CN100382600C - Detection method of moving object under dynamic scene - Google Patents

Abstract

The present invention provides a detection method of a moving object under a dynamic scene. The pixel process is modeled by adopting a kernel density estimation (KDE) function for the state that a scene is not completely static under the existence of movement in a small scale, and the distribution of the pixel grey scale of a video image is calculated by utilizing the theory of nonparameter probability density estimation. A time-domain diversity sample set for model training is obtained from a primitive training sequence in the method of the present invention. Primitive training data does not need to be saved nor used in the processes of background extraction and the detection of the moving object, storage space is saved, time consuming repetitive calculation is avoided, and the real-time position and the shape of the moving object can be obtained. The method of the present invention has the advantages of simple and effective realization, high universality and high pertinence.

Description

Moving object detection method under the dynamic scene

Technical field

The present invention relates to the moving object detection method under a kind of dynamic scene, be mainly used in higher level video analysis such as video monitoring system moving object classification, tracking, belong to technical field of video processing.

Background technology

It is video monitor, man-machine interaction that moving object detects, the key issue that system videos such as traffic monitoring are analyzed, and its result is generally used for higher level analysis and processing such as target following, classification.The validity of detection method and robustness are very crucial to the whole video system.As a kind of effective ways of motion detection, background frames differential technique (Background Subtraction) deducts scene background to obtain sport foreground from present image, has accurate location and does not enlarge advantage such as moving region.Usually hypothesis complete transfixion of background or background can not have the so-called empty background of moving object to obtain by shooting.And in fact, in real application systems such as traffic monitoring, do not comprise that moving object or complete static constant background image can't obtain.Therefore, the background frames differential technique detects moving object to be needed at first dynamically to extract background model from the sequence image of the scene complexity that comprises moving object, and this model must have to the robustness of environmental change with to the high sensitivity of moving object.

Traditional parameter model treats that by prior supposition certain specific character of estimated background obeys certain statistical model, distributes as Gaussian Profile or mixed Gaussian, and the computation model parameter obtains background model then.This will relate to problems such as model parameter estimation and parameter optimization, and these problems are usually directed to the very big value algorithm of the expectation of calculation of complex.In addition, because the actual scene complexity, as have factors such as tree branches and leaves shakes, the real background model profile is unknown and can not suppose in advance, can't obtain treating the priori of estimated background pixel process at all.Therefore, parametric technique is not suitable for video monitoring system.In recent years, nonparametric model method (Elgammal A. by people such as Elgammal proposition, Harwood D., and Davis L., Non-parametric model for backgroundsubtraction, The 6th European Conference on Computer Vision, Dublin, Ireland, 2000, page 751-767) directly from data, estimates unknown density function, avoided the problems such as Estimation Optimization of model form supposition and distributed constant.Yet the nonparametric model method that Elgammal proposes adopts the data in whole sample sets to be used for the model estimation, need preserve all sample datas in testing process.And this method is carried out identical processing to different samples, each sample is considered as identical to the effect of density Estimation, adopts the homogeneous weights in density calculation, needs the many identical or similar samples of double counting like this in the background extracting process.

Summary of the invention

The objective of the invention is to actual needs at the deficiency and the video monitoring system of above-mentioned technology, moving object under a kind of dynamic scene detection method is provided, do not need to suppose in advance the distribution form of background, avoid information redundancy and double counting in background density is estimated, the multi-modal model of being set up of diversity sample nonparametric can be handled the complicated and incomplete static situation of scene, for higher level video analysis is established solid technical foundation as systems such as tracking, classification.

For achieving this end, not exclusively static and have a situation of little motion at scene, the present invention at first passes through the diversity principle, extracts to have higher occurrence frequency in the sample set and have multifarious sample from original training sequence, keeps the important information in the training image sequence.According to nonparametric probability density estimation theory, adopt the plain process of Density Estimator (KDE, kernel density estimation) function object to carry out the distribution that video image pixel gray scale is estimated in modeling then.Last passing thresholdization obtains the two-value mask of moving object, obtains the position and the shape of moving object.

