CN106295548A - A kind of method for tracking target and device - Google Patents

Abstract

The present invention provides a kind of method for tracking target and device, belongs to computer vision and area of pattern recognition.For solving the problem that existing target tracking algorism complexity is high.Including: in the Ith frame, select first object position, the tracking result of first object is defined as the first positive sample, formula (1) is passed through from the multiple first positive sample packages and the first negative sample bag, determine the feature set of the first example collection, by the eigenmatrix formula (2) of the first training data, determine a MIL 1 norm SVM, by formula (3), it is configured to the first minimum support example collection, the first minimum weights supporting example collection are determined by formula (4), the minimum classification mark supporting example collection of corresponding for MIL 1 norm SVM first based on weights distribution is determined by formula (5)；Candidate samples is classified by a MIL 1 norm SVM by I+1 frame, determines the second target location in I+1 two field picture according to the second minimum weights supporting example collection and formula (6).

Description

A kind of method for tracking target and device

Technical field

The invention belongs to computer vision and area of pattern recognition, be specifically related to a kind of method for tracking target and device.

Background technology

Target is one of the important topic in computer vision research field from motion tracking, and it is widely used in intelligence prison Control, traffic flow analysis, onboard system navigation and the various aspects such as man-machine interaction.Study sane Target Tracking System and there is weight The meaning wanted and researching value.Owing to reality tracking system is often by ambient lighting, mixed and disorderly background, block and target self yardstick With the impact of deformation, the decline of tracking accuracy is caused even to follow the tracks of the phenomenon of drift, it is difficult to reach the requirement of Practical Project.

Owing to, in object tracking process, the yardstick of target, rotation and occlusion issue cause the degeneration of target following template, Cause error accumulation, thus affect the performance of tracking system.Therefore, effective To Template is set up for improving tracking system Stability is most important.

Conventional To Template can be divided into template for production (English is: generative model) and discriminant template (English is: discriminative model).Owing to track algorithm based on template for production only utilizes during following the tracks of In first two field picture, target information sets up template, most like with set up To Template by finding during subsequent frame is followed the tracks of Region be tracked, its do not consider follow the tracks of during target self and the change of background environment, therefore when target appearance change and Ambient occlusion, during homologue interference, it is poor that the method follows the tracks of robustness.Track algorithm based on discriminant template is due to the most sharp Grader is trained, by training two classification during following the tracks of with from the positive sample of target area and the negative sample of background area Device realizes target and separates with background, and it is followed the tracks of robustness and is often higher than track algorithm based on template for production.

The typical algorithm that target tracking algorism based on multi-instance learning is followed the tracks of as discriminant template, is following the tracks of in recent years System is widely used and shows preferable tracking performance.In multi-instance learning track algorithm, it trains number Representing according to by positive and negative sample packages, if containing positive sample in Bao, its bag is positive closure, otherwise, wrap as negative bag.In this track algorithm, Bag has positive negative flag, and the example forming bag is unmarked.It obtains grader by the training data study of multiple bag compositions, Example the most correct in test bag is found as following the tracks of result by grader.Many examples follow the tracks of system according to the target position predicted Putting, positive and negative sample instantiation of sampling in its optimal tracing positional certain neighborhood region composition trains positive and negative sample packages.But meet The sample instantiation of above-mentioned sampling request is more, and, its positive sample packages constructed often contains substantial amounts of negative sample.Cause This, comprise example useless to classification in a large number, from the point of view of training angle, substantial amounts of training number during classifier training in bag According to increasing the complexity that algorithm calculates the performance affecting grader.

In sum, it is too many to there is the example in positive and negative sample packages in existing target tracking algorism, causes classifier performance Decline and the high problem of algorithm computation complexity.

Summary of the invention

The embodiment of the present invention provides a kind of method for tracking target and device, just exists in order to solve existing target tracking algorism Example in negative sample bag is too many, causes the problem that classifier performance declines and algorithm computation complexity is high.

The embodiment of the present invention provides a kind of method for tracking target, it is characterised in that including:

In image sequence the Ith frame, select first object position, described first object is tracked, by described first mesh Target is followed the tracks of result and is defined as the first positive sample, by formula (1), determines first become from the multiple described first positive sample architecture The Gradient Features of the first positive and negative sample packages extracted in the first negative sample bag that positive sample packages and the first negative sample are constituted；According to institute State the Gradient Features of the first positive and negative sample packages, determine that the described first positive sample packages and described first negative sample bag include first The feature set of example collection, is defined as the eigenmatrix of the first training data by the feature set of described first example collection；

A MIL is determined by multi-instance learning framework, the eigenmatrix of described first training data, and formula (2) 1-norm SVM, is classified to described first example collection by a described MIL 1-norm SVM, by formula (3) by institute Stating the first example concentrates the described first example collection meeting first threshold to be configured to the first minimum support example collection, passes through formula (4) determine the described first minimum weights supporting example collection, lose according to the described first minimum weights supporting example collection and first Forget the factor, determine the described first corresponding for MIL 1-norm SVM minimum support example based on weights distribution by formula (5) The classification mark of collection；

Candidate samples is selected, by a described MIL 1-norm SVM to described candidate in image sequence I+1 frame Sample is classified, and according to described classification results and described formula (4), determines the second minimum weights supporting example collection, root According to the described second minimum weights supporting example collection, the described first minimum classification mark supporting example collection and formula (6) determine The second target location in described I+1 two field picture；

Described formula (1) is as follows:

G=min (| g_x|+|g_y|,255)

Described formula (2) is as follows:

$y=sign\left(\underset{k\∈\mathrm{\Λ}}{\Σ}{w}_k^{*}g\right(B_i,x_k)+b^{*})$

Described formula (3) is as follows:

${\mathrm{\Λ}}_{j^{*}}=\left\{k:k\∈\mathrm{\Λ},x_{ij}\∈B_i,j^{*}=\arg {\max }_j{e}^{-\frac{\left|\right|x_{ij}-x_k\left|\right|}{{\mathrm{\σ}}^2}}\right\}$

Described formula (4) is as follows:

$h_t\left(x_s\right)=\underset{k\∈{\mathrm{\Λ}}_{j^{*}}}{\Σ}\frac{w_k^{*}s(x_k,x_s)}{m_k}$

Described formula (5) is as follows:

$H\left(x_s\right)=\underset{t}{\Σ}{e}^{-\frac{1}{t}}{h}_t\left(x_s\right)$

Described formula (6) is as follows:

$l_{t+1}^{*}=l\left(\arg {\max }_{x\∈X^{\mathrm{\γ}}}H\right(x\left)\right)$

Wherein, g is the Gradient Features of described first positive and negative sample packages, g_xFor described first positive and negative sample packages horizontal direction Gradient, g_yFor the gradient of described first positive and negative sample packages vertical direction,b^*Optimum that is that be respectively weight vector and that bias Value, x_kFor kth sample, B_iFor arbitrary sample bag, y is discriminant function,For w^*The index of middle nonzero element,For with bag B_iRelevant first is minimum supports example indexed set, x_ijFor arbitrary sample bag B_iMiddle jth example,For referring to Number function, σ²For the predefined factor, h_t(x_s) it is the first minimum classification weights supporting example, m_kRepresent and support example x_kVery big Value number, H (x_s) it is the first minimum classification mark supporting example collection,For corresponding forgetting factor, t is frame number,It is t Target location in+1, l (x) is the position at sample place, x ∈ X^γFor the sample in the γ of region of search, for t+1 Two field picture, with t frame positionCentered by, γ=30 form candidate samples collection X for search radius sampling^γ, i.e.

