CN107844739A - Robustness target tracking method based on adaptive rarefaction representation simultaneously - Google Patents
Robustness target tracking method based on adaptive rarefaction representation simultaneously Download PDFInfo
- Publication number
- CN107844739A CN107844739A CN201710625586.4A CN201710625586A CN107844739A CN 107844739 A CN107844739 A CN 107844739A CN 201710625586 A CN201710625586 A CN 201710625586A CN 107844739 A CN107844739 A CN 107844739A
- Authority
- CN
- China
- Prior art keywords
- template
- adaptive
- target
- tracking
- sparse
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 86
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 29
- 238000000354 decomposition reaction Methods 0.000 claims description 5
- 238000012935 Averaging Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 9
- 238000005286 illumination Methods 0.000 description 9
- 238000002474 experimental method Methods 0.000 description 7
- 238000005457 optimization Methods 0.000 description 5
- 230000000903 blocking effect Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 230000008569 process Effects 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 230000006978 adaptation Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 238000009795 derivation Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 102100024633 Carbonic anhydrase 2 Human genes 0.000 description 1
- 102100024644 Carbonic anhydrase 4 Human genes 0.000 description 1
- 241000282994 Cervidae Species 0.000 description 1
- 101000760643 Homo sapiens Carbonic anhydrase 2 Proteins 0.000 description 1
- 101000760567 Homo sapiens Carbonic anhydrase 4 Proteins 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 239000002537 cosmetic Substances 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000009432 framing Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
- G06V20/42—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items of sport video content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
- G06T7/251—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving models
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The robustness target tracking method based on adaptive rarefaction representation simultaneously of the present invention, comprises the following steps：S1, the size according to Laplacian noise energy, adaptive foundation while sparse tracing model；S2, the tracing model established is solved；S3, template is updated.Method for tracing track identification effect provided by the invention is good, and anti-interference is stronger, can realize that more accurate and real-time target tracking, method for tracing are relatively stable.
Description
Technical field
The invention belongs to computer image processing technology field, is related to a kind of method for tracking target, more specifically, relates to
A kind of and robustness target tracking method based on adaptive rarefaction representation simultaneously.
Background technology
Target tracking in computer vision field in occupation of consequence, with the use of high quality computer and video camera
And video automatically analyze it is required that people to target tracking produce great interest.The main task of target tracking includes：
Detection, frame of video to the continuous tracking between frame and the behavioural analysis for following the trail of target of moving target interested.Currently, target chases after
The relevant application of track includes：Move identification, video frequency searching, man-machine interaction, traffic monitoring, vehicle mounted guidance etc..
At present, it has been suggested in spite of many tracing algorithms, but target tracking technology still suffers from many challenges.
Video noise pollution is frequently run onto in actual tracing process, for example target pose change, illumination variation, background mixes, part hides
Gear and the problems such as blocking completely, these problems frequently can lead to target tracking failure (tracking drift).Especially block for a long time
Influence to target tracking is even more catastrophic, and illumination variation depends on the environment residing for target, and illumination variation is bigger to tracking
Influential effect is also bigger, and background mixes the accuracy for changing with target pose and influenceing tracking, therefore also results in tracking drift
Move, video noise pollution can cause tracking error to accumulate, and will ultimately result in tracking failure.
For target tracking, it is a basic and challenge to solve the problems, such as target shape change, and target shape changes
Inherent profile variation and external profile variation can be divided into.It is a kind of inherent profile variation that posture, which changes, and illumination variation, background are mixed
Miscellaneous, partial occlusion and block belong to external profile variation completely.Solving these cosmetic variations just needs one adaptive to chase after
Track method, i.e. on-line study method.On-line study method is broadly divided into two classes, respectively generation method at present
(Generative Approaches) and method of discrimination (Discriminative Approaches).Generation method (GA) is one
The method in the kind search region most like with tracking target, method of discrimination (DA) can be regarded as a kind of two classification problems, its master
Syllabus is to train a grader using known training sample, for differentiating target and background.Method of discrimination and generation
Method reliable tracking is realized in a certain degree although also having the shortcomings that respective, first, method of discrimination is to requirements for extracting features
It is higher therefore sensitiveer to noise during actual tracking, for the larger target of noise it is possible that tracking fails,
And generation method can not accurately find the region similar to target under the background mixed, therefore easily produce tracking failure；Two
It is that method of discrimination needs enough training sample set, good sample can lift the performance of grader, and bad sample can weaken point
The performance of class device, if bad sample is introduced into grader will influence tracking effect, and generation method is more sensitive to template, and one
The target blocked is introduced template by denier mistake tracking failure may occurs, therefore two methods do not have in reality scene tracking
There are enough robustness.
