CN107798686A - A kind of real-time modeling method method that study is differentiated based on multiple features - Google Patents

A kind of real-time modeling method method that study is differentiated based on multiple features Download PDF

Info

Publication number
CN107798686A
CN107798686A CN201710788553.1A CN201710788553A CN107798686A CN 107798686 A CN107798686 A CN 107798686A CN 201710788553 A CN201710788553 A CN 201710788553A CN 107798686 A CN107798686 A CN 107798686A
Authority
CN
China
Prior art keywords
mrow
feature
msub
represent
characteristic
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
Application number
CN201710788553.1A
Other languages
Chinese (zh)
Inventor
青春美
邓佳丽
徐向民
邢晓芬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
South China University of Technology SCUT
Original Assignee
South China University of Technology SCUT
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by South China University of Technology SCUT filed Critical South China University of Technology SCUT
Priority to CN201710788553.1A priority Critical patent/CN107798686A/en
Publication of CN107798686A publication Critical patent/CN107798686A/en
Pending legal-status Critical Current

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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/277—Analysis of motion involving stochastic approaches, e.g. using Kalman filters

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of real-time modeling method method that study is differentiated based on multiple features, including step：1) the greyscale video frame in video is obtained, the brightness attribute of tracking target is described using Cross bin distribution fields feature；2) the texture diversity for tracking target is modeled using enhancing histogram of gradients feature EHOG；3) by the color video frame of video, color characteristic CN is extracted to keep colour consistency；4) by step 1), 2), 3) in obtained various dimensions feature pass through Hilbert space mapping projections to high-dimensional feature space, obtain inner product mapping；5) obtained confidence map is put into CSK frameworks and be tracked, find tracking position of object, then more new template carries out target tracking.The present invention can effectively solve illumination variation present in target following, ambient interferences, block the problem such as low with real-time.

