CN103106666A - Moving object detection method based on sparsity and smoothness - Google Patents

Abstract

The invention discloses a moving object detection method based on sparsity and smoothness in the technical field of video picture processing. According to the method, a regression model which is suitable for moving object detection is designed, restraints of sparsity and smoothness are exerted on a moving object when the regression model is used for assessing the moving object, and then a final detection result is obtained. By means of the moving object detection method based on sparsity and smoothness, the detection result of the moving object is enabled to be accurate and reliable in the complex environment such as a dynamic background.

Description

Moving target detecting method based on sparse property and flatness

Technical field

What the present invention relates to is a kind of method of technical field of video image processing, specifically a kind of moving target detecting method based on sparse property and flatness.

Background technology

The research of moving target detecting method and application are active branches in computer vision, intelligent video analysis field, play an important role in the practical applications such as video monitoring, control automatically, safety inspection.The moving object detection result is the basis of carrying out the information processing of higher level accurately and reliably, as target following, target identification, behavioural analysis etc.

Present moving target detecting method has been obtained more stable and reliable result under common environment, but the performance of these methods under complex scene often can not be satisfactory.Moving object detection under dynamic background has been subject to paying close attention to widely as one of the difficult point of target detection under some complex scenes for many years, invents a kind of moving target detecting method that is applicable under dynamic background and has great importance.

Existing moving target detecting method classification three classes: optical flow method, frame differential method, background subtraction point-score.

Optical flow method is come the disengaging movement target by calculating pixel motion of point vector, and this class methods calculated amount is larger, and complexity is high, is mainly used at present in dollying headring border.Frame differential method is that the Strength Changes according to corresponding pixel points between consecutive frame detects moving target.Although these class methods are simple, often can only extract the profile of target, and also also more responsive to noise, so this class methods practicality is not strong yet.

The background subtraction point-score is present the most frequently used moving target detecting method, its basic thought is by the study of frame of video being set up the description to background, video image and the background model that then will newly obtain compare calculating, when the pixel in new video frame does not meet the current background description, judge that this point is the foreground point, otherwise belong to background dot, thereby complete detection to moving target by node-by-node algorithm.in the background subtraction point-score, more representational method comprises: the Gaussian mixture model-universal background model method that " the Adaptive background mixture models of real-time tracking " that the people such as C.Stauffer and W.E.L.Grimson delivered at Proc.Conf.Computer Vision and Pattern Recognition (computer vision and pattern-recognition international conference) in 1999 (the adaptive background mixture model that is used for a real-time follow-up) literary composition proposes, the method thinks that the value of pixel meets Gaussian distribution, and the value of each pixel can be obtained by a plurality of adaptive Gaussian mixture model-universal background model weighted arrays, thereby set up the detection that Gaussian mixture model-universal background model has been realized moving target, A.Elgammal, R.Duraiswami, " the Background and foreground modeling using non-parametric kernel density estimation for visual surveillance " that the people such as D.Harwood and L.S.Davis delivered at Proc.IEEE (electronics and the Institution of Electrical Engineers's proceedings) in 2002 (being used for the background prospect modeling based on the norm of nonparametric kernel density estimation of a video monitoring) literary composition proposes to carry out background modeling with the method for Density Estimator, the method is not done any hypothesis to the distribution of pixel point value, but the value of pixel is added up to obtain the parameter estimation of kernel function by the method for time domain, then set up background model according to these estimated parameters and kernel function, realized the detection of target, Mert Dikmen and Thomas S.Huang in 2008 at 19th International Conference on Pattern Recognition, " the Robust Estimation of Foreground in Surveillance Videos by Sparse Error Estimation " that (the pattern-recognition international conference of the 19th boundary) delivered, propose to use based on sparse theory and method in (foreground detection of estimating based on sparse mistake in video monitoring) literary composition and carry out the foreground target detection.The method is regarded the detection of background and prospect as a class signal separation problem: background signal ratio of transformation in time is slower; The foreground signal variation is different from background signal, and foreground signal has sparse character.By the conversion of problem and can realize to sparse signal being the estimation of foreground target by means of existing sparse theory.

Above-mentioned background subtraction separating method effect under the scene of dynamic background is often unsatisfactory, and these methods are the foreground point with part background dot flase drop.Therefore, need to seek a kind of moving target detecting method that can be under the dynamic background environment.

Application number is the Chinese invention patent of CN201110223892.8, this patent provides a kind of object detection method based on linear regression model (LRM), although the method can realize the moving object detection under this complex environment of dynamic background, effect can also further be optimized.

