CN110070563A - Correlation filter method for tracking target and system based on joint perception - Google Patents
Correlation filter method for tracking target and system based on joint perception Download PDFInfo
- Publication number
- CN110070563A CN110070563A CN201910362955.4A CN201910362955A CN110070563A CN 110070563 A CN110070563 A CN 110070563A CN 201910362955 A CN201910362955 A CN 201910362955A CN 110070563 A CN110070563 A CN 110070563A
- Authority
- CN
- China
- Prior art keywords
- target
- sample
- parameter
- tracking
- 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.)
- Pending
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
- 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)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present disclosure proposes a kind of correlation filter method for tracking target and system based on joint perception, it first proposed the objective function differentiated under correlation frame, establish context-aware it is adaptive with regressive object between be associated with, realize the target following of robust.The disclosure is extracted multiple context samples, contains contextual information abundant, helps to position target under complex environment.At the adaptive aspect of regressive object, the filter response by converting sample constructs regressive object, compared to the regressive object of general gaussian-shape, can preferably reflect the current distribution of target and movement tendency.Finally, the objective function proposed has closed solutions, closed solutions show that the method proposed during carrying out model learning can carry out context-aware with associated form and regressive object is adaptive.
Description
Technical field
This disclosure relates to a kind of correlation filter method for tracking target based on context with regressive object joint perception
And system.
Background technique
Only there is provided background technical informations relevant to the disclosure for the statement of this part, it is not necessary to so constitute first skill
Art.
The basic project that target following is computer vision and area of pattern recognition is carried out to moving object.The project is related to
Multiple application fields, such as robotics, biological vision, intelligent transportation, intelligent video monitoring etc. in recent years cause extensively
Concern.However, gradually increasing with tracking target data amount, to the gradually promotion that tracking performance requires, and tracking ring
Border tends to be complicated, and many technical problems in target following still need to further study.
Nearest 5 years, the method for tracking target based on differentiation correlation filter became a master of target tracking domain
Study branch.The differentiation correlation filter theory that such method uses is suggested in field of signal processing earliest, it
Researcher is applied to solution Target Tracking Problem afterwards, and achieves significant performance boost.Its core concept be by when
The convolution of image pattern is converted to the form of image pattern product in frequency domain in domain, steady realizing using Fast Fourier Transform (FFT)
Surely while tracking, reach higher processing speed.However, it is found by the inventors that using only in such algorithm comprising a small amount of
The image pattern of context area, being lost largely being capable of the context that is positioned in complex environment to target of aided algorithm
Information.Moreover, for preventing the Cosine Window of boundary effect from further reducing effective context letter used in such algorithm
Breath.
Target tracking algorism based on differentiation correlation filter generallys use the regressive object of general gaussian-shape.So
And with the progress of target following, the problems such as drift due to target deformation, tracking, the regressive object of gaussian-shape not necessarily meets
The distribution of target morphology in training sample or test sample.If using the regressive object of gaussian-shape always, target can not be adapted to
The continuous variation of shape can interfere target appearance model, influence the accuracy of target appearance model.
Summary of the invention
The disclosure to solve the above-mentioned problems, proposes a kind of correlation based on context with regressive object joint perception
Filter method for tracking target and system.The present disclosure proposes the objective functions under a kind of differentiation correlation filter, establish
Context-aware and regressive object it is adaptive between contact, realize with associated form carry out context perception and return mesh
Target is adaptive.Secondly, the disclosure has used target from tradition based on differentiating that the target tracking algorism of correlation filter is different
Multiple context samples in adjacent area, increase the region of context-aware, and have been dissolved into the target proposed
Function realizes the perception of context.Also, adaptive recurrence mesh is constructed by the classifier response of conversion sample
Mark, compared to the regressive object of general gaussian-shape, can preferably reflect the newest distribution situation of target and movement tendency.This
Outside, the regressive object proposed has closed solutions, contains the regressive object of context sample and building in closed solutions, therefore
The filter learnt can be realized combine perception of the context with regressive object, and then carry out the target following of robust.
In some embodiments, the disclosure adopts the following technical scheme that
Correlation filter method for tracking target based on joint perception, comprising:
In the current frame, centered on previous frame target position, test sample is extracted;
According to target appearance model parameter, the filter response of test sample is calculated, and is merged with histogram response,
The Whole Response of model is obtained, the position using the corresponding position of model Whole Response maximum value as target in the current frame;
In the current frame, training sample, context sample and conversion are extracted in the position according to target in the current frame respectively
Sample, and constraint matrix is constructed using conversion sample;
Learning training sample, context sample, and using building constraint matrix, to filter parameter, model parameter into
Row updates;
Histogram parameter is updated.
