CN105117720A - Object scale self-adaption tracking method based on spatial-temporal model - Google Patents

Abstract

The invention discloses an object scale self-adaption tracking method based on a spatial-temporal model, and the method comprises the following steps: beginning a video, reading in a first frame of image, manually assigning the rectangle position of a tracked object; then based on context airspace, initializing the spatial-temporal model and a multiscale history object template library; next reading in a next frame of image, building the spatial-temporal model by iteration, calculating a confidence map, and estimating an object center position; then according to the history object template library, judging a template optimal scale, determining the rectangle position of the object, completing tracking of the current frame of object, and updating scale parameters of the spatial-temporal model and the multiscale history object template library; finally detecting whether the video is finished or not, continuously reading in a next frame if the video is not finished, or completing tracking. By adopting the object scale self-adaption tracking method, the object appearance scale change is effectively coped with under the conditions of illumination change, partial blocking and swift moving, and the robust tracking is realized.

Description

Based on the target scale adaptive tracking method of space-time model

Technical field:

The invention belongs to field of machine vision, particularly a kind of target scale adaptive tracking method based on space-time model.

Background technology:

Target following is the high-rise visual processes in video monitoring system, namely the correlation technique such as computer vision, image/video process is utilized, when not needing human intervention, target in the image sequence of video camera shooting is described, processes and is analyzed, realize the detection to moving target in dynamic scene, tracking and identification, then according to the target signature processed and analyze, interested movement locus is obtained.Motion target tracking, detection are important research contents of field of video monitoring, and the effect of target in video image tracking, detection directly has influence on the accuracy of the advanced processes such as the behavioral value of the target in subsequent video processing procedure, event understanding and analysis ^[1].

In recent years, for the tracking problem under the disturbing effects such as target scale, illumination variation and partial occlusion in actual environment, Chinese scholars constantly proposes new track algorithm or Improving ways ^[2].But research object is only confined to target itself by a lot of tracking, have ignored the information characteristics around target, causes arithmetic accuracy not high.In order to strengthen tracking accuracy, open China etc. ^[3]introduce target surrounding context spatial domain relation, utilize target and peripheral information relation Modling model thereof to carry out future position, improve the robustness of the interference such as the anti-partial occlusion of algorithm and illumination variation.But when the significant change of target appearance yardstick, the feature extracted of the method cannot effective Modling model, easily occurs that track window departs from target, even lose objects.

The present invention is directed to target scale and change the tracking accuracy decline problem caused, introduce the multiple dimensioned history To Template storehouse of using Clustering and building, propose a kind of target scale adaptive tracking algorithm based on space-time model, achieve the real-time target robust tracking of dimensional variation.

Summary of the invention:

Fundamental purpose of the present invention proposes a kind of target scale adaptive tracking algorithm based on space-time model, under the disturbing effects such as target scale, illumination variation and partial occlusion, and rapid, accurate localizing objects region.

To achieve these goals, the invention provides following technical scheme:

Step one, read in the first two field picture Image ₁, manually specify tracking target rectangle position Ζ;

Step 2, based on context spatial domain Ω _c, initialization space-time model order

$H_1^{stc}\left(X\right)=h_1^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\α}}|{|}^{\mathrm{\β}}}\right)}{F\left(I_1\right(X){\mathrm{\ω}}_{\mathrm{\σ}}(X-x^{*}))}\right);$

Step 3, calculating target area Z ₁direction gradient feature histogram f _{hOG_Z}, make m=<I _z, f _{hOG_Z}>, M=M ∪ m, initialize multi-scale history To Template storehouse M;

Step 4, read in next frame image Image _t+1(t>=1);

Step 5, iteration build space-time model order $H_{t+1}^{stc}=\left(1-\mathrm{\&eta;}\right)H_t^{stc}+{\mathrm{\&eta;h}}_t^{sc},\left(t\&GreaterEqual;1\right),$ Wherein $h_t^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_t\right(X){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-x^{*}))}\right);$

Step 6, calculating confidence map G _t+1, order $G_{t+1}\left(X\right)=H_{t+1}^{stc}\left(X\right)\&CircleTimes;I_{t+1}\left(X\right){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*}),$ Estimating target center $x_{t+1}^{*}=\arg \underset{X\&Element;{\mathrm{\&Omega;}}_C\left({x_t}^{*}\right)}{max}{G}_{t+1}\left(X\right);$

