CN104408401A - Time-sensitive object in-orbit detection method - Google Patents

Abstract

The invention discloses a time-sensitive object in-orbit detection method. The method comprises the following steps: step S1, selecting various time-sensitive object training areas on historical images, extracting high-dimensional multiple-dimensioned concentric circular cluster direction gradient characteristics from each pixel of each training image, and learning structure dictionaries of various time-sensitive objects in an offline mode; step S2, extracting the high-dimensional multiple-dimensioned concentric circular cluster direction gradient characteristics at each pixel on the image of each current in-orbit time phase, by use of the structure dictionaries, solving time-sensitive object type indication vectors, according to the structural sparse characteristics of the time-sensitive object type indication vectors, identifying positions of suspicious objects and types of the suspicious objects, and extracting suspicious object areas; step S3, analyzing the loci of the suspicious object areas detected on the in-orbit images of the different time phases, and according to the singularity of motion loci, identifying the time-sensitive objects in an in-orbit mode; and step S4, taking the images of the in-orbit time-sensitive objects as training images of the structure dictionaries for in-orbit increment updating, and returning to step S1.

Description

A kind of In-flight measurement method of time critical target

Technical field

The present invention relates to a kind of In-flight measurement method of the technical fields, particularly time critical target such as in-orbit imaging processing, target detection, target identification, target monitoring.

Background technology

Compared with common target, time critical target has very strong ageing, and time critical target must identify in limited time window, transient.Meanwhile, time critical target is often all very important target, once lose the chance of identification, will cause heavy losses.Therefore, the detection and Identification of time critical target have important Research Significance, but have more challenge simultaneously.

Along with the development of high spatial resolution, high time resolution remote sensing satellite, utilize satellite image In-flight measurement and identify that time critical target becomes possibility.Compared with other data acquisition means, satellite image scope is large, is conducive to carrying out accurately, for a long time following the tracks of to time critical target.

The difficult point that time critical target detects mainly is the complicacy of time critical target, and time critical target only just presents the feature of time critical target when the change of certain time point generation state or Trajectory Catastrophe, and the time point of this key is difficult to be caught in.For the In-flight measurement of time critical target, seldom, the object module how utilizing latest data automatically to adjust off-line state training is the key that time critical target detects for utilizable priori and data.But above-mentioned gordian technique is very immature at present, limits the practical application of time critical target on-line checkingi.

Summary of the invention

The object of the invention is the demand for the feature processed in-orbit and practical application, a kind of In-flight measurement method of effective time critical target is provided.

To achieve these goals, the In-flight measurement method of time critical target of the present invention, it is as follows that the method comprising the steps of:

Step S1: choose various time critical target training area on history image, the multiple dimensioned donut bunch direction gradient feature of higher-dimension is extracted, the structure dictionary of all kinds of time critical target of off-line learning at each pixel place of every width training image of the time critical target of every type;

Step S2: each pixel place on the image of current each phase in-orbit extracts the multiple dimensioned donut bunch direction gradient feature of higher-dimension, structure dictionary is utilized to solve the instant quick target type instruction vector of projection coefficient of multiple dimensioned donut bunch direction gradient feature, according to the position of structure sparse characteristic identification suspicious object and the type of suspicious object of time critical target type instruction vector, the similarity according to time critical target type instruction vector extracts target area suspicious;

Step S3: analyze the track of the target area suspicious detected on the image in-orbit of different phase, the singularity according to movement locus identifies time critical target in-orbit;

Step S4: using the training image of the image of time critical target in-orbit as structure dictionary incremental update in-orbit, return step S1.

The method of the invention for improving the universality of time critical target In-flight measurement, automaticity has great importance, its major advantage is as follows:

Historic training data and current latest data combine by the present invention, are embodied by the prior-constrained of time critical target by historic training data, have ensured in the processing environment in-orbit of intervening at nobody and demand and data characteristics well can have been combined; The feature of time critical target historical data comprised and the new feature of present image combine, by dictionary in-orbit incremental update improve dictionary sign ability and greatly saved calculated amount.

