CN105513093B - A kind of method for tracking target represented based on low-rank matrix - Google Patents
A kind of method for tracking target represented based on low-rank matrix Download PDFInfo
- Publication number
- CN105513093B CN105513093B CN201510916027.XA CN201510916027A CN105513093B CN 105513093 B CN105513093 B CN 105513093B CN 201510916027 A CN201510916027 A CN 201510916027A CN 105513093 B CN105513093 B CN 105513093B
- Authority
- CN
- China
- Prior art keywords
- template
- target
- matrix
- dictionary
- global
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Abstract
The invention discloses a kind of method for tracking target represented based on low-rank matrix, belong to technical field of image processing, solve the problems, such as in conventional method to track under the conditions of target occlusion etc. unstable.The present invention not only carries out low-rank matrix expression to global characteristics, also low-rank matrix expression is carried out to the local feature for tracking target, so that the description to target, which not only contains global characteristics, also contains local feature so that the tracking to target has more preferable robustness.The method individually handled particle different from rarefaction representation, present invention utilizes interparticle similitude, and the coefficient vector of each particle is formed into coefficient matrix, and has added the minimum limitation of order to coefficient matrix, so as to reduce the operand of algorithm.And in order to suppress the drift phenomenon in object tracking process, the present invention also incorporates the background template with target range farther out, the target as big as possible by finding background template coefficient To Template as small as possible, reaches more preferable tracking effect.
Description
Technical field
The invention belongs to pattern-recognition and computer vision field, more particularly to a kind of target following based on low-rank matrix
Method.
Background technology
As the fast development of computer technology and " intelligent earth ", " intelligent transportation ", " smart home " etc. are general
The it is proposed of thought, people are filled with expectation to following intelligentized life, accelerate the cry of intelligent construction also more and more higher.Intelligence
The main thought of change is to replace manpower using computer, and the tasks such as observation, judgement, early warning are automatically performed under special scenes,
So as to save human resources, convenient daily life.Under intelligentized concept, the task thousand that computer can be completed is poor
Ten thousand are not, however, the specific tasks of computer, are required for first completing to the external world with the knowledge of computer vision field why
Observation, therefore technological accumulation and development of the intelligent construction to computer vision field propose higher requirement.
Important topic of the target following as computer vision field, suffers from many scientific research personnel sight is special always
Note is thereon.In general, the target following based on video either image sequence is by handling video or image, to target
Detected, extracted, identified and tracked, obtain the kinematic parameter (such as speed, position, the anglec of rotation) of target, lock onto target
Movement locus, so as to realize the behavior for further understanding target, or even complete more advanced task.Therefore, target following
It is of great significance in computer vision field tool.The effect of only target following is stabilized, and can just be made follow-up all
Such as the task of behavior understanding higher level is carried out.
Target following also has an extensive prospect in actual applications, and it is in security monitoring, man-machine interaction, military affairs, medical science
Even more it has been widely used Deng field.The today in each corner in life is almost spread in camera, we are to intelligent prison
The demand of control has reached peak since the dawn of human civilization.If each monitoring camera-shooting is required for manpower to go to observe, it will causes huge
Waste of human resource, and intelligent monitoring can be handled automatically the monitor video of these magnanimity, is reached us and is wanted to reach
The purpose arrived.Regardless of whether what the objectives of intelligent monitoring are, the realization of its effect all must be set up successful in target following
On the basis of.Such as the intelligent safety monitoring of old solitary people is needed first to realize the tracking to old man, afterwards again to tracking Objective extraction
The behavior that relevant information carries out old man judges, so as to notify medical personnel in time when accidental falls occur for old man；
, it is necessary to using tracking technique, to make the body language of people, expression etc. turn into controlling soil moist, so as to realize one in terms of man-machine interaction
The function of series, brings the splendid Consumer's Experience of client；Medically, carried out by the cell to patient's diseased region or region
Long-term follow, doctor can be helped more rapidly accurately to treat disease；Militarily, by tracking technique come the flight to guided missile
It is controlled, and quick and precisely positions and play an important roll in unfriendly target.Therefore, the research of tracking technique has very heavy
The realistic meaning wanted.
The content of the invention
For prior art, the present invention provides a kind of method for carrying out target following, and it can be to the mesh under complex scene
Mark is tracked, and has stronger robustness.
To achieve the above object, the following technical scheme that the present invention uses：
A kind of method for tracking target represented based on low-rank matrix, it is characterised in that comprise the following steps：
Step 1：Initial target location is chosen in the 1st frame, To Template and background mould are chosen according to initial target location
Plate, piecemeal processing first is carried out to To Template, affine transformation then is carried out to To Template and background template, completes local dictionary
With the initialization of Global Dictionary；
Step 2：According to the target location of s-1 frames, the tracking target T of selection s frames, then the tracking target by s frames
T affine transformations are the block with atom formed objects in Global DictionaryAnd obtain state parameter Fs, s>=2；
Step 3：With s frame dbjective state parameters FsFor average, δ2Grain is gathered for the Gaussian Profile of variance in s frames at random
Son, piecemeal processing and affine transformation are carried out to particle, matrix in block form and global matrix is built, each piecemeal is represented with local dictionary
Matrix, obtain corresponding local coefficient matrix；Global matrix is represented with Global Dictionary, obtains global coefficient matrix；
Step 4：The total weights omega of each particle is obtained by global coefficient matrix and local coefficient's matrix computationsi, right basis
Particle coordinate positionAnd weights omegai, obtain s frames target locationN represents total number of particles；
Step 5：According to the target location of s frames, local dictionary and Global Dictionary are updated；
Step 6：S=s+1, go to step 2.
