CN103149940B - A kind of unmanned plane target tracking in conjunction with average and variance and particle filter - Google Patents

Abstract

The present invention proposes a kind of unmanned plane target tracking in conjunction with average and variance and particle filter, comprise the following steps: in average and variance tracing process, the average and variance track algorithm based on bandwidth matrices is built, with adaptive updates target scale window in tracing process according to bandwidth matrices; Testing result according to average and variance track algorithm and particle filter algorithm sets up weighted sum data fusion object localization method; According to weighted sum data fusion object localization method determination unmanned plane target position; According to target heavy convergence method, the particle in particle filter algorithm is sampled to generate the particle filter algorithm that based target heavily restrains; Obtain target expanded search strategy according to the particle filter algorithm that target heavily restrains, and accordingly target is followed the tracks of.To the real-time localization and tracking of target under the complex situations such as embodiments of the invention can realize dynamic scene, illumination variation, dimensional variation, block, have that real-time is good, strong adaptability and an advantage such as extensibility is good.

Description

A kind of unmanned plane target tracking in conjunction with average and variance and particle filter

Technical field

The present invention relates to field of computer technology, particularly a kind of unmanned plane target tracking in conjunction with average and variance and particle filter.

Background technology

Target following has important researching value in Science and engineering.In the process of unmanned plane tracking target flight, due to the relative motion between video camera and target, application scenarios complicated and changeable, and the video image gathered generally has the illumination variation features such as obviously, in image foreign material or noise are significantly, target is at least partially obscured or blocks completely, target carriage change is large, makes the target following based on sequence image be difficult to realize.

Target tracking algorism can be divided into determinacy track algorithm and randomness track algorithm two class, and average and variance algorithm is a kind of determinacy track algorithm.This track algorithm can be converted into optimization problem usually, namely finds the Optimum Matching of target.Target tracking algorism based on average and variance is simple, real-time good, but easily converges to Local Extremum, can not carry out Automatic adjusument to tracking window, when target maneuver is comparatively strong, dimensional variation is obvious, there is blocking in various degree, or illumination occurs when changing more by force, tracking effect is undesirable.

Particle filter tracking algorithm is a kind of randomness track algorithm, it adopts multiple particle, effectively have expressed the uncertainty of tracking, stronger robustness is revealed to the Track Table under the tracking of non-rigid object and partial occlusion, but there is sample degeneracy phenomenon, precision of prediction is by the impact of accumulation error effects, and calculated amount is larger, and real-time is poor.

Summary of the invention

The present invention is intended at least one of solve the problems of the technologies described above.

For this reason, the object of the invention is to propose one can realize dynamic scene, illumination variation, dimensional variation, the complex situations such as to block under to the real-time localization and tracking of target, have that real-time is good, a unmanned plane target tracking in conjunction with average and variance and particle filter of strong adaptability and the advantage such as extensibility is good.

To achieve these goals, embodiments of the invention propose a kind of unmanned plane target tracking in conjunction with average and variance and particle filter, comprise the following steps: in average and variance tracing process, the average and variance track algorithm based on described bandwidth matrices is built, with adaptive updates target scale window in tracing process according to bandwidth matrices; Testing result according to described average and variance track algorithm and particle filter algorithm sets up weighted sum data fusion object localization method; Described unmanned plane target position is determined according to described weighted sum data fusion object localization method; Sample to generate the particle filter algorithm of heavily restraining based on described target to the particle in described particle filter algorithm according to the heavy convergence method of target; And obtain target expanded search strategy according to the particle filter algorithm that described target heavily restrains, and according to described target expanded search strategy, target is followed the tracks of.

According to the unmanned plane target tracking in conjunction with average and variance and particle filter of the embodiment of the present invention, can dynamic scene, illumination effect, dimensional variation, block the complex scenes such as interference under, realize the real-time localization and tracking to tracking target, and also can be applicable to different scenes and platform, therefore, embodiments of the invention have that real-time is good, strong adaptability and the advantage such as extensibility is good.

In addition, the unmanned plane target tracking in conjunction with average and variance and particle filter according to the above embodiment of the present invention can also have following additional technical characteristic:

In an embodiment of the present invention, the average and variance track algorithm based on described bandwidth matrices is built according to bandwidth matrices, comprise further: describe object module and candidate family according to the weighted probability density distribution function based on color characteristic, wherein, the feature space of described object module and described candidate family is the one dimension hsv color space vector of 32*32+10; Elliptic region is utilized to represent described target, according to the object initialization way selection rectangle of man-machine interaction as described target, and calculate in described rectangle and connect ellipse, wherein, described each in connect target described in ellipse representation center horizontal coordinate in the picture and vertical coordinate, the major semi-axis of described ellipse and the angle of minor semi-axis, the main shaft of described ellipse and the horizontal coordinate positive dirction of described image.In described tracing process, optimum kernel function window width is calculated according to described bandwidth matrices.

In an embodiment of the present invention, determine described unmanned plane target position according to described weighted sum data fusion object localization method, comprise further: the average and variance track algorithm based on described bandwidth matrices obtains the first center of described target; The second center of described target is obtained according to described particle prediction; Set up Strategy of data fusion according to described first center, described second center and bandwidth matrices parameter, determine the center of described target.

In an embodiment of the present invention, generate the particle filter algorithm of heavily restraining based on described target, comprise further: N number of particle of sampling in the tracking target elliptic region that described average and variance track algorithm obtains; Centered by each particle, using the state transition equation of second-order autoregressive model as described particle; If the region at all particle places is same candidate region after state transfer, then calculate the hsv color proper vector of all particles, and described proper vector size is (10*10+10); Calculate the similarity of each candidate region and object module, the particle that the weight of getting described particle is directly proportional to similarity carries out filtering; Summation is weighted to filtered all particles, and acquisition predicts by particle filter algorithm the target location obtained; In described particle resampling process, determine the center of described target according to described weighted sum data fusion object localization method, and with described center for reference point, the N number of particle of uniform sampling in first area.

In an embodiment of the present invention, according to described target expanded search strategy, target is followed the tracks of, comprise further: the movement locus according to described target carries out one-step prediction; Localized target search is carried out to described target; Global object search is carried out to described target.

In an embodiment of the present invention, described bandwidth matrices is positive definite symmetric matrices, and there is quantitative relation between described bandwidth matrices and described Target ellipse; To probability density function optimizing, calculate optimum bandwidth matrix; By described optimum bandwidth matrix application in average drifting track algorithm, to realize the adaptive tracing to target scale.