The present invention utilizes the target type belonging to time critical target type instruction vector representation pixel in the target detection stage, overcomes the uncertainty of scalar method for expressing; Extract target area according to the similarity of time critical target type instruction vector between pixel, improve noise and the robustness of blocking.

The present invention utilizes the distance metric of the generalized eigenvalue based on covariance matrix of target area to have good robustness to visual angle change in the motion state abnormality detection stage, decrease the false alarm rate of time critical target identification; In the space-time track abnormality detection stage by space-time trail change Curve transform to polar coordinate space, effectively feature the kinematic singularity of time critical target, improve the accuracy rate of time critical target identification.

Have benefited from above-mentioned advantage, the present invention makes the In-flight measurement of time critical target become possibility, drastically increases that time critical target detects, ageing, the robustness that identifies and automaticity, can be widely used in time critical target and find with the system such as monitoring, target monitoring.

Accompanying drawing explanation

Fig. 1 is the In-flight measurement method flow diagram of a kind of time critical target of the present invention.

Fig. 2 is space-time track abnormality detection figure.

Embodiment

Below in conjunction with accompanying drawing, technical matters involved in technical solution of the present invention is described.Be to be noted that described embodiment is only intended to be convenient to the understanding of the present invention, and any restriction effect is not play to it.

As Fig. 1, to illustrate that the present invention proposes a kind of In-flight measurement method performing step of time critical target as follows:

Step S1: choose various time critical target training area on history image, the multiple dimensioned donut bunch direction gradient feature of higher-dimension is extracted, the structure dictionary of all kinds of time critical target of off-line learning at each pixel place of every width training image of the time critical target of every type.

Described multiple dimensioned donut bunch direction gradient feature centered by sampled point, with sampling scale be radius image block on sample and construct the concentric circles loop configuration of 3 different radiis, corresponding sampling spot is positioned on the donut of above-mentioned different radii, on each donut by 45 ° angularly interval extract 8 sampling spots, have identical Gauss's scale-value with the sampling spot on Radius, the sampling spot Gauss scale-value on different radii is different.The detailed process of described multiple dimensioned donut bunch direction gradient feature extraction is as follows:

Step S01: to calculate centered by sampled point, with 8 direction gradients of each pixel (u, v) of the sampling scale ∑ image block that is radius, then, obtain the direction gradient proper vector h at (u, v) place by gaussian kernel convolution _Σ(u, v) represents as follows:

$h_{\mathrm{\Σ}}(u,v)={\left[(g_1^{\mathrm{\Σ}}(u,v),g_2^{\mathrm{\Σ}}(u,v),...,g_m^{\mathrm{\Σ}}(u,v),...,g_8^{\mathrm{\Σ}}(u,v))\right]}^T---\left(1\right)$

U and v is respectively line number and the row number of pixel, and T represents the transposition of vector, represent that the gradient vector that m direction gradient gaussian kernel convolution obtains, m are direction numbering, m=1,2 ..., 8.

Step S02: multiple dimensioned donut bunch direction gradient feature D (u, v) is the union of a series of associated vector describing each position in sampling spot (u, v) local support region, and the representation of D (u, v) is as follows:

$D(u,v)=\left\{h_{{\mathrm{\Σ}}_n}\left(l_{m2}(u,v,R_{n2})\right)\right|n=\mathrm{1,2,3};m2=\mathrm{1,2},...,8;n2=\mathrm{1,2,3}.\}---\left(2\right)$

Wherein, l _m2(u, v, R _n2) represent pixel (u, v) the n-th 2 donuts on the coordinate of m2 sampling spot, the local direction histogram of gradients of m2 sampling spot on the n-th 2 donuts representing pixel (u, v), n is sampling scale sequence number, and n2 is donut sequence number, and m2 is sampling spot sequence number.

