CN110276785A - One kind is anti-to block infrared object tracking method - Google Patents
One kind is anti-to block infrared object tracking method Download PDFInfo
- Publication number
- CN110276785A CN110276785A CN201910547576.2A CN201910547576A CN110276785A CN 110276785 A CN110276785 A CN 110276785A CN 201910547576 A CN201910547576 A CN 201910547576A CN 110276785 A CN110276785 A CN 110276785A
- Authority
- CN
- China
- Prior art keywords
- frame image
- target
- template
- current frame
- center
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
- G06T7/251—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving models
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10048—Infrared image
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
Resist the invention discloses one kind and block infrared object tracking method, solves the problem of that attitudes vibration blocks target long time-tracking under equal complex environments, belong to target following technology and computer vision field.The present invention reads infrared image sequence, target is selected in initial frame image center, obtain center and the size of target, using the target in initial frame image as template, the second frame image is obtained as current frame image, using the template of initial frame image as the template of current frame image, Two-dimensional Cosine window is obtained according to the size of template and cell factory size;The feature and linear fusion that template is extracted based on histograms of oriented gradients and Ha Er are initialized or are updated object module and goal regression coefficient, then obtain multilayer core correlation filtering response set of graphs based on the search box in each frame image and carry out succeeding target tracking.The present invention is used for infrared image target following.
Description
Technical field
One kind is anti-to block infrared object tracking method, is used for infrared image target following, belongs to target following technology and meter
Calculation machine visual field.
Background technique
All the time, target following it is military it is civilian all obtain very big application, nowadays, the various monitoring based on camera
Equipment is throughout many important places, such as airport, market, factory, school etc., for prevent and solve safety problem provide it is important
Effect, but traditional monitoring system still has certain defect to take off such as by attitudes vibration, target occlusion, motion blur, change of scale
It is big from influences such as the visuals field, in the above complex environment, due to target drastic mechanical deformation or be blocked, calculated at present still without good
Method can be automatically performed the tracking of personage interested, it is necessary to by artificial strength.
Method for tracking target mainly has based on area information, such as template matching method at present, and simple correct velocity is fastly but not
The complex environments such as target drastic mechanical deformation are adapted to, easily cause target to lose in such cases;Based on model information, by establishing mesh
Target geometrical model, then model is scanned for, this method is also difficult to solve occlusion issue, and lacks face in infrared environmental
Color information is anti-, and to block ability weaker;Based on Bayesian frame, that is, capturing target original state and by the mesh of feature extraction
It marks in feature base, carries out a kind of Target state estimator of space-time combination, the target position estimation in the case of can be used for being blocked,
But algorithm complexity is higher;There is good robustness based on deep learning class method but be easy to appear shortage of data problem, and net
Network training speed is difficult to reach requirement of real-time;Based on correlation filtering, such methods speed is fast, wherein KCF filtering has fast
The high feature of speed, accuracy, compare the track algorithms such as Struck and TLD, and tracking velocity improves nearly 10 times, compares
The MOSSE algorithm that OBT50 accuracy is 43.1%, and there is high accuracy, the accuracy using HOG feature
Up to 73.2%.
Influenced for the IR imaging target tracking under complex situations by object variations and external environment it is difficult to ensure that with
Track accuracy, finding a kind of anti-long time-tracking algorithm blocked is current urgent problem.And it is more existing based on core
The innovatory algorithm of correlation filtering solve the problems, such as to a certain extent tracking during target be blocked cause tracking failure.No
Target scale is considered when great changes will take place, i.e., under by circumstance of occlusion, such algorithm still largely can be completed accurately
Target following is realized in the matching of region of search and target, but if considering that target scale when great changes will take place, is being blocked feelings
Under condition, it is easy for the problem of causing tracking failure.
Summary of the invention
Aiming at the problem that the studies above, resist the purpose of the present invention is to provide one kind and block infrared object tracking method, solves
The problem of method for tracking target of use certainly in the prior art, being blocked influences, and be easy to cause tracking failure.
In order to achieve the above object, the present invention adopts the following technical scheme:
One kind is anti-to block infrared object tracking method, includes the following steps:
S1: reading infrared image sequence, select target in initial frame image center, obtain center and the size of target,
Using the target in initial frame image as template, the second frame image is obtained as current frame image, by the template of initial frame image
Template as current frame image;
S2: Two-dimensional Cosine window is obtained according to the size of template and cell factory size;
S3: the feature and linear fusion of template are extracted based on histograms of oriented gradients and Ha Er, to the spy after linear fusion
Sign plus Two-dimensional Cosine window obtain fusion feature, then goal regression coefficient is calculated based on fusion feature, if goal regression coefficient
Gained is calculated for the second frame image, with goal regression coefficient initialization object module and goal regression coefficient, if goal regression system
Number is that last frame image calculates gained, does not deal with, otherwise updates object module and goal regression coefficient;
S4: the search box of current frame image is determined using the template center of current frame image as central position of search frame;
S5: being based on template size, traversed in the search box in current frame image, obtain regional ensemble to be matched,
The corresponding fusion feature in multiple regions to be matched is obtained based on regional ensemble to be matched again, is based on fusion feature and corresponding target
Model and goal regression coefficient calculate the corresponding multilayer core correlation filtering response diagram in each region to be matched to get multilayer nuclear phase is arrived
Close filter response set of graphs;
S6: judge whether the maximum response in multilayer core correlation filtering response set of graphs is more than or equal to the first given threshold
Value, if more than going to step S7, otherwise going to S8;
S7: the center of the target of current frame image is calculated with the transverse and longitudinal coordinate of maximum response, if current frame image
It is not last frame, updates the center that the template in step S3 is the template of current frame image and the target of current frame image
Weighting, gone to after update step S3 carry out next frame processing, otherwise terminate to track;
S8: template does not update, and with the dbjective state in Kalman filter prediction current frame image, obtains prediction coordinate, with
Centered on predicting coordinate, the target area for taking size to obtain with the consistent region of template size with current frame image actual match adds
Act temporarily as the matching result for current frame image, wherein target area is with the center of previous frame target or with peak response
Centered on the transverse and longitudinal coordinate of value, size and the consistent region of template size take matching centered on the center of matching result
As a result 3 times of sizes traverse to obtain regional ensemble to be matched as search box, extract each region to be matched in regional ensemble to be matched
Histograms of oriented gradients and Lis Hartel sign, obtain corresponding fusion feature, then the filter of multilayer core correlation is obtained based on each fusion feature
Wave responds set of graphs:
If multilayer core correlation filtering, which responds maximum response in set of graphs, is more than or equal to given second threshold, with maximum value
Transverse and longitudinal coordinate more fresh target center, if current frame image is not last frame, update step S3 in template be work as
The weighting of the center of the template and present frame target of prior image frame goes to step S3 and carries out next frame processing after update, no
Then terminate to track;
If multilayer core correlation filtering responds in set of graphs, maximum response is lower than given second threshold, if current frame image
Then terminate to track for last frame, otherwise, reads next frame image as present frame, go to step S8.
Further, specific step is as follows by the step S1: reading infrared image sequence, selects mesh in initial frame image center
Mark, records center and the size of target, using the target of initial frame image center choosing as template, wherein in template
Heart position and size are center and the size of target;The second frame image is obtained as current frame image, by initial frame figure
Template of the template of picture as current frame image.
Further, specific step is as follows by the step S2:
S2.1: according to the size target_sz of template, search box is determined, the size of search box is window_sz=
Target_sz* (1+padding), wherein padding is the ratio for determining search box size and target sizes;
S2.2: according to the big of given cell factory size cell_size, the size target_sz of template and search box
Small window_sz determines that feature returns label yf, then returns label yf based on feature and obtain Two-dimensional Cosine window cos_window;
Specific step is as follows:
S2.2.1: Gauss is calculated according to the size target_sz of template and cell factory cell_size size and returns label
Bandwidth σ, formula is as follows:
In formula, w and h are the width and height of template, and a is spatial bandwidth, proportional to target sizes;
S2.2.2: returning the bandwidth σ of label and the size window_sz of search box according to Gauss, calculate and return label yf,
Calculation formula is as follows:
Wherein, m, n are respectivelyFirst item and Section 2, after y ' is calculated, cyclic shift make return mark
Label peak value moves on to the upper left corner and obtains y, then carries out Fourier transformation, obtains returning label yf;
S2.2.3: it is calculated according to the size for returning label yf using hann function, obtains Two-dimensional Cosine window cos_window;
Further, specific step is as follows by the step S3:
S3.1: template m is extracted based on HOG and HaartFeature and linear fusion, two dimension is added to the feature after linear fusion
Cosine Window cos_window obtains fusion feature, wherein Haar Ha Er, HOG are histograms of oriented gradients;
S3.2: goal regression coefficient is obtained according to fusion feature g;
S3.3: if current frame image is the second frame image, going to step S3.4, if last frame image, be not processed,
Otherwise step S3.6 is gone to;
S3.4: when target following the second frame image, the fusion feature xf of the frequency domain of current frame image, initialized target are utilized
Model modle_xf, i.e.,
Wherein, t indicates the second frame image,Indicate the fusion feature xf of frequency domain,Indicate object module modle_xf;
S3.5: when target following the second frame image, using the goal regression factor alpha of current frame image, initialized target is returned
Return Coefficient m odle_ α, i.e.,
Wherein, t indicates the second frame image,Indicate goal regression factor alpha,Indicate goal regression Coefficient m odle_ α;
S3.6: when image after target following third frame or third frame, object module is updated by linear interpolation
Modle_xf, i.e.,
Wherein, η is given learning rate,For the object module of current frame image,For the target of previous frame image
Model obtains updatedAs updated modle_xf;
S3.7: when image after target following third frame or third frame, pass through the goal regression system that linear interpolation updates
Number modle_ α, i.e.,
Wherein, η is given learning rate,For the goal regression coefficient of previous frame image,For current frame image
Goal regression Coefficient m odle_ α.