Step 7, with the target's center estimated in step 6 point centered by, extract the sample of n different scale, be all normalized to the image block of 8 × 8, comparison history To Template storehouse M, judge templet optimal scale, determines target rectangle position, completes the target following of t+1 frame;

Step 8, the optimal scale estimated according to step 7, upgrade space-time model scale parameter;

Step 9, according to the optimal objective region Z estimated in step 7 _t+1, upgrade multiple dimensioned history To Template storehouse M;

If step 10 video does not terminate, then proceed to step 4, read in next frame image; Otherwise follow the tracks of and terminate.

Compared with prior art, the present invention has following beneficial effect:

1. utilize the optimal scale of target to upgrade space-time model by step 7, achieve the real-time modeling method of dimensional variation.

2. use Clustering by step 9 and build the multiple dimensioned history To Template storehouse dynamically updated, construct the templatespace that can represent dbjective state, improve the accuracy representing dbjective state.

3. in conjunction with space-time model, that uses Clustering structure dynamically updates multiple dimensioned history To Template storehouse and optimal scale estimation, the present invention is when target is disturbed by illumination variation, partial occlusion and quick movement etc., successfully manage the change of target appearance yardstick, achieve robust tracking.

Therefore, will be with a wide range of applications in the application that the present invention monitors at public safety.

Accompanying drawing illustrates:

Fig. 1 algorithm flow chart of the present invention

The context spatial domain schematic diagram of Fig. 2 establishing target

The key diagram of Fig. 3 different scale sample;

The key diagram in Fig. 4 To Template space;

Fig. 5 algorithm is tracking effect in the experiment of Singer1 sequence image;

Fig. 6 algorithm is at Singer1 sequence image experiment medial error tracing analysis figure;

Fig. 7 algorithm is tracking effect in the experiment of David sequence image;

Fig. 8 algorithm is at David sequence image experiment medial error tracing analysis figure;

Fig. 9 algorithm is tracking effect in the experiment of CarScale sequence image;

Figure 10 algorithm is at CarScale sequence image experiment medial error tracing analysis figure;

Embodiment

In order to object of the present invention, concrete steps and feature are better described, below in conjunction with accompanying drawing, the present invention is further detailed explanation:

With reference to figure 1, a kind of target scale adaptive tracking method based on space-time model that the present invention proposes, mainly comprises following steps:

Step one, read in the first two field picture Image ₁, manually specify tracking target rectangle position Ζ;

Step 2, based on context spatial domain Ω _c, initialization space-time model order

$H_1^{stc}\left(X\right)=h_1^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_1\left(X\right){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}\left(X-x^{*}\right)\right)}\right);$

Step 3, calculating target area Z ₁direction gradient feature histogram f _{hOG_Z}, make m=<I _z, f _{hOG_Z}>, M=M ∪ m, initialize multi-scale history To Template storehouse M;

Step 4, read in next frame image Image _t+1(t>=1);

Step 5, iteration build space-time model order $H_{t+1}^{stc}=\left(1-\mathrm{\&eta;}\right)H_t^{stc}+{\mathrm{\&eta;h}}_t^{sc},\left(t\&GreaterEqual;1\right),$ Wherein $h_t^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_t\right(X){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-x^{*}))}\right);$

Step 6, calculating confidence map G _t+1, order $G_{t+1}\left(X\right)=H_{t+1}^{stc}\left(X\right)\&CircleTimes;I_{t+1}\left(X\right){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*}),$ Estimating target center $x_{t+1}^{*}=\arg \underset{X\&Element;{\mathrm{\&Omega;}}_C\left({x_t}^{*}\right)}{max}{G}_{t+1}\left(X\right);$

Step 7, with the target's center estimated in step 6 point centered by, extract the sample of n different scale, be all normalized to the image block of 8 × 8, comparison history To Template storehouse M, judge templet optimal scale, determines target rectangle position, completes the target following of t+1 frame;

Step 8, the optimal scale estimated according to step 7, upgrade space-time model scale parameter;

Step 9, according to the optimal objective region Z estimated in step 7 _t+1, upgrade multiple dimensioned history To Template storehouse M;

If step 10 video does not terminate, then proceed to step 4, read in next frame image; Otherwise follow the tracks of and terminate.

In technique scheme, the target rectangle position Ζ in step one as shown in solid box in Fig. 2, x ^*for target area geometric center, W _zand H _zrepresent that target area Ζ's is wide and high respectively.

In technique scheme, the context spatial domain Ω in step 2 _cas shown in dotted line frame in Fig. 2, represent target and peripheral information thereof.Context spatial domain Ω _cwith target area geometric center x ^*centered by, Ω _cwide and height be defined as

$W_{{\mathrm{\&Omega;}}_C}=2W_Z,H_{{\mathrm{\&Omega;}}_C}=2H_Z$

In technique scheme, space-time model in step 2 initial method be:

1. objective definition context spatial domain Ω _cfeature X ^c=c (m)=(I (m), m) | m ∈ Ω _c(x ^*), wherein Ω _c(x ^*) represent with x ^*centered by the corresponding context spatial domain of target, m represents spatial domain Ω _c(x ^*) in pixel, I (m) represents the gray-scale value of pixel m;

2. build the Spatial Domain characterizing target and its peripheral reference

$H_1^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_1\left(X\right){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}\left(X-x^{*}\right)\right)}\right)$

Wherein X ∈ R ²represent context spatial domain Ω _c(x ^*) middle set of pixels, F () represents Fourier transform function, F ^-1() represents inverse Fourier transform function, and ‖ ‖ represents Euclidean distance.Parameter recommendation value is α=2.25, β=1.ω _σ() is Gaussian function, is defined as

${\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_1}^{*})=e^{-\frac{\left|\right|X-{x_1}^{*}|{|}^2}{{{\mathrm{\&sigma;}}_1}^2}},{\mathrm{\&sigma;}}_1=\frac{W_{{\mathrm{\&Omega;}}_C}+H_{{\mathrm{\&Omega;}}_C}}{2}$

3. make complete space-time model initialization.

In technique scheme, in step 3, the initial method of multiple dimensioned history To Template storehouse M is:

1. make

2. by target area Z ₁transfer gray-scale map to, and normalization Z ₁be the image block I of 8 × 8 pixels _z;

3. calculate I _zdirection gradient feature histogram f _{hOG_Z};

4. make m=<I _z, f _{hOG_Z}>, M=M ∪ m;

5. multiple dimensioned history To Template storehouse M initialization completes;

In technique scheme, in step 5, iteration builds space-time model method be:

$H_{t+1}^{stc}=\left(1-\mathrm{\&eta;}\right)H_t^{stc}+{\mathrm{\&eta;h}}_t^{sc},\left(t\&GreaterEqual;1\right),$

Wherein with represent t, t+1 moment space-time model respectively, η is the learning rate upgraded, and suggestion adopts η=0.075. represent the target empty domain model of t, characterize target and its peripheral reference, circular is as follows:

$h_t^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_t\right(X){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-x^{*}))}\right)$

Wherein X ∈ R ²represent context spatial domain Ω _c(x ^*) middle set of pixels, F () represents Fourier transform function, F ^-1() represents inverse Fourier transform function, and ‖ ‖ represents Euclidean distance.Parameter recommendation value is α=2.25, β=1.ω _σ() is Gaussian function, is defined as

${\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*})=e^{-\frac{\left|\right|X-{x_t}^{*}|{|}^2}{{{\mathrm{\&sigma;}}_t}^2}}$

In technique scheme, confidence map G in step 6 _t+1computing method be:

$G_{t+1}\left(X\right)=H_{t+1}^{stc}\left(X\right)\&CircleTimes;I_{t+1}\left(X\right){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*})$

Wherein represent convolution algorithm symbol, X ∈ R ²represent context spatial domain Ω _cmiddle set of pixels, G _t+1(X) the context spatial domain Ω of t is represented _cin scope, pixel is in the value of the confidence in t+1 moment, and its value represents that this point drops on the probability of target area Ζ.The point that probable value is the highest is the possible center of t+1 moment target

$x_{t+1}^{*}=\arg \underset{X\&Element;{\mathrm{\&Omega;}}_C\left({x_t}^{*}\right)}{max}{G}_{t+1}\left(X\right)$