In above-mentioned technical proposal, step 1 comprises the following steps：
S11：Choose To Template and background template
Using p pixel as step-length, centered on the L of target location, most step-lengths are moved up and down in vertical direction,
Most step-lengths are moved left and right in horizontal direction, so as to get 8 with target sizes shape identical block as template, target
Itself a template equally is used as, is finally taken at random in target proximity (the length and width minimum value for being no more than target with target location) again
M block forms the To Template set { E formed with m+9 template as template1,E2,…Em+9, not less than 1.5 times of d's
Position randomly selects n size with target identical block as background template set { B1,B2,…,Bn, d is that target is grown with width most
Small value,
S12：Piecemeal processing is carried out to To Template
By To Template Ej(j ∈ [1, m+9]) is divided into 3*3 fritter, and numbers, then by each fritter with column vector
Form represents, obtainsWherein, each vectorial element value is the gray value of corresponding pixel points；
S13：Affine transformation is carried out to To Template and background template
To To Template Ej(j ∈ [1, m+9]) and background template Bk(k ∈ [1, n]) carry out affine transformation, obtain with S12
The block of fritter formed objects, and it is represented in the form of vectors, obtainWith
S14：Complete the initialization of local dictionary and Global Dictionary
With all To Templates and the affine transformation fritter of background templateWithTo initialize Global DictionaryObtain
The Global Dictionary of initializationAgain by each To Template Ej(j∈m+
9) fritterTo initialize local dictionary DS, so as to the local dictionary initialized
In above-mentioned technical proposal, the dbjective state parameter F of step 2sDetermined by six affine coefficients,WhereinTarget x direction displacements, the displacement of y directions, yardstick are represented respectively
Change, the anglec of rotation, the ratio of width to height and chamfer angle.
In above-mentioned technical proposal, step 3 comprises the following steps：
S31：With s frame dbjective state parameters FsFor average, δ2Particle collection is randomly selected for the Gaussian Profile of varianceN is particle collection number；
S32：By To Template Ej(j ∈ [1, m+9]) is divided into 3*3 size identical fritter and numbered, and then will
Each fritter is represented in the form of column vector, is obtainedThe element value of vector is the gray scale of corresponding pixel points
Value；
Reuse and affine transformation is carried out to particle with S13 same operations, obtain
S33：The result after all particle piecemeals, that is, the fritter after dividing equallyExtract,
Wherein i ∈ [1, N] represent the sequence number of particle, and l ∈ [1,9] represent the label of fritter, then willGroup
It is combined into corresponding matrix in block formIt is expressed as with formulaWherein, N represents that particle is total
Number, then all particlesFor forming global matrix
S34：Each matrix in block form is represented with local dictionary, corresponding local coefficient matrix is obtained, represents complete with Global Dictionary
Office's coefficient matrix, obtains global coefficient matrix
In above-mentioned technical proposal, the computational methods of particle weights are as follows in step 4：
S41：From global coefficient matrixIn by row propose that the global of i-th particle represents coefficient column vectorWhereinRepresenting matrixThe i-th row, then useIn preceding m+9 element absolute value and subtract after the vector n it is first
The absolute value sum of element, obtains the global weight of the particle
S42：From local coefficient's matrixIn the local dictionary of i-th particle extracted by row represent
Coefficient column vectorWhereinRepresenting matrixI-th row, then take out the coefficient that fritter is corresponded in column vector, i.e.,In l, 9+l, 9*2+l ..., 9* (m+9-1)+l elements, m+9 is To Template number, seeks their absolute value
Sum, then the absolute value sum of other elements is subtracted, obtain partial weightI.e.
S43：Obtained according to S41 and S42WithCalculate total weight of particle Wherein parameter σ determines the importance accounting of global characteristics and local feature on target similarity measurement；
S44：Total weight of all particles is normalized, then according to particleCoordinate positionAnd weight
ωi, obtain s frames target location LS,
In above-mentioned technical proposal, the update method of dictionary is as follows in step 5：
S51：A cumulative probability sequence is generated to represent the accumulative update probability of each template：
Wherein n represents To Template sum, then random one section of generation [0,1) on number r, judge that r is in above formula
Which section of middle sequence, ifThen delete j-th of template E in dictionaryjCorresponding local dictionary Ds
In fritterThen the tracking result of present frame is divided into 9 fritters, then gone out with column vector form table, i.e.,
{t1,t2,…,t9, finally it is added into local dictionary DsEnd, now
S52：For background template, all strategies updated of every frame are taken, i.e., in each frame, all back ofs the body of the previous frame of deletion
Scape template, then according to the tracking result of present frame, background template is chosen again；
For Global DictionaryFor, the renewal of its forward part To Template is similar with local dictionary updating method, according to
The section where random number r is generated, deletes some template, word is added after the tracking result of present frame then is carried out into affine transformation
Allusion quotation is last, nowWhereinFor tracking result
It is affine to change obtained fritter,Q=1,2 ..., n, background template is achieved for present frame.