Further, the specific steps of the step 3.1 are as follows:
S3.1.1: based on given cell factory size cell_size, the corresponding piotr_toolbox of MATLAB is utilized
Kit extracts template mtFHOG feature, obtain the FHOG feature g of 31 dimensions0, i.e. histograms of oriented gradients feature;
S3.1.2: current frame image is t frame image, calculates the template m of t frame imagetThe integrogram SAT of (x, y), product
The calculation formula of each pixel value in component SAT are as follows:
SAT (x, y)=SAT (x, y-1)+SAT (x-1, y)-SAT (x-1, y-1)+mt(x, y)
Wherein, SAT (x, y-1) indicates current pixel position x, the integrogram pixel value of the top y, and SAT (x-1, y) expression is worked as
The integrogram pixel value on the preceding pixel position left side x, y, SAT (x-1, y-1) indicate current pixel position x, the integrogram in the upper left corner y
Pixel value, the initial boundary of integrogram SAT are SAT (- 1, y)=SAT (x, -1)=SAT (- 1, -1)=0;SAT (- 1, y) is a left side
Border pixel values, SAT (x, -1) are coboundary pixel value, and SAT (- 1, -1) is upper left apex angle pixel value;Initial boundary SAT (- 1,
Y), SAT (x, -1) and SAT (- 1, -1) is for calculating SAT (0, y) and SAT (x, 0);
S3.1.3: integrogram SAT is divided according to cell factory size cell_size, i.e., by any cell factory upper half
Dividing the sum of pixel integration figure is SATA, the sum of cell factory lower half portion pixel integration figure is SATB, cell factory left-half picture
The sum of plain integrogram is SATC, the sum of cell factory right half part pixel integration figure is SATD, corresponding 1 dimension of each cell factory
Vertical direction Haar feature g1 is SATAWith SATBDifference, 1 dimension horizontal direction Haar feature g2 be SATCWith SATDDifference,
1 dimension vertical direction Haar feature g1 of all cell factories and 1 dimension horizontal direction Haar feature g2 is Lis Hartel sign;
S3.1.4: FHOG feature g is tieed up by 310, all 1 dimension vertical direction Haar feature g1, all 1 dimension horizontal directions
Haar feature g2After carrying out linear fusion, with Two-dimensional Cosine window cos_window dot product, one 33 dimension fusion feature g is obtained.
Further, the specific steps of the step 3.2 are as follows:
S3.2.1: fusion feature g is subjected to Fast Fourier Transform (FFT), obtains template mtIn the fusion feature xf of frequency domain,
Calculation template mtIn the formula of the fusion feature xf of frequency domain are as follows:
In formula, g is indicated to template mtThe 33 dimension fusion features extracted,Indicate Fourier transformation,Expression obtains
Frequency domain fusion feature xf;
S3.2.2: according to Gaussian kernel correlation function, the fusion feature xf based on frequency domain calculates the Gauss auto-correlation on frequency domain
Nuclear matrix kf;The formula of Gaussian kernel correlation function:
Wherein, Kxx′The Gaussian kernel correlation matrix of x and x ' is represented, x, x ', which respectively represent calculating Gaussian kernel correlation matrix, to be made
Different characteristic symbol replaces with corresponding feature in practical calculating process, | | x | |2It is characterized the mould of each element in x
Quadratic sum is again the product of two-dimension sizes before matrix x divided by N, N,Form of the representing matrix x in Fourier,It indicates's
Complex conjugate,Indicate inverse Fourier transform, σ is the bandwidth that Gauss returns label, and T indicates transposition;
Using the fusion feature xf of frequency domain, x, the x ' in the formula of Gaussian kernel correlation function are replaced with into xf, calculate frequency
Gauss auto-correlation nuclear matrix kf on domain;
S3.2.3: according to Gauss auto-correlation nuclear matrix kf, goal regression factor alpha, calculation formula are calculated are as follows:
Wherein, λ is regularization parameter, Kxx′Value is kf,To return label yf, it is calculated As mesh
Mark regression coefficient α.
Further, the specific steps of the step S4 are as follows:
According to the template m of current frame imagetDetermine the search box s of current frame imaget, search box stSize be template it is big
1.5 times of small target_sz, search box stCenter be template mtCenter.
Further, specific step is as follows by the step S5:
S5.1: the search box s in traversal current frame imaget, obtain the region to be matched of multiple template size target_sz
piTo get arrive regional ensemble A to be matchedt;
S5.2: based on given cell factory size cell_size, matching area set A is treatedtEach of it is to be matched
Region piSuccessively extract FHOG feature and Haar feature and linear fusion, after linear fusion with Two-dimensional Cosine window cos_window point
Multiply and carry out Fast Fourier Transform (FFT), obtains each region p to be matchediFusion feature zf on corresponding frequency domaini, area to be matched
Domain set AtIn all regions to be matched collection for corresponding to the fusion feature on frequency domain be combined into zf;
S5.3: according to Gaussian kernel correlation function, it is based on each fusion feature zfi, calculate the Gauss cross-correlation core on frequency domain
Matrix obtains each region p to be matchediGauss cross-correlation nuclear matrix kzf on corresponding frequency domaini, wherein Gaussian kernel correlation letter
Several formula are as follows:
For each region p to be matchediFusion feature zf on corresponding frequency domaini, according to Gaussian kernel correlation function formula,
X, x ' are replaced with to object module modle_xf and zf respectivelyi, calculate each region p to be matchediGauss on corresponding frequency domain is mutual
Related nuclear matrix kzfi, regional ensemble A to be matchedtIn the corresponding Gauss cross-correlation nuclear matrix in all regions to be matched set
For kzf;
Scoring function, Gauss cross-correlation nuclear matrix set kzf are responded according to ridge regression, obtains fusion feature core correlation filtering
Set of graphs response is responded, specifically:
Scoring function, which is responded, according to ridge regression obtains each Gauss cross-correlation nuclear matrix kzfiSingle recurrence response,
In, ridge regression responds the formula of scoring function are as follows:
Wherein,Value is the Gauss cross-correlation nuclear matrix kzf in Gauss cross-correlation nuclear matrix set kzfi,For
Goal regression Coefficient m odle_ α,For for a Gauss cross-correlation nuclear matrix kzfiObtained single recurrence response;
By each Gauss cross-correlation nuclear matrix kzfiCorresponding single recurrence response is arranged in matrix according to ranks sequence, and
It carries out Fourier inversion and returns to time domain, retain real part, multilayer core correlation filtering response diagram is obtained, according to Gauss cross-correlation nuclear moment
Battle array set kzf obtains multilayer core correlation filtering response set of graphs response.
Further, specific step is as follows by the step S7:
S7.1: multilayer core correlation filtering responds maximum response max (response) >=T in set of graphs1, rung with maximum
Should value max (response) current frame image coordinate be current frame image target center pos;
S7.2: if current frame image is last frame, end loop, otherwise, updating the template in step S3 is present frame
The weighting of the center of the target of the template and current frame image of image, return step S3 carries out next frame processing after update,
More new formula are as follows:
mt+1=(1-interpfactor)·mt+interpfactor·pmax
Wherein, mt+1For updated template, mtFor current frame image template, pmaxFor maximum response max
(response) corresponding region to be matched, interpfactorFor weight regulatory factor.
