, target recognition and tracking key algorithm; To the characteristics of Weak target under the complex background, take multiple algorithm to improve and realize recognition and tracking Weak target with innovation.

3) compression and transmission equipment:

Compression and transmission equipment make this multimode multi-target accurate tracking apparatus have " people is in the loop " function; Pass all information and the image of discerning detected target automatically back command centre, and the instruction of accepting command centre is adjusted to improve the precision of automatic identification to the target of following the tracks of.This compression and transmission equipment are made up of video compress processor, GPRS transport module.The image input of video compress processor can be digital video or analog video, can require to select different interface protocols based on different output, adopts the video compression algorithm of MEPG-4, and rear end GPRS transport module adopts and transmits based on the GPRS wireless channel.

This device has been broken through the single target detection identification tupe of Target Recognition tracker, can also carry out data interaction, image transmission with other detection system networking, and have servo networking control, people in the circuit controls function.

Details are as follows for relation between this each ingredient of multimode multi-target accurate tracking apparatus:

The annexation of this device comprises ccd video camera, infrared sensor, high accuracy number servo turntable, handle and monitor five parts for this digital servo platform.Wherein ccd video camera and infrared sensor are installed in the two ends on high accuracy number servo turntable top respectively; Both link to each other with information interface through cable and carry out image data transmission, and handle and monitor are placed on the both sides of high accuracy number servo turntable bottom respectively.This integrated information processing platform comprises information interface, high speed digital signal processor and servo control processor three parts, and three parts all are integrated in information processing board and place control box, are placed on high accuracy number servo turntable one side.Wherein the handle in the digital servo platform links to each other with information interface and carries out the transmission of control signal; Monitor in the digital servo platform links to each other with information interface and is used to show the image data information of obtaining; High speed digital signal processor links to each other with information interface; Be used to obtain ccd video camera and infrared sensor image transmitted data; Servo control processor links to each other with information interface, is used to obtain the target detection identifying information of high speed digital signal processor and the positional information of high accuracy number servo turntable feedback, and to high accuracy number servo turntable transmission control command.This compression and transmission equipment; Comprise video compress processor and GPRS transport module two parts; Both are integrated on the information processing board respectively; Video compress processor rear end links to each other with the GPRS transport module, and video compress processor front end links to each other with the high speed digital signal processor in the integrated information processing platform.With regard to the multimode multi-target accurate tracking apparatus generally speaking; The digital servo platform is in the front end of multimode multi-target accurate tracking apparatus; Integrated information processing platform is in the middle-end of multimode multi-target accurate tracking apparatus, and compression and transmission equipment are in the rear end of multimode multi-target accurate tracking apparatus.

A kind of small target with high precision under complex background and low signal-to-noise ratio tracking of the present invention; Be in high speed digital signal processor, to accomplish; Its workflow in whole multimode multi-target accurate tracking apparatus does; At first obtain the visible light and the infrared image of target through ccd video camera and infrared sensor; Send the picture signal under the complex background low signal-to-noise ratio to high speed digital signal processor through information interface then, treated device is accomplished Automatic identification of targets and high precision tracking after image is carried out pre-service, detection; Send the target information of following the tracks of to servo control processor simultaneously, produce control command by servo control processor and give the high accuracy number servo turntable; When picture signal is shown through monitor, send the image of the original image information and the tracking target information that superposeed to the video compress processor and carry out video compress; Carry out wireless transmission through the GPRS transport module then, make command centre in the monitor of control center, observe the target following situation through decoding processor.A kind of small target with high precision under complex background and low signal-to-noise ratio tracking of the present invention the steps include:

(1), the image pre-service under complex background, the low signal-to-noise ratio condition: adopt Weak target algorithm for image enhancement, come image is carried out denoising through on each yardstick of wavelet transformation, choosing different threshold values respectively based on improved Stationary Wavelet Transform and non-linear enhancing operator;

(2), cut apart based on the objective self-adapting thresholding of binomial distribution judgment criterion: the relation between single frames detection probability, single frames false-alarm probability and total detection probability and the total false-alarm probability is set up the model based on theory of probability binomial distribution criterion, has solved definite problem of be correlated with during sequence image detects frame number and thresholding;

(3), infrared and visible data is carried out multimode and is merged: after visible light and infrared image are mated; Adopt the interconnected wave filter of multisensor probability data that the target signature in two kinds of images is mapped to the target signature in the image; Obtain fused data; Thereby improve the degree of confidence of target detection identification, and reject false target;

(4), target travel prediction and estimation: adopt and carry out motion prediction, solve the motion prediction problem under random shake, target overlapping, the memory tracking situation based on the improved curve fitting algorithm of Kalman wave filter thought;

Target's feature-extraction when (5) target shape changes: when target shape changes, adopt the method utilize the normalized shape recognition of edge feature to seek the characteristic invariant of drawing from the electric field angle to reach to the target accurate tracking.

Wherein: based on the improved curve fitting prediction algorithm of Kalman wave filter thought, be to adopt the window index function to come the locus intercepting point data, and control the effect of tracing point in the said step (4) to match through weight coefficient is set; And, choose the length of locus intercepting point through the size of power mobility index through mobility index is set; When being in the state of multiple goal intersection, candidate target of a plurality of tracking chains competitions adopts candidate target is given up, and the chain of competition is remembered tracking separately, the target following problem the when method till not having competition is handled the multiple goal intersection.

Wherein, the target's feature-extraction method when target shape changes in the said step (5) is the implication of giving electric charge point pixel, finds out a characteristic quantity that does not change with shape from the electric field angle and comes irregular figure is carried out effective recognition.

Below, each step is elaborated:

1) the image pre-service under complex background, the low signal-to-noise ratio condition

Weak target under the complex background is detected the necessary effective pre-processing method of selecting, and this is to having very important meaning in the follow-up target detection identifying.In our research in the past, a lot of preprocess methods were all carried out emulation and practical applications.Through a large amount of experiments and analysis, the Weak target algorithm for image enhancement based on improved Stationary Wavelet Transform (DSWT) and non-linear enhancing operator is adopted in the image pre-service of native system.

Wavelet transformation has perfect reconstruction ability; Have localization property (retractility) simultaneously in time domain and frequency domain, can focus on any details of object; Multiple dimensioned, multiresolution characteristic; Directional selectivity is coincide with human visual system's directivity.The multiple dimensioned characteristic of wavelet analysis makes it be suitable under the low environment of signal to noise ratio (S/N ratio), carrying out target detection.Its expansion performance can make the parts of images characteristic under certain yardstick, suppressed effectively, and some interested target (like little target) can be highlighted.Wavelet analysis not only can be used in the image pre-service, also can be used in image segmentation and the target travel estimation.

From a large amount of domestic and foreign literature analyses, under complex background, this field of Weak target recognition and tracking, traditional image pre-service based on wavelet transformation is all operated basically as follows:

(1) selects suitable wavelet basis, and image is carried out N layer wavelet decomposition;

(2) threshold value of high frequency coefficient is selected.For ground floor each layer, select a threshold value to handle to the N layer.

(3), calculate the wavelet reconstruction of image according to the high frequency coefficient of the low frequency coefficient of N layer and process modification from ground floor to the N layer.

Although traditional pre-service based on wavelet transformation can obtain good result, when high frequency coefficient was handled, major part had adopted linear unified threshold value, and details such as edge of image have been suffered weakening in various degree.Native system is being summed up traditionally based on wavelet image on the pretreated basis, utilizes based on Stationary Wavelet Transform (DSWT) and the non-linear enhancing operator image to Weak target to strengthen.On the basis of carrying out DSWT, the high-frequency sub-band that obtains has relatively poor resolution, these high-frequency sub-band is carried out the nonlinear operator computing improve and strengthen high-frequency sub-band, thereby reached the effect that filtering strengthens.Experimental result shows that this algorithm can effectively be eliminated 1/f noise, additive white Gaussian noise and multiplicative noise, the signal to noise ratio (S/N ratio) of raising image.This algorithm mainly comprises following three parts:

(1) suppresses noise

(2) selection of threshold

(3) non-linear enhancing operator

Be elaborated with regard to this three part below

(1) suppresses noise

Adopt traditional " global threshold " that image is come denoising, effect is undesirable.I.M.Johnston has proved that the wavelet transformation of correlation noise all is stably on all yardsticks, we can come image is carried out denoising with different threshold values respectively on each yardstick.

The model of supposing discrete picture is following:

g[i，j]＝f[i，j]+ε[i，j](4.1)

Top formula can be write as the form of matrix:

g＝f+ε (4.2)

Wherein, g={g [i, j] } _{I, j}It is the signal that observes.F={f [i, j] } _{I, j}Expression does not have the original signal of noise pollution, ε={ ε [i, j] } _{I, j}, i=1 ..., M; J=1 ..., N is a stationary signal.

DSWT is carried out in (4.2):

X＝Sf (4.3)

V＝Sε (4.4)

Y＝Sg (4.5)

Y＝X+V (4.6)

Wherein S representes steadily wavelet transformation operator of two dimension, and " soft-threshold " function of in image, quoting the Donoho proposition carries out denoising to image:

Y _δ＝T _δοY(4.7)

T _δ＝diag{t[m，m]}

$t[m,m]=\left\{\begin{array}{cc}0,& \left|Y\right[i,j\left]\right|<\mathrm{\&delta;}\\ 1-\frac{\mathrm{\&delta;}}{\left|Y\right[i,j\left]\right|}& \left|Y\right[i,j\left]\right|\&GreaterEqual;\mathrm{\&delta;}\end{array}\right.$

Wherein, i=1 ..., M, j=1 ..., N, m=1 ..., MN

According to formula (4.5) and (4.7), the contravariant of input signal is changed to:

g _δ＝S ^-1οY _δ(4.8)

T wherein _δRelevant with threshold value δ with signal g.