In an embodiment of the present invention, described target expanded search strategy comprises: use one-step prediction Target Searching Method search target; If described one-step prediction Target Searching Method is searched for unsuccessfully, localized target searching method is used to search for described target; If after described localized target searching method failure, then global object searching method is used to search for.

In an embodiment of the present invention, described one-step prediction Target Searching Method comprises: if when the image of pre-treatment is kth frame, the center of described target is expressed as successively in sequence image: y ₀, y ₁..., y _k-1, y _k, y _k+1..., candidate target position y _k+1estimation formulas be: y _k+1-y _k=y _k-y _k-1; Get y _k+1as the central point of described candidate target, candidate region size, direction are all identical with former frame, extract candidate region proper vector, mate with described To Template proper vector; If matching similarity exceedes first threshold, be then judged as finding target; If matching similarity is less than first threshold, then enter the Local Search stage.

In an embodiment of the present invention, described localized target searching method comprises: with y _kcentered by, the several candidate region of stochastic distribution around, size, the direction of getting described candidate region are identical with former frame, extract each candidate region proper vector and mate with To Template proper vector; Iterative search is carried out, until find described target in the candidate region selecting similarity to exceed Second Threshold; If consecutive numbers frame search is less than described target, then enter the global search stage; The image frames numbers entering Local Search adjusts according to described picture-taken frequency, unmanned plane during flying speed and described target speed.

In an embodiment of the present invention, described global object searching method comprises: the boundary rectangle getting described Target ellipse template, as detection template, carries out traversal search by described detection template to described image; Detect described template long and wide, and detect described template transverse shifting distance and vertically move distance; Calculate the matching similarity between described To Template and described detection template; Retain the surveyed area that all similarities are greater than the 3rd threshold value, and the surveyed area selecting matching similarity maximum is as described candidate target position.

Additional aspect of the present invention and advantage will part provide in the following description, and part will become obvious from the following description, or be recognized by practice of the present invention.

Accompanying drawing explanation

Above-mentioned and/or additional aspect of the present invention and advantage will become obvious and easy understand from accompanying drawing below combining to the description of embodiment, wherein:

Fig. 1 is according to an embodiment of the invention in conjunction with the process flow diagram of the unmanned plane target tracking of average and variance and particle filter;

Fig. 2 is according to an embodiment of the invention in conjunction with the structural drawing of the unmanned plane target tracking of average and variance and particle filter;

Fig. 3 is according to an embodiment of the invention in conjunction with bandwidth matrices and the sample elliptical relationship schematic diagram of the unmanned plane target tracking of average and variance and particle filter;

Fig. 4 is according to an embodiment of the invention in conjunction with the average and variance track algorithm process flow diagram based on bandwidth matrices of the unmanned plane target tracking of average and variance and particle filter;

Fig. 5 heavily restrains sampling schematic diagram in conjunction with the target of the unmanned plane target tracking of average and variance and particle filter according to an embodiment of the invention;

Fig. 6 is according to an embodiment of the invention in conjunction with the particle filter tracking algorithm flow chart that the based target of the unmanned plane target tracking of average and variance and particle filter is heavily restrained;

Fig. 7 is according to an embodiment of the invention in conjunction with the kernel function selecting predictors schematic diagram of the unmanned plane target tracking of average and variance and particle filter;

Fig. 8 is according to an embodiment of the invention in conjunction with the process flow diagram of the target tracking algorism in conjunction with average and variance and particle filter of the unmanned plane target tracking of average and variance and particle filter;

Fig. 9 divides schematic diagram in conjunction with the target control area of the unmanned plane target tracking of average and variance and particle filter according to an embodiment of the invention; And

Figure 10 is according to an embodiment of the invention in conjunction with the unmanned plane target tracker structural drawing of the unmanned plane target tracking of average and variance and particle filter.

Embodiment

Be described below in detail embodiments of the invention, the example of described embodiment is shown in the drawings, and wherein same or similar label represents same or similar element or has element that is identical or similar functions from start to finish.Being exemplary below by the embodiment be described with reference to the drawings, only for explaining the present invention, and can not limitation of the present invention being interpreted as.

In describing the invention, it will be appreciated that, term " " center ", " longitudinal direction ", " transverse direction ", " on ", D score, " front ", " afterwards ", " left side ", " right side ", " vertically ", " level ", " top ", " end ", " interior ", orientation or the position relationship of the instruction such as " outward " are based on orientation shown in the drawings or position relationship, only the present invention for convenience of description and simplified characterization, instead of indicate or imply that the device of indication or element must have specific orientation, with specific azimuth configuration and operation, therefore limitation of the present invention can not be interpreted as.In addition, term " first ", " second " only for describing object, and can not be interpreted as instruction or hint relative importance.

In describing the invention, it should be noted that, unless otherwise clearly defined and limited, term " installation ", " being connected ", " connection " should be interpreted broadly, and such as, can be fixedly connected with, also can be removably connect, or connect integratedly; Can be mechanical connection, also can be electrical connection; Can be directly be connected, also indirectly can be connected by intermediary, can be the connection of two element internals.For the ordinary skill in the art, concrete condition above-mentioned term concrete meaning in the present invention can be understood.

Below in conjunction with accompanying drawing, the unmanned plane target tracking in conjunction with average and variance and particle filter according to the embodiment of the present invention is described.

Fig. 1 is according to an embodiment of the invention in conjunction with the process flow diagram of the unmanned plane target tracking of average and variance and particle filter.

As shown in Figure 1, according to an embodiment of the invention in conjunction with the unmanned plane target tracking of average and variance and particle filter, comprise the following steps:

Step S101, in average and variance tracing process, builds the average and variance track algorithm based on bandwidth matrices, with adaptive updates target scale window in tracing process according to bandwidth matrices.Particularly, describe object module and candidate family according to the weighted probability density distribution function based on color characteristic, wherein, the feature space of object module and candidate family is the one dimension hsv color space vector of 32*32+10.And utilize elliptic region to represent the tracking target of unmanned plane, according to the object initialization way selection rectangle of man-machine interaction as the target that will follow the tracks of, and calculate in rectangle and connect ellipse, wherein, the attitude parameter that ellipse can represent five institute's tracking targets is connect, that is: the center of target horizontal coordinate in the picture and the angle of vertical coordinate, the major semi-axis of ellipse and the horizontal coordinate positive dirction of minor semi-axis, oval main shaft and image in each.Such as: the major semi-axis of note ellipse is a, and minor semi-axis is b, and the angle of oval main shaft and the horizontal coordinate positive dirction of image is φ.And in tracing process, calculate optimum kernel function window width according to bandwidth matrices.