Description

A kind of real-time modeling method method that study is differentiated based on multiple features
Technical field
The present invention relates to the technical field of target following, refers in particular to a kind of real-time target that study is differentiated based on multiple features Tracking, available for intelligent video monitoring, automatic Pilot and man-machine interaction etc..
Background technology
One of the problem of target following is most challenging in computer vision field.It play in different applications to Important effect is closed, particularly has important application in military, medical treatment, monitoring and man-machine interaction.Recent years have Many algorithms are all used for solving the problems, such as target following, but because the deformation of target, the change of illumination, and target are blocked The problems such as influence, receive very big influence for the performance of target tracking algorism.
Current tracking framework generally can be divided into two modules：Expression model and trace model.Table in majority tracking framework Monochrome information is depended on up to model extraction target signature, until Lucas and Kanade are proposed based on the complete of unprocessed information more Office's template, certain methods are using brightness histogram come to the apparent modeling of target, but such a method lost spatial information.For solution Certainly spatial information problem, it is proposed that describe son comprising the multinuclear including spatial information.
When illumination variation, the apparent change of target is more violent.Single illumination intensity information is very easy to by illumination shadow Ring, can not robust tracking target, therefore propose diversified illumination invariant feature, it is such as a variety of not by illumination variation Texture information, as covariance region description, local binary patterns.In the recent period, online increment subspace model also be used to locate Manage target following under illumination variation.
Because color of object contains more rich information, a variety of tracking based on color apparent model are proposed in recent years Algorithm, the average drifting Mean Shift algorithms such as based on color histogram, color space are proposed using multiple color attribute Self-adaptive reduced-dimensions mechanism tracks target.However, it is simple use color characteristic cause track algorithm to background similar in color with The tracking target of low saturation is unstable.
Most experiments show, take a kind of single characteristic information to tackle the change under the conditions of several scenes, do not have The algorithm for having any single feature can handle the tracking problem under all scenes.Therefore, multiple features are merged and carry out target Tracking becomes main flow.Most researchers merge multiple clues, such as dynamic bayesian network, Meng Teka using probabilistic method Sieve algorithm, particle filter etc..The selection of feature and syncretizing mechanism are the challenges of target following in multi thread track algorithm.
The content of the invention
It is an object of the invention to overcome the deficiencies in the prior art, it is proposed that a kind of achievable stabilization, robust based on Multiple features differentiate study real-time modeling method method, can effectively solve illumination variation present in target following, ambient interferences, Block the problem such as low with real-time.
To achieve the above object, technical scheme provided by the present invention is：It is a kind of to differentiate the real-time of study based on multiple features Method for tracking target, comprise the following steps：
1) the greyscale video frame in video is obtained, the brightness category of tracking target is described using Cross-bin distribution fields feature Property；
2) the texture diversity for tracking target is modeled using enhancing histogram of gradients feature EHOG；
3) by the color video frame of video, color characteristic CN is extracted to keep colour consistency；
4) by step 1), 2), 3) in obtained various dimensions feature pass through Hilbert space mapping projections to high dimensional feature Space, inner product mapping is obtained, is comprised the following steps：
4.1) original CSK is extended to multidimensional characteristic by the data of acquisition by multichannel correlation filter, that is, will Multidimensional characteristic in Fourier using the Gauss weights read group total that variance is σ into final feature；
4.2) using differentiate study method learning procedure 1), 2), 3) in various dimensions characteristic response figure adaptively melted Close；
4.3) feature of different clues is accordingly subjected to multi thread weighting using the weights that step 4.2) learns, obtained most The weighting confidence map of all features eventually；
5) confidence map for obtaining step 4.3) is put into CSK frameworks and is tracked, and finds tracking position of object, then More new template carries out target tracking.
In step 1), a video sequence to be tested is chosen from normal video storehouse, the total quantity for acquisition is T Video sequence, since the 1st frame, determine the target data x of t frames, t is 1 to the arbitrary frame between T, and it is labeled as xm,n, m and n represent the ranks value of gray level image；
In step 2), for tape label training data (x1,1,y1,1),...,(xm,n,ym,n), wherein xm,nTo train sample Notebook data, ym,nIt is expected as a result, it is desirable to calculate various dimensions feature Ff(m, n, k), k represent the bin values of gray feature histogram, Various dimensions feature includes brightness CDF, textural characteristics EHOG and color characteristic CN, is specifically calculated as follows：
1. calculate brightness Fcdf(m, n, k), using local correlations approximation and cross-bin distribution field measures To describe brightness；Wherein, distribution field is the probability distribution matrix of each pixel of image, represents that this pixel belongs to a spy The probable value of sign, a distribution field are the probability distribution matrixes of a 3-dimensional, including the width of image, height and signature grey scale value Scope, in gray feature space, w × h sizes picture can generate three-dimensional w × h × b Characteristic Field, and w is the width of image Degree, h is the height of image, and b is the quantity of the bin values of gray feature histogram, is used One original-gray image I (m, n) is split to the different distributions field of different characteristic layer, DkIt is a set, its scope is Dk =[255k/b, 255 (k+1)/b], k ∈ { 0,2 ..., b-1 } represent the bin values of gray feature histogram, and b is that gray feature is straight The quantity of the bin values of square figure；Then cross-bin metric forms are used, i.e. the mode of histogram edge filter calculates brightness Feature, specific formula areWherein,It is the Gaussian kernel of Spatial Dimension, standard Difference is σs,It is the one-dimensional Gaussian kernel of feature space, standard deviation σk, * expression convolution operations；
2. calculate textural characteristics Fehog(m, n, k), texture consistency is described using enhancing histograms of oriented gradients EHOG, Gradient is calculated using Finite-difference Filtering device [- 1,0 ,+1] and its transposition；Wherein, according to gradient to orientation-sensitive whether, Gradient is divided into contrast sensitivity B1With the insensitive B of contrast2, definition
θ (m, n) represents the direction of pixel in image, and p, q represent the direction that contrast is insensitive and contrast is sensitive respectively The number of gradient；Then B is represented with B1Or B2, k ∈ { 0,1 ..., p+q-1 }, k represent the bin values of histogram, then can The characteristic pattern F of grid where calculating (m, n)hog(m, n, k) is：α (m, n) table The range value of pixel in diagram picture, furthermore, it is necessary to which the consistency for obtaining gradient is normalized, use four kinds of different normalizings Change factor N1,-1(m,n,k)、N1,1(m,n,k)、N-1,-1(m,n,k)、N-1,1(m, n, k), obtain the characteristic pattern of gradient consistency Nδx,δY(m, n), wherein δx∈ { -1,1 }, δy∈ { -1,1 }, specific formula for calculation is：
After four kinds of different normalization modes, the characteristic pattern F after being normalizedehog(m, n, k), specific formula for calculation are：
3. calculate color characteristic Fcn(m, n, k), the RGB color used in computer need to be mapped to by mapping matrix In the probability color matrix of 11 dimensions, mapping matrix is by the study of Google picture searching to color probability is expressed as into Fcn(m, N, k)=Map (R (m, n), G (m, n), B (m, n), k), wherein R (m, n), G (m, n), B (m, n) are represented in RGB pictures respectively Three kinds of Color Channels, Map expressions are mapped to the mapping matrix of 11 dimension colors from RGB, and k represents the bin values of histogram, Fcn(m,n, K) value each selected in represents that this yuan of vegetarian refreshments belongs to the probability of each color in 11 kinds of colors；
In step 3), multi thread feature is merged, by the various dimensions feature obtained in step 2) according to formulaMerged, obtain the fusion feature of t framesWherein,Represent t frames Multi thread feature,Represent to carry out Fourier transformation, ∑kRepresent to carry out the accumulating operation on 0 to k, and according to formulaHilbert space mapping projections are carried out to high-dimensional feature space, obtain the interior of t frames Product mappingWherein, κ represents to carry out gaussian kernel function computing,The frequency-domain model of t frames is represented, its computational methods are
In step 4.1), multi thread feature is put into CSK frameworks and is tracked, by original CSK be extended to various dimensions and The tracker of multi thread, regard tracking as classification problem, separate foreground and background；Each pixel value is represented according to multiple clues Fusion standard judges that this position belongs to the probability of prospect or background；Every clue i.e. every kind of feature have a confidence feature Figure；The confidence characteristic pattern for merging all features just obtains final confidence map, this confidence map is put into CSK frameworks carry out with Track；
Multidimensional characteristic is extended by multichannel correlation filter, that is, multidimensional characteristic is used in Fourier Variance be σ Gauss weights read group total into final feature,
Wherein, x(m,n)Represent the target data of t frames, x '(m,n)The target data of t+1 frames is represented, κ represents to carry out Gaussian kernel Functional operation, k represent the bin values of histogram, and exp represents the exponential function using natural constant e the bottom of as,Represent that Fourier becomes Change,Represent inverse Fourier transform；
In step 4.2), for training data x, closed using the kernel function of standardization least squares method RLS loss variances It is A to close solutionf=(Hf+λI)-1Y, wherein I are unit matrix, and λ is regularization coefficient, matrix HfElement be hm,n(x, x'), matrix Y={ ym,nIt is label result；The training sample auto-correlation under higher dimensional space is calculated by kernel function；By above-mentioned closed solution It is transformed into the parameter that CSK correlation pass filters are calculated in FourierSo as to which spy be calculated Sign response⊙ represents point multiplication operation；It is corresponding to differentiate minimum using multiple featuresMethod study to step 2.1), 2.2), 2.3) in each feature weights ωf, wherein rf Represent characteristic response rm,n, argmin () refers to so that function obtains the set of all independents variable of its minimum value；
In step 4.3), the characteristic response of different clues is subjected to multi thread using the weights that step 4.2) learns and added Power,Wherein FfIt is appearance features, MfIt is frequency-domain model, ωfIt is The weighted value of this feature, rm.nThe multi thread confidence map of all features of fusion as finally obtained；
In step 5), confidence map that step 4.3) is obtained, which is put into CSK frameworks, to be tracked, according toFind tracking position of objectIt is calculated by target location renewal frequency-domain model pre- The frequency-domain model M ' of surveyf,So as to which the closed solution A ' of prediction further be calculatedf, A 'f=Y/ (H'f+ λ)=Y/ (κ (M 'f,M′f)+λ), H'fThe inner product mapping of prediction is represented, is then used Renewal formwork calculation obtains the object module of t+1 framesWherein,Represent t The frequency-domain model of frame,The closed solution of t frames is represented, the target data of t+1 frames is found, realizes target tracking, wherein, λ is just Then term coefficient, γ ∈ [0,1] are learning rate.
The present invention compared with prior art, has the following advantages that and beneficial effect：
1st, the present invention combines the robust apparent model of various dimensions feature first and extension CSK (is tied based on kernel function circulation The Vision Tracking of structure) trace model, it is successfully realized target following.
2nd, the present invention finds target by merging three kinds of brightness, texture, colouring information features using the method for differentiating study With the decision boundary of background, the optimal result of study to multiple features fusion, improved model is in complicated and diversified environment Obtain leading tracking result.
3rd, three kinds of features that the present invention uses cope with different scenarios, under different conditions different features Occupy main positions.
Brief description of the drawings
Fig. 1 is the inventive method flow chart.
Embodiment
With reference to specific embodiment, the invention will be further described.
The real-time modeling method method shown in Figure 1, that the present embodiment is provided that study is differentiated based on multiple features, bag Include following steps：
1) the greyscale video frame in video is obtained, is specifically：A video sequence to be tested is chosen from normal video storehouse Row, for the video sequence for including T frames of acquisition, since t=1 frames, determine the target data x of t frames, it are labeled as xm,n, m and n represent the ranks value of gray level image.
2) tracking target is modeled using various dimensions feature, is specifically：For tape label training data (x1,1, y1,1),.......,(xm,n,ym,n), wherein x is training sample data, and y is expectation as a result, it is desirable to calculate various dimensions feature Ff, Various dimensions feature includes brightness CDF, textural characteristics EHOG, color characteristic CN, is specifically calculated as follows:
1. calculate brightness Fcdf(m, n, k), using local correlations approximation and cross-bin distribution field measures To describe brightness；Wherein, distribution field is the probability distribution matrix of each pixel of image, represents that this pixel belongs to a spy The probable value of sign, a distribution field are the probability distribution matrixes of a 3-dimensional, including the width of image, height and signature grey scale value Scope, in gray feature space, w × h sizes picture can generate three-dimensional w × h × b Characteristic Field, and w is the width of image Degree, h is the height of image, and b is the quantity of the bin values of gray feature histogram, is used One original-gray image I (m, n) is split to the different distributions field of different characteristic layer, DkIt is a set, its scope is Dk =[255k/b, 255 (k+1)/b], k ∈ { 0,2 ..., b-1 } represent the bin values of gray feature histogram, and b is that gray feature is straight The quantity of the bin values of square figure；Then cross-bin metric forms are used, i.e. the mode of histogram edge filter calculates brightness Feature, specific formula areWherein,It is the Gaussian kernel of Spatial Dimension, standard Difference is σs,It is the one-dimensional Gaussian kernel of feature space, standard deviation σk, * expression convolution operations；
2. calculate textural characteristics Fehog(m, n, k), texture consistency is described using enhancing histograms of oriented gradients EHOG, Gradient is calculated using Finite-difference Filtering device [- 1,0 ,+1] and its transposition；Wherein, according to gradient to orientation-sensitive whether, Gradient is divided into contrast sensitivity B1With the insensitive B of contrast2, definition
θ (m, n) represents the direction of pixel in image, and p, q represent the direction that contrast is insensitive and contrast is sensitive respectively The number of gradient；Then B is represented with B1Or B2, k ∈ { 0,1 ..., p+q-1 }, k represent the bin values of histogram, then can The characteristic pattern F of grid where calculating (m, n)hog(m, n, k) is：α (m, n) table The range value of pixel in diagram picture, furthermore, it is necessary to which the consistency for obtaining gradient is normalized, use four kinds of different normalizings Change factor N1,-1(m,n,k)、N1,1(m,n,k)、N-1,-1(m,n,k)、N-1,1(m, n, k), obtain the characteristic pattern of gradient consistencyWherein δx∈ { -1,1 }, δy∈ { -1,1 }, specific formula for calculation is：
After four kinds of different normalization modes, the characteristic pattern F after being normalizedehog(m, n, k), specific formula for calculation For：
3. calculate color characteristic Fcn(m, n, k), the RGB color used in computer need to be mapped to by mapping matrix In the probability color matrix of 11 dimensions, mapping matrix is by the study of Google picture searching to color probability is expressed as into Fcn(m, N, k)=Map (R (m, n), G (m, n), B (m, n), k), wherein R (m, n), G (m, n), B (m, n) are represented in RGB pictures respectively Three kinds of Color Channels, Map expressions are mapped to the mapping matrix of 11 dimension colors from RGB, and k represents the bin values of histogram, Fcn(m,n, K) value each selected in represents that this yuan of vegetarian refreshments belongs to the probability of each color in 11 kinds of colors；
3) multi thread feature is merged, by the various dimensions feature obtained in step 2) according to formula Merged, obtain the fusion feature of t framesWherein,The multi thread feature of t frames is represented,Represent into Row Fourier transformation, ∑kRepresent to carry out the accumulating operation on 0 to k, and according to formulaCarry out Hilbert space mapping projections obtain the inner product mapping of t frames to high-dimensional feature spaceWherein, κ represents to carry out Gauss Kernel function computing,The frequency-domain model of t frames is represented, its computational methods are
4) by step 1), 2), 3) in obtained various dimensions feature pass through Hilbert space mapping projections to high dimensional feature Space, inner product mapping is obtained, is comprised the following steps：
4.1) multi thread feature is put into CSK frameworks to be tracked, by original CSK be extended to various dimensions and multi thread with Track device, regard tracking as classification problem, separate foreground and background；Each pixel value represent according to multiple clues merge standard come Judge that this position belongs to the probability of prospect or background；Every clue i.e. every kind of feature have a confidence characteristic pattern；Fusion institute The confidence characteristic pattern for having feature just obtains final confidence map, and this confidence map is put into CSK frameworks and is tracked；
Multidimensional characteristic is extended by multichannel correlation filter, that is, by multidimensional characteristic in Fourier Using the Gauss weights read group total that variance is σ into final feature,
Wherein, x(m,n)Represent the target data of t frames, x '(m,n)The target data of t+1 frames is represented, κ represents to carry out Gaussian kernel Functional operation, k represent the bin values of histogram, and exp represents the exponential function using natural constant e the bottom of as,Represent that Fourier becomes Change,Represent inverse Fourier transform；
4.2) training data x is directed to, the kernel function closed solution using standardization least squares method RLS loss variances is Af= (Hf+λI)-1Y, wherein I are unit matrix, and λ is regularization coefficient, matrix HfElement be hm,n(x, x'), matrix Y= {ym,nIt is label result；The training sample auto-correlation under higher dimensional space is calculated by kernel function；Above-mentioned closed solution is turned Change to Fourier and the parameter of CSK correlation pass filters is calculatedSo as to which spy be calculated Sign response⊙ represents point multiplication operation；It is corresponding to differentiate minimum using multiple featuresMethod study to step 2.1), 2.2), 2.3) in each feature weights ωf, wherein rfTable Show characteristic response rm,n, argmin () refers to so that function obtains the set of all independents variable of its minimum value；
4.3) characteristic response of different clues is subjected to multi thread weighting using the weights that step 4.2) learns,Wherein FfIt is appearance features, MfIt is frequency-domain model, ωfIt is this The weighted value of feature, rm.nThe multi thread confidence map of all features of fusion as finally obtained；
5) confidence map for obtaining step 4.3) is put into CSK frameworks and is tracked, according to Find tracking position of objectThe frequency-domain model M ' of prediction is calculated by target location renewal frequency-domain modelf,So as to which the closed solution A ' of prediction further be calculatedf, A 'f=Y/ (H'f+ λ)=Y/ (κ (M 'f, M′f)+λ), H'fThe inner product mapping of prediction is represented, is then used Renewal formwork calculation obtains the object module of t+1 framesWherein,The frequency-domain model of t frames is represented,Represent The closed solution of t frames, the target data of t+1 frames is found, realizes target tracking, wherein, λ is regularization coefficient, and γ ∈ [0,1] are Learning rate.
Embodiment described above is only the preferred embodiments of the invention, and the implementation model of the present invention is not limited with this Enclose, therefore the change that all shape, principles according to the present invention are made, it all should cover within the scope of the present invention.

Claims (2)

  1. A kind of 1. real-time modeling method method that study is differentiated based on multiple features, it is characterised in that comprise the following steps：
    1) the greyscale video frame in video is obtained, the brightness attribute of tracking target is described using Cross-bin distribution fields feature；
    2) the texture diversity for tracking target is modeled using enhancing histogram of gradients feature EHOG；
    3) by the color video frame of video, color characteristic CN is extracted to keep colour consistency；
    4) by step 1), 2), 3) in obtained various dimensions feature by Hilbert space mapping projections to high-dimensional feature space, Inner product mapping is obtained, is comprised the following steps：
    4.1) original CSK is extended to multidimensional characteristic by the data of acquisition by multichannel correlation filter, that is, by multidimensional Feature in Fourier using the Gauss weights read group total that variance is σ into final feature；
    4.2) using differentiate study method learning procedure 1), 2), 3) in various dimensions characteristic response figure adaptively merged；
    4.3) feature of different clues is accordingly subjected to multi thread weighting using the weights that step 4.2) learns, obtains final institute There is the weighting confidence map of feature；
    5) confidence map for obtaining step 4.3) is put into CSK frameworks and is tracked, and finds tracking position of object, then updates Template carries out target tracking.
  2. A kind of 2. real-time modeling method method that study is differentiated based on multiple features according to claim 1, it is characterised in that：
    In step 1), a video sequence to be tested is chosen from normal video storehouse, the total quantity for acquisition is regarding for T Frequency sequence, since the 1st frame, determine that it, to the arbitrary frame between T, is labeled as x by the target data x of t frames, t for 1m,n, m and N represents the ranks value of gray level image；
    In step 2), for tape label training data (x1,1,y1,1),...,(xm,n,ym,n), wherein xm,nFor number of training According to ym,nIt is expected as a result, it is desirable to calculate various dimensions feature Ff(m, n, k), k represent the bin values of gray feature histogram, multidimensional Degree feature includes brightness CDF, textural characteristics EHOG and color characteristic CN, is specifically calculated as follows：
    1. calculate brightness Fcdf(m, n, k), retouched using local correlations approximation with cross-bin distribution fields measure State brightness；Wherein, distribution field is the probability distribution matrix of each pixel of image, represents that this pixel belongs to feature Probable value, a distribution field are the probability distribution matrixes of a 3-dimensional, include the model of the width of image, height and signature grey scale value Enclose, in gray feature space, w × h sizes picture can generate three-dimensional w × h × b Characteristic Field, and w is the width of image, h It is the height of image, b is the quantity of the bin values of gray feature histogram, is usedWill One original-gray image I (m, n) is split to the different distributions field of different characteristic layer, DkIt is a set, its scope is Dk= [255k/b, 255 (k+1)/b], k ∈ { 0,2 ..., b-1 } represent the bin values of gray feature histogram, and b is gray feature Nogata The quantity of the bin values of figure；Then cross-bin metric forms are used, i.e. the mode of histogram edge filter calculates brightness spy Sign, specific formula areWherein,It is the Gaussian kernel of Spatial Dimension, standard deviation For σs,It is the one-dimensional Gaussian kernel of feature space, standard deviation σk, * expression convolution operations；
    2. calculate textural characteristics Fehog(m, n, k), texture consistency is described using enhancing histograms of oriented gradients EHOG, used Finite-difference Filtering device [- 1,0 ,+1] and its transposition calculate gradient；Wherein, according to gradient to orientation-sensitive whether, by ladder Degree is divided into contrast sensitivity B1With the insensitive B of contrast2, definition
    θ (m, n) represents the direction of pixel in image, and p, q represent the direction that contrast is insensitive and contrast is sensitive respectively The number of gradient；Then B is represented with B1Or B2, k ∈ { 0,1 ..., p+q-1 }, k represent the bin values of histogram, then can The characteristic pattern F of grid where calculating (m, n)hog(m, n, k) is：α (m, n) table The range value of pixel in diagram picture, furthermore, it is necessary to which the consistency for obtaining gradient is normalized, use four kinds of different normalizings Change factor N1,-1(m,n,k)、N1,1(m,n,k)、N-1,-1(m,n,k)、N-1,1(m, n, k), obtain the characteristic pattern of gradient consistencyWherein δx∈ { -1,1 }, δy∈ { -1,1 }, specific formula for calculation is：
    After four kinds of different normalization modes, the characteristic pattern F after being normalizedehog(m, n, k), specific formula for calculation are：
    <mrow> <msub> <mi>F</mi> <mrow> <mi>e</mi> <mi>h</mi> <mi>o</mi> <mi>g</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <msub> <mi>F</mi> <mrow> <mi>h</mi> <mi>o</mi> <mi>g</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>/</mo> <msub> <mi>N</mi> <mrow> <mn>1</mn> <mo>,</mo> <mn>1</mn> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>F</mi> <mrow> <mi>h</mi> <mi>o</mi> <mi>g</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>/</mo> <msub> <mi>N</mi> <mrow> <mn>1</mn> <mo>,</mo> <mo>-</mo> <mn>1</mn> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>F</mi> <mrow> <mi>h</mi> <mi>o</mi> <mi>g</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>/</mo> <msub> <mi>N</mi> <mrow> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mn>1</mn> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>F</mi> <mrow> <mi>h</mi> <mi>o</mi> <mi>g</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>/</mo> <msub> <mi>N</mi> <mrow> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mo>-</mo> <mn>1</mn> </mrow> </msub> <mrow> <mo>(</mo> <mi>m</mi> <mo>,</mo> <mi>n</mi> <mo>,</mo> <mi>k</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> </mtable> </mfenced> </mrow>
    3. calculate color characteristic Fcn(m, n, k), the RGB color used in computer need to be mapped to 11 dimensions by mapping matrix In the probability color matrix of degree, mapping matrix is by the study of Google picture searching to color probability is expressed as into Fcn(m,n,k) =Map (R (m, n), G (m, n), B (m, n), k), wherein R (m, n), G (m, n), B (m, n) represent three kinds in RGB pictures respectively Color Channel, Map expressions are mapped to the mapping matrix of 11 dimension colors from RGB, and k represents the bin values of histogram, FcnIn (m, n, k) The value each selected represents that this yuan of vegetarian refreshments belongs to the probability of each color in 11 kinds of colors；
    In step 3), multi thread feature is merged, by the various dimensions feature obtained in step 2) according to formula Merged, obtain the fusion feature of t framesWherein,The multi thread feature of t frames is represented,Represent to carry out Fourier transformation, ∑kRepresent to carry out the accumulating operation on 0 to k, and according to formulaWished Your Bert space reflection projects to high-dimensional feature space, obtains the inner product mapping of t framesWherein, κ represents to carry out Gaussian kernel Functional operation,The frequency-domain model of t frames is represented, its computational methods are
    In step 4.1), multi thread feature is put into CSK frameworks and is tracked, original CSK is extended to various dimensions and multi-thread The tracker of rope, regard tracking as classification problem, separate foreground and background；Each pixel value represents to be merged according to multiple clues Standard judges that this position belongs to the probability of prospect or background；Every clue i.e. every kind of feature have a confidence characteristic pattern； The confidence characteristic pattern for merging all features just obtains final confidence map, and this confidence map is put into CSK frameworks and is tracked；
    Multidimensional characteristic is extended by multichannel correlation filter, that is, it is σ's that multidimensional characteristic is used into variance in Fourier Gauss weights read group total into final feature,
    Wherein, x(m,n)Represent the target data of t frames, x '(m,n)The target data of t+1 frames is represented, κ represents to carry out Gaussian kernel letter Number computing, k represent the bin values of histogram, and exp represents the exponential function using natural constant e the bottom of as,Represent Fourier transformation,Represent inverse Fourier transform；
    In step 4.2), for training data x, the kernel function closed solution of standardization least squares method RLS loss variances is used For Af=(Hf+λI)-1Y, wherein I are unit matrix, and λ is regularization coefficient, matrix HfElement be hm,n(x, x'), matrix Y= {ym,nIt is label result；The training sample auto-correlation under higher dimensional space is calculated by kernel function；Above-mentioned closed solution is converted The parameter of CSK correlation pass filters is calculated to FourierRung so as to which feature be calculated Should⊙ represents point multiplication operation；It is corresponding to differentiate minimum using multiple featuresMethod study to step 2.1), 2.2), 2.3) in each feature weights ωf, wherein rf Represent characteristic response rm,n, argmin () refers to so that function obtains the set of all independents variable of its minimum value；
    In step 4.3), the characteristic response of different clues is subjected to multi thread weighting using the weights that step 4.2) learns,Wherein FfIt is appearance features, MfIt is frequency-domain model, ωfIt is this The weighted value of feature, rm.nThe multi thread confidence map of all features of fusion as finally obtained；
    In step 5), confidence map that step 4.3) is obtained, which is put into CSK frameworks, to be tracked, according to Find tracking position of objectThe frequency-domain model M ' of prediction is calculated by target location renewal frequency-domain modelf,So as to which the closed solution A ' of prediction further be calculatedf, A 'f=Y/ (H'f+ λ)=Y/ (κ (M 'f, M′f)+λ), H'fThe inner product mapping of prediction is represented, is then used Renewal formwork calculation obtains the object module of t+1 framesWherein,The frequency-domain model of t frames is represented,Represent the The closed solution of t frames, the target data of t+1 frames is found, realizes target tracking, wherein, λ is regularization coefficient, and γ ∈ [0,1] are Habit rate.
CN201710788553.1A 2017-09-04 2017-09-04 A kind of real-time modeling method method that study is differentiated based on multiple features Pending CN107798686A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710788553.1A CN107798686A (en) 2017-09-04 2017-09-04 A kind of real-time modeling method method that study is differentiated based on multiple features

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710788553.1A CN107798686A (en) 2017-09-04 2017-09-04 A kind of real-time modeling method method that study is differentiated based on multiple features

Publications (1)

Publication Number Publication Date
CN107798686A true CN107798686A (en) 2018-03-13

Family

ID=61531542

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710788553.1A Pending CN107798686A (en) 2017-09-04 2017-09-04 A kind of real-time modeling method method that study is differentiated based on multiple features

Country Status (1)

Country Link
CN (1) CN107798686A (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109255304A (en) * 2018-08-17 2019-01-22 西安电子科技大学 Method for tracking target based on distribution field feature
CN109410245A (en) * 2018-09-13 2019-03-01 北京米文动力科技有限公司 A kind of video target tracking method and equipment
CN109816694A (en) * 2019-01-28 2019-05-28 北京旷视科技有限公司 Method for tracking target, device and electronic equipment
CN109934853A (en) * 2019-03-21 2019-06-25 云南大学 Correlation filtering tracking based on the fusion of response diagram confidence region self-adaptive features
CN110009660A (en) * 2019-03-06 2019-07-12 浙江大学 Object space method for tracing based on correlation algorithm filter
CN110766723A (en) * 2019-10-22 2020-02-07 湖南大学 Unmanned aerial vehicle target tracking method and system based on color histogram similarity
CN110852335A (en) * 2019-11-19 2020-02-28 燕山大学 Target tracking system based on multi-color feature fusion and depth network
CN111145121A (en) * 2019-12-27 2020-05-12 安徽工业大学 Confidence term filter target tracking method for strengthening multi-feature fusion

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103218628A (en) * 2013-03-22 2013-07-24 中国科学技术大学 Abnormal behavior description method based on characteristics of block mass and track
CN104992451A (en) * 2015-06-25 2015-10-21 河海大学 Improved target tracking method
CN105872859A (en) * 2016-06-01 2016-08-17 深圳市唯特视科技有限公司 Video compression method based on moving target trajectory extraction of object
CN106530325A (en) * 2016-10-26 2017-03-22 合肥润客软件科技有限公司 Multi-target visual detection and tracking method
CN106570486A (en) * 2016-11-09 2017-04-19 华南理工大学 Kernel correlation filtering target tracking method based on feature fusion and Bayesian classification

