CN103824297B - In complicated high dynamic environment, background and the method for prospect is quickly updated based on multithreading - Google Patents

Abstract

The present invention relates to image processing field, be specifically related to a kind of in complicated high dynamic environment, quickly update background and the method for prospect based on multithreading, comprise the following steps: pixel and vicinity points thereof are carried out Haar like feature extraction；According to the complexity of scene, the characteristic vector of front 10～35 two field pictures is directly incorporated into the end of background model matrix；Obtain a certain pixel in image, calculate this pixel distance to its background model, differentiate that the most whether this pixel is as foreground point with this；If showing that current pixel point is background dot by step 3, then update eigenvalue minimum with current P2M distance in background model matrix；If showing that current pixel point is background dot by step 3, from the pixel that current pixel point is neighbouring, randomly selecting a point, updating feature maximum with current pixel characteristic distance in its background model.The present invention can update background in real time, exactly, adapts to various complex environment, is effectively improved accuracy and the adaptability of foreground detection.

Description

In complicated high dynamic environment, background and the method for prospect is quickly updated based on multithreading

Technical field

The present invention relates to image processing field, be specifically related to one and in complicated high dynamic environment, quickly update background based on multithreading Method with prospect.

Background technology

Along with developing rapidly of the emerging fields such as video monitoring, man-machine interaction, picture coding and retrieval increases with the quick of technical need Many, the correlation technique of computer vision also achieves huge breakthrough, and image procossing becomes the basis of industry and the technology core in field The heart, and background modeling is as common process means therein, also achieves significant progress.

Background modeling method the most popular, most popular surely belongs to mixed Gauss model.Each pixel in image is clicked on by it Row modeling, the distributed model defining each pixel is the set being made up of multiple single Gauss models, according to the pixel that each is new Value updates model parameter, according to certain criterion judge which pixel be background dot, which is as foreground point；When illumination occurs big When changing rapidly of scale, mixed Gauss model will for its newly-built Gauss body, but still using former pixel value as background (because of The stage that can replace original main Gauss body it is less than for " strength " of new Gauss body), until after certain frame number, new Gauss body Replace original background.But for dynamic, illumination variation complex environment high in city, foreground target quantity is very big and moves Slowly, the little change such as illumination in background, shade, leaves shake emerges in an endless stream, and usually occurs that background does not has enough time to have updated The situation that complete environment changes again, mixed Gauss model has occurred as soon as and has constantly built new Gauss body, tired in meter between various changes The result calculated, does not reaches the purpose of detection prospect real-time, accurate.

Bayes method, as the succedaneum of Gaussian modeling, the method employing Density Estimator, recursively uses Bayes Study approximates the probability density distribution of each pixel, instead of the accurate parameters method of estimation of mixed Gauss model.But pattra leaves This method still cannot solve prospect and move discrimination (prospect slowly that will move is mistakenly considered background) slowly；Then, code This algorithm is by pixel value quantization encoding so that the information of neighborhood is added into model, the problem solving slow prospect.But code book Algorithm requires a great deal of time on off-line training, it is difficult to meet the requirement that complicated high dynamic environment is changeable.

The problem that background modeling problem is regarded as a signal reconstruction is the most popular a kind of way.When the prospect in environment When only accounting for very small part, using compressive sensing theory that prospect carries out detection is a kind of effective method, is i.e. by front Scape regards that the noise in signal reconstruction, such background modeling have just become the problem that a main signal amount is recovered as.Meanwhile, also one Planting way is that prospect is regarded as the amount of a temporary transient appearance in the environment, uses sparse expression, prospect is regarded as a sparse amount, Thus from some images in past, recover current background.But, under complicated high dynamic environment, any time or sky Sparse hypothesis between is all non-existent, prospect likely time occupy the biggest ratio in the air.Therefore, there has been one The method recovering Gaussian modeling and signal to be together in series, uses principal component analysis, background is recovered from main feature. But, this method has clearly a need for spending substantial amounts of resource in training and parameter estimation, it is difficult to reach real-time purpose.

Summary of the invention

It is an object of the invention to provide and in complicated high dynamic environment, quickly update background and the method for prospect based on multithreading, solve Certainly existing modeling method cannot tackle complicated high dynamic environment, it is impossible to updates background model, and prospect inspection in real time, exactly The accuracy surveyed and the highest problem of adaptability.