Further, in above-mentioned steps S101, bandwidth matrices is positive definite symmetric matrices, and there is quantitative relation between bandwidth matrices and described Target ellipse, and namely when the kernel function factor is known, the size and Orientation of target area depends on bandwidth matrices completely.And to probability density function optimizing in tracing process, calculate optimum bandwidth matrix, then by optimum bandwidth matrix application in average drifting track algorithm, to realize the adaptive tracing to tracking target yardstick.Wherein, bandwidth matrices can be used for describing the shape of searching for sample.At two-dimensional space, if bandwidth matrices is positive definite matrix, then the shape of sample areas is planar elliptical; When bandwidth matrices is positive definite symmetric matrices, mean shift vector points to the direction that target function value increases.Traditional average and variance method supposition bandwidth is constant in an iterative process, just selects initial bandwidth.But the situation that sample distribution constantly changes may be there is in reality, if bandwidth does not adjust in good time, the speed of meeting limit algorithm convergence.

Step S102, the testing result according to average and variance track algorithm and particle filter algorithm sets up weighted sum data fusion object localization method.Wherein, average and variance algorithm is actual is carry out optimizing with gradient descent method, once cannot jump out after being absorbed in locally optimal solution.And data fusion method can be seen current solution as and addition of a disturbance variable, makes algorithm jump out current local optimum state.

Step S103, according to weighted sum data fusion object localization method determination unmanned plane target position.Particularly, first obtaining the first center of tracking target based on the average and variance track algorithm of bandwidth matrices, such as, is y _ms, and the second center of tracking target is obtained according to particle prediction, be such as y _pf, then according to the first center (y _ms), the second center (y _pf) and step S101 in bandwidth matrices parameter, set up Strategy of data fusion, finally determine the center of target, such as, be .In other words, namely adaptive bandwidth average and variance algorithm detects a dbjective state, particle filter algorithm also obtains a dbjective state, finally use weighted sum data fusion method to merge two testing results, judge the accurate location of target, whether occur blocking or whether target the situation such as has lost.

Step S104, samples to generate to the particle in particle filter algorithm the particle filter algorithm that based target heavily restrains according to the heavy convergence method of target.Particularly, to sample in the tracking target elliptic region that average and variance track algorithm obtains N number of particle, and centered by each particle, using the state transition equation of second-order autoregressive model as these particles, if after state turns, the region at all particle places is all in same candidate region, then calculate the hsv color proper vector of all particles, and the size of this color feature vector is (10*10+10), and calculate the similarity of each candidate region and object module, such as, be .And get the particle that in particle, weight is directly proportional to similarity and carry out filtering, being weighted summation to filtered all particles, and obtaining and predict by particle filter algorithm the target location obtained, such as, is y _pf.Wherein, in the resampling process of particle, according to weighted sum data fusion object localization method determine institute's tracking target center ( ), and with the center of this target for reference point, the N number of particle of uniform sampling in first area.Wherein, first area i.e. this center ( ) zonule of around certain limit.In an embodiment of the present invention, preferably, N is 20.

In above-mentioned steps S104, in other words, namely in particle filter algorithm, when getting tracking particle in each frame, first all N number of particles are converged to the target's center position obtained by data fusion method .Then with for reference point, the N number of particle of uniform sampling in certain zonule around.This method of sampling is different from traditional particle method for resampling, and it combines with " weighted sum data fusion object localization method ", effectively can avoid the error accumulation in predicting, and prevent sample degeneracy phenomenon.

Step S105, obtains target expanded search strategy according to the particle filter algorithm that target heavily restrains, and follows the tracks of tracking target according to target expanded search strategy.Namely according to target expanded search strategy, that can give loss fast for change follows middle target, and again follows the tracks of.Particularly, first carry out one-step prediction according to the movement locus of tracking target, then localized target search is carried out to tracking target, finally global object search is carried out to tracking target.In other words, namely one-step prediction Target Searching Method search and track target is first used, if one-step prediction Target Searching Method is searched for unsuccessfully, localized target searching method is then used to search for this tracking target, if after the failure of localized target searching method, then global object searching method is used to search for further.

Wherein, one-step prediction Target Searching Method namely: suppose that the image that current process image is is kth frame, the center of tracked target is expressed as successively in sequence image: y ₀, y ₁..., y _k-1, y _k, y _k+1..., then according to the continuity of target travel, the target after loss should at y _knear, then candidate target position y _k+1estimation formulas be: y _k+1-y _k=y _k-y _k-1, get y _k+1the alternatively central point of target, candidate region size, direction are all identical with former frame, extract candidate region proper vector, mate, if matching similarity exceedes first threshold, be then judged as finding target with To Template proper vector; If matching similarity is less than first threshold, then enter the Local Search stage, localized target search is carried out to tracking target.Wherein, first threshold is preset by technician, does not repeat herein.

Localized target searching method comprises: with y _kcentered by, several candidate region of stochastic distribution around it, size, the direction of getting candidate region are identical with former frame, extract each candidate region proper vector to mate with To Template proper vector, and iterative search is carried out in the candidate region selecting similarity to exceed Second Threshold, until find tracking target; If consecutive numbers frame search is less than this tracking target, then enter the global search stage, global object search is carried out to this tracking target.Wherein, the image frames numbers entering Local Search adjusts according to picture-taken frequency, unmanned plane during flying speed and tracking target movement velocity.

Global object searching method specifically comprises: the boundary rectangle getting Target ellipse template, as detection template, carries out traversal search by detection template to image, and detection template is long and wide, such as: the length of note detection template is l, and wide is w.And in search procedure, detection template transverse shifting distance and vertically move distance, such as: note detection template transverse shifting distance is w/2, vertically moves distance l/2.And the matching similarity calculated between To Template and detection template, retain the surveyed area that all similarities are greater than the 3rd threshold value, select the surveyed area alternatively target location that matching similarity is maximum.Wherein, the 3rd threshold value is preset by technician, does not repeat herein.

In above-mentioned steps S105, search element strategy for target expansion, common searching method travels through image, search characteristics region, then carry out characteristic matching with To Template, and the calculated amount traveled through image is very large.Embodiments of the invention divide three levels to search for lost target: one-step prediction is searched for, Local Search and global search.This searching method can improve the specific aim of searching algorithm, improves search efficiency.

As concrete example, describe the unmanned plane target tracking in conjunction with average and variance and particle filter according to the embodiment of the present invention in detail below in conjunction with accompanying drawing 2-10.

Fig. 2 is according to an embodiment of the invention in conjunction with the structural drawing of the unmanned plane target tracking of average and variance and particle filter.

As shown in Figure 2, first sample and transform is carried out to the image sequence of tracked target.Particularly, unmanned plane camera and land station WIFI can be used to complete real time sequence collection to image and transmission.And combine the particle filter tracking algorithm of heavily restraining based on average and variance track algorithm and the based target of bandwidth matrices.Tool, utilizes bandwidth matrices to calculate target window, to adapt to the change of target scale; Introduce particle filter, adopt weight-sum method to merge detection and localization result, make not easily to be absorbed in local optimum state in tracing process.Heavily sampling is restrained to particle, the diversity of particle can be maintained, prevent sample degeneracy, reduce cumulative errors.And in tracing process, judge whether tracking target loses, if target is not lost, then export control command, control objectives object center is positioned at image zone line; If track rejection, then use target expanded search decision search target, lost target can be given for change fast, and after again searching lose objects, export control command.Therefore, embodiments of the invention can realize at dynamic scene, look after change, dimensional variation, the complex situations such as to block under to the real-time localization and tracking of target, have that real-time is good, strong adaptability and an advantage such as extensibility is good.

Fig. 3 is according to an embodiment of the invention in conjunction with bandwidth matrices and the sample elliptical relationship schematic diagram of the unmanned plane target tracking of average and variance and particle filter.

As shown in Figure 3, the quantitative relationship between bandwidth matrices and sample elliptic region, and the bandwidth matrices proposed based on gaussian kernel and correlation parameter initialization rule.Particularly, bandwidth matrices H is positive definite symmetric matrices, can be used for describing the shape of searching for sample.

Positive definite symmetric matrices H can be decomposed into: $H={\mathrm{AA}}^T=\left[\begin{array}{cc}{h}_{11}& h_{12}\\ h_{12}& h_{22}\end{array}\right],$ Wherein A can be expressed as:

$A=\left[\begin{array}{cc}\cos \mathrm{\φ}& -\sin \mathrm{\φ}\\ \sin \mathrm{\φ}& \cos \mathrm{\φ}\end{array}\right]\mathrm{diag}(a,b)$

Can obtain the parameter a in bandwidth matrices H and Target ellipse S, b, φ have following relation:

$\begin{array}{c}a=\sqrt{\frac{1}{2}[h_{11}+h_{22}+\sqrt{{4h}_{12}^2+{(h_{11}-h_{12})}^2}]}\\ b=\sqrt{\frac{1}{2}[h_{11}+h_{22}-\sqrt{{4h}_{12}^2+{(h_{11}-h_{12})}^2}]}\\ \mathrm{\φ}=\frac{1}{2}a\tan 2({2h}_{12},h_{11}-h_{22})\end{array}$

And further, H and kernel function factor sigma determine major semi-axis a, minor semi-axis b and the rotationangleφ of target area S jointly.

Fig. 4 is according to an embodiment of the invention in conjunction with the average and variance track algorithm process flow diagram based on bandwidth matrices of the unmanned plane target tracking of average and variance and particle filter.

As shown in Figure 4, according to an embodiment of the invention in conjunction with the average and variance track algorithm based on bandwidth matrices of the unmanned plane target tracking of average and variance and particle filter, each iteration is carried out in two steps, the first step is average and variance, second step asks optimum bandwidth matrix, and the continuous iteration of algorithm is until convergence.Initial bandwidth is by sample point set S ₀determine, in successive iterations process, automatically upgrade bandwidth, except comprising the position x obtaining sample extreme in Output rusults ₁outward, the bandwidth matrices H describing region shape is also comprised ₁.The method comprises the steps:

Step S401, the oval S of given initial sample ₀, obtain oval two-semiaxle a ₀, b ₀with angle φ ₀.Namely elliptic region S is gathered according to initial sample point ₀with kernel function factor sigma, obtain oval half a ₀, b ₀, angle φ ₀, and initial bandwidth matrix H ₀.

The method that target area initialization is commonly used has two kinds: one to be manually choose target area, and two is automatically detect target area according to priori.Embodiments of the invention end user machine interaction method manually selects tracking target, operator uses mouse to select one piece of rectangular area as tracking target on the image plane, after target area is selected, calculate in rectangle and connect ellipse, initialization is carried out to the parameter in Target ellipse region, obtains target signature point set S:

S={s|(s-s ₀) ^TH ^-1(s-s ₀)<σ ²}(1)

Wherein, s is the pixel in elliptic region S; The center of S is s ₀, rotationangleφ, two halves axial length is respectively σ a and σ b; σ is the factor determined by kernel function; H represents bandwidth matrices.

Wherein, kernel function factor of determination σ is a very important parameter in average and variance algorithm, but average and variance algorithm itself does not have the mechanism of self-adaptative adjustment kernel function bandwidth.In an embodiment of the invention, use standard Gauss kernel function, therefore can get σ ∈ [-3 ,+3] and σ ≈ 3, preferably imitate, get σ=2.5.

Step S402, calculates initial position, according to kernel function factor sigma determination bandwidth matrices H ₀.Particularly, bandwidth matrices H is positive definite symmetric matrices, can be used for describing the shape of searching for sample.

Positive definite symmetric matrices H is to be decomposed into: $H={\mathrm{AA}}^T=\left[\begin{array}{cc}{h}_{11}& h_{12}\\ h_{12}& h_{22}\end{array}\right],$ Wherein A can be expressed as:

$A=\left[\begin{array}{cc}\cos \mathrm{\&phi;}& -\sin \mathrm{\&phi;}\\ \sin \mathrm{\&phi;}& \cos \mathrm{\&phi;}\end{array}\right]\mathrm{diag}(a,b)---\left(2\right)$

Then can obtain the parameter a in bandwidth matrices H and Target ellipse S, b, φ have following relation:

$\begin{array}{c}a=\sqrt{\frac{1}{2}[h_{11}+h_{22}+\sqrt{{4h}_{12}^2+{(h_{11}-h_{12})}^2}]}\\ b=\sqrt{\frac{1}{2}[h_{11}+h_{22}-\sqrt{{4h}_{12}^2+{(h_{11}-h_{12})}^2}]}\\ \mathrm{\&phi;}=\frac{1}{2}a\tan 2({2h}_{12},h_{11}-h_{22})\end{array}---\left(3\right)$

From the formula (1) in step S401, H and kernel function factor sigma determine major semi-axis a, minor semi-axis b and the rotationangleφ of target area S jointly.