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103218628A (en) * 2013-03-22 2013-07-24 中国科学技术大学 Abnormal behavior description method based on characteristics of block mass and track
CN104992451A (en) * 2015-06-25 2015-10-21 河海大学 Improved target tracking method
CN105872859A (en) * 2016-06-01 2016-08-17 深圳市唯特视科技有限公司 Video compression method based on moving target trajectory extraction of object
CN106530325A (en) * 2016-10-26 2017-03-22 合肥润客软件科技有限公司 Multi-target visual detection and tracking method
CN106570486A (en) * 2016-11-09 2017-04-19 华南理工大学 Kernel correlation filtering target tracking method based on feature fusion and Bayesian classification

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
徐桂从: ""人脸的跟踪识别与年龄估计"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109255304B (en) * 2018-08-17 2021-07-27 西安电子科技大学 Target tracking method based on distribution field characteristics
CN109255304A (en) * 2018-08-17 2019-01-22 西安电子科技大学 Method for tracking target based on distribution field feature
CN109410245A (en) * 2018-09-13 2019-03-01 北京米文动力科技有限公司 A kind of video target tracking method and equipment
CN109410245B (en) * 2018-09-13 2021-08-10 北京米文动力科技有限公司 Video target tracking method and device
CN109816694A (en) * 2019-01-28 2019-05-28 北京旷视科技有限公司 Method for tracking target, device and electronic equipment
CN110009660A (en) * 2019-03-06 2019-07-12 浙江大学 Object space method for tracing based on correlation algorithm filter
CN109934853A (en) * 2019-03-21 2019-06-25 云南大学 Correlation filtering tracking based on the fusion of response diagram confidence region self-adaptive features
CN110766723B (en) * 2019-10-22 2020-11-24 湖南大学 Unmanned aerial vehicle target tracking method and system based on color histogram similarity
CN110766723A (en) * 2019-10-22 2020-02-07 湖南大学 Unmanned aerial vehicle target tracking method and system based on color histogram similarity
CN110852335A (en) * 2019-11-19 2020-02-28 燕山大学 Target tracking system based on multi-color feature fusion and depth network
CN110852335B (en) * 2019-11-19 2023-06-20 燕山大学 Target tracking system based on multi-color feature fusion and depth network
CN111145121A (en) * 2019-12-27 2020-05-12 安徽工业大学 Confidence term filter target tracking method for strengthening multi-feature fusion
CN111145121B (en) * 2019-12-27 2023-02-28 安徽工业大学 Confidence term filter target tracking method for strengthening multi-feature fusion

Similar Documents

Publication Publication Date Title
CN107798686A (en) A kind of real-time modeling method method that study is differentiated based on multiple features
CN108986140B (en) Target scale self-adaptive tracking method based on correlation filtering and color detection
CN108549891B (en) Multi-scale diffusion well-marked target detection method based on background Yu target priori
US9025880B2 (en) Visual saliency estimation for images and video
CN109753975A (en) Training sample obtaining method and device, electronic equipment and storage medium
Yu et al. Context-based hierarchical unequal merging for SAR image segmentation
CN105761238B (en) A method of passing through gray-scale statistical data depth information extraction well-marked target
CN110929593A (en) Real-time significance pedestrian detection method based on detail distinguishing and distinguishing
CN106778687A (en) Method for viewing points detecting based on local evaluation and global optimization
Ma et al. Saliency detection based on singular value decomposition
CN110222572A (en) Tracking, device, electronic equipment and storage medium
CN109146925A (en) Conspicuousness object detection method under a kind of dynamic scene
Xiao et al. Pedestrian object detection with fusion of visual attention mechanism and semantic computation
Xu et al. Bayberry image segmentation based on manifold ranking salient object detection method
CN107610136B (en) Salient object detection method based on convex hull structure center query point sorting
CN107423771B (en) Two-time-phase remote sensing image change detection method
CN110135435B (en) Saliency detection method and device based on breadth learning system
CN106846377A (en) A kind of target tracking algorism extracted based on color attribute and active features
Lou et al. Hierarchical co-salient object detection via color names
Yuan et al. Explore double-opponency and skin color for saliency detection
Xu et al. Covariance descriptor based convolution neural network for saliency computation in low contrast images
You et al. Salient object detection via point-to-set metric learning
Zhang et al. A unified saliency detection framework for visible and infrared images
Sun et al. Study on subtractive clustering video moving object locating method with introduction of eigengap
CN110472653A (en) A kind of semantic segmentation method based on maximization region mutual information

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20180313

WD01 Invention patent application deemed withdrawn after publication