For solving above-mentioned technical problem, the present invention by the following technical solutions:

A kind of in complicated high dynamic environment, quickly update background and the method for prospect based on multithreading, comprise the following steps:

Step one, carries out Haar-like feature extraction to pixel and vicinity points thereof；

Step 2, according to the complexity of scene, is directly incorporated into background model matrix by the characteristic vector of front 10～35 two field pictures End；

Step 3, obtains a certain pixel in image, calculates this pixel P2M distance to its background model, sentence with this The most not this pixel is foreground point；

Step 4, if showing that current pixel point is background dot by step 3, then update in background model matrix with current P2M The eigenvalue that distance is minimum；

Step 5, if showing that current pixel point is background dot by step 3, random from the pixel that current pixel point is neighbouring Choose a point, update feature maximum with current pixel feature P2M distance in its background model.

Further technical scheme is, in described step one, Haar-like feature extraction concrete grammar is:

Obtain the block of pixels vector P that in current frame image, kth pixel and its neighbor point are constituted_k, utilize integrogram, by image Arrive each rectangular area pixel sum of being formed of point from the off to save, when to calculate certain as the element of an array The pixel in individual region and time direct index array in the value of corresponding point, obtain multiplier by adding to cut algorithm, by compressed sensing matrix with The problem of block of pixels vector multiplication is transformed into the multiplier that integrogram obtains and is multiplied with weights the problem sued for peace again, thus gets compression Perception matrix A, and then obtain characteristic vector v of kth pixel_k=AP_k。

Further technical scheme is, described compressed sensing matrix(n ＞ m), wherein n is the dimension of object vector, Being i.e. the length of image block centered by current pixel, m is the characteristic dimension after compressed sensing, is i.e. background model square The line number of battle array.

Further technical scheme is, in described step 2, it is judged that if current frame image belongs to front N frame, then kth picture The background model matrix of vegetarian refreshmentsIt is expressed as M_k={ v_k,1,v_k,2,…,v_k,N, whereinRepresent kth pixel The characteristic vector of point the i-th frame.

Further technical scheme is, in described step 3, it determines the most whether this pixel is that the method for foreground point is:

Use I_kRepresent the pixel value of kth pixel in current frame image, useRepresent its background model, use v_kRepresent The characteristic vector of this pixel, uses v_k,lRepresent the value (1≤l≤m) of this pixel characteristic vector l dimension, use v_k,i,lRepresent this picture The value (1≤i≤N, 1≤l≤m) of the characteristic vector l dimension of i-th in vegetarian refreshments background model.So minimum P2M distance definition is

$Min\_P2M(I_k,M_k)=\underset{l=1}{\overset{m}{\Σ}}\underset{i\∈\{1,2,\mathrm{...},N\}}{\min }{(v_{k,l}-v_{k,i,l})}^2$

If met

Min_P2M(I_k, M_k) ＞ Threshold

So, it is believed that this pixel is foreground point, wherein Threshold is the constant manually specified, by international background The F-Measure of modelling effect assesses:

$F=2\left(\frac{precision\·recall}{precision+recall}\right),0\≤F\≤1$

Wherein, precision is the accuracy rate of foreground detection, and recall is the capture rate of foreground detection, and F-Measure is the biggest, prospect Detection results is the best.

Further technical scheme is, when described current pixel point is background dot, be updated be divided into pixel context update and Neighborhood territory pixel point context update,

Wherein said pixel background update method is: use v_k,i,lRepresent in this kth pixel background model the feature of i-th to The value (1≤i≤N, 1≤l≤m) of flow control l dimension, uses v_k,lRepresent the value (1≤l≤m) that this pixel characteristic vector l ties up, utilize Formula

${v_{k,i,l}}^{new}=v_{k,l},(i=\underset{i\∈\{1,2,\mathrm{...},N\}}{\arg \min }{\left(v_{k,l}-v_{k,i,l}\right)}^2)$

Draw characteristic vector v that this pixel and its neighbouring block are new_k,i,l ^new；

Wherein said neighborhood territory pixel point background update method is: we randomly select a picture from 8 neighborhoods of kth pixel Vegetarian refreshments, if it is jth pixel in image, uses I_jRepresent the pixel value of this pixel, useRepresent its background mould Type, uses v_jRepresent the characteristic vector of this pixel, use v_j,lRepresent the value (1≤l≤m) of this pixel characteristic vector l dimension, use v_j,i,lRepresent the value (1≤i≤N, 1≤l≤m) of the characteristic vector l dimension of i-th in this pixel background model, maximum P2M away from From for

$Max\_P2M(I_j,M_j)=\underset{l=1}{\overset{m}{\Σ}}\underset{i\∈\{1,2,\mathrm{...},N\}}{\max }{(v_{j,l}-v_{j,i,l})}^2$

The context update of neighborhood random point, utilizes formula,

${v_{j,i,l}}^{new}=v_{j,l},(i=\underset{i\∈\{1,2,\mathrm{...},N\}}{\arg \min }{\left(v_{j,l}-v_{j,i,l}\right)}^2)$

Obtain new pixel this vector v special_k,i,l ^new, wherein v_j,i,lFor the characteristic vector l of i-th in this jth pixel background model The value (1≤i≤N, 1≤l≤m) of dimension, v_j,lValue (1≤l≤m) for this pixel characteristic vector l dimension.

Compared with prior art, the invention has the beneficial effects as follows:

Use pixel to the thought of model, by single pixel with a series of partial descriptions device character representations based on compressed sensing, And weigh whether pixel is foreground point by the distance of point to class；Meanwhile, carrying out background model renewal when, also using Model smoothing that local describer is constituted by point to the distance of class and be effectively updated, so that background modeling and prospect Detection, no matter at indoor or outdoor complex environment, suffers from quickly efficient performance；Background can be updated in real time, exactly, Adapt to various complex environment, be effectively improved accuracy and the adaptability of foreground detection.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the present invention.

Fig. 2 is the schematic diagram of Haar-like feature extraction.

Fig. 3 is that smallest point updates the schematic diagram of background model to class distance.

Fig. 4 is the schematic diagram that maximum point arrives class distance renewal background model.

Detailed description of the invention

In order to make the purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, to this Bright it is further elaborated.Should be appreciated that specific embodiment described herein, and need not only in order to explain the present invention In limiting the present invention.

The present invention can realize on Windows and Linux platform, and programming language is also alternatively, C++ can be used Realize, the method employing multithreading.

Fig. 1 shows the one of a kind of method quickly updating background and prospect based on multithreading in complicated high dynamic environment of the present invention Individual embodiment: a kind of quickly update background and the method for prospect in complicated high dynamic environment based on multithreading, comprises the following steps:

Step one, carries out Haar-like feature extraction to pixel and vicinity points thereof；

Step 2, according to the complexity of scene, is directly incorporated into background model matrix by the characteristic vector of front 10～35 two field pictures End；

Step 3, obtains a certain pixel in image, calculates this pixel P2M distance to its background model, sentence with this The most not this pixel is foreground point；

Step 4, if showing that current pixel point is background dot by step 3, then update in background model matrix with current P2M The eigenvalue that distance is minimum；

Step 5, if showing that current pixel point is background dot by step 3, random from the pixel that current pixel point is neighbouring Choose a point, update feature maximum with current pixel feature P2M distance in its background model.

Fig. 2 shows the one of a kind of method quickly updating background and prospect based on multithreading in complicated high dynamic environment of the present invention Individual preferred embodiment, in described step one, Haar-like feature extraction concrete grammar is:

Obtain the block of pixels vector P that in current frame image, kth pixel and its neighbor point are constituted_k, utilize integrogram (integrogram It is used to calculate compressed sensing matrix A), image is arrived from the off the rectangular area pixel sum work that each point is formed Be that the element of an array saves, when to calculate certain region pixel and time direct index array in the value of corresponding point, logical Cross that adding cuts algorithm obtains multiplier, the problem of compressed sensing matrix Yu block of pixels vector multiplication is transformed into multiplier that integrogram obtains and Weights are multiplied the problem sued for peace again, thus get compressed sensing matrix A, and then obtain characteristic vector v of kth pixel_k =AP_k。

Another of method quickly updating background and prospect according to the present invention based on multithreading in complicated high dynamic environment is the most real Execute example, described compressed sensing matrix(n ＞ m), wherein n is the dimension of object vector, is i.e. to be with current pixel The length of the image block of the heart, m is the characteristic dimension after compressed sensing, is i.e. the line number of background model matrix；According to experiment, For outdoor scene, it is preferable that m preferably greater than 5, n preferably greater than 32, m are set to 3, n is set to effect when 25.

Another of method quickly updating background and prospect according to the present invention based on multithreading in complicated high dynamic environment is the most real Execute example, in described step 2, it is judged that if current frame image belongs to front N frame, then the background model matrix of kth pixelIt is expressed as M_k={ v_k,1,v_k,2,…,v_k,N, whereinRepresent the characteristic vector of kth pixel the i-th frame； According to background modeling methods such as such as CodeBook and PBAS currently existed, and the experiment at scene, background model needs Carry out the training of 10～35 frames, for common outdoor scene and traffic scene, train 20 frames enough.

Another according to a kind of method quickly updating background and prospect based on multithreading in complicated high dynamic environment of the present invention is excellent Select embodiment, in described step 3, it determines the most whether this pixel is that the method for foreground point is:

Use I_kRepresent the pixel value of kth pixel in current frame image, useRepresent its background model, use v_kRepresent The characteristic vector of this pixel, uses v_k,lRepresent the value (1≤l≤m) of this pixel characteristic vector l dimension, use v_k,i,lRepresent this pixel The value (1≤i≤N, 1≤l≤m) of the characteristic vector l dimension of i-th in some background model.So minimum P2M distance definition is

$Min\_P2M(I_k,M_k)=\underset{l=1}{\overset{m}{\Σ}}\underset{i\∈\{1,2,\mathrm{...},N\}}{\min }{(v_{k,l}-v_{k,i,l})}^2$

If met

Min_P2M(I_k, M_k) ＞ Threshold

So, it is believed that this pixel is foreground point, wherein Threshold is the constant manually specified, by international background The F-Measure of modelling effect assesses:

$F=2\left(\frac{precision\·recall}{precision+recall}\right),0\≤F\≤1$

Wherein, precision is the accuracy rate of foreground detection, and recall is the capture rate of foreground detection, and F-Measure is the biggest, front Scape Detection results is the best, according to experiment, as Threshold=3000, can obtain reasonable effect.

Another of method quickly updating background and prospect according to the present invention based on multithreading in complicated high dynamic environment is the most real Execute example, when described current pixel point is background dot, is updated and is divided into pixel context update and neighborhood territory pixel point context update,

Wherein said pixel background update method is: as it is shown on figure 3, use v_k,i,lRepresent in this kth pixel background model The value (1≤i≤N, 1≤l≤m) of the characteristic vector l dimension of i-th, uses v_k,lRepresent the value that this pixel characteristic vector l ties up (1≤l≤m), utilizes formula

${v_{k,i,l}}^{new}=v_{k,l},(i=\underset{i\∈\{1,2,\mathrm{...},N\}}{\arg \min }{\left(v_{k,l}-v_{k,i,l}\right)}^2)$

Draw characteristic vector v that this pixel and its neighbouring block are new_k,i,l ^new；

Wherein said neighborhood territory pixel point background update method is: as shown in Figure 4, we from 8 neighborhoods of kth pixel with A pixel chosen by machine, if it is jth pixel in image, uses I_jRepresent the pixel value of this pixel, useGeneration Its background model of table, uses v_jRepresent the characteristic vector of this pixel, use v_j,lRepresent the value of this pixel characteristic vector l dimension (1≤l≤m), uses v_j,i,lRepresent the value (1≤i≤N, 1≤l≤m) of the characteristic vector l dimension of i-th in this pixel background model, Maximum P2M distance is

$Max\_P2M(I_j,M_j)=\underset{l=1}{\overset{m}{\Σ}}\underset{i\∈\{1,2,\mathrm{...},N\}}{\max }{(v_{j,l}-v_{j,i,l})}^2$

The context update of neighborhood random point, utilizes formula,

${v_{j,i,l}}^{new}=v_{j,l},(i=\underset{i\∈\{1,2,\mathrm{...},N\}}{\arg \min }{\left(v_{j,l}-v_{j,i,l}\right)}^2)$

Obtain new pixel this vector v special_k,i,l ^new, wherein v_j,i,lFor the characteristic vector of i-th in this jth pixel background model The value (1≤i≤N, 1≤l≤m) of l dimension, v_j,lValue (1≤l≤m) for this pixel characteristic vector l dimension.

Claims (4)

1. one kind quickly updates background and the method for prospect based on multithreading in complicated high dynamic environment, it is characterised in that: include Following steps:

Step one, carries out Haar-like feature extraction to pixel and vicinity points thereof；

Step 2, according to the complexity of scene, is directly incorporated into background model matrix by the characteristic vector of front 10～35 two field pictures End；

Step 3, obtains a certain pixel in image, calculates this pixel P2M distance to its background model, sentence with this The most not this pixel is foreground point；

Step 4, if showing that current pixel point is background dot by step 3, then update in background model matrix with current P2M The eigenvalue that distance is minimum；

Step 5, if showing that current pixel point is background dot by step 3, random from the pixel that current pixel point is neighbouring Choose a point, update feature maximum with current pixel feature P2M distance in its background model；

In described step one, Haar-like feature extraction concrete grammar is:

Obtain the block of pixels vector P that in current frame image, kth pixel and its neighbor point are constituted_k, utilize integrogram, by image Arrive each rectangular area pixel sum of being formed of point from the off to save, when to calculate certain as the element of an array The pixel in individual region and time direct index array in the value of corresponding point, obtain multiplier by adding to cut algorithm, by compressed sensing matrix with The problem of block of pixels vector multiplication is transformed into the multiplier that integrogram obtains and is multiplied with weights the problem sued for peace again, thus gets compression Perception matrix A, and then obtain characteristic vector v of kth pixel_k=AP_k；

In described step 3, it determines the most whether this pixel is that the method for foreground point is:

Use I_kRepresent the pixel value of kth pixel in current frame image, useRepresent its background model, use v_kRepresent The characteristic vector of this pixel, uses v_k,lRepresent the value (1≤l≤m) of this pixel characteristic vector l dimension, use v_k,i,lRepresent this picture The value (1≤i≤N, 1≤l≤m) of the characteristic vector l dimension of i-th in vegetarian refreshments background model；So minimum P2M distance definition is $Min\_P2M(I_k,M_k)=\underset{l=1}{\overset{m}{\Σ}}\underset{i\∈\left\{1,2,...,N\right\}}{\min }{(v_{k,l}-v_{k,i,l})}^2$

If met

Min_P2M(I_k, M_k) ＞ Threshold

So, it is believed that this pixel is foreground point, wherein Threshold is the constant manually specified, by international background The F-Measure of modelling effect assesses:

$\begin{array}{cc}F=2\left(\frac{precision\·recall}{precision+recall}\right)& 0\≤F\≤1\end{array}$

Wherein, precision is the accuracy rate of foreground detection, and recall is the capture rate of foreground detection, and F-Measure is the biggest, prospect Detection results is the best.

The most according to claim 1 in complicated high dynamic environment, quickly update background and the method for prospect based on multithreading, It is characterized in that: described compressed sensing matrix(n ＞ m), wherein n is the dimension of object vector, is i.e. with current picture The length of the image block centered by element, m is the characteristic dimension after compressed sensing, is i.e. the line number of background model matrix.

The most according to claim 1 in complicated high dynamic environment, quickly update background and the method for prospect based on multithreading, It is characterized in that: in described step 2, it is judged that if current frame image belongs to front N frame, the then background model of kth pixel MatrixIt is expressed as M_k={ v_k,1,v_k,2,…,v_k,N, whereinRepresent the feature of kth pixel the i-th frame to Amount.

The most according to claim 1 in complicated high dynamic environment, quickly update background and the method for prospect based on multithreading, It is characterized in that: when described current pixel point is background dot, be updated and be divided into pixel context update and neighborhood territory pixel point background Update,

Wherein said pixel background update method is: use v_k,i,lRepresent in this kth pixel background model the feature of i-th to The value (1≤i≤N, 1≤l≤m) of flow control l dimension, uses v_k,lRepresent the value (1≤l≤m) that this pixel characteristic vector l ties up, utilize Formula

$\begin{array}{cc}{v_{k,i,l}}^{new}=v_{k,l}& \left(i=\underset{i\∈\left\{1,2,...,N\right\}}{\arg \min }{\left(v_{k,l}-v_{k,i,l}\right)}^2\right)\end{array}$

Draw characteristic vector v that this pixel and its neighbouring block are new_k,i,l ^new；

Wherein said neighborhood territory pixel point background update method is: we randomly select a picture from 8 neighborhoods of kth pixel Vegetarian refreshments, if it is jth pixel in image, uses I_jRepresent the pixel value of this pixel, useRepresent its background mould Type, uses v_jRepresent the characteristic vector of this pixel, use v_j,lRepresent the value (1≤l≤m) of this pixel characteristic vector l dimension, use v_j,i,lRepresent the value (1≤i≤N, 1≤l≤m) of the characteristic vector l dimension of i-th in this pixel background model, maximum P2M away from From for

$Max\_P2M(I_j,M_j)=\underset{l=1}{\overset{m}{\mathrm{\Σ}}}\underset{i\∈\left\{1,2,...,N\right\}}{\max }{(v_{j,l}-v_{j,i,l})}^2$

The context update of neighborhood random point, utilizes formula,

$\begin{array}{cc}{v_{j,i,l}}^{new}=v_{j,l}& \left(i=\underset{i\∈\left\{1,2,...,N\right\}}{\arg \min }{\left(v_{j,l}-v_{j,i,l}\right)}^2\right)\end{array}$

Obtain new pixel characteristic vector v_k,i,l ^new, wherein v_j,i,lFor the characteristic vector l of i-th in this jth pixel background model The value (1≤i≤N, 1≤l≤m) of dimension, v_j,lValue (1≤l≤m) for this pixel characteristic vector l dimension.