After selecting the rectangular area of tracking target, suppose that this region is wide for w1, height is h1, then as shown in Figure 3:

$H_0=\left[\begin{array}{cc}{b}^2& 0\\ 0& a^2\end{array}\right],$ $a_0=\frac{h1}{2\mathrm{\&sigma;}},$ $b_0=\frac{w1}{2\mathrm{\&sigma;}},$ φ ₀=90°(4)

And the parameter of formula (4) is used for determining object module and as iteration initial value.

Step S403, is transformed into HSV space image from rgb space.Particularly, colouring information has unchangeability to the translation of target, rotation and dimensional variation, to block and attitudes vibration insensitive, thus become a key character in target following.For the tracking (such as human body tracking) of non-rigid object, especially applicable color characteristic is followed the tracks of.And histogram is the common method setting up color probability density distribution characteristics.Mathematically image histogram is the function of each gray-scale value statistical property of image and image intensity value, the number of times that in its statistics piece image, each gray level occurs or probability; From figure, it is an X-Y scheme, and horizontal ordinate represents the gray level of each pixel in image, and ordinate is number of times or the probability of the appearance of each each pixel of gray level epigraph.

RGB and hsv color histogram are two kinds of the most frequently used color space model.The digital picture RGB color space of camera acquisition is expressed, but rgb space structure does not also meet the subjective judgement of people to color similarity, hsv color space is closer to the subjective understanding of people to color, and its three components represent respectively: color (Hue), saturation degree (saturation) and brightness (value).Therefore, in one embodiment of the invention, often obtain a two field picture, first image is transformed into HSV space from rgb space, in HSV space, use the even 32*32 joint histogram of H, channel S to add the even 10bin histogram-modeling color characteristic of V passage, use kernel function weighted feature histogram to describe target signature simultaneously.

Step S404, calculates object module particularly, object module is by the probability distribution function of feature space represent.And

${\hat{q}}_u=\underset{s\&Element;S_0}{\mathrm{\&Sigma;}}K\left[{(s-s_0)}^T{H}_0^{-1}(s-s_0)\right]\mathrm{\&delta;}(b\left(s\right)-u)---\left(5\right)$

Wherein, s is any point of target area; U is color component, u=1 ..., m; M is the dimension of HSV histogram space, according to the division of HSV space histogram bins, and m=32*32+10; b (s) is Dirac function; δ () function is the color index value of pixel s in respective histogram space.And it is right further be normalized.

Step S405, input current frame image and previous frame result of calculation: target's center y ₀, sample set S ₀, bandwidth matrices H ₀.

Step S406, performs average and variance y ₁=m (y ₀).I.e. current goal center y ₀, candidate target center y ₁, target's center is by y ₀y is obtained after a step average and variance ₁:

$y_1=\frac{\underset{s\&Element;S}{\mathrm{\&Sigma;}}{G}_H(y-s)\mathrm{\&omega;}\left(s\right)s}{\underset{s\&Element;S}{\mathrm{\&Sigma;}}{G}_H(y-s)s}---\left(6\right)$

Wherein, G (s) is gaussian kernel function, relevant with the derivative of exponential function K (s), and

G(s)=-K′(s)(7)

$\mathrm{\&omega;}\left(s\right)=\underset{u=1}{\overset{m}{\mathrm{\&Sigma;}}}\sqrt{\frac{{\hat{q}}_u}{{\hat{p}}_u\left(y_0\right)}}\mathrm{\&delta;}(b\left(s\right)-u)---\left(8\right)$

Wherein, ω (s) is the weighted value of pixel s.

Step S407, calculated candidate model particularly, the characteristic density of candidate region is distributed as

And it is right further be normalized.

Step S408, calculates the matching similarity of object module and candidate family .The matching similarity of object module and candidate family calculated by Bhattachayya coefficient:

Step S409, judges whether be greater than if so, then perform step S411, otherwise perform step S410.

Step S410, performs backtracking y ₁=0.5 (y ₁+ y ₀).And return continuation execution step S409.

Step S411, according to bandwidth matrices and new center y ₁more new sample point collection S.Namely according to bandwidth matrices H before ₀with new center y ₁more new sample point collection is such as S ₀.

Step S412, bandwidth optimizing, upgrades bandwidth matrices according to kernel function formula.Optimum bandwidth matrix meets the following conditions: through a step average and variance, bandwidth matrices H can make obtain maximum value.Obtain optimum bandwidth matrix H thus _gcomputing formula:

$H_g=\frac{\underset{s\&Element;S}{\mathrm{\&Sigma;}}\mathrm{\&omega;}\left(s\right)(s-y_1){(s-y_1)}^T}{\underset{s\&Element;S}{\mathrm{\&Sigma;}}\mathrm{\&omega;}\left(s\right)}---\left(11\right)$

Step S413, according to new bandwidth matrices and new center y ₁more new sample point collection S.Namely the more new sample point collection S in above-mentioned steps S411 ₀new bandwidth matrices H is asked for according to formula ₁, and by y ₁and H ₁asking for more new sample point collection according to correlation formula, such as, is S ₁.

Step S414, judges whether center and sample point set no longer change.If it is perform step S415, otherwise perform step S406.If namely || y ₁-y ₀|| < ε, S ₀=S ₁, then iteration is stopped, Output rusults y ₁and H ₁, and perform step S415; Otherwise y ₀← y ₁, S ₀← S ₁, H ₀← H ₁, and return execution step S406, enter new round iteration.

Step S415, obtains the long a of oval two halves according to bandwidth matrices and σ ₀, b ₀and rotationangleφ ₀.Particularly, according to H ₁and formula (3) calculates a in above-mentioned steps S402 ₀, b ₀and φ ₀.

Step S416, y _ms=y ₁=y ₀, a=a ₀, b=b ₀, φ=φ ₀.

Step S417, exports y _ms, H ₀, a, b, φ.

Step S408, selects next frame image, and starts to process next frame image.Namely return and perform step S405.