The content of the invention
It is an object of the present invention to solve the above problems, there is provided one kind cope with destination object illumination variation in video,
Dimensional variation, block, deform, motion blur, the quickly various challenges such as motion, rotation, background clutter, low resolution, can be right
Destination object carries out the method for tracing of robustness target that is continuous, accurately following the trail of.
The object of the present invention is achieved like this, based on the robustness target tracking method of adaptive rarefaction representation simultaneously,
Comprise the following steps：
S1, the size according to Laplacian noise energy, adaptive foundation while sparse tracing model；
S2, the tracing model established is solved；
S3, template is updated.
Further, the execution method of the step S1 is：Contrast Laplce's average noise | | S | |2With given noise
Energy threshold τ size, and the foundation adaptive according to comparing result while sparse tracing model:
When | | S | |2During≤τ, while sparse tracing model is：
When | | S | |2>During τ, while sparse tracing model is：
Wherein, D is tracking template, is expressed as D=[T, I], and T is To Template,
T=[T1,T2,…,Tn]∈Rd×n(d>>N), d represents the dimension of image, and n represents the number of template base vector, T
Each row be all that tracking template is represented by vectorial D after zero averaging, Y is candidate target collection, λ1And λ2For the canonical of model
Parameter, X are sparse coefficient, and S is Laplacian noise,For the fidelity item of sparse model, | | X | |1,1For coefficient square
Battle array, | | S | |1,1Characterize Laplacian noise energy.
Further, the step S2 includes：
S21, calculate tracking target ytWith the similitude of template T averages, sim is designated as；
S22, the size for judging sim and cosine angle threshold value α, and be adaptive selected model and be tracked：
Work as sim<During α, solved using the tracing model of above formula (1), and by alternating direction multiplier (ADMM) method
And obtain sparse coefficient X；
As sim >=α, asked using the tracing model of above formula (2), and by alternating direction multiplier (ADMM) method
Solve and obtain sparse coefficient X and Laplacian noise S.
Further, the similitude of tracking target and template average is calculated in the step S21 according to below equation：
Wherein, c is template T average, and y is the target traced into, then | | τ | |2It is equal with template that target can be equivalent to
The anticosine of value, i.e. cosine angle.
Further, when | | s | |2During≤τ, template T is updated.
Further, the template renewal method includes:
S31, to carry out singular value decomposition to current template T and the target y tracked respectively as follows：
S32, using singular vector u, s, v incremental update U, S, V are removed, so as to obtain new singular vector U*,S*,V*, then newly
Template representation be：
T*=U*S*V* T (5)
Further, in addition to step S33：Template is trained using unsupervised learning K-means methods, gives initial classes
Number is k, then
Wherein, i represents i-th of sample, whenWhen belonging to class k, then rik=1, otherwise rik=0, ukFor all category
In the average value of class k sample, J represent sample point to such sample point average distance with；Kth then can obtain by formula (6)
The average value of all samples of class, reduces original template dimension；Thus, the template for updating to obtain is：
Tnew=[u1,u2,…,uk] (7)
The present invention also provides the fast target method for tracing that a kind of adaptive sparse represents, comprises the following steps：
A, target tracking model is established：
B, using the method for alternating iteration, optimal tracking target is obtained；
C, To Template T is updated, and obtains new template T*；
D, best tracking target and To Template set are returned, continues the target tracking of next frame.