Compared with prior art, beneficial effects of the present invention are shown：
Global characteristics are not only carried out rarefaction representation by the first, method of the invention based on rarefaction representation, also to tracking target
Local feature carries out rarefaction representation, so that the description to target, which not only contains global characteristics, also contains local feature,
So that the tracking to target has more preferable robustness.Particularly in the case where locally blocking, shield portions its is removed
His fritter also can be good at being indicated target, so as to greatly improve the stability of tracking.
2nd, the present invention is individually handled particle different from general sparse representation method, but between make use of particle
Potential similitude, it is that a matrix considers by all particle combinations.Because interparticle similitude is in dictionary expression system
The low-rank characteristic of coefficient matrix is shown as on matrix number, our coefficient matrixes minimum by solving order, to obtain each particle
Dictionary represents coefficient.Than the sparse representation method to the independent computing of each particle, present invention reduces the operand of algorithm, and
Obtain the likelihood ratio measurement between more accurate candidate target and time of day.
3rd, dictionary of the invention not use only the To Template for having actively impact to target, also incorporate with target away from
From background template farther out.Under the background for emphasizing rarefaction representation, it is desirable to which the coefficient of background template is as small as possible, so as to suppressing
Drift phenomenon in object tracking process has very obvious action.
Brief description of the drawings
Fig. 1 is six steps of the target following provided by the invention represented based on low-rank matrix；
Fig. 2 To Templates gather schematic diagram；
Fig. 3 background templates gather schematic diagram；
Fig. 4 target segment schematic diagrames；
The good Global Dictionary index contrasts with the candidate target of difference of Fig. 5；
The local dictionary index contrast of Fig. 6 same target difference fritters；
The Global Dictionary index contrast of the good fritters corresponding with the candidate target of difference of Fig. 7；
Pay attention to：Fig. 2,5,6,7 are a secondary accompanying drawing, the situation in the absence of several accompanying drawings using 1 label.
Embodiment
To describe the technology contents of the present invention, construction feature, the objects and the effects in detail, below in conjunction with embodiment
And accompanying drawing is coordinated to be explained in detail.
Embodiment
As shown in figure 1, the method for tracking target represented based on low-rank matrix, is specifically comprised the following steps：
Step 1：In the 1st frame, To Template and background template are chosen according to initial target location, first To Template entered
The processing of row piecemeal, then carries out affine transformation to To Template and background template, completes the initial of local dictionary and Global Dictionary
Change.The tracking result of initial target location initial frame the most.Afterwards, it is switched to the 2nd frame；
Step 2：Tracking target T using the tracking result of s-1 frames as s frames, then the tracking target T of s frames is imitated
Penetrate the block being transformed to atom formed objects in dictionaryAnd obtain state parameter Fs, s>=2；
Step 3：With s frame dbjective state parameters FsFor average, δ2Grain is gathered for the Gaussian Profile of variance in s frames at random
Son, piecemeal processing and affine transformation are carried out to particle, matrix in block form and global matrix is built, each piecemeal is represented with local dictionary
Matrix, obtain corresponding local coefficient matrix；Global matrix is represented with Global Dictionary, obtains global coefficient matrix；
Step 4：The total weights omega of each particle is obtained by global coefficient matrix and local coefficient's matrix computationsi, right basis
Particle coordinate positionAnd weights omegai, obtain s frames target location Ls,
Step 5：According to the tracking result of s frames, local dictionary and Global Dictionary are updated；
Step 6：S=s+1, go to step 2.
In the present embodiment, step 1 includes step S11, S12, S13 and S14, described in text specific as follows：
S11：Choose To Template
In this step, as shown in Fig. 2 using 3 pixels as step-length, centered on the L of target location, in vertical direction
On move up and down most step-lengths, move left and right most step-lengths in the horizontal direction, it is big with target so as to get eight
Small shape identical block is equally used as a template, finally takes a block at random in target proximity again in itself as template, target
As template, the To Template set { E formed with ten templates is formed1,E2,…,E10}.Again in the place remote enough from target
10 sizes are randomly selected with target identical block as background template set { B1,B2,…,B10, as shown in Figure 3.
S12：Piecemeal processing is carried out to To Template
In this step, by To Template EjIt is divided into 3*3 fritter as shown in Figure 4, and numbers.Then by each fritter
Represented, obtained in the form of column vectorThe element value of vector is the gray value of corresponding pixel points
S13：Affine transformation is carried out to To Template and background template
In this step, to To Template EjWith background template BkAffine transformation is carried out, is obtained identical with fritter in S12
The block of size, and it is represented in the form of vectors, obtainWith
S14：Complete the initialization of local dictionary and Global Dictionary
In this step, with all To Templates and the affine transformation fritter of background templateWithTo initialize the overall situation
DictionaryAgain by each To Template EjFritter
To initialize local dictionary DS, so as to obtain local dictionary
In the present embodiment, the dbjective state parameter F of step 2sDetermined by six affine coefficients,WhereinTarget x direction displacements, the displacement of y directions, yardstick are represented respectively
Change, the anglec of rotation, the ratio of width to height and chamfer angle.
In the present embodiment, step 3 includes step S31, S32, S33 and S34, and concrete operations are as described below：
S31：In s frames, particle is gathered
In this step, if directly according to the station acquisition particle of target, target can not be taken into account and rotated
Situation, and if particle is gathered according to its affine parameter, the situations of change such as the stretching of target, rotation can all be taken into account,
So as to preferably gather candidate target.Therefore, herein with s frame dbjective state parameters FsFor average, δ2For the Gauss of variance
S frames are distributed in randomly select particle collection(N is particle collection number).And for population N, if N is bigger,
Particle diversity will be enriched all the more, and so as to can more describe the time of day of present frame target, final tracking effect also will more
It is good.But then, particle number is more, longer the time required to program operation, so as to influence the real-time of target following.Cause
This, the choosing selection of population must take into account two aspects of effect and efficiency.It our experiments show that, it is not that population N is set into 600
Wrong selection.
S32：Piecemeal processing and affine transformation are carried out to particle
In this step, take and particle is divided into 9 fritters with S12 same operations, obtain
Reuse and affine transformation is carried out to particle with S13 same operations, obtain
S33：Build matrix in block form and global matrix
In this step, all particlesExtract, be combined as corresponding piecemeal
MatrixIt is expressed as with formulaAgain all particlesFor forming the overall situation
Matrix
S34：Each matrix in block form is represented with local dictionary, obtains corresponding local coefficient matrix；Represent complete with Global Dictionary
Office's matrix, global coefficient matrix is obtained (for Unify legislation, next using X come the matrix in block form in representative instanceAnd the overall situation
MatrixD represents local dictionary D1And Global Dictionary
Represent that the coefficient matrix that particle obtains is not unique because dictionary was complete, therefore by dictionary.And
According to the optic nerve principle of the mankind, it is desirable to which coefficient matrix is sparse.Again due to particle be all around target it is close away from
From selection, there is similitude between particle, the similitude shows as the low-rank of matrix on coefficient matrix, considered further that noise situations
Under little deviation, the solution to coefficient matrix is to following formula subrepresentation：
Wherein, ‖ A ‖0,0The non-zero entry number of representing matrix A each columns, Rank (A) representing matrixs A order, Z are X on D's
Coefficient matrix, ∈ represent to represent X error by dictionary D.By some processing mathematically, the problem is converted into as follows most
The solution of optimization problem：
Wherein, ‖ A ‖*The nuclear norm of representing matrix, ‖ A ‖1,1The absolute value sum of representing matrix A all elements, i.e.,
‖A‖1,1=| a11|+|a21|+…+|am1|+|a12|+…+|am2|+…+|amn|
Formula (2) can be solved with IALM (Inexact Augmented Lagrange Multiplier) method,
Idiographic flow is as follows：
(1) X, D and ρ, μ are inputted1, μ2, μ3, and initialize Z3=0, ∈=0, Y1=0, Y2=0, Y3=0, ρ are taken more than 0
Meaning number, μ1, μ2, μ3Take more than 0 any number.Wherein ρ is that Lagrange multiplier updates coefficient, Y1, Y2, Y3Multiply for Lagrange
Son, μ1, μ2, μ3For penalty coefficient
(2) keep other specification constant, update Z1：
Wherein,
Decompose to obtain by the SVD of matrix A,
(3) keep other specification constant, update Z2：
(4) keep other specification constant, update ∈：
(5) keep other specification constant, update Z3：
Wherein,
(6) each multiplier and parameter are updated：
Y1=Y1+μ1(X-DZ3) formula (7)
Y2=Y2+μ2(Z3-Z1) formula (8)
Y3=Y3+μ3(Z3-Z2) formula (9)
μ1=ρ μ1Formula (10)
μ2=ρ μ2Formula (11)
μ3=ρ μ3Formula (12)
(7) Z=Z is made3, judge whether to restrain, if not restraining, jump to (2), if convergence, is solved
, can be according to matrix in block form according to this flowWith global matrixSo as to obtain sparse and low-rank local coefficient
MatrixWith global coefficient matrix
In the present embodiment, step 4 contains step S41, S42, S43 and S44, described in text specific as follows：
S41：Calculate particle overall situation weight
Due to Global DictionaryIt is made up of To Template and background template, therefore its coefficient is also classified into target factor and background system
Number.Difference between candidate target and the candidate target of difference that Fig. 5 has been disclosed between background coefficient and target factor.In figure,
The candidate target that solid box has represented, dotted line frame represent the candidate target of difference, the word of the good candidate target of upper right statistics with histogram
Allusion quotation represents coefficient, and the dictionary of the candidate target of lower right histogram statistical difference represents coefficient, and in histogram first 10 be target mould
Plate coefficient, latter 10 are background template coefficient.The candidate target that can be seen that according to histogram in Fig. 5 should be all by mesh
Mark template representation and unrelated with background template, i.e. nonzero coefficient is all distributed in To Template correspondence position, and background template is corresponding
Coefficient is almost 0.Therefore, we are first from global coefficient matrixIn by row propose i-th particle it is global represent coefficient arrange to
AmountWherein (A)iRepresenting matrix A the i-th row.Use againIn preceding m element absolute value and subtract the vector
The absolute value sum of n element afterwards, obtain the global weight of the particleWherein n is To Template number, and m is background template
Number, m=n=10 in this example
S42：Calculate particle partial weight
Local dictionary DsIncluding all fritters of To Template are all included, therefore the local dictionary of each fritter represents coefficient
All it is a longer column vector.Fig. 6 is disclosed between the local dictionary expression coefficient of different fritters in same candidate target
The similarities and differences.In figure, left figure is a good candidate target, wherein No. 1 frame and No. 2 frames represent two fritters in target respectively,
Its coefficient is represented with local dictionary, and coefficient is showed with two histograms in the right, wherein, upper right histogram represents No. 1
Fritter coefficient, lower right histogram represent No. 2 fritter coefficients.Contrasted by two histograms it can be found that for a preferably tracking
Target, the coefficient of its fritter is all than sparse, but different fritters, all differences in position that its larger coefficient occurs.And pass through reason
By analysis, ideally, l-th of fritter of i-th of particleShould only as corresponding to To Template fritter, i.e., each mesh
L-th of fritter for marking template represents that therefore, its larger coefficient should only appear in l-th of fritter of each To Template
On position, and coefficient corresponding to other fritters should be smaller or for 0, therefore what different fritters its larger coefficient occurred in histogram
Position is different, obtains and Fig. 6 identical conclusions.