(2) selection of threshold

Suppose that (x y) can use its neighborhood territory pixel linear expression to original signal f.If make is g [k; L] linear expression; With the neighborhood average image is carried out smoothly can removing a part of noise.

after level and smooth can be used for calculating the threshold value of removing noise.G [i; J] [the i of expression among the g; J] element, it is substituted by :

$\tilde{g}=Z{\left(g\right[\mathrm{1,1}],...,g[i,j],...g[M,N\left]\right)}^T---\left(4.9\right)$

We think

than g [i, the j] threshold value that can better be optimized.

If threshold value δ is too little, mainly show as noise in

; If threshold value δ is too big, can filter a lot of useful signals.

All pixels are implemented same operation, and best threshold value can obtain through following computing:

$\mathrm{OCV}\left(\mathrm{\&delta;}\right)=\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(g\right[i,j]-{\tilde{g}}_{\mathrm{\&delta;}}[i,j\left]\right)}^2---\left(4.10\right)$

Form have a variety of, here, the order ${\tilde{g}}_{\mathrm{\&delta;}}[i,j]=\tilde{g}[i,j],$ Then:

$g[i,j]-{\tilde{g}}_{\mathrm{\&delta;}}[i,j]=\frac{g[i,j]-g_{\mathrm{\&delta;}}[i,j]}{1-\tilde{z}[i,j]}---\left(4.11\right)$

Wherein:

$\tilde{z}[i,j]=\frac{g_{\mathrm{\&delta;}}[i,j]-{\tilde{g}}_{\mathrm{\&delta;}}[i,j]}{g[i,j]-{\tilde{g}}_{\mathrm{\&delta;}}[i,j]}\&ap;z^{\&prime;}[m,n]=\frac{{\&PartialD;g}_{\mathrm{\&delta;}}[i,j]}{\&PartialD;g_{\mathrm{\&delta;}}[k,l]}$

Wherein, i, k=1 ..., M, j, l=1 ..., N, m, n=1 ..., MN

Yet in (4.11), z ' [m, m] is 1 or 0, and is unavailable in actual computation.Therefore providing following formula substitutes (4.10):

$\mathrm{SGCV}\left(\mathrm{\&delta;}\right)=\frac{\frac{1}{\mathrm{MN}}{\left|\right|Y-Y_{\mathrm{\&delta;}}\left|\right|}^2}{{\left[\frac{\mathrm{trace}(I-Z_{\mathrm{\&delta;}}^{\&prime;})}{\mathrm{MN}}\right]}^2}---\left(4.12\right)$

Wherein, the mark of trace representing matrix, || || the expression euclideam norm.I is the unit matrix of M * N.Make δ ^*=argminMSE (δ), $\widetilde{\mathrm{\&delta;}}=\mathrm{Arg}\mathrm{Min}\mathrm{SGCV}\left(\mathrm{\&delta;}\right),$ M.Jansen has proved

It is progressive optimal threshold.

(3) non-linear enhancing operator

1994, A.Laine once provided the non-linear enhancing operator based on DSWT, strengthened the local contrast of image.For convenience's sake, on each yardstick, define the transforming function transformation function of each high-frequency sub-band images respectively:

g[i，j]＝MAG{f[i，j]}(4.13)

Wherein, g [i, j] is the subband that strengthens, and f [i, j] is original sub-band images, and MAG is non-linear enhancing operator.

Make f _s ^r[i, j] is r high-frequency sub-band coefficient on the s decomposition scale, s=1 wherein, and 2 ..., L; R=1,2,3.max f _s ^rBe all pixel f _s ^rMaximum gradation value in [i, j].f _s ^r[i, j] can be from [max f _s ^r, max f _s ^r] be mapped to [1,1].

Therefore, the scope of a, b and c can be set respectively.The contrast Enhancement Method can be described below:

$g_s^r[i,j]=\left\{\begin{array}{cc}{f}_s^r[i,j],& \left|f_s^r\right[i,j\left]\right|<T_s^r\\ a\&CenterDot;\max f_s^r\left\{\mathrm{sigm}\right[c\left(y_s^r\right[i,j]-b)]-\mathrm{sigm}[-c\left(y_s^r\right[i,j]+b)\left]\right\},& \left|f_s^r\right[i,j\left]\right|\&GreaterEqual;T_s^r\end{array}\right.---\left(4.14\right)$

Wherein,

$y_s^r[i,j]=f_s^r[i,j]/\max f_s^r$

Carry out inverse wavelet transform at last, promptly obtain pretreated image.

Native system is on the characteristic and background model basis of the infrared small object of fully analyzing and researching, and the algorithm of employing Stationary Wavelet Transform (DSWT) and non-linear enhancing operator carries out the pre-service of figure image intensifying to infrared small object.Experimental result show this algorithm not only to the background of Weak target suppress, target strengthens that effect is preferably arranged, and noise, the background of general objective suppressed to have equally effect preferably.

2) cut apart based on the objective self-adapting thresholding of binomial distribution judgment criterion

Based on the characteristics of Weak target (low SNR, complex background), rely on single frames not reach to its detection, must rely on image sequence.In the present Weak target of doing detected, the detection identification probability that provides was generally more than or equal to 98%, and false-alarm probability is smaller or equal to 10 ^-6,, choose noise gate with them exactly if these two parameters are embodied in the detection.On the one hand, if according to false alarm rate smaller or equal to 10 ^-6Choose noise gate, then thresholding is very high, and a lot of targets will be lost; On the other hand, if choose noise gate according to detecting identification probability more than or equal to 98%, then choosing of thresholding should be very low, to guarantee taking out all impact points, can cause very big false alarm rate like this.Weak target under this low signal-to-noise ratio condition detects, and can't realize by single frames, must utilize the relevant information of multiple image, continuity, the consistance of the motion feature of target and movement locus is combined consider.Calculate the detection probability and the false-alarm probability of single frames from total detection probability and false-alarm probability, thereby choose the key issue of rational noise gate for solving.

With the theory of probability is the basis; We are in entire image detects; According to the relation of single frames detection probability, single frames false-alarm probability and total detection probability and total false-alarm probability, solve definite problem of be correlated with in the sequence image detection frame number and thresholding, for low false alarm rate, high detection rate provide theoretical assurance.

Regard the target detection in every two field picture as independently repeated experiments; According to the principle in the theory of probability; Detection probability should be obeyed binomial distribution, sets up mathematical model in view of the above, supposes that picture noise is the white noise of Gaussian distribution after the pre-service; According to statistical theory, when image carries out the single frames Threshold detection [derivation is referring to publishing an article]:

$p_d={\&Integral;}_v^{+\&infin;}p\left(x\right)\mathrm{dx}={\&Integral;}_v^{+\&infin;}\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\exp (-\frac{{(x-\mathrm{\&mu;}-\mathrm{\&sigma;S})}^2}{2{\mathrm{\&sigma;}}^2})\mathrm{dx}=1-\mathrm{\&Phi;}\left(\frac{v-\mathrm{\&mu;}-\mathrm{\&sigma;S}}{\mathrm{\&sigma;}}\right)$

$=\mathrm{\&Phi;}\left(\frac{\mathrm{\&mu;}+\mathrm{\&sigma;S}-v}{\mathrm{\&sigma;}}\right)---\left(4.15\right)$

$p_f={\&Integral;}_v^{+\&infin;}\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\exp (-\frac{{(x-\mathrm{\&mu;})}^2}{2{\mathrm{\&sigma;}}^2})\mathrm{dx}=1-\mathrm{\&Phi;}\left(\frac{v-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)=\mathrm{\&Phi;}\left(\frac{\mathrm{\&mu;}-v}{\mathrm{\&sigma;}}\right)$

$\frac{\mathrm{\&mu;}+\mathrm{\&sigma;S}-v}{\mathrm{\&sigma;}}={\mathrm{\&Phi;}}^{-1}\left(p_d\right)---\left(4.16\right)$

ν=μ+σ S-σ Φ wherein ^-1(p _d) or ν=μ-σ Φ ^-1(p _f)

Can release the relation of single frames detection probability and single frames false-alarm probability:

Φ ^-1(P _f)-Φ ^-1(P _d)＝S (4.17)

In the formula: ν---detection threshold, σ ²---the noise mean square deviation

μ---be the noise average after K frame background offsets

S---signal to noise ratio snr is defined as the ratio of amplitude with the mean square deviation of noise of signal

The relation of total detection probability and single frames detection probability:

$P_D(i\&GreaterEqual;k)=1-P_D=1-\underset{k=0}{\overset{i-1}{\mathrm{\&Sigma;}}}{C}_n^k{p}_d^k{(1-p_d)}^{n-k}---\left(4.18\right)$

In above-mentioned experiment, suppose that the single frames detection probability is P _d=0.90, require total detection probability P _D=0.98.When in n width of cloth image, when target has occurred k time, can reach requirement so.For example, when each collection 16 two field pictures made a decision, if target occurs 11 times, then total detection probability can reach requirement.

Therefore can be by single frames detection probability, single frames false-alarm probability and always detection probability and the always relation of false-alarm probability, definite problem of relevant frame number and thresholding during the solution sequence image detects.

3) infrared and visible data is carried out the multimode fusion

Detect with the visible light sensor fusion goal based on infrared, can improve detection probability, the reduction false-alarm probability of target.

Need carry out images match to visible light and infrared image fusion, at first the target of different sensors carried out feature extraction, can draw the transformation relation between visible light and the infrared image through the object matching result then, like translation, rotation, convergent-divergent etc.The concrete method for registering that uses is based on the reference mark coupling of least square, phase correlation method, template matching method etc.

