CN109461166A - A kind of fast-moving target tracking based on KCF mixing MFO - Google Patents
A kind of fast-moving target tracking based on KCF mixing MFO Download PDFInfo
- Publication number
- CN109461166A CN109461166A CN201811255830.3A CN201811255830A CN109461166A CN 109461166 A CN109461166 A CN 109461166A CN 201811255830 A CN201811255830 A CN 201811255830A CN 109461166 A CN109461166 A CN 109461166A
- Authority
- CN
- China
- Prior art keywords
- target
- image block
- tracking
- candidate
- kcf
- 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.)
- Pending
Links
- 230000004044 response Effects 0.000 claims abstract description 46
- 238000000034 method Methods 0.000 claims abstract description 44
- 239000003795 chemical substances by application Substances 0.000 claims description 35
- 239000011159 matrix material Substances 0.000 claims description 26
- 230000006870 function Effects 0.000 claims description 23
- 238000012549 training Methods 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 8
- 238000006243 chemical reaction Methods 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 7
- 238000013507 mapping Methods 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- 230000009467 reduction Effects 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 5
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 230000015572 biosynthetic process Effects 0.000 claims description 3
- 230000021615 conjugation Effects 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 238000009499 grossing Methods 0.000 claims description 3
- 238000005259 measurement Methods 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 125000004122 cyclic group Chemical group 0.000 claims description 2
- 230000007246 mechanism Effects 0.000 abstract description 9
- 238000004422 calculation algorithm Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 4
- 238000007796 conventional method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000010845 search algorithm Methods 0.000 description 2
- 102100038784 Carbohydrate sulfotransferase 4 Human genes 0.000 description 1
- 101710158423 Carbohydrate sulfotransferase 4 Proteins 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000035772 mutation Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- 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
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention proposes a kind of fast-moving target trackings based on KCF mixing MFO, and its step are as follows: initialized target state parameter and Optimized model parameter；It executes KCF tracking and calculates maximum response of the target in former frames, and confidence threshold value initial value is found out with this；According to the relationship of present frame maximum response and confidence threshold value, different base sample image producing methods is determined: when maximum response is higher than confidence threshold value, randomly selecting basic image sample, execute KCF method and track target；When maximum response is lower than confidence threshold value, execution MFO search mechanisms search for globally optimal solution, and as new basic image sample, then execute KCF method and track target；Dynamic updates confidence threshold value, repeats effective tracking that aforesaid operations realize fast-moving target.The present invention can effectively realize duration tracking to the target quickly moved in video, improve the adaptability of the tracking under complex scene.
Description
Technical field
The present invention relates to the technical field of target following more particularly to a kind of fast-moving targets based on KCF mixing MFO
Tracking realizes the duration tracking of target, further includes that target generates under quick motion conditions in video adjacent image interframe
State space search mechanism and target Continuous tracking ability.
Background technique
The emphasis of vision tracking problem is the track for predicting mobile object in video.Currently, vision tracking has been applied
In many practical applications, such as robot service, autonomous driving vehicle, monitoring, human-computer interaction.Vision tracking is one and opens
It puts and challenging problem, tracker allows for tracking target object within a very long time, even in complexity
Scene in, such as target drift and background block.In initial frame, the initial position of target object is given, tracker should be able to
The uncertain variation of enough appearances and background that target is overcome during tracking, and target object is found in subsequent frames.
In recent decades, target following research has been achieved for very big progress, but due to many factors, such as illumination variation is blocked, carried on the back
Scape is mixed and disorderly, dimensional variation, deformation and the factors such as quickly moves, and in terms of computer vision is still one and challenging asks
Topic.In addition, since tracker is in a local area search target centered on the position that target is located at former frame, if mesh
Quickly movement or mutation movement occur for mark, the local search range beyond track algorithm, and without redefining for target and
In the case where the Restoration Mechanism of repositioning, when the tracking result of tracker output error, tracking of the tracking from mistake
As a result an inaccurate model is formed in, this may fundamentally influence tracking performance.
Summary of the invention
For the technical issues of operational efficiency of existing motion target tracking method is not high and poor universality, the present invention is proposed
Moth-flame (MFO) search strategy is introduced into nuclear phase and closed by a kind of fast-moving target tracking based on KCF mixing MFO
In filter tracker (KCF) design, combined based on traditional local search algorithm and optimization full search algorithm, and using dynamic
State threshold determination redefines target, improves operational efficiency, can adapt to the video frequency object tracking quickly moved.
In order to achieve the above object, the technical scheme of the present invention is realized as follows: it is a kind of based on the fast of KCF mixing MFO
Fast motion target tracking method, its step are as follows:
Step 1: the state parameter of the data information initialized target of first frame image is read, setting MFO is searched for and KCF
The Optimized model parameter of tracking；
Step 2: obtaining maximum response using KCF tracking, determines tracking target；
Step 3: the set element length of confidence threshold value is calculated according to setting in KCF tracking, most according to target
Big response calculates confidence threshold value；
Step 4: according to the maximum response of present frame and the relationship of confidence threshold value, different base sample images is determined
Producing method: if the maximum response R of the n-th f framemaxnf< Thrnf-1, Thrnf-1For the confidence threshold value of the n-th f frame, using flying
Moth-flame global search obtains base sample image, determines this image block areas of basic pattern, realizes target following in conjunction with KCF method；Such as
Fruit Rmaxnf≥Thrnf-1, then base sample image is obtained at random near the mapping position in (nf-1) frame according to the n-th f frame target
Block executes KCF method and is effectively tracked to the target of motion smoothing；Cyclic Moment is constructed using this image block of basic pattern of update
Battle array；
Step 5: updating confidence threshold value according to peak response, determines new basic image sample, and return step two determines
Track target.