Fig. 5 heavily restrains sampling schematic diagram in conjunction with the target of the unmanned plane target tracking of average and variance and particle filter according to an embodiment of the invention.

As shown in Figure 5, in particle filter algorithm, when getting tracking particle in each frame, first all N number of particles are converged to the target's center position obtained by data fusion method , then with for reference point, the N number of particle of uniform sampling in certain zonule around.These are different from traditional particle method for resampling, and it combines with " weighted sum data fusion object localization method ", effectively can avoid the error accumulation in predicting, and prevent sample degeneracy phenomenon.

Fig. 6 is according to an embodiment of the invention in conjunction with the particle filter tracking algorithm flow chart that the based target of the unmanned plane target tracking of average and variance and particle filter is heavily restrained.

As shown in Figure 6, according to an embodiment of the invention in conjunction with the particle filter tracking algorithm that the based target of the unmanned plane target tracking of average and variance and particle filter is heavily restrained, comprise the following steps:

Step S601, starts.Tracing process starts.

Step S602, setting number of particles, selects motion model.Namely the initial time followed the tracks of, select target model, the simultaneously N number of particle of heart position sampling in the target area, in an embodiment of the invention, preferably, N is 20.

Step S603, gathers piece image to buffer memory.Namely, after determining object module and number of particles, gather a sub-picture of tracking target and be saved in buffer memory.

Step S604, rgb space goes to HSV space.

Does step S605, judge whether to there is target feature vector? namely the particle that whether there is target feature vector in the particle of the position at center, target area is judged.If it is perform step S607, otherwise perform step S606.

Step S606, according to target area, sets up color of object histogram.Even there is not target feature vector, then set up color of object histogram.Particularly, be reduce computation complexity, the color histogram using HSV space (10*10+10) is that each particle sets up object module, namely H, channel S add the even 10bin histogram of V passage for uniform 10*10 joint histogram.

Step S607, particle initialization, weights are set to 1.Namely the initial time followed the tracks of, carries out initialization to N number of particle.Wherein, represent the state value of k moment i-th particle in sequence, represent the observed reading of k moment i-th particle in sequence, i-th particle initial size is taken as the size of target rectangle window, and long is L ⁱ, wide is M ⁱ; Particle weights is set to 1.

Step S608, sets up color probability distribution figure.Namely the histogram of color of object probability distribution is set up.

Step S609, systematic state transfer (prediction).Particularly, systematic state transfer equation adopts second-order autoregressive model, to adapt to the target following of rapid movement.

$y_k^i-y_{k-1}^i=y_{k-1}^i-y_{k-2}^i+v---\left(12\right)$

Wherein, v is the random number of value between [0,1].

The variable upgraded by state transition equation has three: { y ⁱ, L ⁱ, M ⁱ.

Step S610, systematic observation (renewal).The i.e. observation of particle state and the renewal of particle weights.Specifically, namely each particle obtains the candidate target region of particle after state transfer.Extract the object module of candidate region, use Bhattacharyya coefficient calculations candidate family and object module similarity

Definition observation probability density function is:

$p\left(z_k\right|y_k)=\frac{1}{\sqrt{2\mathrm{\&pi;\&sigma;}}}\exp [-\frac{1-{\mathrm{\&zeta;}}_k}{{2\mathrm{\&sigma;}}^2}]---\left(13\right)$

Wherein σ is Gauss's variance, and preferably, getting σ is 0.2.

The weight of particle be expressed as

${\mathrm{\&omega;}}_k^i={\mathrm{\&omega;}}_{k-1}^{i}p\left(z_k\right|y_k^i)---\left(14\right)$

Step S611, particle weighting, calculates posterior probability.Such as: the posterior probability in k moment, namely desired in target following target component { y _k, L _k, M _k, represented by the weighted sum of each particle:

$y_k=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_k^i{y}_k^i,L_k=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_k^i{L}_k^i;M_k=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_k^i{M}_k^i---\left(15\right)$

Export the target-like state value that posterior probability obtains:

y _pf=y _k(16)

Step S612, exports dbjective state.Namely the target-like state value y tried to achieve in above-mentioned steps S611 is exported _pf=y _k.

Step S613, obtains target location according to data fusion method, and use target heavily to restrain method of sampling resampling, weights are set to 1.Namely use weight sum according to fusion method, to particle filtering result y _pfwith the target localization result y that average and variance method obtains _msmerge, obtain target location use the based target N number of particle of particle sampler method resampling of heavily restraining, namely with for reference point, the N number of particle of uniform sampling in certain zonule around, particle weights are 1.

Does step S614, judge to follow the tracks of and terminates? if so, then perform step S615, otherwise return execution step S603.

Step S615, terminates.Namely tracing process terminates.

Fig. 7 is according to an embodiment of the invention in conjunction with the kernel function selecting predictors schematic diagram of the unmanned plane target tracking of average and variance and particle filter.

As shown in Figure 7, be the curve map of an Epanechnikov function.At x, in y-axis in-10 ~ 10 intervals, kernel function has value, and outside above-mentioned interval, the value of kernel function is 0, and therefore kernel function bandwidth σ can be taken as 10.The search window comprising target generally also selects this region, or more slightly smaller than it.And in one embodiment of the invention, use standard Gauss kernel function, therefore can get σ ∈ [-3 ,+3] and σ ≈ 3, preferably, get σ=2.5.

Fig. 8 is according to an embodiment of the invention in conjunction with the process flow diagram of the target tracking algorism in conjunction with average and variance and particle filter of the unmanned plane target tracking of average and variance and particle filter.

As shown in Figure 8, according to an embodiment of the invention in conjunction with the target tracking algorism in conjunction with average and variance and particle filter of the unmanned plane target tracking of average and variance and particle filter, comprise the steps:

Step S801, input piece image.

Step S802, rgb space goes to HSV space.

Does step S803, judge whether to there is target model vectors? if so, then perform step S805 and step S806, otherwise perform step S804.

Step S804, the initialization of Target ellipse region.Select target region, determines target's center position, calculates initial bandwidth matrices, object module, weight, and the N number of particle of sampling in the heart band of position in the target, and to get N be 20, initialization particle weights ω ⁱ=1.

Step S805, uses the average and variance algorithm based on bandwidth matrices.

Step S806, uses the particle filter algorithm that based target is heavily restrained.And this step is combined with step S805 simultaneously.

Step S807, exports y _ms, a, b, φ.Namely target's center position y is obtained according to bandwidth matrices average and variance algorithm _ms, and the major semi-axis a of target area S, minor semi-axis b and rotationangleφ.

Step S808, exports y _pf.Namely heavily restrain particle filter algorithm according to target and obtain target's center position y _pf.

Step S809, carries out data fusion target localization.Namely heavily restrain particle filter algorithm according to Strategy of data fusion in conjunction with bandwidth matrices average and variance algorithm and target, thus determine the position of target average and variance algorithm is avoided to be absorbed in local optimum state.

Does step S810, judge that target is lost? if it is perform step S812, otherwise perform step S811.

Step S811, obtains target's center position and perform step S815.Particularly, if y _mswith y _pfeuclidean distance between 2 is d.D is less, shows that target location that ABMS with PF two kinds of methods obtain is close to consistent.According to parameter a, b and the scale factor σ design weighted sum data anastomosing algorithm of Target ellipse in d and Fig. 3. represent the matching similarity of target area model and candidate family in average and variance method, preferably, setting matching threshold (first threshold) is threshold1=0.8, according to value blending algorithm is divided into three kinds of situations.

(1) if >=threshold1:

(2) if Second Threshold threshold2< <threshold1 (preferably, getting threshold2=0.5), think that dbjective state changes, or target is at least partially obscured:

(3) if 0≤ ≤ threshold2, think that target is blocked completely or target is lost, the forecast function of the particle filter that now places one's entire reliance upon is followed the tracks of target:

$\hat{y}=y_{\mathrm{pf}}---\left(19\right)$

Step S812, uses target expanded search decision search target.Namely, when losing tracking target, target expanded search strategy is used again to search for target.Specifically, when following the tracks of the target in sequence image, if run into the situation that can not confirm target or track rejection, enable target expanded search program, point three levels are searched for lost target;

(1) one-step prediction is carried out according to target trajectory:

Suppose that the image when pre-treatment is kth frame, the center of tracked target is expressed as successively in sequence image: y ₀, y ₁..., y _k-1, y _k, y _k+1..., according to the continuity of target travel, the target after loss should at y _knear, y _k+1estimation formulas be:

y _k+1-y _k=y _k-y _k-1(20)

Get y _k+1the alternatively central point of target, candidate region size, direction are all identical with former frame, extract candidate region proper vector, mate with To Template.If matching similarity exceedes the judgment threshold of setting, be judged as finding target.If matching similarity is lower, then enter the Local Search stage.

(2) local searching strategy:

With y _kcentered by, the several candidate region of stochastic distribution around, size, the direction of getting candidate region are identical with former frame, extract each candidate region feature and mate with To Template.

Select, iterative search is carried out, until find target in the candidate region that similarity exceedes setting threshold value.

If consecutive numbers frame search is less than moving target, then enter the global search stage; The image frames numbers wherein entering search adjusts according to picture-taken frequency, unmanned plane during flying speed and target speed.

(3) global search plan: the boundary rectangle getting Target ellipse template, as detection template, uses this detection template to carry out traversal search to image, and detection template length is l, wide is w.In search procedure, detection template transverse shifting distance is w/2, vertically move distance for l/2, use the matching similarity between Bhattacharyya coefficient calculations To Template and detection template, retain the surveyed area that all similarities are greater than setting threshold value, select the surveyed area alternatively target location that matching similarity is maximum.

Does step S813, judge whether to find target? if it is perform step S815, otherwise perform step S814.

Step S814, the lower width image of input.Namely start to process lower piece image, and return execution step S812.

Step S815, exports target location with target scale parameter, and return execution step S801, lower piece image is processed, carries out the circulation of next round.

Fig. 9 divides schematic diagram in conjunction with the target control area of the unmanned plane target tracking of average and variance and particle filter according to an embodiment of the invention.

As shown in Figure 9, image-region is divided into nine districts, according to tracking target position in the picture, sends corresponding steering order, the center controlling tracking target is positioned at the central area of image, i.e. region five.Below regional is introduced particularly:

5th district: picture centre region.

Firstth district: if target's center is positioned at this district, or be positioned at this district and the trend of oriented image border movement, then control unmanned plane and suitably accelerate, and forward left.

Secondth district: if target's center is positioned at this district, or be positioned at this district and the trend of oriented image border movement, then control unmanned plane and suitably accelerate.

3rd district: if target's center is positioned at this district, or be positioned at this district and the trend of oriented image border movement, then control unmanned plane and suitably accelerate, and forward to the right.

4th district: if target's center is positioned at this district, or be positioned at this district and the trend of oriented image border movement, then control unmanned plane and move to the left.

6th district: if target's center is positioned at this district, or be positioned at this district and the trend of oriented image border movement, then control unmanned plane and move to the right.

SECTOR-SEVEN: if target's center is positioned at this district, or be positioned at this district and the trend of oriented image border movement, then control unmanned plane and suitably slow down, and to left movement.

Section Eight: if target's center is positioned at this district, or be positioned at this district and the trend of oriented image border movement, then control unmanned plane and suitably slow down.

9th district: if target's center is positioned at this district, or be positioned at this district and the trend of oriented image border movement, then control unmanned plane and suitably slow down, and move right.

Figure 10 is according to an embodiment of the invention in conjunction with the unmanned plane target tracker structural drawing of the unmanned plane target tracking of average and variance and particle filter.

As shown in Figure 10, unmanned plane target tracker comprises: four rotor wing unmanned aerial vehicles, WIFI communication and surface work station (Fig. 9).Wherein, unmanned plane primary clustering comprises: a Hokuyo laser range finder, for location estimation with keep away barrier; The UEye monocular-camera of a USB interface, for gathering image; One piece of airborne processor; One piece of AscTec unmanned plane flies to control chip.Because the processing power of airborne processor is limited, therefore, the image transmitting arrived by unmanned plane camera acquisition by WIFI communication is to the enterprising row relax of computing machine at surface work station.Surface work station computing machine takes out image frame by frame, uses track algorithm to detect target location and scale parameter; Finally export unmanned aerial vehicle (UAV) control instruction according to result, control the zone line of tracking target at image.The steering order that computing machine sends carries out wireless transmission by WIFI, received by unmanned aerial vehicle onboard processor, after unmanned aerial vehicle onboard processor confirms the movement instruction received and processes, send to system for flight control computer, control the motion of unmanned plane tracking target.

According to the unmanned plane target tracking in conjunction with average and variance and particle filter of the embodiment of the present invention, can dynamic scene, illumination effect, dimensional variation, block the complex scenes such as interference under, realize the real-time localization and tracking to tracking target, and also can be applicable to different scenes and platform, therefore, embodiments of the invention have that real-time is good, strong adaptability and the advantage such as extensibility is good.

In the description of this instructions, specific features, structure, material or feature that the description of reference term " embodiment ", " some embodiments ", " example ", " concrete example " or " some examples " etc. means to describe in conjunction with this embodiment or example are contained at least one embodiment of the present invention or example.In this manual, identical embodiment or example are not necessarily referred to the schematic representation of above-mentioned term.And the specific features of description, structure, material or feature can combine in an appropriate manner in any one or more embodiment or example.

Although illustrate and describe embodiments of the invention, those having ordinary skill in the art will appreciate that: can carry out multiple change, amendment, replacement and modification to these embodiments when not departing from principle of the present invention and aim, scope of the present invention is by claim and equivalency thereof.

Claims (7)

1., in conjunction with a unmanned plane target tracking for average and variance and particle filter, it is characterized in that, comprise the following steps:

In average and variance tracing process, build the average and variance track algorithm based on described bandwidth matrices, with adaptive updates target scale window in tracing process according to bandwidth matrices;

Testing result according to described average and variance track algorithm and particle filter algorithm sets up weighted sum data fusion object localization method;

Described unmanned plane target position is determined according to described weighted sum data fusion object localization method;

Sample to generate the particle filter algorithm of heavily restraining based on described target to the particle in described particle filter algorithm according to the heavy convergence method of target; And

Obtain target expanded search strategy according to the particle filter algorithm that described target heavily restrains, and follow the tracks of target according to described target expanded search strategy, described tracking target according to described target expanded search strategy comprises further:

Movement locus according to described target carries out one-step prediction;

Localized target search is carried out to described target;

Global object search is carried out to described target;

Wherein, described target expanded search strategy comprises:

Use one-step prediction Target Searching Method search target, specifically comprise:

If when the image of pre-treatment is kth frame, the center of described target is expressed as successively in sequence image: y ₀, y ₁..., y _k-1, y _k, y _k+1..., candidate target position y _k+1estimation formulas be: y _k+1-y _k=y _k-y _k-1;

Get y _k+1as the central point of described candidate target, candidate region size, direction are all identical with former frame, extract candidate region proper vector, mate with To Template proper vector;

If matching similarity exceedes first threshold, be then judged as finding target;

If matching similarity is less than first threshold, then enter the Local Search stage;

If described one-step prediction Target Searching Method is searched for unsuccessfully, localized target searching method is used to search for described target;

If after described localized target searching method failure, then global object searching method is used to search for.

2. the unmanned plane target tracking in conjunction with average and variance and particle filter according to claim 1, is characterized in that, builds the average and variance track algorithm based on described bandwidth matrices, comprise further according to bandwidth matrices:

Describe object module and candidate family according to the weighted probability density distribution function based on color characteristic, wherein, the feature space of described object module and described candidate family is the one dimension hsv color space vector of 32*32+10;

Elliptic region is utilized to represent described target, according to the object initialization way selection rectangle of man-machine interaction as described target, and calculate in described rectangle and connect ellipse, wherein, the center connecing target described in ellipse representation in each horizontal coordinate in the picture and vertical coordinate, the major semi-axis of described ellipse and the angle of minor semi-axis, the main shaft of described ellipse and the horizontal coordinate positive dirction of described image;

In described tracing process, optimum kernel function window width is calculated according to described bandwidth matrices.

3. the unmanned plane target tracking in conjunction with average and variance and particle filter according to claim 1, is characterized in that, determine described unmanned plane target position, comprise further according to described weighted sum data fusion object localization method:

Average and variance track algorithm based on described bandwidth matrices obtains the first center of described target;

The second center of described target is obtained according to particle prediction;

Set up Strategy of data fusion according to described first center, described second center and bandwidth matrices parameter, determine the center of described target.

4. the unmanned plane target tracking in conjunction with average and variance and particle filter according to claim 1, is characterized in that, generate the particle filter algorithm of heavily restraining based on described target, comprise further:

To sample in the tracking target elliptic region that described average and variance track algorithm obtains N number of particle;

Centered by each particle, using the state transition equation of second-order autoregressive model as described particle;

If the region at all particle places is same candidate region after state transfer, then calculate the hsv color proper vector of all particles, and described proper vector size is (10*10+10);

Calculate the similarity of each candidate region and object module, the particle that the weight of getting described particle is directly proportional to similarity carries out filtering;

Summation is weighted to filtered all particles, and acquisition predicts by particle filter algorithm the target location obtained;

In particle resampling process, determine the center of described target according to described weighted sum data fusion object localization method, and with described center for reference point, the N number of particle of uniform sampling in first area.

5. the unmanned plane target tracking in conjunction with average and variance and particle filter according to claim 2, is characterized in that,

Described bandwidth matrices is positive definite symmetric matrices, and there is quantitative relation between described bandwidth matrices and Target ellipse;

To probability density function optimizing, calculate optimum bandwidth matrix;

By described optimum bandwidth matrix application in average drifting track algorithm, to realize the adaptive tracing to target scale.

6. the unmanned plane target tracking in conjunction with average and variance and particle filter according to claim 1, it is characterized in that, described localized target searching method comprises:

With y _kcentered by, the several candidate region of stochastic distribution around, size, the direction of getting described candidate region are identical with former frame, extract each candidate region proper vector and mate with To Template proper vector;

Iterative search is carried out, until find described target in the candidate region selecting similarity to exceed Second Threshold;

If consecutive numbers frame search is less than described target, then enter the global search stage;

The image frames numbers entering Local Search adjusts according to picture-taken frequency, unmanned plane during flying speed and described target speed.

7. the unmanned plane target tracking in conjunction with average and variance and particle filter according to claim 1, it is characterized in that, described global object searching method comprises:

The boundary rectangle getting Target ellipse template, as detection template, carries out traversal search by described detection template to described image;

Detect described detection template long and wide, and detect described detection template transverse shifting distance and vertically move distance;

Calculate the matching similarity between To Template and described detection template;

Retain the surveyed area that all similarities are greater than the 3rd threshold value, and the surveyed area selecting matching similarity maximum is as described candidate target position.