CN109544600A - It is a kind of based on it is context-sensitive and differentiate correlation filter method for tracking target - Google Patents
It is a kind of based on it is context-sensitive and differentiate correlation filter method for tracking target Download PDFInfo
- Publication number
- CN109544600A CN109544600A CN201811403138.0A CN201811403138A CN109544600A CN 109544600 A CN109544600 A CN 109544600A CN 201811403138 A CN201811403138 A CN 201811403138A CN 109544600 A CN109544600 A CN 109544600A
- Authority
- CN
- China
- Prior art keywords
- correlation filter
- tracking
- filter
- target
- context
- 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.)
- Withdrawn
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
- G06T7/248—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving reference images or patches
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- 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
- G06T2207/20024—Filtering details
-
- 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
-
- 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
- G06T2207/20084—Artificial neural networks [ANN]
Abstract
The invention discloses a kind of based on context-sensitive and differentiation correlation filter method for tracking target, comprising steps of the end-to-end tracking network of S1, building based on correlation filter, the tracker of pursuit tracking object is used for using the tracking network as benchmark network struction;S2, characteristic pattern is generated using the three first layers convolutional layer of VGG16 model, and based on the characteristic pattern and contextual information training study context correlation filter and scale correlation filter;S3, filter is translated in conjunction with the context-sensitive filter and characteristic pattern training, with the position of the translation filter locating and tracking object;S4, the position based on tracking object calculate the ratio of tracking object using the scale correlation filter, and in conjunction with the position of tracking object in the translation filter, scale correlation filter, characteristic pattern and contextual information positioning next frame;The present invention can effectively improve the accuracy and robustness of target following.
Description
Technical field
The invention belongs to Visual Tracking fields, and in particular to one kind is based on context-sensitive and differentiation correlation filter
Method for tracking target.
Background technique
Target following is a hot issue in computer vision, is had a wide range of applications, such as human-computer interaction, robot skill
Art, autonomous driving, intellectual traffic control etc..Currently, existing method for tracking target can be mainly divided into discriminate method and life
Accepted way of doing sth method.
Production method for tracking target is mainly the appearance features that mobile target is described using model is generated, i.e., only focuses on
Ignore background information to portraying for target itself, core how is studied using including rarefaction representation, particle filter, mean-
Object module including shift etc. searches for candidate target, and finding target to minimize reconstructed error is located at next frame position.Sentence
Other formula method is that the discrimination of target and background is realized by training classifier, also referred to as based on the target following of detection.With generation
The maximum difference of class method is that classifier training has used background information in the process, and such classifier can be absorbed in differentiation prospect
And background.
In recent years, deep learning method is widely used in target following research, that is, utilizes the characteristic pattern of deep learning training
Clarification of objective can be preferably indicated, to improve target following effect.Currently, the method for tracking target master based on deep learning
It is divided into two classes, one kind is the method for tracking target of real-time online study, and this method utilizes auxiliary image data pre-training depth
Model, and pre-training model is finely adjusted using the finite sample information of tracking target in real-time modeling method, to reduce
The demand of target training sample is tracked, such as deep learning tracker, structuring export deep learning tracker;Another kind of is benefit
Feature is extracted with the convolutional neural networks of existing categorized data set pre-training, recycles observation model to carry out classification and obtains satisfaction
Tracking result, such as full convolutional network, tree construction convolutional neural networks.
It was proposed by P.Martins in 2012 based on the method for tracking target of correlation filter, is in the nature that one kind is based on
The core tracking of circular matrix realizes the quick detection of target using Fourier transformation.However, correlation filter tracker
Usually there is very limited contextual information.The interesting target in the initial position estimation next frame using interesting target
Possible position during, estimated result is affected by many factors, including illumination variation, target occlusion, cosmetic variation,
Motion blur, quickly movement, dimensional variation etc. reduce so as to cause tracking precision, while cannot be guaranteed the stability of tracking.
Summary of the invention
Cause tracking precision low object of the present invention is to be changed for the tracking precision of the prior art among the above by external factor
With unstable problem, provide a kind of based on context-sensitive and differentiation correlation filter method for tracking target, this method benefit
With characteristic pattern and contextual information training study context correlation filter and scale correlation filter, specific technical solution is such as
Under:
It is a kind of based on it is context-sensitive and differentiate correlation filter method for tracking target, the method includes the steps:
S1, end-to-end tracking network of the building based on correlation filter, using the tracking network as benchmark network struction
Tracker for pursuit tracking object;
S2, characteristic pattern is generated using the three first layers convolutional layer of VGG16 model, and is believed based on the characteristic pattern and context
Breath training study context correlation filter and scale correlation filter;
S3, filter is translated in conjunction with the context-sensitive filter and characteristic pattern training, with translation filtering
The position of device locating and tracking object;
S4, the position based on tracking object calculate the ratio of tracking object using the scale correlation filter, and
In conjunction with tracking object in the translation filter, scale correlation filter, characteristic pattern and contextual information positioning next frame
Position.
Further, the building of the tracking network includes: to learn correlation filter using the characteristic pattern of training image,
And the similar image in test set is searched for by cross-correlation using identical characteristic pattern.
Further, context-sensitive filter is learnt based on the characteristic pattern and contextual information training described in step S2
Wave device and scale correlation filter include that the contextual information is combined in the correlation filter:
S21, in the image of each frame tracking object, m context block a is taken centered on tracking objecti∈Rn*nMake
For hard negative sample, and obtain the corresponding circular matrix A of the hard negative samplei∈Rn*n;
S22, the context block formation regularizer is added in normalized form, acquires and has to tracking object patch
There are high response and the correlation filter ω ∈ R to the context block without responsen;
S23, according to formulaThe object block generated during tracking is returned
It is grouped into the actual position x of tracking object;Wherein, λ1、λ2For parameter, and λ1It is a weighting parameters, λ2For controlling above and below described
The recurrence of literary block is zero, and the actual position x is the vectorial images of dimensional Gaussian.
Further, the method also includes steps: with formulaDefine correlation filter ω
∈Rn, and the differential of each dependent variable is provided as correlation filter ω ∈ RnLine of the differential of input variable in Fourier
Property function, realize pass through correlation filter ω ∈ RnObtain the Linear Mapping of backpropagation.
Further, the context block is comprising the various interference sources around tracking target and to track different around target
Background.
Further, the correlation filter ω ∈ RnIncluded: by characteristic pattern training
By minimizing formulaKeep tracking object image x × δ-uEach circulation
The inner product of displacement and desired response yuIt is close, wherein δ-uIt is the Dirac delta function of image area, u=0,1 ... .., n-1}2,
yuIt is uthElement;× indicate that cyclic convolution symbol, * indicate circulation cross-correlation symbol.
It further, the use of the ratio that the scale correlation filter calculates tracking object include step described in step S4
It is rapid:
S41, the scale correlation filter is utilized to extract pericentral piece of the feature centered on tracking object
Construct training sample;
S42, building more new formulaAnd with the training sample be it is described more
New formulaInput update the scale correlation filter, wherein using one-dimensional
Gauss as expectation the correlation output g, f of the more new formula be in rectangular domain at each position n by d dimensional feature vector f (l) ∈
RdThe tracking target of compositionBe tracking object f scaling filter time t the molecule μ of fisrt feature channel be study speed
Rate parameter,It is the complex conjugate for it is expected correlation output g in Fourier, F is the discrete Fourier transform of f.
Further, target is set as psam, and set the image I of target t momentt, and assume image ItThe mesh of previous frame
Cursor position pt-1, scale st-1, scale correlation filter At-1And Bt-1, the target position p that is estimated using scale correlation filtertWith
Scale st, updated scale correlation filter is AtAnd Bt, then further include the ruler at time step t in the step S41
Degree filter iteration include:
In target position pt-1With scale st-1From image ItIn target psamCenter extraction sample block;
Use formulaCalculate associated score Bsam, and select ptMaximize BsamAs t
The target position at moment, wherein BtIt is denominator of the t moment new samples f in scale correlation filter, in T with new sample
F;
In target position pt-1With scale st-1From image ItIn target psamCenter extraction sample block;
Use formulaCalculate associated score Bsam;
Selecting scale stMaximized associated score BsamAs the target scale of t moment, and use more new formulaUpdate scale correlation filter AtAnd Bt。
Further, the baseline network is a kind of asymmetrical Siam's network.
It is of the invention based on context-sensitive and differentiate the method for tracking target of correlation filter, firstly, building is to be based on
Tracker on the basis of the tracking network of the end-to-end net of correlation filter;Then, characteristic pattern is generated using VGG16 model, and
It is trained based on characteristic pattern and contextual information, learns or obtain context-sensitive filter and scale correlation filter;Meanwhile
Translation filter is obtained in conjunction with context-sensitive filter and characteristic pattern training, and acquires tracking pair by translating filter
The position of elephant;Finally, the position based on tracking object is in conjunction with translation filter, scale correlation filter and characteristic pattern and up and down
Literary information realization positions the position of tracking target;Compared with prior art, the present invention can effectively be promoted to tracking object with
Track precision is applicable to the target following operation of the more changeable environment of environment, has good robustness.
Detailed description of the invention
Fig. 1 is the stream based on method for tracking target that is context-sensitive and differentiating correlation filter described in the embodiment of the present invention
The signal of journey figure;
Fig. 2 is the frame diagram signal of target following model described in the embodiment of the present invention;
Fig. 3 is the frame diagram signal of the target following end to end based on correlation filtering network described in the embodiment of the present invention;
Fig. 4 is the embodiment of the present invention and mean accuracy of the different trackings on VOT2017 data set in the prior art
Meaning is illustrated compared with robustness;
Fig. 5 is that the embodiment of the present invention is average from expection of the different trackings on VOT2017 data set in the prior art
Overlay chart signal.
Fig. 6 (a)~6 (o) is partial results example of the embodiment of the present invention on VOT2017 data set.
Specific embodiment
In order to enable those skilled in the art to better understand the solution of the present invention, below in conjunction in the embodiment of the present invention
Attached drawing, technical scheme in the embodiment of the invention is clearly and completely described.
Embodiment one
In conjunction with FIG. 1 to FIG. 3, in embodiments of the present invention, the present invention provides one kind based on context-sensitive and differentiation phase
The method for tracking target for closing filter, the method includes the steps:
S1, end-to-end tracking network of the building based on correlation filter, are used for by benchmark network struction of tracking network
The tracker of pursuit tracking object;Specifically, in embodiments of the present invention, learning line by the characteristic pattern using training image
Property template, and the similar image in test set is searched for using identical characteristic pattern by cross-correlation, tracking network is realized with this
Building operation.
Preferably, the baseline network of building is a kind of asymmetrical Siam's network;Also, method of the invention is with formulaDefine correlation filter ω ∈ Rn, obtained by characteristic pattern training;And provide the micro- of each dependent variable
It is allocated as correlation filter ω ∈ RnLinear function of the differential of input variable in Fourier, correlation can be passed through by being realized with this
Filter ω ∈ RnObtain the Linear Mapping of backpropagation.
Wherein, correlation filter ω ∈ R is being obtained by characteristic pattern trainingnWhile need by minimize formulaKeep tracking object image x × δ-uEach cyclic shift inner product and desired response
yuIt is close, wherein δ-uIt is the Dirac delta function of image area, u=0,1 ... .., n-1 }2, yuIt is uthElement;× indicate circulation
Convolution symbol, * indicate circulation cross-correlation symbol;Method of the invention can be according to tracking object image x* δ-uEach cyclic shift
Inner product and desired response yuMutually judge that correlation filter tests the distance of new frame picture locations of real targets recently, this
Sample, so that it may improve tracking accuracy according to the distance, improve tracking accurate rate.
S2, characteristic pattern is generated using the three first layers of VGG16 model, and based on characteristic pattern and contextual information training study
Context-sensitive filter and scale correlation filter;Specifically, middle school's acquistion of the present invention is to context-sensitive filter and ruler
Spend the detailed process of correlation filter are as follows:
Firstly, taking m context block a centered on tracking object in the image of each frame tracking objecti∈Rn*n
As hard negative sample, and obtain the corresponding circular matrix A of hard negative samplei∈Rn*n;Then, context is added in normalized form
Block forms regularizer, acquires the correlation filter to tracking object patch with high response and to context block without response
ω∈Rn;Finally, according to specified formulaObject block is revert to recurrence mesh
Mark x;Wherein, λ1、λ2For parameter, and λ1It is a weighting parameters, λ2Recurrence for controlling the context block is zero, is returned
Target x is the vectorial images of dimensional Gaussian.
In a particular embodiment, the context block in the present invention is comprising tracking various interference sources and tracking around target
Different background around target.
S3, filter is translated in conjunction with the context-sensitive filter and characteristic pattern training, it is fixed to translate filter
The position of position tracking object.
S4, the position based on tracking object calculate the ratio of tracking object using the scale correlation filter, and
In conjunction with tracking object in the translation filter, scale correlation filter, characteristic pattern and contextual information positioning next frame
Position;Wherein, the ratio of tracking object is calculated comprising steps of firstly, being with tracking object using the scale correlation filter
Pericentral piece of latent structure training sample is extracted using scale correlation filter in center;Then, building scale correlation filter
The more new formula of wave deviceAnd pass through more new formula using training sample as inputThe scale correlation filter updated, wherein more new formulaExpectation correlation output be one-dimensional Gauss g, f is in rectangular domain at each position n
By d dimensional feature vector f (l) ∈ RdThe tracking target of compositionBe tracking object f scaling filter in time t in fisrt feature
The molecule μ of channel is learning rate parameter,It is the complex conjugate for it is expected correlation output g in Fourier, F is discrete Fu of f
In leaf transformation.
In embodiments of the present invention, target is set as psam, and set the image I of target t momentt, and assume image ItOn
The target position p of one framet-1, scale st-1, Scale Model At-1And Bt-1;And the target position estimated using scale correlation filter
ptWith scale st, according to more new formulaUpdate Scale Model AtAnd Bt;To realize ruler
Iteration of the filter at time step t is spent, process is specifically included:
Firstly, by target position pt-1With scale st-1From image ItIn target psamCenter extraction sample block;It uses
FormulaCalculate associated score Bsam, and select ptMaximize BsamTarget position as t moment
It sets;Then, in target position pt-1With scale st-1From image ItIn target psamCenter extraction sample block;And use formulaCalculate associated score Bsam;Finally, selecting scale stMaximized associated score BsamAs
The target scale of t moment, and use more new formulaUpdate Scale Model AtAnd Bt。
Embodiment two
Based on described in embodiment one based on it is context-sensitive and differentiate correlation filter method for tracking target, pass through
Experiment is to be specifically described method of the invention;Specifically, the present embodiment specifically uses VOT2017 data set, VOT2017
Video sequence data collection is as popular tracking benchmark, with all ground truth annotations and visual attributes, including camera_
Motion, empty, illum_change, motion_change, occlusion, size_change, mean,
Weightedmean and pooled;Preferably, the present embodiment is averagely overlapped these three assessments using accuracy, robustness and expection
Standard assesses the performance of proposed method, and the details of these three evaluation criterias are described as follows:
Accuracy: when the bounding box that tracker is predicted is zero overlapping with ground truth, a mistake will be detected.
Under given conditions, average overlapping of the accuracy during successfully tracking between prediction and ground truth bounding box is by including phase
Machine movement, illumination variation, motion change, blocks influence with dimensional variation at the free time.Robustness measure statistical tracker is to tracking
The number of target drift (failure), and reset during tracking.
Robustness: in order to reduce each variance for obtaining result, all trackers are run 18 times on each video sequence,
The precision of every frame is extracted in the result that these are obtained as average value, and each sequence is obtained by the precision of averagely every frame
Precision;Then by different operations vision response test calculate the robustness of each sequence.
It is expected that average overlapping: combining every frame accuracy rate and error rate in a manner of criterion, transported in short-term sequence with measurement
The expection of capable tracker is averagely overlapped;Compared with average overlapping is estimated, it is expected that not big variance is estimated in average overlapping, and
There is apparent relationship with the definition tracked in short term.
Next, based on three above-mentioned evaluation criterias, the detailed process of the present embodiment are as follows:
Firstly, the setting specific parameter setting of target following model;4 are set by the quantity of context block, and utilizes life
At characteristic pattern and the context block of sampling train translation filter and scaling filter;By parameter lambda1And λ2It is respectively set to
0.01 and 0.2, learning rate μ is set as 0.025;For correlation filter, the 1/ of translation dimension target size is arranged in standard deviation
16, the initial target size for being dimensioned to twice in space.
Then, by it is of the invention based on contextual information and differentiate correlation filter method for tracking target and it is existing with
Track method compares, and specifically includes: learning one-dimensional differentiation scaling filter based on the tracking for differentiating scale space tracker
To estimate target size, and strength characteristic is combined with the histogram of orientation Gradient Features to learn to translate filter;It is based on
Tracker can generate expert's set, and tracker by storage historical snapshot in the multi-expert tracking of entropy minimization
It can be restored by using entropy minimization criterion;It is proposed based on consensus matching and tracking a kind of based on consensus
Scheme is positioned each tracking object by putting to the vote to current key point series, and is realized in fact using binary system descriptor
When target following;Based on spatial regularization differentiation correlation filter tracker (SRDCF) introduce spatial regularization function come
Punishment resides in the filter coefficient except target area, to mitigate training in effective filter size and detect sample increase
Influence;Target appearance is resolved into submodel by anchor template tracker, describes target with different level of detail, these submodels exist
It interacts in target position fixing process, and is updated for the supervision across submodel;And based on square error and context with
Tracking performance of the minimum output of track in shows combining context even if a very simple correlation filter tracker
Also reasonable performance can be obtained when information.
Refering to Tables 1 and 2, table 1 gives the comparison result based on accuracy evaluation criterion, and table 2 gives based on robustness
The comparison result of assessment level.
Table 1: accuracy of each method on VOT2017 data set compares.
Table 2: robustness of each method on VOT2017 data set compares.
As can be seen from Table 1 and Table 2, method of the invention illustrates optimum performance in different standards;It follows that
Using of the invention based on context-sensitive and differentiation correlation filter method for tracking target, on the one hand, in asymmetric Siam
It is added in neural network and estimates that can preferably handle dimensional variation with the differentiation correlation filter of size estimation asks for translating
Topic;On the other hand, contextual information combination scale peace shifting correlation filter may be implemented to the more exact position of tracking object.
Show that averaged accuracies-robust linearity curve of distinct methods, Fig. 5 are shown in conjunction with Fig. 4 and Fig. 5, Fig. 4 simultaneously
The expection of distinct methods is averagely overlapped, it can be seen that, compared with existing tracking, method proposed by the present invention is realized more
Good detection and tracking performance, this demonstrate that in vision tracking field contextual information and size estimation importance.
In the case of Fig. 6 gives the several of target detection on VOT2017 data set as a result, wherein Fig. 6 (a) is illumination variation
Testing result in fernando video sequence, wherein Fig. 6 (b) is the detection in the case of illumination variation in fish3 video sequence
As a result, wherein Fig. 6 (c) is the testing result in the case of illumination variation in racing video sequence, wherein Fig. 6 (d) is that background is multiple
Testing result in miscellaneous situation in godfather video sequence, wherein Fig. 6 (e) is fish1 video sequence under background complex situations
In testing result, wherein Fig. 6 (f) is the testing result under background complex situations in sheep video sequence, and wherein Fig. 6 (g) is
Testing result under background situation of change in gymnastics2 video sequence, wherein Fig. 6 (h) is under background situation of change
Testing result in helicopter video sequence, wherein Fig. 6 (i) is under background situation of change in iceskater1 video sequence
Testing result, wherein Fig. 6 (j) is the testing result under color similar situation in graduate video sequence, wherein Fig. 6 (k)
It is the testing result under color similar situation in soldier video sequence, wherein Fig. 6 (l) is that wiper is regarded under color similar situation
Testing result in frequency sequence, wherein Fig. 6 (m) is the testing result under circumstance of occlusion in handball1 video sequence, wherein scheming
6 (n) be the testing result under circumstance of occlusion in road video sequence, and wherein Fig. 6 (o) is traffic video sequence under circumstance of occlusion
Testing result in column.
In summary, of the invention based on context-sensitive and differentiation correlation filter method for tracking target, firstly, structure
Build the tracker on the basis of the tracking network of the end-to-end net based on correlation filter;Then, it is generated using VGG16 model special
Sign figure, and be trained based on characteristic pattern and contextual information, learn or obtain context-sensitive filter and scale correlation filtering
Device;Meanwhile translation filter is obtained in conjunction with context-sensitive filter and characteristic pattern training, and obtain by translating filter
To the position of tracking object;Finally, the position based on tracking object combines translation filter, scale correlation filter and feature
Figure and contextual information realize that the position to tracking target positions;Compared with prior art, the present invention can effectively be promoted to tracking
The tracking accuracy of object is applicable to the target following operation of the more changeable environment of environment, has good robustness.
The foregoing is merely a prefered embodiment of the invention, is not intended to limit the scope of the patents of the invention, although referring to aforementioned reality
Applying example, invention is explained in detail, still can be to aforementioned each tool for coming for those skilled in the art
Technical solution documented by body embodiment is modified, or carries out equivalence replacement to part of technical characteristic.All benefits
The equivalent structure made of description of the invention and accompanying drawing content is directly or indirectly used in other related technical areas,
Similarly within the invention patent protection scope.
Claims (9)
1. a kind of based on context-sensitive and differentiation correlation filter method for tracking target, which is characterized in that the method packet
Include step:
S1, end-to-end tracking network of the building based on correlation filter, are used for by benchmark network struction of the tracking network
The tracker of pursuit tracking object;
S2, characteristic pattern is generated using the three first layers convolutional layer of VGG16 model, and is instructed based on the characteristic pattern and contextual information
Practice study context correlation filter and scale correlation filter;
S3, filter is translated in conjunction with the context-sensitive filter and characteristic pattern training, it is fixed with the translation filter
The position of position tracking object;
S4, the position based on tracking object calculate the ratio of tracking object using the scale correlation filter, and combine
The position of tracking object in the translation filter, scale correlation filter, characteristic pattern and contextual information positioning next frame.
2. according to claim 1 based on context-sensitive and differentiation correlation filter method for tracking target, feature
It is, the building of the tracking network includes: to learn correlation filter using the characteristic pattern of training image, and utilize identical
Characteristic pattern searches for the similar image in test set by cross-correlation.
3. according to claim 1 based on context-sensitive and differentiation correlation filter method for tracking target, feature
It is, based on the characteristic pattern and contextual information training study context correlation filter and scale phase described in step S2
Closing filter includes that the contextual information is combined in the correlation filter:
S21, in the image of each frame tracking object, m context block a is taken centered on tracking objecti∈Rn*nAs hard
Negative sample, and obtain the corresponding circular matrix A of the hard negative samplei∈Rn*n;
S22, the context block formation regularizer is added in normalized form, acquiring has height to tracking object patch
Response and the correlation filter ω ∈ R to the context block without responsen;
S23, according to formulaThe object block generated during tracking is revert to
The actual position x of tracking object;Wherein, λ1、λ2For parameter, and λ1It is a weighting parameters, λ2For controlling the context block
Recurrence be zero, the actual position x be dimensional Gaussian vectorial images.
4. according to claim 3 based on context-sensitive and differentiation correlation filter method for tracking target, feature
It is, the method also includes steps: with formulaDefine correlation filter ω ∈ Rn, and provide every
The differential of a dependent variable is as correlation filter ω ∈ RnLinear function of the differential of input variable in Fourier is realized logical
Cross correlation filter ω ∈ RnObtain the Linear Mapping of backpropagation.
5. according to claim 3 based on context-sensitive and differentiation correlation filter method for tracking target, feature
It is, the context block is comprising the various interference sources around tracking target and to track background different around target.
6. according to claim 3 based on context-sensitive and differentiation correlation filter method for tracking target, feature
It is, the correlation filter ω ∈ RnIncluded: by characteristic pattern training
By minimizing formulaKeep tracking object image x × δ-uEach cyclic shift
Inner product and desired response yuIt is close, wherein δ-uIt is the Dirac delta function of image area, u=0,1 ... .., n-1 }2, yuIt is
uthElement;× indicate that cyclic convolution symbol, * indicate circulation cross-correlation symbol.
7. according to claim 4 based on context-sensitive and differentiation correlation filter method for tracking target, feature
Be, described in step S4 using the scale correlation filter calculate tracking object ratio comprising steps of
S41, the scale correlation filter is utilized to extract pericentral piece of the latent structure centered on tracking object
Training sample;
S42, building more new formulaAnd it is public for the update with the training sample
FormulaInput update the scale correlation filter, wherein use one-dimensional Gauss
Expectation correlation output g, f as the more new formula be in rectangular domain at each position n by d dimensional feature vector f (l) ∈ RdGroup
At tracking targetBe tracking object f scaling filter time t the molecule μ of fisrt feature channel be learning rate ginseng
Number,It is the complex conjugate for it is expected correlation output g in Fourier, F is the discrete Fourier transform of f.
8. according to claim 7 based on context-sensitive and differentiation correlation filter method for tracking target, feature
It is, sets target as psam, and set the image I of target t momentt, and assume image ItThe target position p of previous framet-1, ruler
Spend st-1, scale correlation filter At-1And Bt-1, the target position p that is estimated using scale correlation filtertWith scale st, after update
Scale correlation filter be AtAnd Bt, then further include in the step S41 at time step t the scaling filter repeatedly
Generation include:
In target position pt-1With scale st-1From image ItIn target psamCenter extraction sample block;
Use formulaCalculate associated score Bsam, and select ptMaximize BsamAs t moment
Target position, wherein BtIt is denominator of the t moment new samples f in scale correlation filter, in T with new sample F;
In target position pt-1With scale st-1From image ItIn target psamCenter extraction sample block;
Use formulaCalculate associated score Bsam;
Selecting scale stMaximized associated score BsamAs the target scale of t moment, and use more new formulaUpdate scale correlation filter AtAnd Bt。
9. described in any item based on context-sensitive and differentiation correlation filter target following side according to claim 1~8
Method, which is characterized in that the baseline network is a kind of asymmetrical Siam's network.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811403138.0A CN109544600A (en) | 2018-11-23 | 2018-11-23 | It is a kind of based on it is context-sensitive and differentiate correlation filter method for tracking target |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811403138.0A CN109544600A (en) | 2018-11-23 | 2018-11-23 | It is a kind of based on it is context-sensitive and differentiate correlation filter method for tracking target |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109544600A true CN109544600A (en) | 2019-03-29 |
Family
ID=65849265
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811403138.0A Withdrawn CN109544600A (en) | 2018-11-23 | 2018-11-23 | It is a kind of based on it is context-sensitive and differentiate correlation filter method for tracking target |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109544600A (en) |
Cited By (7)
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 |
CN110070562A (en) * | 2019-04-02 | 2019-07-30 | 西北工业大学 | A kind of context-sensitive depth targets tracking |
CN110211157A (en) * | 2019-06-04 | 2019-09-06 | 重庆邮电大学 | A kind of target long time-tracking method based on correlation filtering |
CN110298402A (en) * | 2019-07-01 | 2019-10-01 | 国网内蒙古东部电力有限公司 | A kind of small target deteection performance optimization method |
CN110570454A (en) * | 2019-07-19 | 2019-12-13 | 华瑞新智科技(北京)有限公司 | Method and device for detecting foreign matter invasion |
CN110738685A (en) * | 2019-09-09 | 2020-01-31 | 桂林理工大学 | space-time context tracking method with color histogram response fusion |
CN114140501A (en) * | 2022-01-30 | 2022-03-04 | 南昌工程学院 | Target tracking method and device and readable storage medium |
-
2018
- 2018-11-23 CN CN201811403138.0A patent/CN109544600A/en not_active Withdrawn
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110070562A (en) * | 2019-04-02 | 2019-07-30 | 西北工业大学 | A kind of context-sensitive depth targets tracking |
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 |
CN110211157A (en) * | 2019-06-04 | 2019-09-06 | 重庆邮电大学 | A kind of target long time-tracking method based on correlation filtering |
CN110298402A (en) * | 2019-07-01 | 2019-10-01 | 国网内蒙古东部电力有限公司 | A kind of small target deteection performance optimization method |
CN110570454A (en) * | 2019-07-19 | 2019-12-13 | 华瑞新智科技(北京)有限公司 | Method and device for detecting foreign matter invasion |
CN110570454B (en) * | 2019-07-19 | 2022-03-22 | 华瑞新智科技(北京)有限公司 | Method and device for detecting foreign matter invasion |
CN110738685A (en) * | 2019-09-09 | 2020-01-31 | 桂林理工大学 | space-time context tracking method with color histogram response fusion |
CN114140501A (en) * | 2022-01-30 | 2022-03-04 | 南昌工程学院 | Target tracking method and device and readable storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109544600A (en) | It is a kind of based on it is context-sensitive and differentiate correlation filter method for tracking target | |
Liu et al. | Visual tracking in complex scenes: A location fusion mechanism based on the combination of multiple visual cognition flows | |
Wang et al. | Learning attentional recurrent neural network for visual tracking | |
CN104574445A (en) | Target tracking method and device | |
Gong et al. | Pagerank tracker: From ranking to tracking | |
CN108520530A (en) | Method for tracking target based on long memory network in short-term | |
Wang et al. | A self-attention-based destruction and construction learning fine-grained image classification method for retail product recognition | |
CN110533691A (en) | Method for tracking target, equipment and storage medium based on multi-categorizer | |
Dou et al. | Robust visual tracking based on interactive multiple model particle filter by integrating multiple cues | |
CN109146911A (en) | A kind of method and device of target following | |
Han et al. | Fast learning of spatially regularized and content aware correlation filter for visual tracking | |
Xiao et al. | MeMu: Metric correlation Siamese network and multi-class negative sampling for visual tracking | |
Shuang et al. | Scale-balanced loss for object detection | |
Gong et al. | A two-level framework for place recognition with 3D LiDAR based on spatial relation graph | |
Ma et al. | Study on computer vision target tracking algorithm based on sparse representation | |
Lian et al. | A novel scale insensitive KCF tracker based on HOG and color features | |
Qin et al. | Structure-aware feature disentanglement with knowledge transfer for appearance-changing place recognition | |
Hu et al. | Siamese network object tracking algorithm combining attention mechanism and correlation filter theory | |
Jiang et al. | Automated detection of multitype landforms on mars using a light-weight deep learning-based detector | |
Jiang et al. | A robust end-to-end deep learning framework for detecting Martian landforms with arbitrary orientations | |
Yin et al. | Fast scale estimation method in object tracking | |
Zhang et al. | A framework for long-term tracking based on a global proposal network | |
Zhou et al. | Robust visual tracking using discriminative sparse collaborative map | |
Wang et al. | Deep convolutional correlation filters for forward-backward visual tracking | |
Sun et al. | Correlation filter tracking algorithm based on multiple features and average peak correlation energy |
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 | ||
WW01 | Invention patent application withdrawn after publication | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190329 |