The method that KCF tracking in the step 2 obtains maximum response are as follows:
A, the init state according to tracking target determines base sample image, generates target candidate sample set, and construction follows
Ring matrix X: the width and height of candidate region are respectively β * width and β * high, wherein and width is the width of target area,
High is the height of target area, and 1≤β < 2 is the scale factor of candidate region and target area；In column by target area conversion
Vector indicates, obtains target sample x=[x1,x2,…xn], wherein n=width*high；Using target sample x as basic pattern sheet
Vector, the circulative shift operation according to basic pattern sheet generate final candidate samples set, and candidate samples set contains basic pattern sheet
(n-1) a candidate samples of its generation of vector sum, the circular matrix of formation are as follows:
The first row of circular matrix X is the transposition x of base sample vectorT；
B, the time-frequency domain conversion of tracking problem, realizes the solution of classification problem: trained target is exactly to ask minimum flat
Weight w under square error, to obtain decision function f (z)=wTZ separates the target sample of tracking from candidate samples, power
Value w is obtained by following formula:
Wherein, xiFor i-th of training sample, yiIt is training sample xiCorresponding regressand value, z are observation samples or are to wait
Sampling sheet, λ are the regularization factors for controlling overfitting；
The vector description form of weight w solution under complex field are as follows:
W=(XHX+λI)-1XHy；
Wherein, I is recognition matrix, XHIt is the conjugation conversion of circular matrix X, and XH=(X*)T, X*It is answering for circular matrix X
Conjugate matrices；
The weight w of time domain is converted to the expression formula in frequency domain are as follows:
Wherein,For the Fourier Transform vector of x,For vectorComplex conjugate,For the Fourier transformation value of y, ⊙ table
Show that representation vector corresponding element is multiplied；
C, maximum response is obtained, determines tracking target:
Candidate samples z is chosen, candidate samples z is identical with the dimension of target sample x, according to formula:
Wherein, kxzIt is the core correlation of target sample x and candidate samples z, decision function value f (z) is a dimension and returns
The vector for returning value y the same, the position for finding out maximum element in vector at this time seek to the position of detection, and maximum element is most
Big response Rmax。
It is described that the weight w of time domain is converted to the method that frequency domain solves are as follows: using geo-nuclear tracin4 with target sample x's
Nonlinear combination indicates weightNon-linear Solve problems are converted into solving factor alphaiAnd Nonlinear MappingRelationship, enable K=C (kxx), the final frequency domain representation for obtaining vector α:
Wherein, xiIndicate that i-th of training sample, α are factor alphasiThe vector of composition, kxx=[k (x1,x1),k(x1,x2),k
(x1,x3),...,k(x1,xn)], kxxIt is the first row vector of circular matrix,Indicate vector kxxFourier transformation value；
Then the classification problem of target following is transformed into the solution in frequency domain to vector α, then through inversefouriertransform
Decision function form in time domain is obtained, realizes efficient object tracking process.
According to the maximum response of target in the step 3, setting k is the preceding k frame image of the n-th f frame image, and calculating is set
The method of confidence threshold are as follows:
If number of image frames nf < k, return step two re-execute；
If number of image frames nf=k, it is determined that initial confidence level threshold value Thr0Method are as follows: according to obtain preceding k frame k
A maximum response, construction set Rnf={ Rmax1,Rmax2…Rmaxk, then initial confidence level threshold value Thr0Are as follows:
Thr0=median (Rmax1,Rmax2…Rmaxk),
Wherein, median () is indicated to set RnfElement take intermediate value；
If number of image frames nf > k, according to the n-th f, the maximum response that nf-1, nf-2 ... ... nf- (k-1) frame obtain is dynamic
State updates set Rnf={ Rmax(nf),Rmax(nf-1)…Rmax(nf-(k-1)), then the confidence threshold value of the n-th f frame are as follows:
Thrnf=median { Rmaxnf,Rmax(nf-1)…Rmax(nf-(k-1))}。
Moth-flame search strategy method in the step 4 are as follows:
A1, initial position is generated, constructs objective function: randomly selects num initialization search agent position M1,M2…
Mnum, upper at various locations to choose image block, size is identical with tracking target, the objective function that constitution optimization process uses；
B1, candidate image block (flame) position and similarity value are updated.Calculate separately search agent image block and target figure
As the similarity metric function value OM=(OM of block1,OM2…OMnum)。
A. when current iteration number cur_iter is equal to 1, position and the similarity of search agent image block will be initialized
Value saves respectively as initial candidate tile location and similarity magnitude；
B. when current iteration number cur_iter is not equal to 1, by the similarity of search agent image block in cur_iter
It is updated to choose preferably image block composition compared with the candidate image block similarity value in (cur_iter) -1 iteration for value
Candidate image block position and similarity magnitude simultaneously save；Using the maximum correspondence image block of similarity value in candidate image block as
The best candidate image block of current iteration is stored in (Xbest,Ybest) in；
C1, update search agent position: m-th of search agent image block is around n-th of candidate image block according to logarithm spiral shell
Spin line formula is updated the position of search agent image block:
P(Mm,Fn)=Dm·eφt·cos(2πt)+Fn；
Wherein, MmRefer to m-th of search agent image block, FnRefer to n-th of candidate image block, DmRefer to m-th of search agent
Image block before n-th of candidate image block at a distance from, φ is the constant of logarithmic spiral wire shaped；T is between section [- 1,1]
Random number, t is defined as search agent image block in next degree for being closely located to candidate image block, and t=-1 is closest to
The position of flame, and t=1 illustrates the position farthest apart from flame；Distance DmAre as follows: Dm=| Fn-Mm|；
D1, candidate image number of blocks: adaptive reduction candidate image number of blocks is updated:
Wherein, iter is maximum number of iterations；Num is Maximum alternative image number of blocks；Round () expression rounds up
It is rounded；Each time in iteration in sequence reduction candidate image block corresponding to search agent image block then according to current suitable
The candidate image block for answering angle value worst updates its own position；
E1, an iteration are completed, and judge whether to reach maximum the number of iterations, and reaching termination condition terminates, and are exported optimal
Position as this image block of basic pattern.
The method of objective function is constructed in the step A1 are as follows: the HOG feature for extracting image block, the HOG of image block is special
Sign is used as stochastic variable, obtains the similarity measurement between image block:
Wherein, D () indicates variance, and Cov () indicates covariance, and M and N are the HOG feature of two image blocks respectively；
Objective function is defined as energy value: E=2+2* ρ (M, N).
The maximum response R that the n-th f+1 frame is calculated in step 4 is utilized in the step 5max(nf+1), update new set
Confidence threshold are as follows:
Thr(nf+1)=median { Rmax(nf+1),Rmaxnf…Rmax(nf-3)}。
Beneficial effects of the present invention: moth-flame (MFO) search strategy is introduced into nuclear phase and closes filter tracker (KCF)
In design, using the search mechanisms of global optimum, is determined according to confidence threshold value and generate base sample image, guarantee circular matrix energy
Enough coverage goal state spaces, enhance the ability for assessing quick motion state, compensating for traditional tracking can not adapt to fastly
The problem of speed movement；Using the feature of spatial domain and frequency domain processing data, the regression problem in spatial domain is equivalent to correlation
Filtering operation replaces time-consuming convolution algorithm with point multiplication operation, realizes the determination of candidate target sample responses value in frequency domain,
Improve operational efficiency；According to the adaptively selected different tracing modes of confidence threshold value, the same of quick motion problems can adapt to
When also take into account operational efficiency, avoid traditional tracking because using global optimization method solve motion problems due to bring calculating cost
Excessive phenomenon is of great significance to identification, understanding and the analysis of tracking target.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is flow chart of the invention.
Fig. 2 is the range accuracy comparison schematic diagram of the present invention with conventional method, wherein (a) is BLURCAR3 video, (b)
(c) it is FLEETFACE video for BLURFACE video, (d) is MAN video.
Fig. 3 is the Duplication comparison schematic diagram of the present invention with conventional method, wherein (a) is BLURCAR3 video, (b) is
BLURFACE video (c) is FLEETFACE video, (d) is MAN video.
Fig. 4 is the tracking effect schematic diagram of the present invention with conventional method, wherein (a) is BLURCAR3 video, (b) is
BLURFACE video (c) is FLEETFACE video, (d) is MAN video.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art are obtained every other under that premise of not paying creative labor
Embodiment shall fall within the protection scope of the present invention.
A kind of fast-moving target tracking based on KCF mixing MFO, main thought are: (1) by moth-flame
(MFO) search technique is introduced into KCF tracking problem, points out that a new research is thought to the research of fast-moving target tracking
Road；(2) search mechanisms that log spiral can be utilized based on moth-flame search mechanisms realize that global search and part are opened
The balance of hair, the motion model that target quickly moves can effectively be assessed by proposing；(3) a kind of new track algorithm frame is established
Frame can take into account the tracking of Target Sustainability existing for smooth and quick two kinds of motor patterns.Moth-flame search mechanisms are introduced
In KCF tracking problem, the novel tracking frame that can adapt to smooth and quick two kinds of motor patterns is proposed；It is proposed fortune of overall importance
Dynamic evaluation mechanism, can predict uncertainty of objective motion state, improve the acquisition capability of base sample image；It utilizes
Confidence threshold value switches different models, and the reliability of circular matrix is improved by the interaction of information, target is can adapt to and is scheming
The quick motion problems as existing for interframe.As shown in Figure 1, its step are as follows.
Step 1: the state parameter of the data information initialized target of first frame image is read, setting MFO is searched for and KCF
The Optimized model parameter of tracking.
Hardware environment of the present invention for implementation are as follows: Intel (R) Core (TM) i3 CPU 3.2G computer, 4GB memory
Software environment with 1G video card, operation is: Matlab R2014a and Windows 7, and the data used are the view of laboratory shooting
The database that frequency and teacher Wu Yi announce.The data information for reading first frame image, determines shape of the target in first frame image
State parameter [px,py, width, high], wherein (px,py) be target top left corner pixel point coordinate value, width be target it is wide
Degree, high is object height.Initial pictures block positional number num=50, the iteration optimization number iter=500 of MFO search are set,
Log spiral constant φ=2；Scale factor β=1.3 of candidate region and target area in KCF tracking are set, are arranged
Calculate the set element length k=5 of confidence threshold value.
Step 2: obtaining maximum response using KCF tracking, determines tracking target.
When new image arrives, maximum response is obtained using KCF tracking, positions mesh according to maximum response
Mark.The realization of KCF method is broadly divided into three steps:
A, the init state according to tracking target determines base sample image, generates target candidate sample set, and construction follows
Ring matrix X.
Base sample image carries out random sampling in the position that present frame maps according to previous frame target image.It is assumed that target
The width in region is width and height is high, and candidate samples are as the size of target template.The width of specified candidate region
It is respectively the scale factor that β * width and β * high, 1≤β < 2 is candidate region and target area with height.By target area
It is converted into column vector, obtains target sample x=[x1,x2,…xn], wherein n=width*high；Using vector x as basic pattern sheet
Vector, the circulative shift operation according to base sample image generate other candidate samples, the candidate samples set packet of circular matrix X
This vector x of basic pattern and its n-1 candidate samples generated are contained.The circular matrix X of formation is as follows:
The first row of circular matrix X is the transposition x of base sample vectorT。
B, the time-frequency domain conversion of tracking problem, realizes the solution of classification problem.
The behaviour that the process that KCF tracking carries out target following can be considered as target sample and candidate samples are persistently classified
Make.Known xiFor i-th of training sample, yiIt is training sample xiCorresponding regressand value, trained target are exactly to seek minimum square
Weight w under error, to obtain decision function f (z)=wTZ, z are observation sample, candidate samples or are test sample, are incited somebody to action
The target sample of tracking is separated from candidate samples.Weight w is obtained by following formula:
Wherein, λ is the regularization factors for controlling overfitting.This linear regression problem has closing solution, in complex field
The vector description form of lower solution are as follows:
W=(XHX+λI)-1XHy
Wherein, I is recognition matrix, XHIt is the conjugation conversion of circular matrix X, and XH=(X*)T, X*It is answering for circular matrix X
Conjugate matrices.
In order to obtain weight w, above formula calculation amount is larger, and efficiency is lower.Therefore, time domain complex calculation is transformed at frequency domain
Reason is to improve operational efficiency.Weight w is converted to the expression formula in frequency domain are as follows:
Wherein,For the Fourier Transform vector of x,ForComplex conjugate,For the Fourier transformation value of y, ⊙ indicates generation
Table vector corresponding element is multiplied.
Inverting in time domain is converted into frequency domain by element division, and computational efficiency greatly improves, and is become by anti-Fourier
Change acquisition weight w.Many problems belong to non-linear Solve problems in actual tracking, non-thread with target sample x using geo-nuclear tracin4
Property combination to indicate weightNon-linear Solve problems are converted into solving factor alphaiAnd Nonlinear Mapping's
Relationship, wherein xiIndicate i-th of training sample.Mathematical derivation under similar linear case, obtains the frequency domain representation of vector α:
Wherein, α is factor alphaiThe vector of composition.Wherein, kxx=[k (x1,x1),k(x1,x2),k(x1,x3),...,k(x1,
xn)], kxxIt is the first row vector of circular matrix, the Fourier transformation of ^ representation vector,Indicate vector kxxFourier transformation
Value.Then the classification problem of target following is transformed into the solution in frequency domain to vector α, when then obtaining through inversefouriertransform
Decision function form in domain realizes efficient object tracking process.
C, maximum response is obtained, determines tracking target.
Target detection is realized by the mode classification of classifier.Choose candidate samples z, candidate samples z and target sample x
Dimension is identical, according to formula:
Wherein, kxzIt is the core correlation of target sample x and candidate samples z, decision function value f (z) is a dimension and returns
The vector for returning value y the same finds out maximum element (maximum response R in vector at this timemax), determine that it is tracking result, i.e. mesh
Mark is repositioned.
Step 3: according to the maximum response of target, confidence threshold value is calculated；
If number of image frames nf < 5, return step two is re-executed；
If number of image frames nf=5, initial confidence level threshold value Thr is determined according to the following rules0:
5 maximum response R according to preceding 5 frame obtainedmax1,Rmax2…Rmax5, construction set Rnf={ Rmax1,Rmax2…
Rmax5, calculate initial confidence level threshold value Thr0It is as follows:
Thr0=median (Rmax1,Rmax2…Rmax5),
Wherein, median () is indicated to set RnfElement take intermediate value.
If number of image frames nf > 5, according to the n-th f, nf-1, nf-2, k-3, the maximum response dynamic that nf-4 frame obtains is more
New set Rnf={ Rmaxnf,Rmax(nf-1)…Rmax(nf-4), then the confidence threshold value of the n-th f frame are as follows:
Thrnf=median { Rmaxnf,Rmax(nf-1)…Rmax(nf-4)}。
Step 4: according to the maximum response of present frame and the relationship of confidence threshold value, different base sample images is determined
Producing method.
When (nf+1) frame image arrives, according to maximum response Rmax(nf+1)With confidence threshold value ThrnfCompare, with determination
The generation form of base sample image.
If Rmax(nf+1)< Thrnf, base sample image is obtained using moth-flame global search, determines base sample image
Target following is realized in conjunction with KCF method in block region.Moth-flame search strategy is as follows:
A1, initial position is generated, constructs objective function.Randomly select num=50 initialization search agent (moth) position
Set M1,M2…Mnum.Upper at various locations to choose image block, size is identical with tracking target, the target that constitution optimization process uses
Function.In the present invention, the HOG feature for extracting image block obtains image block using the HOG feature of image block as stochastic variable
Between similarity measurement:
Wherein, D () indicates variance, and Cov () indicates covariance, and M and N are the HOG feature of two image blocks respectively.Mesh
Scalar functions can be defined as foloows:
E=2+2* ρ (M, N)
Its functional value, that is, energy value E characterizes the similitude of two image blocks, energy value E two image blocks of bigger explanation more phase
Seemingly.
B1, candidate image block (flame) position and similarity value are updated.Calculate separately search agent image block and target figure
As the similarity metric function value OM=(OM of block1,OM2…OMnum)。
A. when current iteration number cur_iter is equal to 1, position and the similarity of search agent image block will be initialized
Value is stored in F=(F as initial candidate tile location and similarity magnitude respectively1,F2…Fnum) and OF=(OF1,
OF2…OFnum)。
B. when current iteration number cur_iter is not equal to 1, by the similarity of search agent image block in cur_iter
It is updated to choose preferably image block composition compared with the candidate image block similarity value in (cur_iter) -1 iteration for value
Candidate image block position and similarity magnitude are saved to F=(F1,F2…Fnum) and OF=(OF1,OF2…OFnum)。
It is protected the maximum correspondence image block of similarity value in candidate image block as the best candidate image block of current iteration
In the presence of (Xbest,Ybest) in.
C1, search agent position is updated.Search agent image block is actually the Search of Individual moved in search space,
And candidate image block is then the optimal location that search agent image block corresponding so far can reach.Each search generation
Reason image block is looped around around a corresponding candidate image block, wherein m-th of search agent image block surrounds n-th
Candidate image block is updated the position of search agent image block according to the log spiral formula in following formula:
P(Mm,Fn)=Dm·eφt·cos(2πt)+Fn
Wherein, MmRefer to m-th of search agent image block；FnRefer to n-th of candidate image block；DmRefer to m-th of search agent
Image block before n-th of candidate image block at a distance from；φ is the constant for defining logarithmic spiral wire shaped；T be section [-
1,1] random number between, t are defined as search agent image block in next degree (t=-1 for being closely located to candidate image block
It is closest to the position of flame, and t=1 illustrates the position farthest apart from flame)；Distance DmIt can be calculated by following formula:
Dm=| Fn-Mm|。
D1, candidate image number of blocks is updated.Adaptive reduction candidate image number of blocks, formula are as follows:
Iter is maximum number of iterations；Num is Maximum alternative image number of blocks；Round () indicates round.
With the increase of the number of iterations, candidate image number of blocks is gradually decreased, at this point, in every generation in sequence the candidate of reduction scheme
The search agent image block as corresponding to block then updates its own position according to the worst candidate image block of current fitness value.
E1, an iteration are completed, and judge whether to reach maximum the number of iterations, and reaching termination condition terminates, and are exported optimal
Position as this image block of basic pattern.Otherwise, B step is gone to.
If Rmax(nf+1)≥Thrnf, then random near the mapping position in (nf+1) frame according to the n-th f frame target to obtain
This image block of basic pattern executes KCF method and is effectively tracked to the target of motion smoothing.
Circular matrix X is constructed using this image block of basic pattern of update, maximum response is obtained using KCF tracking, it is real
Now to effective tracking of fast-moving target.
Step 5: confidence threshold value is updated according to peak response, determines new basic image sample, returns to second step, is determined
Track target.
(nf+1) frame maximum response R is calculated in step 4max(nf+1), then new confidence threshold value calculates as follows:
Thr(nf+1)=median { Rmax(nf+1),Rmaxnf…Rmax(nf-3)}。
According to the maximum response R of acquisitionmax, determine that candidate samples are new basic image sample.It is real to repeat step 2-five
Effective tracking of existing fast-moving target.
Implementation steps of the invention are as follows: initialized target state parameter and Optimized model parameter；Using KCF tracking
The maximum response in the former frames of target is obtained, and calculates confidence threshold value initial value；According to present frame maximum response with set
The relationship of confidence threshold determines different base sample image producing methods: when maximum response is higher than confidence threshold value, random choosing
Basic image sample is taken, KCF method is executed and tracks target；When maximum response be lower than confidence threshold value, obtained using MFO search mechanisms
Globally optimal solution is taken, new basic image sample is generated, KCF method is executed and tracks target；More using new maximum response dynamic
New confidence threshold value repeats effective tracking that aforesaid operations realize fast-moving target.The present invention quickly moves energy for target
Duration tracking is enough effectively realized, the adaptability of the tracking under complex scene is improved.
Effectiveness of the invention is evaluated using qualitative and quantitative two kinds of evaluation methods.Quantitative assessment mode uses center
Location error rate and target Duplication are evaluated.Center error is the center and true position by calculating tracking target
Euclidean distance between setting, generally its value is smaller illustrates that tracking result is more excellent.Target Duplication refers to tracking result and true
The ratio of target area area and operation, value is bigger, illustrates that tracking result is better.
Fig. 2 illustrates the range accuracy of the present invention with existing representative track algorithm KCF, CACF, DSST, FCT and LSST
DP value comparison result.Range accuracy DP refers to the frame number that target can be successfully tracked in the threshold binary image sequence according to setting and total
The ratio of frame number.Wherein, threshold value is determined by tracking result and the errors of centration value of legitimate reading, and threshold value is arranged in the present invention
It is 0.5.Fig. 3 shows the comparison result of the target Duplication OP value of corresponding track algorithm.
Qualitative evaluation mode uses the tracking effect figure of the method for the present invention and existing various exemplary process in partial frame,
As shown in Figure 4.Fig. 4 is tracking effect of tetra- videos of Blurcar3, Blurface, Fleetface and Man in partial frame respectively
Fruit figure, wherein there is the largest motion of 65 pixels to be displaced in first video, and second, third video and the 4th view
Frequency is after falling frame processing, they have the largest motion of 202 pixels, 125 pixels and 36 pixels to be displaced respectively, these are due to fast
The interframe big displacement that speed is moved and generated makes other track algorithms show to be not suitable with, and method proposed by the present invention obtains
Tracking effect.Complex chart 2 and Fig. 3 result indicate it is found that method for tracking target provided by the invention can well solve mesh
Quick motion problems are marked, preferably tracking performance is obtained.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (7)
1. a kind of fast-moving target tracking based on KCF mixing MFO, which is characterized in that its step are as follows:
Step 1: the state parameter of the data information initialized target of first frame image is read, setting MFO is searched for and KCF tracking
The Optimized model parameter of method；
Step 2: obtaining maximum response using KCF tracking, determines tracking target；
Step 3: calculating the set element length of confidence threshold value according to setting in KCF tracking, and the maximum according to target is rung
It should be worth, calculate confidence threshold value；
Step 4: according to the maximum response of present frame and the relationship of confidence threshold value, determine that different base sample images generates
Mode: if the maximum response R of the n-th f framemaxnf< Thrnf-1, Thrnf-1For the confidence threshold value of the n-th f frame, using moth-
Flame global search obtains base sample image, determines this image block areas of basic pattern, realizes target following in conjunction with KCF method；If
Rmaxnf≥Thrnf-1, then obtain this image block of basic pattern at random near the mapping position in (nf-1) frame according to the n-th f frame target,
KCF method is executed effectively to track the target of motion smoothing；Circular matrix is constructed using this image block of basic pattern of update；
Step 5: updating confidence threshold value according to peak response, determines new basic image sample, and return step two determines tracking
Target.
2. the fast-moving target tracking according to claim 1 based on KCF mixing MFO, which is characterized in that described
The method that KCF tracking in step 2 obtains maximum response are as follows:
A, the init state according to tracking target determines base sample image, generates target candidate sample set, constructs Cyclic Moment
Battle array X: the width and height of candidate region are respectively β * width and β * high, wherein width is the width of target area, high
For the height of target area, 1≤β < 2 is the scale factor of candidate region and target area；Target area is converted into column vector
It indicates, obtains target sample x=[x1,x2,…xn], wherein n=width*high；Using target sample x as base sample vector,
Circulative shift operation according to basic pattern sheet generates final candidate samples set, and candidate samples set contains this vector sum of basic pattern
Its (n-1) a candidate samples generated, the circular matrix of formation are as follows:
The first row of circular matrix X is the transposition x of base sample vectorT；
B, the time-frequency domain conversion of tracking problem, realizes the solution of classification problem: trained target is exactly to ask to minimize square mistake
Weight w under difference, to obtain decision function f (z)=wTZ separates the target sample of tracking, weight w from candidate samples
It is obtained by following formula:
Wherein, xiFor i-th of training sample, yiIt is training sample xiCorresponding regressand value, z are observation samples or are candidate sample
This, λ is the regularization factors for controlling overfitting；
The vector description form of weight w solution under complex field are as follows:
W=(XHX+λI)-1XHy；
Wherein, I is recognition matrix, XHIt is the conjugation conversion of circular matrix X, and XH=(X*)T, X*It is the complex conjugate of circular matrix X
Matrix；
The weight w of time domain is converted to the expression formula in frequency domain are as follows:
Wherein,For the Fourier Transform vector of x,For vectorComplex conjugate,For the Fourier transformation value of y, ⊙ indicates generation
Table vector corresponding element is multiplied；
C, maximum response is obtained, determines tracking target:
Candidate samples z is chosen, candidate samples z is identical with the dimension of target sample x, according to formula:
Wherein, kxzIt is the core correlation of target sample x and candidate samples z, decision function value f (z) is a dimension and regressand value y
The same vector, the position for finding out maximum element in vector at this time seek to the position of detection, and maximum element is maximum rings
It should value Rmax。
3. the fast-moving target tracking according to claim 2 based on KCF mixing MFO, which is characterized in that described
The weight w of time domain is converted into the method that frequency domain solves are as follows: using geo-nuclear tracin4 with the nonlinear combination of target sample x come
Indicate weightNon-linear Solve problems are converted into solving factor alphaiAnd Nonlinear MappingRelationship, enable K
=C (kxx), the final frequency domain representation for obtaining vector α:
Wherein, xiIndicate that i-th of training sample, α are factor alphasiThe vector of composition, kxx=[k (x1,x1),k(x1,x2),k(x1,
x3),...,k(x1,xn)], kxxIt is the first row vector of circular matrix,Indicate vector kxxFourier transformation value；
Then the classification problem of target following is transformed into the solution in frequency domain to vector α, then obtains through inversefouriertransform
Decision function form in time domain realizes efficient object tracking process.
4. the fast-moving target tracking according to claim 1 based on KCF mixing MFO, which is characterized in that described
According to the maximum response of target in step 3, setting k is the preceding k frame image of the n-th f frame image, calculates the side of confidence threshold value
Method are as follows:
If number of image frames nf < k, return step two re-execute；
If number of image frames nf=k, it is determined that initial confidence level threshold value Thr0Method are as follows: the k according to the preceding k frame obtained is most
Big response, construction set Rnf={ Rmax1,Rmax2…Rmaxk, then initial confidence level threshold value Thr0Are as follows:
Thr0=median (Rmax1,Rmax2…Rmaxk),
Wherein, median () is indicated to set RnfElement take intermediate value；
If number of image frames nf > k, according to the n-th f, the maximum response dynamic that nf-1, nf-2 ... ... nf- (k-1) frame obtain is more
New set Rnf={ Rmax(nf),Rmax(nf-1)…Rmax(nf-(k-1)), then the confidence threshold value of the n-th f frame are as follows:
Thrnf=median { Rmaxnf,Rmax(nf-1)…Rmax(nf-(k-1))}。
5. the fast-moving target tracking according to claim 1 based on KCF mixing MFO, which is characterized in that described
Moth-flame search strategy method in step 4 are as follows:
A1, initial position is generated, constructs objective function: randomly selects num initialization search agent position M1,M2…Mnum,
Image block is chosen on each position, size is identical with tracking target, the objective function that constitution optimization process uses；
B1, candidate image block (flame) position and similarity value are updated.Calculate separately search agent image block and target image block
Similarity metric function value OM=(OM1,OM2…OMnum)。
A. when current iteration number cur_iter is equal to 1, the position for initializing search agent image block and similarity value are made
It is saved respectively for initial candidate tile location and similarity magnitude；
B. when current iteration number cur_iter be not equal to 1 when, by the similarity value of search agent image block in cur_iter with
(cur_iter) the candidate image block similarity value in -1 iteration compares, and chooses preferably image block and forms updated candidate
Tile location and similarity magnitude simultaneously save；Using the maximum correspondence image block of similarity value in candidate image block as this
The best candidate image block of iteration is stored in (Xbest,Ybest) in；
C1, update search agent position: m-th of search agent image block is around n-th of candidate image block according to log spiral
Formula is updated the position of search agent image block:
P(Mm,Fn)=Dm·eφt·cos(2πt)+Fn；
Wherein, MmRefer to m-th of search agent image block, FnRefer to n-th of candidate image block, DmRefer to m-th of search agent image
Block before n-th of candidate image block at a distance from, φ is the constant of logarithmic spiral wire shaped；T be between section [- 1,1] with
Machine number, t are defined as search agent image block in next degree for being closely located to candidate image block, and t=-1 is closest to flame
Position, and t=1 illustrates the position farthest apart from flame；Distance DmAre as follows: Dm=| Fn-Mm|；
D1, candidate image number of blocks: adaptive reduction candidate image number of blocks is updated:
Wherein, iter is maximum number of iterations；Num is Maximum alternative image number of blocks；Round () indicates round；
Each time in iteration in sequence reduction candidate image block corresponding to search agent image block then according to current fitness
It is worth worst candidate image block and updates its own position；
E1, an iteration are completed, and judge whether to reach maximum the number of iterations, reaching termination condition terminates, and exports optimal position
It sets as this image block of basic pattern.
6. the fast-moving target tracking according to claim 5 based on KCF mixing MFO, which is characterized in that described
The method of objective function is constructed in step A1 are as follows: the HOG feature for extracting image block becomes using the HOG feature of image block as random
Amount obtains the similarity measurement between image block:
Wherein, D () indicates variance, and Cov () indicates covariance, and M and N are the HOG feature of two image blocks respectively；
Objective function is defined as energy value: E=2+2* ρ (M, N).
