CN109993776A - A kind of correlation filtering method for tracking target and system based on multistage template - Google Patents
A kind of correlation filtering method for tracking target and system based on multistage template Download PDFInfo
- Publication number
- CN109993776A CN109993776A CN201910270129.7A CN201910270129A CN109993776A CN 109993776 A CN109993776 A CN 109993776A CN 201910270129 A CN201910270129 A CN 201910270129A CN 109993776 A CN109993776 A CN 109993776A
- Authority
- CN
- China
- Prior art keywords
- response
- template
- translation
- max
- target
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/262—Analysis of motion using transform domain methods, e.g. Fourier domain methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- 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
- G06T2207/20048—Transform domain processing
- G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]
-
- 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)
- Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of correlation filtering method for tracking target and system based on multistage template.Its method includes the size for calculating multistage translation filter template, determines multistage translation filter template, determines multistage translation filter output response output, judge to translate filter template whether meet peak value of response requirement, in multistage translation filter template the suitable response output of selection, according to translating filter prediction target's center in the position of present frame.Method and system of the invention, which are solved, quickly to be moved in face of target and when the mixed and disorderly situation of background, the ineffective technical problem of tracker.
Description
Technical field
The invention belongs to computation vision target tracking domain, in particular to a kind of correlation filtering target based on multistage template
Tracking and system.
Background technique
Target following is the basic research of computer vision field, in sides such as precision control, automatic Pilot and robots
Face has a wide range of applications.Nearest lot of domestic and foreign researcher achieves many significant achievements in succession in this direction,
But in such methods application process, there is various challenges, including partial occlusion, quickly movement, motion blur, background are miscellaneous
Disorderly, situations such as illumination change, this kind of challenge causes tracker that can not accurately track target or even can track failure.To overcome view
The challenge content being likely to occur in frequency sequence, the tracking tool for designing a kind of robustness acquire a certain degree of difficulty.
Correlation filtering tracking is the MOSSE proposed by Bolme earliest, and this method is that the research of subsequent related fields is established
Basis is determined.Henriques is by introducing circulating sampling and utilizing Fourier's diagonalization property of Cyclic Moment speed-raising training and inspection
Survey process proposes CSK.Later, Henriques is further improved CSK, introduces kernel function and extends multichannel direction gradient feature
(HOG) feature extraction KCF.Danelljan M uses multichannel color characteristic Color Names (CN) on the basis of CSK,
The colouring information for efficiently using target, proposes CN.
Although these methods have been achieved with all well and good effect in target tracking domain, all only wrapped in their frame
Containing a filter template, which immobilizes, and the detection range of target must be consistent with the size of target,
This causes target search range limited.When there is the too fast equal peculair motions situation of movement velocity in target, the correlation of single template
Filtering method tends not to well adapt to, and leads to tracking drift even tracking failure.A kind of target that faces is needed quickly to move
Guarantee the technical solution of tracker effect when situation mixed and disorderly with background, for this purpose, proposing a kind of double-template dimension self-adaption correlation filter
Wave method for real time tracking and system.
Summary of the invention
When the technical problem to be solved by the present invention is to quickly move situation mixed and disorderly with background in face of target, tracker effect
Bad problem proposes a kind of correlation filtering method for tracking target and system based on multistage template.
The x-y coordinate system for indicating image pixel positions, target's center position (x are established in advancen,yn) indicate, wherein
N indicates frame number.Target's center position (the x of video sequence first frame1,y1) setting in advance, target size (high, width) is in advance
Setting, the velocity and acceleration of prior estimating target motion.
Correlation filtering method for tracking target based on multistage template of the invention, comprising the following steps:
Calculate the size of multistage translation filter template: reading the 1st frame of video sequence, according to target size (high,
) and the estimation peak value v of target speed widthmaxThe full-size for calculating translation filter, with variable sizemax=
(size_highmax,size_widthmax) indicate, according to the estimated acceleration a of target movement, target size and translation filter
Full-size calculates the rank m of translation filter template, calculates filter templates at different levels according to translation filter template rank m
Size is denoted as window_sz_i, wherein 1≤i≤m.
It is described
The rank of the translation filter template
Wherein γ is step-size factor set in advance, TminIt is minimum step set in advance.
The window_sz_1=Tmin(high, width), wherein TminIt is minimum step set in advance,
2≤i≤m, wherein γ is step-size factor set in advance.
Determine multistage translation filter template: heart position (x in the targetn,yn), the ruler of filter template is translated according to m grades
Very little window_sz_i determines Gaussian label yf_i, interception image block patch_for_train_i_n, and wherein n is indicated
Frame number；Image block characteristics are extracted respectively, and addition Cosine Window obtains translation features sample xf_for_train_i_n, high using translation
This type label and translation feature samples obtain m various sizes of translation filter templates, are indicated with α _ i；
The translation filter templateWherein, α indicates α _ i,It indicates in inverse Fu
Leaf transformation, ()*Indicate conjugation,Indicating the Fourier transformation of Gaussian label, λ is regularization parameter,It is nuclear matrix K
Generation sample Fourier transformation, nuclear matrix K is a circular matrix, the generation sample of matrix the first behavior nuclear matrix.
It determines the output response output of multistage translation filter: enabling n=n+1, video sequence n-th frame is read, in the (n-1)th frame
Target's center position (xn-1,yn-1), according to translation filter template size window_sz_i interception image block patch_for_
Det_i_n extracts characteristics of image and adds Cosine Window and obtain translation features sample zf_for_det_i_n to be detected, utilizes translation
Response output matrix response_i and peak value of response max_response_i is calculated in template α _ i.
It is described Indicate inverse Fourier transform,Indicate Fourier
Transformation, ⊙ representing matrix element point multiplication operation symbol, kxzIndicate the generator matrix of the nuclear matrix of sample x and sample z to be detected.
Judge translate filter template whether meet peak value of response requirement: judge peak value of response max_response_i (i's
Initial value is 1) whether to be greater than output response threshold value R set in advance, if so, determining i-stage translation translation filter response
Peak value meet demand enables response output matrix response=response_i, peak value of response max_response=max_
Response_i is entered step: otherwise prediction target's center determines that i-stage translates filter response peak in the position of present frame
Value is unsatisfactory for demand, and i=i+1 is entered step: the suitable response output of selection in multistage translation filter template.
The suitable response output of selection in multistage translation filter template: compare the translation filter template of two ranks
The peak value of response being calculated selects (i-1)-th grade of translation filter if max_response_i-1 > max_response_i
The response of template exports, even response output matrix response=response_i-1, peak value of response max_response=
max_response_i-1；Otherwise return step: judge to translate whether filter template meets peak value of response requirement.
According to translation filter prediction target's center in the position of present frame: exporting peak value according to the response of translation filter
Position of the max_response in response output matrix response, position (x of the prediction target's center in current n-th framen,
yn).Return step: multistage translation filter template is determined.
Correlation filtering Target Tracking System based on multistage template of the invention, comprising:
Video sequence；
Computer；
And
One or more programs, wherein one or more of programs are stored in the memory of computer, and by
It is configured to be executed by the processor of the computer, described program includes:
Calculate the size module of multistage translation filter template: reading the 1st frame of video sequence, according to target size (high,
) and the estimation peak value v of target speed widthmaxThe full-size for calculating translation filter, with variable sizemax=
(size_highmax,size_widthmax) indicate, according to the estimated acceleration a of target movement, target size and translation filter
Full-size calculates the rank m of translation filter template, calculates filter templates at different levels according to translation filter template rank m
Size is denoted as window_sz_i, wherein 1≤i≤m.
It is described
The rank of the translation filter template
Wherein γ is step-size factor set in advance, TminIt is minimum step set in advance.
The window_sz_1=Tmin(high, width), wherein TminIt is minimum step set in advance,
2≤i≤m, wherein γ is step-size factor set in advance.
Determine multistage translation filter template module: heart position (x in the targetn,yn), according to m grades of translation filter templates
Size window_sz_i, determine Gaussian label yf_i, interception image block patch_for_train_i_n, wherein n
Indicate frame number；Extract image block characteristics respectively, addition Cosine Window obtains translation features sample xf_for_train_i_n, using flat
It moves Gaussian label and translation feature samples obtains m various sizes of translation filter templates, indicated with α _ i；
The translation filter templateWherein, α indicates α _ i,It indicates in inverse Fu
Leaf transformation, ()*Indicate conjugation,Indicating the Fourier transformation of Gaussian label, λ is regularization parameter,It is nuclear matrix K
Generation sample Fourier transformation, nuclear matrix K is a circular matrix, the generation sample of matrix the first behavior nuclear matrix.
It determines the output response output module of multistage translation filter: enabling n=n+1, read video sequence n-th frame, the
N-1 frame target's center position (xn-1,yn-1), according to translation filter template size window_sz_i interception image block patch_
For_det_i_n extracts characteristics of image and adds Cosine Window and obtain translation features sample zf_for_det_i_n to be detected, utilizes
Response output matrix response_i and peak value of response max_response_i is calculated in translation template α _ i.
It is described Indicate inverse Fourier transform,Indicate Fourier
Transformation, ⊙ representing matrix element point multiplication operation symbol, kxzIndicate the generator matrix of the nuclear matrix of sample x and sample z to be detected.
Judge to translate whether filter template meets peak value of response requirement module: judging peak value of response max_response_i
Whether (initial value of i is 1) is greater than output response threshold value R set in advance, if so, determining i-stage translation translation filter
Peak value of response meet demand enables response output matrix response=response_i, peak value of response max_response=
Otherwise max_response_i determines i-stage translation filter response into prediction target's center in the position module of present frame
Peak value is unsatisfactory for demand, i=i+1, into the suitable response output module of selection in multistage translation filter template.
The suitable response output module of selection in multistage translation filter template: compare the translation filter of two ranks
The peak value of response that formwork calculation obtains selects (i-1)-th grade of translation filter if max_response_i-1 > max_response_i
The response of wave device template exports, even response output matrix response=response_i-1, peak value of response max_
Response=max_response_i-1；Otherwise it returns and judges whether translation filter template meets peak value of response and want modulus
Block.
According to translation filter prediction target's center present frame position module: according to translation filter response export
Position of the peak value max_response in response output matrix response, predicts target's center in the position of current n-th frame
(xn,yn).It returns and determines multistage translation filter template module.
Present invention has the advantage that
(1) when facing fast-moving target, suitable filter template quantity and filter template size are chosen, even if
Target is moved with maximum speed, and detection range still may include complete target；
(2) using multistage translation filter template predicted motion target position, make target detection process searches variable range,
It is adapted to the changeable movement velocity of target；
(3) in final future position, the testing result of multiple filter templates is compared, until response exports peak value
It is more other than upper level low, can to avoid filter template it is excessive caused by include excessive mixed and disorderly background and response in detection range
Export influence of the fluctuation generated to final predicted position；
Detailed description of the invention
Fig. 1 is the correlation filtering method for tracking target flow chart based on multistage template of the embodiment of the present invention；
Fig. 2 is the correlation filtering Target Tracking System structural schematic diagram based on multistage template of the embodiment of the present invention.
Specific embodiment
It elaborates below to the preferred embodiment of the present invention.
The x-y coordinate system for indicating image pixel positions, target's center position (x are established in advancen,yn) indicate, wherein
N indicates frame number.Target's center position (the x of video sequence first frame1,y1) setting in advance, target size (high, width) is in advance
Setting, the velocity and acceleration of prior estimating target motion.In the present embodiment, in image pixel positions coordinate system, image upper left
The position of angle pixel is (1,1), provides target's center position (x in first frame image1,y1)=(48,65), target size is
20 pixels × 20 pixels, i.e. high=20, width=20, the velocity peak values v of prior estimating target motionmax=2.5pix/ms,
Acceleration a=0.025pix/ms2。
Correlation filtering method for tracking target based on multistage template of the invention, comprising the following steps:
Calculate the size of multistage translation filter template: reading the 1st frame of video sequence, according to target size (high,
) and the estimation peak value v of target speed widthmaxThe full-size for calculating translation filter, with variable sizemax=
(size_highmax,size_widthmax) indicate, according to the estimated acceleration a of target movement, target size and translation filter
Full-size calculates the rank m of translation filter template, calculates filter templates at different levels according to translation filter template rank m
Size is denoted as window_sz_i, wherein 1≤i≤m.
It is described
The rank of the translation filter template
Wherein γ is step-size factor set in advance, TminIt is minimum step set in advance.
The window_sz_1=Tmin(high, width), wherein TminIt is minimum step set in advance,
2≤i≤m, wherein γ is step-size factor set in advance.In the present embodiment, video frame rate 50fps is then calculated Step-size factor γ=16 set in advance, minimum step T set in advancemin=
1.2, then calculate the rank of translation filter templateAnd it is full
Foot Then m=5, root
The size of filter templates at different levels, respectively window_sz_1=T are calculated according to translation filter template rank mmin·(high,
Width)=(24,24), window_sz_2=(Tmin+ (i-1) γ a) (high, width)=(1.2+16 ×
0.025) × 20=(32,32), and so on, window_sz_3=(40,40), window_sz_4=(48,48),
Window_sz_5=(56,56).
Determine multistage translation filter template: heart position (x in the targetn,yn), the ruler of filter template is translated according to m grades
Very little window_sz_i determines Gaussian label yf_i, interception image block patch_for_train_i_n, and wherein n is indicated
Frame number；Image block characteristics are extracted respectively, and addition Cosine Window obtains translation features sample xf_for_train_i_n, high using translation
This type label and translation feature samples obtain m various sizes of translation filter templates, are indicated with α _ i；
The translation filter templateWherein, α indicates α _ i,It indicates in inverse Fu
Leaf transformation, ()*Indicate conjugation,Indicating the Fourier transformation of Gaussian label, λ is regularization parameter,It is nuclear matrix K
Generation sample Fourier transformation, nuclear matrix K is a circular matrix, the generation sample of matrix the first behavior nuclear matrix.This
In embodiment, according to template size size, Gaussian label yf_i is determined, it is 1 that tag hub, which is maximized, and surrounding numerical value is gradually
Reduce, edge 0, numerical value is at Gaussian Profile.Heart position (48,65) in the target, according to search box size, (i.e. template size is big
Bear) interception image block patch_for_train_i_1, extracts image block characteristics respectively, then adds Cosine Window and obtain feature
Sample xf_for_train_i_1, size be respectively (24,24), (32,32), (40,40), (48,48), (56,56), here
Cosine Window is equivalent to a weight matrix, assigns focus target region bigger weight, smaller closer to edge weights.Final root
According to ridge regression training pattern, using feature samples and translation Gaussian label be calculated 5 ranks filter template α _ 1,
α _ 2, α _ 3, α _ 4 and α _ 5.
It determines the output response output of multistage translation filter: enabling n=n+1, video sequence n-th frame is read, in the (n-1)th frame
Target's center position (xn-1,yn-1), according to translation filter template size window_sz_i interception image block patch_for_
Det_i_n extracts characteristics of image and adds Cosine Window and obtain translation features sample zf_for_det_i_n to be detected, utilizes translation
Response output matrix response_i and peak value of response max_response_i is calculated in template α _ i.
It is described Indicate inverse Fourier transform,Indicate Fourier
Transformation, ⊙ representing matrix element point multiplication operation symbol, kxzIndicate the generator matrix of the nuclear matrix of sample x and sample z to be detected.This reality
It applies in example, n=1+1=2, in the 1st frame target's center position (48,65), according to translation filter template size window_sz_i
Interception image block patch_for_det_i_n extracts characteristics of image and adds Cosine Window and obtain translation features sample zf_ to be detected
Response output matrix response_i and peak value of response max_ is calculated using translation template α _ i in for_det_i_n
Response_i, wherein max_response_1=0.7, max_response_2=0.7, max_response_3=0.6,
Max_response_4=0.6, max_response_5=0.5.
Judge translate filter template whether meet peak value of response requirement: judge peak value of response max_response_i (i's
Initial value is 1) whether to be greater than output response threshold value R set in advance, if so, determining i-stage translation translation filter response
Peak value meet demand enables response output matrix response=response_i, peak value of response max_response=max_
Response_i is entered step: otherwise prediction target's center determines that i-stage translates filter response peak in the position of present frame
Value is unsatisfactory for demand, and i=i+1 is entered step: the suitable response output of selection in multistage translation filter template.This implementation
In example, output response threshold value R=0.8 set in advance judge max_response_1=0.7 < 0.8 in the present embodiment, judgement
1st grade of translation filter peak value of response is unsatisfactory for demand, and i=i+1=2 is entered step: selecting in multistage translation filter template
Select suitable response output.
The suitable response output of selection in multistage translation filter template: compare the translation filter template of two ranks
The peak value of response being calculated selects (i-1)-th grade of translation filter if max_response_i-1 > max_response_i
The response of template exports, even response output matrix response=response_i-1, peak value of response max_response=
max_response_i-1；Otherwise return step: judge to translate whether filter template meets peak value of response requirement.The present embodiment
In, then return step: max_response_1=max_response_2 judges to translate whether filter template meets response peak
Value requires.
In step: judge to translate whether filter template meets in peak value of response requirement, max_response_2=0.7 <
0.8, determine that the 2nd grade of translation filter peak value of response is unsatisfactory for demand, i=i+1=3 is entered step: translating filter in multistage
Suitable response output is selected in template.
In step: in multistage translation filter template in the suitable response output of selection, max_response_2=0.7
> max_response_3=0.6 then selects the response output of the 2nd grade of translation filter template, even response output matrix
Response=response_2, peak value of response max_response=max_response_2.
According to translation filter prediction target's center in the position of present frame: exporting peak value according to the response of translation filter
Position of the max_response in response output matrix response, position (x of the prediction target's center in current n-th framen,
yn).Return step: multistage translation filter template is determined.In the present embodiment, peak value is exported according to the response of translation filter
Position of the max_response in response output matrix response, position (x of the prediction target's center in current 2nd frame2,
y2)=(50,67), return step: determine multistage translation filter template.
Update translation filter after, read video sequence next frame, executed according to above-mentioned steps, until video last
Frame.
The correlation filtering method for tracking target flow chart based on multistage template of the present embodiment, as shown in Figure 1.
The correlation filtering Target Tracking System based on multistage template of the present embodiment, comprising:
Video sequence；
Computer；
And
One or more programs, wherein one or more of programs are stored in the memory of computer, and by
It is configured to be executed by the processor of the computer, described program includes:
Calculate the size module of multistage translation filter template: reading the 1st frame of video sequence, according to target size (high,
) and the estimation peak value v of target speed widthmaxThe full-size for calculating translation filter, with variable sizemax=
(size_highmax,size_widthmax) indicate, according to the estimated acceleration a of target movement, target size and translation filter
Full-size calculates the rank m of translation filter template, calculates filter templates at different levels according to translation filter template rank m
Size is denoted as window_sz_i, wherein 1≤i≤m.
It is described
The rank of the translation filter template
Wherein γ is step-size factor set in advance, TminIt is minimum step set in advance.
The window_sz_1=Tmin(high, width), wherein TminIt is minimum step set in advance,
2≤i≤m, wherein γ is step-size factor set in advance.In the present embodiment, video frame rate 50fps is then calculated Step-size factor γ=16 set in advance, minimum step T set in advancemin=
1.2, then calculate the rank of translation filter templateAnd it is full
Foot Then m=5, root
The size of filter templates at different levels, respectively window_sz_1=T are calculated according to translation filter template rank mmin·(high,
Width)=(24,24), window_sz_2=(Tmin+ (i-1) γ a) (high, width)=(1.2+16 ×
0.025) × 20=(32,32), and so on, window_sz_3=(40,40), window_sz_4=(48,48),
Window_sz_5=(56,56).
Determine multistage translation filter template module: heart position (x in the targetn,yn), according to m grades of translation filter templates
Size window_sz_i, determine Gaussian label yf_i, interception image block patch_for_train_i_n, wherein n
Indicate frame number；Extract image block characteristics respectively, addition Cosine Window obtains translation features sample xf_for_train_i_n, using flat
It moves Gaussian label and translation feature samples obtains m various sizes of translation filter templates, indicated with α _ i；
The translation filter templateWherein, α indicates α _ i,It indicates in inverse Fu
Leaf transformation, ()*Indicate conjugation,Indicating the Fourier transformation of Gaussian label, λ is regularization parameter,It is nuclear matrix K
Generation sample Fourier transformation, nuclear matrix K is a circular matrix, the generation sample of matrix the first behavior nuclear matrix.This
In embodiment, according to template size size, Gaussian label yf_i is determined, it is 1 that tag hub, which is maximized, and surrounding numerical value is gradually
Reduce, edge 0, numerical value is at Gaussian Profile.Heart position (48,65) in the target, according to search box size, (i.e. template size is big
Bear) interception image block patch_for_train_i_1, extracts image block characteristics respectively, then adds Cosine Window and obtain feature
Sample xf_for_train_i_1, size be respectively (24,24), (32,32), (40,40), (48,48), (56,56), here
Cosine Window is equivalent to a weight matrix, assigns focus target region bigger weight, smaller closer to edge weights.Final root
According to ridge regression training pattern, using feature samples and translation Gaussian label be calculated 5 ranks filter template α _ 1,
α _ 2, α _ 3, α _ 4 and α _ 5.
It determines the output response output module of multistage translation filter: enabling n=n+1, read video sequence n-th frame, the
N-1 frame target's center position (xn-1,yn-1), according to translation filter template size window_sz_i interception image block patch_
For_det_i_n extracts characteristics of image and adds Cosine Window and obtain translation features sample zf_for_det_i_n to be detected, utilizes
Response output matrix response_i and peak value of response max_response_i is calculated in translation template α _ i.
It is described Indicate inverse Fourier transform,Indicate Fourier
Transformation, ⊙ representing matrix element point multiplication operation symbol, kxzIndicate the generator matrix of the nuclear matrix of sample x and sample z to be detected.This reality
It applies in example, n=1+1=2, in the 1st frame target's center position (48,65), according to translation filter template size window_sz_i
Interception image block patch_for_det_i_n extracts characteristics of image and adds Cosine Window and obtain translation features sample zf_ to be detected
Response output matrix response_i and peak value of response max_ is calculated using translation template α _ i in for_det_i_n
Response_i, wherein max_response_1=0.7, max_response_2=0.7, max_response_3=0.6,
Max_response_4=0.6, max_response_5=0.5.
Judge to translate whether filter template meets peak value of response requirement module: judging peak value of response max_response_i
Whether (initial value of i is 1) is greater than output response threshold value R set in advance, if so, determining i-stage translation translation filter
Peak value of response meet demand enables response output matrix response=response_i, peak value of response max_response=
Otherwise max_response_i determines i-stage translation filter response into prediction target's center in the position module of present frame
Peak value is unsatisfactory for demand, i=i+1, into the suitable response output module of selection in multistage translation filter template.This implementation
In example, output response threshold value R=0.8 set in advance judge max_response_1=0.7 < 0.8 in the present embodiment, judgement
1st grade of translation filter peak value of response is unsatisfactory for demand, and i=i+1=2 selects to close into multistage translation filter template
Suitable response output module.
The suitable response output module of selection in multistage translation filter template: compare the translation filter of two ranks
The peak value of response that formwork calculation obtains selects (i-1)-th grade of translation filter if max_response_i-1 > max_response_i
The response of wave device template exports, even response output matrix response=response_i-1, peak value of response max_
Response=max_response_i-1；Otherwise it returns and judges whether translation filter template meets peak value of response and want modulus
Block.In the present embodiment, max_response_1=max_response_2 is then returned and is judged to translate whether filter template meets
Peak value of response requires module.
Judging to translate whether filter template meets in peak value of response requirement module, max_response_2=0.7 <
0.8, determine that the 2nd grade of translation filter peak value of response is unsatisfactory for demand, i=i+1=3 translates filter template into multistage
The middle suitable response output module of selection.
It is being selected in suitable response output module in multistage translation filter template, max_response_2=0.7 >
Max_response_3=0.6 then selects the response output of the 2nd grade of translation filter template, even response output matrix
Response=response_2, peak value of response max_response=max_response_2.
According to translation filter prediction target's center present frame position module: according to translation filter response export
Position of the peak value max_response in response output matrix response, predicts target's center in the position of current n-th frame
(xn,yn).It returns and determines multistage translation filter template module.In the present embodiment, peak value is exported according to the response of translation filter
Position of the max_response in response output matrix response, position (x of the prediction target's center in current 2nd frame2,
y2)=(50,67), it returns and determines multistage translation filter template module.
Update translation filter after, read video sequence next frame, executed according to above-mentioned steps, until video last
Frame.
The correlation filtering Target Tracking System structural schematic diagram based on multistage template of the present embodiment, as shown in Figure 2.
Certainly, those of ordinary skill in the art is it should be appreciated that above embodiments are intended merely to illustrate this hair
It is bright, and be not intended as limitation of the invention, as long as within the scope of the invention, all to the variations of above embodiments, modification
Protection scope of the present invention will be fallen into.
Claims (10)
1. a kind of correlation filtering method for tracking target based on multistage template, it is characterised in that the following steps are included:
Calculate the size of multistage translation filter template: reading the 1st frame of video sequence, according to target size (high, width) and
The estimation peak value v of target speedmaxThe full-size for calculating translation filter, with variable sizemax=(size_highmax,
size_widthmax) indicate, it is calculated according to the estimated acceleration a of target movement, target size and translation filter full-size
The rank m for translating filter template calculates the size of filter templates at different levels according to translation filter template rank m, is denoted as
Window_sz_i, wherein 1≤i≤m；
Determine multistage translation filter template: heart position (x in the targetn,yn), the size of filter template is translated according to m grades
Window_sz_i determines Gaussian label yf_i, interception image block patch_for_train_i_n, and wherein n indicates frame
Number；Image block characteristics are extracted respectively, and addition Cosine Window obtains translation features sample xf_for_train_i_n, utilizes translation Gauss
Type label and translation feature samples obtain m various sizes of translation filter templates, are indicated with α _ i；
It determines the output response output of multistage translation filter: enabling n=n+1, video sequence n-th frame is read, in the (n-1)th frame target
Center (xn-1,yn-1), according to translation filter template size window_sz_i interception image block patch_for_det_i_
N extracts characteristics of image and adds Cosine Window and obtain translation features sample zf_for_det_i_n to be detected, using translation template α _
Response output matrix response_i and peak value of response max_response_i is calculated in i；
Judge to translate whether filter template meets peak value of response requirement: judging whether peak value of response max_response_i is greater than
Output response threshold value R set in advance, wherein the initial value of i is 1, if so, determining i-stage translation translation filter response peak
It is worth meet demand, enables response output matrix response=response_i, peak value of response max_response=max_
Response_i is entered step: otherwise prediction target's center determines that i-stage translates filter response peak in the position of present frame
Value is unsatisfactory for demand, and i=i+1 is entered step: the suitable response output of selection in multistage translation filter template；
The suitable response output of selection in multistage translation filter template: the translation filter template for comparing two ranks calculates
Obtained peak value of response selects (i-1)-th grade of translation filter template if max_response_i-1 > max_response_i
Response output, even response output matrix response=response_i-1, peak value of response max_response=max_
response_i-1；Otherwise return step: judge to translate whether filter template meets peak value of response requirement；
According to translation filter prediction target's center in the position of present frame: exporting peak value max_ according to the response of translation filter
Position of the response in response output matrix response, position (x of the prediction target's center in current n-th framen,yn)；It returns
It returns step: determining multistage translation filter template.
2. the correlation filtering method for tracking target according to claim 1 based on multistage template, which is characterized in that described
3. the correlation filtering method for tracking target according to claim 1 based on multistage template, which is characterized in that described flat
The rank of shift filter template
Wherein γ is step-size factor set in advance, TminIt is minimum step set in advance；The window_sz_1=Tmin·
(high, width), wherein TminIt is minimum step set in advance,
2≤i≤m, wherein γ is step-size factor set in advance.
4. the correlation filtering method for tracking target according to claim 1 based on multistage template, which is characterized in that described flat
Shift filter templateWherein, α indicates α _ i,Indicate inverse Fourier transform, ()*Table
Show conjugation,Indicating the Fourier transformation of Gaussian label, λ is regularization parameter,It is Fu of the generation sample of nuclear matrix K
In leaf transformation, nuclear matrix K is a circular matrix, the generation sample of matrix the first behavior nuclear matrix.
5. the correlation filtering method for tracking target according to claim 1 based on multistage template, which is characterized in that described Indicate inverse Fourier transform,Indicate Fourier transformation, ⊙ indicates square
Array element vegetarian refreshments multiplication symbol, kxzIndicate the generator matrix of the nuclear matrix of sample x and sample z to be detected.
6. a kind of correlation filtering Target Tracking System based on multistage template, characterized by comprising:
Video sequence；
Computer
And
One or more programs wherein one or more of programs are stored in the memory of computer, and are configured
At the processor execution by the computer, described program includes:
Calculate the size module of multistage translation filter template: reading the 1st frame of video sequence, according to target size (high,
) and the estimation peak value v of target speed widthmaxThe full-size for calculating translation filter, with variable sizemax=
(size_highmax,size_widthmax) indicate, according to the estimated acceleration a of target movement, target size and translation filter
Full-size calculates the rank m of translation filter template, calculates filter templates at different levels according to translation filter template rank m
Size is denoted as window_sz_i, wherein 1≤i≤m；
Determine multistage translation filter template module: heart position (x in the targetn,yn), the ruler of filter template is translated according to m grades
Very little window_sz_i determines Gaussian label yf_i, interception image block patch_for_train_i_n, and wherein n is indicated
Frame number；Image block characteristics are extracted respectively, and addition Cosine Window obtains translation features sample xf_for_train_i_n, high using translation
This type label and translation feature samples obtain m various sizes of translation filter templates, are indicated with α _ i；
It determines the output response output module of multistage translation filter: enabling n=n+1, video sequence n-th frame is read, in the (n-1)th frame
Target's center position (xn-1,yn-1), according to translation filter template size window_sz_i interception image block patch_for_
Det_i_n extracts characteristics of image and adds Cosine Window and obtain translation features sample zf_for_det_i_n to be detected, utilizes translation
Response output matrix response_i and peak value of response max_response_i is calculated in template α _ i；
Judge to translate whether filter template meets peak value of response requirement: judging whether peak value of response max_response_i is greater than
Output response threshold value R set in advance, wherein the initial value of i is 1, if so, determining i-stage translation translation filter response peak
It is worth meet demand, enables response output matrix response=response_i, peak value of response max_response=max_
Response_i is entered step: otherwise prediction target's center determines that i-stage translates filter response peak in the position of present frame
Value is unsatisfactory for demand, i=i+1, into the suitable response output module of selection in multistage translation filter template；
The suitable response output module of selection in multistage translation filter template: compare the translation filter template of two ranks
The peak value of response being calculated selects (i-1)-th grade of translation filter if max_response_i-1 > max_response_i
The response of template exports, even response output matrix response=response_i-1, peak value of response max_response=
max_response_i-1；Otherwise it returns and judges to translate whether filter template meets peak value of response requirement module；
According to translation filter prediction target's center present frame position module: according to translation filter response export peak value
Position of the max_response in response output matrix response, position (x of the prediction target's center in current n-th framen,
yn), it returns and determines multistage translation filter template module.
7. the correlation filtering target following according to claim 6 based on multistage template, which is characterized in that described
8. the correlation filtering target following according to claim 6 based on multistage template, which is characterized in that the translation filter
The rank of wave device template
Wherein γ is step-size factor set in advance, TminIt is minimum step set in advance；The window_sz_1=Tmin·
(high, width), wherein TminIt is minimum step set in advance,
2≤i≤m, wherein γ is step-size factor set in advance.
9. the correlation filtering target following according to claim 6 based on multistage template, which is characterized in that the translation filter
Wave device templateWherein, α indicates α _ i,Indicate inverse Fourier transform, ()*It indicates altogether
Yoke,Indicating the Fourier transformation of Gaussian label, λ is regularization parameter,It is the Fourier of the generation sample of nuclear matrix K
Transformation, nuclear matrix K is a circular matrix, the generation sample of matrix the first behavior nuclear matrix.
10. the correlation filtering target following according to claim 6 based on multistage template, which is characterized in that described Indicate inverse Fourier transform,Indicate Fourier transformation, ⊙ indicates square
Array element vegetarian refreshments multiplication symbol, kxzIndicate the generator matrix of the nuclear matrix of sample x and sample z to be detected.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910270129.7A CN109993776B (en) | 2019-04-04 | 2019-04-04 | Related filtering target tracking method and system based on multi-level template |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910270129.7A CN109993776B (en) | 2019-04-04 | 2019-04-04 | Related filtering target tracking method and system based on multi-level template |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109993776A true CN109993776A (en) | 2019-07-09 |
| CN109993776B CN109993776B (en) | 2020-12-11 |
Family
ID=67132366
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910270129.7A Active CN109993776B (en) | 2019-04-04 | 2019-04-04 | Related filtering target tracking method and system based on multi-level template |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109993776B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490899A (en) * | 2019-07-11 | 2019-11-22 | 东南大学 | A kind of real-time detection method of the deformable construction machinery of combining target tracking |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108846851A (en) * | 2018-04-25 | 2018-11-20 | 河北工业职业技术学院 | Motion target tracking method and terminal device |
| CN109087333A (en) * | 2018-06-14 | 2018-12-25 | 中国科学院福建物质结构研究所 | Target scale estimation method and its device based on correlation filter tracking algorithm |
| CN109543615A (en) * | 2018-11-23 | 2019-03-29 | 长沙理工大学 | A kind of double learning model method for tracking target based on multi-stage characteristics |
-
2019
- 2019-04-04 CN CN201910270129.7A patent/CN109993776B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108846851A (en) * | 2018-04-25 | 2018-11-20 | 河北工业职业技术学院 | Motion target tracking method and terminal device |
| CN109087333A (en) * | 2018-06-14 | 2018-12-25 | 中国科学院福建物质结构研究所 | Target scale estimation method and its device based on correlation filter tracking algorithm |
| CN109543615A (en) * | 2018-11-23 | 2019-03-29 | 长沙理工大学 | A kind of double learning model method for tracking target based on multi-stage characteristics |
Non-Patent Citations (2)
| Title |
|---|
| CONGMEI ZHAO: "An algorithm of adaptive scale object tracking in occlusion", 《 AIP CONFERENCE PROCEEDINGS》 * |
| 韩光星，李崇荣: "网络视频监控中运动目标跟踪方法改进", 《通信学报》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490899A (en) * | 2019-07-11 | 2019-11-22 | 东南大学 | A kind of real-time detection method of the deformable construction machinery of combining target tracking |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109993776B (en) | 2020-12-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108734723B (en) | Relevant filtering target tracking method based on adaptive weight joint learning | |
| CN109087337B (en) | Long-time target tracking method and system based on hierarchical convolution characteristics | |
| Zhang et al. | Learning scale-adaptive tight correlation filter for object tracking | |
| CN112949408B (en) | Real-time identification method and system for target fish passing through fish channel | |
| CN112183537B (en) | Model training method and device, and text region detection method and device | |
| Zhang et al. | Sparse learning-based correlation filter for robust tracking | |
| CN111429481B (en) | Target tracking method, device and terminal based on adaptive expression | |
| CN111914878A (en) | Feature point tracking training and tracking method and device, electronic equipment and storage medium | |
| CN116665095B (en) | Method and system for detecting motion ship, storage medium and electronic equipment | |
| CN112990298A (en) | Key point detection model training method, key point detection method and device | |
| Zhao et al. | Effective local and global search for fast long-term tracking | |
| CN109993776A (en) | A kind of correlation filtering method for tracking target and system based on multistage template | |
| CN107566826B (en) | The test method and device of grating image processor | |
| CN111507135A (en) | Face detection method and device, computer equipment and storage medium | |
| KR20220103370A (en) | Method and apparatus for detecting out-of-distribution using noise filter | |
| Kuipers et al. | Hard occlusions in visual object tracking | |
| Jiang et al. | MultiBSP: multi-branch and multi-scale perception object tracking framework based on siamese CNN | |
| JP2018120402A (en) | Event detecting device, event detecting method, and program | |
| CN108154522A (en) | Target tracking system | |
| CN116433722A (en) | Target tracking method, electronic device, storage medium, and program product | |
| Huang et al. | Visual attention learning and antiocclusion-based correlation filter for visual object tracking | |
| CN113658217B (en) | Self-adaptive target tracking method, device and storage medium | |
| CN113706580B (en) | Target tracking method, system, equipment and medium based on relevant filtering tracker | |
| CN112614158B (en) | Sampling frame self-adaptive multi-feature fusion online target tracking method | |
| CN109993777A (en) | A kind of method for tracking target and system based on double-template adaptive threshold |
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 |