In technique scheme, in step 7, template optimal scale determination methods is:

1. with the target's center estimated in step 6 point centered by, extract the sample of n different scale, be all normalized to the image block (as shown in Figure 3) of 8 × 8, proposed parameter n=20 in the present invention;

2. build sample space D={d to be matched _{j ∈ [1, n]}, wherein d _jrepresent a jth sample, the direction gradient feature histogram of its correspondence is expressed as s _jrepresent the yardstick of corresponding sample;

3. the histograms of oriented gradients similarity of k template in the sample of calculated crosswise n different scale and multiple dimensioned history To Template storehouse M, obtains similar matrix S _dM∈ R ^{n × k},

$S_{DM}=\left\{s(m_i,d_j)\right|s(m_i,d_j)=1-\left|\right|f_{{\mathrm{HOG}}_{m_i}}-f_{{\mathrm{HOG}}_{d_j}}|{|}^2,m_i\&Element;M,d_j\&Element;D\}$

Wherein the highest with template similarity sample be t+1 moment target area Z _t+1optimal estimation, the yardstick of its correspondence

In technique scheme, space-time model in step 8 scale parameter update method is:

1. target area Z _t+1wide and high:

W _Z(t+1)＝W _Z(t)*s

H _Z(t+1)＝H _Z(t)*s,

2. target context spatial domain Ω _cwide and high:

$W_{{\mathrm{\&Omega;}}_C}(t+1)=W_{{\mathrm{\&Omega;}}_C}\left(t\right)*s$

$H_{{\mathrm{\&Omega;}}_C}(t+1)=H_{{\mathrm{\&Omega;}}_C}\left(t\right)*s$

3. Gaussian function ${\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*})=e^{-\frac{\left|\right|X-x*|{|}^2}{{{\mathrm{\&sigma;}}_t}^2}}$ In scale parameter σ _t:

σ _t+1＝σ _t*s

In technique scheme, in step 9, multiple dimensioned history To Template storehouse M update method is:

1. the optimal objective region Z will estimated in step 7 _t+1transfer gray-scale map to, and normalization Z _t+1be the image block I of 8 × 8 pixels _z;

2. calculate I _zdirection gradient feature histogram f _{hOG_Z};

3. make m=<I _z, f _{hOG_Z}>, M=M ∪ m;

4. if, t≤k (parameter recommendation value is k=10), multiple dimensioned history To Template storehouse M has upgraded, and algorithm terminates; Otherwise, proceed to step 5;

5. calculate similarity matrix S _m

$S_M=\left\{s(m_i,m_j)\right|s(m_i,m_j)=1-\left|\right|f_{{\mathrm{HOG}}_{m_i}}-f_{{\mathrm{HOG}}_{m_j}}|{|}^2,m_i\&Element;M,m_j\&Element;M\}$

Wherein represent template m _i/ m _jdirection gradient feature histogram;

6. obtain the minimum template pair of similarity $(m_{min1},m_{min2})=\underset{m_i\&Element;M,m_j\&Element;M}{\mathrm{argmin}}s(m_i,m_j);$

7. calculate m respectively _min1, m _min2with the similarity of other templates and, S _{sum_p}=∑ _{mj ∈ M}s (m _p, m _j), m _p∈ { m _min1, m _min2;

8. if S _{sum_min1}>=S _{sum_min2}, adjustment templatespace M=M-m _min1otherwise, M=M-m _min2;

9. multiple dimensioned history To Template storehouse M has upgraded;

In technique scheme, in step 9, multiple dimensioned history To Template storehouse M is after K >=k renewal, as shown in Figure 4, template number in template base | M| will remain k, and k the representational dbjective state of most remained in template base from the t=1 moment, along with the continuation followed the tracks of, template base dynamically updates continuing.

