CN106952288B - Based on convolution feature and global search detect it is long when block robust tracking method - Google Patents
Based on convolution feature and global search detect it is long when block robust tracking method Download PDFInfo
- Publication number
- CN106952288B CN106952288B CN201710204379.1A CN201710204379A CN106952288B CN 106952288 B CN106952288 B CN 106952288B CN 201710204379 A CN201710204379 A CN 201710204379A CN 106952288 B CN106952288 B CN 106952288B
- Authority
- CN
- China
- Prior art keywords
- target
- scale
- init
- feature
- tracking
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 25
- 238000001514 detection method Methods 0.000 claims abstract description 25
- 238000001914 filtration Methods 0.000 claims abstract description 19
- 238000004364 calculation method Methods 0.000 claims description 26
- 239000011159 matrix material Substances 0.000 claims description 26
- 230000004044 response Effects 0.000 claims description 11
- 239000000284 extract Substances 0.000 claims description 8
- 230000004048 modification Effects 0.000 claims description 7
- 238000012986 modification Methods 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 230000002146 bilateral effect Effects 0.000 claims description 3
- 230000004043 responsiveness Effects 0.000 claims description 3
- 238000005286 illumination Methods 0.000 abstract description 3
- 230000005923 long-lasting effect Effects 0.000 abstract description 2
- 230000007246 mechanism Effects 0.000 abstract description 2
- 238000013527 convolutional neural network Methods 0.000 description 8
- 230000008569 process Effects 0.000 description 2
- 230000001154 acute effect 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
- 238000012512 characterization method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- 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]
Landscapes
- Image Analysis (AREA)
Abstract
The present invention relates to it is a kind of based on convolution feature and global search detect it is long when block robust tracking method, by using convolution feature and multiple dimensioned correlation filtering method in tracking module, the feature representation ability of tracking target appearance model is enhanced, so that tracking result has very strong robustness for factors such as illumination variation, target scale variation, target rotations;Further through the global search testing mechanism of introducing, so that when target by it is long when block cause tracking failure when, detection module can detect target again, recover tracking module from mistake, accordingly even when can also be tracked long lasting for ground in the case where target appearance variation.
Description
Technical field
The invention belongs to computer vision fields, are related to a kind of method for tracking target, and in particular to one kind is based on convolution feature
With global search detection it is long when block robust tracking method.
Background technique
The main task of target following is to obtain the position of specific objective and motion information in video sequence, is supervised in video
The fields such as control, human-computer interaction have a wide range of applications.During tracking, illumination variation, background are complicated, target is rotated or contracted
The complexity of Target Tracking Problem can all be increased by the factors such as putting, especially when target by it is long when block when, then be easier to cause with
Track failure.
Document " Tracking-learning-detection, IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2012,34 (7): the tracking that 1409-1422 " is proposed is (referred to as
TLD) traditional track algorithm and detection algorithm are combined for the first time, improve tracking result using testing result, improves and is
The reliability and robustness of system.Its track algorithm is based on optical flow method, and detection algorithm generates a large amount of detection window, for each inspection
Survey window, it is necessary to last testing result could be become by three detector receiving.For occlusion issue, TLD provides one
A effective resolving ideas can carry out long time-tracking (Long-term Tracking) to target.But TLD is used
Be shallow-layer manual features, it is limited to the characterization ability of target, and the design of detection algorithm is also complex, there is certain change
Into space.
Summary of the invention
Technical problems to be solved
In order to avoid the shortcomings of the prior art, the present invention proposes a kind of to detect based on convolution feature and global search
Block robust tracking method when long, solve video frequency motion target during tracking due to it is long when block or target translates out the visual field
Except etc. factors cause display model to drift about, thus easily lead to tracking failure the problem of.
Technical solution
It is a kind of based on convolution feature and global search detect it is long when block robust tracking method, it is characterised in that step is such as
Under:
Step 1 reads the first frame image data and the initial position message [x, y, w, h] where target in video, wherein
X, y indicate that the abscissa and ordinate of target's center, w, h indicate the width and height of target.(x, y) corresponding coordinate points are denoted as P,
Centered on P, size is that the target prime area of w × h is denoted as Rinit, then the scale of target is denoted as scale, it is initialized as 1.
Step 2 determines the region R comprising target and background information centered on Pbkg, RbkgSize be M × N, M
=2w, N=2h.Using VGGNet-19 as CNN model, convolution characteristic pattern is extracted to R' in the 5th layer of convolutional layer (conv5-4)
ztarget_init.Then according to ztarget_initConstruct the object module of tracking moduleT ∈ { 1,2 ..., T }, T CNN
Molded passage number, calculation method are as follows:
Wherein, the variable of capitalization is expression of the corresponding non-capitalized variables on frequency domain, gaussian filtering templateM, n be Gaussian function independent variable, m ∈ { 1,2 ..., M }, n ∈ { 1,2 ..., N },
σtargetFor the bandwidth of Gaussian kernel,⊙ representative element multiplication operation, upper scribing line indicate complex conjugate, λ1To adjust
Whole parameter (in order to avoid denominator is 0 and is introduced), is set as 0.0001.
Step 3 extracts the image subblock of S different scale centered on P, and S is set as 33.The size of each sub-block is w
× h × s, variable s are the scale factor of image subblock, s ∈ [0.7,1.4].Then the HOG for extracting each image subblock respectively is special
Sign becomes the HOG feature vector of S dimension, is named as scale feature vector here, is denoted as z after mergingscale_init.Again
According to zscale_initConstruct the Scale Model W of tracking modulescale, calculation method calculates with step 2Similar to (by scale
Feature vector replaces convolution characteristic pattern), specific as follows:
Wherein,S' is Gaussian function independent variable, s' ∈ { 1,2 ..., S }, σscaleFor Gaussian kernel
Bandwidth,λ2For adjusting parameter, it is set as 0.0001.
Step 4 is to target prime area RinitGray feature is extracted, obtained gray feature expression is a two-dimensional matrix,
Here the matrix is named as target appearance representing matrix, is denoted as Ak, the current frame number of subscript k expression, k=1 when initial.Then will
The Filtering Model D of detection module is initialized as A1, i.e. D=A1, the history that reinitializes target expression set of matrices Ahis。AhisWork
With being current and each frame before the target appearance representing matrix of storage, i.e. Ahis={ A1,A2,...,Ak, A when initialhis=
{A1}。
Step 5 reads next frame image, and still centered on P, extraction size is RbkgThe process scaling of × scale
Target search region afterwards.Then the convolution feature in target search region is extracted, and by the CNN network in step 2 with bilateral
The mode of interpolation samples RbkgSize obtain the convolution characteristic pattern z of present frametarget_cur, recycle object moduleMeter
Calculate target confidence map ftarget, calculation method is as follows:
Wherein,For inverse Fourier transform.(x, y) is modified to f by the coordinate of final updating PtargetIn peak response
The corresponding coordinate of value:
Step 6 is extracted the image subblock of S different scale, is then extracted each image subblock respectively centered on P
HOG feature obtains the scale feature vector z of present frame after mergingscale_cur(with z in step 3scale_initCalculation method).Again
Utilize Scale Model WscaleCalculate scale confidence map:
The scale scale of final updating target, calculation method are as follows:
So far, output of the available tracking module in present frame (kth frame): with coordinate be (x, y) P centered on, greatly
Small is RinitThe image subblock TPatch of × scalek.In addition, the f that completion will have been calculatedtargetIn maximum response brief note
For TPeakk, i.e. TPeakk=ftarget(x,y)。
The entire image of Filtering Model D and present frame is carried out convolution, meter in a manner of global search by step 7 detection module
Calculate the similarity degree of Filtering Model D and each position of present frame.The highest preceding j value (j is set as 10) of responsiveness is taken, and respectively
Centered on the corresponding location point of this j value, extraction size is RinitThe j image subblock of × scale.By this j image
Block generates an image subblock set DPatches as elementk, i.e. output of the detection module in kth frame.
The set DPatches that step 8 exports detection modulekIn each image subblock, it is defeated with tracking module to calculate separately it
TPatch outkBetween pixel Duplication, it is available j value, will wherein highest value be denoted asIfLess than threshold value(It is set as 0.05), being determined as that target is blocked completely, the learning rate β for needing to inhibit tracking module in model modification,
And go to step 9;Otherwise initial learning rate β is pressedinit(βinitIt is set as 0.02) being updated, and goes to step 10.The calculation formula of β
It is as follows:
Step 9 is according to DPatcheskIn each image subblock center, respectively extract size be RbkgThe j mesh of × scale
Region of search is marked, to each target search extracted region convolution characteristic pattern and calculates target confidence according to the method in step 5
Scheme, the maximum response on available j target search region.It is compared again in this j response, by maximum value
It is denoted as DPeakk.If DPeakkGreater than TPeakk, then the coordinate of P is updated again, and (x, y) is modified to DPeakkCorresponding
Coordinate.And recalculate target scale feature vector and target scale scale (with the calculation in step 6).
Step 10 target is determined as P in the optimal place-centric of present frame, and optimal scale is determined as scale.In the picture
Indicate new target area Rnew, i.e., centered on P, the rectangle frame of wide and high respectively w × scale, h × scale.In addition,
It will calculate and completed and can obtain the convolution characteristic pattern of optimal objective place-centric P and be abbreviated as ztarget;Equally, will
The scale feature vector for accessing optimal objective scale scale is abbreviated as zscale。
Step 11 utilizes ztarget、zscaleAnd the object module in the tracking module of previous frame foundationAnd scale
Model Wscale, model modification is carried out in a manner of weighted sum respectively, calculation method is as follows:
Wherein, β is the learning rate after step 8 calculates.
Step 12 is to new target area RnewThe target appearance representing matrix A of present frame is obtained after extracting gray featurek,
By AkBeing added to history target indicates set of matrices Ahis.If set AhisMiddle element number is greater thanc(cBe set as 20), then from
AhisMiddle random selectioncOne three-dimensional matrice C of a Element generationk, CkCorresponding (:, i) is AhisIn any one element it is (i.e. two-dimentional
Matrix Ak);Otherwise A is usedhisMiddle all elements generator matrix Ck.Then to CkAverage to obtain two-dimensional matrix, by this two
The matrix Filtering Model D new as detection module is tieed up, calculation method is as follows:
Step 13 judges whether to have handled picture frame all in video, and algorithm terminates if having handled, and otherwise goes to step 5
It continues to execute.
Beneficial effect
It is proposed by the present invention it is a kind of based on convolution feature and global search detect it is long when block robust tracking method, respectively
Tracking module and detection module are devised, during tracking, two module cooperative work: tracking module mainly utilizes convolutional Neural
The convolution feature that network (Convolutional Neural Network, CNN) extracts target is used to construct the target mould of robust
Type, and pass through histograms of oriented gradients (Histogram of Oriented Gradient, HOG) feature construction Scale Model,
Determine the place-centric and scale of target respectively in conjunction with correlation filtering method;Detection module extracts gray feature building target
Filtering Model, is used for quickly detecting in entire image to target by the way of global search and judges the generation blocked, one
Denier target is blocked (or other factors lead to target appearance acute variation) completely, and detection module utilizes testing result amendment tracking
The position of target, and inhibit the model modification of tracking module, prevent from introducing unnecessary noise drift about so as to cause model and with
Track failure.
Superiority: by using convolution feature and multiple dimensioned correlation filtering method in tracking module, tracking mesh is enhanced
The feature representation ability of display model is marked, so that tracking result is for factors such as illumination variation, target scale variation, target rotations
With very strong robustness;Further through the global search testing mechanism of introducing so that when target by it is long when block cause tracking lose
When losing, detection module can detect target again, recover tracking module from mistake, accordingly even when in target appearance
In the case where variation, it can also be tracked long lasting for ground.
Detailed description of the invention
Fig. 1: based on convolution feature and global search detect it is long when block robust tracking method flow chart
Specific embodiment
Now in conjunction with embodiment, attached drawing, the invention will be further described:
Step 1 reads the first frame image data and the initial position message [x, y, w, h] where target in video, wherein
X, y indicate that the abscissa and ordinate of target's center, w, h indicate the width and height of target.(x, y) corresponding coordinate points are denoted as P,
Centered on P, size is that the target prime area of w × h is denoted as Rinit, then the scale of target is denoted as scale, it is initialized as 1.
Step 2 determines the region R comprising target and background information centered on Pbkg, RbkgSize be M × N, M
=2w, N=2h.Using VGGNet-19 as CNN model, convolution characteristic pattern is extracted to R' in the 5th layer of convolutional layer (conv5-4)
ztarget_init.Then according to ztarget_initConstruct the object module of tracking moduleT ∈ { 1,2 ..., T }, T CNN
Molded passage number, calculation method are as follows:
Wherein, the variable of capitalization is expression of the corresponding non-capitalized variables on frequency domain, gaussian filtering templateM, n be Gaussian function independent variable, m ∈ { 1,2 ..., M }, n ∈ { 1,2 ..., N },
σtargetFor the bandwidth of Gaussian kernel,⊙ representative element multiplication operation, upper scribing line indicate complex conjugate, λ1To adjust
Whole parameter (in order to avoid denominator is 0 and is introduced), is set as 0.0001.
Step 3 extracts the image subblock of S different scale centered on P, and S is set as 33.The size of each sub-block is w
× h × s, variable s are the scale factor of image subblock, s ∈ [0.7,1.4].Then the HOG for extracting each image subblock respectively is special
Sign becomes the HOG feature vector of S dimension, is named as scale feature vector here, is denoted as z after mergingscale_init.Again
According to zscale_initConstruct the Scale Model W of tracking modulescale, calculation method calculates with step 2Similar to (by scale
Feature vector replaces convolution characteristic pattern), specific as follows:
Wherein,S' is Gaussian function independent variable, s' ∈ { 1,2 ..., S }, σscaleFor Gaussian kernel
Bandwidth,λ2For adjusting parameter, it is set as 0.0001.
Step 4 is to target prime area RinitGray feature is extracted, obtained gray feature expression is a two-dimensional matrix,
Here the matrix is named as target appearance representing matrix, is denoted as Ak, the current frame number of subscript k expression, k=1 when initial.Then will
The Filtering Model D of detection module is initialized as A1, i.e. D=A1, the history that reinitializes target expression set of matrices Ahis。AhisWork
With being current and each frame before the target appearance representing matrix of storage, i.e. Ahis={ A1,A2,...,Ak, A when initialhis=
{A1}。
Step 5 reads next frame image, and still centered on P, extraction size is RbkgThe process scaling of × scale
Target search region afterwards.Then the convolution feature in target search region is extracted, and by the CNN network in step 2 with bilateral
The mode of interpolation samples RbkgSize obtain the convolution characteristic pattern z of present frametarget_cur, recycle object module
Calculate target confidence map ftarget, calculation method is as follows:
Wherein,For inverse Fourier transform.(x, y) is modified to f by the coordinate of final updating PtargetIn maximum ring
Corresponding coordinate should be worth:
Step 6 is extracted the image subblock of S different scale, is then extracted each image subblock respectively centered on P
HOG feature obtains the scale feature vector z of present frame after mergingscale_cur(with z in step 3scale_initCalculation method).Again
Utilize Scale Model WscaleCalculate scale confidence map:
The scale scale of final updating target, calculation method are as follows:
So far, output of the available tracking module in present frame (kth frame): with coordinate be (x, y) P centered on, greatly
Small is RinitThe image subblock TPatch of × scalek.In addition, the f that completion will have been calculatedtargetIn maximum response brief note
For TPeakk, i.e. TPeakk=ftarget(x,y)。
The entire image of Filtering Model D and present frame is carried out convolution, meter in a manner of global search by step 7 detection module
Calculate the similarity degree of Filtering Model D and each position of present frame.The highest preceding j value (j is set as 10) of responsiveness is taken, and respectively
Centered on the corresponding location point of this j value, extraction size is RinitThe j image subblock of × scale.By this j image
Block generates an image subblock set DPatches as elementk, i.e. output of the detection module in kth frame.
The set DPatches that step 8 exports detection modulekIn each image subblock, it is defeated with tracking module to calculate separately it
TPatch outkBetween pixel Duplication, it is available j value, will wherein highest value be denoted asIfLess than threshold
Value(It is set as 0.05), being determined as that target is blocked completely, needs to inhibit learning rate of the tracking module in model modification
β, and go to step 9;Otherwise initial learning rate β is pressedinit(βinitIt is set as 0.02) being updated, and goes to step 10.The calculating of β is public
Formula is as follows:
Step 9 is according to DPatcheskIn each image subblock center, respectively extract size be RbkgThe j mesh of × scale
Region of search is marked, to each target search extracted region convolution characteristic pattern and calculates target confidence according to the method in step 5
Scheme, the maximum response on available j target search region.It is compared again in this j response, by maximum value
It is denoted as DPeakk.If DPeakkGreater than TPeakk, then the coordinate of P is updated again, and (x, y) is modified to DPeakkCorresponding
Coordinate.And recalculate target scale feature vector and target scale scale (with the calculation in step 6).
Step 10 target is determined as P in the optimal place-centric of present frame, and optimal scale is determined as scale.In the picture
Indicate new target area Rnew, i.e., centered on P, the rectangle frame of wide and high respectively w × scale, h × scale.In addition,
It will calculate and completed and can obtain the convolution characteristic pattern of optimal objective place-centric P and be abbreviated as ztarget;Equally, will
The scale feature vector for accessing optimal objective scale scale is abbreviated as zscale。
Step 11 utilizes ztarget、zscaleAnd the object module in the tracking module of previous frame foundationAnd scale
Model Wscale, model modification is carried out in a manner of weighted sum respectively, calculation method is as follows:
Wherein, β is the learning rate after step 8 calculates.
Step 12 is to new target area RnewThe target appearance representing matrix A of present frame is obtained after extracting gray featurek,
By AkBeing added to history target indicates set of matrices Ahis.If set AhisMiddle element number is greater thanc(cBe set as 20), then from
AhisMiddle random selectioncOne three-dimensional matrice C of a Element generationk, CkCorresponding (:, i) is AhisIn any one element it is (i.e. two-dimentional
Matrix Ak);Otherwise A is usedhisMiddle all elements generator matrix Ck.Then to CkAverage to obtain two-dimensional matrix, by this two
The matrix Filtering Model D new as detection module is tieed up, calculation method is as follows:
Step 13 judges whether to have handled picture frame all in video, and algorithm terminates if having handled, and otherwise goes to step 5
It continues to execute.
Claims (6)
1. it is a kind of based on convolution feature and global search detect it is long when block robust tracking method, it is characterised in that step is such as
Under:
Step 1: reading the first frame image data and the initial position message [x, y, w, h] where target in video, wherein x, y
Indicate that the abscissa and ordinate of target's center, w, h indicate the width and height of target;(x, y) corresponding coordinate points are denoted as P, with P
Centered on, size is that the target prime area of w × h is denoted as Rinit, then the scale of target is denoted as scale, it is initialized as 1;
Step 2: centered on P, determining the region R comprising target and background informationbkg, RbkgSize be M × N, M=
2w, N=2h;Using VGGNet-19 as CNN model, at the 5th layer convolutional layer, that is, conv5-4 layers to RbkgExtract convolution characteristic pattern
ztarget_init;Then according to ztarget_initConstruct the object module of tracking moduleT ∈ { 1,2 ..., T }, T CNN
Molded passage number, calculation method are as follows:
Wherein: the variable of capitalization is expression of the corresponding non-capitalized variables on frequency domain, gaussian filtering templateM, n be Gaussian function independent variable, m ∈ { 1,2 ..., M }, n ∈ { 1,2 ..., N },
σtargetFor the bandwidth of Gaussian kernel,⊙ representative element multiplication operation, upper scribing line indicate complex conjugate, λ1To adjust
Whole parameter;
Step 3: centered on P, extracting the image subblock of S different scale, S is set as 33;The size of each sub-block is w × h
× s, variable s are the scale factor of image subblock, s ∈ [0.7,1.4];Then the HOG feature of each image subblock is extracted respectively,
Become the HOG feature vector of S dimension after merging, and be named as scale feature vector, is denoted as zscale_init;Further according to
zscale_initConstruct the Scale Model W of tracking modulescale, calculation method is as follows:
Wherein,S' is Gaussian function independent variable, s' ∈ { 1,2 ..., S }, σscaleFor the band of Gaussian kernel
Width,λ2For adjusting parameter;
Step 4: to target prime area RinitGray feature is extracted, the two-dimensional matrix of gray feature expression is obtained, is named as target
Appearance representing matrix, is denoted as Ak, the current frame number of subscript k expression, k=1 when initial;Then at the beginning of will test the Filtering Model D of module
Beginning turns to A1, i.e. D=A1, the history that reinitializes target expression set of matrices Ahis;AhisStore current and each frame before target
Appearance representing matrix, i.e. Ahis={ A1,A2,...,Ak, A when initialhis={ A1};
Step 5: reading next frame image, still centered on P, extraction size is Rbkg× scale after scaling
Target search region;Then the convolution feature in target search region is extracted, and by the CNN network in step 2 with bilateral interpolation
Mode sample RbkgSize obtain the convolution characteristic pattern z of present frametarget_cur, recycle object moduleIt calculates
Target confidence map ftarget, calculation method is as follows:
Wherein,For inverse Fourier transform;(x, y) is modified to f by the coordinate of final updating PtargetIn maximum response institute
Corresponding coordinate:
Step 6: centered on P, extracting the image subblock of S different scale, the HOG for then extracting each image subblock respectively is special
Sign, obtains the scale feature vector z of present frame after mergingscale_cur, with z in step 3scale_initCalculation method;Recycle ruler
Spend model WscaleCalculate scale confidence map:
The scale scale of final updating target, calculation method are as follows:
Tracking module is obtained in the output of current kth frame: with coordinate be (x, y) P centered on, size RinitThe figure of × scale
As sub-block TPatchk;In addition, the f that completion will have been calculatedtargetIn maximum response be abbreviated as TPeakk, i.e. TPeakk=
ftarget(x,y);
Step 7: the entire image of Filtering Model D and present frame is carried out convolution in a manner of global search by detection module, is calculated
The similarity degree of Filtering Model D and each position of present frame;The highest preceding j value of responsiveness is taken, and corresponding with j value respectively
Centered on location point, extraction size is RinitThe j image subblock of × scale;Using j image subblock as element, one is generated
A image subblock set DPatchesk, i.e. output of the detection module in kth frame;
Step 8: calculating separately the set DPatches of detection module outputkIn each image subblock and tracking module output
TPatchkBetween pixel Duplication, obtain j value, will wherein highest value be denoted asIfLess than threshold valueDetermine
It is blocked completely for target, the learning rate β for needing to inhibit tracking module in model modification, and goes to step 9;Otherwise it is learned by initial
Habit rate βinitIt is updated, and goes to step 10;
The calculation formula of the β is as follows:
Step 9: according to DPatcheskIn each image subblock center, respectively extract size be RbkgThe j target of × scale is searched
Rope region to each target search extracted region convolution characteristic pattern and calculates target confidence map according to the method in step 5, obtains
Maximum response onto j target search region;Maximum value in j response is denoted as DPeakk;If DPeakkGreatly
In TPeakk, then the coordinate of P is updated again, and (x, y) is modified to DPeakkCorresponding coordinate;And recalculate target scale
Feature vector and target scale scale, using the calculation in step 6;
Step 10: target is determined as P in the optimal place-centric of present frame, and optimal scale is determined as scale;It indicates in the picture
New target area R outnew, centered on P, the rectangle frame of wide and high respectively w × scale, h × scale;In addition, by
It calculates and completes and can obtain the convolution characteristic pattern of optimal objective place-centric P and be abbreviated as ztarget;Equally, it will access
The scale feature vector of optimal objective scale scale is abbreviated as zscale;
Step 11: utilizing ztarget、zscaleAnd the object module in the tracking module of previous frame foundationWith scale mould
Type Wscale, model modification is carried out in a manner of weighted sum respectively, calculation method is as follows:
Wscale=Wscale_new;
Step 12: to new target area RnewThe target appearance representing matrix A of present frame is obtained after extracting gray featurek, by Ak
Being added to history target indicates set of matrices Ahis;If set AhisMiddle element number is greater than c, then from AhisMiddle random selection c
One three-dimensional matrice C of Element generationk, CkCorresponding (:, i) is AhisIn any one element, i.e. two-dimensional matrix Ak;Otherwise A is usedhis
Middle all elements generator matrix Ck;Then to CkIt averages to obtain two-dimensional matrix, using this two-dimensional matrix as detection module
New Filtering Model D, calculation method are as follows:
Step 13: algorithm terminates if having handled picture frame all in video, otherwise goes to step 5 and continues to execute.
2. according to claim 1 based on convolution feature and global search detect it is long when block robust tracking method, it is special
Sign is: the adjusting parameter λ1And λ2It is set as 0.0001.
3. according to claim 1 based on convolution feature and global search detect it is long when block robust tracking method, it is special
Sign is: the j value is set as 10.
4. according to claim 1 based on convolution feature and global search detect it is long when block robust tracking method, it is special
Sign is: the threshold valueIt is set as 0.05.
5. according to claim 1 based on convolution feature and global search detect it is long when block robust tracking method, it is special
Sign is: the initial learning rate βinitIt is set as 0.02.
6. according to claim 1 based on convolution feature and global search detect it is long when block robust tracking method, it is special
Sign is: the c is set as 20.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710204379.1A CN106952288B (en) | 2017-03-31 | 2017-03-31 | Based on convolution feature and global search detect it is long when block robust tracking method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710204379.1A CN106952288B (en) | 2017-03-31 | 2017-03-31 | Based on convolution feature and global search detect it is long when block robust tracking method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106952288A CN106952288A (en) | 2017-07-14 |
CN106952288B true CN106952288B (en) | 2019-09-24 |
Family
ID=59475259
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710204379.1A Expired - Fee Related CN106952288B (en) | 2017-03-31 | 2017-03-31 | Based on convolution feature and global search detect it is long when block robust tracking method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106952288B (en) |
Families Citing this family (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107452022A (en) * | 2017-07-20 | 2017-12-08 | 西安电子科技大学 | A kind of video target tracking method |
CN107644430A (en) * | 2017-07-27 | 2018-01-30 | 孙战里 | Target following based on self-adaptive features fusion |
CN107491742B (en) * | 2017-07-28 | 2020-10-23 | 西安因诺航空科技有限公司 | Long-term stable target tracking method for unmanned aerial vehicle |
CN108734151B (en) * | 2018-06-14 | 2020-04-14 | 厦门大学 | Robust long-range target tracking method based on correlation filtering and depth twin network |
CN110276782B (en) * | 2018-07-09 | 2022-03-11 | 西北工业大学 | Hyperspectral target tracking method combining spatial spectral features and related filtering |
CN109271865B (en) * | 2018-08-17 | 2021-11-09 | 西安电子科技大学 | Moving target tracking method based on scattering transformation multilayer correlation filtering |
CN109308469B (en) * | 2018-09-21 | 2019-12-10 | 北京字节跳动网络技术有限公司 | Method and apparatus for generating information |
CN109410249B (en) * | 2018-11-13 | 2021-09-28 | 深圳龙岗智能视听研究院 | Self-adaptive target tracking method combining depth characteristic and hand-drawn characteristic |
CN109596649A (en) * | 2018-11-29 | 2019-04-09 | 昆明理工大学 | A kind of method and device that host element concentration is influenced based on convolutional network coupling microalloy element |
CN109754424B (en) * | 2018-12-17 | 2022-11-04 | 西北工业大学 | Correlation filtering tracking algorithm based on fusion characteristics and self-adaptive updating strategy |
CN109740448B (en) * | 2018-12-17 | 2022-05-10 | 西北工业大学 | Aerial video target robust tracking method based on relevant filtering and image segmentation |
CN111260687B (en) * | 2020-01-10 | 2022-09-27 | 西北工业大学 | Aerial video target tracking method based on semantic perception network and related filtering |
CN111652910B (en) * | 2020-05-22 | 2023-04-11 | 重庆理工大学 | Target tracking algorithm based on object space relationship |
CN112762841A (en) * | 2020-12-30 | 2021-05-07 | 天津大学 | Bridge dynamic displacement monitoring system and method based on multi-resolution depth features |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105631895A (en) * | 2015-12-18 | 2016-06-01 | 重庆大学 | Temporal-spatial context video target tracking method combining particle filtering |
CN105741316A (en) * | 2016-01-20 | 2016-07-06 | 西北工业大学 | Robust target tracking method based on deep learning and multi-scale correlation filtering |
CN106326924A (en) * | 2016-08-23 | 2017-01-11 | 武汉大学 | Object tracking method and object tracking system based on local classification |
-
2017
- 2017-03-31 CN CN201710204379.1A patent/CN106952288B/en not_active Expired - Fee Related
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105631895A (en) * | 2015-12-18 | 2016-06-01 | 重庆大学 | Temporal-spatial context video target tracking method combining particle filtering |
CN105741316A (en) * | 2016-01-20 | 2016-07-06 | 西北工业大学 | Robust target tracking method based on deep learning and multi-scale correlation filtering |
CN106326924A (en) * | 2016-08-23 | 2017-01-11 | 武汉大学 | Object tracking method and object tracking system based on local classification |
Non-Patent Citations (3)
Title |
---|
CNNTracker: Online discriminative object tracking via deep convolutional neural network;Yan Chen等;《Applied Soft Computing》;20160131;第38卷;第1088-1098页 * |
Hierarchical convolutional features for visual tracking;Chao Ma等;《2015 IEEE International Conference on Computer》;20151213;第3074-3082页 * |
Tracking Human-like Natural Motion Using Deep Recurrent Neural Networks;Youngbin Park等;《arXiv: Computer Vision and Pattern Recognition》;20160415;第1-8页 * |
Also Published As
Publication number | Publication date |
---|---|
CN106952288A (en) | 2017-07-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106952288B (en) | Based on convolution feature and global search detect it is long when block robust tracking method | |
CN109800628B (en) | Network structure for enhancing detection performance of SSD small-target pedestrians and detection method | |
Yuan et al. | Robust visual tracking with correlation filters and metric learning | |
CN104574347B (en) | Satellite in orbit image geometry positioning accuracy evaluation method based on multi- source Remote Sensing Data data | |
CN110119728A (en) | Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network | |
CN111795704A (en) | Method and device for constructing visual point cloud map | |
CN107316316A (en) | The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features | |
CN106570893A (en) | Rapid stable visual tracking method based on correlation filtering | |
CN103886325B (en) | Cyclic matrix video tracking method with partition | |
CN104517095B (en) | A kind of number of people dividing method based on depth image | |
CN112967341B (en) | Indoor visual positioning method, system, equipment and storage medium based on live-action image | |
CN112232134B (en) | Human body posture estimation method based on hourglass network and attention mechanism | |
CN111882546A (en) | Weak supervised learning-based three-branch convolutional network fabric defect detection method | |
Li et al. | Deformable dictionary learning for SAR image change detection | |
CN105469428B (en) | A kind of detection method of small target based on morphologic filtering and SVD | |
CN115527050A (en) | Image feature matching method, computer device and readable storage medium | |
Gao et al. | Occluded person re-identification based on feature fusion and sparse reconstruction | |
Xie et al. | A method of small face detection based on CNN | |
CN110276782A (en) | A kind of EO-1 hyperion method for tracking target of combination sky spectrum signature and correlation filtering | |
CN113096104B (en) | Training method and device of target segmentation model and target segmentation method and device | |
CN116385915A (en) | Water surface floater target detection and tracking method based on space-time information fusion | |
CN115082854A (en) | Pedestrian searching method oriented to security monitoring video | |
CN113112522A (en) | Twin network target tracking method based on deformable convolution and template updating | |
CN112862002A (en) | Training method of multi-scale target detection model, target detection method and device | |
Luo et al. | Real-time pedestrian detection method based on improved YOLOv3 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190924 |