Dynamic scene moving object detection method proposed by the invention mainly comprises four parts: the extraction of diversity sample set, nuclear estimate that window width calculating, Density Estimator and moving object two-value mask calculate.Concrete steps are as follows:

1) the diversity sample set is chosen.Gather one group of continuous shooting and comprise the video sequence image (N frame) of moving object as the original training sample collection, from N time domain value histogram of each pixel, alternately choose and have the highest occurrence frequency and form new sample set with selected sample has a maximum difference under Euclidean distance sample, be the former sample number in the positive and negative unit of the center calculation gray scale interval simultaneously with the new samples, the different weights that obtain new samples are used for Density Estimator;

2) nuclear estimates that window width calculates.After obtaining new samples collection and corresponding weights, also need to obtain examining the window width of each pixel correspondence in the estimation.In background model was estimated, window width mainly will reflect the pixel gray scale because the image blurring jump variation that waits the localized variation rather than the gray scale of generation.Utilize original sample to concentrate the sample absolute difference median (MAD) of each pixel in consecutive frame to calculate, obtain the relation of this pixel window width and sample absolute difference median, thereby try to achieve the window width value of different picture elements.

3) Density Estimator.After utilizing resulting diversity sample, weights and window width, can carry out Density Estimator to current frame image.With the gray scale value substitution Density Estimator function of each pixel of present image, calculate the estimation density of present image pixel.

4) moving object two-value mask calculates.For different image sequences, Xuan Ding a certain threshold value is carried out thresholding to the estimation density that calculates in the step 3 and is handled by experiment.When estimating density greater than threshold value, corresponding picture element is considered as background dot and composes 0, otherwise be considered as the foreground moving object point, tax is 1.The two-value mask that obtains thus can characterize the position and the shape thereof of current time moving object well.

The inventive method does not need to suppose in advance any form of background, has avoided complicated parameter calculation and optimization.No longer need to preserve and use the total data of original training sequence in background extracting and the moving object testing process, saved memory space, avoided double counting consuming time.The realization of the inventive method is simply effective, has good versatility and specific aim.

Description of drawings

Fig. 1 is the FB(flow block) of the moving object detection method under the dynamic scene of the present invention.

Fig. 2 extracts FB(flow block) for diversity sample set of the present invention.

Fig. 3 is the traffic scene original image that the embodiment of the invention adopted.

The moving object testing result that Fig. 4 obtains from original image for the embodiment of the invention.

Embodiment

In order to understand technical scheme of the present invention better, be described in further detail below in conjunction with drawings and Examples.

Fig. 1 is the FB(flow block) of the inventive method.In order to set up the dynamic background model that moving object detects, need N frame continuous images sequence to be used for model training as sample.(x y), needs to extract new diversity sample (M for a certain pixel _{X, y}Individual), and obtain window width σ simultaneously _{X, y}Then current frame image is carried out Density Estimator, and estimated result is carried out thresholding handle, obtain the moving object testing result at last.Fig. 2 is the FB(flow block) that the diversity sample set extracts among Fig. 1.From the histogram of original training sample, at first obtain having the gray value of maximum occurrence frequency, then from the remaining sample of former sample set, choose and selected gray value farthest gray value under Euclidean distance, and then never choose obtain in the sample maximum frequency gray scale and with the gray value farthest of this distance of sampling, so repeatedly up to obtaining required sample number.

The current time original image of a certain traffic scene that the embodiment of the invention adopts as shown in Figure 3, moving object detects specifically to be carried out according to the following steps:

1) extraction of diversity sample set

Gather one group of video sequence image of taking continuously and comprising moving object (N frame) as the original training sample collection, (x, y) the gray scale value in the N frame is S to a certain pixel ₁={ y ₁, y ₂..., y _N.Because S ₁Middle some similar even identical values of existence are so available M is individual at S ₁In have the highest occurrence frequency and have maximum variational value and represent.Specific practice is: at first, calculate S ₁The middle the highest gray value g of the frequency of occurrences ₁:

$g_1=y_{q_1}=\underset{q_1}{\arg \max }(n_{y_1},n_{y_2},\·\·\·,n_{y_p})$

N in the formula _YiExpression gray scale value is y _iTotal sample number, P is the different gray scale value numbers of N sample.Secondly, choose and g ₁Under Euclidean distance, differ gray value g farthest ₂:

$g_2=y_{q_2}=\underset{q_2}{\arg \max }\left(\right|g_1-y_k\left|\right)k=\mathrm{1,2},\·\·\·,P$

Then, from S ₁Obtain the highest gray value g of frequency in the value that is not selected ₃:

$g_3=y_{q_3}=\underset{q_3\≠q_1,q_2}{\arg \max }(n_{y_1},n_{y_2},\·\·\·,n_{y_P})$

New samples { the g that then chooses and obtained ₁, g ₂, g ₃At a distance of farthest value g ₄:

$g_4=y_{q_4}=\underset{q_4\≠q_1,q_2,q_3}{\arg \max }\left(\underset{l=q_1,q_2,q_3}{\min }\left(\right|y_k-g_l\left|\right)\right)k=\mathrm{1,2},\·\·\·,P$

So repeatedly, alternately do not choose frequency in the sample set maximum and with selected to such an extent that sample differs farthest value, until obtaining M _{X, y}The new samples collection of individual sample $S_2=\{g_1,\·\·\·,g_{M_{x,y}}\}.$ Obviously, work as M _{X, y}Be during=N and choose S ₁In whole gray scale values.

For new samples g _i, calculate its weights α by following formula _i

${\mathrm{\α}}_i=\frac{N_i}{N},i=1,\·\·\·,M_{x,y}$

N in the formula _iBe at [g _i-Δ g, g _i+ Δ g] former sample number, M _{X, y}Can (x, the gray scale value in N frame y) be counted P and is obtained by pixel

$M_{x,y}=\left\{\begin{array}{cc}P,& 1\&le;P\&le;K_1\\ \&lsqb;P/(2\mathrm{\&Delta;g}+1)\&rsqb;,& K_1<P\&le;K_2\\ M_{\max },& P>K_2\end{array}\right.$

[P/ (2 Δ g+1)] is for being not more than the smallest positive integral of P/ (2 Δ g+1) in the formula.K ₁, K ₂For testing given parameter, M _MaxMaximum sample number for the new samples collection.

2) nuclear estimates that window width calculates

In background density nuclear was estimated, window width σ reflected that mainly the pixel gray scale waits the localized variation of generation rather than the jump of gray scale to change owing to image blurring.Neighboring pixels is to (y on the time domain _i, y _I+1) derive from identical local distribution usually and have only few cases to derive from cross-distribution.Suppose its local distribution obedience N (μ, σ ²) Gaussian Profile, difference (y so _i-y _I+1) be distributed as Gaussian Profile N (μ, 2 σ ²).Can obtain the sample absolute difference by the symmetry of Gaussian Profile and the definition of sample median | y _i-y _I+1| intermediate value m satisfy

${\&Integral;}_{-\&infin;}^m\frac{1}{\sqrt{2\mathrm{\&pi;}2{\mathrm{\&sigma;}}^2}}{e}^{\frac{u^2}{2{\mathrm{\&sigma;}}^2}}\mathrm{du}=0.25$

Can check in its upside 0.25 quantile Φ (u by the standardized normal distribution table _0.25) be 0.68, so

m＝0+u _0.25(σ)＝0.68σ

Window width σ _{X, y}Can be by sample median m _{X, y}Obtain σ _{X, y}=m _{X, y}/ (0.68 ).

3) Density Estimator

Density Estimator is estimated unknown density distribution by the local function that the weighted average central point is positioned at sampled value.By 1) 2) step obtains the diversity new samples collection { g in pixel features space ₁, g ₂..., g _{Mx, y}, weights α _iAnd window width σ _{X, y}, (x, gray scale value y) is y to pixel in the present image (accompanying drawing 3) _tDensity distribution be p (y _t):

$p\left(y_t\right)={\mathrm{\&Sigma;}}_{i=1}^{M_{x,y}}{\mathrm{\&alpha;}}_i{K}_{{\mathrm{\&sigma;}}_{x,y}}(y_t-g_i)$

K in the formula _σFor window width is the kernel function of σ and satisfies $K_{\mathrm{\&sigma;}}=\left(x\right)=\frac{1}{\mathrm{\&sigma;}}K\left(\frac{x}{\mathrm{\&sigma;}}\right),$ α _iBe normalization weights coefficient ${\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&alpha;}}_i=1.$ Calculate if adopt standardized normal distribution to examine, then the formula of gradation of image distribution is:

$p\left(y_t\right)={\mathrm{\&Sigma;}}_{i=1}^{M_{x,y}}{\mathrm{\&alpha;}}_i\frac{1}{\sqrt{2\mathrm{\&pi;}{\mathrm{\&sigma;}}_{x,y}^2}}{e}^{\frac{{(y_t-g_i)}^2}{2{\mathrm{\&sigma;}}_{x,y}^2}}$

4) moving object two-value mask calculates

After calculating the density Estimation of each pixel among Fig. 3, the two-value mask that can passing thresholdization obtains obtains the position and the shape of moving object.For a certain pixel gray value y _tIf Density Estimator result is less than a certain threshold value th, then this picture element is classified as the foreground point, otherwise is classified as background dot.The moving object testing result can be represented by the two-value mask

$M_t=\left\{\begin{array}{cc}1& \mathrm{if\; p}\left(y_t\right)<\mathrm{th}\\ 0& \mathrm{otherwise}\end{array}\right.$

Fig. 4 is the moving object testing result that is obtained by Fig. 3.Though the contrast of overhead hypograph is very low, moving vehicle has still obtained quite good detecting.Though the pedestrian on Fig. 3 the right is blocked by trees and leaf existence motion among a small circle, pedestrian's position and shape have also obtained good detection.Noise in the testing result mainly is because some state of background is not included in the background model of being set up, and can it be removed to obtain better testing result by noise filtering technique.

Claims (1)

1. the moving object detection method under the dynamic scene is characterized in that comprising the steps:

1) the diversity sample set is chosen: gather one group of continuous shooting and comprise the video sequence image of moving object as the original training sample collection, from the time domain value histogram of each edge element, alternately choose and have the highest occurrence frequency and form new sample set with selected sample has a maximum difference under Euclidean distance sample, be the former sample number in the positive and negative unit of the center calculation gray scale interval simultaneously with the new samples, the different weights that obtain new samples are used for Density Estimator;

2) nuclear estimates that window width calculates: utilize original sample to concentrate the sample absolute difference median of each pixel in consecutive frame, obtain the relation of this pixel window width and sample absolute difference median, thereby try to achieve the window width value of different picture elements;

3) Density Estimator: utilize resulting diversity sample, weights and window width, current frame image is carried out Density Estimator,, calculate the estimation density of present image pixel with the gray scale value substitution Density Estimator function of each pixel of present image;

4) moving object two-value mask calculates: nuclear estimation density is carried out thresholding handle, when estimating density greater than selected a certain threshold value, corresponding picture element is considered as background dot and composes 0, otherwise be considered as the foreground moving object point, tax is 1, and the two-value mask that obtains thus characterizes the position and the shape thereof of current time moving object.

Images