It is limited as further, further includes: be updated to next frame image, method changes constantly in progress claim 1
Generation, until all image procossings are completed.
It is limited as further, extracts the process of test sample specifically: with previous frame target in current frame image
Centered on position, extract the image pattern of N times of previous frame target scale of scale, N is greater than 1, and by image pattern adjust to
Specified pixel, the test sample as present frame.
It is limited as further, the filter response of the test sample specifically:
Filter response and histogram by test sample are accordingly merged, and the Whole Response of model is obtained, specifically:
F=(1- δ) fc+δ·fh;
Wherein, x0For test sample, ^ indicates discrete Fourier transform, ⊙ representing matrix dot product, and ω is target appearance model
Parameter, f indicate the Whole Response of model, fcIndicate filter response, fhIndicate histogram response, δ indicates fusion parameters.
It is limited as further, the process for extracting training sample includes: in current frame image with current goal
Centered on position, the image pattern of N times of current goal scale of scale is extracted, N is greater than 1, and image pattern is adjusted to finger
Determine pixel, the training sample x as present frame0。
It is limited as further, the process for extracting context sample includes: with B+onCentered on, extraction and training sample
Multiple context sample x of same scale1:k;
Wherein, onThe position for being target in n-th frame, the value of matrix B are B=[- size (x0,1),0;0,-size(x0,
2);size(x0,1),0;0,size(x0,2)]。
It is limited as further, the process constructed to constraint matrix includes:
J conversion sample m is extracted using transition matrix T1:j, the center for converting sample is (p, q)=T+on, scale
It is identical as training sample;
Wherein, onThe position for being target in n-th frame;
The filter response of each conversion sample is calculated, and takes the corresponding response in conversion center of a sample position, as about
The value of beam matrix corresponding position;
Assuming that constraint matrix y0Central element coordinate be located at origin (0,0);
Based on constraint matrix y0In known element value, generate unknown element value by Gauss interpolation, obtain constraint square
Battle array and regressive object.
It is limited as further, filter parameter is updated, specifically:
Wherein, x0For training sample, x1:kFor context sample, y0For constraint matrix, w is filter parameter, θ1、θ2、θ3?
For parameter;
According to filter parameter w, the update to model parameter ω is realized, specifically:
Wherein, λ is the parameter for indicating learning rate.
It is limited as further, includes: to the process that histogram parameter is updated
Histogram parameter h is calculated using the method in Staple algorithm;
According to histogram parameter h, realize to histogram receptance function fhThe update of middle parameter, renewal process are joined with to model
The update method of number ω is identical.
In other embodiments, the disclosure adopts the following technical scheme that
Based on the correlation filter Target Tracking System of joint perception, including server, the server includes storage
Device, processor and storage on a memory and the computer program that can run on a processor, the processor execution journey
Above-mentioned any method is realized when sequence.
Compared with prior art, the beneficial effect of the disclosure is:
(1) context-aware can not be carried out for most target tracking algorisms based on differentiation correlation filter and return
The problem of returning objective self-adapting, propose differentiate correlation filter frame under objective function, establish context-aware with
Association between regressive object is adaptive increases robustness of the algorithm under complicated tracking environmental;
(2) in terms of context-aware, differentiate that correlation algorithm filter has only used small-scale upper and lower region with tradition
Domain is different, and the disclosure is extracted multiple context samples, can perceive to large-scale target context region;
(3) at the adaptive aspect of regressive object, constraint matrix can be constructed according to conversion sample, and then is able to reflect
The regressive object of current goal distribution and movement tendency;
(4) objective function proposed has closed solutions, and closed solutions show to be proposed during carrying out model learning
Tracking can with associated form carry out context-aware and regressive object it is adaptive.
Detailed description of the invention
The accompanying drawings constituting a part of this application is used to provide further understanding of the present application, and the application's shows
Meaning property embodiment and its explanation are not constituted an undue limitation on the present application for explaining the application.
Fig. 1 is the correlation filter method for tracking target in embodiment one based on context with regressive object joint perception
Chief component schematic diagram;
Fig. 2 is effect schematic diagram of contextual information during target positioning and model modification in embodiment one;
Fig. 3 is the building process schematic diagram of constraint matrix in embodiment one;
Fig. 4 is constraint matrix and common Gaussian regressive object comparison diagram in embodiment one;
Fig. 5 (a) and Fig. 5 (b) is the experimental result in embodiment one on OTB50 data set respectively;
Fig. 6 (a) and Fig. 6 (b) is the experimental result in embodiment one on OTB2013 data set respectively;
Fig. 7 (a) and Fig. 7 (b) is the experimental result in embodiment one on OTB2015 data set respectively;
Fig. 8 is the experimental result for carrying out Real-time in embodiment one on VOT2017 data set and testing.