After image registration, according to nearest neighbouring rule, the coordinate of same target in images after registration should be coincidence or very approaching, can be mapped the target signature in two kinds of images through proximity or similarity measure.

We are used for the fusion of visible light and infrared image with the interconnected wave filter of multisensor probability data, reduce false alarm rate.The interconnected basic thought of probability data is: so long as effectively detect information, all possibly come from target, it is different that just each information comes from the probability of target.This method has utilized all information in the track window obtaining possible posterior information, and provides each probability weight coefficient and weighted sum thereof according to a large amount of correlation computations, upgrades dbjective state with it then.This research adopts the method to carry out the target signature fusion.

After the Target Fusion processing, enriched clarification of objective information, improved target detection probability, reduced false alarm rate.

The treatment step that fusion infrared and visible data is adopted is following:

A) visible light and infrared image are carried out target's feature-extraction respectively;

B) characteristic of extracting is carried out the characteristic matching of different sensors, confirm the transformation relation between visible light and the infrared image;

To mating the target signature in the image of back, the interconnected filter process of probability of use obtains fused data, thereby improves the degree of confidence of target detection identification, and rejects false target.

4) target travel prediction and estimation

Multiple goal in the target following process, intersection, shake, memory tracking etc. all relate to the movement locus of the estimation forecasting problem Weak target of target; There is shake between frame and the frame; Do not have the direction consistance; And the area of other target such as bird is bigger slightly than target, and its movement locus has more directivity than Weak target.This is the shake that has caused image owing to reasons such as imaging system, spatial light interference, air vibrations; In the time of shake the geometric center of target is changed between neighbor; If data are without processing; Target location between the successive frame is because the existence of shake, and changes not according to the consistance of direction, and this just brings error to the target accurate tracking that retrains (like the filter's prediction of routine) with directivity.

Influence how to eliminate the jitter phenomenon generation is the major issue that actual tracking will solve.In project, solve these class methods and adopted two kinds of methods:

● based on the motion prediction of the improved curve fitting of Kalman wave filter thought.Solve the motion prediction problem under the situation such as random shake, target overlapping, memory tracking.

● after being identified as target, will slightly following the tracks of and transfer accurate tracking to, and convert position of form center into centroid position, the jitter phenomenon when reducing to follow the tracks of improves tracking accuracy.

Because the Kalman wave filter can not reflect the motion change of target immediately, so when target maneuver property was strong, the error that Kalman follows the tracks of can be bigger.We have defined the maneuverability parameter from the Kalman kinetics equation, simultaneously the match mode are improved, and strengthen the performance of its tracking prediction.

Shown in Figure 7 is target detection, recognition and tracking process flow diagram based on trajectory predictions.The target travel prediction is the pre-service of image process with the process of estimating, carries out clarification of objective and extracts, and obtains potential target; Judgment criterion through binomial distribution judges whether it is fresh target, sets up equation of locus through the kalman theory, if equation of locus is set up; Then need upgrade equation of locus; Target of prediction obtains predicted position at the track of next frame simultaneously, and with match correction Kalman tracking filter correction predicted position.

Below be several Points Concerning and improvement based on the theoretical improved curve fitting target of prediction running orbit research of Kalman:

(1) improved least-squares line approximating method

Criterion function in the classical least square fitting is the quadratic sum of the distance of each data point along ordinate y to matched curve.Tentation data point (x _i, y _i), y=f (x) wherein.Do the curve that least square fitting obtains and be based on following criterion:

$\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{{\mathrm{\&delta;}}^2}_i=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{\&omega;}\left(x_i\right){[s^{*}\left(x\right)-f\left(x_i\right)]}^2=\underset{s\left(x\right)\&Element;\mathrm{\&Phi;}}{\mathrm{Min}}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{\&omega;}\left(x_i\right){[s\left(x\right)-f\left(x_i\right)]}^2,$ But quadratic sum with distance $\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_i^2=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{\&omega;}\left(x_i\right){\left(\frac{|y-a_{i}x-b_i|}{\sqrt{1+a_i^2}}\right)}^2=\mathrm{Min}$ It is more suitable to do criterion function, and the latter is the least moment of inertia in the mechanics.

(2) the locus intercepting point is done match

In application, only consider to close on several or tens tracing points, if weighting coefficient ω is (x _i) ≡ 1, be equivalent to obtain recent tracing point with the rectangular window locus intercepting.Rectangular window can not reflect the variation of importance, and can not smoothly block; Exponential function can reflect data point importance over time, and exponential function is smooth, so adopt the window index function to come data intercept.

Consider the continuity and the consistance of target travel; The size of every frame instantaneous velocity, direction and mass motion trend are compared and big saltus step should do not arranged; When this change is big, wherein should comprise bigger error, suppress this error should for the less weights of this data points.

According to above criterion, through experiment and theoretical derivation, the value of definition weight coefficient:

$\mathrm{\&omega;}\left(x_i\right)=q^n\&CenterDot;p^{d_i}$

Wherein | q|＜1, | p|＜1, d _iBe the distance of data point to fitting a straight line Ax+By+C=0, $d_i=\frac{|{\mathrm{Ax}}_i+{\mathrm{By}}_i+C|}{\sqrt{A^2+{\mathrm{<figure-callout\; id="2"\; label="B"\; filenames="H2009101807801E00011.png,H2009101807801E00022.png"\; state="\{\{state\}\}">B</figure-callout>}}^2}};$ N=N-i, N are current frame numbers, and N-M≤i＜N, M are the numbers of fitting data point.The value of q and p can obtain through experiment, generally gets between 0.7～0.8.

The tracing point number M can be selected to fix, and controls the effect of tracing point to match through adjustment weight coefficient size, for example selects little attenuation coefficient, is equivalent to the short track of intercepting, selects bigger attenuation coefficient, is equivalent to the long track of intercepting.

(3) mobility index

The front has defined weight coefficient $\mathrm{\&omega;}\left(x_i\right)=q^n\&CenterDot;p^{d_i},$ Wherein q size choose relevantly with locus intercepting point length, and when requiring mobility strong, locus intercepting is lacked, when maneuverability was hanged down, locus intercepting was longer.

The system equation of setting up departments is: X (k+1)=F (k) X (k)+w (k) (4.19)

Observation equation: z (k)=H (k) X (k)+u (k) (4.20)

Motor-driven coefficient is defined as: $\mathrm{\&lambda;}=\frac{{\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="2"\; label="w\; T"\; filenames="H2009101807801E00011.png,H2009101807801E00022.png"\; state="\{\{state\}\}">w</figure-callout>}}{T}^{}{}^2}{{\mathrm{\&sigma;}}_u}---\left(4.21\right)$

σ wherein _uBe the observation noise variance, σ _wBe the system noise variance, T is the sampling period.σ in practical application _uAnd σ _wCan't obtain, can replace with associated arguments.Because coordinates of targets has three kinds in following the tracks of: predicted value, actual observed value and current estimated value.Predicted value is meant the prediction to the present frame target location that in the previous frame processing procedure, obtains, the expression with

; Observed reading representes that present image cuts apart the target location that obtains, and representes with z (k); Current estimated value is meant the actual coordinate of target in the space that estimates according to current observed reading, representes with X (k).

In Kalman filtering $\stackrel{\&OverBar;}{X}\left(k\right)=\hat{X}\left(k\right|k);$ And in match was followed the tracks of, X (k) was the projection of z (k) on fitting a straight line.Because can be used as the true coordinate of target to X (k), then the observed reading noise can be expressed as u (k)=z (k)-X (k), and predicted value and actual value is poor

Embodied the uncertainty of system, represented unpredictable factor in the tracker in other words, so can think that system noise does $w\left(n\right)=\hat{X}(k-1)-\stackrel{\&OverBar;}{X}\left(k\right).$ σ like this _uAnd σ _wCan use following parameter to replace:

${\mathrm{\&sigma;}}_u\&ap;\frac{1}{N}\sqrt{\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(z\left(k\right)-\stackrel{\&OverBar;}{X}\left(k\right)){(z\left(k\right)-\stackrel{\&OverBar;}{X}\left(k\right))}^T}---\left(4.22\right)$

${\mathrm{\&sigma;}}_w\&ap;\frac{1}{N}\sqrt{\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(\hat{X}(k-1)-\stackrel{\&OverBar;}{X}\left(k\right)){(\hat{X}(k-1)-\stackrel{\&OverBar;}{X}\left(k\right))}^T}---\left(4.23\right)$

T is exactly the gap periods of infrared image acquisition, is known parameters, can calculate motor-driven coefficient lambda.Can find σ through test _wValue size and target maneuver property are coincide, in other words σ _wBig more, target travel changes violent; On the contrary, σ _wLittle, then target travel changes mild.

(4) data association after the match

If candidate target of a plurality of tracking chain competitions normally is in the state that multiple goal is intersected.Plan is taked such disposal route: candidate target is given up, and the chain of competition is remembered tracking separately, till not competing.

Also can regard candidate target as suspicious fresh target, it is followed the tracks of, judge whether to be fresh target by continuity; When the track chain can not find the matched candidate target, should remember tracking, up to giving target again for change; Still can't give target for change if surpass certain hour, think that then tracking target loses, this moment, the relation of data association became simple.

In tracking prediction, the single order fitting of a polynomial is that fitting a straight line can reflect movement tendency.The error function of classical least square is the distance function of y coordinate distance curve, and it is more suitable for straight line, square to make error function with distance between beeline and dot; In match, introduce the time weight factor in addition, solve ageing problem; Through the adjustment weight coefficient, the interative computation amount of match is fixed, thus the operand of whole tracking can control, be convenient to system design and realization.

5) target's feature-extraction under the different situations

When target shape changes, utilize the normalized shape recognition of edge feature to seek the characteristic invariant of drawing and reach the target accurate tracking from the electric field angle.