The study of described structure dictionary be from the target type numbering learning low-dimensional of the multiple dimensioned donut bunch direction gradient proper vector set of higher-dimension and correspondence, dictionary that separability is good.If image for the i-th width training image of target type j, number of pixels be N _i, then from image n can be extracted _iindividual multiple dimensioned donut bunch direction gradient proper vector, this N _ithe union of individual proper vector is as the feature of target type j.For convenience of describing, the set of jth classification target training characteristics is designated as X _j={ x _k={ j, f _k| 1≤k≤A _j, x _k={ j, f _kan expression kth training sample wherein, j is target type numbering, f _kfor the multiple dimensioned donut bunch direction gradient proper vector that a kth training sample is corresponding, A _jrepresent element number in the set of jth class target training characteristics.Structure dictionary learning model of the present invention is as follows:

$<D,Z>=\arg \underset{\text{D,Z}}{\min }{\left|\right|X-\mathrm{DZ}\left|\right|}_F^2+{\mathrm{\&lambda;}}_1{\left|\right|Z\left|\right|}_1^2+{\mathrm{\&lambda;}}_2\underset{i1,i2}{\mathrm{\&Sigma;}}{\left|\right|z_{i1}-z_{i2}\left|\right|}_2^2{w}_{i1,i2}---\left(3\right)$

Wherein, matrix X is the multiple dimensioned donut bunch direction gradient proper vector set that all training images obtain, and the dimension of matrix X is the capable N row of M, and M is the dimension of multiple dimensioned donut bunch direction gradient proper vector, and N is the number of all training samples.Matrix D is structure dictionary, and the dimension of structure dictionary D is the capable K row of M, and each row of structure dictionary D are called a dictionary atom, and K is structure dictionary atom number.Matrix for the projection coefficient matrix that matrix X utilizes structure dictionary to try to achieve, T represents vector or transpose of a matrix.K dimensional vector z _athe a row of representing matrix Z, 1≤a≤N.|| || _f, || || ₁with || || ₂the Frobenius norm of representing matrix, 1 norm and 2 norms, λ ₁and λ ₂for regularization coefficient, control degree of rarefication and the separability of projection coefficient respectively.W _{i1, i2}represent training sample z _i1and z _i2similar weight, if z _i1and z _i2for the projection coefficient of the training sample of same type target, then w _{i1, i2}=1, if z _i1and z _i2for the projection coefficient of the training sample of not same type target, then w _{i1, i2}=0.

It is as follows that described structure dictionary learning model solves detailed process:

Step S11 initial value sets.Setting regularization coefficient λ ₁and λ ₂, λ in the present invention ₁=λ ₂=0.01.Carry out principal component analysis (PCA) to a multiple dimensioned donut bunch direction gradient characteristic set for the target of every type to try to achieve and notable feature value characteristic of correspondence vector, the union of dissimilar clarification of objective vector is as the initial value D of structure dictionary D ⁽⁰⁾, the initial value Z of projection coefficient Z ⁽⁰⁾=([D ⁽⁰⁾] ^td ⁽⁰⁾) ^-1[D ⁽⁰⁾] ^tx.Notable feature value exceedes L eigenwert before all eigenwert energy 90% after referring to and carrying out descending sort to eigenwert.All notable feature value numbers are dictionary atom number K.

The alternating iteration of step S12 structure dictionary and projection coefficient matrix upgrades.Make D ^(t) and Z ^(t)for solution when structure dictionary D and projection coefficient matrix Z the t time iteration, according to following formula, alternating iteration renewal is carried out to structure dictionary and projection coefficient matrix:

${D_{r,c}}^{(t+1)}={D_{r,c}}^{\left(t\right)}\frac{{\left(X{\left[Z^{\left(t\right)}\right]}^T\right)}_{r,c}}{{\left(D^{\left(t\right)}{Z}^{\left(t\right)}{\left[Z^{\left(t\right)}\right]}^T\right)}_{r,c}}$

${Z_{r,c}}^{(t+1)}={Z_{r,c}}^{\left(t\right)}\frac{{({\left[D^{(t+1)}\right]}^{T}X+{\mathrm{\&lambda;}}_1{Z}^{\left(t\right)}W)}_{r,c}}{{(Z^{\left(t\right)}{\left[D^{(t+1)}\right]}^T{Z}^{\left(t\right)}+{\mathrm{\&lambda;}}_1{Z}^{\left(t\right)}S+{\mathrm{\&lambda;}}_2)}_{r,c}}---\left(4\right)$