Further, specific step is as follows by the step S8:
S8.1: multilayer core correlation filtering responds maximum response max (response) the < T in set of graphs1, then target quilt
It blocks, template does not update, mt+1=mt;
S8.2: according to the dbjective state x in previous frame imaget-1The dbjective state of current frame image is predicted, from prediction
The coordinate of the center of target, i.e. prediction coordinate are taken out in dbjective state, wherein dbjective state includes the center of target
And speed, since template does not update, i.e., adjacent two interframe template variation is little, it is believed that target moves with uniform velocity, and predicts present frame
The specific formula of the dbjective state of image calculates as follows:
xt=Axt-1+B·ut-1+wt-1
Wherein, A is state-transition matrix, and B is the matrix of contact external control parameter, xt-1It is the target in t-1 frame image
State, ut-1It is the acceleration of target in t-1 frame image, makees uniform motion, as 0, wt-1For describing process noise, obey high
This distribution wt-1~N (0, Qt-1),pxAnd pyFor the coordinate of the center of target in t frame image, vx, vyFor t
Speed of the center of target in x-axis and y-axis in frame image, according to uniform motion model, state-transition matrix is set asTherefore predict the formula of the dbjective state of current frame image are as follows:
S8.3: centered on the center for the target for predicting to obtain, size and the consistent region of template and present frame are taken
The target area that image actual match obtains weights the matching result as current frame image, the specific steps are as follows:
S8.3.1: if the center of the target in previous frame image is prediction gained, current frame image actual match
The center of obtained target areaFor the center of the target in previous frame image, target area and template size one
It causes;If obtained by the center of the target in previous frame image and nonanticipating, the target that current frame image actual match obtains
The center in regionIn using position of the maximum response in current frame image as present frame actual match target
Heart position, target area are consistent with template size;
S8.3.2: calculating the current frame image i.e. covariance matrix of the prior estimate of t frame image, specific formula is as follows:
For the posteriori error of t-1 frame image, initial value is specified value, ATFor the transposition of A, Q gives for frame image
Process noise covariance;
S8.3.3: the filtering gain matrix of current frame image are calculatedIts calculation formula is as follows:
Wherein,It is state transition matrix, It is the transposition of state transition matrix, RtIt is to see
Noise covariance is surveyed, is definite value R, (X)-1Indicate that X's is inverse;
S8.3.4: according to the filtering gain matrix of current frame imageWith the dbjective state x of predictiontGenerate posteriority state
Best estimate positionThat is matching result, calculation formula are as follows:
Indicate the center for the target area that current frame image actual match obtains, as measured value,
For measured valueWith prediction coordinateBetween error, use vtIt indicates, vtMeet Gaussian Profile, vt~N (0, Rt);
S8.3.5: if present frame is not last frame, according to filtering gain matrixState transition matrixEstimate with priori
The covariance matrix P of metertThe posteriori error of current frame image is updated, calculation formula is as follows:
S8.4: according to best estimate positionThe center more new formula for updating the target of current frame image is as follows:
Wherein, posxAnd posyFor the center of updated target, pxAnd pyFor best estimate positionCoordinate;
S8.5: centered on the center of matching result, i.e., centered on the center of updated target, working as
It takes the rectangle frame of 3 times of sizes of matching result as search box in prior image frame, regional ensemble to be matched is traversed in search box, is mentioned
It takes in regional ensemble to be matched after the histograms of oriented gradients in each region to be matched and the feature of Ha Er, then carries out subsequent processing and obtain
Set of graphs is responded to correlation filtering, the specific steps are as follows:
S8.5.1: with the center (pos of matching resultx, posy) centered on, determine that search box is long, width is template length and width
3 times ofWith template size wm, lmEntire search box is traversed, regional ensemble to be matched is obtained;
S8.5.2: obtaining the corresponding fusion feature in multiple regions to be matched based on regional ensemble to be matched again, based on fusion
Feature and corresponding object module and goal regression coefficient calculate multilayer core correlation filtering and respond set of graphs response_c;
S8.6: if maximum response max (response_c) >=T in multilayer core correlation filtering response set of graphs2, update
The center of the target of present frame, update mode are as follows:
Wherein, posx, posyFor the center of the target in updated current frame image, px, pyFor in step S8.4
The center of obtained target, wm, lmFor template size, vertx, vertyRelative search frame is left respectively at maximum response
Upper angle (posx-1.5·wm, posy-1.5·lm) movement pixel number;
S8.7: if present frame is last frame, end loop, otherwise, updating the template in step S3 is current frame image
Template mtWith the weighting of the center of the target in current frame image, calculation formula is as follows:
mt+1=(1-interpfactor)·mt+interpfactor·p_cmax
Wherein, mt+1For updated template, mtFor the template of current frame image, p_cmaxFor maximum response max
(response_c) corresponding region to be matched, interpfactorFor weight regulatory factor;
S8.8: if maximum response max (response_c) the < T in multilayer core correlation filtering response set of graphs2If working as
Prior image frame is that last frame then terminates to track, and otherwise, reads next frame image as present frame, goes to step S8.
The present invention compared with the existing technology, its advantages are shown in:
One, core correlation filtering is used in the present invention, is had and the identical low complexity of linearly related filter
Property, lines of code is few, and speed quickly, can be run compared with other track class algorithm with the speed of hundreds of frames per second, better than all
Such as the tracker of Struke or TLD etc, completely competent real-time tracking;
Two, the present invention levies the fusion feature combined using FHOG feature and Lis Hartel, the former describes the edge of image
With change of gradient, and occupy that storage space is small, arithmetic speed is fast, the latter describes edge with low volume data, and the two is combined and held in the palm
The characteristics of fusion feature effectively describes the edge and change of gradient of regional area, more accurately embodies infrared target, subtracts
Few original image information is lost, and tracking accuracy is improved;
Three, the present invention solves lasting tracking under occlusion, and excessive lead of drifting about by introducing Kalman filter
The tracking failure problem of cause increases the anti-power of blocking of tracking, can search for target in bigger search box when tracking failure, realizes
Target detection function again, is greatly improved tracking accuracy.
Detailed description of the invention
Fig. 1 is flow diagram of the invention;
Fig. 2 is the target image of the initial frame center choosing of infrared image sequence in the present invention;
Fig. 3 is the response fusion figure of a certain frame during infrared image sequence tracking in the present invention;
Fig. 4 is 3 frame image sequence original images and tracking effect figure in the present invention.
Specific embodiment
Below in conjunction with the drawings and the specific embodiments, the invention will be further described.
The present invention is based on KCF frames, after reading in image sequence, select target in first frame frame, extract and be based on Ha Er (Haar)
With the feature and linear fusion of histograms of oriented gradients (HOG), first frame template is obtained, the second frame image is pre-processed,
The feature based on Ha Er and histograms of oriented gradients is extracted in search box, filter forecasting target position is closed using nuclear phase, calculates mesh
Similarity between mark and template is greater than threshold value then more new template, reads next frame predicted position, until last frame or confidence level
Less than threshold value, confidence level is less than threshold value, is predicted using Kalman filter target position, and template no longer updates, with prediction
Target trajectory is motion profile, calculates similarity between predicted position and template, thinks that target reappears if more than certain threshold value and open
Otherwise dynamic correlation filtering carries out Kalman filter until last frame.
One kind is anti-to block infrared object tracking method, includes the following steps:
S1: reading infrared image sequence, select target in initial frame image center, obtain center and the size of target,
Using the target in initial frame image as template, the second frame image is obtained as current frame image, by the template of initial frame image
Template as current frame image;Specific step is as follows: reading infrared image sequence, selects target in initial frame image center, remembers
Center and the size for recording target, using the target of initial frame image center choosing as template, wherein the center of template
It is center and the size of target with size;The second frame image is obtained as current frame image, by the mould of initial frame image
Template of the plate as current frame image.
S2: Two-dimensional Cosine window is obtained according to the size of template and cell factory size;Specific step is as follows:
S2.1: according to the size target_sz of template, search box is determined, the size of search box is window_sz=
Target_sz* (1+padding), wherein padding is the ratio for determining search box size and target sizes;
S2.2: according to the big of given cell factory size cell_size, the size target_sz of template and search box
Small window_sz determines that feature returns label yf, then returns label yf based on feature and obtain Two-dimensional Cosine window cos_window;
Specific step is as follows:
S2.2.1: Gauss is calculated according to the size target_sz of template and cell factory cell_size size and returns label
Bandwidth σ, formula is as follows:
In formula, w and h are the width and height of template, and a is spatial bandwidth, proportional to target sizes;
S2.2.2: returning the bandwidth σ of label and the size window_sz of search box according to Gauss, calculate and return label yf,
Calculation formula is as follows:
Wherein, m, n are respectivelyFirst item and Section 2, after y ' is calculated, cyclic shift make return mark
Label peak value moves on to the upper left corner and obtains y, then carries out Fourier transformation, obtains returning label yf;
S2.2.3: it is calculated according to the size for returning label yf using hann function, obtains Two-dimensional Cosine window cos_window;
S3: the feature and linear fusion of template are extracted based on histograms of oriented gradients and Ha Er, to the spy after linear fusion
Sign plus Two-dimensional Cosine window obtain fusion feature, then goal regression coefficient is calculated based on fusion feature, if goal regression coefficient
Gained is calculated for the second frame image, with goal regression coefficient initialization object module and goal regression coefficient, if goal regression system
Number is that last frame image calculates gained, does not deal with, otherwise updates object module and goal regression coefficient;Specific steps are such as
Under:
S3.1: template m is extracted based on HOG and HaartFeature and linear fusion, two dimension is added to the feature after linear fusion
Cosine Window cos_window obtains fusion feature, wherein Haar Ha Er, HOG are histograms of oriented gradients;Specific steps are as follows:
S3.1.1: based on given cell factory size cell_size, the corresponding piotr_toolbox of MATLAB is utilized
Kit extracts template mtFHOG feature, obtain the FHOG feature g of 31 dimensions0, i.e. histograms of oriented gradients feature;
S3.1.2: current frame image is t frame image, calculates the template m of t frame imagetThe integrogram SAT of (x, y), product
The calculation formula of each pixel value in component SAT are as follows:
SAT (x, y)=SAT (x, y-1)+SAT (x-1, y)-SAT (x-1, y-1)+mt(x, y)
Wherein, SAT (x, y-1) indicates current pixel position x, the integrogram pixel value of the top y, and SAT (x-1, y) expression is worked as
The integrogram pixel value on the preceding pixel position left side x, y, SAT (x-1, y-1) indicate current pixel position x, the integrogram in the upper left corner y
Pixel value, the initial boundary of integrogram SAT are SAT (- 1, y)=SAT (x, -1)=SAT (- 1, -1)=0;SAT (- 1, y) is a left side
Border pixel values, SAT (x, -1) are coboundary pixel value, and SAT (- 1, -1) is upper left apex angle pixel value;Initial boundary SAT (- 1,
Y), SAT (x, -1) and SAT (- 1, -1) is for calculating SAT (0, y) and SAT (x, 0);
S3.1.3: integrogram SAT is divided according to cell factory size cell_size, i.e., by any cell factory upper half
Dividing the sum of pixel integration figure is SATA, the sum of cell factory lower half portion pixel integration figure is SATB, cell factory left-half picture
The sum of plain integrogram is SATC, the sum of cell factory right half part pixel integration figure is SATD, corresponding 1 dimension of each cell factory
Vertical direction Haar feature g1 is SATAWith SATBDifference, 1 dimension horizontal direction Haar feature g2 be SATCWith SATDDifference,
1 dimension vertical direction Haar feature g1 of all cell factories and 1 dimension horizontal direction Haar feature g2 is Lis Hartel sign;
S3.1.4: FHOG feature g is tieed up by 310, all 1 dimension vertical direction Haar feature g1, all 1 dimension horizontal directions
Haar feature g2After carrying out linear fusion, with Two-dimensional Cosine window cos_window dot product, one 33 dimension fusion feature g is obtained.
S3.2: goal regression coefficient is obtained according to fusion feature g;Specific steps are as follows:
S3.2.1: fusion feature g is subjected to Fast Fourier Transform (FFT), obtains template mtIn the fusion feature xf of frequency domain,
Calculation template mtIn the formula of the fusion feature xf of frequency domain are as follows:
In formula, g is indicated to template mtThe 33 dimension fusion features extracted,Indicate Fourier transformation,Expression obtains
Frequency domain fusion feature xf;
S3.2.2: according to Gaussian kernel correlation function, the fusion feature xf based on frequency domain calculates the Gauss auto-correlation on frequency domain
Nuclear matrix kf;The formula of Gaussian kernel correlation function:
Wherein, Kxx′The Gaussian kernel correlation matrix of x and x ' is represented, x, x ', which respectively represent calculating Gaussian kernel correlation matrix, to be made
Different characteristic symbol replaces with corresponding feature in practical calculating process, | | x | |2It is characterized the mould of each element in x
Quadratic sum is again the product of two-dimension sizes before matrix x divided by N, N,Form of the representing matrix x in Fourier,It indicates's
Complex conjugate,Indicate inverse Fourier transform, σ is the bandwidth that Gauss returns label, and T indicates transposition;
Using the fusion feature xf of frequency domain, x, the x ' in the formula of Gaussian kernel correlation function are replaced with into xf, calculate frequency
Gauss auto-correlation nuclear matrix kf on domain;
S3.2.3: according to Gauss auto-correlation nuclear matrix kf, goal regression factor alpha, calculation formula are calculated are as follows:
Wherein, λ is regularization parameter, Kxx' value is kf,To return label yf, it is calculated As mesh
Mark regression coefficient α.
S3.3: if current frame image is the second frame image, going to step S3.4, if last frame image, be not processed,
Otherwise step S3.6 is gone to;
S3.4: when target following the second frame image, the fusion feature xf of the frequency domain of current frame image, initialized target are utilized
Model modle_xf, i.e.,
Wherein, t indicates the second frame image,Indicate the fusion feature xf of frequency domain,Indicate object module modle_xf;
S3.5: when target following the second frame image, using the goal regression factor alpha of current frame image, initialized target is returned
Return Coefficient m odle_ α, i.e.,
Wherein, t indicates the second frame image,Indicate goal regression factor alpha,Indicate goal regression Coefficient m odle_ α;
S3.6: when image after target following third frame or third frame, object module is updated by linear interpolation
Modle_xf, i.e.,
Wherein, η is given learning rate, value 0.02,For the object module of current frame image,For former frame
The object module of image obtains updatedAs updated modle_xf;
S3.7: when image after target following third frame or third frame, pass through the goal regression system that linear interpolation updates
Number modle_ α, i.e.,
Wherein, η is given learning rate, value 0.02,For the goal regression coefficient of previous frame image,To work as
The goal regression Coefficient m odle_ α of prior image frame.
S4: the search box of current frame image is determined using the template center of current frame image as central position of search frame;Specifically
Step are as follows:
According to the template m of current frame imagetDetermine the search box s of current frame imaget, search box stSize be template it is big
1.5 times of small target_sz, search box stCenter be template mtCenter.
S5: being based on template size, traversed in the search box in current frame image, obtain regional ensemble to be matched,
The corresponding fusion feature in multiple regions to be matched is obtained based on regional ensemble to be matched again, is based on fusion feature and corresponding target
Model and goal regression coefficient calculate the corresponding multilayer core correlation filtering response diagram in each region to be matched to get multilayer nuclear phase is arrived
Close filter response set of graphs;Specific step is as follows:
S5.1: the search box s in traversal current frame imaget, obtain the region to be matched of multiple template size target_sz
piTo get arrive regional ensemble A to be matchedt;
S5.2: based on given cell factory size cell_size, matching area set A is treatedtEach of it is to be matched
Region piSuccessively extract FHOG feature and Haar feature and linear fusion, after linear fusion with Two-dimensional Cosine window cos_window point
Multiply and carry out Fast Fourier Transform (FFT), obtains each region p to be matchediFusion feature zf on corresponding frequency domaini, area to be matched
Domain set AtIn all regions to be matched collection for corresponding to the fusion feature on frequency domain be combined into zf;
S5.3: according to Gaussian kernel correlation function, it is based on each fusion feature zfi, calculate the Gauss cross-correlation core on frequency domain
Matrix obtains each region p to be matchediGauss cross-correlation nuclear matrix kzf on corresponding frequency domaini, wherein Gaussian kernel correlation letter
Several formula are as follows:
For each region p to be matchediFusion feature zf on corresponding frequency domaini, according to Gaussian kernel correlation function formula,
X, x ' are replaced with to object module modle_xf and zf respectivelyi, calculate each region p to be matchediGauss on corresponding frequency domain is mutual
Related nuclear matrix kzfi, regional ensemble A to be matchedtIn the corresponding Gauss cross-correlation nuclear matrix in all regions to be matched set
For kzf;
Scoring function, Gauss cross-correlation nuclear matrix set kzf are responded according to ridge regression, obtains fusion feature core correlation filtering
Set of graphs response is responded, specifically:
Scoring function, which is responded, according to ridge regression obtains each Gauss cross-correlation nuclear matrix kzfiSingle recurrence response,
In, ridge regression responds the formula of scoring function are as follows:
Wherein,Value is the Gauss cross-correlation nuclear matrix kzf in Gauss cross-correlation nuclear matrix set kzfi,For
Goal regression Coefficient m odle_ α,For for a Gauss cross-correlation nuclear matrix kzfiObtained single recurrence response;
By each Gauss cross-correlation nuclear matrix kzfiCorresponding single recurrence response is arranged in matrix according to ranks sequence, and
It carries out Fourier inversion and returns to time domain, retain real part, multilayer core correlation filtering response diagram is obtained, according to Gauss cross-correlation nuclear moment
Battle array set kzf obtains multilayer core correlation filtering response set of graphs response.
S6: judge whether the maximum response in multilayer core correlation filtering response set of graphs is more than or equal to the first given threshold
Value, if more than going to step S7, otherwise going to S8;
S7: the center of the target of current frame image is calculated with the transverse and longitudinal coordinate of maximum response, if current frame image
It is not last frame, updates the center that the template in step S3 is the template of current frame image and the target of current frame image
Weighting, gone to after update step S3 carry out next frame processing, otherwise terminate to track;Specific step is as follows:
S7.1: multilayer core correlation filtering responds maximum response max (response) >=T in set of graphs1, rung with maximum
Should value max (response) current frame image coordinate be current frame image target center pos;
S7.2: if current frame image is last frame, end loop, otherwise, updating the template in step S3 is present frame
The weighting of the center of the target of the template and current frame image of image, return step S3 carries out next frame processing after update,
More new formula are as follows:
mt+1=(1-interpfactor)·mt+interpfactor·pmax
Wherein, mt+1For updated template, mtFor current frame image template, pmaxFor maximum response max
(response) corresponding region to be matched, interpfactorFor weight regulatory factor.
S8: template does not update, and with the dbjective state in Kalman filter prediction current frame image, obtains prediction coordinate, with
Centered on predicting coordinate, the target area for taking size to obtain with the consistent region of template size with current frame image actual match adds
Act temporarily as the matching result for current frame image, wherein target area is with the center of previous frame target or with peak response
Centered on the transverse and longitudinal coordinate of value, size and the consistent region of template size take matching centered on the center of matching result
As a result 3 times of sizes traverse to obtain regional ensemble to be matched as search box, extract each region to be matched in regional ensemble to be matched
Histograms of oriented gradients and Lis Hartel sign, obtain corresponding fusion feature, then the filter of multilayer core correlation is obtained based on each fusion feature
Wave responds set of graphs (calculation method is identical as the calculation method in step S3);
If multilayer core correlation filtering, which responds maximum response in set of graphs, is more than or equal to given second threshold, with maximum value
Transverse and longitudinal coordinate more fresh target center, if current frame image is not last frame, update step S3 in template be work as
The weighting of the center of the template and present frame target of prior image frame goes to step S3 and carries out next frame processing after update, no
Then terminate to track;
If multilayer core correlation filtering responds in set of graphs, maximum response is lower than given second threshold, if current frame image
Then terminate to track for last frame, otherwise, reads next frame image as present frame, go to step S8.
Specific step is as follows:
S8.1: multilayer core correlation filtering responds maximum response max (response) the < T in set of graphs1, then target quilt
It blocks, template does not update, mt+1=mt;
S8.2: according to the dbjective state x in previous frame imaget-1The dbjective state of current frame image is predicted, from prediction
The coordinate of the center of target, i.e. prediction coordinate are taken out in dbjective state, wherein dbjective state includes the center of target
And speed, since template does not update, i.e., adjacent two interframe template variation is little, it is believed that target moves with uniform velocity, and predicts present frame
The specific formula of the dbjective state of image calculates as follows:
xt=Axt-1+B·ut-1+wt-1
Wherein, A is state-transition matrix, and B is the matrix of contact external control parameter, xt-1It is the target in t-1 frame image
State, ut-1It is the acceleration of target in t-1 frame image, makees uniform motion, as 0, wt-1For describing process noise, obey high
This distribution wt-1~N (0, Qt-1),pxAnd pyFor the coordinate of the center of target in t frame image, vx, vyFor t
Speed of the center of target in x-axis and y-axis in frame image, according to uniform motion model, state-transition matrix is set asTherefore predict the formula of the dbjective state of current frame image are as follows:
S8.3: centered on the center for the target for predicting to obtain, size and the consistent region of template and present frame are taken
The target area that image actual match obtains weights the matching result as current frame image, the specific steps are as follows:
S8.3.1: if the center of the target in previous frame image is prediction gained, current frame image actual match
The center of obtained target areaFor the center of the target in previous frame image, target area and template size one
It causes;If obtained by the center of the target in previous frame image and nonanticipating, the target that current frame image actual match obtains
The center in regionIn using position of the maximum response in current frame image as present frame actual match target
Heart position, target area are consistent with template size;
S8.3.2: calculating the current frame image i.e. covariance matrix of the prior estimate of t frame image, specific formula is as follows:
For the posteriori error of t-1 frame image, initial value is specified value, ATFor the transposition of A, Q gives for frame image
Process noise covariance;
S8.3.3: the filtering gain matrix of current frame image are calculatedIts calculation formula is as follows:
Wherein,It is state transition matrix, It is the transposition of state transition matrix, RtIt is to see
Noise covariance is surveyed, is definite value R, (X)-1Indicate that X's is inverse;
S8.3.4: according to the filtering gain matrix of current frame imageWith the dbjective state x of predictiontGenerate posteriority state
Best estimate positionThat is matching result, calculation formula are as follows:
Indicate the center for the target area that current frame image actual match obtains, as measured value,
For measured valueWith prediction coordinateBetween error, use vtIt indicates, vtMeet Gaussian Profile, vt~N (0, Rt);
S8.3.5: if present frame is not last frame, according to filtering gain matrixState transition matrixEstimate with priori
The covariance matrix P of metertThe posteriori error of current frame image is updated, calculation formula is as follows:
S8.4: according to best estimate positionThe center more new formula for updating the target of current frame image is as follows:
Wherein, posxAnd posyFor the center of updated target, pxAnd pyFor best estimate positionCoordinate;
S8.5: centered on the center of matching result, i.e., centered on the center of updated target, working as
It takes the rectangle frame of 3 times of sizes of matching result as search box in prior image frame, regional ensemble to be matched is traversed in search box, is mentioned
It takes in regional ensemble to be matched after the histograms of oriented gradients in each region to be matched and the feature of Ha Er, then carries out subsequent processing and obtain
Respond set of graphs to correlation filtering (calculation method is identical as the calculation method in step S3), the specific steps are as follows:
S8.5.1: with the center (pos of matching resultx, posy) centered on, determine that search box is long, width is template length and width
3 times ofWith template size wm, lmEntire search box is traversed, regional ensemble to be matched is obtained;
S8.5.2: obtaining the corresponding fusion feature in multiple regions to be matched based on regional ensemble to be matched again, based on fusion
Feature and corresponding object module and goal regression coefficient calculate multilayer core correlation filtering and respond set of graphs response_c;
S8.6: if maximum response max (response_c) >=T in multilayer core correlation filtering response set of graphs2, update
The center of the target of present frame, update mode are as follows:
Wherein, posx, posyFor the center of the target in updated current frame image, px, pyFor in step S8.4
The center of obtained target, wm, lmFor template size, vertx, vertyRelative search frame is left respectively at maximum response
Upper angle (posx-1.5·wm, posy-1.5·lm) movement pixel number;
S8.7: if present frame is last frame, end loop, otherwise, updating the template in step S3 is current frame image
Template mtWith the weighting of the center of the target in current frame image, calculation formula is as follows:
mt+1=(1-interpfactor)·mt+interpfactor·p_cmax
Wherein, mt+1For updated template, mtFor the template of current frame image, p_cmaxFor maximum response max
(response_c) corresponding region to be matched, interpfactorFor weight regulatory factor;
S8.8: if maximum response max (response_c) the < T in multilayer core correlation filtering response set of graphs2If working as
Prior image frame is that last frame then terminates to track, and otherwise, reads next frame image as present frame, goes to step S8.
The above is only the representative embodiment in the numerous concrete application ranges of the present invention, to protection scope of the present invention not structure
At any restrictions.It is all using transformation or equivalence replacement and the technical solution that is formed, all fall within rights protection scope of the present invention it
It is interior.
Claims (10)
1. one kind is anti-to block infrared object tracking method, which comprises the steps of:
S1: reading infrared image sequence, select target in initial frame image center, obtain center and the size of target, will be first
Target in beginning frame image obtains the second frame image as current frame image as template, using the template of initial frame image as
The template of current frame image;
S2: Two-dimensional Cosine window is obtained according to the size of template and cell factory size;
S3: the feature and linear fusion of template are extracted based on histograms of oriented gradients and Ha Er, the feature after linear fusion is added
Two-dimensional Cosine window obtains fusion feature, then goal regression coefficient is calculated based on fusion feature, if goal regression coefficient is the
Two frame images calculate gained, with goal regression coefficient initialization object module and goal regression coefficient, if goal regression coefficient is
Last frame image calculates gained, does not deal with, and otherwise updates object module and goal regression coefficient;
S4: the search box of current frame image is determined using the template center of current frame image as central position of search frame;
S5: it is based on template size, is traversed in the search box in current frame image, obtains regional ensemble to be matched, then base
The corresponding fusion feature in multiple regions to be matched is obtained in regional ensemble to be matched, is based on fusion feature and corresponding object module
With goal regression coefficient, calculates the corresponding multilayer core correlation filtering response diagram in each region to be matched and filtered to get to multilayer core correlation
Wave responds set of graphs;
S6: judging whether the maximum response in multilayer core correlation filtering response set of graphs is more than or equal to given first threshold,
If more than going to step S7, otherwise go to S8;
S7: the center of the target of current frame image is calculated with the transverse and longitudinal coordinate of maximum response, if current frame image is not
Last frame updates adding for the center that the template in step S3 is the template of current frame image and the target of current frame image
Power goes to step S3 and carries out next frame processing, otherwise terminates to track after update;
S8: template does not update, and with the dbjective state in Kalman filter prediction current frame image, prediction coordinate is obtained, with prediction
Centered on coordinate, the target area weighting for taking size and the consistent region of template size and current frame image actual match to obtain is made
For the matching result of current frame image, wherein target area is with the center of previous frame target or with maximum response
Centered on transverse and longitudinal coordinate, size and the consistent region of template size take matching result centered on the center of matching result
3 times of sizes traverse to obtain regional ensemble to be matched as search box, extract the side in each region to be matched in regional ensemble to be matched
It is levied to histogram of gradients and Lis Hartel, obtains corresponding fusion feature, then multilayer core correlation filtering is obtained based on each fusion feature and is rung
Answer set of graphs;
If multilayer core correlation filtering, which responds maximum response in set of graphs, is more than or equal to given second threshold, with the cross of maximum value
The center of ordinate more fresh target, if current frame image is not last frame, updating the template in step S3 is present frame
The weighting of the center of the template and present frame target of image goes to step S3 and carries out next frame processing, otherwise ties after update
Beam tracking;
If multilayer core correlation filtering responds in set of graphs, maximum response is lower than given second threshold, if current frame image is most
A later frame then terminates to track, and otherwise, reads next frame image as present frame, goes to step S8.
2. one kind according to claim 1 is anti-to block infrared object tracking method, which is characterized in that the tool of the step S1
Steps are as follows for body: infrared image sequence is read, target is selected in initial frame image center, records center and the size of target,
Using the target of initial frame image center choosing as template, wherein the center of template and size are the centre bit of target
It sets and size;The second frame image is obtained as current frame image, using the template of initial frame image as the template of current frame image.
3. one kind according to claim 1 is anti-to block infrared object tracking method, which is characterized in that the tool of the step S2
Steps are as follows for body:
S2.1: according to the size target_sz of template, search box is determined, the size of search box is window_sz=target_
Sz* (1+padding), wherein padding is the ratio for determining search box size and target sizes;
S2.2: according to given cell factory size cell_size, the size target_sz of template and the size of search box
Window_sz determines that feature returns label yf, then returns label yf based on feature and obtain Two-dimensional Cosine window cos_window;
Specific step is as follows:
S2.2.1: the band that Gauss returns label is calculated according to the size target_sz of template and cell factory cell_size size
Wide σ, formula are as follows:
In formula, w and h are the width and height of template, and a is spatial bandwidth, proportional to target sizes;
S2.2.2: the bandwidth σ of label and the size window_sz of search box are returned according to Gauss, calculates and returns label yf, is calculated
Formula is as follows:
Wherein, m, n are respectivelyFirst item and Section 2, after y ' is calculated, cyclic shift make return label peak
Value moves on to the upper left corner and obtains y, then carries out Fourier transformation, obtains returning label yf;
S2.2.3: it is calculated according to the size for returning label yf using hann function, obtains Two-dimensional Cosine window cos_window.
4. one kind according to claim 1 is anti-to block infrared object tracking method, which is characterized in that the tool of the step S3
Steps are as follows for body:
S3.1: template m is extracted based on HOG and HaartFeature and linear fusion, Two-dimensional Cosine is added to the feature after linear fusion
Window cos_window obtains fusion feature, wherein Haar Ha Er, HOG are histograms of oriented gradients;
S3.2: goal regression coefficient is obtained according to fusion feature 9;
S3.3: if current frame image is the second frame image, step S3.4 is gone to, if last frame image, is not processed, otherwise
Go to step S3.6;
S3.4: when target following the second frame image, the fusion feature xf of the frequency domain of current frame image, initialized target model are utilized
Modle_xf, i.e.,
Wherein, t indicates the second frame image,Indicate the fusion feature xf of frequency domain,Indicate object module modle_xf;
S3.5: when target following the second frame image, using the goal regression factor alpha of current frame image, initialized target returns system
Number modle_ α, i.e.,
Wherein, t indicates the second frame image,Indicate goal regression factor alpha,Indicate goal regression Coefficient m odle_ α;
S3.6: when image after target following third frame or third frame, pass through linear interpolation and update object module modle_xf.
I.e.
Wherein, η is given learning rate,For the object module of current frame image,For the object module of previous frame image,
It obtains updatedAs updated modle_xf;
S3.7: when image after target following third frame or third frame, pass through the goal regression coefficient that linear interpolation updates
Modle_ α, i.e.,
Wherein, η is given learning rate,For the goal regression coefficient of previous frame image,For the target of current frame image
Regression coefficient modle_ α.
5. blocking infrared object tracking method according to any described one kind of claim 4 is anti-, which is characterized in that the step
3.1 specific steps are as follows:
S3.1.1: based on given cell factory size cell_size, the corresponding piotr_toolbox tool of MATLAB is utilized
Packet extracts template mtFHOG feature, obtain the FHOG feature g of 31 dimensions0, i.e. histograms of oriented gradients feature;
S3.1.2: current frame image is t frame image, calculates the template m of t frame imagetThe integrogram SAT of (x, y), integrogram
The calculation formula of each pixel value in SAT are as follows:
SAT (x, y)=SAT (x, y-1)+SAT (x-1, y)-SAT (x-1, y-1)+mt(x,y)
Wherein, SAT (x, y-1) indicates that current pixel position x, the integrogram pixel value of the top y, SAT (x-1, y) indicate current picture
Plain position x, the integrogram pixel value on the left side y, SAT (x-1, y-1) indicate current pixel position x, the integrogram pixel in the upper left corner y
Value, the initial boundary of integrogram SAT are SAT (- 1, y)=SAT (x, -1)=SAT (- 1, -1)=0;SAT (- 1, y) is left margin
Pixel value, SAT (x, -1) are coboundary pixel value, and SAT (- 1, -1) is upper left apex angle pixel value;Initial boundary SAT (- 1, y),
SAT (x, -1) and SAT (- 1, -1) is for calculating SAT (0, y) and SAT (x, 0);
S3.1.3: integrogram SAT is divided according to cell factory size cell_size, i.e., by any cell factory top half picture
The sum of plain integrogram is SATA, the sum of cell factory lower half portion pixel integration figure is SATB, cell factory left-half pixel product
The sum of component is SATC, the sum of cell factory right half part pixel integration figure is SATD, each cell factory corresponding 1 is tieed up vertical
Direction Haar feature g1 is SATAWith SATBDifference, 1 dimension horizontal direction Haar feature g2 be SATCWith SATDDifference, own
1 dimension vertical direction Haar feature g1 of cell factory and 1 dimension horizontal direction Haar feature g2 is Lis Hartel sign;
S3.1.4: FHOG feature g is tieed up by 310, all 1 dimension vertical direction Haar feature g1, all 1 dimension horizontal direction Haar
Feature g2After carrying out linear fusion, with Two-dimensional Cosine window cos_window dot product, one 33 dimension fusion feature g is obtained.
6. blocking infrared object tracking method according to any described one kind of claim 4 is anti-, which is characterized in that the step
3.2 specific steps are as follows:
S3.2.1: fusion feature g is subjected to Fast Fourier Transform (FFT), obtains template mtIn the fusion feature xf of frequency domain,
Calculation template mtIn the formula of the fusion feature xf of frequency domain are as follows:
In formula, g is indicated to template mtThe 33 dimension fusion features extracted,Indicate Fourier transformation,Indicate obtained frequency
The fusion feature xf in domain;
S3.2.2: according to Gaussian kernel correlation function, the fusion feature xf based on frequency domain calculates the Gauss auto-correlation nuclear moment on frequency domain
Battle array kf;The formula of Gaussian kernel correlation function:
Wherein, Kxx′The Gaussian kernel correlation matrix of x and x ' is represented, x, x ', which are respectively represented, to be calculated used in Gaussian kernel correlation matrix
Different characteristic symbol replaces with corresponding feature in practical calculating process, | | x | |2It is characterized square of the mould of each element in x
It is again the product of two-dimension sizes before matrix x divided by N, N,Form of the representing matrix x in Fourier,It indicatesIt is multiple altogether
Yoke,Indicate inverse Fourier transform, σ is the bandwidth that Gauss returns label, and T indicates transposition;
Using the fusion feature xf of frequency domain, x, the x ' in the formula of Gaussian kernel correlation function are replaced with into xf, calculated on frequency domain
Gauss auto-correlation nuclear matrix kf;
S3.2.3: according to Gauss auto-correlation nuclear matrix kf, goal regression factor alpha, calculation formula are calculated are as follows:
Wherein, enter for regularization parameter, KxxValue is kf,To return label yf, it is calculatedAs target
Regression coefficient α.
7. -6 any described one kind are anti-according to claim 1 blocks infrared object tracking method, which is characterized in that the step
The specific steps of S4 are as follows:
According to the template m of current frame imagetDetermine the search box s of current frame imaget, search box stSize be template size
1.5 times of target_sz, search box stCenter be template mtCenter.
8. one kind according to claim 7 is anti-to block infrared object tracking method, which is characterized in that the tool of the step S5
Steps are as follows for body:
S5.1: the search box s in traversal current frame imaget, obtain the region p to be matched of multiple template size target_szi, i.e.,
Obtain regional ensemble A to be matchedt;
S5.2: based on given cell factory size cell_size, matching area set A is treatedtEach of region to be matched
piSuccessively extract FHOG feature and Haar feature and linear fusion, after linear fusion simultaneously with Two-dimensional Cosine window cos_window dot product
Fast Fourier Transform (FFT) is carried out, each region p to be matched is obtainediFusion feature zf on corresponding frequency domaini, region collection to be matched
Close AtIn all regions to be matched collection for corresponding to the fusion feature on frequency domain be combined into zf;
S5.3: according to Gaussian kernel correlation function, it is based on each fusion feature zfi, the Gauss cross-correlation nuclear matrix on frequency domain is calculated,
Obtain each region p to be matchediGauss cross-correlation nuclear matrix kzf on corresponding frequency domaini, wherein the public affairs of Gaussian kernel correlation function
Formula are as follows:
For each region p to be matchediFusion feature zf on corresponding frequency domaini, according to Gaussian kernel correlation function formula, by x,
X ' replaces with object module modle_xf and zf respectivelyi, calculate each region p to be matchediGauss cross-correlation on corresponding frequency domain
Nuclear matrix kzfi, regional ensemble A to be matchedtIn the collection of the corresponding Gauss cross-correlation nuclear matrix in all regions to be matched be combined into
kzf;
Scoring function, Gauss cross-correlation nuclear matrix set kzf are responded according to ridge regression, fusion feature nuclear phase is obtained and closes filter response
Set of graphs response, specifically:
Scoring function, which is responded, according to ridge regression obtains each Gauss cross-correlation nuclear matrix kzfiSingle recurrence response, wherein ridge return
Return the formula of response scoring function are as follows:
Wherein,Value is the Gauss cross-correlation nuclear matrix kzf in Gauss cross-correlation nuclear matrix set kzfi,For target
Regression coefficient modle_ α,For for a Gauss cross-correlation nuclear matrix kzfiObtained single recurrence response;
By each Gauss cross-correlation nuclear matrix kzfiCorresponding single recurrence response is arranged in matrix according to ranks sequence, and carries out
Fourier inversion returns to time domain, retains real part, multilayer core correlation filtering response diagram is obtained, according to Gauss cross-correlation nuclear matrix collection
It closes kzf and obtains multilayer core correlation filtering response set of graphs response.
9. one kind according to claim 8 is anti-to block infrared object tracking method, which is characterized in that the tool of the step S7
Steps are as follows for body:
S7.1: multilayer core correlation filtering responds maximum response max (response) >=T in set of graphs1, with maximum response
Center pos of the max (response) in the target that the coordinate of current frame image is current frame image;
S7.2: if current frame image is last frame, end loop, otherwise, updating the template in step S3 is current frame image
Template and current frame image target center weighting, after update return step S3 carry out next frame processing, update
Formula are as follows:
mt+1=(1-interpfactor)·mt+interpfactor·pmax
Wherein, mt+1For updated template, mtFor current frame image template, pmaxIt is corresponding for maximum response max (response)
Region to be matched, interpfactorFor weight regulatory factor.
10. one kind according to claim 9 is anti-to block infrared object tracking method, which is characterized in that the step S8's
Specific step is as follows:
S8.1: multilayer core correlation filtering responds maximum response max (response) the < T in set of graphs1, then target is blocked,
Template does not update, mt+1=mt;
S8.2: according to the dbjective state x in previous frame imaget-1The dbjective state of current frame image is predicted, from the target of prediction
The coordinate of the center of target, i.e. prediction coordinate are taken out in state, wherein dbjective state includes center and the speed of target
Degree, since template does not update, i.e., adjacent two interframe template variation is little, it is believed that target moves with uniform velocity, and predicts current frame image
Dbjective state specific formula calculate it is as follows:
xt=Axt-1+B·ut-1+wt-1
Wherein, A is state-transition matrix, and B is the matrix of contact external control parameter, xt-1It is the target-like in t-1 frame image
State, ut-1It is the acceleration of target in t-1 frame image, makees uniform motion, as 0, wt-1For describing process noise, Gauss is obeyed
It is distributed wt-1~N (0, Qt-1),pxAnd pyFor the coordinate of the center of target in t frame image, vx, vyFor t
Speed of the center of target in x-axis and y-axis in frame image, according to uniform motion model, state-transition matrix is set asTherefore predict the formula of the dbjective state of current frame image are as follows:
S8.3: centered on the center for the target for predicting to obtain, size and the consistent region of template and current frame image are taken
The target area that actual match obtains weights the matching result as current frame image, the specific steps are as follows:
S8.3.1: if the center of the target in previous frame image is prediction gained, current frame image actual match is obtained
Target area centerFor the center of the target in previous frame image, target area is consistent with template size;
If obtained by the center of the target in previous frame image and nonanticipating, the target area that current frame image actual match obtains
CenterFor the centre bit using position of the maximum response in current frame image as present frame actual match target
It sets, target area is consistent with template size;
S8.3.2: calculating the current frame image i.e. covariance matrix of the prior estimate of t frame image, specific formula is as follows:
For the posteriori error of t-1 frame image, initial value is specified value, ATFor the transposition of A, Q is the mistake that frame image gives
Journey noise covariance;
S8.3.3: the filtering gain matrix of current frame image are calculatedIts calculation formula is as follows:
Wherein,It is state transition matrix, It is the transposition of state transition matrix, RtIt is that observation is made an uproar
Sound covariance is definite value R, (X)-1Indicate that X's is inverse;
S8.3.4: according to the filtering gain matrix of current frame imageWith the dbjective state x of predictiontGenerate the best of posteriority state
Estimated locationThat is matching result, calculation formula are as follows:
Indicate the center for the target area that current frame image actual match obtains, as measured value,To survey
MagnitudeWith prediction coordinateBetween error, use vtIt indicates, vtMeet Gaussian Profile, vt~N (0, Rt);
S8.3.5: if present frame is not last frame, according to filtering gain matrixState transition matrixWith prior estimate
Covariance matrix PtThe posteriori error of current frame image is updated, calculation formula is as follows:
S8.4: according to best estimate positionThe center more new formula for updating the target of current frame image is as follows:
Wherein, posxAnd posyFor the center of updated target, pxAnd pyFor best estimate positionCoordinate;
S8.5: centered on the center of matching result, i.e., centered on the center of updated target, in present frame
Take the rectangle frame of 3 times of sizes of matching result as search box in image, traverse regional ensemble to be matched in search box, extract to
In matching area set after the histograms of oriented gradients in each region to be matched and the feature of Ha Er, then carries out subsequent processing and obtain phase
Close filter response set of graphs, the specific steps are as follows:
S8.5.1: with the center (pos of matching resultx, posy) centered on, determine that search box is long, width is the 3 of template length and width
TimesWith template size wm,lmEntire search box is traversed, regional ensemble to be matched is obtained;
S8.5.2: obtaining the corresponding fusion feature in multiple regions to be matched based on regional ensemble to be matched again, is based on fusion feature
And corresponding object module and goal regression coefficient, calculating multilayer core correlation filtering respond set of graphs response_c;
S8.6: if maximum response max (response_c) >=T in multilayer core correlation filtering response set of graphs2, update current
The center of the target of frame, update mode are as follows:
Wherein, posx, posyFor the center of the target in updated current frame image, px, pyTo be obtained in step S8.4
Target center, wm, lmFor template size, vertx, vertyThe relative search frame upper left corner respectively at maximum response
(posx-1.5·wm,posy-1.5·lm) movement pixel number;
S8.7: if present frame is last frame, otherwise end loop updates the mould that the template in step S3 is current frame image
Plate mtWith the weighting of the center of the target in current frame image, calculation formula is as follows:
mt+1=(1-interpfactor)·mt+interpfactor·p_cmax
Wherein, mt+1For updated template, mtFor the template of current frame image, p_cmaxFor maximum response max (response_
C) corresponding region to be matched, interpfactorFor weight regulatory factor;
S8.8: if maximum response max (response_c) the < T in multilayer core correlation filtering response set of graphs2, if present frame
Image is that last frame then terminates to track, and otherwise, reads next frame image as present frame, goes to step S8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910547576.2A CN110276785B (en) | 2019-06-24 | 2019-06-24 | Anti-shielding infrared target tracking method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910547576.2A CN110276785B (en) | 2019-06-24 | 2019-06-24 | Anti-shielding infrared target tracking method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110276785A true CN110276785A (en) | 2019-09-24 |
CN110276785B CN110276785B (en) | 2023-03-31 |
Family
ID=67961532
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910547576.2A Active CN110276785B (en) | 2019-06-24 | 2019-06-24 | Anti-shielding infrared target tracking method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110276785B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110728697A (en) * | 2019-09-30 | 2020-01-24 | 华中光电技术研究所(中国船舶重工集团有限公司第七一七研究所) | Infrared dim target detection tracking method based on convolutional neural network |
CN110796687A (en) * | 2019-10-30 | 2020-02-14 | 电子科技大学 | Sky background infrared imaging multi-target tracking method |
CN111563919A (en) * | 2020-04-03 | 2020-08-21 | 深圳市优必选科技股份有限公司 | Target tracking method and device, computer readable storage medium and robot |
CN111721420A (en) * | 2020-04-27 | 2020-09-29 | 浙江智物慧云技术有限公司 | Semi-supervised artificial intelligence human body detection embedded algorithm based on infrared array time sequence |
CN112991394A (en) * | 2021-04-16 | 2021-06-18 | 北京京航计算通讯研究所 | KCF target tracking method based on cubic spline interpolation and Markov chain |
CN113076949A (en) * | 2021-03-31 | 2021-07-06 | 成都唐源电气股份有限公司 | Method and system for quickly positioning parts of contact net |
CN114066934A (en) * | 2021-10-21 | 2022-02-18 | 华南理工大学 | Anti-shielding cell tracking method facing targeted micro-operation |
CN115631216A (en) * | 2022-12-21 | 2023-01-20 | 中航金城无人系统有限公司 | Holder target tracking system and method based on multi-feature filter fusion |
CN115631359A (en) * | 2022-11-17 | 2023-01-20 | 诡谷子人工智能科技(深圳)有限公司 | Image data processing method and device for machine vision recognition |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120189162A1 (en) * | 2009-07-31 | 2012-07-26 | Fujitsu Limited | Mobile unit position detecting apparatus and mobile unit position detecting method |
CN103700112A (en) * | 2012-09-27 | 2014-04-02 | 中国航天科工集团第二研究院二O七所 | Sheltered target tracking method based on mixed predicting strategy |
CN105981075A (en) * | 2013-12-13 | 2016-09-28 | 英特尔公司 | Efficient facial landmark tracking using online shape regression method |
CN106887012A (en) * | 2017-04-11 | 2017-06-23 | 山东大学 | A kind of quick self-adapted multiscale target tracking based on circular matrix |
CN108550161A (en) * | 2018-03-20 | 2018-09-18 | 南京邮电大学 | A kind of dimension self-adaption core correlation filtering fast-moving target tracking method |
CN108665481A (en) * | 2018-03-27 | 2018-10-16 | 西安电子科技大学 | Multilayer depth characteristic fusion it is adaptive resist block infrared object tracking method |
-
2019
- 2019-06-24 CN CN201910547576.2A patent/CN110276785B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120189162A1 (en) * | 2009-07-31 | 2012-07-26 | Fujitsu Limited | Mobile unit position detecting apparatus and mobile unit position detecting method |
CN103700112A (en) * | 2012-09-27 | 2014-04-02 | 中国航天科工集团第二研究院二O七所 | Sheltered target tracking method based on mixed predicting strategy |
CN105981075A (en) * | 2013-12-13 | 2016-09-28 | 英特尔公司 | Efficient facial landmark tracking using online shape regression method |
CN106887012A (en) * | 2017-04-11 | 2017-06-23 | 山东大学 | A kind of quick self-adapted multiscale target tracking based on circular matrix |
CN108550161A (en) * | 2018-03-20 | 2018-09-18 | 南京邮电大学 | A kind of dimension self-adaption core correlation filtering fast-moving target tracking method |
CN108665481A (en) * | 2018-03-27 | 2018-10-16 | 西安电子科技大学 | Multilayer depth characteristic fusion it is adaptive resist block infrared object tracking method |
Non-Patent Citations (6)
Title |
---|
HAICHAO, ZHENG: "a novel method for quantifying target tracking difficulty of the infrared of the image sequence", 《INFRARED PHYSICS & TECHNOLOGY》 * |
LI,GUN: "target tracking based on biological-like vision identity via improved sparse representation and particle filtering", 《GOGNITIVE COMPUTATION》 * |
刘兆雄: "复杂背景中的红外人体目标检测技术研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
包晓安等: "基于KCF和SIFT特征的抗遮挡目标跟踪算法", 《计算机测量与控制》 * |
姜丹: "基于视频监控的目标检测与跟踪算法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
许俊平: "强背景下模糊闪烁成像目标探测与跟踪", 《强激光与粒子束》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110728697A (en) * | 2019-09-30 | 2020-01-24 | 华中光电技术研究所(中国船舶重工集团有限公司第七一七研究所) | Infrared dim target detection tracking method based on convolutional neural network |
CN110796687A (en) * | 2019-10-30 | 2020-02-14 | 电子科技大学 | Sky background infrared imaging multi-target tracking method |
CN111563919A (en) * | 2020-04-03 | 2020-08-21 | 深圳市优必选科技股份有限公司 | Target tracking method and device, computer readable storage medium and robot |
CN111563919B (en) * | 2020-04-03 | 2023-12-29 | 深圳市优必选科技股份有限公司 | Target tracking method, device, computer readable storage medium and robot |
CN111721420A (en) * | 2020-04-27 | 2020-09-29 | 浙江智物慧云技术有限公司 | Semi-supervised artificial intelligence human body detection embedded algorithm based on infrared array time sequence |
CN111721420B (en) * | 2020-04-27 | 2021-06-29 | 浙江智物慧云技术有限公司 | Semi-supervised artificial intelligence human body detection embedded algorithm based on infrared array time sequence |
CN113076949A (en) * | 2021-03-31 | 2021-07-06 | 成都唐源电气股份有限公司 | Method and system for quickly positioning parts of contact net |
CN112991394A (en) * | 2021-04-16 | 2021-06-18 | 北京京航计算通讯研究所 | KCF target tracking method based on cubic spline interpolation and Markov chain |
CN112991394B (en) * | 2021-04-16 | 2024-01-19 | 北京京航计算通讯研究所 | KCF target tracking method based on cubic spline interpolation and Markov chain |
CN114066934A (en) * | 2021-10-21 | 2022-02-18 | 华南理工大学 | Anti-shielding cell tracking method facing targeted micro-operation |
CN114066934B (en) * | 2021-10-21 | 2024-03-22 | 华南理工大学 | Anti-occlusion cell tracking method for targeting micro-operation |
CN115631359A (en) * | 2022-11-17 | 2023-01-20 | 诡谷子人工智能科技(深圳)有限公司 | Image data processing method and device for machine vision recognition |
CN115631216A (en) * | 2022-12-21 | 2023-01-20 | 中航金城无人系统有限公司 | Holder target tracking system and method based on multi-feature filter fusion |
Also Published As
Publication number | Publication date |
---|---|
CN110276785B (en) | 2023-03-31 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110276785A (en) | One kind is anti-to block infrared object tracking method | |
CN106875424B (en) | A kind of urban environment driving vehicle Activity recognition method based on machine vision | |
Sidla et al. | Pedestrian detection and tracking for counting applications in crowded situations | |
CN104899590B (en) | A kind of unmanned plane sensation target follower method and system | |
Li et al. | A generic approach to simultaneous tracking and verification in video | |
CN109949340A (en) | Target scale adaptive tracking method based on OpenCV | |
CN108010067A (en) | A kind of visual target tracking method based on combination determination strategy | |
CN103886325B (en) | Cyclic matrix video tracking method with partition | |
CN106204638A (en) | A kind of based on dimension self-adaption with the method for tracking target of taking photo by plane blocking process | |
CN101324956A (en) | Method for tracking anti-shield movement object based on average value wander | |
CN111311647B (en) | Global-local and Kalman filtering-based target tracking method and device | |
CN108198201A (en) | A kind of multi-object tracking method, terminal device and storage medium | |
CN108776975A (en) | Visual tracking method based on semi-supervised feature and filter joint learning | |
CN112488057A (en) | Single-camera multi-target tracking method utilizing human head point positioning and joint point information | |
CN105279769A (en) | Hierarchical particle filtering tracking method combined with multiple features | |
CN102156995A (en) | Video movement foreground dividing method in moving camera | |
CN109977971A (en) | Dimension self-adaption Target Tracking System based on mean shift Yu core correlation filtering | |
CN101408983A (en) | Multi-object tracking method based on particle filtering and movable contour model | |
CN104574401A (en) | Image registration method based on parallel line matching | |
CN109767454A (en) | Based on Space Time-frequency conspicuousness unmanned plane video moving object detection method | |
CN110827262B (en) | Weak and small target detection method based on continuous limited frame infrared image | |
CN109063549A (en) | High-resolution based on deep neural network is taken photo by plane video moving object detection method | |
Li et al. | Simultaneous tracking and verification via sequential posterior estimation | |
CN109448023A (en) | A kind of satellite video Small object method for real time tracking of combination space confidence map and track estimation | |
CN110006444A (en) | A kind of anti-interference visual odometry construction method based on optimization mixed Gauss model |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |