CN105261043A - Video motion object detection method on the basis of significance detection - Google Patents

Abstract

The present invention provides a video motion object detection method on the basis of significance detection. The video motion object detection method on the basis of the significance detection comprises the following steps: if incremental video motion object detection is performed, inputted images are rearranged in real time to obtain an input vector o; and if a batch video motion object detection method is employed, each frame image of the video is rearranged to be a vector, and a matrix O is obtained through combination. In the incremental video motion object detection method, a plurality of frames in the video is required to be randomly selected for the training of a background base vector U, an initial U is obtained, and at the same time parameters and variables are initialized. In the batch video motion object detection method, parameters and variables are also initialized. Through the adoption of the video motion object detection method on the basis of the significance detection, corresponding significance detection result vectors s of the inputted images are calculated; when the batch video motion object detection method is employed for processing, a matrix S is formed by the significance detection result vectors of the images; and an iterative calculation is carried out. According to the invention, the false detection problem caused by the background of the motion may be effectively solved, and the missing detection problem may be also solved.

Description

Based on the video moving object detection method that conspicuousness detects

Technical field

To the present invention relates in the fields such as computer vision video moving object detection method fast and efficiently, particularly relate to the video moving object detection method adopting conspicuousness to detect.

Background technology

It is the pith in computer vision research field that video moving object detects, and moving object segmentation is an important link in video automatic analysis technology [1], and main task is exactly the target of interested motion separated from video sequence.Video moving object detects at video brainpower watch and control, and video compress, the aspect such as video frequency searching and image enhaucament is commonly employed.In actual video process, due to intensity of illumination change, background object rocks (as the water surface, leaf etc.) affects for video moving object detection.How real-time and completely detect that the object of motion becomes the problem that video moving object detects most critical.Therefore, this patent mainly studies the accuracy and integrality that how to improve video moving object detection.

Video moving object detection algorithm mainly contains object detection operator and background subtraction method.Object detection operator [3] is a sorter normally, and it is scanned each frame by a moving window, thus marks the subimage in each moving window, judges that subimage is moving target or background.But for background subtraction algorithm [4], it is by each two field picture and the background model set up in advance being compared, motion target area being regarded as in the vicissitudinous region detected.But this method requires that front some width images of video sequence do not comprise motion target area usually, could train background model well.T.Haines [4] etc. proposes a kind of gauss hybrid models algorithm based on Di Li Cray process, uses Di Li Cray gauss hybrid models to background modeling, and then video sequence and gained background model are subtracted each other and obtained moving target.Can find out, the demand of these training sets or background model in fact limits the use of Detection for Moving Target in practical application.

The method that another kind realizes Detection for Moving Target is based drive method.This method only utilizes movable information to make moving target and background separation, thus avoids the training stage.This method is based on a hypothesis: in the video sequence, and the motion of moving target is the motion being different from background.Therefore we automatically can isolate moving target from background.Usually first calculate the movement locus of each pixel, carry out motion segmentation, thus make moving target and background separation, be generally referred to as " optical flow method " [7].2010, the people such as T.Brox [7] proposed a kind of moving object detection algorithm based on pixel movement locus.This Algorithm Analysis in the video sequence each pixel movement locus and to its segmentation cluster.The advantage of motion segmentation method is that it can process the situation of violent camera motion, but it has following shortcomings: the first, the motion that the treatment conditions of this method are based on moving target in camera angles is all the hypothesis of rigid motion or smooth motion.But this does not conform to often with actual conditions.In a practical situation, the motion of moving target is very complicated often, may with nonrigid body shape changes.The second, in the video sequence of reality, change of background is also very complicated.The structure of change may be contain, such as: the wave in the trees of swing or water in background.Meanwhile, background also can be subject to the impact of illumination variation.These factors affect the estimation of optical flow method to background motion track, make motion segmentation produce mistake.

Another video frequency motion target detection algorithm is the algorithm based on matrix representation.This method is usually based on a hypothesis: in video, between each image background pixels point, correlativity is comparatively large, and the background matrix be therefore made up of each two field picture of vectorization is low-rank; And Moving Object in Video Sequences is often less, so prospect matrix is sparse.And then can image array resolve into low-rank matrix and sparse matrix and.In recent years, the moving object detection algorithm based on matrix representation that Chinese scholars proposes is broadly divided into two classes: batch type moving object detection algorithm and increment type moving object detection algorithm.Batch type detection algorithm needs to carry out batch type process to some frames of input video, and increment type detection algorithm processes frame by frame to video sequence.

For batch type moving object detection algorithm, more classical is 2011, RobustPrincipalComponentAnalysis (RPCA) algorithm proposed by E.Candes [8], article mainly proposes to be background low-rank matrix and prospect sparse matrix by video sequence matrix decomposition.And for background low-rank matrix, can approach with convex being similar to, thus can retrain it with nuclear norm.For prospect sparse matrix, usually by 1 norm, it is retrained.Moving object detection can be realized by solving so constrained optimization problem.But this algorithm does not consider the impact that noise detects motion.In order to reach better accuracy of detection, 2013, on RPCA basis, the people such as X.Zhou [2] proposed DetectingContiguousOutliersintheLow-rankRepresentation (DECOLOR) algorithm.This algorithm carries out sparse constraint to prospect matrix except continuing to use 1 norm, also add the impact of neighbor, namely retrains prospect matrix with markov random file.In addition, this algorithm also contemplates the impact of noise.This algorithm adds camera motion Matrix Estimation simultaneously, effectively can process the situation of camera motion.This algorithm, as batch type moving object detection algorithm, achieves good Detection results.2015, the people such as BoXin [6] propose a kind of new batch type moving object segmentation algorithm on the basis of RPCA, this algorithm retrains (thegeneralizedfusedlassoregularization) prospect matrix drag-line method, for background matrix, it is carried out linear expression with the pure image array of moving target that do not have, meanwhile, sparse constraint is carried out to sparse matrix.Solve optimization with Lagrangian method, finally realize moving object detection.

About increment type moving object detection algorithm, equally based on the hypothesis of RPCA, usually recover to realize with sparse signal.2012, the people such as J.He [10] proposed GrassmannianRobustAdaptiveSubspaceTrackingAlgorithm (GRASTA) algorithm on RPCA basis, and this algorithm is for prospect, and same base sparse is supposed.For background, by the subspace at estimated background low-rank matrix, the background of each two field picture of input video all can be regarded as and linearly be made up of the base vector of background subspace.Solve optimization procedure thus reach separate moving objects and background, input video can be processed in real time.But this algorithm does not consider the impact of the noise in video, do not utilize any priori about moving target yet.Same 2014, X.Guo [9] etc. also proposes a kind of New Algorithm on the basis of RPCA algorithm, based on the hypothesis that moving target is all motion continuously, set the base vector of background matrix equally, and in objective function, with the addition of the level and smooth item of moving target matrix in time domain and spatial domain.May be simultaneously the hypothesis of random color or intensity based on the outward appearance of moving target, add the item constraint item about moving target outward appearance from the angle of probability.

From the evolution of video moving object detection algorithm, the video sequence moving object detection algorithm based on matrix representation has following advantage relative to other motion detection algorithms: one, do not need the various training stage and set up background model; Two, energy sweetly disposition illumination and noise are on the impact of testing result; Three, the various type of sports of moving target can be processed, comprise non-rigid motion.Video moving object based on matrix representation detects and also becomes the focus studied in recent years.But there is Railway Project in existing video moving object detection algorithm: be moving object and prospect by background (as the leaf rocked, the ripple of the water surface, the elevator etc. of the motion) flase drop of motion; The moving target detected is imperfect, has hollow out phenomenon.This patent, for above two problems, proposes a kind of video moving object detection algorithm detected based on conspicuousness.On the basis of matrix representation, by adding that conspicuousness detects bound term, thus reduce flase drop and undetected problem.

List of references:

[1]A.Yilmaz,O.Javed,etal.,“Objecttracking:Asurvey,”ACMTransactiononComputingSurveys,vol.38,no.4,pp.1–45,2006.

[2]X.Zhou,C.Yang,etal.,“Movingobjectdetectionbydetectingcontiguousoutliersinthelow-rankrepresentation,”IEEETransactionsonPatternAnalysisandMachineIntelligence,vol.35,no.3,pp.597–610,2013.

[3]P.Viola,M.Jones,etal.,“Detectingpedestriansusingpatternsofmotionandappearance,”InternationalJournalofComputerVision,vol.63,no.2,pp.153–161,2005.

[4]T.Haines,TaoXiang,“BackgroundSubtractionwithDirichletProcessMixtureModels,”IEEETransactiononPatternAnalysisandMachineIntelligence,vol.36,no.4,2014.

[5]W.Zhu,S.Liang,Y.Wei,andJ.Sun,“Saliencyoptimizationfromrobustbackgrounddetection,”Proc.IEEEInternationalConferenceonComputerVisionandPatternRecognition,2014.

[6]B.Xin,Y.Tian,Y.Wang,andW.Gao,“BackgroundSubtractionviaGeneralizedFusedLassoForegroundModeling,”Proc.IEEEInternationalConferenceonComputerVisionandPatternRecognition,2015.

[7]T.Brox,J.Malik,“Objectsegmentationbylongtermanalysisofpointtrajectories,”EuropeanConferenceonComputerVision,2010.

[8]E.Candes,X.Li,etal.,“RobustPrincipalComponentAnalysis？”JournaloftheACM,2011.

[9]X.Guo,X.Wang,etal.,“RobustforegroundDetectionUsingSmoothnessandArbitrarinessConstraints,”EuropeanConferenceonComputerVision,2014.

[10]J.He,L.Balzano,etal.,“IncrementalGradientontheGrassmannianforOnlineForegroundandBackgroundSeparationinSubsampledVideo,”IEEEInternationalConferenceonComputerVisionandPatternRecognition,2012.

Summary of the invention

The object of the invention is to reduce the flase drop in video moving object detection and undetected problem, propose the video moving object detection method detected based on conspicuousness.Technical scheme of the present invention is as follows:

Based on the video moving object detection method that conspicuousness detects, comprise the following steps:

Step 1: if increment type video moving object detect, then the real-time image by input is rearranged into vector, obtains input vector o, if batch like process, then by every two field picture of video all permutatation become a vector, combination obtain a matrix O;

Step 2: increment method needs the some frames in Stochastic choice video to carry out the training of background base vector U, obtains an initial U, simultaneously some parameters of initialization and variable, and batch like process needs some parameters of first initialization and variable equally;

Step 3: according to adopted conspicuousness detection method, for every width image of input, calculates its corresponding conspicuousness testing result vector s; During batch methods process, by the conspicuousness testing result of an every width image vector composition matrix S;

Step 4: setting iterations, starts iterative computation;

Step 5: for each variable and parameter, calculate successively according to method for solving, until net result convergence.

Step 6: after iteration terminates, the net result of output variable.

Adopt the method for the invention, detect bound term by increasing conspicuousness, Background suppression motion detects on video moving object the impact caused effectively, and undetected problem.For existing video moving object detection method, the video moving object detection method detected based on conspicuousness effectively improves background (as the leaf rocked due to motion, the ripple of the water surface and the elevator etc. of operation) the flase drop problem that produces, can reduce undetected problem simultaneously.Meanwhile, the method is being carried out based on the basis of matrix decomposition, comparatively simple for conventional video moving object segmentation algorithm.

Accompanying drawing explanation

Fig. 1 is institute of the present invention extracting method block diagram, and batch like process performs by bracket.

Embodiment

Technical matters to be solved by this invention is on the basis of matrix representation, detecting bound term, realizing video moving object and detecting by increasing conspicuousness.

The existing video moving object detection algorithm based on matrix representation generally comprises four: Background Reconstruction item, background low-rank bound term, prospect sparse constraint item, level and smooth item or connection item.Video moving object based on matrix representation detects increment method and is generally formulated as follows:

$\underset{b,U,v}{\min }\underset{i=1}{\overset{N}{\Σ}}\[\frac{1}{2}{b}_i{(U_{i}v-o_i)}^2\]+\underset{i=1}{\overset{N}{\Σ}}\mathrm{\β}(1-b_i)+\mathrm{\λ}\left|\right|Db|{|}_1.---\left(1\right)$

In formula (1), b _i∈ [0,1] is background vector, b _ibe worth larger, represent that this point may be more background dot; D is a forward difference matrix; represent Background Reconstruction item; (1-b _i) be the sparse constraint item of prospect; || Db|| ₁namely be communicated with item, consider the impact of neighbor pixel.

For the moving object segmentation algorithm of batch, be typically expressed as:

$\underset{B,F}{\min }\underset{ij}{\Σ}\frac{1}{2}{F}_{ij}{\left(O_{ij}-B_{ij}\right)}^2+\mathrm{\β}\underset{ij}{\Σ}{F}_{ij}+\mathrm{\γ}\underset{ij,kl\∈\mathrm{\ϵ}}{\Σ}\left|F_{ij}-F_{kl}\right|,---\left(2\right)$

s.t.rank(B)≤K.

Or be expressed as:

$\underset{B,F}{\min }\underset{ij}{\Σ}\frac{1}{2}{F}_{ij}{(O_{ij}-B_{ij})}^2+\mathrm{\α}\left|\right|B|{|}_{*}+\mathrm{\β}|\left|F\right|{|}_1+\mathrm{\γ}\left|\right|Avec\left(F\right)|{|}_1.---\left(3\right)$

In formula (3), F _ij∈ [0,1] is prospect matrix, and the larger expression of value more may become foreground point; it is Background Reconstruction item; || B|| _*represent background low-rank bound term, with the low-rank of nuclear norm representing matrix; || F|| ₁the sparse constraint item of expression prospect; || Avec (F) || ₁represent and be communicated with item.

There are some problems in the existing video moving object detection algorithm based on matrix representation.But increment type video moving object detection algorithm is due to the simple robust not of model, the impact of the background of easily being moved, thus causes flase drop, and the moving target detected exists hollow out simultaneously.Batch type video moving object detection algorithm is comparatively complicated, and robustness is good, but the motion target area often detected is than actual large, then produces flase drop problem.In order to solve above-mentioned flase drop and undetected problem, this patent adds that on above-mentioned matrix decomposition basis conspicuousness detects bound term, conspicuousness detection method can detect interested motion target area to a certain extent, and conspicuousness detection method processes single image, by the impact of movement background, thus can not can improve the problems referred to above.

It is as follows that increment type video moving object based on conspicuousness constraint detects general formulae, formula (1) basis adds conspicuousness and detects [5] bound term:

$\underset{b,U,v}{\min }\underset{i=1}{\overset{N}{\Σ}}\[\frac{1}{2}{b}_i{(U_{i}v-o_i)}^2+\mathrm{\β}(1-b_i)-{\mathrm{\αb}}_i(1-s_i)\]+\mathrm{\λ}\left|\right|Db|{|}_1.---\left(4\right)$

In formula (4), b _i(1-s _i) be that conspicuousness detects bound term, wherein s _i∈ [0,1] is the result that conspicuousness detects.Namely with conspicuousness testing result s, background vector b is retrained, background vector b _ivalue along with s _ireduction and increase, so the inapparent point of conspicuousness testing result just more may be detected by video moving object detection algorithm and become background dot, thus improves the flase drop problem that background motion produces.

Batch type video moving object is detected, increases conspicuousness detection bound term as follows:

$\underset{B,F}{\min }\underset{ij}{\Σ}\frac{1}{2}{F}_{ij}{\left(O_{ij}-B_{ij}\right)}^2+\mathrm{\α}\left|\right|B|{|}_{*}+\mathrm{\β}|\left|F\right|{|}_1+\mathrm{\γ}\left|\right|Avec\left(F\right)|{|}_1-\mathrm{\λ}\underset{ij}{\Σ}{S}_{ij}{F}_{ij}.---\left(5\right)$

In formula (5), S _ij∈ [0,1] is conspicuousness testing result, that conspicuousness detects bound term.Namely with conspicuousness testing result S, prospect matrix F is retrained, reduce the problem of the moving target profile expansion detected in original batch like process, make testing result more accurate.

In formula (4) and formula (5), by conspicuousness testing result, background or prospect matrix are retrained, because conspicuousness testing result is not by the impact of movement background, can be relatively complete detect interested moving target in video, then effectively can improve the flase drop problem that movement background causes.Can find out, conspicuousness detects and has certain effect for video moving object detection.

The present invention, from whether reducing the flase drop and undetected problem that exist in video moving object detection algorithm, proposes the video moving object detection method detected based on conspicuousness.No matter be batch or increment method, all can realize.In incremental method, each variable is vector operation, and batch type algorithm is matrix operation.Its concrete steps are as follows:

Step 1: if increment type video moving object detects, then the real-time image by input is rearranged into vector, obtains input vector o.If batch like process, then by every two field picture of video all permutatation become a vector, combination obtain a matrix O.

Step 2: increment method needs the some frames in Stochastic choice video to carry out the training of background base vector U, obtains an initial U, simultaneously some parameters of initialization and variable (as α, β, λ, v).Batch like process needs some parameters of first initialization and variable equally, α, β, λ, γ.

Step 3: according to adopted conspicuousness detection method [5], for every width image of input, calculates its corresponding conspicuousness testing result vector s; During batch methods process, by the conspicuousness testing result of an every width image vector composition matrix S.

Step 4: setting iterations, starts iterative computation.

Step 5: for each variable and parameter, calculate successively according to method for solving, until net result convergence.

Step 6: after iteration terminates, the net result b (or F) of output variable.

Claims (1)

1., based on the video moving object detection method that conspicuousness detects, comprise the following steps:

Step 1: if increment type video moving object detect, then the real-time image by input is rearranged into vector, obtains input vector o, if batch like process, then by every two field picture of video all permutatation become a vector, combination obtain a matrix O;

Step 2: increment method needs the some frames in Stochastic choice video to carry out the training of background base vector U, obtains an initial U, simultaneously some parameters of initialization and variable, and batch like process needs some parameters of first initialization and variable equally;

Step 3: according to adopted conspicuousness detection method, for every width image of input, calculates its corresponding conspicuousness testing result vector s; During batch methods process, by the conspicuousness testing result of an every width image vector composition matrix S;

Step 4: setting iterations, starts iterative computation;

Step 5: for each variable and parameter, calculate successively according to method for solving, until net result convergence.

Step 6: after iteration terminates, the net result of output variable.