Preferably, described according to the described second minimum weights supporting example collection, the described first minimum support example collection After classification mark and formula (6) determine the second target location in described I+1 two field picture, also include:

Described second target is tracked, the tracking result of described second target is defined as the second positive sample, passes through The second negative sample that formula (1), the second positive sample packages become from the multiple described second positive sample architecture and the second negative sample are constituted Bag extracts the Gradient Features of the second positive and negative sample packages；According to the Gradient Features of described second positive and negative sample packages, determine described The feature set of all second examples that two positive and negative sample packages include, is defined as second by the feature set of all described second examples The eigenmatrix of training data；By a described MIL 1-norm SVM, described second example collection is classified, by institute State formula (3) concentrates the described second example collection meeting Second Threshold to be configured to the second minimum support example by described second example Collection, and support example collection with the described second minimum support example collection renewal described first is minimum；

Eigenmatrix according to described second training data and formula (7), carried out a described MIL 1-norm SVM Update；

Formula (7) is as follows:

$\underset{w,b,\mathrm{\ϵ},\mathrm{\η}}{min}\mathrm{\λ}\underset{k=1}{\overset{N+L-1}{\Σ}}\left|w_k\right|+C_1\underset{i=0}{\overset{l^{+}}{\Σ}}{\mathrm{\ϵ}}_i+C_2\underset{j=1}{\overset{l^{-}}{\Σ}}{\mathrm{\η}}_j$

$\begin{array}{c}s.t.(w^T{g}_i^{+}+b)\≥1-{\mathrm{\ϵ}}_i,i=0,\mathrm{...},l^{+}\\ -(w^T{g}_j^{-}+b)\≥1-{\mathrm{\η}}_j,j=0,\mathrm{...},l^{-}\\ {\mathrm{\ϵ}}_i,{\mathrm{\η}}_j\≥0\end{array}$

Wherein, λ is scale factor, ε_iFor the slack variable of positive sample packages feature, η_jLax change for negative sample bag feature Amount,For the constraint total amount of positive sample packages,For the constraint total amount of negative sample bag, C₁Punishment for lookup error sample is joined Number, C₂Penalty term parameter for negative error sample.

Preferably, before the described weights being determined the described first minimum support example collection by formula (4), also include:

Determine that the described first minimum support example collection supports the similar of example Gradient Features to existing by following equation Angle value:

$s(B_i,x_k)=s\left(\right\{x_{{\mathrm{ij}}^{*}}\},x_k)=max\left(e^{-\frac{\left|\right|g_{{\mathrm{ij}}^{*}}-g_k|{|}^2}{{\mathrm{\σ}}^2}}\right)$

Wherein, example x_kWith bag B_iSimilarity s (B_i,x_k) it is defined as x_kWith bag B_iIn any exampleNearest example,And g_kRepresent bag B respectively_iMiddle jth example and example x_kGradient Features.

Preferably, described determine a described MIL 1-norm SVM before, also include:

1-norm SVM is determined by following equation:

Y=sign (w^T g+b)

Wherein, w, b are corresponding weight vectors and biasing.

The embodiment of the present invention also provides for a kind of target tracker, including:

First determines unit, for selecting first object position in image sequence the Ith frame, carries out described first object Follow the tracks of, the tracking result of described first object is defined as the first positive sample, by formula (1), determines from multiple described first The the first positive negative sample extracted in the first negative sample bag that first positive sample packages of positive sample architecture one-tenth and the first negative sample are constituted The Gradient Features of bag；According to the Gradient Features of described first positive and negative sample packages, determine the described first positive sample packages and described first The feature set of the first example collection that negative sample bag includes, is defined as the first training data by the feature set of described first example collection Eigenmatrix；

Second determines unit, for by multi-instance learning framework, the eigenmatrix of described first training data, Yi Jigong Formula (2) determines a MIL 1-norm SVM, is carried out described first example collection point by a described MIL 1-norm SVM Class, is concentrated described first example by formula (3) and meets the described first example collection of first threshold and be configured to the first ramuscule Hold example collection, determine the described first minimum weights supporting example collection by formula (4), according to the described first minimum support example Collection weights and the first forgetting factor, by formula (5) determine based on weights distribution MIL 1-norm SVM corresponding described in The first minimum classification mark supporting example collection；

3rd determines unit, for selecting candidate samples in image sequence I+1 frame, by a described MIL 1- Described candidate samples is classified by norm SVM, according to described classification results and described formula (4), determines the second ramuscule Hold the weights of example collection, according to the described second minimum weights supporting example collection, the described first minimum classification supporting example collection Mark and formula (6) determine the second target location in described I+1 two field picture；

Described formula (1) is as follows:

G=min (| g_x|+|g_y|,255)

Described formula (2) is as follows:

$y=sign\left(\underset{k\∈\mathrm{\Λ}}{\Σ}{w}_k^{*}g\right(B_i,x_k)+b^{*})$

Described formula (3) is as follows:

${\mathrm{\Λ}}_{j^{*}}=\left\{k:k\∈\mathrm{\Λ},j^{*}=\arg {\max }_j{e}^{-\frac{\left|\right|x_{ij}-x_k\left|\right|}{{\mathrm{\σ}}^2}}\right\}$

Described formula (4) is as follows:

$h_t\left(x_s\right)=\underset{k\∈{\mathrm{\Λ}}_{j^{*}}}{\Σ}\frac{w_k^{*}s(x_k,x_s)}{m_k}$

Described formula (5) is as follows:

$H\left(x_s\right)=\underset{t}{\Σ}{e}^{-\frac{1}{t}}{h}_t\left(x_s\right)$

Described formula (6) is as follows:

$l_{t+1}^{*}=l\left(\arg {\max }_{x\∈X^{\mathrm{\γ}}}H\right(x\left)\right)$

Wherein, g is the Gradient Features of described first positive and negative sample packages, g_xFor described first positive and negative sample packages horizontal direction Gradient, g_yFor the gradient of described first positive and negative sample packages vertical direction,b^*Optimal value that is that be respectively weight vector and that bias, x_kFor kth sample, B_iFor arbitrary sample bag, y is discriminant function,For w^*The index of middle nonzero element,For With bag B_iRelevant first is minimum supports example indexed set, x_ijFor arbitrary sample bag B_iMiddle jth example,For index letter Number, σ²For the predefined factor, h_t(x_s) it is the first minimum classification weights supporting example, m_kRepresent and support example x_kMaximum Number, H (x_s) it is the first minimum classification mark supporting example collection,For corresponding forgetting factor, t is frame number,It is in t+1 Target location, l (x) is the position at sample place, x ∈ X^γFor the sample in the γ of region of search, for t+1 frame figure Picture, with t frame positionCentered by, γ=30 form candidate samples collection X for search radius sampling^γ, i.e.