The corresponding fritter part dictionary between the candidate target of difference that Fig. 7 is used to disclose represents the difference of coefficient.Figure
In, candidate target that No. 1 frame has represented, No. 2 frames represent the candidate target of difference, and the 2nd fritter for calculating two targets respectively is local
Dictionary coefficient, and contrasted with statistics with histogram, wherein the histogram in upper right side represents that the fritter of good candidates target is local
Dictionary represents coefficient, and the histogram of lower right represents that the fritter part dictionary of the candidate target of difference represents coefficient.It is straight by contrasting
Square graph discovery, the local dictionary of good its fritter of candidate target represent that coefficient is sparse, and the distribution of its coefficient of the candidate target of difference compared with
It is average, and fritter is corresponded in its dictionary, i.e., the coefficient sum a of the 2nd fritter subtracts other coefficient sums b and obtained in each template
The candidate target that the number c arrived has been much smaller than.
Therefore, we are obtained as drawn a conclusion：Good candidate target, the local dictionary coefficient of its l-th of fritter is sparse, and compared with
Big coefficient can only be distributed in the l, 9+l, 9*2+l ... of local dictionary coefficient column vector, on 9* (n-1)+l element positions, and this
Coefficient sum on a little positions is much larger than poor candidate target.Thus summarize the local dictionary of a utilization and represent that coefficient weighs time
Select the rule of target quality：First from local coefficient's matrixIn by row extract the part of i-th of particle
Dictionary represents coefficient column vectorWherein (A)iRepresenting matrix A the i-th row.Then take out fritter is corresponded in column vector be
Number, i.e.,In l, 9+l, 9*2+l ..., 9* (n-1)+l elements (n is To Template number), seek the absolute of them
It is worth sum, then subtracts the absolute value sum of other elements, obtains partial weight
S43：Calculate total weight of particle
Total weight metric of particle similitude of particle and target, in order to more accurately represent this similitude, so
Local feature should be considered, also to consider global characteristics, i.e., to combine partial weight and global weight to determine total power of particle
Weight.Therefore, obtained according to S41 and S42WithCalculate total weight of particleIts
Middle parameter σ determines the importance accounting of global characteristics and local feature on target similarity measurement.Through experiment, energy during σ=3
Play good coordinative role；
S44：Calculate s+1 frames target location
In this step, first total weight of all particles is normalized for we, then according to particleCoordinate
PositionAnd weights omegai, obtain s+1 frames target location LS+1,
In the present embodiment, step 5 includes step S51 and S52, and concrete operations are as follows：
S51：Update local dictionary
Under normal circumstances, track the firm incipient stage it is selected template drift error it is smaller, tracking result is more accurate.But
When the situations such as target occlusion, deformation occur, the template at more early moment can not represent current goal well, therefore need tracking
During carry out dictionary updating.In summary two aspect, we discharge the sequencing of template temporally, then to compared with
The template progress frequent updating at late moment, and the renewal of the template progress more low frequency to the more early moment.To realize this renewal side
Method, we generate a cumulative probability sequence to represent the accumulative update probability of each template：
Wherein n represents To Template sum.Then random one section of generation [0,1) on number r, judge that r is in formula
(13) which section of sequence in, ifThen delete j-th of template E in dictionaryjCorresponding DsIn
FritterThen the tracking result of present frame according to being divided into 9 fritters shown in Fig. 2, then with column vector form table
Go out, i.e. { t1,t2,…,t9, finally it is added into dictionary DsEnd, now
S52：Update Global Dictionary
For background template, we take the strategy of all renewals of every frame, i.e., each frame deletion previous frame all back ofs the body
Scape template, then according to the tracking result of present frame, background template is chosen again.Therefore for Global DictionaryFor, before it
The renewal of partial target template is similar with local dictionary updating method, according to the section where generation random number r, deletes some mould
Plate, addition dictionary is last after the tracking result of present frame then is carried out into affine transformation, nowWhereinChange for tracking result is affine
The fritter arrived,Q=1,2 ..., n, background template is achieved for present frame
Global characteristics are not only carried out rarefaction representation by the method for the invention based on rarefaction representation, also to the office of tracking target
Portion's feature carries out rarefaction representation, so that the description to target, which not only contains global characteristics, also contains local feature, makes
There must be more preferable robustness to the tracking of target.Particularly in the case where locally blocking, shield portions other are removed
Fritter also can be good at being indicated target, so as to greatly improve the stability of tracking.And it is dilute to be different from general
The method for representing individually to be handled particle is dredged, present invention utilizes interparticle similitude, by the coefficient vector of each particle
Coefficient matrix is formed, and has added the minimum limitation of order to coefficient matrix, so as to reduce the operand of algorithm, and optimizes tracking
As a result.And in order to suppress the drift phenomenon in object tracking process, the present invention also incorporates the background mould with target range farther out
Plate adds Global Dictionary.Under the background for emphasizing rarefaction representation, it is desirable to which the coefficient of background template is as small as possible, so as to suppress to drift about
Phenomenon, obtain more preferably effect.