In technique scheme, effect such as the Fig. 5 based on the target scale adaptive tracking method of space-time model shows.Fig. 5 gives algorithm in the experiment of Singer1 sequence image, and target object experiences the visual tracking effect continuing to diminish etc. under disturbed condition of violent illumination variation and yardstick.Fig. 6 is the algorithm center position of tracking and graph of errors analysis chart of standard-track central point in the experiment of Singer1 sequence image.Fig. 7 gives algorithm in the experiment of David sequence image, the visual tracking effect under the disturbed conditions such as target object experience illumination variation, plane internal rotation turn, non-linear deformation and the irregular change of yardstick.Fig. 8 is the algorithm center position of tracking and graph of errors analysis chart of standard-track central point in the experiment of David sequence image.Fig. 9 gives algorithm in the experiment of CarScale sequence image, the visual tracking effect under the disturbed conditions such as the significant change of target object experience target appearance yardstick, partial occlusion and rapid movement.Figure 10 is the algorithm center position of tracking and graph of errors analysis chart of standard-track central point in the experiment of CarScale sequence image.By three groups of sequential tests, experimental result is described with tracking effect figure and quantitative graph of errors qualitatively, the precision of verification algorithm and robustness.

This patent utilizes target and its surrounding context spatial information (si), and this target serial relation on a timeline, sets up space-time model.Effective extraction target signature, that avoids the interference such as partial occlusion and illumination variation to cause departs from.Use Clustering and build screening rule, dynamically update the representational template of most and build templatespace, accurately represent the state of target.Incoming direction histogram of gradients signature analysis template and Sample Similarity, improve the accuracy of coupling further.The target optimal scale finally obtained according to coupling upgrades space-time model, realizes the real-time modeling method of dimensional variation, the robustness of boosting algorithm.Experimental verification, the method that this patent proposes, when target is disturbed by illumination variation and partial occlusion etc., successfully manages the change of target appearance yardstick.

By reference to the accompanying drawings the specific embodiment of the present invention is elaborated above, but the present invention is not limited to above-mentioned embodiment, in the ken that those of ordinary skill in the art possess, can also make a variety of changes under the prerequisite not departing from present inventive concept.

[1]Felzenszwalb,P.F.,Girshick,R.B.,McAllester,D.,etal.ObjectDetectionwithDiscriminativelyTrainedPart-BasedModels[J].PatternAnalysisandMachineIntelligence,2010,32(9):1627-1645.

[2]WUYi,LimJongwoo,YangM-H.ObjectTrackingBenchmark[J].IEEETransactionsonPatternAnalysisandMachineIntelligence,InPress,2014.

[3]ZHANGKai-hua,ZHANGLei,,YangM-H.FastVisualTrackingviaDenseSpatio-TemporalContextLearning[C].Proceeding.of13thEuropeanConferenceonComputerVision,2014.

Claims (9)

1., based on the target scale adaptive tracking method of space-time model, it is characterized in that, comprise the following steps:

Step one, read in the first two field picture Image ₁, manually specify tracking target rectangle position Ζ;

Step 2, based on context spatial domain Ω _c, initialization space-time model order

$H_1^{stc}\left(X\right)=h_1^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_1\right(X){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-x^{*}))}\right);$

Step 3, calculating target area Z ₁direction gradient feature histogram f _{hOG_Z}, make m=<I _z, f _{hOG_Z}>, M=M ∪ m, initialize multi-scale history To Template storehouse M;

Step 4, read in next frame image Image _t+1(t>=1);

Step 5, iteration build space-time model order $H_{t+1}^{stc}=(1-\mathrm{\&eta;})H_t^{stc}+{\mathrm{\&eta;h}}_t^{sc}(t\&GreaterEqual;1),$ Wherein $h_t^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_t\right(X){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-x^{*}))}\right);$

Step 6, calculating confidence map G _t+1, order $G_{t+1}\left(X\right)=H_{t+1}^{stc}\left(X\right)\&CircleTimes;I_{t+1}\left(X\right){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*}),$ Estimating target center $x_{t+1}^{*}=\arg \underset{X\&Element;{\mathrm{\&Omega;}}_c\left({x_t}^{*}\right)}{max}{G}_{t+1}\left(X\right);$

Step 7, with the target's center estimated in step 6 point centered by, extract the sample of n different scale, be all normalized to the image block of 8 × 8, comparison history To Template storehouse M, judge templet optimal scale, determines target rectangle position, completes the target following of t+1 frame;

Step 8, the optimal scale estimated according to step 7, upgrade space-time model scale parameter;

Step 9, according to the optimal objective region Z estimated in step 7 _t+1, upgrade multiple dimensioned history To Template storehouse M;

If step 10 video does not terminate, then proceed to step 4, read in next frame image; Otherwise follow the tracks of and terminate.