Preferably, the 4th unit and updating block is also included；

The tracking result of described second target, for being tracked described second target, is defined as by described Unit the 4th Second positive sample, by formula (1), the second positive sample packages and the second negative sample structure become from the multiple described second positive sample architecture The the second negative sample bag become extracts the Gradient Features of the second positive and negative sample packages；Gradient according to described second positive and negative sample packages is special Levy, determine the feature set of all second examples that described second positive and negative sample packages includes, by the spy of all described second examples Collection is defined as the eigenmatrix of the second training data；By a described MIL 1-norm SVM, described second example collection is entered Row classification, is concentrated described second example by described formula (3) and meets the described second example collection of Second Threshold and be configured to the Two minimum example collection of supporting, and support example collection with the described second minimum support example collection renewal described first is minimum；

Described updating block is for the eigenmatrix according to described second training data and formula (7), to a described MIL 1-norm SVM is updated；

Formula (7) is as follows:

$\begin{array}{c}\underset{w,b,\mathrm{\ϵ},\mathrm{\η}}{min}\mathrm{\λ}\underset{k=1}{\overset{N+L-1}{\Σ}}\left|w_k\right|+C_1\underset{i=0}{\overset{l^{+}}{\Σ}}{\mathrm{\ϵ}}_i+C_2\underset{j=1}{\overset{l^{-}}{\Σ}}{\mathrm{\η}}_j\\ s.t.(w^T{g}_i^{+}+b)\≥1-{\mathrm{\ϵ}}_i,i=0,\mathrm{...},l^{+}\\ -(w^T{g}_j^{-}+b)\≥1-{\mathrm{\η}}_j,j=0,\mathrm{...},l^{-}\\ {\mathrm{\ϵ}}_i,{\mathrm{\η}}_j\≥0\end{array}.$

Wherein, λ is scale factor, ε_iFor the slack variable of positive sample packages feature, η_jLax change for negative sample bag feature Amount,For the constraint total amount of positive sample packages,For the constraint total amount of negative sample bag, C₁Punishment for lookup error sample is joined Number, C₂Penalty term parameter for negative error sample.

Preferably, described second determines that unit is additionally operable to:

Determine that the described first minimum support example collection supports the similar of example Gradient Features to existing by following equation Angle value:

$s(B_i,x_k)=s\left(\right\{x_{{\mathrm{ij}}^{*}}\},x_k)=max\left(e^{-\frac{\left|\right|g_{{\mathrm{ij}}^{*}}-g_k|{|}^2}{{\mathrm{\σ}}^2}}\right)$

Wherein, example x_kWith bag B_iSimilarity s (B_i,x_k) it is defined as x_kWith bag B_iIn any exampleNearest example,And g_kRepresent bag B respectively_iMiddle jth example and example x_kGradient Features.

Preferably, described second determines that unit is additionally operable to:

1-norm SVM is determined by following equation:

Y=sign (w^T g+b)

Wherein, w, b are corresponding weight vectors and biasing.

In the embodiment of the present invention, it is provided that a kind of method for tracking target and device, it is included in image sequence the Ith frame selection First object position, is tracked described first object, and the tracking result of described first object is defined as the first positive sample, By formula g=min (| g_x|+|g_y|, 255), determine the first positive sample packages become from the multiple described first positive sample architecture and the The Gradient Features of the first positive and negative sample packages extracted in the first negative sample bag that one negative sample is constituted；According to described first positive and negative sample The Gradient Features of this bag, determines the feature of the first example collection that the described first positive sample packages and described first negative sample bag include Collection, is defined as the eigenmatrix of the first training data by the feature set of described first example collection；By multi-instance learning framework, institute State the eigenmatrix of the first training data, and formulaDetermine a MIL1-norm SVM, is classified to described first example collection by a described MIL 1-norm SVM, passes through formulaDescribed first example concentration is met described the first of first threshold show Example collection is configured to the first minimum support example collection, passes through formulaDetermine the described first minimum support The weights of example collection, according to the described first minimum weights supporting example collection and the first forgetting factor, pass through formulaDetermine the described first corresponding for MIL 1-norm SVM minimum support example collection based on weights distribution Classification mark；Candidate samples is selected, by a described MIL 1-norm SVM to described time in image sequence I+1 frame Sampling is originally classified, according to described classification results and described formulaDetermine that second is minimum The weights of support example collection, according to the described second minimum weights supporting example collection, dividing of the described first minimum support example collection Class mark and formulaDetermine the second target location in described I+1 two field picture；Wherein, g For the Gradient Features of described first positive and negative sample packages, g_xFor the gradient of described first positive and negative sample packages horizontal direction, g_yFor described The gradient of one positive and negative sample packages vertical direction,b^*Optimal value that is that be respectively weight vector and that bias, x_kFor kth sample, B_iFor arbitrary sample bag, y is discriminant function,For w^*The index of middle nonzero element,For with bag B_iRelevant the One minimum support example indexed set, x_ijFor arbitrary sample bag B_iMiddle jth example,For exponential function, σ²For predefined because of Son, h_t(x_s) it is the first minimum classification weights supporting example, m_kRepresent and support example x_kMaximum number, H (x_s) it is first The minimum classification mark supporting example collection,For corresponding forgetting factor, t is frame number,It is the target location in t+1, l X () is the position at sample place, x ∈ X^γFor the sample in the γ of region of search, for t+1 two field picture, with t framing bit PutCentered by, γ=30 form candidate samples collection X for search radius sampling^γ, i.e.In the present invention In embodiment, targeted measure based on Gradient Features estimates the importance of the first example, utilizes multi-instance learning framework, On-line training 1-norm SVM, selects important support example, thus eliminates the sample instantiation useless to classification, thus effectively Solve owing to a large amount of useless examples cause the problem that classifier performance declines and algorithm computation complexity is high.

Accompanying drawing explanation

In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing In having technology to describe, the required accompanying drawing used is briefly described, it should be apparent that, the accompanying drawing in describing below is only this Some embodiments of invention, for those of ordinary skill in the art, on the premise of not paying creative work, it is also possible to Other accompanying drawing is obtained according to these accompanying drawings.

A kind of method for tracking target schematic flow sheet that Fig. 1 provides for the embodiment of the present invention；

A kind of target tracker structural representation that Fig. 2 provides for the embodiment of the present invention.

Detailed description of the invention

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete Describe, it is clear that described embodiment is only a part of embodiment of the present invention rather than whole embodiments wholely.Based on Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under not making creative work premise Embodiment, broadly falls into the scope of protection of the invention.

A kind of method for tracking target schematic flow sheet that Fig. 1 provides for the embodiment of the present invention, as it is shown in figure 1, the present invention is real Execute a kind of method for tracking target that example is provided, comprise the following steps:

Step 101, selects first object position in image sequence the Ith frame, is tracked described first object, by institute The tracking result stating first object is defined as the first positive sample, by formula (1), determines from the multiple described first positive sample architecture The gradient of the first positive and negative sample packages extracted in the first negative sample bag that the first positive sample packages become and the first negative sample are constituted is special Levy；According to the Gradient Features of described first positive and negative sample packages, determine in the described first positive sample packages and described first negative sample bag Including the feature set of the first example collection, the feature set of described first example collection is defined as the feature square of the first training data Battle array；

Step 102, is determined by multi-instance learning framework, the eigenmatrix of described first training data, and formula (2) Oneth MIL 1-norm SVM, is classified, by public affairs to described first example collection by a described MIL 1-norm SVM Described first example is concentrated the described first example collection meeting first threshold to be configured to the first minimum support example collection by formula (3), The described first minimum weights supporting example collection are determined, according to the described first minimum weights supporting example collection by formula (4) With the first forgetting factor, determine described first minimum corresponding to MIL 1-norm SVM based on weights distribution by formula (5) Support the classification mark of example collection；

Step 103, selects candidate samples, by described MIL 1-norm SVM couple in image sequence I+1 frame Described candidate samples is classified, and according to described classification results and described formula (4), determines the second minimum support example collection Weights, according to the described second minimum weights supporting example collection, the described first minimum classification mark supporting example collection and formula (6) the second target location in described I+1 two field picture is determined；

