CN103077521B - A kind of area-of-interest exacting method for video monitoring - Google Patents

Abstract

The invention provides the extracting method of area-of-interest in a kind of video image, be applicable to field of video monitoring.Described method can by area-of-interest automatic mark out, specifically comprise the following steps: 1) adopt binocular camera (in advance demarcate) to carry out video acquisition 2) correct image to left and right two viewpoint, eliminate distortion 3) left and right viewpoint selection one, its depth map is calculated by binocular parallax, and build its Gaussian mixture model-universal background model 4 by consecutive image) use the Depth contrasts of the overall situation to obtain degree of depth salient region, use the method for background difference to obtain moving region simultaneously, obtain two width bianry images 5) XOR is done to the two width bianry images that step 4 obtains, then connected region (denoising) is screened, and expansive working 6 is done to remaining connected region) finally mark profile and boundary rectangle, obtain final area-of-interest.The present invention effectively overcomes color and brightness changes the impact detected area-of-interest, has very high accuracy and robustness.

Description

A kind of area-of-interest exacting method for video monitoring

Technical field

The invention belongs to technical field of image detection, relate to a kind of area-of-interest exacting method that may be used for video monitoring.

Background technology

For video and image, user is only interested in subregion wherein, and this part region comprises larger quantity of information, can more significant attracting attention.Use computing machine to obtain area-of-interest automatically and accurately, have a wide range of applications in fields such as image detection, target identification, tracking.The extracting method of existing a lot of area-of-interest at present, such as frame difference method and the background subtraction method [1] of moving target recognition, based on the method for extracting region [2] that saliency is analyzed, and some detection algorithms for certain objects (as HOG pedestrian detection [3]), these methods can extract the area-of-interest with certain discrimination.But frame difference and background differential technique can only detect foreground moving object, lack the recognition capability to static target; The region detection of conspicuousness is easily subject to the impact of color and brightness, and Detection results is also unstable; Detection algorithm for certain objects identifies often through the sorter trained, and it is relatively more accurate to detect, but often can only for specific objective, and range of application is narrower and be difficult in real time.

List of references

【1】Zivkovic,Z.and F.van der Heijden(2006).Efficient adaptive density estimation per image pixelfor thetask ofbackground subtraction[J].Patternrecognition letters 27(7):773-780.

【2】Itti L,Koch C andNiebur E.Amodel ofsaliency-based visual attention forrapid scene analysis[J].IEEETransaction on PatternAnalysis and Machine Intelligence.1998,20(11):1254-1259.

【3】N.Dalal,B.Triggs,Histograms oforiented gradients forhuman detection[C].In:ComputerVisionand Pattern Recognition,San Diego,CA,June 20–25,2005.

Summary of the invention

The object of the invention is to the above-mentioned deficiency overcoming prior art, provide a kind of new area-of-interest exacting method, it can extract the area-of-interest of dynamic and static state in video image simultaneously, and effectively can overcome the impact of color and luminance factor.

Technical scheme of the present invention is:

For an area-of-interest exacting method for video monitoring, comprise the following steps:

1) use the binocular camera demarcated in advance, or binocular camera is connected the binocular video image that computer collection obtains left and right viewpoint;

2) according to the inside and outside parameter of video camera or binocular camera, binocular is done to the binocular video image of left and right viewpoint and corrects, eliminate distortion, make the alignment that two width images can be strict;

3) one of them selecting left and right viewpoint is as referenced viewpoints, BM pattern matching algorithm is used to calculate binocular parallax to obtain depth map, and utilize the consecutive image sequence of this viewpoint, re-use the background model that gauss hybrid models (GMM) sets up this viewpoint;

4) according to obtained depth map, use the degree of depth squared difference sum of overall weighting to calculate all degree of depth and significantly scheme wherein, DS _ifor the degree of depth conspicuousness that pixel i is corresponding, d _iand d _jfor the degree of depth that pixel i and pixel j is corresponding, p _i=(x _i, y _i) and p _j=(x _j, y _j) be pixel i, j position, || d _i-d _j|| in image between pixel i and pixel j Euclidean distance, ω (p _i, p _j) be the weight relevant with two pixel Euclidean distances; Threshold value T is greater than in the remarkable figure of selected depth ₁pixel DS _i> T ₁, obtain bianry image I ₁;

5) according to obtained background model, background subtraction method is used to obtain the bianry image I representing target ₂, obtain prospect area-of-interest;

6) by I ₁, I ₂do XOR, obtain a secondary new image I ₃;

7) to I ₃in all connected regions screen, remove and be less than the isolated area of certain pixel count.And expansive working is done to remaining area, obtain regional ensemble R;

8) R is area-of-interest set ROI, mark profile and boundary rectangle, and result is indicated in display.

Preferably, step 5) in, when meeting following formula, think that pixel belongs to prospect area-of-interest:

|i(t)-bg(t)|≥T ₂

Wherein, i (t) represents the pixel in t frame, and bg (t) is the background model of t frame, T ₂for threshold value.

The present invention, by extracting degree of depth salient region and motion target area, both merging, marks final region after screening.Can not only moving region be extracted, and can stagnant zone be extracted, avoid in classic method color and luminance factor to the impact of area-of-interest simultaneously.The present invention can mark various foreground target region accurately, also compares to meet human eye and select the subjectivity of area-of-interest.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the area-of-interest exacting method for video monitoring of the present invention.

Binocular camera example in Fig. 2 the present invention.

Embodiment

For making object of the present invention, implementation and advantage more clear, below the present invention is described in further detail.Method flow of the present invention as shown in Figure 1.