Embodiments of the invention are the foregoing is only, are not intended to limit the scope of the invention, it is every to utilize this hair
The equivalent structure or equivalent flow conversion that bright specification and accompanying drawing content are made, or directly or indirectly it is used in other related skills
Art field, is included within the scope of the present invention.
Claims (4)
1. a kind of method for tracking target represented based on low-rank matrix, it is characterised in that comprise the following steps：
Step 1：Initial target location is chosen in the 1st frame, To Template and background template are chosen according to initial target location, first
Piecemeal processing is carried out to To Template, affine transformation then is carried out to To Template and background template, completes local dictionary and complete
The initialization of office's dictionary；Comprise the following steps：
S11：Choose To Template and background template
Using p pixel as step-length, centered on the L of target location, most step-lengths are moved up and down in vertical direction, in level
Most step-lengths are moved left and right on direction, so as to get 8 with target sizes shape identical block as template, target is in itself
A template equally is used as, finally again with taking m at random in the range of length and width minimum value of the target location distance no more than target
It is individual, form the To Template set { E formed with m+9 template1,E2,…Em+9, randomly selected in the position not less than 1.5 times of d
N size is with target identical block as background template set { B1,B2,…,Bn, d is grown for target and wide minimum value,
S12：Piecemeal processing is carried out to To Template
By To Template Ej(j ∈ [1, m+9]) is divided into 3*3 fritter, and numbers, then by each fritter in the form of column vector
Represent, obtainWherein, each vectorial element value is the gray value of corresponding pixel points；
S13：Affine transformation is carried out to To Template and background template
To To Template Ej(j ∈ [1, m+9]) and background template Bk(k ∈ [1, n]) carries out affine transformation, obtains and fritter in S12
The block of formed objects, and it is represented in the form of vectors, obtainWith
S14：Complete the initialization of local dictionary and Global Dictionary
With all To Templates and the affine transformation fritter of background templateWithTo initialize Global DictionaryInitialized
Global DictionaryAgain by each To Template Ej(j ∈ [1, m+9])
FritterTo initialize local dictionary Ds, so as to the local dictionary initialized
Step 2：According to the target location of s-1 frames, the tracking target T of s frames is chosen, then the tracking target T of s frames is imitated
Penetrate the block being transformed to atom formed objects in Global DictionaryAnd obtain state parameter Fs, s>=2；
Step 3：With s frame dbjective state parameters FsFor average, δ2Particle is gathered in s frames at random for the Gaussian Profile of variance,
Piecemeal processing and affine transformation are carried out to particle, matrix in block form and global matrix is built, each piecemeal square is represented with local dictionary
Battle array, obtain corresponding local coefficient matrix；Global matrix is represented with Global Dictionary, obtains global coefficient matrix；Comprise the following steps：
S31：With s frame dbjective state parameters FsFor average, δ2Particle collection is randomly selected for the Gaussian Profile of varianceN is particle collection number；
S32：By To Template Ej(j ∈ [1, m+9]) is divided into 3*3 size identical fritter and numbered, then will be each small
Block is represented in the form of column vector, is obtainedThe element value of vector is the gray value of corresponding pixel points；
Reuse and affine transformation is carried out to particle with S13 same operations, obtain
S33：The result after all particle piecemeals, that is, the fritter after dividing equallyExtract, wherein i
∈ [1, N] represents the sequence number of particle, and l ∈ [1,9] represent the label of fritter, then willIt is combined as pair
The matrix in block form answeredIt is expressed as with formulaWherein, N represents total number of particles, then
All particlesFor forming global matrix
S34：Each matrix in block form is represented with local dictionary, corresponding local coefficient matrix is obtained, global system is represented with Global Dictionary
Matrix number, obtain global coefficient matrix
Step 4：The total weights omega of each particle is obtained by global coefficient matrix and local coefficient's matrix computationsi, then according to grain
Subcoordinate positionAnd weights omegai, obtain s frames target location Ls,N represents total number of particles；
Step 5：According to the target location of s frames, local dictionary and Global Dictionary are updated；
Step 6：S=s+1, go to step 2.
A kind of 2. method for tracking target represented based on low-rank matrix according to claim 1, it is characterised in that step 2
Dbjective state parameter FsDetermined by six affine coefficients,Wherein dx,dy,s,θ,γ,Point
Biao Shi not target x direction displacements, the displacement of y directions, dimensional variation, the anglec of rotation, the ratio of width to height and chamfer angle.
A kind of 3. method for tracking target represented based on low-rank matrix according to claim 1, it is characterised in that step 4
The computational methods of middle particle weights are as follows：
S41：From global coefficient matrixIn by row propose that the global of i-th particle represents coefficient column vectorWhereinRepresenting matrixThe i-th row, then useIn preceding m+9 element absolute value and subtract after the vector n it is first
The absolute value sum of element, obtains the global weight of the particle
S42：From local coefficient's matrixIn the local dictionary of i-th particle extracted by row represent coefficient
Column vectorWhereinRepresenting matrixI-th row, then take out the coefficient that fritter is corresponded in column vector, i.e.,
In l, 9+l, 9*2+l ..., 9* (m+9-1)+l elements, m+9 is To Template number, seeks their absolute value sum,
The absolute value sum of other elements is subtracted again, obtains partial weightI.e.
S43：Obtained according to S41 and S42WithCalculate total weight of particle Wherein
Parameter σ determines the importance accounting of global characteristics and local feature on target similarity measurement；
S44：Total weight of all particles is normalized, then according to particleCoordinate positionAnd weights omegai, obtain
S frames target location Ls,
A kind of 4. method for tracking target represented based on low-rank matrix according to claim 1, it is characterised in that step 5
The update method of middle dictionary is as follows：
S51：A cumulative probability sequence is generated to represent the accumulative update probability of each template：
<mrow>
<msub>
<mi>L</mi>
<mi>p</mi>
</msub>
<mo>=</mo>
<mo>{</mo>
<mn>0</mn>
<mo>,</mo>
<mfrac>
<mn>1</mn>
<mrow>
<msup>
<mn>2</mn>
<mrow>
<mi>n</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
<mo>-</mo>
<mn>1</mn>
</mrow>
</mfrac>
<mo>,</mo>
<mfrac>
<mn>3</mn>
<mrow>
<msup>
<mn>2</mn>
<mrow>
<mi>n</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
<mo>-</mo>
<mn>1</mn>
</mrow>
</mfrac>
<mo>,</mo>
<mo>...</mo>
<mo>,</mo>
<mn>1</mn>
<mo>}</mo>
</mrow>
Wherein n represents To Template sum, then random one section of generation [0,1) on number r, judge that r is in sequence in above formula
Which section of row, ifThen delete j-th of template E in dictionaryjCorresponding local dictionary DsIn
FritterThen the tracking result of present frame is divided into 9 fritters, then gone out with column vector form table, i.e. { t1,
t2,…,t9, finally it is added into local dictionary DsEnd, now
S52：For background template, all strategies updated of every frame are taken, i.e., in each frame, the mould that has powerful connections of the previous frame of deletion
Plate, then according to the tracking result of present frame, background template is chosen again；
For Global DictionaryFor, the renewal of its forward part To Template is similar with local dictionary updating method, according to generation
Section where random number r, some template is deleted, dictionary is added most after the tracking result of present frame then is carried out into affine transformation
Afterwards, nowWhereinIt is affine for tracking result
Change obtained fritter,Background template is achieved for present frame.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510916027.XA CN105513093B (en) | 2015-12-10 | 2015-12-10 | A kind of method for tracking target represented based on low-rank matrix |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510916027.XA CN105513093B (en) | 2015-12-10 | 2015-12-10 | A kind of method for tracking target represented based on low-rank matrix |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105513093A CN105513093A (en) | 2016-04-20 |
| CN105513093B true CN105513093B (en) | 2018-02-16 |
Family
ID=55721050
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510916027.XA Expired - Fee Related CN105513093B (en) | 2015-12-10 | 2015-12-10 | A kind of method for tracking target represented based on low-rank matrix |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105513093B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105957026B (en) * | 2016-04-22 | 2019-02-05 | 温州大学 | Based on the denoising method inside non local similar image block recessiveness low-rank structure between block |
| CN106296740B (en) * | 2016-08-11 | 2019-05-14 | 电子科技大学 | A kind of target fine definition tracking based on low-rank sparse expression |
| CN106384356A (en) * | 2016-09-22 | 2017-02-08 | 北京小米移动软件有限公司 | Method and apparatus for separating foreground and background of video sequence |
| CN106503652A (en) * | 2016-10-21 | 2017-03-15 | 南京理工大学 | Based on the accident detection method that low-rank adaptive sparse is rebuild |
| CN106503647A (en) * | 2016-10-21 | 2017-03-15 | 南京理工大学 | The accident detection method that structural sparse is represented is approached based on low-rank |
| CN106530330B (en) * | 2016-12-08 | 2017-07-25 | 中国人民解放军国防科学技术大学 | Video target tracking method based on low-rank sparse |
| CN107301652B (en) * | 2017-05-04 | 2020-07-14 | 西北工业大学 | Robust target tracking method based on local sparse representation and particle swarm optimization |
| CN108399620B (en) * | 2018-03-19 | 2022-04-19 | 嘉兴学院 | Image quality evaluation method based on low-rank sparse matrix decomposition |
| CN108957395B (en) * | 2018-04-18 | 2022-06-10 | 江苏省送变电有限公司 | Three-dimensional positioning method for noise immunity moving target in tunnel |
| CN110648351B (en) * | 2019-09-19 | 2022-02-01 | 安徽大学 | Multi-appearance model fusion target tracking method and device based on sparse representation |
| CN113658227B (en) * | 2021-08-26 | 2024-02-20 | 安徽大学 | RGBT target tracking method and system based on collaborative low-rank graph model |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103440645A (en) * | 2013-08-16 | 2013-12-11 | 东南大学 | Target tracking algorithm based on self-adaptive particle filter and sparse representation |
| CN104408708A (en) * | 2014-10-29 | 2015-03-11 | 兰州理工大学 | Global-local-low-rank-based image salient target detection method |
| CN104766343A (en) * | 2015-03-27 | 2015-07-08 | 电子科技大学 | Vision target tracking method based on sparse representation |
| CN105046717A (en) * | 2015-05-25 | 2015-11-11 | 浙江师范大学 | Robust video object tracking method |
| CN105139422A (en) * | 2015-08-14 | 2015-12-09 | 中国联合网络通信集团有限公司 | Self-interpreting target tracking method and self-interpreting target tracking device |
-
2015
- 2015-12-10 CN CN201510916027.XA patent/CN105513093B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103440645A (en) * | 2013-08-16 | 2013-12-11 | 东南大学 | Target tracking algorithm based on self-adaptive particle filter and sparse representation |
| CN104408708A (en) * | 2014-10-29 | 2015-03-11 | 兰州理工大学 | Global-local-low-rank-based image salient target detection method |
| CN104766343A (en) * | 2015-03-27 | 2015-07-08 | 电子科技大学 | Vision target tracking method based on sparse representation |
| CN105046717A (en) * | 2015-05-25 | 2015-11-11 | 浙江师范大学 | Robust video object tracking method |
| CN105139422A (en) * | 2015-08-14 | 2015-12-09 | 中国联合网络通信集团有限公司 | Self-interpreting target tracking method and self-interpreting target tracking device |
Non-Patent Citations (3)
| Title |
|---|
| Low-Rank Sparse Learning for Robust Visual Tracking;TianZhu Zhang等;《In Proceedings of the 12th European conference on Computer Vision》;20121013;第4卷;第473-476页 * |
| Motion object tracking based on the low-rank matrix representation;Xiaofang Kong 等;《Optical Review》;20151031;第22卷(第5期);第786–801页 * |
| 基于全局和局部信息的目标跟踪算法;周明珠 等;《光电工程》;20150430;第42卷(第4期);第19-22页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105513093A (en) | 2016-04-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105513093B (en) | A kind of method for tracking target represented based on low-rank matrix | |
| Zhang et al. | An object tracking framework with recapture based on correlation filters and Siamese networks | |
| CN110021051B (en) | Human image generation method based on generation of confrontation network through text guidance | |
| Shi et al. | Adaptive image-based visual servoing with temporary loss of the visual signal | |
| US10175662B2 (en) | Method of constructing navigation map by robot using mouse hippocampal place cell model | |
| CN109800689B (en) | Target tracking method based on space-time feature fusion learning | |
| Duan et al. | Template matching using chaotic imperialist competitive algorithm | |
| CN110874578B (en) | Unmanned aerial vehicle visual angle vehicle recognition tracking method based on reinforcement learning | |
| CN102722697B (en) | Unmanned aerial vehicle autonomous navigation landing visual target tracking method | |
| CN105678338B (en) | Target tracking method based on local feature learning | |
| CN103440667B (en) | The automaton that under a kind of occlusion state, moving target is stably followed the trail of | |
| CN105930767A (en) | Human body skeleton-based action recognition method | |
| Gundavarapu et al. | Structured Aleatoric Uncertainty in Human Pose Estimation. | |
| CN109785385A (en) | Visual target tracking method and system | |
| CN110334584B (en) | Gesture recognition method based on regional full convolution network | |
| CN106127776A (en) | Based on multiple features space-time context robot target identification and motion decision method | |
| CN109508686B (en) | Human behavior recognition method based on hierarchical feature subspace learning | |
| CN110555404A (en) | Flying wing unmanned aerial vehicle ground station interaction device and method based on human body posture recognition | |
| Scorsoglio et al. | Safe Lunar landing via images: A Reinforcement Meta-Learning application to autonomous hazard avoidance and landing | |
| CN110163132B (en) | Correlation filtering tracking method based on maximum response value change rate updating strategy | |
| CN103985141B (en) | Method for tracking target based on hsv color covariance feature | |
| Ding et al. | Simultaneous body part and motion identification for human-following robots | |
| Teng et al. | Generative robotic grasping using depthwise separable convolution | |
| Chu et al. | Target tracking via particle filter and convolutional network | |
| CN106530319A (en) | Video object cooperative segmentation method based on track directed graph |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180216 Termination date: 20181210 |