Described formula (1) is as follows:

G=min (| g_x|+|g_y|,255)

Described formula (2) is as follows:

$y=sign\left(\underset{k\∈\mathrm{\Λ}}{\Σ}{w}_k^{*}g\right(B_i,x_k)+b^{*})$

Described formula (3) is as follows:

${\mathrm{\Λ}}_{j^{*}}=\left\{k:k\∈\mathrm{\Λ},x_{ij}\∈B_i,j^{*}=\arg {\max }_j{e}^{-\frac{\left|\right|x_{ij}-x_k\left|\right|}{{\mathrm{\σ}}^2}}\right\}$

Described formula (4) is as follows:

$h_t\left(x_s\right)=\underset{k\∈{\mathrm{\Λ}}_{j^{*}}}{\Σ}\frac{w_k^{*}s(x_k,x_s)}{m_k}$

Described formula (5) is as follows:

$H\left(x_s\right)=\underset{t}{\Σ}{e}^{-\frac{1}{t}}{h}_t\left(x_s\right)$

Described formula (6) is as follows:

$l_{t+1}^{*}=l\left(\arg {\max }_{x\∈X^{\mathrm{\γ}}}H\right(x\left)\right)$

Wherein, g is the Gradient Features of described first positive and negative sample packages, g_xFor described first positive and negative sample packages horizontal direction Gradient, g_yFor the gradient of described first positive and negative sample packages vertical direction,b^*Optimal value that is that be respectively weight vector and that bias, x_kFor kth sample, B_iFor arbitrary sample bag, y is discriminant function,For w^*The index of middle nonzero element,For With bag B_iRelevant first is minimum supports example indexed set, x_ijFor arbitrary sample bag B_iMiddle jth example,For index letter Number, σ²For the predefined factor, h_t(x_s) it is the first minimum classification weights supporting example, m_kRepresent and support example x_kMaximum Number, H (x_s) it is the first minimum classification mark supporting example collection,For corresponding forgetting factor, t is frame number,It is in t+1 Target location, l (x) is the position at sample place, x ∈ X^γFor the sample in the γ of region of search, for t+1 frame figure Picture, with t frame positionCentered by, γ=30 form candidate samples collection X for search radius sampling^γ, i.e.

Below as a example by 5 two field pictures, specifically introduce the method for tracking target that the embodiment of the present invention is provided.

In a step 101, initial training sample set, i.e. in image sequence I frame, institute according to embodiments of the present invention are constructed The method provided, for determining one training sample of I+1 frame target formation.

Initial frame at video sequence selects first object position, is tracked, the first object selected by the first mesh Target is followed the tracks of result and is defined as the first positive sample.

It should be noted that in embodiments of the present invention, initial frame in video sequence can be chosen as first object, also Second frame in video sequence can be chosen as first object, in embodiments of the present invention, to selecting first from video sequence The particular location of target does not do concrete restriction.

It should be noted that during owing to following the tracks of at every frame, only exist one and follow the tracks of the positive sample of result, i.e. only one of which This.Therefore, rotating the first positive sample, scaling change produces the first positive sample packages that the first positive sample is constituted；Further Ground, centered by present frame first object position, some each and every one bearing of sampling in the annular region in setting radius at random is shown Example, then can form the first negative sample bag.

Such as, the initial frame t=1 at video sequence manually chooses initial target locationFrom the 2nd frame of video sequence to 5th frame, i.e. t=2～n, (n=5), utilize tracker that target is tracked, using obtained tracking result as positive sample. During following the tracks of at every frame, only exist one and follow the tracks of the positive sample of result, i.e. only one of which.Therefore, to produced by every frame Positive sample carries out rotating, scaling change generation N number of positive sample composition positive closureWith present frame Tracing positionalCentered by, L the negative bag of negative example composition of sampling in the annular region of radius α ＜ ξ ＜ β at random

Further, from the first positive sample packages and the first negative sample bag, extract the Gradient Features of the first positive and negative sample packages.

In actual applications, it is that a kind of image that calculates converts in both horizontally and vertically gradient due to the Gradient Features of image The feature of information.Owing to the gradient information of target area and background area exists obvious difference, the therefore Gradient Features of image It it is the feature of a kind of good differentiation target and background.

In embodiments of the present invention, Gradient Features is got up with targeted measurement results, follow the tracks of during choose to The support example that track is important, thus realize the robust tracking of target.

In order to quickly estimate the targeted of all training examples in the tracking of many examples, in the embodiment of the present invention, first will just The unified image block being converted to 8 × 8 sizes of negative sample example, uses 1_D cover module [-1,0,1] to calculate each exemplary horizontal And gradient g of vertical direction_xAnd g_y, by formula (1), it is thus achieved that the Gradient Features of the first positive and negative sample packages.

G=min (| g_x|+|g_y|,255) (1)

Due to the first positive and negative sample packagesIn all examples Gradient Features form corresponding first example collection feature Collection And the feature set of the first example collection is determined It it is the eigenmatrix of the first training data.

Through above-mentioned initialization procedure, for front n frame video image, l the first positive sample bag and l the first negative sample can be constructed This bag, each first positive sample packages is made up of the N number of first positive sample instantiation, and the first negative sample bag is by L the first negative sample example Composition.Therefore, initialization procedure forms (N+L) × l the first example sample altogether.Corresponding by these the first example samples Gradient Features construct the eigenmatrix of the first training data by following equation (8):

$\left[\begin{array}{cccccc}g(B_1^{+},x_1)& \mathrm{...}& g(B_l^{+},x_1),& g(B_1^{-},x_N)& \mathrm{...}& g(B_l^{-},x_N)\\ g(B_1^{+},x_2)& \mathrm{...}& g(B_l^{+},x_2),& g(B_1^{-},x_{N+1})& \mathrm{...}& g(B_l^{-},x_{N+1})\\ \mathrm{...}& & \mathrm{...}& \mathrm{...}& & \mathrm{...}\\ g(B_1^{+},x_{N-1})& \mathrm{...}& g(B_l^{+},x_m),& g(B_1^{-},x_{N+L-1})& \mathrm{...}& g(B_l^{-},x_{N+L-1})\end{array}\right]---\left(8\right)$

In a step 102, based on multi-instance learning framework, the Gradient Features training 1-in the first positive and negative sample packages is utilized Norm SVM classifier, i.e. following equation (9) determine 1-norm SVM:

Y=sign (w^Tg+b) (9)

Wherein, w, b are the model parameter of grader, and g is the Gradient Features in positive and negative sample packages.

Further, in embodiments of the present invention, according to eigenmatrix and the formula (2) of the first training data, permissible Determine a MIL 1-norm SVM:

$y=sign\left(\underset{k\∈\mathrm{\Λ}}{\Σ}{w}_k^{*}g\right(B_i,x_k)+b^{*})---\left(2\right)$

In actual applications, for arbitrarily wrapping B_iIn sample instantiation x_k, can be by a MIL 1-norm SVM to it Classify.

Specifically, if wrapping B_iIn sample instantiation x_kTo in formula (2)Contribution more than or equal to certain (threshold value is set to one threshold value), then to example x_kGive positive label, otherwise, give negative label.In order to reduce many examples with The quantity of sample instantiation in track, it is achieved to target efficient and quickly follow the tracks of, uses the support extracted by 1-norm SVM The minimum of example constructions supports example collection.Therefore, minimum supports that example collection can pass through following equation (10) and determine:

$\mathrm{\ψ}=\left\{j^{*}:j^{*}=\arg {\max }_j{e}^{-\frac{\left|\right|g_{ij}-g_k\left|\right|}{{\mathrm{\σ}}^2}},k\∈\mathrm{\Λ},g_{ij}\∈B_i\right\}---\left(10\right)$

Wherein, g_ijRepresent bag B_iGradient Features corresponding to middle jth example, for bag B_iIn sample instantiation, if Example x_ij(j^*∈ ψ), then retain, it is for effectively supporting example.Otherwise,Then it is rightWithout contribution, lose Abandon x_ijExample.

In embodiments of the present invention, for any first example x_ij(j^*∈ ψ), formula (3) can be passed through, from the first example collection In meet the described first example collection of first threshold and be configured to the first minimum example collection of supporting:

${\mathrm{\Λ}}_{j^{*}}=\left\{k:k\∈\mathrm{\Λ},x_{ij}\∈B_i,j^{*}=\arg {\max }_j{e}^{-\frac{\left|\right|x_{ij}-x_k\left|\right|}{{\mathrm{\σ}}^2}}\right\}---\left(3\right)$

Through the above-mentioned first minimum selection supporting example collection, for example x_k, can determine that first is minimum further Support the Similarity value of example collection and already present support example Gradient Features, in embodiments of the present invention, can pass through following Formula (11) determines:

$s(B_i,x_k)=s\left(\right\{x_{{\mathrm{ij}}^{*}}\},x_k)=max\left(e^{-\frac{\left|\right|g_{{\mathrm{ij}}^{*}}-g_k|{|}^2}{{\mathrm{\σ}}^2}}\right)---\left(11\right)$

Wherein, example x_kWith bag B_iSimilarity s (B_i,x_k) it is defined as x_kWith bag B_iIn any exampleNearest example,And g_kRepresent bag B respectively_iMiddle jth example and example x_kGradient Features.

Further, the described first minimum weights supporting example collection are determined by following equation (4):

$h_t\left(x_s\right)=\underset{k\∈{\mathrm{\Λ}}_{j^{*}}}{\Σ}\frac{w_k^{*}s(x_k,x_s)}{m_k}---\left(4\right)$

Wherein, m_kRepresent and support example x_kSelected number of times.Can obtain from above formula, support example x calculating_sWeights time Consider x_kExample x is supported with existing_kSimilarity.

Further, based on the first minimum weights supporting example collection and the first forgetting factor, following equation can be passed through (5), the described first corresponding for MIL 1-norm SVM minimum classification mark supporting example collection based on weights distribution is determined:

$H\left(x_s\right)=\underset{t}{\Σ}{e}^{-\frac{1}{t}}{h}_t\left(x_s\right)---\left(5\right)$

Wherein, H (x_s) it is the first minimum classification mark supporting example collection, t represents the frame number relevant to supporting example, In the embodiment of the present invention, use the method (here, t=5) periodically selecting to support example,For forgetting factor, by forgeing The factor can realize the newly selected support example is given more high weight and retain tracking frame in the purpose of important support example.

In step 103, on-line tracing target.

In image sequence I+1 frame, with t frame positionCentered by, γ be radius region in sampling formed candidate's sample This collection X^γ, i.e.Wherein l_t+1X () is candidate samples x position at place in t+1 two field picture.Root The Gradient Features of all candidate samples is calculated according to formula (1).

Further, utilize a MIL 1-norm SVM of training in advance to all candidate example x ∈ X^γClassify, According to classification results and formula (4), determine the two the second minimum weights supporting example collection, minimum support example collection according to second With formula (6), determine the second target location in I+1 two field picture.

Wherein, formula (6) is as follows:

$l_{t+1}^{*}=l\left(\arg {\max }_{x\∈X^{\mathrm{\γ}}}H\right(x\left)\right)---\left(6\right)$

In embodiments of the present invention, after the second target location determining I+1 frame, in addition it is also necessary to according to I+1 frame Second target location, supports example collection to first determined according to I frame first object is minimum, and a MIL 1-norm SVM enters Row updates.

Specifically, the second target is tracked, the tracking result of the second target is defined as the second positive sample, from multiple The second negative sample bag that second positive sample packages of the second positive sample architecture one-tenth and the second negative sample are constituted extracts the second positive and negative sample The Gradient Features of this bag.

By formula (1), determine the feature set of all second examples that the second positive and negative sample packages includes, by all second The feature set of example is defined as the eigenmatrix of the second training data.

By a MIL 1-norm SVM, the second example collection is classified, by formula (3), the second example is concentrated Meet the second example collection of Second Threshold and be configured to the second minimum example collection of supporting, and minimum support example collection renewal the with second One minimum support example collection.

Eigenmatrix according to the second training data and formula (7), be updated a MIL 1-norm SVM.

Wherein, formula (7) is as follows:

$\underset{w,b,\mathrm{\ϵ},\mathrm{\η}}{min}\mathrm{\λ}\underset{k=1}{\overset{N+L-1}{\Σ}}\left|w_k\right|+C_1\underset{i=0}{\overset{l^{+}}{\Σ}}{\mathrm{\ϵ}}_i+C_2\underset{j=1}{\overset{l^{-}}{\Σ}}{\mathrm{\η}}_j$

$\begin{array}{c}s.t.(w^T{g}_i^{+}+b)\≥1-{\mathrm{\ϵ}}_i,i=0,\mathrm{...},l^{+}\\ -(w^T{g}_j^{-}+b)\≥1-{\mathrm{\η}}_j,j=0,\mathrm{...},l^{-}\\ {\mathrm{\ϵ}}_i,{\mathrm{\η}}_j\≥0\end{array}---\left(7\right)$

Wherein, λ is scale factor, ε_iFor the slack variable of positive sample packages feature, η_jLax change for negative sample bag feature Amount,For the constraint total amount of positive sample packages,For the constraint total amount of negative sample bag, C₁Punishment for lookup error sample is joined Number, C₂Penalty term parameter for negative error sample.

It should be noted that in embodiment of the present invention power, in each renewal process, retain l positive sample packages and l all the time Individual negative sample bag training 1-norm SVM, its update cycle is t=5.

In sum, in embodiments of the present invention, targeted measure based on Gradient Features estimates the first example Importance, utilizes multi-instance learning framework, on-line training 1-norm SVM, selects important support example, thus eliminate right Classify useless sample instantiation, thus effectively solve to cause classifier performance to decline due to a large amount of useless examples and algorithm calculates multiple The problem that miscellaneous degree is high.

Based on same inventive concept, embodiments provide a kind of target tracker, owing to this device solves skill The principle of art problem is similar to a kind of method for tracking target, and therefore the enforcement of this device may refer to the enforcement of method, repeats it Place repeats no more.

A kind of target tracker structural representation that Fig. 2 provides for the embodiment of the present invention.As in figure 2 it is shown, the present invention is real A kind of target tracker that executing example provides includes: first determines unit 21, and second determines unit 22, and the 3rd determines unit 23, 4th unit 24 and updating block 25.

First determines unit 21, for selecting first object position in image sequence the Ith frame, enters described first object Line trace, is defined as the first positive sample by the tracking result of described first object, by formula (1), determines from multiple described The the first positive and negative sample extracted in the first negative sample bag that first positive sample packages of one positive sample architecture one-tenth and the first negative sample are constituted The Gradient Features of this bag；According to the Gradient Features of described first positive and negative sample packages, determine the described first positive sample packages and described The feature set of the first example collection that one negative sample bag includes, is defined as the first training number by the feature set of described first example collection According to eigenmatrix；

Second determines unit 22, for by multi-instance learning framework, and the eigenmatrix of described first training data, and Formula (2) determines a MIL 1-norm SVM, is carried out described first example collection by a described MIL 1-norm SVM Classification, concentrates the described first example collection meeting first threshold to be configured to first described first example by formula (3) minimum Support example collection, determine the described first minimum weights supporting example collection by formula (4), show according to the described first minimum support The weights of example collection and the first forgetting factor, determine institute corresponding for MIL 1-norm SVM based on weights distribution by formula (5) State the first minimum classification mark supporting example collection；

3rd determines unit 23, for selecting candidate samples in image sequence I+1 frame, by a described MIL 1- Described candidate samples is classified by norm SVM, according to described classification results and described formula (4), determines the second ramuscule Hold the weights of example collection, according to the described second minimum weights supporting example collection, the described first minimum classification supporting example collection Mark and formula (6) determine the second target location in described I+1 two field picture；

Described formula (1) is as follows:

G=min (| g_x|+|g_y|,255)

Described formula (2) is as follows:

$y=sign\left(\underset{k\∈\mathrm{\Λ}}{\Σ}{w}_k^{*}g\right(B_i,x_k)+b^{*})$

Described formula (3) is as follows:

${\mathrm{\Λ}}_{j^{*}}=\left\{k:k\∈\mathrm{\Λ},j^{*}=\arg {\max }_j{e}^{-\frac{\left|\right|x_{ij}-x_k\left|\right|}{{\mathrm{\σ}}^2}}\right\}$

Described formula (4) is as follows:

$h_t\left(x_s\right)=\underset{k\∈{\mathrm{\Λ}}_{j^{*}}}{\Σ}\frac{w_k^{*}s(x_k,x_s)}{m_k}$

Described formula (5) is as follows:

$H\left(x_s\right)=\underset{t}{\Σ}{e}^{-\frac{1}{t}}{h}_t\left(x_s\right)$

Described formula (6) is as follows:

$l_{t+1}^{*}=l\left(\arg {\max }_{x\∈X^{\mathrm{\γ}}}H\right(x\left)\right)$

Wherein, g is the Gradient Features of described first positive and negative sample packages, g_xFor described first positive and negative sample packages horizontal direction Gradient, g_yFor the gradient of described first positive and negative sample packages vertical direction,b^*Optimal value that is that be respectively weight vector and that bias, x_kFor kth sample, B_iFor arbitrary sample bag, y is discriminant function,For w^*The index of middle nonzero element,For With bag B_iRelevant first is minimum supports example indexed set, x_ijFor arbitrary sample bag B_iMiddle jth example,For index letter Number, σ²For the predefined factor, h_t(x_s) it is the first minimum classification weights supporting example, m_kRepresent and support example x_kMaximum Number, H (x_s) it is the first minimum classification mark supporting example collection,For corresponding forgetting factor, t is frame number,It is in t+1 Target location, l (x) is the position at sample place, x ∈ X^γFor the sample in the γ of region of search, for t+1 frame figure Picture, with t frame positionCentered by, γ=30 form candidate samples collection X for search radius sampling^γ, i.e.

Preferably, the 4th unit 24 and updating block 25 are also included；

The tracking result of described second target, for being tracked described second target, is determined by described 4th unit 24 It is the second positive sample, by formula (1), the second positive sample packages and the second negative sample become from the multiple described second positive sample architecture The the second negative sample bag constituted extracts the Gradient Features of the second positive and negative sample packages；Gradient according to described second positive and negative sample packages Feature, determines the feature set of all second examples that described second positive and negative sample packages includes, by all described second examples Feature set is defined as the eigenmatrix of the second training data；By a described MIL 1-norm SVM to described second example collection Classify, concentrate the described second example collection meeting Second Threshold to be configured to described second example by described formula (3) Second minimum example collection of supporting, and support example collection with the described second minimum support example collection renewal described first is minimum；

Described updating block 25 is for the eigenmatrix according to described second training data and formula (7), to described first MIL 1-norm SVM is updated；

Formula (7) is as follows:

$\begin{array}{c}\underset{w,b,\mathrm{\ϵ},\mathrm{\η}}{min}\mathrm{\λ}\underset{k=1}{\overset{N+L-1}{\Σ}}\left|w_k\right|+C_1\underset{i=0}{\overset{l^{+}}{\Σ}}{\mathrm{\ϵ}}_i+C_2\underset{j=1}{\overset{l^{-}}{\Σ}}{\mathrm{\η}}_j\\ s.t.(w^T{g}_i^{+}+b)\≥1-{\mathrm{\ϵ}}_i,i=0,\mathrm{...},l^{+}\\ -(w^T{g}_j^{-}+b)\≥1-{\mathrm{\η}}_j,j=0,\mathrm{...},l^{-}\\ {\mathrm{\ϵ}}_i,{\mathrm{\η}}_j\≥0\end{array}.$

Wherein, λ is scale factor, ε_iFor the slack variable of positive sample packages feature, η_jLax change for negative sample bag feature Amount,For the constraint total amount of positive sample packages,For the constraint total amount of negative sample bag, C₁Punishment for lookup error sample is joined Number, C₂Penalty term parameter for negative error sample.

Preferably, described second determines that unit 22 is additionally operable to:

Determine that the described first minimum support example collection supports the similar of example Gradient Features to existing by following equation Angle value:

$s(B_i,x_k)=s\left(\right\{x_{{\mathrm{ij}}^{*}}\},x_k)=max\left(e^{-\frac{\left|\right|g_{{\mathrm{ij}}^{*}}-g_k|{|}^2}{{\mathrm{\σ}}^2}}\right)$

Wherein, example x_kWith bag B_iSimilarity s (B_i,x_k) it is defined as x_kWith bag B_iIn any exampleNearest example,And g_kRepresent bag B respectively_iMiddle jth example and example x_kGradient Features.

Preferably, described second determines that unit 22 is additionally operable to:

1-norm SVM is determined by following equation:

Y=sign (w^T g+b)

Wherein, w, b are corresponding weight vectors and biasing.

Should be appreciated that the function that according to the unit included by a kind of target tracker is only, this apparatus realizes is carried out Logical partitioning, in actual application, superposition or the fractionation of said units can be carried out.And a kind of target that this embodiment provides Function and a kind of method for tracking target one_to_one corresponding of above-described embodiment offer that device is realized are provided, real for this device institute Existing more detailed handling process, is described in detail in said method embodiment one, is not described in detail herein.

Those skilled in the art are it should be appreciated that embodiments of the invention can be provided as method, system or computer program Product.Therefore, the reality in terms of the present invention can use complete hardware embodiment, complete software implementation or combine software and hardware Execute the form of example.And, the present invention can use at one or more computers wherein including computer usable program code The upper computer program product implemented of usable storage medium (including but not limited to disk memory, CD-ROM, optical memory etc.) The form of product.