Wherein, S is diagonal matrix, the element on diagonal line the value of the capable l row of b of matrix W is w _{b, 1}, 1≤b≤N, 1≤l≤N.D _{r, c} ^(t+1)and D _{r, c} ^(t)the value of the capable c row of r of solution when representing (t+1) of structure dictionary secondary and the t time iteration respectively; Z _{r, c} ^(t)represent Z ^(t)the value of r capable c row, r and c is line number and the row number of matrix, 1≤r≤M, 1≤c≤K.[Z ^(t)] ^trepresenting matrix Z ^(t)transposition.It is t < 100 or mse (D that alternating iteration upgrades the criterion stopped ^(t+1)z ^(t+1), D ^(t)z ^(t)) < ε, mse (D ^(t+1)z ^(t+1), D ^(t)z ^(t)) representing the square error of adjacent twice iteration, ε is a threshold value, ε=0.1 in the present invention.

Step S2: each pixel place on the image of current each phase in-orbit extracts the multiple dimensioned donut bunch direction gradient feature of higher-dimension, structure dictionary is utilized to solve the instant quick target type instruction vector of projection coefficient of multiple dimensioned donut bunch direction gradient feature, according to the position of structure sparse characteristic identification suspicious object and the type of suspicious object of time critical target type instruction vector, similarity according to time critical target type instruction vector extracts target area suspicious, and detailed process is as follows:

Step S21 builds new projection matrix according to structure dictionary, the multiple dimensioned donut of each pixel of the image of current each a phase in-orbit bunch direction gradient proper vector is projected, obtain the time critical target type instruction vector of each pixel, detailed process is as follows:

If F _{r2, c2}for the multiple dimensioned donut bunch direction gradient proper vector of the capable c2 row of r2 on present image, then F _{r2, c2}corresponding time critical target type target type instruction vector is CE _{r2, c2}=PF _{r2, c2}, CE _{r2, c2}for K dimensional vector.Wherein P=(DTD+ λ I) ^-1d ^t, λ is correction factor, λ=0.1 of the present invention.I is unit matrix.

Step S22 determines suspicious object according to the structure of the time critical target type instruction vector of the type information of the target embodied belonging to pixel is openness, and detailed process is as follows:

If the notable feature number of jth type target is K _j, then corresponding in the dictionary obtained K _jwhat individual dictionary atom represented is jth type clarification of objective.Correspondingly, time critical target type indicates vectorial CE _{r2, c2}in contain the type information of target.Concrete, vectorial Ce _{r2, c2}in the K of correspondence _jthe energy of section component and degree of rarefication reflect the probability that this target belongs to jth type target.If Ce _{r2, c2}k _jthe energy of section component is maximum and the component of other section is very sparse, then represent that on present image, r2 capable c2 row pixel is a part for jth type target.If energy and the degree of rarefication of each section of component are more or less the same, then represent that this pixel is background area.Compared with traditional scalar formula target type method for expressing, the target type method for expressing robust more of this vector mode.

Step S23 has the Image Segmentation Using of similar target type instruction vector to each phase in-orbit according to the pixel being positioned at the same area, extracts target area suspicious.Iamge Segmentation of the present invention is the partitioning algorithm based on graph theory.For this reason, non-directed graph G=(V is first built; E), a vertex v i in each pixel pi and the V in image is corresponding, and V is the vertex set of non-directed graph G, and E is the set on the limit of non-directed graph G; Between the time critical target type instruction vector that the weights d ((vi, vj)) on limit (vi, vj) is pixel p i and pj difference.The step that target area extraction carries out the detailed process of merge and split to non-directed graph G is exactly as follows:

S231: initialization.E is carried out ascending order arrangement by weights, obtains orderly limit set π=(o ₁..., o _m3), limit o ₁weight minimum, limit p _m3weight maximum; Calculate initial segmentation S ⁰, each region only comprises the number that vertex v i, a m3 are limit in the set E on the limit of described non-directed graph G.

S232: make S ^q-1and S ^qrepresent the segmentation result comprising q-1 limit and q limit respectively, q represents that the sequence number on limit in π is gathered on orderly limit, 1≤q≤m3.To q (1≤q≤m3), perform and operate as follows: given S ^q-1with the vertex v i, the vj that are connected its q article of limit, order with be respectively S ^q-1in comprise the regional ensemble of vi and vj, construct S as follows ^qif: region and weights $d\left((\mathrm{vi},\mathrm{vj})\right)\&le;\mathrm{MInt}(C_{\mathrm{vi}}^{q-1},C_{\mathrm{vj}}^{q-1}),$ Then combined region and region if region or weights $d\left((\mathrm{vi},\mathrm{vj})\right)>\mathrm{MInt}(C_{\mathrm{vi}}^{q-1},C_{\mathrm{vj}}^{q-1}),$ Then S ^q=S ^q-1; Wherein:

$\mathrm{MInt}(C_{\mathrm{vi}}^{q-1},C_{\mathrm{vj}}^{q-1})=\min (\mathrm{Int}\left(C_{\mathrm{vi}}^{q-1}\right)+\mathrm{\&tau;}\left(C_{\mathrm{vi}}^{q-1}\right),\mathrm{Int}\left(C_{\mathrm{vj}}^{q-1}\right)+\mathrm{\&tau;}\left(C_{\mathrm{vj}}^{q-1}\right))$ Represent with minimum value, function control area is gathered global similarity, in the present invention represent region the number of middle pixel, tt=vi or tt=vj.The value of α reflects observing result, α=200 in the present invention. represent region the inside diversity represented based on minimum spanning tree weight limit, namely represent region based on minimum spanning tree, d (e) represents the weight of minimum spanning tree top.

The final segmentation result of S233 is S=S ^m3.

Step S3: analyze the track of the target area suspicious detected on the image in-orbit of different phase, the singularity according to movement locus identifies time critical target in-orbit.The singularity of movement locus is mainly manifested in the following two kinds situation: the exception of motion state abnormal (the violent change of the unexpected appearance of suspicious object or disappearance, outward appearance), space-time track.The present invention identifies time critical target according to these two kinds of singularitys.The identification detailed process in-orbit of described time critical target is as follows:

Step S31: utilize the time critical target type of target area suspicious to indicate the distance difference of the covariance matrix of vector set to be that the searching of each suspicious object is closing on the arest neighbors in moment, if the covariance matrix between bunch direction gradient proper vector set of the multiple dimensioned donut between arest neighbors region is still arest neighbors, then represent that the motion state no exceptions of this target is non-time critical target; If the covariance matrix between bunch direction gradient proper vector set of the multiple dimensioned donut between arest neighbors region is not arest neighbors, then the motion state of this target occurs extremely to be time critical target, and detailed process is as follows:

Be located at two adjacent moment t ₁and t ₂image on the number of jth type target that detects be respectively with with represent t respectively _cthe multiple dimensioned donut bunch direction gradient character vector set of β the target of (c=1,2) moment jth type and the covariance matrix of time critical target type instruction vector set, if time critical target type instruction vector covariance matrix in covariance matrix set in arest neighbors and time next-door neighbour be respectively with if $\mathrm{dis}(x_{j,\mathrm{\&beta;}}^{t_1},x_{j,p}^{t_2})/\mathrm{dis}(x_{j,\mathrm{\&beta;}}^{t_1},x_{j,q}^{t_2})>{\mathrm{\&tau;}}_1,$ Then represent the dbjective state represented or outward appearance change, and this suspicious object is judged to time critical target; If then represent the target represented is non-time critical target.In the present invention, τ ₁=0.9, τ ₂=0.6.For two covariance matrix A and B, δ _ηη the generalized eigenvalue of (A, B) representing matrix A and B, μ represents the number of the generalized eigenvalue of covariance matrix A and B.

Step S32: by the space-time trail change curve projection of non-for different phase time critical target under polar coordinates, the direction change according to space-time track adjacent in polar coordinates identifies time critical target.Detailed process is as follows:

If the number of the non-time critical target obtained on current multi-temporal image in-orbit according to step S31 is N1, γ target wherein in the center-of-mass coordinate in g moment is ζ is the number of current multi-temporal image in-orbit, then its space-time trail change curve is expressed as:

$({\mathrm{ZX}}_{\mathrm{\&gamma;}}^2-{\mathrm{ZX}}_{\mathrm{\&gamma;}}^1,{\mathrm{ZY}}_{\mathrm{\&gamma;}}^2-{\mathrm{ZY}}_{\mathrm{\&gamma;}}^1),({\mathrm{ZX}}_{\mathrm{\&gamma;}}^3-{\mathrm{ZX}}_{\mathrm{\&gamma;}}^3-{\mathrm{ZY}}_{\mathrm{\&gamma;}}^3-{\mathrm{ZY}}_{\mathrm{\&gamma;}}^2),...,({\mathrm{ZX}}_{\mathrm{\&gamma;}}^{\mathrm{\&zeta;}}-{\mathrm{ZX}}_{\mathrm{\&gamma;}}^{\mathrm{\&zeta;}-1},{\mathrm{ZY}}_{\mathrm{\&gamma;}}^{\mathrm{\&zeta;}}-{\mathrm{ZY}}_{\mathrm{\&gamma;}}^{\mathrm{\&zeta;}-1})\left(5\right)$

Space-time track is abnormal is usually expressed as direction of motion sudden change, and its space-time trail change curve is converted into polar form { (ρ by the present invention for this reason ², θ ²), (ρ ³, θ ³) ..., (ρ ^ζ, θ ^ζ), wherein ${\mathrm{\&rho;}}^{k1}=\sqrt{{({\mathrm{ZX}}_{\mathrm{\&gamma;}}^{k1}-{\mathrm{ZX}}_{\mathrm{\&gamma;}}^{k1-1})}^2+{({\mathrm{ZY}}_{\mathrm{\&gamma;}}^{k1}-{\mathrm{ZY}}_{\mathrm{\&gamma;}}^{k1-1})}^2}$ Represent the speed of 2≤k1≤ζ moment space-time trail change, represent the direction of k1 moment space-time trail change.

If θ ^k1> pi/2, shows that this suspicious object is the opposite direction of former direction of motion in the direction of motion in kth 1 moment, and space-time track is abnormal, and this suspicious object is time critical target.Fig. 2 is space-time track abnormality detection figure, and wherein, the target that variation track a is corresponding is normal target, and the target that variation track b is corresponding is time critical target.

Step S4: using the training image of the image of time critical target in-orbit as structure dictionary incremental update in-orbit, return step S1.Detailed process is as follows:

Step S41: when training sample is more is meet the ageing requirement that processes in-orbit, using the multiple dimensioned donut bunch direction gradient feature of the time critical target in-orbit that detected and suspicious object as new training sample.

Step S42: according to the training image of the structure dictionary incremental update in-orbit that present image, current time critical target obtain, utilize new training sample, to the structure dictionary incremental update in-orbit of previous moment.The dictionary learning method of structure dictionary increment updating method and step S1 is in-orbit similar, and difference is the setting of initial dictionary.The initial dictionary of step S11 comes from the base vector of principal component analytical method, and the initial dictionary of this step is from existing dictionary.Dictionary after renewal combines latest image characteristic sum time critical target feature, has stronger sign ability.

The above; be only the embodiment in the present invention; but protection scope of the present invention is not limited thereto; any people being familiar with this technology is in the technical scope disclosed by the present invention; the conversion or replacement expected can be understood; all should be encompassed within protection scope of the present invention, therefore, protection scope of the present invention should be as the criterion with the protection domain of claims.

Claims (10)

1. an In-flight measurement method for time critical target, it is as follows that the method comprising the steps of:

Step S1: choose various time critical target training area on history image, the multiple dimensioned donut bunch direction gradient feature of higher-dimension is extracted, the structure dictionary of all kinds of time critical target of off-line learning at each pixel place of every width training image of the time critical target of every type;

Step S2: each pixel place on the image of current each phase in-orbit extracts the multiple dimensioned donut bunch direction gradient feature of higher-dimension, structure dictionary is utilized to solve the instant quick target type instruction vector of projection coefficient of multiple dimensioned donut bunch direction gradient feature, according to the position of structure sparse characteristic identification suspicious object and the type of suspicious object of time critical target type instruction vector, the similarity according to time critical target type instruction vector extracts target area suspicious;

Step S3: analyze the track of the target area suspicious detected on the image in-orbit of different phase, the singularity according to movement locus identifies time critical target in-orbit;

Step S4: using the training image of the image of time critical target in-orbit as structure dictionary incremental update in-orbit, return step S1.

2. method according to claim 1, it is characterized in that, described structure dictionary is the multiple dimensioned donut bunch direction gradient feature of the low-dimensional obtained from the multiple dimensioned donut bunch direction gradient proper vector set of higher-dimension and the time critical target type number learning of correspondence.

3. method according to claim 2, is characterized in that, indicates the similarity between vector to build the learning model of structure dictionary according to the openness and different pixels place time critical target type of time critical target type instruction vector.

4. method according to claim 2, it is characterized in that, principal component analysis (PCA) is carried out to a multiple dimensioned donut bunch direction gradient characteristic set for all types of time critical target, to obtain and using the union with notable feature value characteristic of correspondence vector as initial dictionary, alternating iteration upgrades initial dictionary and projection coefficient again, obtains structure dictionary.

5. method according to claim 1, is characterized in that, the step of described extraction target area suspicious comprises as follows:

Step S21: build new projection matrix according to structure dictionary, projects the multiple dimensioned donut of each pixel of the image of current each a phase in-orbit bunch direction gradient proper vector, obtains the time critical target type instruction vector of each pixel;

Step S22: determine suspicious object according to the structure of the time critical target type instruction vector of the type information of the target embodied belonging to pixel is openness;

Step S23: according to the pixel being positioned at the same area, there is the Image Segmentation Using of similar target type instruction vector to each phase in-orbit, extract target area suspicious.

6. method according to claim 1, is characterized in that, the described time critical target step that identifies in-orbit comprises as follows:

Step S31: utilize the time critical target type of target area suspicious to indicate the distance difference of the covariance matrix of vector set to be that the searching of each suspicious object is closing on the arest neighbors in moment, if the covariance matrix between bunch direction gradient proper vector set of the multiple dimensioned donut between arest neighbors region is still arest neighbors, then represent that the motion state no exceptions of this target is non-time critical target; If the covariance matrix between bunch direction gradient proper vector set of the multiple dimensioned donut between arest neighbors region is not arest neighbors, then the motion state of this target occurs extremely to be time critical target;

Step S32: by the space-time trail change curve projection of non-for different phase time critical target under polar coordinates, the direction change according to space-time track adjacent in polar coordinates identifies time critical target.

7. method according to claim 6, is characterized in that, the distance difference had between described target area suspicious, utilizes the time critical target type of target area suspicious to indicate the difference between the covariance matrix of vector set to carry out metric range difference; Utilize the quadratic sum of the generalized eigenvalue of covariance matrix to measure the distance difference between described covariance matrix.

8. method according to claim 6, it is characterized in that, the distance difference had between described target area suspicious, utilizes the difference between the covariance matrix of the multiple dimensioned donut of target area suspicious bunch direction gradient proper vector set to carry out metric range difference; Utilize the quadratic sum of the generalized eigenvalue of covariance matrix to measure the distance difference between described covariance matrix.

9. method according to claim 6, it is characterized in that, the described time critical target of identification is in-orbit under polar coordinates by space-time trail change curve projection, the velocity variations of non-for different phase time critical target and direction are changed and is separated, obtain the curve of space-time trajectory direction change, for better describing the abnormality of space-time track.

10. method according to claim 1, is characterized in that, the step of described structure dictionary incremental update training image in-orbit comprises as follows:

Step S41: when training sample is more is meet the ageing requirement that processes in-orbit, using the multiple dimensioned donut bunch direction gradient feature of the time critical target in-orbit that detected and suspicious object as new training sample;

Step S42: according to the training image of the structure dictionary incremental update in-orbit that present image, current time critical target obtain, utilize new training sample, to the structure dictionary incremental update in-orbit of previous moment.