7. the fast-moving target tracking according to claim 5 based on KCF mixing MFO, which is characterized in that described
The maximum response R that the n-th f+1 frame is calculated in step 4 is utilized in step 5max(nf+1), update new confidence threshold value are as follows:
Thr(nf+1)=median { Rmax(nf+1),Rmaxnf…Rmax(nf-3)}。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811255830.3A CN109461166A (en) | 2018-10-26 | 2018-10-26 | A kind of fast-moving target tracking based on KCF mixing MFO |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811255830.3A CN109461166A (en) | 2018-10-26 | 2018-10-26 | A kind of fast-moving target tracking based on KCF mixing MFO |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109461166A true CN109461166A (en) | 2019-03-12 |
Family
ID=65608558
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811255830.3A Pending CN109461166A (en) | 2018-10-26 | 2018-10-26 | A kind of fast-moving target tracking based on KCF mixing MFO |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109461166A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109993777A (en) * | 2019-04-04 | 2019-07-09 | 杭州电子科技大学 | A kind of method for tracking target and system based on double-template adaptive threshold |
| CN110136165A (en) * | 2019-05-17 | 2019-08-16 | 河南科技学院 | A kind of mutation movement method for tracking target based on the optimization of adaptive whale |
| CN110717930A (en) * | 2019-08-28 | 2020-01-21 | 河南中烟工业有限责任公司 | Mutation moving target tracking method based on extended Wang-landau Monte Carlo and KCF |
| CN111739055A (en) * | 2020-06-10 | 2020-10-02 | 新疆大学 | Infrared point-like target tracking method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106557774A (en) * | 2015-09-29 | 2017-04-05 | 南京信息工程大学 | The method for real time tracking of multichannel core correlation filtering |
| CN107705306A (en) * | 2017-10-26 | 2018-02-16 | 中原工学院 | A kind of fabric defect detection method based on multiple features matrix low rank decomposition |
| CN107341820B (en) * | 2017-07-03 | 2018-04-20 | 郑州轻工业学院 | A kind of fusion Cuckoo search and the mutation movement method for tracking target of KCF |
-
2018
- 2018-10-26 CN CN201811255830.3A patent/CN109461166A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106557774A (en) * | 2015-09-29 | 2017-04-05 | 南京信息工程大学 | The method for real time tracking of multichannel core correlation filtering |
| CN107341820B (en) * | 2017-07-03 | 2018-04-20 | 郑州轻工业学院 | A kind of fusion Cuckoo search and the mutation movement method for tracking target of KCF |
| CN107705306A (en) * | 2017-10-26 | 2018-02-16 | 中原工学院 | A kind of fabric defect detection method based on multiple features matrix low rank decomposition |
Non-Patent Citations (1)
| Title |
|---|
| ZIDAN AM ET AL: "《Moth-flame Optimization Based Segmentation for MRI Liver Images》", 《INTERNATIONAL CONFERENCE ON ADVANCED INTELLIGENT SYSTEMS&INFORMATICS SPRINGER》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109993777A (en) * | 2019-04-04 | 2019-07-09 | 杭州电子科技大学 | A kind of method for tracking target and system based on double-template adaptive threshold |
| CN110136165A (en) * | 2019-05-17 | 2019-08-16 | 河南科技学院 | A kind of mutation movement method for tracking target based on the optimization of adaptive whale |
| CN110717930A (en) * | 2019-08-28 | 2020-01-21 | 河南中烟工业有限责任公司 | Mutation moving target tracking method based on extended Wang-landau Monte Carlo and KCF |
| CN110717930B (en) * | 2019-08-28 | 2024-02-20 | 河南中烟工业有限责任公司 | Mutation moving target tracking method based on extended Wang-Landau Monte Carlo and KCF |
| CN111739055A (en) * | 2020-06-10 | 2020-10-02 | 新疆大学 | Infrared point-like target tracking method |
| CN111739055B (en) * | 2020-06-10 | 2022-07-05 | 新疆大学 | Infrared point-like target tracking method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107341820B (en) | A kind of fusion Cuckoo search and the mutation movement method for tracking target of KCF | |
| CN109461166A (en) | A kind of fast-moving target tracking based on KCF mixing MFO | |
| Zhong et al. | Hierarchical tracking by reinforcement learning-based searching and coarse-to-fine verifying | |
| Zhang et al. | Learning regression and verification networks for robust long-term tracking | |
| CN107545582B (en) | Video multi-target tracking method and device based on fuzzy logic | |
| Chen et al. | SiamBAN: Target-aware tracking with Siamese box adaptive network | |
| CN104200237B (en) | One kind being based on the High-Speed Automatic multi-object tracking method of coring correlation filtering | |
| Yang et al. | SiamAtt: Siamese attention network for visual tracking | |
| CN110135314A (en) | A kind of multi-object tracking method based on depth Trajectory prediction | |
| CN110120064B (en) | Depth-related target tracking algorithm based on mutual reinforcement and multi-attention mechanism learning | |
| CN112651995B (en) | Online multi-target tracking method based on multifunctional aggregation and tracking simulation training | |
| Yuan et al. | Accurate bounding-box regression with distance-IoU loss for visual tracking | |
| CN112001428B (en) | Anchor frame-free target detection network training method based on feature matching optimization | |
| CN111292355A (en) | Nuclear correlation filtering multi-target tracking method fusing motion information | |
| CN106981071A (en) | A kind of method for tracking target applied based on unmanned boat | |
| CN113255611A (en) | Twin network target tracking method based on dynamic label distribution and mobile equipment | |
| Zhu et al. | Tiny object tracking: A large-scale dataset and a baseline | |
| CN108305272A (en) | A kind of mutation movement method for tracking target based on ALO search | |
| Gündoğdu et al. | The visual object tracking VOT2016 challenge results | |
| CN112802061A (en) | Robust target tracking method and system based on hierarchical decision network | |
| CN111105442B (en) | Switching type target tracking method | |
| Yang et al. | Adadet: An adaptive object detection system based on early-exit neural networks | |
| Zuo et al. | Double DQN method for object detection | |
| Li et al. | Progressive multistage learning for discriminative tracking | |
| CN110070120B (en) | Depth measurement learning method and system based on discrimination sampling strategy |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190312 |
|
| RJ01 | Rejection of invention patent application after publication |