Preferably, the step C includes：
C1, the similitude for calculating tracking target and template average, similitude are designated as sim；
C2, by similitude sim and cosine angle threshold value α, β is compared, if sim<α, perform step C3 and continue to update
Template, if α≤sim≤β, then y ← m, step C3 more new templates are performed, if sim>β, now follow the trail of target and template
Substantially it is dissimilar, target is represented by serious noise pollution, not more new template；
C3, T=U ∑s V is obtained to template T progress singular value decompositionT, incremental update template T left singular vector U simultaneously obtains
New singular vector U*, convolution (5), (6), new template is calculated
Compared with prior art, beneficial effects of the present invention are embodied in：
(1) present invention proposes new stencil-chosen and update method under the framework of rarefaction representation, and this method is more strengthened
The real-time update of adjustment mould plate, simultaneously because real-time update be template left singular vector, the error for updating introducing can be controlled
It is very low, relative to carrying out noise eliminating and then being re-introduced into new To Template to target, also draw in template renewal of the invention
Enter K-means technologies, it reduce the dimension of template, is effectively reduced redundancy template vector, improves the real-time of tracking,
And reduce the influence of noise.And traditional template renewal method has then been readily incorporated larger noise error, to next frame
The tracking of target causes many uncertainties；
(2) influence of Gaussian noise and Laplacian noise has been taken into full account in target tracking model of the invention, according to
The regular terms of the preference pattern of the size adaptation of Laplacian noise energy, which not only improves the precision of tracking, also improve
The real-time of tracking；
(3) ADMM algorithms are attached in the solution of tracing model by the present invention, pass through the control to regular terms parameter
The solution of model is more stable.
Embodiment
The robustness target tracking based on adaptive rarefaction representation simultaneously below in conjunction with specific embodiment to the present invention
Method makes explanation further elucidated above.
Based on the robustness target tracking method of adaptive rarefaction representation simultaneously, comprise the following steps：
S1, the size according to Laplacian noise energy, adaptive foundation while sparse tracing model
Contrast Laplce's average noise | | S | |2With the size of given noise energy threshold tau, and according to comparing result from
The foundation of adaptation while sparse tracing model:
When | | S | |2During≤τ, while sparse tracing model is：
When | | S | |2>During τ, while sparse tracing model is：
Wherein, define D=[T, I] and represent tracking template, I represents trifling template, gives the image collection of To Template
T=[T1,T2,…,Tn]∈Rd×n(d>>N),
D represents the dimension of image, and n represents the number of template base vector, T each row be all by after zero averaging to
Amount.Wherein, Y=[y1,y2,…,ym], represent all candidate targets, it is assumed that noise obeys Laplacian distribution, then：
Y=TZ+S+E (9)
S represents Laplacian noise, and E represents Gaussian noise, and X is sparse coefficient,λ1And λ2For the canonical of model
Parameter,For the fidelity item of sparse model, model has this deformation after considering Laplacian noise, | | X | |1,1For
Coefficient matrix, it can more preferably extract similitude between particle and effectively go removing template redundancy, | | S | |1,1Characterize Laplce
Noise energy.
Although existing rarefaction representation target tracking algorithm solves partial occlusion, illumination variation, appearance to a certain extent
Gesture changes and background mixes etc. and influenceed, but model is all based on what is established in the case of noise Gaussian distributed, too simply
Consider noise profile situation, therefore it is possible that tracking failure during in face of some complicated noise profile situations.The present invention's
The influence of Gaussian noise and Laplacian noise is taken into full account in target tracking model, according to the big of Laplacian noise energy
The regular terms of small adaptive preference pattern, which not only improves the precision of tracking, also improve the real-time of tracking.
S2, the tracing model established is solved
Preferably, the step S2 includes：
S21, calculate tracking target ytWith the similitude of template average, sim is designated as；
As preferable, the similitude of tracking target and template average is calculated according to below equation：
S22, the size for judging sim and cosine angle threshold value α, and be adaptive selected model and be tracked：
Work as sim<During α, using the tracing model of above formula (1), solved simultaneously by alternating direction multiplier (ADMM) method
Obtain sparse coefficient X；
As sim >=α, using the tracing model of above formula (2), solved by alternating direction multiplier (ADMM) method
And obtain sparse coefficient X and noise S.
Illustrated by the solution scheme of the clearer tracing model to establishing, above formula (2) is exemplified below
Solution throughway：
Above formula (1) and the object function of (2) are a convex optimization problems, therefore can use Optimization without restriction to mesh
Scalar functions are solved, and alternating direction Multiplier Method (ADMM) is a kind of classical way of no constraint solving, and it, which has, solves surely
It is the advantages that fixed and fast convergence rate, as follows using ADMM method solving-optimizing problems (2)：
First, restricted problem is changed into unconstrained problem is
WhereinIt is an indicator function (xiRepresent X the i-th row, and if xiIt is non-negative, then τ+(xi)
Equal to 0；Otherwise τ+(xi) be equal to+∞).Therefore, optimization problem (2) has the following equivalent form of value：
Wherein V1,V2,V3For dual variable, formula (11) is further optimized for
Here,
The Augmented Lagrangian Functions of formula (11) are
Wherein β represents Lagrange multiplier, U=[U1,U2,U3]T.
Formula (15) can be decomposed into three sub- optimization problems, respectively X subproblem, and S subproblem and V are asked
Topic.It is as follows to this little optimization problem separately below：
Therefore, according to extremum principle, it is only necessary to seek first derivative to above-mentioned subproblem, the optimal solution of equation (15) can be obtained
It is as follows：
V1*=[β (TX-U1)+(Y-S)]/(1+β)
V2*=shrink (X-U2,λ1/β)
V3*=max (0, X-U3)
S*=shrink (Y-V1,λ2)
X*=(TTT+2I)-1[TT(V1+U1)+V2+U2+V3+U3]
Wherein shrink is a deflation operator, i.e., for a non-negative vector p, then has
ShrinkI (x, p)=sgn (x) ο max | x |-p, 0 }
Similarly, it can still be solved for above formula (1) using ADMM methods, and obtain the form of following solution：
V1*=[β (TX-U1)+Y]/(1+β)
V2*=shrink (X-U2,λ1/β)
V3*=max (0, X-U3)
X*=(TTT+2I)-1[TT(V1+U1)+V2+U2+V3+U3]
So by the analysis and solution to subproblem, the general type of formula (1) and (2) solution is obtained.If input to
Fixed Lagrange multiplier β, formula (1) and the optimum value solution of (2) can be obtained by alternating direction iteration, wherein, table 1 is
The ADMM derivation algorithm processes of formula (1), table 2 are the ADMM derivation algorithm processes of formula (2).
Table 1
Table 2
S3, template is updated
As a kind of preferable scheme, when | | s | |2During≤τ, template T is updated.
Further, the template renewal method includes：
S31, singular value decomposition is carried out such as to the current template T and target y tracked respectively
Under：
T=USVT
Y=usvT (4)
S32, using singular vector u, s, v, remove incremental update U, S, V, so as to obtain new singular vector U*, S*, V*, then
New template representation is：
T*=U*S*V* T (5)
Preferably, in addition to step S33：Template is trained using unsupervised learning K-means methods, gives of initial classes
Number is k, then K-means learning methods are as follows：
Wherein i represents i-th of sample, whenWhen belonging to class k, then rik=1, otherwise rik=0, ukBelong to be all
The average value of class k sample, J represent sample point to such sample point average distance with；Kth class then can obtain by formula (6)
The average value of all samples, reduce original template dimension.Thus, the template for updating to obtain is：
Tnew=[u1,u2,…,uk] (7)
Template renewal method provided by the invention shows stronger robustness, this method for blocking with illumination variation
Different from traditional template renewal, it is it is emphasised that selection has the template of significant contribution to target tracking, and avoids using trivial
Broken template, and unsupervised training is carried out to template by K-means algorithms, the redundancy of template is eliminated significantly, so as to
Improve the real-time of tracking.
It is adaptive while rarefaction representation the target tracking algorithmic procedure of the embodiment of the present invention as shown in table 3：
Table 3
Below, will be by experiment that target tracking method provided by the invention (Ours) is existing with very with other five kinds
The method of good tracking performance is compared, and this five kinds of tracing algorithms are respectively that the circular matrix of geo-nuclear tracin4 follows the trail of (CSK), acceleration pair
Even gradient follows the trail of (L1APG), multitask tracking (MTT), sparse prototype tracking (SPT) and sparse joint tracking (SCM).Below
Experiment is all based on Matlab 2012a, and calculator memory 2GB, CPU are to be carried out on Intel (R) Core (TM) i3 platform.
Data and description of test：
The different videos with tracking challenge of 14 kinds of experimental selection, including blocking, illumination variation, background is mixed
Miscellaneous, posture changes, and the factor of the influence tracking result such as low resolution and quick motion, video sequence attribute brief introduction is shown in Table 4.Table 4
Middle video contains different noises, and wherein OV represents that target is lost, and BC represents that background mixes, and OCC represents to block completely, OCP
Partial occlusion is represented, OPR represents to rotate out of plane, and LR represents low resolution, and FM represents quick motion, and SV represents size variation.
The evaluation method that the experiment of the embodiment of the present invention uses has three kinds, and every kind of evaluation method can interpretive tracing to a certain extent
The quality of performance, respectively local center error (Center Local Error), Duplication (Overlap Ratio) and curve
Under area (Area Under Curve).Real goal frame R to framingg(ground truth) and tracking target frame Rt
(tracked target bounding), the center that might as well set them is respectively:pg=(xg,yg) and pt=(xt,yt),
Then local center error is CLE=| | pg-pt||2, Duplication is
Area () represents all pixels in the region, and the value of the curve of areas (AUC) every bit represents that Duplication is more than
The success rate of video frequency tracking during given threshold value η.Especially, we set η=0.5, as Duplication OR>Then think to chase after when 0.5
Track frame success.Correlation tracking result, is shown in Table shown in 4,5,6.
Table 4
| Video sequence | Frame number | Noise (s) | Video sequence | Frame number | Noise (s) |
| Walking2 | 495 | SV,OCP,LR | Suv | 945 | OCC,OV,BC |
| Car4 | 659 | IV,SV | CarDark | 393 | IV,BC,LR |
| Car2 | 913 | IV,SV,BC | Deer | 71 | FM,LR,BC |
| Girl | 500 | OPR,OCC,LR | Singr2 | 366 | IV,OPR,BC |
| FaceOcc2 | 812 | OCC,OPR,IV | Skater2 | 435 | SV,OPR |
| Football | 362 | OCC,OPR,BC | Dudek | 1145 | OCC,BC,OV |
| FaceOcc1 | 892 | OCC | Subway | 175 | OCC,BC |
Table 5 is the correction data of the various algorithms of different performances based on average Duplication, and wherein AOR represents total and is averaged
Duplication, wherein, average Duplication is bigger to represent that tracking performance is better；Parameter setting is as follows in an experiment:Regular parameter λ1=
0.1, λ2=0.1, penalty factor β=0.1, the minimum α of complementary chord angle threshold valuemin=20, it is up to αmax=35, template maximum base
Vectorial number is 15, and particle sampler number is 600, and the size of image block is 25 × 25, tests maximum iteration Loop=20, receives
Hold back error tol=0.001.Parameter lambda in experiment1,λ2It is to be obtained by cross validation method, and λ2The regulation of parameter meets such as
Lower rule, if Laplacian noise S energy is larger (i.e. target is blocked by larger, profile variation or illumination variation), this
When λ2Value should be smaller, it is on the contrary then larger.From Experimental comparison's data of table 5, method for tracing provided by the invention
(Ours) no matter it is substantially higher in other tracking sides from the other average Duplication of different video classes or total average Duplication
Method, i.e. method for tracing of the invention achieve best tracking performance effect.
Table 5
It is the contrast of the various algorithms of different performances based on average local center error as shown in table 6, wherein ACLE is represented
Total mean center error, mean center error is smaller to represent that tracking performance is better.By the data of table 6 it is clear that the present invention
No matter the method for tracing (Ours) of offer is from the other mean center error of different video classes or total mean center error,
Other method for tracing are significantly lower than, i.e. method for tracing of the invention achieves optimal final performance resultss.
Table 6
It is the contrast of the various algorithms of different performances based on average success rate as shown in table 7, wherein, ASR represents total and put down
Equal success rate；
Table 7
From the correction data in table 7, no matter method for tracing (Ours) provided by the invention is to different video classification
Average success rate is still evaluated from total average success rate, and it is intended to due to other method for tracing.
In order to further understand tracing algorithm proposed by the present invention, the Laplce mentioned in model is described below
The specific influence of noise and template renewal criterion on tracking effect.
Traditional template renewal method is updated by following the trail of the similarity of target and template, if similarity
More than given threshold value, then it is assumed that target has met with larger noise pollution, it is therefore desirable to the original weights of target substitution will be followed the trail of
Less template vector, it is relatively rough in fact so to replace, because larger noise error is introduced, thus to next
The tracking of frame target causes many uncertainties, and new template renewal method proposed by the present invention then weakens noise shadow
Ring.Specific manifestation is as follows：
(1) weight that new template renewal method is effectively weighed between the new tracking target of primary template vector sum, passes through
Forgetting factor realizes template renewal；
(2) new template renewal method introduces K-means methods, and this can be effectively reduced redundancy template vector, carry
The real-time of height tracking, and the calculating at class center is obtained by weighted average, therefore can also effectively weaken noise.
Specific experiment is given below and is respectively compared the influence of template renewal and Laplce to experiment effect, experimental subjects choosing
Select：MTT Algorithm, ASSAT algorithms (only Laplce), ASSAT (only template renewal), ASSAT (Laplce+template renewal),
Experimental data selection sequence Skater2, Dudek, SUV, Walking2, Subway, Deer etc..
Table 8 is influence contrast of the Laplce to experimental result, as can be seen from Table 8 except Walking2 sequences, is added
Its tracking effect is better than MTT Algorithm after Laplacian noise, but the method for primary template renewal limits its traceability
Can, and the new template renewal method proposed promotes the tracking performance of ASSAT algorithms.
Table 8
Table 9 is that the influence that different templates are updated to experimental result contrasts, as can be seen from Table 9 using only template renewal
The tracking effect of ASSAT methods and IVT methods is similar, and how many is not lifted, for Skater2, two kinds of Subway sequences
Method effect is all bad, reason be both sequences contain it is larger block, for only consider template renewal without consider draw
The ASSAT algorithms of this noise of pula can not effectively track target, and IVT is also the same.In fact, for this containing larger
Situation about blocking, if not considering, Laplacian noise can sum up in the point that noise factor have impact on the sparse knot of solution X in formula (1)
Structure.
Table 9
Target selects the influence to solution to be also manifested by the different noise of circumstance of occlusion, when consideration Laplacian noise when institute
Obtained solution is sparse, and now solution is optimal, and it is dense non-optimal not consider resulting solution during Laplacian noise
Solution, therefore keep the sparsity structure of solution to directly affect the tracking performance of algorithm.
The present invention also provides the fast target method for tracing that a kind of adaptive sparse represents, comprises the following steps：
A, target tracking model is established：
B, using the method for alternating iteration, optimal tracking target is obtained；
C, To Template T is updated, and obtains new template T*
D, best tracking target and To Template set are returned, continues the target tracking of next frame.
As preferable, the step C includes：
C1, the similitude for calculating tracking target and template average, similitude are designated as sim；
C2, by similitude sim and cosine angle threshold value α, β is compared, if sim<α, perform step C3 and continue to update
Template, if α≤sim≤β, then y ← m, step C3 more new templates are performed, if sim>β, now follow the trail of target and template
Substantially it is dissimilar, target is represented by serious noise pollution, not more new template；
C3, T=U ∑s V is obtained to template T progress singular value decompositionT, incremental update template T left singular vector U simultaneously obtains
New singular vector U*, convolution (5), (6), new template is calculated
Although an embodiment of the present invention has been shown and described, for the ordinary skill in the art, can be with
A variety of changes, modification can be carried out to these embodiments, replace without departing from the principles and spirit of the present invention by understanding
And modification, the scope of the present invention is defined by the appended.
Claims (8)
1. the robustness target tracking method based on adaptive rarefaction representation simultaneously, it is characterised in that comprise the following steps：
S1, the size according to Laplacian noise energy, adaptive foundation while sparse tracing model；
S2, the tracing model established is solved；
S3, template is updated.
2. the robustness target tracking method according to claim 1 based on adaptive rarefaction representation simultaneously, its feature exist
In the execution method of the step S1 is：Contrast Laplce's average noise | | S | |2It is big with given noise energy threshold tau
Small and adaptive according to comparing result foundation while sparse tracing model:
When | | S | |2During≤τ, while sparse tracing model is：
When | | S | |2>During τ, while sparse tracing model is：
Wherein, D represents tracking template, and Y is candidate target collection, λ1And λ2For the regular parameter of model, X is sparse coefficient, and S is drawing
This noise of pula,For the fidelity item of sparse model, | | X | |1,1For coefficient matrix, | | S | |1,1Characterize Laplce
Noise energy.
3. the robustness target tracking method according to claim 2 based on adaptive rarefaction representation simultaneously, its feature exist
In, the tracking template D is expressed as D=[T, I], wherein, T is To Template, T=[T1,T2,…,Tn]∈Rd×n(d>>N),
D represents the dimension of image, and n represents the number of template base vector, and T each row are all by the vector after zero averaging.
4. the robustness target tracking method according to claim 3 based on adaptive rarefaction representation simultaneously, its feature exist
In the step S2 includes：
S21, calculate tracking target ytWith the similitude of template T averages, sim is designated as；
S22, the size for judging sim and cosine angle threshold value α, and be adaptive selected model and be tracked：
Work as sim<During α, using the tracing model of above formula (1), solved and obtained by alternating direction multiplier (ADMM) method
Sparse coefficient X；
As sim >=α, using the tracing model of above formula (2), solved and obtained by alternating direction multiplier (ADMM) method
Sparse coefficient X and Laplacian noise S.