When target became big by remote several pixels, because the detector image-forming angle is different, the shape of same target can change, and how to guarantee that at this moment trace point is constant, and continuing accurate tracking is key issue.In the identification of the coupling of image object, people hope is to locate an amount that can characterize the targeted graphical characteristic, and whether characterize two figures through this characteristic quantity then is same target.Traditional coupling, to the identification of irregular figure following several method commonly used: Fourier describe son coupling identification, reach in the last few years the coupling identification of passing through network learning method etc. based on the coupling identification of invariant moment features method.The recognition methods that Fourier describes son is to mate identification from the angle of frequency field; The invariant moment features method is to characterize piece image with square, and matees identification through this approach of similar features in extraction and statistics and the mechanics; Network learning method is to learn to discern through sample characteristics; The research method of this application project is the implication of giving electric charge point pixel, comes irregular figure is carried out effective recognition from this brand-new angle of electric field, through theoretical derivation, research and experiment, finds out a characteristic quantity that does not change with shape.Below introduce electric field intensity and electromotive force identification shape algorithms in detail

In electricity, the distribution of the equally distributed any electrified body of electric charge electrostatic field that produces in self surrounding space is unique, and this electrified body and it are one to one at the electrostatic field that three dimensions produced.Electrostatic field is only relevant with size, electric density size and the shape of this conductor.Above notion obtains a conclusion: variform uniform charged body is inequality in the Electric Field Distribution that three dimensions produces.

Utilize this conclusion, derive the characteristic invariant of identification shape.When derivation method, exceed the distribution of the surface charge density of considering electrified body, consider the size of electrified body and the influence that shape produces Electric Field Distribution emphatically.

Because the shape information of figure mainly on the edge of; Therefore the edge detection method through Flame Image Process; Can access the marginal information of figure after the binaryzation; If regard these edge pixel points as electrified body, so just can calculate this figure in three-dimensional electromotive force and electric-field intensity distribution, also just can be with electric field and Potential Distribution as differentiating whether same or analogous foundation of two figures.

Method as shown in Figure 8, that this research adopts polygon to approach with " directly " generation " song ", is represented pattern edge with approximate polygon.

Because electromotive force is scalar, so the direct algebraic addition of electromotive force of a bit producing arbitrarily in the space of each bar limit.Though electric field intensity is vector, project on the z direction, the electric field intensity on each bar limit also can algebraic addition.After approaching through polygon like this,

The edge of arbitrary graphic is made up of line segment, and (is observation station to call this point in the following text) electromotive force of being produced and electric field intensity size formula do pattern edge more arbitrarily to the space so

Electromotive force size formula:

$\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{u}_i=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\left|\ln \right|\frac{\tan \frac{{\mathrm{\&theta;}}_{i,1}}{2}}{\tan \frac{{\mathrm{\&theta;}}_{i,2}}{2}}\left|\right|---\left(4.24\right)$

Electric field intensity size formula:

$\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{E}_i=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\frac{\left|{\mathrm{<figure-callout\; id="1"\; label="PO"\; filenames="H2009101807801E00022.png"\; state="\{\{state\}\}">PO</figure-callout>}}_1\right|}{{\left|\mathrm{OP}\right|}^2}\left|(\cos {\mathrm{\&theta;}}_{i,1}-{\cos \mathrm{\&theta;}}_{i,2})\right|---\left(4.25\right)$

θ wherein _{I, 1}, θ _{I, 2}The θ of the i bar line segment of pattern edge is formed in expression respectively ₁With θ ₂

The normalization of observation station is following:

The observation station of each figure is chosen requirement " unanimity "; Its objective is owing to will guarantee that identical shaped figure is identical with electric field intensity at the electromotive force that each self-corresponding observation station produced; Change so will carry out " unanimity " to the observation station of figure earlier, be referred to as the normalization of observation station.

For seeking the normalization observation station; In this research, original formula has been made certain modification; Corresponding normalization observation station should satisfy: do ray from the center of each figure to each observation station; Each bar ray should be perpendicular with graphics plane or the projection on graphics plane consistent with the drift angle of the principal direction that need discern figure with each; And ray is identical with the angle of graphics plane, and then this moment, corresponding observation station equaled the ratio [proof see publish an article] of each graphics area evolution to the ratio of the distance of each centre of figure.

As shown in Figure 9 like this, formula (4.24) (4.25) is improved to as follows

Electromotive force size conversion formula:

$U=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{u}_i=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\left|\ln \right|\frac{\tan \frac{{\mathrm{\&theta;}}_{i,1}}{2}}{\tan \frac{{\mathrm{\&theta;}}_{i,2}}{2}}\left|\right|---\left(4.26\right)$

Electric field intensity size conversion formula:

$E=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{E}_{z,i}\sqrt{S}=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\frac{\left|{\mathrm{<figure-callout\; id="1"\; label="PO"\; filenames="H2009101807801E00022.png"\; state="\{\{state\}\}">PO</figure-callout>}}_1\right|}{{\left|\mathrm{OP}\right|}^2}\left|(\cos {\mathrm{\&theta;}}_{i,1}-\cos {\mathrm{\&theta;}}_{i,2})\right|\sqrt{S}---\left(4.27\right)$

θ wherein _{I, 1}, θ _{I, 2}(as shown in Figure 8) θ of the i bar line segment of pattern edge is formed in expression respectively ₁With θ ₂

Though revised above formula, electric field intensity has still kept the physical significance of needs.Through the normalization of corresponding observation station, just solved the comparability problem of the identical figure of shape that varies in size like this.

Adopt technology and device among the present invention, total system can reach following technical indicator to the detection identification and the accurate tracking of the Weak target under the complex background:

1) input signal: infrared, television video frequency signal (gps signal, laser range finder signal input interface and processing power);

2) output error signal: target is with respect to the angle offset value at center, visual field;

3) the minimum contrast of following the tracks of :≤3%;

4) capture ability: can catch automatically simultaneously with tracking field of view in 4 targets;

5) memory is followed the tracks of: when target temporarily is blocked, tracker should be able to change the memory tracking mode automatically over to, and output keeps track rejection value constantly constant; When target occurs again, can catch automatically again;

6) error output delay :≤20ms; Interface: parallel port, RS232, PCI

7) ATR (automatic checkout system parameter): full frame (768 * 576) are detected in real time, and control information can be by field, frame output.Target property: sky, terrain object.

8) environmental baseline: working temperature :-35 ℃～+ 55 ℃, relative humidity: 95%.

The present invention is a kind of multimode multi-target accurate tracking apparatus and method, and its advantage is: the present invention detects is small and weak under the complex background, spot and general objective, has adaptive tracing and antijamming capability under the good complicated battlefield surroundings, and it is mainly reflected in:

1) preprocessing part:

(1) solved because motion or platform shake cause fuzzy image restoration problem;

(2) having accomplished the ringing effect in the image recovery process suppresses;

(3) adopted improved Weak target algorithm for image enhancement, solved in the practical engineering application under the complex background low signal-to-noise ratio condition signal to noise ratio (S/N ratio) Enhancement problem of Weak target based on Stationary Wavelet Transform and non-linear enhancing operator.

2) based on the detection and the accurate tracking technique of the Weak target of motion compensation:

(1) based on the tracking of trajectory predictions, combination Kalman filtering thought improved curve-fitting method target of prediction direction of motion and speed have been proposed, reduced the complicacy of data association, selected suitable judgment models and decision rule.

(2) study complex background down to the motion prediction and the backoff algorithm of different target, solved the accurate tracking problem under the complex situations such as target overlapping, shake, multiple goal.

(3) in the accurate tracking process, the searching of characteristic invariant.When target shape changes, how to find a kind of characteristic invariant to go to characterize difficult point and the gordian technique that target is an image processing field always.In this project, it does not change the characteristic invariant that utilizes the normalized shape recognition of edge feature to seek to draw from the electric field angle with target rotation, displacement, distortion, and then has effectively reached the purpose of target accurate tracking.

3) set up visible light, infrared target fusion detection model:

The single-sensor target detection probability is lower, and false-alarm probability is high, and it is most important to set up the fusion structure that is fit to.In this project:

(1) having accomplished Multi-sensor Fusion structural system selects;

(2) multisensor time alignment, spacial alignment, target signature matching problem have been solved;

4) systematization and through engineering approaches are used:

Take into full account the requirement that systematization and through engineering approaches are used; In design, consider multiple general requirment; Multiple information interface, integrated powerful soft, hardware resource need to change software and can realize the processing to different target such as marine, aerial not changing following of hardware case.

(4) description of drawings:

Shown in Figure 1 is the formation block diagram of multimode multi-target accurate tracking apparatus

Shown in Figure 2 is the formation block diagram of digital servo platform

Shown in Figure 3 is the formation block diagram of integrated information processing platform

Formation block diagram for compression and transmission equipment shown in Figure 4

Shown in Figure 5 is multimode multi-target accurate tracking apparatus connection layout

Shown in Figure 6 is small target with high precision under complex background and low signal-to-noise ratio tracking process flow diagram

Shown in Figure 7 is target detection, recognition and tracking process flow diagram based on trajectory predictions

Fig. 8 (a) is depicted as the former figure that polygon approaches synoptic diagram

Fig. 8 (b) is depicted as the straight-line segment that polygon approaches synoptic diagram and approaches design sketch

Shown in Figure 9ly concern synoptic diagram for the normalization observation station

Label declaration is following among the figure:

1, digital servo platform 11, ccd video camera 12, infrared sensor

13, high accuracy number servo turntable 14, handle 15, monitor

2, integrated information processing platform 21, information interface 22, high speed digital signal processor

23, servo control processor

3, compression and transmission equipment 31, video compress processor 32, GPRS transport module

(5) embodiment:

With small target with high precision under complex background and low signal-to-noise ratio tracking of the present invention, be applied on the multimode multi-target accurate tracking apparatus of independent development, with the performance index of verification system.This multimode multi-target accurate tracking apparatus, as shown in Figure 1: as to constitute: digital servo platform 1, integrated information processing platform 2, compression and transmission equipment 3 by following three parts; Small target with high precision under complex background and low signal-to-noise ratio tracking of the present invention mainly is achieved in integrated information processing platform.In this device:

1) digital servo platform

As shown in Figure 2; Digital servo platform 1 is by CCD (Charge Coupled Device; Being the CCD imageing sensor) video camera 11, infrared sensor 12, high accuracy number servo turntable 13, handle 14 and monitor 15 form, and also can select two ccd video cameras 11 or two infrared sensors 12 as required for use.The ccd video camera 11 that the present invention adopts can be simulating signal input or digital signal input, and the infrared sensor of employing is that resolution is 768 * 576.

