CN108549905A - A kind of accurate method for tracking target under serious circumstance of occlusion - Google Patents

A kind of accurate method for tracking target under serious circumstance of occlusion Download PDF

Info

Publication number
CN108549905A
CN108549905A CN201810310348.9A CN201810310348A CN108549905A CN 108549905 A CN108549905 A CN 108549905A CN 201810310348 A CN201810310348 A CN 201810310348A CN 108549905 A CN108549905 A CN 108549905A
Authority
CN
China
Prior art keywords
sparse
target
formula
llc
histogram
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
CN201810310348.9A
Other languages
Chinese (zh)
Inventor
戴林旱
聂桂芝
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.)
SHANGHAI FERLY DIGITAL TECHNOLOGIES Co Ltd
Original Assignee
SHANGHAI FERLY DIGITAL TECHNOLOGIES Co Ltd
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 SHANGHAI FERLY DIGITAL TECHNOLOGIES Co Ltd filed Critical SHANGHAI FERLY DIGITAL TECHNOLOGIES Co Ltd
Priority to CN201810310348.9A priority Critical patent/CN108549905A/en
Publication of CN108549905A publication Critical patent/CN108549905A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/513—Sparse representations

Abstract

The present invention proposes a kind of Gauss sparse expression coordination model, for the target following under seriously blocking.The place of method core the most is to carry out rarefaction representation to candidate samples in such a way that LLC codings are combined using sparse coding.But also with highly accurately reconstructed error while method easily obtains sparse solution, and prior probability is added in model, the sample around next frame target is set to be more easy to as final tracking result, by carrying out many experiments comparison with other methods, method of the invention can preferably track target under serious circumstance of occlusion.

Description

A kind of accurate method for tracking target under serious circumstance of occlusion
Technical field
The invention belongs to pattern-recognitions and machine learning field, can be used for the target following of object, in particular for tracking The case where object is seriously blocked.
Background technology
There are many kinds of the mode classifications of method for tracking target, and in chronological sequence sequence, is segmented into last century and this century The track algorithm just occurred, as Cam Shift, Mean Shift, Kalman Filtering, Optical Flow and , there are many non-deep learning track algorithms in Particle Filtering etc. in last decade, such as KCF, SCM, TLD, Struck etc..Until one of the founder of deep learning in 2012 Hinton has pushed the development of deep learning, occurred in recent years Many deep learning target tracking algorisms, such as MDnet, CNT.The problem of different types of method for tracking target exposure differs Sample.
1. wherein to carry out the tracking of pre-training in advance it is easy to appear future position shift phenomenon, and such Method, which has training dataset, to be relied on；And off-line tracking method can do track path global optimization, it can be in test sequence It is scanned forward or backward on row, but application scenario is relatively fewer, is online main research side into the method for line trace therefore To.
2. using the method performance that generation model is combined with discriminative model more outstanding in a practical situation, mainly Because generating model following method is easy error in the case where background is complex, disturbing factor is more, it is difficult to correctly judge The moving direction of target.
3. some deep learning target tracking algorisms are the further feature that target study is extracted target.When target is serious When blocking, the primitive character partial disappearance of target, these depth characteristics cannot match, therefore tracker can be searched in wide scope Middle tracking target, this easilys lead to the loss of target or with larger central point mistake.
4. Fig. 8 be target following research field it is most common be also most intractable occlusion issue, when tracking target be blocked Afterwards, target primitive character becomes imperfect, and many trackers are difficult to capture target at present, and after target is with losing, track algorithm It is difficult to give target for change, fail so as to cause tracking.
Invention content
The present invention proposes one kind under serious circumstance of occlusion, using sparse distinction grader (SDC) and sparse generation The Gauss conjunctive model (GSCM) of model (SGM) carries out target following, can be applied to unmanned plane, service robot,
The first-class "smart" products of intelligent camera.
In SDC models, candidate samples are weighted using the priori of Gaussian Profile, according to the variance of previous frame target With the weight of mean prediction present frame candidate samples.Meanwhile using sparse coding and LLC points simultaneously in SDC and SGM models Not Ji Suan candidate samples confidence level and sample and template similitude, and obtain two coefficients are combined.Finally, with power Weight, confidence level, similarity carry out the sample of decision maximum likelihood.
Mainly sparse coding technology and LLC coding techniques are used in combination by the present invention, and entire target following frame is divided into SDC (sparsity-based discriminative classifier, the identification and classification device based on sparse expression) SGM moulds Type (sparsity-based generative model, sparse generation model).
In SDC models, the feature of identification is extracted according to following formula (1)：
Herein, NpRefer to positive sample, NnRefer to negative sample, simultaneouslyIt is by NpAnd NnComposition.Fig. 3 is shown Positive and negative sample form, template picture size are 32 × 32.
K be feature selecting before characteristic dimension, vector elementIndicate any one template in training set A In attribute.Sparse features vector s, its non-zero element are select with distinctive feature from K luv spaces.
It is the regular terms of formula on the right of formula (2), uses l here1- norm enhances the generalized ability of formula and is easier Obtain sparse solution.Original feature space is projected to the feature space of selection by projection matrix S, S is removed by diagonal matrix S ' All zero row are constituted, and the element in S ' is determined by following formula.Therefore, the candidate target of training template and projector space is A ' =SA and x '=Sx.
In SDC module, target can by positive masterplate linear combination more preferably indicate and background can by negative norm plate extension more It is good to indicate.Given candidate target can be indicated by training template and the calculated factor alpha of following formula.
Therefore, the trust value H of structure candidate target xcPass through following formula (4)：
Hc=exp (- (εf-εb)/σ) (4)
HereIt is candidate samples x and foreground template collection A+Reconstructed error.Similarly, It is candidate samples x and background template collection A-Reconstructed error, and α is related sparse coefficient vector.
M block diagram pictures are obtained with the sliding window of overlapping on normalized image, y is converted into per block diagram picturei∈RG×1To Amount, G indicate the size of image block.The sparse coefficient vector β of each patch is calculated by formula (5).
Dictionary D ∈ R hereinG×JIt is the most representative model of target, it is generated by k-means cluster centres.Fig. 4 Illustrate the illustration of this method.In Fig. 4, dictionary D ∈ RG×JIt is a according to J from first frame image by k-mean algorithms Cluster centre generates M fritter.It is stitched together the acquired β sparse coefficient vectors of each image block to form histogram.
ρ=[β1,β2,β3,...,βm]T(6)
Here ρ ∈ R(J×M)×1It is the histogram of each candidate target, such as Fig. 4.Means clustering method generates in method Histogram be effective, but this method may lose the spatial information of each image block.Modified reconstruct histogram Exclude the image block blocked.The larger image block of reconstructed error is counted as blocking, and 0 is set to fixed sparse coefficient vector, Here calculating passes through formula (7):
ρ=[β1,β2,β3,...,βm]T (7)
Here the dot product between ⊙ representing matrixes element, each element in o are the indicator of corresponding shielded image block, meter It calculates and passes through formula (8):
In formula (8),It is image fritter yiReconstructed error, and yiIt is that candidate samples pass through sliding window Mouth scanning is formed, ε0It is that preset threshold value is used to judge whether patch is blocked.Then, then with histogram intersection letter Number calculates the similitude between goal histogram and THE TEMPLATE HYSTOGRAM, and calculation formula is
THE TEMPLATE HYSTOGRAM ψjIt is calculated by formula (5,6), since template only uses in the first frame of image sequence, So template only calculates once in each image sequence.
Utilize the acquired trust value H in SDC modelscGoal histogram and THE TEMPLATE HYSTOGRAM are acquired in SGM models Between similitude LcIt is combined, is calculated as formula (10)
This combination contributes to the robustness of track algorithm.The appearance of target constantly changes during tracking Become, therefore every five frame is updated the image-region negative norm plate of current tracking result in SDC models, and entire positive template is protected Hold it is constant, this is because positive sample template generation target a small range variation and it is little.Due in this way, SDC models are with adaptive and distinction.And dictionary is fixed in whole image sequence calculating process, in order to catch New appearance and restore target in blocking, masterplate histogram is updated to the update of SGM modules, is calculated as formula (11)
New histogram ψnBy according to first frame histogram ψfThe histogram ψ finally storedlIt is constituted, μ is for distributing formula The weight size on the left side and the right is used for updated histogram to track the target in next frame image.
In cooperation model, angle value H is trustedcThink to be to confer to the higher weight of positive sample and punish other candidate samples This, distinguishes target from background, so it must assure that positive and negative masterplate is correct and different.Trust angle value HcPass through Domain of walker very little such as Fig. 5 of calculation formula (1~4), therefore selected in most suitable positive sample and do not have very strong differentiation Ability.
For example, the value of most numerical example confidence level is relatively sparse is distributed between 0.95 to 1.15, and only one Fraction sample has very high confidence level (difference between them is less than 0.01).Therefore, the sample of these similitudes is simultaneously There is no any apparent difference.
SDC models and SGM models form a coordination model, and the robustness of target following is improved by model.However, The candidate samples of current algorithm are equally distributed, and whether coordination model can add priori, then in conjunction with current association The robustness of tracking is further improved as model.It is understood that when target is seriously stopped, some external appearance characteristics of target disappear It loses, next frame tracking displacement of targets is little, is most possibly distributed in around current tracking sample.Therefore, sample is being generated at random When, can be the weight of target according to the candidate samples of the position prediction present frame of target, sample distribution is that former frame generates.
The Bayesian filter frame p (x that tracking target is determined by prior probabilityt|xt-1) determined.Here, xtIt is that target exists The state of time t moment.Allow xt=[lx,ly,θ,s,α,Φ]T, l herex,ly, θ, s, α, Φ refer respectively to x coordinate, y-coordinate, Central point, rotation angle, scale, length-width ratio are crooked.Assuming that this 6 affine parameters independently of each other and Gaussian distributed, producing Using the method for particle filter when raw candidate samples.Sample can be by mean value, variance σ2Gaussian function described：
HereinRefer to state of i-th of affine parameter at the t frame moment, allowsIndicate that j-th of sample exists Weight on i-th of affine parameter is then to the weight calculation formula of each candidate samples
Here n refers to the quantity of affine parameter, while candidate samples collectionIt is that height is used by formula (12) This function generates, while obtaining the weight of each candidate samples using formula (13), also means that the target of previous frame tracking The status candidate samples that surrounding is distributed in the current frame will obtain more weights, on the contrary, from previous frame sample position Sample farther out obtains smaller weight.By this priori and consonance models coupling before, pass through formula：
Here, pcIt is the likelihood probability value of each sample obtained by above-mentioned formula (10), is finally chosen at t MomentFor the sample of maximum likelihood value,In consider Prior ProbabilityModel can be further increased in this way Robustness.The block schematic illustration of the method for the present invention is as shown in Figure 2.
Figure it is seen that predicting 5 affine parameters of target in t frames by Gauss weighting, each waited in t+1 frames The weight of sampling sheet is w.Then, these candidate samples will be by SDC and SGD resume modules.In SDC and SGD models, make herein The reconstructed error between each candidate samples and template is constructed respectively with SC and LLC methods.By SDC models, can obtain every Confidence level h between a sample and template, the similarity between each sample and template are obtained by SGD models.Finally, t+1 frames Tracking result be made of this three parts.
Compared with SDC module, in SGM modules, by the way that all candidate samples are divided into mesh by being overlapped sliding window The local feature of template rarefaction representation is marked, histogram is formed.Fig. 6 shows the distribution situation of Sample Similarity, these samples point Cloth relative discrete, and not in some region integrated distribution.The higher sample size of similarity is higher less than confidence level shown in Fig. 5 Sample size, be more advantageous to differentiation foreground and background in this way.Therefore, SGM modules have stronger differentiation similar than SDC module The ability of sample.As shown in fig. 6, distribution of similarity is discrete in figure, compared with SDC module, it has stronger differentiation Ability.
In addition, either carry out feature selecting by equation (1), or with the coefficient between trained template and candidate samples It indicates to use formula (3), or it is all to use sparse coding to calculate the histogram of candidate samples and dictionary similitude by formula (5) Solve sparse expression coefficient.This method is readily available sparse solution, while l1Norm regularization is a kind of Embedded feature Selection method, it helps to pick out with differentiating characteristics.However compared to LLC methods, there is no high for sparse coding (SC) Spend accurately reconstructed error.Therefore, the method for the present invention proposes LLC algorithms and applies in target error reconstruct, and the present invention passes through Combination S C and LLC method, in SDC models, feature selecting is not only to compile based on formula (1) while also based on local limit Code LLC methods, and LLC codings use following formula：
Herein,Refer to the dot product between element, it is expressly noted that
Here, dist (xi, B) and=[dist (xi,b1),…,dist(xi,bM)]T, while dist (xi,bi) it is xiWith biIt Between Euclidean distance.σ is the rate of decay for adjusting local adapter weight.
Code coefficient between training template and candidate template can be calculated by following formula：
Then pass through formula (4), candidate samples and target LLC sparse coding trust values H equally can be obtainedllc.Pass through public affairs again Formula：
H=Hc+ρHllc (18)
Here, ρ refers to HcWith HllcBetween attachment coefficient, be in the present invention 0.01, this is because regular terms norm is dilute Thin coding more lays particular emphasis on global optimum, and LLC codings tend to local optimum.For target following, local sparse solution should It is added in global sparse solution.
In SGM models, each candidate samples are scanned through overlapping sliding window and obtain m image block and dictionary D, then lead to It crosses calculation formula (19) and obtains the sparse coefficient vector β of each patch.
Similarly, the sparse coefficient β vectors obtained from each image block are cascaded to form histogram ρ=[β1,β2, β3,…βm]T, while each sample can be calculated and track the similarity L of targetllcThe acquired L of last formula (9)cPass through Formula：
L=Lc+ρLllc (20)
Obtain final sample similarity.
It utilizes put forward GSCM models to be brought into formula (14) trust value H and similarity L herein, is regarded same It frequency sequence and is contrasted such as Fig. 7 in the video frame of identical t moment and the obtained sample likelihood value of SCM models:
Clearly it can find out that sample likelihood value is in GSCM models from the sample likelihood value comparison diagram of SCM and GSCM models In than in SCM model have higher amplitude.Meanwhile in GSCM models, most numerical example dense distribution, likelihood value compared with It is small, but sample likelihood value relative discrete in the distribution of SCM models.And compared with SCM models, there is highest seemingly in GSCM models The sample being so worth can be distinguished with other samples very well.Further, want to choose the sample with highest likelihood value herein This is as final tracking result, this means that the sample distribution ratio SCM models in GSCM models have more distinction, this returns Work(has used local l in LLC methods2Norm has the characteristics that height accurately reconstructed error.
The method of the present invention considers the rarefaction representation ability between sample and template, such as above-mentioned invention content, adds in method LLC codings are entered, this coding regular terms uses l2Norm, the l in machine learning sparse coding2The use of norm determines Coding has an accurately reconstructed error characteristic, avoids because of l1The too small accumulation of error brought of norm reconstructed error, to draw Target is played with losing or tracking failure.
The method of the present invention increases prior probability in step 2, and using Gaussian Profile in probability theory to target around Candidate samples assign different size of weight, and prior probability is added to improve the robustness of model.
In addition, the method for the present invention combines l1Norm and l2Norm carries out calculating feature selecting, and template matches and sample are known Not, detailed process such as above-mentioned formula (20) is described.This method is not only easy to get as a result of existing sparse coding mode Sparse solution, and reduce reconstructed error.Therefore, the sample for sample likelihood ratio other models acquisition that model of the present invention obtains is more With identification.It is used to solve the target following under seriously blocking using this method so that method for tracking target has more robustness.
Sparse coding and LLC coding this methods of combined coding are applied to target following by the present invention, due to sparse coding In regular terms use l1Norm, this coding characteristic determine that method easily obtains sparse solution, the canonical in LLC codings Item uses part l2Norm, coding characteristic determine that code coefficient has the characteristics of accurately reconstructed error.Both are encoded Mode combines, and while easily obtaining sparse solution, also height is accurate for reconstructed error.The present invention utilizes this characteristic, in structure sample When this sparse coefficient between template, otherness can be had more between sample, invention is more made using picking out optimal candidate sample For final goal tracking result.Second, method adds priori, using the Gaussian Profile in probability theory to target around Sample assigns different size of weight, i.e., closer to the easier acquisition greater weight of the sample of target, the addition of prior probability meets The objective law of sample distribution so that method has more robustness.
In conclusion the present invention is to encode two kinds of coding methods using sparse coding and LLC to be applied to target tracking domain, Method only utilizes sparse coding method on feature selecting, template matches and specimen discerning, because of sparse coding It is a kind of coding mode of the overall situation, it has the advantage for easily obtaining sparse solution, but ignores the importance of local code.In It is that experiment effect has obtained very big improvement after being encoded invention introduces LLC, it is a kind of local volume that this, which has benefited from LLC codings, Code mode, and use l2The regular terms of norm compensates for the deficiency of sparse coding.This method to the present invention play to Close important role.Simultaneously, it is proposed by the invention code coefficient addition by way of, make LLC coding how with it is dilute Thin coding is used in combination.Meanwhile the present invention is found that next frame target is very likely distributed in present frame mesh during the experiment Objective fact around marking is found through experiments that using the Gaussian Profile in probability theory to attempt to introduce priori to sample The different size of weight of this imparting, this method enable to result to have more robustness, and one of the innovative point of the present invention.
Description of the drawings
Fig. 1 is expression effect of the method for the present invention under serious circumstance of occlusion.
Fig. 2 is inventive method general frame schematic diagram.
Fig. 3 is the positive and negative template picture that 32 × 32 sizes are obtained from original image.
Fig. 4 is that sliding window scanning K-means clusters to form dictionary schematic diagram.
150 candidate samples confidence value distribution maps of Fig. 5.
Fig. 6 is 150 samples and target similarity value distribution map.
Fig. 7 SCM models and GSCM model sample likelihood value distribution situation comparison diagrams.
Fig. 8 seriously blocks the performance figure on video sequence at 17 groups for before effect 10 tracker.
Fig. 9 is to use OPE and SRE evaluation index experimental results.
Figure 10, which is other 32 groups, has the video sequence experimental result for blocking attribute.
Figure 11 is 49 groups of overall performance results blocked on video sequence.
Specific implementation mode
The present invention proposes the accurate method for tracking target under a kind of serious circumstance of occlusion, and the method includes following steps Suddenly：
Step 1：Target following is carried out using sparse distinction grader and the sparse Gauss conjunctive model for generating model；
Step 2：In sparse distinction grader, candidate samples are weighted using the priori of Gaussian Profile, according to The weight of the variance and mean prediction present frame candidate samples of previous frame target；
Step 3：Use l simultaneously in sparse distinction grader and sparse generation model1Norm is calculated separately with LLC The similitude of candidate samples confidence level and sample and template, and the coefficient that they are obtained is combined；
Step 4：With the sample of weight, confidence level, similarity decision maximum likelihood.
Method proposed by the present invention is very effective in blocking for OPE or SRE assessment performances in video sequence, such as long The TLD track algorithms of time track, and Struck algorithms are general.Due to using ridge regression training objective detector and using cycle Simplifying for matrix Fourier space diagonalization calculates, and KCF algorithm performances are good, show that this method counterweight blocks video sequence more It is effective.Meanwhile even if using deep learning algorithm using OPE assess it is serious block the overall performance of video sequence if not Such as some traditional machine learning algorithms, such as methods herein, KCF, DFT, SCM.The experiment assessed using SRE, performance are not so good as Methods herein and KCF algorithms.This absolutely proves that serious target following under blocking is a challenging task, and The performance of deep learning algorithm is for solving the problems, such as that this might not be effective.MDnet and CNT algorithms are in different video sequence On performance be not fine, wherein MDnet is Bird2, has minimum mean center positioning to miss in Girl video sequences Difference, CNT are only good in the performance of panda video sequences.Reason is that some deep learning target tracking algorisms are that target study carries The further feature of target is taken.When target is seriously blocked, the primitive character partial disappearance of target, these depth characteristics are not Can matching, therefore tracker can wide scope search in track target, this easily lead to target loss or with compared with Big central point mistake.Particularly, the algorithm of this paper using OPE evaluation indexes compared with original SCM algorithms, in centralized positioning Precision in error increases 6.7%, and Duplication precision increases 1.9%.Equally, the algorithm of this paper uses SRE evaluation indexes Compared with original SCM algorithms, precision increases 4.7% in centralized positioning error, and Duplication precision improves 2.6%.
In fact, OTB2013 and OTB2015 target tracking data collection has altogether has the video sequence for blocking attribute comprising 49 groups Row, including general block and block two kinds by force.The main purpose of this paper is the target following solved under seriously blocking, because first Eclipse phenomena is the most common situation of target following, and is that track algorithm institute is most intractable, is secondly seriously blocked than general Block more challenge.But for the robustness of verification algorithm, equally test herein other 32 groups have commonly block category The video sequence of property is simultaneously compared with other most advanced algorithms.Meanwhile also OPE and SRE evaluation indexes being used to judge in fact herein It tests as a result, experimental result is as shown in Figure 9.
The experimental result of Figure 10 and Fig. 9 compares, and the performance of MDnet algorithms is greatly improved, and target is commonly blocking feelings Still a greater part of part primitive character is exposed under condition, deep learning algorithm has very strong feature extraction and characteristic matching energy Power, it can accurately capture target.Therefore, method can do well on this 32 groups of video sequences.The algorithm of this paper 11.6% is improved with SCM algorithm comparisons, central point deviation ratio in OPE evaluation indexes, Duplication improves 11%.Equally, Methods herein improves 9.5% in SRE evaluation indexes with SCM algorithm comparisons, central point deviation ratio, and Duplication improves 5%.On this 32 groups of video sequences, methods herein, the performance of MDnet, KCF algorithm are all very close.But MDnet is more Add with robustness.
In order to test overall performance of this paper total algorithms on blocking video sequence, the present invention completely uses OTB2013 Upper 49 groups there is the video sequence for blocking attribute to be tested with OTB2015, experimental result such as Figure 11.The algorithm of this paper is in OPE 9.7% is improved with SCM algorithm comparisons, central point deviation ratio in evaluation index, Duplication improves 7.4%.Equally, this paper Method improves 7.6% in SRE evaluation indexes with SCM algorithm comparisons, central point deviation ratio, and Duplication improves 3.7%.Always On body surface is existing, context of methods central point deviation is minimum, can more be accurately positioned target.Meanwhile context of methods also has compared with Gao Chong Folded rate, can precisely frame tracked target.
The protection content of the present invention is not limited to above example.Without departing from the spirit and scope of the invention, originally Field technology personnel it is conceivable that variation and advantage be all included in the present invention, and with appended claims be protect Protect range.

Claims (6)

1. the accurate method for tracking target under a kind of serious circumstance of occlusion, which is characterized in that the described method comprises the following steps：
Step 1：Target following is carried out using sparse distinction grader and the sparse Gauss conjunctive model for generating model；
Step 2：In sparse distinction grader, candidate samples are weighted using the priori of Gaussian Profile, according to upper one The weight of the variance and mean prediction present frame candidate samples of frame target；
Step 3：Sparse distinction grader with counted respectively with LLC methods using sparse coding simultaneously in sparse generation model Candidate samples confidence level and sample, the similitude of template are calculated, and obtain two coefficients are combined；
Step 4：With the sample of weight, confidence level, similarity decision maximum likelihood.
2. the accurate method for tracking target under serious circumstance of occlusion as described in claim 1, which is characterized in that in sparse differentiation Property sorter model in, the feature of identification is extracted according to following formula (1)：
NpRefer to positive sample, NnRefer to negative sample, simultaneouslyIt is by NpAnd NnComposition, before K is feature selecting Characteristic dimension, vector elementIndicate attribute of any one template in training set A；Sparse features vector s, it Non-zero element be from K luv spaces it is select have distinctive feature.
3. the accurate method for tracking target under serious circumstance of occlusion as described in claim 1, which is characterized in that in sparse differentiation In property sorter model, target can be indicated by the linear combination of positive masterplate, and background can be indicated by the extension of negative norm plate；Given candidate Target can be indicated by training template and the calculated factor alpha of following formula：
Therefore, the trust value H of structure candidate target xcPass through following formula (4)：
Hc=exp (- (εf-εb)/σ) (4)
Wherein,It is candidate samples x and foreground template collection A+Reconstructed error；It is to wait The reconstructed error of this x of sampling and background template collection A-, and α is related sparse coefficient vector.
4. the accurate method for tracking target under serious circumstance of occlusion as described in claim 1, which is characterized in that normalized M block diagram pictures are obtained with the sliding window of overlapping on image, y is converted into per block diagram picturei∈RG×1Vector, G indicate the big of image block It is small, the sparse coefficient vector β of each patch is calculated by formula (5)：
Dictionary D ∈ RG×JIt is to be generated by k-means cluster centres；The acquired β sparse coefficient vectors of each image block are stitched together Form histogram：
ρ=[β1,β2,β3,...,βm]T (6)
ρ∈R(J×M)×1It is the histogram of each candidate target；Modified reconstruct histogram excludes the image block blocked；
The larger image block of reconstructed error is counted as blocking, and is set to 0 to fixed sparse coefficient vector, calculating passes through formula (7):
ρ=[β1,β2,β3,...,βm]T (7)
Dot product between representing matrix element, each element in o are the indicators of corresponding shielded image block, and calculating passes through public affairs Formula (8):
In formula (8),It is image fritter yiReconstructed error, and yiIt is that candidate samples are swept by sliding window It retouches to be formed, ε0It is that preset threshold value is used to judge whether patch is blocked；
Histogram intersection function is used to calculate the similitude between goal histogram and THE TEMPLATE HYSTOGRAM again, calculation formula is
THE TEMPLATE HYSTOGRAM ψjIt is calculated by formula (5), (6)；Utilize the acquired trust value H in point property sorter modelcWith The similitude L between goal histogram and THE TEMPLATE HYSTOGRAM is acquired in sparse generation modelcIt is combined, is calculated as formula (10)
5. the accurate method for tracking target under serious circumstance of occlusion as claimed in claim 4, which is characterized in that new in order to catch Appearance and restore target in blocking, masterplate histogram is updated to sparse generation model modification, is calculated as formula (11)
New histogram ψnBy according to first frame histogram ψfThe histogram ψ finally storedlIt is constituted, μ is for distributing the formula left side With the weight size on the right, it is used for updated histogram to track the target in next frame image.
6. the accurate method for tracking target under serious circumstance of occlusion as described in claim 1, which is characterized in that LLC codings are adopted With following formula：
Refer to the dot product between element,
In formula (16), dist (xi, B) and=[dist (xi,b1),…,dist(xi,bM)]T, while dist (xi,bi) it is xiWith biIt Between Euclidean distance；σ is the rate of decay for adjusting local adapter weight；
Code coefficient between training template and candidate template can be calculated by following formula：
Then candidate samples and target LLC sparse coding trust values H are obtained by following formulallc
Hllc=exp (- (εf-εb)/σ)
Wherein,It is candidate samples x and foreground template collection A+Reconstructed error；It is candidate samples x With background template collection A-Reconstructed error, and α is related sparse coefficient vector；
Pass through formula (18) again：
H=Hc+ρHllc (18)
Wherein, ρ refers to HcWith HllcBetween attachment coefficient, be in the present invention 0.01, this is because the sparse volume of regular terms norm Code more lays particular emphasis on global optimum, and LLC codings tend to local optimum；For target following, local sparse solution should be added Into global sparse solution；
In sparse generation model, each candidate samples are scanned through overlapping sliding window and obtain m image block and dictionary D, then The sparse coefficient vector β of each patch is obtained by calculation formula (19)；
Similarly, the sparse coefficient β vectors obtained from each image block are cascaded to form histogram ρ=[β1,β2,β3,…βm ]T, while each sample can be calculated and track the similarity L of targetllc.Last formula (9) acquired LcPass through formula：
L=Lc+ρLllc (20)
Obtain final sample similarity.
CN201810310348.9A 2018-04-09 2018-04-09 A kind of accurate method for tracking target under serious circumstance of occlusion Pending CN108549905A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810310348.9A CN108549905A (en) 2018-04-09 2018-04-09 A kind of accurate method for tracking target under serious circumstance of occlusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810310348.9A CN108549905A (en) 2018-04-09 2018-04-09 A kind of accurate method for tracking target under serious circumstance of occlusion

Publications (1)

Publication Number Publication Date
CN108549905A true CN108549905A (en) 2018-09-18

Family

ID=63514390

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810310348.9A Pending CN108549905A (en) 2018-04-09 2018-04-09 A kind of accurate method for tracking target under serious circumstance of occlusion

Country Status (1)

Country Link
CN (1) CN108549905A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110544266A (en) * 2019-09-11 2019-12-06 陕西师范大学 traffic target tracking method based on structure sparse representation
CN110659566A (en) * 2019-08-15 2020-01-07 重庆特斯联智慧科技股份有限公司 Target tracking method and system in shielding state
CN112837342A (en) * 2021-02-04 2021-05-25 闽南师范大学 Target tracking method, terminal equipment and storage medium
CN114152912A (en) * 2021-11-22 2022-03-08 吉林大学 Jet aircraft positioning method based on double-cross function

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZHIJIAN YIN ET AL.: "Accurate target tracking via Gaussian sparsity and locality-constrained coding in heavy occlusion", 《MULTIMEDIA TOOLS AND APPLICATIONS》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110659566A (en) * 2019-08-15 2020-01-07 重庆特斯联智慧科技股份有限公司 Target tracking method and system in shielding state
CN110544266A (en) * 2019-09-11 2019-12-06 陕西师范大学 traffic target tracking method based on structure sparse representation
CN110544266B (en) * 2019-09-11 2022-03-18 陕西师范大学 Traffic target tracking method based on structure sparse representation
CN112837342A (en) * 2021-02-04 2021-05-25 闽南师范大学 Target tracking method, terminal equipment and storage medium
CN112837342B (en) * 2021-02-04 2023-04-25 闽南师范大学 Target tracking method, terminal equipment and storage medium
CN114152912A (en) * 2021-11-22 2022-03-08 吉林大学 Jet aircraft positioning method based on double-cross function

Similar Documents

Publication Publication Date Title
US10929649B2 (en) Multi-pose face feature point detection method based on cascade regression
Hu et al. Deep metric learning for visual tracking
CN108549905A (en) A kind of accurate method for tracking target under serious circumstance of occlusion
CN111795704A (en) Method and device for constructing visual point cloud map
CN106204651B (en) A kind of method for tracking target based on improved judgement with generation conjunctive model
CN110033002A (en) Detection method of license plate based on multitask concatenated convolutional neural network
CN104598890A (en) Human body behavior recognizing method based on RGB-D video
US8761510B2 (en) Object-centric spatial pooling for image classification
CN104616319B (en) Multiple features selection method for tracking target based on support vector machines
CN103020986A (en) Method for tracking moving object
WO2022218396A1 (en) Image processing method and apparatus, and computer readable storage medium
CN107784284B (en) Face recognition method and system
CN105574545B (en) The semantic cutting method of street environment image various visual angles and device
CN104732247B (en) A kind of human face characteristic positioning method
Yang et al. Multiple heterogeneous P-DCNNs ensemble with stacking algorithm: A novel recognition method of space target ISAR images under the condition of small sample set
CN109508674A (en) Airborne lower view isomery image matching method based on region division
Chu et al. Target tracking via particle filter and convolutional network
Sun et al. Visual tracking via joint discriminative appearance learning
CN112084353A (en) Bag-of-words model method for rapid landmark-convolution feature matching
CN110046590A (en) It is a kind of one-dimensional as recognition methods based on particle group optimizing deep learning feature selecting
Sun et al. Feature space fusion classification of remote sensing image based on ant colony optimisation algorithm
Song et al. An improved TLD target tracking algorithm based on Mean Shift
CN113724325A (en) Multi-scene monocular camera pose regression method based on graph convolution network
CN113325379A (en) Ship radar matching method based on target attribute and topological characteristic
Qu et al. Visual tracking with genetic algorithm augmented logistic regression

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: 20180918

WD01 Invention patent application deemed withdrawn after publication