Specific embodiment
It is noted that described further below be all exemplary, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms that the disclosure uses have logical with the application person of an ordinary skill in the technical field
The identical meanings understood.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular
Also it is intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or their combination.
Embodiment one
A kind of correlation based on context with regressive object joint perception is disclosed in one or more embodiments
Filter method for tracking target, as shown in Figure 1, can be during carrying out target positioning and model learning, with associated form
It is adaptive with regressive object to carry out context-aware, enhances the accuracy and robustness of target following.
Firstly, extract a certain frame image in video sequence as the first frame image tracked, video sequence can be with
For the video shot in the scenes such as video monitoring, intelligent transportation.Later, the target object tracked in image is carried out
Calibration, and target appearance model parameter is initialized, tracking target can be any object in image.Later, enter
Next frame carries out target following using the method that the disclosure proposes, specific step is as follows for tracking:
(1) in the current frame, centered on previous frame target position, test sample is extracted;
(2) according to target appearance model parameter, the filter response of test sample is calculated, and is melted with histogram response
It closes, obtains the Whole Response of model, the position using the corresponding position of model Whole Response maximum value as target in the current frame;
(3) in the current frame, training sample is extracted in the position according to target in the current frame;
(4) in the current frame, multiple context samples are extracted in the position according to target in the current frame;
(5) in the current frame, multiple conversion samples are extracted, and use conversion sample in the position according to target in the current frame
This constructs constraint matrix;
(6) learning training sample, context sample, and using the constraint matrix of building, to filter parameter, model parameter
It is updated;
(7) histogram parameter is updated.
Above-mentioned steps are described in detail below.
In the first frame tracked, with the target position o of calibration1Centered on, the target scale of N times of calibration is range
(N > 1) acquires target sample x0.Using target sample x0Calculate obtain filter parameter w to target appearance model parameter ω into
Row initialization, formula is as follows, and next frame is entered after the completion of initialization, is tracked to target,
Wherein, ^ indicates discrete Fourier transform, x0Conjugate complex number beygFor scale height identical with target sample
This type regressive object, θ1It is the parameter for preventing over-fitting.
In n-th frame (n > 1), in order to realize the positioning to target, with previous frame target position on-1Centered on, on N times
One frame target scale is range (N > 1), collecting test sample x0。
By current goal display model parameter ω, the filter that can get test sample responds fc, formula is as follows:
Wherein, ⊙ representing matrix dot product.
F is responded according to the histogram that histogram method obtains test sampleh, and it is merged with filter response,
The final response diagram f of test sample is obtained, formula is as follows:
F=(1- δ) fc+δ·fh
Wherein, δ is the parameter for indicating integration percentage.
Take position o of the corresponding position of maximum value as target in n-th frame in final response diagramn。
In the study stage, with current goal position onCentered on, N times of current goal scale is range (N > 1), extracts instruction
Practice sample x0。
In n-th frame, with B+onCentered on extract context sample x1:k, context sample is identical as training sample scale.
Wherein, matrix B indicates offset of the context sample center compared to target position, B=[- size (x0,1),0;0,-
size(x0,2);size(x0,1),0;0,size(x0,2)]。
Using transition matrix T, conversion sample is extracted, and constructs constraint matrix y0.The construction method of constraint matrix will pass through
Fig. 3 is introduced in detail.
The present disclosure proposes context-aware and the adaptive objective function of regressive object, public affairs can be carried out with associated form
Formula is as follows:
Wherein, X0For training sample, XiFor context sample, y is to pass through constraint matrix y0The adaptive recurrence mesh of building
Mark, θ1、θ2And θ3It is the parameter for preventing over-fitting.In the objective function, X0And XiIndicate circulation sample, x0And xiRespectively
X0And XiBasic sample.
For the ease of solving the optimization problem in objective function, it is assumed that
Then objective function can be rewritten as following form:
In above formula, z=[wy '] is enabledT, can following form further be converted by the optimization problem in above formula:
Objective function after conversion is convex function, therefore it is zero that the closed solutions of parameter z, which can enable the first derivative of objective function,
It obtains:
Known circular matrix A has the property that
Circular matrix property is substituted into the deformation of objective function first derivative, be can be obtained:
Above formula is simplified, can be obtained:
Wherein,N=diag (1+ θ2)。
According to above formula,Value can easily obtain:
Matrix for Inverse Problem in above formula can be by way of matrix inversion formula be expressed as:
Therefore, the disclosure has further obtained the solution of filter parameter w according to Inversion Formula:
Wherein, (V-DN-1C)-1Deployable expression are as follows:
Later, the solution of filter parameter w is further simplified, it is as follows that the closed solutions of filter parameter, which can be obtained,
Form:
According to known training sample x0, context sample x1:k, constraint matrix y0, filtering can be calculated using above formula
Device parameter w.By above formula as can be seen that containing contextual information and constructed adaptive in the filter parameter learnt
Regressive object can be improved the accuracy and robustness of target following.
Finally, realizing the update to model parameter ω according to filter parameter, formula is as follows, while carrying out to histogram
It updates,
Wherein, λ is the parameter for indicating learning rate.
After the completion of update, into next frame, the target in next frame is tracked using above step, until video sequence
The last frame of column.
Contextual information can provide important clue under complicated tracking environmental for target following, can assist pair
The positioning of target significantly improves the performance of method for tracking target.Fig. 2 be in the disclosure contextual information target positioning and mould
Effect schematic diagram in type renewal process.
During model modification, only use that small-scale target context region is different, and the disclosure makes from conventional method
Filter parameter and model parameter are updated with the context sample comprising a large amount of context areas.The filter learnt
Wave device parameter contains the contextual information around target, is able to reflect the background of target adjacent domain.
When carrying out target positioning using the filter parameter learnt, the contextual information around target can be in target
Generation is blocked, on a large scale in the complex environments such as cosmetic variation, and important auxiliary information is provided.Track algorithm can utilize target
The information such as spatial position realize more accurate target positioning.
The present disclosure proposes the construction methods of adaptive regressive object.Traditional target based on differentiation correlation filter
The regressive object Gaussian distributed that tracking uses, however, with the continuous variation of target appearance during tracking, target
Appearance is difficult to obey the distribution of gaussian shape.Therefore, if all using static gaussian-shape regressive object during entire tracking,
Accurately target can not accurately be marked in tracking process or training process, and then tracking drift is gradually caused to be asked
Topic.To solve the above-mentioned problems, the adaptive regressive object that the disclosure proposes can construct energy according to the response of conversion sample
Enough reflect that the regressive object that target is currently distributed, building process are as shown in Figure 3.In the disclosure, regressive object obeys noise
Model, y=y+n,Wherein, y0Building for constraint matrix, for regressive object.
In Fig. 3, transition matrix T=[t is used1,t2,...,tj]=[0,0;0,1;1,1;0,-1;-1,0;- 1, -1] it mentions
Take j conversion sample m1:j.In the disclosure, when using HOG feature, it is considered as the cell value of HOG feature;In multiple dimensioned mesh
It marks in tracking, is also considered as influence of the target scale variation to transition matrix.Convert sample center be (p, q)=
T+on, wherein onThe position for being target in n-th frame.The scale for converting sample is identical as training sample.
Later, the filter response of each conversion sample is calculated, and takes the corresponding response in conversion center of a sample position, is made
For the value of constraint matrix corresponding position, formula is as follows:
y0(t1:j)=fc(m1:j)center
Where it is assumed that constraint matrix y0Central element coordinate be located at origin (0,0).
In the disclosure, it is based on constraint matrix y0In known element value, unknown member can be generated by Gauss interpolation
Element value, and then can get constraint matrix and regressive object.Fig. 4 show constraint matrix and common Gaussian regressive object comparison diagram,
Compared to gaussian-shape regressive object used in conventional method, constructed constraint matrix can better adapt to work as in the disclosure
The distribution situation of preceding target and the motion information for reflecting target.
To being proposed based on context and regressive object on OTB50, OTB2013, OTB2015, VOT2017 data set
The correlation filter method for tracking target of joint perception is assessed.The experiment of the disclosure is being furnished with tetra- core processor of i7,
It is carried out on the PC of 2.59GHz CPU, 8GB RAM.
The disclosure performs two versions of proposed method for tracking target: using the version of HOG and color names feature
This Ours and version Ours_deep for using depth characteristic.Ours, can be with 12 frames in the MATLAB environment that GPU is not used
Processing speed per second tracks target;Ours_deep using it is tall and handsome reach GTX 1070GPU when, in PyTorch environment
In can carry out target following with 43 frames speed per second.
In an experiment, the value of all parameters is all fixed value.Regularization parameter θ1,θ2,θ3Value be θ1=1e-3,θ2=1,
θ3=0.5;Context sample number is k=4;Number of samples is converted as j=4;Fusion parameters are δ=1/11 in Ours,
Fusion parameters are δ=0 in Ours_deep;Learning parameter is λ=0.02 in Ours, and learning parameter is λ in Ours_deep
=0.012.
Ours_deep has used the convolution feature that second layer convolutional layer exports in VGG-16, and eliminates whole
Maxpooling layers.The disclosure has carried out network by stochastic gradient descent method using 2015 data set of ILSVRC to use 50
The training of secondary complete data set, batch value are 32.In the disclosure, it is 1.5 extraction image patterns to fill size, and is adjusted
Whole is 128*128 pixel, and each pair of frame is acquired in 10 neighbouring frames.In stochastic gradient descent method, momentum value is 0.9,
Weight decaying is set as 0.005, learning rate 1e-2.The loss function of network training, which uses, returns loss function.
On OTB data set, the disclosure by the Ours of proposition and Ours_deep and 10 kinds based on deep learning it is advanced with
Track algorithm (MCPF, TRACA, SiamFC, CFNet, ACFN, DeepSRDCF, HCF, HDT, CNN-SVM, DLS-SVM) and 8 kinds it is non-
Based on deep learning advanced track algorithm (i.e. CSRDCF, Staple_CA, SAMF_AT, SRDCF, Staple, MUSTer,
LCT, SAMF) carry out comparative analysis.Wherein, TRACA, CSRDCF, Staple_CA and SAMF_AT are context-aware tracking
Algorithm.MCPF,TRACA,CFNet,ACFN,DeepSRDCF,HCF,HDT,CSRDCF,Staple_CA,SAMF_AT,SRDCF,
Staple, MUSTer, LCT, SAMF are to differentiate correlation filter tracks algorithm.The baseline track algorithm of Ours is Staple.
The baseline track algorithm of Ours_deep is KCF.MCPF,TRACA,ACFN,DeepSRDCF,DLS-SVM,HDT,CNN-SVM,
The data that the tracking result of SiamFC, CFNet, CSRDCF, SRDCF, HCF, Staple, LCT have used its author to provide,
The data that the tracking result of MUSTer has used the author of LCT to provide.The tracking knot of Staple_CA, SAMF_AT, SAMF algorithm
Fruit is that the Open Source Code that its author provides generates.The speed of service target tracking algorism per second greater than 20 frames is seen and is put into effect by the disclosure
When track algorithm.
The disclosure mainly in terms of the tracking success rate of OTB data set, is analyzed the tracking result for experiment,
Tracking precision when area under the curve (area under the curve, AUC) and threshold value are 20 is respectively adopted to algorithm
Tracking success rate and tracking precision are measured.
It is accurate that Fig. 5 (a)-(b) show on OTB50 data set the tracking success rate of 14 algorithms and tracking before ranking
Degree.In tracking success rate figure, the Ours_deep algorithm that the disclosure proposes comes second, lower than the side MCPF for being located at first place
Method 0.69%, but the speed of service of the speed of service ratio MCPF of Ours_deep algorithm is 86 times high.DeepSRDCF method comes
Three, success rate ratio Ours_deep low 3.39%, slow 43 times of the speed of service.The success rate of Ours algorithm is located at the 4th, than
Ours_deep method low 5.08%, but it is better than SRDCF algorithm 2.23%.In terms of real-time, in preceding 14 algorithms only
Ours_deep, TRACA, Staple_CA, CFNet and SiamFC are able to carry out real-time target following, and the disclosure proposes
The tracking success rate of Ours_deep rank the first.The success rate of context-aware track algorithm for comparative experiments is by height
Ours_deep, Ours, TRACA, Staple_CA, CSRDCF, SAMF_AT are followed successively by low ranking.
Fig. 6 (a)-(b) show on OTB2013 data set the tracking success rate of 14 algorithms and tracking essence before ranking
Exactness.In tracking success rate figure, the tracking success rate for the Ours_deep method that the disclosure proposes comes second, than being located at
The first MCPF method low 2.42%, but Ours_deep has apparent advantage in terms of the speed of service.The tracking of Ours algorithm
Success rate ratio Ours_deep method low 5.25%, this is because having used depth characteristic in Ours_deep algorithm, improves calculation
The performance of method.Although the tracking success rate of Ours method is located at the 7th, it is better than the tracking of baseline track algorithm Staple
Success rate, success rate of the Staple method on OTB2013 data set are located at sixteen bit.In terms of real-time, first 14
Only Ours_deep, TRACA, LCT, Staple_CA, CFNet and SiamFC algorithm have reached real-time processing speed in tracking
Degree, and the tracking success rate of Ours_deep is located at first place.The success rate of context-aware track algorithm for comparative experiments by
High to Low ranking is followed successively by Ours_deep, TRACA, Ours, SAMF_AT, Staple_CA, CSRDCF.
Fig. 7 (a)-(b) show on OTB2015 data set the tracking success rate of 14 algorithms and tracking essence before ranking
Exactness.In tracking success rate figure, the tracking success rate for the Ours_deep algorithm that the disclosure proposes comes third position, than
DeepSRDCF method low 3.92%, lower than MCPF method 2.61%.But DeepSRDCF and MCPF are not that real-time tracking is calculated
Method, Ours_deep have a clear superiority in the speed of service.The success rate of Ours algorithm comes the 5th, is higher than SRDCF method
0.84%, it is higher than baseline algorithm Staple algorithm 3.79%.In terms of real-time, only Ours_ in preceding 14 trackings
Deep, TRACA, Staple_CA, SiamFC, Staple and CFNet meet requirement of real-time, and the tracking of Ours_deep at
Power comes first in these four algorithms.The success rate of context-aware track algorithm for comparative experiments is from high to low
Ranking is followed successively by Ours_deep, TRACA, Ours, Staple_CA, CSRDCF, SAMF_AT.
Fig. 8 is that Ours_deep carries out the result schematic diagram that Real-time is tested on VOT2017 data set.As shown in the figure, grey
Color horizontal line indicates that 10 kinds of real-time expections for being listed in " advanced " target tracking algorism are averagely overlapped (expected in VOT2017
Average overlap, EAO) mean value (0.106), and the disclosure proposes that the real-time EAO of algorithm Ours_deep is higher by compared with it
53.77%.In VOT2017 in all context-aware algorithms, other than the CSRDCF++ that real-time EAO makes number one,
The real-time EAO of Ours_deep algorithm is higher than other target tracking algorisms for carrying out context-aware, i.e. CSRDCFf, MOSSE_
CA, CSRDCF, SPCT and FSTC.In addition, the real-time EAO of Ours_deep is better than baseline algorithm KCF method 22.56%.
Embodiment two
A kind of correlation based on context with regressive object joint perception disclosed in one or more embodiments
Filter Target Tracking System, including server, the server include memory, processor and storage on a memory and can
The computer program run on a processor, the processor are realized described in embodiment one when executing described program based on upper
The correlation filter method for tracking target hereafter perceived with regressive object joint.
Embodiment three
A kind of computer readable storage medium disclosed in one or more embodiments, is stored thereon with computer journey
Sequence executes the correlation described in embodiment one based on context with regressive object joint perception when the program is executed by processor
Filter method for tracking target.
Although above-mentioned be described in conjunction with specific embodiment of the attached drawing to the disclosure, model not is protected to the disclosure
The limitation enclosed, those skilled in the art should understand that, on the basis of the technical solution of the disclosure, those skilled in the art are not
Need to make the creative labor the various modifications or changes that can be made still within the protection scope of the disclosure.
Claims (10)
1. the correlation filter method for tracking target based on joint perception characterized by comprising
In the current frame, centered on previous frame target position, test sample is extracted;
According to target appearance model parameter, the filter response of test sample is calculated, and is merged with histogram response, is obtained
The Whole Response of model, the position using the corresponding position of model Whole Response maximum value as target in the current frame;
In the current frame, training sample, context sample and conversion sample are extracted in the position according to target in the current frame respectively
This, and constraint matrix is constructed using conversion sample;
Learning training sample, context sample, and using the constraint matrix of building, filter parameter, model parameter are carried out more
Newly;
Histogram parameter is updated.
2. the correlation filter method for tracking target as described in claim 1 based on joint perception, which is characterized in that also wrap
It includes: being updated to next frame image, constantly carry out the iteration of method in claim 1, until all image procossings are completed.
3. the correlation filter method for tracking target as described in claim 1 based on joint perception, which is characterized in that extract
The process of test sample specifically: in current frame image centered on previous frame target position, extract N times of scale upper one
The image pattern of frame target scale, N is greater than 1, and image pattern is adjusted to specified pixel, the test sample as present frame.
4. the correlation filter method for tracking target as described in claim 1 based on joint perception, which is characterized in that described
The filter of test sample responds specifically:
Filter response and histogram by test sample are accordingly merged, and the Whole Response of model is obtained, specifically: f=
(1-δ)·fc+δ·fh;
Wherein, x0For test sample, ^ indicates discrete Fourier transform, ⊙ representing matrix dot product, and ω is target appearance model parameter,
F indicates the Whole Response of model, fcIndicate filter response, fhIndicate histogram response, δ indicates fusion parameters.
5. the correlation filter method for tracking target as described in claim 1 based on joint perception, which is characterized in that described
Extract training sample process include: in current frame image centered on current goal position, it is current to extract N times of scale
The image pattern of target scale, N is greater than 1, and image pattern is adjusted to specified pixel, the training sample x as present frame0。
6. the correlation filter method for tracking target as described in claim 1 based on joint perception, which is characterized in that extract
The process of context sample includes: with B+onCentered on, extract multiple context sample x with training sample same scale1:k;
Wherein, onThe position for being target in n-th frame, the value of matrix B are B=[- size (x0,1),0;0,-size(x0,2);
size(x0,1),0;0,size(x0,2)]。
7. the correlation filter method for tracking target as described in claim 1 based on joint perception, which is characterized in that about
The process that beam matrix is constructed includes:
J conversion sample m is extracted using transition matrix T1:j, the center for converting sample is (p, q)=T+on, scale and training
Sample is identical;
Wherein, onThe position for being target in n-th frame;
The filter response of each conversion sample is calculated, and takes the corresponding response in conversion center of a sample position, as constraint square
The value of battle array corresponding position;
Assuming that constraint matrix y0Central element coordinate be located at origin (0,0);
Based on constraint matrix y0In known element value, generate unknown element value by Gauss interpolation, obtain constraint matrix and return
Return target.
8. the correlation filter method for tracking target as described in claim 1 based on joint perception, which is characterized in that filter
Wave device parameter is updated, specifically:
Wherein, x0For training sample, x1:kFor context sample, y0For constraint matrix, w is filter parameter, θ1、θ2、θ3It is ginseng
Number;
According to filter parameter w, the update to model parameter ω is realized, specifically:
Wherein, λ is the parameter for indicating learning rate.
9. the correlation filter method for tracking target as described in claim 1 based on joint perception, which is characterized in that straight
The process that square graph parameter is updated includes:
Histogram parameter h is calculated using the method in Staple algorithm;
According to histogram parameter h, realize to histogram receptance function fhThe update of middle parameter, renewal process with to model parameter ω
Update method it is identical.
10. the correlation filter Target Tracking System based on joint perception, which is characterized in that including server, the service
Device include memory, processor and storage on a memory and the computer program that can run on a processor, the processor
The described in any item methods of claim 1-9 are realized when executing described program.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910362955.4A CN110070563A (en) | 2019-04-30 | 2019-04-30 | Correlation filter method for tracking target and system based on joint perception |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910362955.4A CN110070563A (en) | 2019-04-30 | 2019-04-30 | Correlation filter method for tracking target and system based on joint perception |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110070563A true CN110070563A (en) | 2019-07-30 |
Family
ID=67370101
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910362955.4A Pending CN110070563A (en) | 2019-04-30 | 2019-04-30 | Correlation filter method for tracking target and system based on joint perception |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110070563A (en) |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015163830A1 (en) * | 2014-04-22 | 2015-10-29 | Aselsan Elektronik Sanayi Ve Ticaret Anonim Sirketi | Target localization and size estimation via multiple model learning in visual tracking |
CN107346548A (en) * | 2017-07-06 | 2017-11-14 | 电子科技大学 | A kind of tracking for electric transmission line isolator |
KR101837407B1 (en) * | 2017-11-03 | 2018-03-12 | 국방과학연구소 | Apparatus and method for image-based target tracking |
US10009062B1 (en) * | 2014-10-21 | 2018-06-26 | Marvell International Ltd. | Systems and methods for a twisted pair transceiver with correlation detection |
CN108447078A (en) * | 2018-02-28 | 2018-08-24 | 长沙师范学院 | The interference of view-based access control model conspicuousness perceives track algorithm |
CN108549839A (en) * | 2018-03-13 | 2018-09-18 | 华侨大学 | The multiple dimensioned correlation filtering visual tracking method of self-adaptive features fusion |
CN108734723A (en) * | 2018-05-11 | 2018-11-02 | 江南大学 | A kind of correlation filtering method for tracking target based on adaptive weighting combination learning |
CN109035290A (en) * | 2018-07-16 | 2018-12-18 | 南京信息工程大学 | A kind of track algorithm updating accretion learning based on high confidence level |
CN109461172A (en) * | 2018-10-25 | 2019-03-12 | 南京理工大学 | Manually with the united correlation filtering video adaptive tracking method of depth characteristic |
CN109685831A (en) * | 2018-12-20 | 2019-04-26 | 山东大学 | Method for tracking target and system based on residual error layering attention and correlation filter |
-
2019
- 2019-04-30 CN CN201910362955.4A patent/CN110070563A/en active Pending
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015163830A1 (en) * | 2014-04-22 | 2015-10-29 | Aselsan Elektronik Sanayi Ve Ticaret Anonim Sirketi | Target localization and size estimation via multiple model learning in visual tracking |
US10009062B1 (en) * | 2014-10-21 | 2018-06-26 | Marvell International Ltd. | Systems and methods for a twisted pair transceiver with correlation detection |
CN107346548A (en) * | 2017-07-06 | 2017-11-14 | 电子科技大学 | A kind of tracking for electric transmission line isolator |
KR101837407B1 (en) * | 2017-11-03 | 2018-03-12 | 국방과학연구소 | Apparatus and method for image-based target tracking |
CN108447078A (en) * | 2018-02-28 | 2018-08-24 | 长沙师范学院 | The interference of view-based access control model conspicuousness perceives track algorithm |
CN108549839A (en) * | 2018-03-13 | 2018-09-18 | 华侨大学 | The multiple dimensioned correlation filtering visual tracking method of self-adaptive features fusion |
CN108734723A (en) * | 2018-05-11 | 2018-11-02 | 江南大学 | A kind of correlation filtering method for tracking target based on adaptive weighting combination learning |
CN109035290A (en) * | 2018-07-16 | 2018-12-18 | 南京信息工程大学 | A kind of track algorithm updating accretion learning based on high confidence level |
CN109461172A (en) * | 2018-10-25 | 2019-03-12 | 南京理工大学 | Manually with the united correlation filtering video adaptive tracking method of depth characteristic |
CN109685831A (en) * | 2018-12-20 | 2019-04-26 | 山东大学 | Method for tracking target and system based on residual error layering attention and correlation filter |
Non-Patent Citations (8)
Title |
---|
LUCA BERTINETTO: "Staple:Complementary Learners for Real-Time Tracking", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
MATTHIAS MUELLER: "Context-Aware Correlation Filter Tracking", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
MING TANG: "High-Speed Tracking with Multi-kernel Correlation Filters", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
QIAN ZOU: "High Confidence Staple for Visual Tracking", 《2017 INTERNATIONAL CONFERENCE ON COMPUTER SYSTEMS, ELECTRONICS AND CONTROL (ICCSEC)》 * |
XIAOGUANG NIU: "Robust Visual Tracking via Occlusion Detection Based on Staple algorithm", 《2017 11TH ASIAN CONTROL CONFERENCE (ASCC)》 * |
XIAOHUI YANG: "A Joint Multi-Feature and Scale-Adaptive Correlation Filter Tracker", 《IEEE ACCESS》 * |
YANG LI: "A Scale Adaptive Kernel Correlation Filter Tracker with Feature Integration", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 * |
李红梅: "基于相关滤波的尺度自适应目标跟踪方法的研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110111366A (en) | A kind of end-to-end light stream estimation method based on multistage loss amount | |
CN112233147B (en) | Video moving target tracking method and device based on two-way twin network | |
CN109299701A (en) | Expand the face age estimation method that more ethnic group features cooperate with selection based on GAN | |
CN102592112B (en) | Method for determining gesture moving direction based on hidden Markov model | |
CN106910176A (en) | A kind of facial image based on deep learning removes occlusion method | |
CN111695457B (en) | Human body posture estimation method based on weak supervision mechanism | |
CN108921879A (en) | The motion target tracking method and system of CNN and Kalman filter based on regional choice | |
CN109859241B (en) | Adaptive feature selection and time consistency robust correlation filtering visual tracking method | |
CN108416266A (en) | A kind of video behavior method for quickly identifying extracting moving target using light stream | |
CN111160294B (en) | Gait recognition method based on graph convolution network | |
CN102663416A (en) | Segmentation method of viscera and internal blood vessels thereof in surgical planning system | |
CN109508787A (en) | Neural network model training method and system for ultrasound displacement estimation | |
CN109377555A (en) | Autonomous underwater robot prospect visual field three-dimensional reconstruction target's feature-extraction recognition methods | |
CN102592135B (en) | Visual tracking method of subspace fusing target space distribution and time sequence distribution characteristics | |
CN104899896B (en) | Multi-task learning target tracking method based on subspace characteristics | |
CN115375737B (en) | Target tracking method and system based on adaptive time and serialized space-time characteristics | |
CN109508686A (en) | A kind of Human bodys' response method based on the study of stratification proper subspace | |
CN114821764A (en) | Gesture image recognition method and system based on KCF tracking detection | |
CN110544267B (en) | Correlation filtering tracking method for self-adaptive selection characteristics | |
CN113223055A (en) | Image target tracking model establishing method and image target tracking method | |
CN102222231B (en) | Visual attention information computing device based on guidance of dorsal pathway and processing method thereof | |
CN102930250B (en) | A kind of action identification method of multi-scale random field models | |
CN112819930A (en) | Real-time role garment fabric animation simulation method based on feedforward neural network | |
CN110070563A (en) | Correlation filter method for tracking target and system based on joint perception | |
CN111539985A (en) | Self-adaptive moving target tracking method fusing multiple features |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190730 |
|
RJ01 | Rejection of invention patent application after publication |