2. the target scale adaptive tracking method based on space-time model according to claim 1, is characterized in that, space-time model in step 2 initial method be:

1) objective definition context spatial domain Ω _cfeature X ^c=c (m)=(I (m), m) | m ∈ Ω _c(x ^*), wherein Ω _c(x ^*) represent with x ^*centered by the corresponding context spatial domain of target, m represents spatial domain Ω _c(x ^*) in pixel, I (m) represents the gray-scale value of pixel m;

2) Spatial Domain characterizing target and its peripheral reference is built

$h_1^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_1\right(X){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-x^{*}))}\right)$

Wherein X ∈ R ²represent context spatial domain Ω _c(x ^*) middle set of pixels, F () represents Fourier transform function, F ^-1() represents inverse Fourier transform function, and ‖ ‖ represents Euclidean distance; Parameter recommendation value is α=2.25, β=1; ω _σ() is Gaussian function, is defined as

${\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_1}^{*})=e^{-\frac{\left|\right|X-{x_1}^{*}|{|}^2}{{{\mathrm{\&sigma;}}_1}^2}},{\mathrm{\&sigma;}}_1=\frac{W_{{\mathrm{\&Omega;}}_C}+H_{{\mathrm{\&Omega;}}_C}}{2}$

3) make complete space-time model initialization.

3. the target scale adaptive tracking method based on space-time model according to claim 1, is characterized in that, the initial method of the multiple dimensioned history To Template storehouse M described in step 3 is:

1) make

2) by target area Z ₁transfer gray-scale map to, and normalization Z ₁be the image block I of 8 × 8 pixels _z;

3) I is calculated _zdirection gradient feature histogram f _{hOG_Z};

4) m=<I is made _z, f _{hOG_Z}>, M=M ∪ m;

5) multiple dimensioned history To Template storehouse M initialization completes.

4. the target scale adaptive tracking method based on space-time model according to claim 1, is characterized in that, the iteration described in step 5 builds space-time model method be:

$H_{t+1}^{stc}=(1-\mathrm{\&eta;})H_t^{stc}+{\mathrm{\&eta;h}}_t^{sc}(t\&GreaterEqual;1)$

Wherein with represent t, t+1 moment space-time model respectively, η is the learning rate upgraded, and suggestion adopts η=0.075; represent the target empty domain model of t, characterize target and its peripheral reference, circular is as follows:

$h_t^{SC}\left(X\right)=F^{-1}\left(\frac{F\left(e^{-\left|\right|\frac{X-x^{*}}{\mathrm{\&alpha;}}|{|}^{\mathrm{\&beta;}}}\right)}{F\left(I_t\right(X){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-x^{*}))}\right)$

Wherein X ∈ R ²represent context spatial domain Ω _c(x ^*) middle set of pixels, F () represents Fourier transform function, F ^-1() represents inverse Fourier transform function, and ‖ ‖ represents Euclidean distance; Parameter recommendation value is α=2.25, β=1; ω _σ() is Gaussian function, is defined as

${\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*})=e^{-\frac{\left|\right|X-{x_t}^{*}|{|}^2}{{{\mathrm{\&sigma;}}_t}^2}}.$

5. the target scale adaptive tracking method based on space-time model according to claim 1, is characterized in that, the confidence map G described in step 6 _t+1computing method be:

$G_{t+1}\left(X\right)=H_{t+1}^{stc}\left(X\right)\&CircleTimes;I_{t+1}\left(X\right){\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*})$

Wherein represent convolution algorithm symbol, X ∈ R ²represent context spatial domain Ω _cmiddle set of pixels, G _t+1(X) the context spatial domain Ω of t is represented _cin scope, pixel is in the value of the confidence in t+1 moment, and its value represents that this point drops on the probability of target area Ζ; The point that probable value is the highest is the possible center of t+1 moment target

$x_{t+1}^{*}=\arg \underset{X\&Element;{\mathrm{\&Omega;}}_c\left({x_t}^{*}\right)}{max}{G}_{t+1}\left(X\right).$