The present invention is with reference to method, equipment (system) and the flow process of computer program according to embodiments of the present invention Figure and/or block diagram describe.It should be understood that can the most first-class by computer program instructions flowchart and/or block diagram Flow process in journey and/or square frame and flow chart and/or block diagram and/or the combination of square frame.These computer programs can be provided Instruction arrives the processor of general purpose computer, special-purpose computer, Embedded Processor or other programmable data processing device to produce A raw machine so that the instruction performed by the processor of computer or other programmable data processing device is produced for real The device of the function specified in one flow process of flow chart or multiple flow process and/or one square frame of block diagram or multiple square frame now.

These computer program instructions may be alternatively stored in and computer or other programmable data processing device can be guided with spy Determine in the computer-readable memory that mode works so that the instruction being stored in this computer-readable memory produces and includes referring to Make the manufacture of device, this command device realize at one flow process of flow chart or multiple flow process and/or one square frame of block diagram or The function specified in multiple square frames.

These computer program instructions also can be loaded in computer or other programmable data processing device so that at meter Perform sequence of operations step on calculation machine or other programmable devices to produce computer implemented process, thus at computer or The instruction performed on other programmable devices provides for realizing at one flow process of flow chart or multiple flow process and/or block diagram one The step of the function specified in individual square frame or multiple square frame.

Although preferred embodiments of the present invention have been described, but those skilled in the art once know basic creation Property concept, then can make other change and amendment to these embodiments.So, claims are intended to be construed to include excellent Select embodiment and fall into all changes and the amendment of the scope of the invention.

Obviously, those skilled in the art can carry out various change and the modification essence without deviating from the present invention to the present invention God and scope.So, if these amendments of the present invention and modification belong to the scope of the claims in the present invention and equivalent technologies thereof Within, then the present invention is also intended to comprise these change and modification.

Claims (8)

1. a method for tracking target, it is characterised in that including:

In image sequence the Ith frame, select first object position, described first object is tracked, by described first object Follow the tracks of result and be defined as the first positive sample, by formula (1), determine the first positive sample become from the multiple described first positive sample architecture The Gradient Features of the first positive and negative sample packages extracted in the first negative sample bag that this bag and the first negative sample are constituted；According to described The Gradient Features of one positive and negative sample packages, determines the first example that the described first positive sample packages and described first negative sample bag include The feature set of collection, is defined as the eigenmatrix of the first training data by the feature set of described first example collection；

A MIL 1-is determined by multi-instance learning framework, the eigenmatrix of described first training data, and formula (2) Norm SVM, is classified to described first example collection by a described MIL 1-norm SVM, by formula (3) by described First example concentrates the described first example collection meeting first threshold to be configured to the first minimum support example collection, by formula (4) Determine the described first minimum weights supporting example collection, according to the described first minimum weights and first supporting example collection forget because of Son, determines the described first corresponding for MIL 1-norm SVM minimum support example collection based on weights distribution by formula (5) Classification mark；

Candidate samples is selected, by a described MIL 1-norm SVM to described candidate samples in image sequence I+1 frame Classify, according to described classification results and described formula (4), determine the second minimum weights supporting example collection, according to institute Stating the second minimum weights supporting example collection, the described first minimum classification mark supporting example collection and formula (6) determine described The second target location in I+1 two field picture；

Described formula (1) is as follows:

G=min (| g_x|+|g_y|,255)

Described formula (2) is as follows:

$y=sign\left(\underset{k\∈\mathrm{\Λ}}{\Σ}{w}_k^{*}g\right(B_i,x_k)+b^{*})$

Described formula (3) is as follows:

${\mathrm{\Λ}}_{j^{*}}=\{k:k\∈\mathrm{\Λ},x_{ij}\∈B_i,j^{*}=\arg {\max }_j{e}^{-\frac{\left|\right|x_{ij}-x_k\left|\right|}{{\mathrm{\σ}}^2}}\}$

Described formula (4) is as follows:

$h_t\left(x_s\right)=\underset{k\∈{\mathrm{\Λ}}_{j^{*}}}{\Σ}\frac{w_k^{*}s(x_k,x_s)}{m_k}$

Described formula (5) is as follows:

$H\left(x_s\right)=\underset{t}{\Σ}{e}^{-\frac{1}{t}}{h}_t\left(x_s\right)$

Described formula (6) is as follows:

$l_{t+1}^{*}=l\left(\arg {\max }_{x\∈X^{\mathrm{\γ}}}H\right(x\left)\right)$

Wherein, g is the Gradient Features of described first positive and negative sample packages, g_xFor the gradient of described first positive and negative sample packages horizontal direction, g_yFor the gradient of described first positive and negative sample packages vertical direction,b^*Optimal value that is that be respectively weight vector and that bias, x_kFor Kth sample, B_iFor arbitrary sample bag, y is discriminant function,For w^*The index of middle nonzero element,For with bag B_iRelevant first is minimum supports example indexed set, x_ijFor arbitrary sample bag B_iMiddle jth example,For exponential function, σ² For the predefined factor, h_t(x_s) it is the first minimum classification weights supporting example, m_kRepresent and support example x_kMaximum number, H (x_s) it is the first minimum classification mark supporting example collection,For corresponding forgetting factor, t is frame number,It it is the mesh in t+1 Cursor position, l (x) is the position at sample place, x ∈ X^γFor the sample in the γ of region of search, for t+1 two field picture, With t frame positionCentered by, γ=30 form candidate samples collection X for search radius sampling^γ, i.e.

2. the method for claim 1, it is characterised in that described according to the described second minimum weights supporting example collection, The described first minimum classification mark supporting example collection and formula (6) determine the second target location in described I+1 two field picture Afterwards, also include:

Described second target is tracked, the tracking result of described second target is defined as the second positive sample, passes through formula (1) the second negative sample bag that the second positive sample packages, become from the multiple described second positive sample architecture and the second negative sample are constituted Extract the Gradient Features of the second positive and negative sample packages；According to the Gradient Features of described second positive and negative sample packages, just determining described second The feature set of all second examples that negative sample bag includes, is defined as the second training by the feature set of all described second examples The eigenmatrix of data；Described second example collection is classified, by described public affairs by a described MIL 1-norm SVM Described second example is concentrated the described second example collection meeting Second Threshold to be configured to the second minimum support example collection by formula (3), And update the described first minimum support example collection with the described second minimum support example collection；

Eigenmatrix according to described second training data and formula (7), be updated a described MIL 1-norm SVM；

Formula (7) is as follows:

$\underset{w,b,\mathrm{\ϵ},\mathrm{\η}}{min}\mathrm{\λ}\underset{k=1}{\overset{N+L-1}{\Σ}}\left|w_k\right|+C_1\underset{i=0}{\overset{l^{+}}{\Σ}}{\mathrm{\ϵ}}_i+C_2\underset{j=1}{\overset{l^{-}}{\Σ}}{\mathrm{\η}}_j$

$\begin{array}{c}s.t.(w^T{g}_i^{+}+b)\≥1-{\mathrm{\ϵ}}_i,i=0,...,l^{+}\\ -\left(w^T{g}_j^{-}+b\right)\≥1-{\mathrm{\η}}_j,j=0,...,l^{-}\end{array}$

ε_i,η_j≥0 (7)

Wherein, λ is scale factor, ε_iFor the slack variable of positive sample packages feature, η_jFor the slack variable of negative sample bag feature, For the constraint total amount of positive sample packages,For the constraint total amount of negative sample bag, C₁For the punishment parameter of lookup error sample, C₂It is negative The penalty term parameter of error sample.

3. the method for claim 1, it is characterised in that described determine that described first minimum support is shown by formula (4) Before the weights of example collection, also include:

Determine the described first minimum support example collection by following equation and there is the Similarity value supporting example Gradient Features:

$s(B_i,x_k)=s\left(\right\{x_{{\mathrm{ij}}^{*}}\},x_k)=max\left(e^{-\frac{\left|\right|g_{{\mathrm{ij}}^{*}}-g_k|{|}^2}{{\mathrm{\σ}}^2}}\right)$

Wherein, example x_kWith bag B_iSimilarity s (B_i,x_k) it is defined as x_kWith bag B_iIn any exampleNearest example,And g_k Represent bag B respectively_iMiddle jth example and example x_kGradient Features.

4. the method for claim 1, it is characterised in that described determine a described MIL 1-norm SVM before, also Including:

1-norm SVM is determined by following equation:

Y=sign (w^Tg+b)

Wherein, w, b are corresponding weight vectors and biasing.

5. a target tracker, it is characterised in that including:

First determines unit, in image sequence the Ith frame select first object position, described first object is carried out with Track, is defined as the tracking result of described first object the first positive sample, by formula (1), from multiple described first just determines The the first positive and negative sample packages extracted in the first negative sample bag that first positive sample packages of sample architecture one-tenth and the first negative sample are constituted Gradient Features；According to the Gradient Features of described first positive and negative sample packages, determine the described first positive sample packages and described first negative The feature set of the first example collection that sample packages includes, is defined as the first training data by the feature set of described first example collection Eigenmatrix；

Second determines unit, for by multi-instance learning framework, the eigenmatrix of described first training data, and formula (2) determine a MIL 1-norm SVM, by a described MIL 1-norm SVM, described first example collection classified, Concentrate the described first example collection meeting first threshold to be configured to the first minimum support described first example by formula (3) to show Example collection, determines the described first minimum weights supporting example collection by formula (4), according to the described first minimum support example collection Weights and the first forgetting factor, determine corresponding for MIL 1-norm SVM described first based on weights distribution by formula (5) The minimum classification mark supporting example collection；

3rd determines unit, for selecting candidate samples in image sequence I+1 frame, by a described MIL 1-norm Described candidate samples is classified by SVM, according to described classification results and described formula (4), determines that the second minimum support is shown The weights of example collection, according to the described second minimum weights supporting example collection, the described first minimum classification mark supporting example collection The second target location in described I+1 two field picture is determined with formula (6)；

Described formula (1) is as follows:

G=min (| g_x|+|g_y|,255)

Described formula (2) is as follows:

$y=sign\left(\underset{k\∈\mathrm{\Λ}}{\Σ}{w}_k^{*}g\right(B_i,x_k)+b^{*})$

Described formula (3) is as follows:

${\mathrm{\Λ}}_{j^{*}}=\{k:k\∈\mathrm{\Λ},j^{*}=\arg {\max }_j{e}^{-\frac{\left|\right|x_{ij}-x_k\left|\right|}{{\mathrm{\σ}}^2}}\}$

Described formula (4) is as follows:

$h_t\left(x_s\right)=\underset{k\∈{\mathrm{\Λ}}_{j^{*}}}{\Σ}\frac{w_k^{*}s(x_k,x_s)}{m_k}$

Described formula (5) is as follows:

$H\left(x_s\right)=\underset{t}{\Σ}{e}^{-\frac{1}{t}}{h}_t\left(x_s\right)$

$H\left(x_s\right),e^{-\frac{1}{t}},$

Described formula (6) is as follows:

$l_{t+1}^{*}=l\left(\arg {\max }_{x\∈X^{\mathrm{\γ}}}H\right(x\left)\right)$

Wherein, g is the Gradient Features of described first positive and negative sample packages, g_xFor the gradient of described first positive and negative sample packages horizontal direction, g_yFor the gradient of described first positive and negative sample packages vertical direction,b^*Optimal value that is that be respectively weight vector and that bias, x_kFor Kth sample, B_iFor arbitrary sample bag, y is discriminant function,For w^*The index of middle nonzero element,For with bag B_iRelevant first is minimum supports example indexed set, x_ijFor arbitrary sample bag B_iMiddle jth example,For exponential function, σ² For the predefined factor, h_t(x_s) it is the first minimum classification weights supporting example, m_kRepresent and support example x_kMaximum number, H (x_s) it is the first minimum classification mark supporting example collection,For corresponding forgetting factor, t is frame number,It it is the mesh in t+1 Cursor position, l (x) is the position at sample place, x ∈ X^γFor the sample in the γ of region of search, for t+1 two field picture, With t frame positionCentered by, γ=30 form candidate samples collection X for search radius sampling^γ, i.e.

6. device as claimed in claim 5, it is characterised in that also include the 4th unit and updating block；

The tracking result of described second target, for being tracked described second target, is defined as second by described Unit the 4th Positive sample, is consisted of formula (1), the second positive sample packages become from the multiple described second positive sample architecture and the second negative sample Second negative sample bag extracts the Gradient Features of the second positive and negative sample packages；According to the Gradient Features of described second positive and negative sample packages, Determine the feature set of all second examples that described second positive and negative sample packages includes, by the feature set of all described second examples It is defined as the eigenmatrix of the second training data；Described second example collection carried out point by a described MIL 1-norm SVM Class, is concentrated described second example by described formula (3) and meets the described second example collection of Second Threshold and be configured to second Little support example collection, and update the described first minimum support example collection with the described second minimum support example collection；

Described updating block is for the eigenmatrix according to described second training data and formula (7), to a described MIL 1- Norm SVM is updated；

Formula (7) is as follows:

$\underset{w,b,\mathrm{\ϵ},\mathrm{\η}}{min}\mathrm{\λ}\underset{k=1}{\overset{N+L-1}{\Σ}}\left|w_k\right|+C_1\underset{i=0}{\overset{l^{+}}{\Σ}}{\mathrm{\ϵ}}_i+C_2\underset{j=1}{\overset{l^{-}}{\Σ}}{\mathrm{\η}}_j$

$\begin{array}{c}s.t.(w^T{g}_i^{+}+b)\≥1-{\mathrm{\ϵ}}_i,i=0,...,l^{+}\\ -\left(w^T{g}_j^{-}+b\right)\≥1-{\mathrm{\η}}_j,j=0,...,l^{-}\end{array}.$

ε_i,η_j≥0

Wherein, λ is scale factor, ε_iFor the slack variable of positive sample packages feature, η_jFor the slack variable of negative sample bag feature, For the constraint total amount of positive sample packages,For the constraint total amount of negative sample bag, C₁For the punishment parameter of lookup error sample, C₂It is negative The penalty term parameter of error sample.

7. device as claimed in claim 5, it is characterised in that described second determines that unit is additionally operable to:

Determine the described first minimum support example collection by following equation and there is the Similarity value supporting example Gradient Features:

$s(B_i,x_k)=s\left(\right\{x_{{\mathrm{ij}}^{*}}\},x_k)=max\left(e^{-\frac{\left|\right|g_{{\mathrm{ij}}^{*}}-g_k|{|}^2}{{\mathrm{\σ}}^2}}\right)$

Wherein, example x_kWith bag B_iSimilarity s (B_i,x_k) it is defined as x_kWith bag B_iIn any exampleNearest example,And g_k Represent bag B respectively_iMiddle jth example and example x_kGradient Features.

8. device as claimed in claim 5, it is characterised in that described second determines that unit is additionally operable to: true by following equation Determine 1-norm SVM:

Y=sign (w^Tg+b)

Wherein, w, b are corresponding weight vectors and biasing.