Summary of the invention

The present invention is directed to the prior art above shortcomings, a kind of moving target detecting method based on sparse property and flatness is provided, the target detection result under this complex environment of dynamic scene is accurate and reliable.

The present invention is achieved by the following technical solutions, at first the present invention designs a kind of regression model that is suitable for target detection, then when using this model that moving target is estimated, according to the characteristic of moving target, it is applied the constraint of sparse property and flatness, thereby obtain final testing result.

The described regression model that is suitable for target detection refers to: this regression model can be expressed as form:

y=Xw+t+n (1)

Wherein: y=(y ₁, y ₂..., y _m) ^TThe expression dependent variable, X=(x ₁, x ₂..., x _m) ^TExpression independent variable and x ₁, x ₂..., x _mBe the p dimensional vector, p represents the element number that each vector comprises, and m represents number and the m of observed quantity〉p, w=(w ₁, w ₂..., w _p) ^TThe coefficient of expression regression model, t=(t ₁, t ₂..., t _m) ^TRepresent difference in the part of independent variable, n=(n ₁, n ₂..., n _m) ^TThe expression stochastic error is according to the hypothesis of classical linear regression model (LRM) to stochastic error, n ₁, n ₂..., n _mAll obeying average is 0, and variance is σ ²Gaussian distribution and this Gaussian distribution be designated as Ν (0, σ ²).

The detection that this model of described utilization carries out moving target refers to: corresponding to formula (1), with present frame as dependent variable y, as independent variable X, the prospect part as the part t of difference in independent variable, is regarded some historical frames as noise n with the motion of background self.Regression coefficient is estimated according to the data of historical frames and present frame according to the original idea needs of linear regression model (LRM), in order to estimate more accurately, two characteristics utilizing the prospect part usually to have are sparse property and flatness, define a new objective function and obtain the coefficient w of regression model, and then estimate difference in the part t of independent variable.

Described new objective function refers to: suppose that prospect is partly sparse in level and smooth, objective function is expressed as:

$(\hat{w},\hat{t})=\underset{w,t}{\arg \min }{\left|\right|y-\mathrm{Xw}-t\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|t\left|\right|}_0+{\mathrm{\λ}}_2{\left|\right|t^{\′}\left|\right|}_0---\left(2\right)$

Wherein: || || ₀The 0-norm of expression vector, t '=(t ₂-t ₁, t ₃-t ₂..., t _m-t _m-1) ^TExpression foreground signal adjacent element value difference divides a new vector that consists of, in order to describing its flatness,

Represent respectively model coefficient and difference in the estimated value of the part of independent variable, λ ₁And λ ₂Expression adjustment factor, wherein λ ₁Control the sparse degree of foreground signal, λ ₂Control the level and smooth degree of foreground signal.More pay close attention to difference in the target detection practical problems in the part of independent variable owing to two independents variable being arranged in formula (2) simultaneously, being incorporated into

Use a kind of disposal route of simplification to simplify.

Preferably, the processing of a kind of simplification of said method is: at first computing formula (2) is set as 0 to the partial derivative of variable w and with the value of resulting expression formula, like this can be in the hope of the value of objective function (2) corresponding w when obtaining minimum value, even:

$\frac{\∂({\left|\right|y-\mathrm{Xw}-t\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|t\left|\right|}_0+{\mathrm{\λ}}_2{\left|\right|t^{\′}\left|\right|}_0)}{\∂w}=0---\left(3\right)$

Can obtain following formula by the calculating to formula (3):

$\hat{w}=X^{+}(y-t)---\left(4\right)$

X wherein ⁺Pseudo inverse matrix for X is expressed as X ⁺=(X ^TX) ^-1X ^TThen be similar to the 1-norm with formula (4) substitution formula (2), and with the 0-norm of vector, can obtain the objective function of following simplification:

$\left(\hat{t}\right)=\underset{t}{\arg \min }{\left|\right|W(y-t)\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|t\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|t^{\′}\left|\right|}_1---\left(5\right)$

W=(I wherein _m-XX ⁺), I _mIt is the unit matrix of m for size.

Can estimate according to existing sparse theory and derivation algorithm

Can obtain the estimated value of difference in the part of independent variable, consider simultaneously the loss that causes spatial information after image vector, adopt by different way with the view data vector quantization, will estimate that then the result that obtains merges.

Described by different way with the view data vector quantization, then the result of estimating to obtain is merged and refer to: respectively image is carried out that horizontal vector is stretching and vertical vector is stretching, and respectively foreground target is estimated according to the data under these two kinds of processing modes, then these two estimated results carry out logical OR and merge to obtain net result.

Compared with prior art, the present invention has following beneficial effect:

The invention provides the moving target detecting method under a kind of dynamic background, make in the environment of background perturbation and relatively sparse moving target can be extracted, simultaneously in leaching process with the divergent setting noise, utilize the character of moving target local smoothing method, with noise separation out and be removed, in the testing result that obtains, the clear in structure of moving target is complete, and accurately and reliably, and ground unrest is substantially suppressed falls.CN201110223892.8 compares with patent of invention, and the present invention has not only utilized the sparse attribute of target in the process of estimating moving target with linear regression model (LRM), also target has been carried out the constraint of local smoothing method.The present invention provides effective solution for this technical barrier of detection of moving target under dynamic background.

Description of drawings

By reading the detailed description of non-limiting example being done with reference to the following drawings, it is more obvious that other features, objects and advantages of the present invention will become:

Fig. 1 is the process flow diagram of one embodiment of the invention.

Fig. 2 is one embodiment of the invention effect schematic diagram.

Embodiment

The present invention is described in detail below in conjunction with specific embodiment.Following examples will help those skilled in the art further to understand the present invention, but not limit in any form the present invention.Should be pointed out that to those skilled in the art, without departing from the inventive concept of the premise, can also make some distortion and improvement.These all belong to protection scope of the present invention.

As shown in Figure 1, the present embodiment provides a kind of moving target detecting method based on sparse property and flatness, and concrete implementation detail is as follows, and the part that following examples do not have to describe in detail is carried out with reference to summary of the invention:

(1). structure independent variable matrix: extract piece image every 10 frames from the front 200 width images of current video frame sequence, extract altogether 20 width images as the training video image of present frame, every two field picture is consisted of a vector by the stretching mode of level (vertically), this vector is joined in the i row of corresponding independent variable matrix X wherein (0＜i≤20).

(2). calculate the pseudo-inverse matrix X of independent variable matrix X ⁺

(3). for non-training frames, be translated into vector y according to the stretching mode of step (1).

(4). according to the parameter lambda in data with existing X and y estimation formulas (5) ₁And λ ₂Concrete steps are as follows:

(a). estimate a regression coefficient with classical linear regression model (LRM)

(b). according to the coefficient calculations error vector that obtains

(c). estimate an initial deviate according to absolute meta deviation approach:

Wherein median operation is got in med () expression;

(d). λ is set ₁And λ ₂Parameter be ${\mathrm{\λ}}_1=\frac{\widehat{\mathrm{\σ}}}{15}*\sqrt{2\log \left(m\right)},$ ${\mathrm{\λ}}_2=\frac{25*\widehat{\mathrm{\σ}}}{15}*\sqrt{2\log \left(m\right)}.$

(5). differ from the part t of independent variable according to formula (5) estimated difference, the image of estimating to obtain is carried out thresholding process, be judged to be the foreground point greater than the pixel of threshold value, otherwise be background dot.Be made as 25 in this threshold value.

(6). by the stretching and vertical stretching vector constituted mode of level, obtain corresponding estimated result under these two kinds of stretching modes according to step (1)-(5) respectively.

(7). above-mentioned two kinds of estimated results are carried out the logical OR fusion obtain final result.

Implementation result

According to above-mentioned steps, to being tested by the open dynamic background cycle tests that provides on the internet.The sequence scene is a campus environment, motor vehicle and people process successively then, and in whole process, leaf can be shaken along with wind, the testing result of this method test moving target.

As shown in Figure 2, figure (a) is sequence incoming frame, namely training image; Figure (b) and figure (c) are respectively test pattern and the testing result thereof of the 1204th frame of moving target; Figure (d) and figure (e) are respectively test pattern and the testing result of the 1385th frame of moving target; Figure (f) and figure (g) are respectively test pattern and the testing result of the 1668th frame of moving target.Figure (h) and figure (i) are respectively test pattern and the testing result of the 1812nd frame of moving target.The inventive method is accurate and reliable to the testing result of moving target under dynamic background as seen from the figure, has embodied validity of the present invention and value.

In order to embody progressive of the present invention, the inventive method and the traditional Gaussian mixture model-universal background model method (proposition such as C.Stauffer, be called for short GMM), Density Estimator method propositions such as (, be called for short KDE) A.Elgammal and the patent No. be that the patent of invention " based on the object detection method of linear regression model (LRM) " (being called for short the linear regression model (LRM) method) of CN201110223892.8 has carried out quantizing relatively.It is that measurement index is estimated the testing result of three kinds of methods that the present invention adopts F_score:

Wherein precision ratio and recall ratio are orientated as respectively:

The F_score index is higher shows that method is more effective.

Test by the 10 frame results that above-mentioned three kinds of methods are chosen arbitrarily at this cycle tests, and the result that will estimate is compared as follows:

Quantitative evaluation result comparative illustration this method all is better than above-mentioned traditional two kinds of methods and linear regression model (LRM) method at the detection aspect of performance, has further embodied the value of the inventive method.

Above specific embodiments of the invention are described.It will be appreciated that, the present invention is not limited to above-mentioned specific implementations, and those skilled in the art can make various distortion or modification within the scope of the claims, and this does not affect flesh and blood of the present invention.

Claims (2)

1. moving target detecting method based on sparse property and flatness, it is characterized in that, by designing a kind of regression model that is suitable for moving object detection, then when using this model that moving target is estimated, the moving target item is applied the constraint of sparse property and flatness, thereby obtain final testing result;

The described regression model that is suitable for moving object detection, this regression model is expressed as form:

y=Xw+t+n (1)

Wherein: y=(y ₁, y ₂..., y _m) ^TThe expression dependent variable, X=(x ₁, x ₂..., x _m) ^TExpression independent variable and x ₁, x ₂..., x _mBe the p dimensional vector, p represents the element number that each vector comprises, and m represents number and the m of observed quantity〉p, w=(w ₁, w ₂..., w _p) ^TThe coefficient of expression regression model, t=(t ₁, t ₂..., t _m) ^TRepresent difference in the part of independent variable, n=(n ₁, n ₂..., n _m) ^TThe expression stochastic error is according to the hypothesis of classical linear regression model (LRM) to stochastic error, n ₁, n ₂..., n _mAll obeying average is 0, and variance is σ ²Gaussian distribution and this Gaussian distribution be designated as Ν (0, σ ²);

the detection that this model of described utilization carries out moving target refers to: corresponding to formula (1), with present frame as dependent variable y, with some historical frames as independent variable X, with prospect part as the part t of difference in independent variable, regard the motion of background self as noise n, regression coefficient is estimated according to the data of historical frames and present frame according to the original idea needs of linear regression model (LRM), in order to estimate more accurately, utilizing two characteristics that prospect partly has is sparse property and flatness, define a new objective function and obtain the coefficient w of regression model, and then estimate difference in the part t of independent variable, consider the loss that causes spatial information after image vector, adopt by different way with the view data vector quantization, then will estimate that the result that obtains merges,

Described new objective function refers to: suppose that prospect is partly sparse in level and smooth, objective function is expressed as:

$(\hat{w},\hat{t})=\underset{w,t}{\arg \min }{\left|\right|y-\mathrm{Xw}-t\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|t\left|\right|}_0+{\mathrm{\λ}}_2{\left|\right|t^{\′}\left|\right|}_0---\left(2\right)$

Wherein: || || ₀The 0-norm of expression vector, t '=(t ₂-t ₁, t ₃-t ₂..., t _m-t _m-1) ^TExpression foreground signal adjacent element value difference divides a new vector that consists of, in order to describing its flatness,

Represent respectively model coefficient and difference in the estimated value of the part of independent variable, λ ₁And λ ₂Expression adjustment factor, wherein λ ₁Control the sparse degree of foreground signal, λ ₂Control the level and smooth degree of foreground signal;

Described employing is by different way with the view data vector quantization, then the result of estimating to obtain is merged and refer to: respectively image is carried out that horizontal vector is stretching and vertical vector is stretching, and respectively foreground target is estimated according to the data under these two kinds of processing modes, then these two estimated results carry out logical OR and merge to obtain net result.

2. the moving target detecting method based on sparse property and flatness according to claim 1, it is characterized in that, the processing of a kind of simplification of described method is: at first computing formula (2) is set as 0 to the partial derivative of variable w and with the value of resulting expression formula, try to achieve like this value of objective function (2) corresponding w when obtaining minimum value, even:

$\frac{\∂({\left|\right|y-\mathrm{Xw}-t\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|t\left|\right|}_0+{\mathrm{\λ}}_2{\left|\right|t^{\′}\left|\right|}_0)}{\∂w}=0---\left(3\right)$

By the following formula that calculates to formula (3):

$\hat{w}=X^{+}(y-t)---\left(4\right)$

X wherein ⁺Pseudo inverse matrix for X is expressed as X ⁺=(X ^TX) ^-1X ^T, then with formula (4) substitution formula (2), and the 0-norm of vector is similar to the 1-norm, obtain the objective function of following simplification:

$\left(\hat{t}\right)=\underset{t}{\arg \min }{\left|\right|W(y-t)\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|t\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|t^{\′}\left|\right|}_1---\left(5\right)$

W=(I wherein _m-XX ⁺), I _mIt is the unit matrix of m for size.

Images