This digital servo platform 1 is the support platform of image acquiring device, and the ccd video camera 11 in the said apparatus is installed in high accuracy number servo turntable 13 two ends respectively with infrared sensor 12, can be with 13 motions of high accuracy number servo turntable.High accuracy number servo turntable 13 can rotate according to the control command that receives simultaneously, and target is carried out accurate tracking, makes target remain on the center, visual field of image acquiring device.

Ccd video camera 11 is used with infrared sensor 12 combinations, obtain visible light and infrared target image information simultaneously, the target signature in comprehensive two kinds of information, thereby the detection probability and the accurate tracking precision of raising target.

2) integrated information processing platform

As shown in Figure 3, this integrated information processing platform 2 is made up of information interface 21, high speed digital signal processor 22, servo control processor 23.The signal processing system that this high speed digital signal processor 22 adopts based on DSP (digital signal processor).High speed digital signal processor 22 receives the image information of importing into from ccd video camera 11 or infrared sensor 12, accomplishes the realization to the target's feature-extraction in visible light and the infrared image under the complex background low signal-to-noise ratio, characteristic matching, target travel prediction and estimation, precise tracking method.Servo control processor 23 is according to the result of target prediction and tracking; Confirm the direction of motion of high accuracy number servo turntable 13; And send control command to high accuracy number servo turntable 13, high accuracy number servo turntable 13 is followed the tracks of target with the result who follows the tracks of according to prediction.

This information processing platform adopts two separate signal processors; Be high speed digital signal processor 22 and servo control processor 23; Respectively the control information of image information and high accuracy number servo turntable 13 is handled; In image pre-service, target recognition and tracking key algorithm,, take multiple algorithm to improve and realize recognition and tracking Weak target with innovation to the characteristics of Weak target under the complex background.

3) compression and transmission equipment

This compression and transmission equipment 3 make this multimode multi-target accurate tracking apparatus have " people is in the loop " function; Pass all information and the image of discerning detected target automatically back command centre, and the instruction of accepting command centre is adjusted to improve the precision of automatic identification to the target of following the tracks of.As shown in Figure 4, compression and transmission equipment 3 are made up of video compress processor 31, GPRS transport module 32.The image input of this video compress processor 31 can be digital video or analog video; Can require to select different interface protocols according to different output; Adopt the video compression algorithm of MEPG-4, rear end GPRS transport module adopts and transmits based on the GPRS wireless channel.

This device has been broken through the single target detection identification tupe of Target Recognition tracker, can also carry out data interaction, image transmission with other detection system networking, and have servo networking control, people in functions such as circuit controls.

The annexation of the relation multimode multi-target accurate tracking apparatus as shown in Figure 5 between the multimode multi-target accurate tracking apparatus each several part; This digital servo platform 1 comprises ccd video camera 11, infrared sensor 12, high accuracy number servo turntable 13, handle 14 and monitor 15 5 parts.Wherein ccd video camera 11 and infrared sensor 12 are installed in the two ends on high accuracy number servo turntable 13 tops respectively; Both link to each other with information interface 21 through cable and carry out image data transmission, and handle 14 and monitor 15 are placed on the both sides of high accuracy number servo turntable 13 bottoms respectively.This integrated information processing platform 2 comprises information interface 21, high speed digital signal processor 22 and servo control processor 23 3 parts, and three parts all are integrated in information processing board and place control box, are placed on high accuracy number servo turntable 13 1 sides.Wherein the handle 14 in this digital servo platform 1 links to each other with information interface 21 and carries out the transmission of control signal; Monitor 15 in this digital servo platform 1 links to each other with information interface 21 and is used to show the image data information of obtaining; High speed digital signal processor 22 links to each other with information interface 21; Be used to obtain ccd video camera 11 and infrared sensor 12 image transmitted data; Servo control processor 23 links to each other with information interface 21, is used to obtain the target detection identifying information of high speed digital signal processor and the positional information of high accuracy number servo turntable 13 feedbacks, and to high accuracy number servo turntable 13 transmission control commands.This compression and transmission equipment 3; Comprise video compress processor 31 and GPRS transport module 32 two parts; Both are integrated on the information processing board respectively; Video compress processor 31 rear ends link to each other with GPRS transport module 32, and video compress processor 31 front ends link to each other with the high speed digital signal processor in the integrated information processing platform 2 22.With regard to the multimode multi-target accurate tracking apparatus generally speaking; Digital servo platform 1 is in the front end of multimode multi-target accurate tracking apparatus; Integrated information processing platform 2 is in the middle-end of multimode multi-target accurate tracking apparatus, and compression and transmission equipment 3 are in the rear end of multimode multi-target accurate tracking apparatus.As shown in Figure 5; A kind of small target with high precision under complex background and low signal-to-noise ratio tracking of the present invention; Be in high speed digital signal processor, to accomplish; Its workflow in whole multimode multi-target accurate tracking apparatus does, at first obtains the visible light and the infrared image of target through ccd video camera 11 and infrared sensor 12, sends the picture signal under the complex background low signal-to-noise ratio to high speed digital signal processor 22 through information interface 21 then; After treated device carries out pre-service, detection to image; Completion sends the target information of following the tracks of to servo control processor 23 simultaneously to Automatic identification of targets and tracking, produces control command by servo control processor 23 and gives high accuracy number servo turntable 13; When picture signal is shown through monitor 15, send the image of the original image information and the tracking target information that superposeed to video compress processor 31 and carry out video compress; Carry out wireless transmission through GPRS transport module 32 then, make command centre in the monitor 15 of control center, observe the target following situation through decoding processor.

As shown in Figure 6; A kind of small target with high precision under complex background and low signal-to-noise ratio tracking of the present invention is used for that Weak target detects automatically under complex background, recognition and tracking, and the concrete steps of its method are 1, the image pre-service under the complex background, low signal-to-noise ratio condition; 2, cut apart based on the objective self-adapting thresholding of binomial distribution judgment criterion; 3, infrared and visible data is carried out multimode and is merged, 4, the target travel prediction with estimate 5, the target's feature-extraction under the different situations.Being described in detail as follows of each step:

1) the image pre-service under complex background, the low signal-to-noise ratio condition

Weak target under the complex background is detected the necessary effective pre-processing method of selecting, and this is to having very important meaning in the follow-up target detection identifying.In our research in the past, a lot of preprocess methods were all carried out emulation and practical applications.Through a large amount of experiments and analysis, the Weak target algorithm for image enhancement based on improved Stationary Wavelet Transform (DSWT) and non-linear enhancing operator is adopted in the image pre-service of native system.

Wavelet transformation has perfect reconstruction ability; Have localization property (retractility) simultaneously in time domain and frequency domain, can focus on any details of object; Multiple dimensioned, multiresolution characteristic; Directional selectivity is coincide with human visual system's directivity.The multiple dimensioned characteristic of wavelet analysis makes it be suitable under the low environment of signal to noise ratio (S/N ratio), carrying out target detection.Its expansion performance can make the parts of images characteristic under certain yardstick, suppressed effectively, and some interested target (like little target) can be highlighted.Wavelet analysis not only can be used in the image pre-service, also can be used in image segmentation and the target travel estimation.

From a large amount of domestic and foreign literature analyses, under complex background, this field of Weak target recognition and tracking, traditional image pre-service based on wavelet transformation is all operated basically as follows:

(1) selects suitable wavelet basis, and image is carried out N layer wavelet decomposition;

(2) threshold value of high frequency coefficient is selected.For ground floor each layer, select a threshold value to handle to the N layer.

(3), calculate the wavelet reconstruction of image according to the high frequency coefficient of the low frequency coefficient of N layer and process modification from ground floor to the N layer.

Although traditional pre-service based on wavelet transformation can obtain good result, when high frequency coefficient was handled, major part had adopted linear unified threshold value, and details such as edge of image have been suffered weakening in various degree.Native system is being summed up traditionally based on wavelet image on the pretreated basis, utilizes based on Stationary Wavelet Transform (DSWT) and the non-linear enhancing operator image to Weak target to strengthen.On the basis of carrying out DSWT, the high-frequency sub-band that obtains has relatively poor resolution, these high-frequency sub-band is carried out the nonlinear operator computing improve and strengthen high-frequency sub-band, thereby reached the effect that filtering strengthens.Experimental result shows that this algorithm can effectively be eliminated 1/f noise, additive white Gaussian noise and multiplicative noise, the signal to noise ratio (S/N ratio) of raising image.This algorithm mainly comprises following three parts:

(1) suppresses noise

(2) selection of threshold

(3) non-linear enhancing operator

Be elaborated with regard to this three part below

(1) suppresses noise

Adopt traditional " global threshold " that image is come denoising, effect is undesirable.I.M.Johnston has proved that the wavelet transformation of correlation noise all is stably on all yardsticks, we can come image is carried out denoising with different threshold values respectively on each yardstick.

The model of supposing discrete picture is following:

g[i，j]＝f[i，j]+ε[i，j](4.1)

Top formula can be write as the form of matrix:

g＝f+ε(4.2)

Wherein, g={g [i, j] } _{I, j}It is the signal that observes.F={f [i, j] } _{I, j}Expression does not have the original signal of noise pollution, ε={ ε [i, j] } _{I, j}, i=1 ..., M; J=1 ..., N is a stationary signal.

DSWT is carried out in (4.2):

X＝Sf (4.3)

V＝Sε (4.4)

Y＝Sg (4.5)

Y＝X+V (4.6)

Wherein S representes steadily wavelet transformation operator of two dimension, and " soft-threshold " function of in image, quoting the Donoho proposition carries out denoising to image:

Y _δ＝T _δοY(4.7)

T _δ＝diag{t[m，m]}

$t[m,m]=\left\{\begin{array}{cc}0,& \left|Y\right[i,j\left]\right|<\mathrm{\&delta;}\\ 1-\frac{\mathrm{\&delta;}}{\left|Y\right[i,j\left]\right|}& \left|Y\right[i,j\left]\right|\&GreaterEqual;\mathrm{\&delta;}\end{array}\right.$

Wherein, i=1 ..., M, j=1 ..., N, m=1 ..., MN

According to formula (4.5) and (4.7), the contravariant of input signal is changed to:

g _δ＝S ^-1οY _δ(4.8)

T wherein _δRelevant with threshold value δ with signal g.

(2) selection of threshold

Suppose that (x y) can use its neighborhood territory pixel linear expression to original signal f.If make

is g [k; L] linear expression; With the neighborhood average image is carried out smoothly can removing a part of noise.

after level and smooth can be used for calculating the threshold value of removing noise.G [i; J] [the i of expression among the g; J] element, it is substituted by

:

$\tilde{g}=Z{\left(g\right[\mathrm{1,1}],...,g[i,j],...g[M,N\left]\right)}^T---\left(4.9\right)$

We think than g [i, the j] threshold value that can better be optimized.

If threshold value δ is too little, mainly show as noise in

; If threshold value δ is too big, can filter a lot of useful signals.

All pixels are implemented same operation, and best threshold value can obtain through following computing:

$\mathrm{OCV}\left(\mathrm{\&delta;}\right)=\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(g\right[i,j]-{\tilde{g}}_{\mathrm{\&delta;}}[i,j\left]\right)}^2---\left(4.10\right)$

Form have a variety of, here, the order ${\tilde{g}}_{\mathrm{\&delta;}}[i,j]=\tilde{g}[i,j],$ Then:

$g[i,j]-{\tilde{g}}_{\mathrm{\&delta;}}[i,j]=\frac{g[i,j]-g_{\mathrm{\&delta;}}[i,j]}{1-\tilde{z}[i,j]}---\left(4.11\right)$

Wherein:

$\tilde{z}[i,j]=\frac{g_{\mathrm{\&delta;}}[i,j]-{\tilde{g}}_{\mathrm{\&delta;}}[i,j]}{g[i,j]-{\tilde{g}}_{\mathrm{\&delta;}}[i,j]}\&ap;z^{\&prime;}[m,n]=\frac{{\&PartialD;g}_{\mathrm{\&delta;}}[i,j]}{\&PartialD;g_{\mathrm{\&delta;}}[k,l]}$

Wherein, i, k=1 ..., M, j, l=1 ..., N, m, n=1 ..., MN

Yet in (4.11), z ' [m, m] is 1 or 0, and is unavailable in actual computation.Therefore providing following formula substitutes (4.10):

$\mathrm{SGCV}\left(\mathrm{\&delta;}\right)=\frac{\frac{1}{\mathrm{MN}}{\left|\right|Y-Y_{\mathrm{\&delta;}}\left|\right|}^2}{{\left[\frac{\mathrm{trace}(I-Z_{\mathrm{\&delta;}}^{\&prime;})}{\mathrm{MN}}\right]}^2}---\left(4.12\right)$

Wherein, the mark of trace representing matrix, || || the expression euclideam norm.I is the unit matrix of M * N.Make δ ^*=argminMSE (δ), $\widetilde{\mathrm{\&delta;}}=\mathrm{Arg}\mathrm{Min}\mathrm{SGCV}\left(\mathrm{\&delta;}\right),$ M.Jansen has proved

It is progressive optimal threshold.

(3) non-linear enhancing operator

1994, A.Laine once provided the non-linear enhancing operator based on DSWT, strengthened the local contrast of image.For convenience's sake, on each yardstick, define the transforming function transformation function of each high-frequency sub-band images respectively:

g[i，j]＝MAG{f[i，j]}(4.13)

Wherein, g [i, j] is the subband that strengthens, and f [i, j] is original sub-band images, and MAG is non-linear enhancing operator.

Make f _s ^r[i, j] is r high-frequency sub-band coefficient on the s decomposition scale, s=1 wherein, and 2 ..., L; R=1,2,3.maxf _s ^rBe all pixel f _s ^rMaximum gradation value in [i, j].f _s ^r[i, j] can be from [max f _s ^r, max f _s ^r] be mapped to [1,1].

Therefore, the scope of a, b and c can be set respectively.The contrast Enhancement Method can be described below:

$g_s^r[i,j]=\left\{\begin{array}{cc}{f}_s^r[i,j],& \left|f_s^r\right[i,j\left]\right|<T_s^r\\ a\&CenterDot;\max f_s^r\left\{\mathrm{sigm}\right[c\left(y_s^r\right[i,j]-b)]-\mathrm{sigm}[-c\left(y_s^r\right[i,j]+b)\left]\right\},& \left|f_s^r\right[i,j\left]\right|\&GreaterEqual;T_s^r\end{array}\right.---\left(4.14\right)$

Wherein,

$y_s^r[i,j]=f_s^r[i,j]/\max f_s^r$

Carry out inverse wavelet transform at last, promptly obtain pretreated image.

Native system is on the characteristic and background model basis of the infrared small object of fully analyzing and researching, and the algorithm of employing Stationary Wavelet Transform (DSWT) and non-linear enhancing operator carries out the pre-service of figure image intensifying to infrared small object.Experimental result show this algorithm not only to the background of Weak target suppress, target strengthens that effect is preferably arranged, and noise, the background of general objective suppressed to have equally effect preferably.

2) cut apart based on the objective self-adapting thresholding of binomial distribution judgment criterion

Based on the characteristics of Weak target (low SNR, complex background), rely on single frames not reach to its detection, must rely on image sequence.In the present Weak target of doing detected, the detection identification probability that provides was generally more than or equal to 98%, and false-alarm probability is smaller or equal to 10 ^-6,, choose noise gate with them exactly if these two parameters are embodied in the detection.On the one hand, if according to false alarm rate smaller or equal to 10 ^-6Choose noise gate, then thresholding is very high, and a lot of targets will be lost; On the other hand, if choose noise gate according to detecting identification probability more than or equal to 98%, then choosing of thresholding should be very low, to guarantee taking out all impact points, can cause very big false alarm rate like this.Weak target under this low signal-to-noise ratio condition detects, and can't realize by single frames, must utilize the relevant information of multiple image, continuity, the consistance of the motion feature of target and movement locus is combined consider.Calculate the detection probability and the false-alarm probability of single frames from total detection probability and false-alarm probability, thereby choose the key issue of rational noise gate for solving.

With the theory of probability is the basis; We are in entire image detects; According to the relation of single frames detection probability, single frames false-alarm probability and total detection probability and total false-alarm probability, solve definite problem of be correlated with in the sequence image detection frame number and thresholding, for low false alarm rate, high detection rate provide theoretical assurance.

Regard the target detection in every two field picture as independently repeated experiments; According to the principle in the theory of probability; Detection probability should be obeyed binomial distribution, sets up mathematical model in view of the above, supposes that picture noise is the white noise of Gaussian distribution after the pre-service; According to statistical theory, when image carries out the single frames Threshold detection [derivation is referring to publishing an article]:

$p_d={\&Integral;}_v^{+\&infin;}p\left(x\right)\mathrm{dx}={\&Integral;}_v^{+\&infin;}\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\exp (-\frac{{(x-\mathrm{\&mu;}-\mathrm{\&sigma;S})}^2}{2{\mathrm{\&sigma;}}^2})\mathrm{dx}=1-\mathrm{\&Phi;}\left(\frac{v-\mathrm{\&mu;}-\mathrm{\&sigma;S}}{\mathrm{\&sigma;}}\right)$

$=\mathrm{\&Phi;}\left(\frac{\mathrm{\&mu;}+\mathrm{\&sigma;S}-v}{\mathrm{\&sigma;}}\right)---\left(4.15\right)$

$p_f={\&Integral;}_v^{+\&infin;}\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\exp (-\frac{{(x-\mathrm{\&mu;})}^2}{2{\mathrm{\&sigma;}}^2})\mathrm{dx}=1-\mathrm{\&Phi;}\left(\frac{v-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)=\mathrm{\&Phi;}\left(\frac{\mathrm{\&mu;}-v}{\mathrm{\&sigma;}}\right)$

$\frac{\mathrm{\&mu;}+\mathrm{\&sigma;S}-v}{\mathrm{\&sigma;}}={\mathrm{\&Phi;}}^{-1}\left(p_d\right)---\left(4.16\right)$

ν=μ+σ S-σ Φ wherein ^-1(p _d) or ν=μ-σ Φ ^-1(p _f)

Can release the relation of single frames detection probability and single frames false-alarm probability:

Φ ^-1(P _f)-Φ ^-1(P _d)＝S (4.17)

In the formula: ν---detection threshold, σ ²---the noise mean square deviation

μ---be the noise average after K frame background offsets

S---signal to noise ratio snr is defined as the ratio of amplitude with the mean square deviation of noise of signal

The relation of total detection probability and single frames detection probability:

$P_D(i\&GreaterEqual;k)=1-P_D=1-\underset{k=0}{\overset{i-1}{\mathrm{\&Sigma;}}}{C}_n^k{p}_d^k{(1-p_d)}^{n-k}---\left(4.18\right)$

In above-mentioned experiment, suppose that the single frames detection probability is P _d=0.90, require total detection probability P _D=0.98.When in n width of cloth image, when target has occurred k time, can reach requirement so.For example, when each collection 16 two field pictures made a decision, if target occurs 11 times, then total detection probability can reach requirement.

Therefore can be by single frames detection probability, single frames false-alarm probability and always detection probability and the always relation of false-alarm probability, definite problem of relevant frame number and thresholding during the solution sequence image detects.

3) infrared and visible data is carried out the multimode fusion

Detect with the visible light sensor fusion goal based on infrared, can improve detection probability, the reduction false-alarm probability of target.

Need carry out images match to visible light and infrared image fusion, at first the target of different sensors carried out feature extraction, can draw the transformation relation between visible light and the infrared image through the object matching result then, like translation, rotation, convergent-divergent etc.The concrete method for registering that uses is based on the reference mark coupling of least square, phase correlation method, template matching method etc.

After image registration, according to nearest neighbouring rule, the coordinate of same target in images after registration should be coincidence or very approaching, can be mapped the target signature in two kinds of images through proximity or similarity measure.

We are used for the fusion of visible light and infrared image with the interconnected wave filter of multisensor probability data, reduce false alarm rate.The interconnected basic thought of probability data is: so long as effectively detect information, all possibly come from target, it is different that just each information comes from the probability of target.This method has utilized all information in the track window obtaining possible posterior information, and provides each probability weight coefficient and weighted sum thereof according to a large amount of correlation computations, upgrades dbjective state with it then.This research adopts the method to carry out the target signature fusion.

After the Target Fusion processing, enriched clarification of objective information, improved target detection probability, reduced false alarm rate.

The treatment step that fusion infrared and visible data is adopted is following:

A) visible light and infrared image are carried out target's feature-extraction respectively;

B) characteristic of extracting is carried out the characteristic matching of different sensors, confirm the transformation relation between visible light and the infrared image;

To mating the target signature in the image of back, the interconnected filter process of probability of use obtains fused data, thereby improves the degree of confidence of target detection identification, and rejects false target.

4) target travel prediction and estimation

Multiple goal in the target following process, intersection, shake, memory tracking etc. all relate to the movement locus of the estimation forecasting problem Weak target of target; There is shake between frame and the frame; Do not have the direction consistance; And the area of other target such as bird is bigger slightly than target, and its movement locus has more directivity than Weak target.This is the shake that has caused image owing to reasons such as imaging system, spatial light interference, air vibrations; In the time of shake the geometric center of target is changed between neighbor; If data are without processing; Target location between the successive frame is because the existence of shake, and changes not according to the consistance of direction, and this just brings error to the target accurate tracking that retrains (like the filter's prediction of routine) with directivity.

Influence how to eliminate the jitter phenomenon generation is the major issue that actual tracking will solve.In project, solve these class methods and adopted two kinds of methods:

● based on the motion prediction of the improved curve fitting of Kalman wave filter thought.Solve the motion prediction problem under the situation such as random shake, target overlapping, memory tracking.

● after being identified as target, will slightly following the tracks of and transfer accurate tracking to, and convert position of form center into centroid position, the jitter phenomenon when reducing to follow the tracks of improves tracking accuracy.

Because the Kalman wave filter can not reflect the motion change of target immediately, so when target maneuver property was strong, the error that Kalman follows the tracks of can be bigger.We have defined the maneuverability parameter from the Kalman kinetics equation, simultaneously the match mode are improved, and strengthen the performance of its tracking prediction.

Shown in Figure 7 is target detection, recognition and tracking process flow diagram based on trajectory predictions.The target travel prediction is the pre-service of image process with the process of estimating, carries out clarification of objective and extracts, and obtains potential target; Judgment criterion through binomial distribution judges whether it is fresh target, sets up equation of locus through the kalman theory, if equation of locus is set up; Then need upgrade equation of locus; Target of prediction obtains predicted position at the track of next frame simultaneously, and with match correction Kalman tracking filter correction predicted position.

Below be several Points Concerning and improvement based on the theoretical improved curve fitting target of prediction running orbit research of Kalman:

1) improved least-squares line approximating method

Criterion function in the classical least square fitting is the quadratic sum of the distance of each data point along ordinate y to matched curve.Tentation data point (x _i, y _i), y=f (x) wherein.Do the curve that least square fitting obtains and be based on following criterion:

$\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{{\mathrm{\&delta;}}^2}_i=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{\&omega;}\left(x_i\right){[s^{*}\left(x\right)-f\left(x_i\right)]}^2=\underset{s\left(x\right)\&Element;\mathrm{\&Phi;}}{\mathrm{Min}}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{\&omega;}\left(x_i\right){[s\left(x\right)-f\left(x_i\right)]}^2,$ But quadratic sum with distance $\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_i^2=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{\&omega;}\left(x_i\right){\left(\frac{|y-a_{i}x-b_i|}{\sqrt{1+a_i^2}}\right)}^2=\mathrm{Min}$ It is more suitable to do criterion function, and the latter is the least moment of inertia in the mechanics.

2) the locus intercepting point is done match

In application, only consider to close on several or tens tracing points, if weighting coefficient ω is (x _i) ≡ 1, be equivalent to obtain recent tracing point with the rectangular window locus intercepting.Rectangular window can not reflect the variation of importance, and can not smoothly block; Exponential function can reflect data point importance over time, and exponential function is smooth, so adopt the window index function to come data intercept.

Consider the continuity and the consistance of target travel; The size of every frame instantaneous velocity, direction and mass motion trend are compared and big saltus step should do not arranged; When this change is big, wherein should comprise bigger error, suppress this error should for the less weights of this data points.

According to above criterion, through experiment and theoretical derivation, the value of definition weight coefficient:

$\mathrm{\&omega;}\left(x_i\right)=q^n\&CenterDot;p^{d_i}$

Wherein | q|＜1, | p|＜1, d _iBe the distance of data point to fitting a straight line Ax+By+C=0, $d_i=\frac{|{\mathrm{Ax}}_i+{\mathrm{By}}_i+C|}{\sqrt{A^2+{\mathrm{<figure-callout\; id="2"\; label="B"\; filenames="H2009101807801E00011.png,H2009101807801E00022.png"\; state="\{\{state\}\}">B</figure-callout>}}^2}};$ N=N-i, N are current frame numbers, and N-M≤i＜N, M are the numbers of fitting data point.The value of q and p can obtain through experiment, generally gets between 0.7～0.8.

The tracing point number M can be selected to fix, and controls the effect of tracing point to match through adjustment weight coefficient size, for example selects little attenuation coefficient, is equivalent to the short track of intercepting, selects bigger attenuation coefficient, is equivalent to the long track of intercepting.

3) mobility index

The front has defined weight coefficient $\mathrm{\&omega;}\left(x_i\right)=q^n\&CenterDot;p^{d_i},$ Wherein q size choose relevantly with locus intercepting point length, and when requiring mobility strong, locus intercepting is lacked, when maneuverability was hanged down, locus intercepting was longer.

The system equation of setting up departments is: X (k+1)=F (k) X (k)+w (k) (4.19)

Observation equation: z (k)=H (k) X (k)+u (k) (4.20)

Motor-driven coefficient is defined as: $\mathrm{\&lambda;}=\frac{{\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="2"\; label="w\; T"\; filenames="H2009101807801E00011.png,H2009101807801E00022.png"\; state="\{\{state\}\}">w</figure-callout>}}{T}^{}{}^2}{{\mathrm{\&sigma;}}_u}---\left(4.21\right)$

σ wherein _uBe the observation noise variance, σ _wBe the system noise variance, T is the sampling period.σ in practical application _uAnd σ _wCan't obtain, can replace with associated arguments.Because coordinates of targets has three kinds in following the tracks of: predicted value, actual observed value and current estimated value.Predicted value is meant the prediction to the present frame target location that in the previous frame processing procedure, obtains, the expression with

; Observed reading representes that present image cuts apart the target location that obtains, and representes with z (k); Current estimated value is meant the actual coordinate of target in the space that estimates according to current observed reading, representes with X (k).

In Kalman filtering $\stackrel{\&OverBar;}{X}\left(k\right)=\hat{X}\left(k\right|k);$ And in match was followed the tracks of, X (k) was the projection of z (k) on fitting a straight line.Because can be used as the true coordinate of target to X (k), then the observed reading noise can be expressed as u (k)=z (k)-X (k), and predicted value and actual value is poor

Embodied the uncertainty of system, represented unpredictable factor in the tracker in other words, so can think that system noise does $w\left(n\right)=\hat{X}(k-1)-\stackrel{\&OverBar;}{X}\left(k\right).$ σ like this _uAnd σ _wCan use following parameter to replace:

${\mathrm{\&sigma;}}_u\&ap;\frac{1}{N}\sqrt{\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(z\left(k\right)-\stackrel{\&OverBar;}{X}\left(k\right)){(z\left(k\right)-\stackrel{\&OverBar;}{X}\left(k\right))}^T}---\left(4.22\right)$

${\mathrm{\&sigma;}}_w\&ap;\frac{1}{N}\sqrt{\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}(\hat{X}(k-1)-\stackrel{\&OverBar;}{X}\left(k\right)){(\hat{X}(k-1)-\stackrel{\&OverBar;}{X}\left(k\right))}^T}---\left(4.23\right)$

T is exactly the gap periods of infrared image acquisition, is known parameters, can calculate motor-driven coefficient lambda.Can find σ through test _wValue size and target maneuver property are coincide, in other words σ _wBig more, target travel changes violent; On the contrary, σ _wLittle, then target travel changes mild.

4) data association after the match

If candidate target of a plurality of tracking chain competitions normally is in the state that multiple goal is intersected.Plan is taked such disposal route: candidate target is given up, and the chain of competition is remembered tracking separately, till not competing.

Also can regard candidate target as suspicious fresh target, it is followed the tracks of, judge whether to be fresh target by continuity; When the track chain can not find the matched candidate target, should remember tracking, up to giving target again for change; Still can't give target for change if surpass certain hour, think that then tracking target loses, this moment, the relation of data association became simple.

In tracking prediction, the single order fitting of a polynomial is that fitting a straight line can reflect movement tendency.The error function of classical least square is the distance function of y coordinate distance curve, and it is more suitable for straight line, square to make error function with distance between beeline and dot; In match, introduce the time weight factor in addition, solve ageing problem; Through the adjustment weight coefficient, the interative computation amount of match is fixed, thus the operand of whole tracking can control, be convenient to system design and realization.

5) target's feature-extraction under the different situations

When target shape changes, utilize the normalized shape recognition of edge feature to seek the characteristic invariant of drawing and reach the target accurate tracking from the electric field angle.

When target became big by remote several pixels, because the detector image-forming angle is different, the shape of same target can change, and how to guarantee that at this moment trace point is constant, and continuing accurate tracking is key issue.In the identification of the coupling of image object, people hope is to locate an amount that can characterize the targeted graphical characteristic, and whether characterize two figures through this characteristic quantity then is same target.Traditional coupling, to the identification of irregular figure following several method commonly used: Fourier describe son coupling identification, reach in the last few years the coupling identification of passing through network learning method etc. based on the coupling identification of invariant moment features method.The recognition methods that Fourier describes son is to mate identification from the angle of frequency field; The invariant moment features method is to characterize piece image with square, and matees identification through this approach of similar features in extraction and statistics and the mechanics; Network learning method is to learn to discern through sample characteristics; The research method of this application project is the implication of giving electric charge point pixel, comes irregular figure is carried out effective recognition from this brand-new angle of electric field, through theoretical derivation, research and experiment, finds out a characteristic quantity that does not change with shape.Below introduce electric field intensity and electromotive force identification shape algorithms algorithm in detail

In electricity, the distribution of the equally distributed any electrified body of electric charge electrostatic field that produces in self surrounding space is unique, and this electrified body and it are one to one at the electrostatic field that three dimensions produced.Electrostatic field is only relevant with size, electric density size and the shape of this conductor.Above notion obtains a conclusion: variform uniform charged body is inequality in the Electric Field Distribution that three dimensions produces.

Utilize this conclusion, derive the characteristic invariant of identification shape.When derivation method, exceed the distribution of the surface charge density of considering electrified body, consider the size of electrified body and the influence that shape produces Electric Field Distribution emphatically.

Because the shape information of figure mainly on the edge of; Therefore the edge detection method through Flame Image Process; Can access the marginal information of figure after the binaryzation; If regard these edge pixel points as electrified body, so just can calculate this figure in three-dimensional electromotive force and electric-field intensity distribution, also just can be with electric field and Potential Distribution as differentiating whether same or analogous foundation of two figures.

Method as shown in Figure 8, that this research adopts polygon to approach with " directly " generation " song ", is represented pattern edge with approximate polygon.

Because electromotive force is scalar, so the direct algebraic addition of electromotive force of a bit producing arbitrarily in the space of each bar limit.Though electric field intensity is vector, project on the z direction, the electric field intensity on each bar limit also can algebraic addition.After approaching through polygon like this,

The edge of arbitrary graphic is made up of line segment, and (is observation station to call this point in the following text) electromotive force of being produced and electric field intensity size formula do pattern edge more arbitrarily to the space so

Electromotive force size formula:

$\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{u}_i=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\left|\ln \right|\frac{\tan \frac{{\mathrm{\&theta;}}_{i,1}}{2}}{\tan \frac{{\mathrm{\&theta;}}_{i,2}}{2}}\left|\right|---\left(4.24\right)$

Electric field intensity size formula:

$\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{E}_i=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\frac{\left|{\mathrm{<figure-callout\; id="1"\; label="PO"\; filenames="H2009101807801E00022.png"\; state="\{\{state\}\}">PO</figure-callout>}}_1\right|}{{\left|\mathrm{OP}\right|}^2}\left|(\cos {\mathrm{\&theta;}}_{i,1}-{\cos \mathrm{\&theta;}}_{i,2})\right|---\left(4.25\right)$

θ wherein _{I, 1}, θ _{I, 2}The θ of the i bar line segment of pattern edge is formed in expression respectively ₁With θ ₂

The normalization of observation station is following:

The observation station of each figure is chosen requirement " unanimity "; Its objective is owing to will guarantee that identical shaped figure is identical with electric field intensity at the electromotive force that each self-corresponding observation station produced; Change so will carry out " unanimity " to the observation station of figure earlier, be referred to as the normalization of observation station.

For seeking the normalization observation station; In this research, original formula has been made certain modification; Corresponding normalization observation station should satisfy: do ray from the center of each figure to each observation station; Each bar ray should be perpendicular with graphics plane or the projection on graphics plane consistent with the drift angle of the principal direction that need discern figure with each; And ray is identical with the angle of graphics plane, and then this moment, corresponding observation station equaled the ratio [proof see publish an article] of each graphics area evolution to the ratio of the distance of each centre of figure.

As shown in Figure 9 like this, formula (4.24) (4.25) is improved to as follows

Electromotive force size conversion formula:

$U=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{u}_i=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\left|\ln \right|\frac{\tan \frac{{\mathrm{\&theta;}}_{i,1}}{2}}{\tan \frac{{\mathrm{\&theta;}}_{i,2}}{2}}\left|\right|---\left(4.26\right)$

Electric field intensity size conversion formula:

$E=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{E}_{z,i}\sqrt{S}=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\frac{\left|{\mathrm{<figure-callout\; id="1"\; label="PO"\; filenames="H2009101807801E00022.png"\; state="\{\{state\}\}">PO</figure-callout>}}_1\right|}{{\left|\mathrm{OP}\right|}^2}\left|(\cos {\mathrm{\&theta;}}_{i,1}-\cos {\mathrm{\&theta;}}_{i,2})\right|\sqrt{S}---\left(4.27\right)$

θ wherein _{I, 1}, θ _{I, 2}(as shown in Figure 8) θ of the i bar line segment of pattern edge is formed in expression respectively ₁With θ ₂

Though revised above formula, electric field intensity has still kept the physical significance of needs.Through the normalization of corresponding observation station, just solved the comparability problem of the identical figure of shape that varies in size like this.

Adopt technology and device among the present invention, total system can reach following technical indicator to the detection identification and the accurate tracking of the Weak target under the complex background:

1) input signal: infrared, television video frequency signal (gps signal, laser range finder signal input interface and processing power);

2) output error signal: target is with respect to the angle offset value at center, visual field;

3) the minimum contrast of following the tracks of :≤3%;

4) capture ability: can catch automatically simultaneously with tracking field of view in 4 targets;

5) memory is followed the tracks of: when target temporarily is blocked, tracker should be able to change the memory tracking mode automatically over to, and output keeps track rejection value constantly constant; When target occurs again, can catch automatically again;

6) error output delay :≤20ms; Interface: parallel port, RS232, PCI

7) ATR (automatic checkout system parameter): full frame (768 * 576) are detected in real time, and control information can be by field, frame output.

Target property: sky, terrain object.

8) environmental baseline: working temperature :-35 ℃～+ 55 ℃, relative humidity: 95%.

Claims (2)

1. small target with high precision under complex background and low signal-to-noise ratio tracking, it is characterized in that: its step is following:

(1), the image pre-service under complex background, the low signal-to-noise ratio condition: adopt Weak target algorithm for image enhancement, come image is carried out denoising through on each yardstick of wavelet transformation, choosing different threshold values respectively based on improved Stationary Wavelet Transform and non-linear enhancing operator;

(2), cut apart based on the objective self-adapting thresholding of binomial distribution judgment criterion: the relation between single frames detection probability, single frames false-alarm probability and total detection probability and the total false-alarm probability is set up the model based on theory of probability binomial distribution criterion, has solved definite problem of be correlated with during sequence image detects frame number and thresholding;

(3), infrared and visible data is carried out multimode and is merged: after visible light and infrared image are mated; Adopt the interconnected wave filter of multisensor probability data that the target signature in two kinds of images is mapped to the target signature in the image; Obtain fused data; Thereby improve the degree of confidence of target detection identification, and reject false target;

(4), target travel prediction and estimation: adopt and carry out motion prediction, solve the motion prediction problem under random shake, target overlapping, the memory tracking situation based on the improved curve fitting algorithm of Kalman wave filter thought;

Target's feature-extraction when (5) target shape changes: when target shape changes, adopt the method utilize the normalized shape recognition of edge feature to seek the characteristic invariant of drawing from the electric field angle to reach to the target accurate tracking.

2. a kind of small target with high precision under complex background and low signal-to-noise ratio tracking according to claim 1; It is characterized in that: the target's feature-extraction method when target shape changes in the said step (5); Be the implication of giving electric charge point, find out a characteristic quantity that does not change from the electric field angle and come irregular figure is carried out effective recognition with shape pixel.

Images