6. the target scale adaptive tracking method based on space-time model according to claim 1, is characterized in that, the template optimal scale determination methods described in step 7 is:

1) with the target's center estimated in step 6 point centered by, extract the sample of n different scale, be all normalized to the image block (as shown in Figure 3) of 8 × 8, proposed parameter n=20 in the present invention;

2) sample space D={d to be matched is built _{j ∈ [1, n]}, wherein d _jrepresent a jth sample, the direction gradient feature histogram of its correspondence is expressed as represent the yardstick of corresponding sample;

3) the histograms of oriented gradients similarity of k template in the sample of calculated crosswise n different scale and multiple dimensioned history To Template storehouse M, obtains similar matrix S _dM∈ R ^{n × k},

$S_{DM}=\left\{s(m_i,d_j)\right|s(m_i,d_j)=1-\left|\right|f_{{\mathrm{HOG}}_{m_i}}-f_{{\mathrm{HOG}}_{d_j}}|{|}^2,m_i\&Element;M,d_j\&Element;D\}$

Wherein the highest with template similarity sample be t+1 moment target area Z _t+1optimal estimation, the yardstick of its correspondence

7. the target scale adaptive tracking method based on space-time model according to claim 1, is characterized in that, space-time model in described step 8 scale parameter update method is:

1) target area Z _t+1wide and high:

W _Z(t+1)＝W _Z(t)*s

H _Z(t+1)＝H _Z(t)*s,

2) target context spatial domain Ω _cwide and high:

$W_{{\mathrm{\&Omega;}}_c}(t+1)=W_{{\mathrm{\&Omega;}}_c}\left(t\right)*s$

$H_{{\mathrm{\&Omega;}}_c}(t+1)=H_{{\mathrm{\&Omega;}}_c}\left(t\right)*s$

3) Gaussian function ${\mathrm{\&omega;}}_{\mathrm{\&sigma;}}(X-{x_t}^{*})=e^{-\frac{\left|\right|X-x^{*}|{|}^2}{{{\mathrm{\&sigma;}}_t}^2}}$ In scale parameter σ _t:

σ _t+1＝σ _t*s。

8. the target scale adaptive tracking method based on space-time model according to claim 1, is characterized in that, in described step 9, multiple dimensioned history To Template storehouse M update method is:

1) the optimal objective region Z will estimated in step 7 _t+1transfer gray-scale map to, and normalization Z _t+1be the image block I of 8 × 8 pixels _z;

2) I is calculated _zdirection gradient feature histogram f _{hOG_Z};

3) m=<I is made _z, f _{hOG_Z}>, M=M ∪ m;

4) if t≤k (parameter recommendation value is k=10), multiple dimensioned history To Template storehouse M has upgraded, and algorithm terminates; Otherwise, proceed to step 5;

5) similarity matrix S is calculated _m

$S_M=\left\{s(m_i,m_j)\right|s(m_i,m_j)=1-\left|\right|f_{{\mathrm{HOG}}_{m_i}}-f_{{\mathrm{HOG}}_{m_j}}|{|}^2,m_i\&Element;M,m_j\&Element;M\}$

Wherein represent template m _i/ m _jdirection gradient feature histogram;

6) the minimum template pair of similarity is obtained $(m_{min1},m_{min2})=\underset{m_i\&Element;M,m_j\&Element;M}{\mathrm{argmin}}s(m_i,m_j);$

7) m is calculated respectively _min1, m _min2with the similarity of other templates and, $S_{sum\_p}=$ ${\mathrm{\&Sigma;}}_{m_j\&Element;M}s(m_p,m_j),m_p\&Element;\{m_{min1},m_{min2}\};$

8) if S _{sum_min1}>=S _{sum_min2}, adjustment templatespace M=M-m _min1otherwise, M=M-m _min2;

9) multiple dimensioned history To Template storehouse M has upgraded.

9. the target scale adaptive tracking method based on space-time model according to claim 1, it is characterized in that, in described step 9, multiple dimensioned history To Template storehouse M is after K >=k renewal, template number in template base | M| will remain k, and k the representational dbjective state of most remained in template base from the t=1 moment, along with the continuation followed the tracks of, template base dynamically updates continuing.