5. the robustness target tracking method according to claim 4 based on adaptive rarefaction representation simultaneously, its feature exist
According to the similitude of below equation calculating tracking target and template T averages in the step S21：
Wherein, c is template T average, and y is the target traced into, then | | τ | |2The anti-of target and template average can be equivalent to
Cosine, i.e. cosine angle.
6. the robustness target tracking method based on adaptive rarefaction representation simultaneously according to claim any one of 1-5,
Characterized in that, work as | | s | |2During≤τ, template T is updated.
7. the robustness target tracking method according to claim 6 based on adaptive rarefaction representation simultaneously, its feature exist
In the template renewal method includes:
S31, to carry out singular value decomposition to current template T and the target y tracked respectively as follows：
S32, using singular vector u, s, v incremental update U, S, V are removed, so as to obtain new singular vector U*,S*,V*, then new mould
Plate is expressed as：
T*=U*S*V* T (5)
8. the robustness target tracking method according to claim 7 based on adaptive rarefaction representation simultaneously, its feature exist
In, in addition to step S33：Template is trained using unsupervised learning K-means methods, the number for giving initial classes is k, then
Wherein, i represents i-th of sample, whenWhen belonging to class k, then rik=1, otherwise rik=0, ukBelong to class k's to be all
The average value of sample, J represent sample point to such sample point average distance with；All samples of kth class then can obtain by formula (6)
This average value, thus, the template for updating to obtain is：
Tnew=[u1,u2,…,uk] (7)。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710625586.4A CN107844739B (en) | 2017-07-27 | 2017-07-27 | Robust target tracking method based on self-adaptive simultaneous sparse representation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710625586.4A CN107844739B (en) | 2017-07-27 | 2017-07-27 | Robust target tracking method based on self-adaptive simultaneous sparse representation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107844739A true CN107844739A (en) | 2018-03-27 |
| CN107844739B CN107844739B (en) | 2020-09-04 |
Family
ID=61683198
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710625586.4A Expired - Fee Related CN107844739B (en) | 2017-07-27 | 2017-07-27 | Robust target tracking method based on self-adaptive simultaneous sparse representation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107844739B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109858546A (en) * | 2019-01-28 | 2019-06-07 | 北京工业大学 | A kind of image-recognizing method based on rarefaction representation |
| CN110610527A (en) * | 2019-08-15 | 2019-12-24 | 苏州瑞派宁科技有限公司 | SUV calculation method, device, equipment, system and computer storage medium |
| CN110738683A (en) * | 2018-07-19 | 2020-01-31 | 中移（杭州）信息技术有限公司 | computer vision tracking method and device |
| CN112630774A (en) * | 2020-12-29 | 2021-04-09 | 北京润科通用技术有限公司 | Target tracking data filtering processing method and device |
| CN115861379A (en) * | 2022-12-21 | 2023-03-28 | 山东工商学院 | Video tracking method for twin network to update target template based on local credible template |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2012078702A1 (en) * | 2010-12-10 | 2012-06-14 | Eastman Kodak Company | Video key frame extraction using sparse representation |
| CN103440645A (en) * | 2013-08-16 | 2013-12-11 | 东南大学 | Target tracking algorithm based on self-adaptive particle filter and sparse representation |
| CN105930812A (en) * | 2016-04-27 | 2016-09-07 | 东南大学 | Vehicle brand type identification method based on fusion feature sparse coding model |
| CN106203495A (en) * | 2016-07-01 | 2016-12-07 | 广东技术师范学院 | A kind of based on the sparse method for tracking target differentiating study |
| CN106651912A (en) * | 2016-11-21 | 2017-05-10 | 广东工业大学 | Compressed sensing-based robust target tracking method |
-
2017
- 2017-07-27 CN CN201710625586.4A patent/CN107844739B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2012078702A1 (en) * | 2010-12-10 | 2012-06-14 | Eastman Kodak Company | Video key frame extraction using sparse representation |
| CN103440645A (en) * | 2013-08-16 | 2013-12-11 | 东南大学 | Target tracking algorithm based on self-adaptive particle filter and sparse representation |
| CN105930812A (en) * | 2016-04-27 | 2016-09-07 | 东南大学 | Vehicle brand type identification method based on fusion feature sparse coding model |