Video acquisition can use the binocular camera on market, can designed, designed binocular camera, and camera parallax range can be set to 100mm ~ 200mm, as Fig. 2.Binocular calibration can adopt plane gridiron pattern calibrating template, coordinate corresponding relation according to template angle point and image mid point obtains (focal length, imaging initial point, distortion factor) in binocular camera, outer (rotation matrix and translation vector) parameter.

It is the monocular internal reference data (focal length, imaging initial point, distortion factor) and binocular relative position relation (rotation matrix and translation vector) that obtain after calibrating according to camera that binocular corrects, respectively elimination distortion is carried out to left and right view and row is aimed at, such that the imaging origin of left and right view is consistent, two camera optical axises are parallel, left and right imaging plane is coplanar, to polar curve row alignment.Binocular corrects and uses BOUGUET method, more specifically adopts the cvStereoRectify in opencv, cvInitUndistortRectifyMap and cvRemap function.

The present invention tests and adopts left viewpoint as referenced viewpoints (also can use right viewpoint), uses BM pattern matching algorithm to calculate binocular parallax to obtain depth map., as Matching power flow, SAD window size is 5 to use absolute value difference sum (sum of absolute differences, SAD) in the present invention; Use the parallax between the viewpoint of BM (Boyer-Moore) searching algorithm calculating left and right, minimum parallax is set to 25 pixels.More specifically be adopt the cvFindStereoCorrespondenceBM function in opencv.

Consecutive image based on left viewpoint calculates Gaussian Mixture background, and Gaussian mixture model-universal background model adopts 5 Gaussian distribution, and initial variance is 30, and the threshold value of Gaussian distribution weight sum is 0.7, and learning rate is 0.05; The parameter using EM algorithm to carry out Gauss model upgrades.

Obtaining after based on the depth map of left viewpoint and Background, calculating and significantly scheming based on the degree of depth of global depth contrast, calculate sport foreground.

Significantly schemed by formula 1 compute depth: ${\mathrm{DS}}_i=\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\left|\right|d_i-d_j\left|\right|}^2\mathrm{\ω}(p_i,p_j)$ Formula 1

Wherein DS _ifor the degree of depth conspicuousness that pixel i is corresponding, d _iand d _jfor the degree of depth that pixel i and pixel j is corresponding, p _i=(x _i, y _i) and p _j=(x _j, y _j) be pixel i, j position, ω (p _i, p _j) be the weight relevant with two pixel Euclidean distances. the Euclidean distance definition of pixel i, j is as formula 2: $D(i,j)=\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}$ Formula 2

Use threshold method after obtaining degree of depth conspicuousness, selected depth conspicuousness is greater than threshold value T ₁all pixels, obtain the bianry image I representing degree of depth conspicuousness ROI ₁.

Sport foreground pixel is obtained by formula 3: | i (t)-bg (t) |>=T ₂formula 3

Wherein, i (t) represents the pixel in t frame, and bg (t) is the background model of t frame, T ₂for threshold value. obtain all pixels meeting formula 2, form the bianry image I representing motion target area ₂

The expression degree of depth salient region image I that the present invention is obtaining ₁with sport foreground area image I ₂after, both are done XOR, obtains the bianry image I merged ₃.

Use 5 to take advantage of the mean filter of 5 to remove noise spot, use expansive working to obtain final connected region set R.More particularly adopt the blur function in opencv and dilate function, then the threshold function check edge in opencv is used, use findContours function to find profile, use boundingRect function to obtain the boundary rectangle of all connected regions.

Claims (2)

1., for an area-of-interest exacting method for video monitoring, comprise the following steps:

1) use the binocular camera demarcated in advance, or binocular camera is connected the binocular video image that computer collection obtains left and right viewpoint;

2) according to the inside and outside parameter of video camera or binocular camera, binocular is done to the binocular video image of left and right viewpoint and corrects, eliminate distortion, make the alignment that two width images can be strict;

3) one of them selecting left and right viewpoint is as referenced viewpoints, BM pattern matching algorithm is used to calculate binocular parallax to obtain depth map, and utilize the consecutive image sequence of this viewpoint, re-use the background model that gauss hybrid models (GMM) sets up this viewpoint;

4) according to obtained depth map, use the degree of depth squared difference sum of overall weighting to calculate all degree of depth and significantly scheme wherein, DS _ifor the degree of depth conspicuousness that pixel i is corresponding, d _iand d _jfor the degree of depth that pixel i and pixel j is corresponding, p _i=(x _i, y _i) and p _j=(x _j, y _j) be pixel i, j position, || d _i-d _j|| in image between pixel i and pixel j Euclidean distance, ω (p _i, p _j) be the weight relevant with two pixel Euclidean distances; Threshold value T is greater than in the remarkable figure of selected depth ₁pixel DS _i> T ₁, obtain bianry image I ₁;

5) according to obtained background model, background subtraction method is used to obtain the bianry image I representing target ₂, obtain prospect area-of-interest;

6) by I ₁, I ₂do XOR, obtain a secondary new image I ₃;

7) to I ₃in all connected regions screen, remove and be less than the isolated area of certain pixel count, and expansive working is done to remaining area, obtain regional ensemble R;

8) R is area-of-interest set ROI, mark profile and boundary rectangle, and result is indicated in display.

2. the area-of-interest exacting method for video monitoring according to claim 1, is characterized in that, step 5) in, when meeting following formula, think that pixel belongs to prospect area-of-interest:

|i(t)-bg(t)|≥T ₂

Wherein, i (t) represents the pixel in t frame, and bg (t) is the background model of t frame, T ₂for threshold value.