| CN106203495A (en) * | 2016-07-01 | 2016-12-07 | 广东技术师范学院 | A kind of based on the sparse method for tracking target differentiating study |
| CN106651912A (en) * | 2016-11-21 | 2017-05-10 | 广东工业大学 | Compressed sensing-based robust target tracking method |
Non-Patent Citations (3)
| Title |
|---|
| TIANZHU ZHANG ETAL.: "Robust visual tracking via multi-task sparse learning", 《2012 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 张晓伟 等: "一种利用稀疏表示的距离扩展目标检测新方法", 《西安电子科技大学学报（自然科学版）》 * |
| 黄晶晶: "基于压缩感知的多选择正交匹配追踪改进算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110738683A (en) * | 2018-07-19 | 2020-01-31 | 中移（杭州）信息技术有限公司 | computer vision tracking method and device |
| CN109858546A (en) * | 2019-01-28 | 2019-06-07 | 北京工业大学 | A kind of image-recognizing method based on rarefaction representation |
| CN109858546B (en) * | 2019-01-28 | 2021-03-30 | 北京工业大学 | Image identification method based on sparse representation |
| CN110610527A (en) * | 2019-08-15 | 2019-12-24 | 苏州瑞派宁科技有限公司 | SUV calculation method, device, equipment, system and computer storage medium |
| CN110610527B (en) * | 2019-08-15 | 2023-09-22 | 苏州瑞派宁科技有限公司 | SUV computing method, device, equipment, system and computer storage medium |
| CN112630774A (en) * | 2020-12-29 | 2021-04-09 | 北京润科通用技术有限公司 | Target tracking data filtering processing method and device |
| CN115861379A (en) * | 2022-12-21 | 2023-03-28 | 山东工商学院 | Video tracking method for twin network to update target template based on local credible template |
| CN115861379B (en) * | 2022-12-21 | 2023-10-20 | 山东工商学院 | Video tracking method for updating templates based on local trusted templates by twin network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107844739B (en) | 2020-09-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107844739A (en) | Robustness target tracking method based on adaptive rarefaction representation simultaneously | |
| US10929649B2 (en) | Multi-pose face feature point detection method based on cascade regression | |
| CN109800689B (en) | Target tracking method based on space-time feature fusion learning | |
| CN107633226B (en) | Human body motion tracking feature processing method | |
| CN106446933B (en) | Multi-target detection method based on contextual information | |
| CN111563915B (en) | KCF target tracking method integrating motion information detection and Radon transformation | |
| CN105809672B (en) | A kind of image multiple target collaboration dividing method constrained based on super-pixel and structuring | |
| CN107146237B (en) | Target tracking method based on online state learning and estimation | |
| CN103488993B (en) | A kind of crowd's abnormal behaviour recognition methods based on FAST | |
| CN105320917B (en) | A kind of pedestrian detection and tracking based on head-shoulder contour and BP neural network | |
| CN103886325B (en) | Cyclic matrix video tracking method with partition | |
| CN106570490B (en) | A kind of pedestrian's method for real time tracking based on quick clustering | |
| CN107194408A (en) | A kind of method for tracking target of the sparse coordination model of mixed block | |
| CN101867699A (en) | Real-time tracking method of nonspecific target based on partitioning | |
| Yang et al. | MGC-VSLAM: A meshing-based and geometric constraint VSLAM for dynamic indoor environments | |
| CN104484890A (en) | Video target tracking method based on compound sparse model | |
| CN104599291B (en) | Infrared motion target detection method based on structural similarity and significance analysis | |
| CN106056627B (en) | A kind of robust method for tracking target based on local distinctive rarefaction representation | |
| CN105825233A (en) | Pedestrian detection method based on random fern classifier of online learning | |
| CN111652836A (en) | Multi-scale target detection method based on clustering algorithm and neural network | |
| Chen et al. | Robust vehicle detection and viewpoint estimation with soft discriminative mixture model | |
| CN109272036B (en) | Random fern target tracking method based on depth residual error network | |
| Ikram et al. | Real time hand gesture recognition using leap motion controller based on CNN-SVM architechture | |
| Zheng et al. | Attention assessment based on multi‐view classroom behaviour recognition | |
| CN103996207A (en) | Object tracking method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for 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 |
Granted publication date: 20200904 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |