CN106203429A - Based on the shelter target detection method under binocular stereo vision complex background - Google Patents

Abstract

The invention discloses a kind of based on the shelter target detection method under binocular stereo vision complex background, for solving the technical problem of existing shelter target detection method accuracy of detection difference.Technical scheme is first to demarcate binocular camera to obtain the correction chart picture of row alignment, then obtain disparity map by Stereo matching and carry out background modeling, calculate scene three-dimensional coordinate and generate downward projection figure, finally by MeanShift method, downward projection figure is carried out cluster and obtain testing result.The present invention utilizes space three-dimensional information, efficiently solves the technical problem in the monocular vision such as interference of target occlusion, the change of scene light, shade and complex background, improves accuracy of detection.

Description

Based on the shelter target detection method under binocular stereo vision complex background

Technical field

The present invention relates to a kind of shelter target detection method, particularly relate to a kind of based on binocular stereo vision complex background Under shelter target detection method.

Background technology

Traditional moving object detection is mostly based on the method for monocular vision, and for stereoscopic vision, monocular vision has Its advantage, but there is also the biggest defect.The quantity of information of monocular vision is little, the most only need to process piece image, and arithmetic speed is relative Comparatively fast, but image lost the three-dimensional information of actual scene in projection process, therefore has irremediable defect.Adopting Carry out in moving object detection by method based on monocular vision, such as document " Effective Gaussian mixture learning for video background subtraction.Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2005,27 (5): 827-832 " it is usually present target occlusion and week How the problems such as the change of enveloping field scape light and the interference of shade, solve the difficult point of these problems always research.For these Problem many scholars have done numerous studies, put forward algorithm based on target characteristic coupling, multi-tool plate for occlusion issue The algorithm joined, for light change and the problem such as shadow interference, proposes many Gaussian Background model and shadow removing algorithm etc., but this A little methods are affected relatively big by environmental factors, are easy to the phenomenon that target detection is failed occur in actual applications.

Summary of the invention

In order to overcome the deficiency of existing shelter target detection method accuracy of detection difference, the present invention provides a kind of and stands based on binocular Shelter target detection method under body vision complex background.First the method is demarcated binocular camera and is obtained the correction chart of row alignment Picture, then obtains disparity map by Stereo matching and carries out background modeling, calculates scene three-dimensional coordinate and generates downward projection figure, finally By MeanShift method, downward projection figure is carried out cluster and obtain testing result.The present invention utilizes space three-dimensional information, effectively Solve the technical problem in the monocular vision such as interference of target occlusion, the change of scene light, shade and complex background, improve Accuracy of detection.

The technical solution adopted for the present invention to solve the technical problems is: a kind of based under binocular stereo vision complex background Shelter target detection method, be characterized in comprising the following steps:

Step one, binocular camera are demarcated.

Initially with Zhang Zhengyou gridiron pattern scaling method, shoot multiple gridiron pattern images respectively, in order to demarcate two cameras Inner parameter M₁, external parameter M₂The scaling board image calculation on ground it is placed on by shooting.Image coordinate (u, v) and the world Coordinate (X_w,Y_w,Z_w) homogeneous transformational relation as follows:

$\begin{array}{c}{z}_c\left[\begin{array}{c}u\\ v\\ 1\end{array}\right]=\left[\begin{array}{ccc}\frac{1}{d_x}& 0& u_0\\ 0& \frac{1}{d_y}& v_0\\ 0& 0& 1\end{array}\right]\left[\begin{array}{cccc}f& 0& 0& 0\\ 0& f& 0& 0\\ 0& 0& 1& 0\end{array}\right]\left[\begin{array}{cc}{R}_C& T_C\\ \stackrel{\→}{0}& 1\end{array}\right]\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]=\\ =\left[\begin{array}{cccc}{f}_u& 0& u_0& 0\\ 0& f_v& v_0& 0\\ 0& 0& 1& 0\end{array}\right]\left[\begin{array}{cc}{R}_C& T_C\\ \stackrel{\→}{0}& 1\end{array}\right]\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]=M_1{M}_2\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]\end{array}---\left(1\right)$

Wherein, f is camera focus, (u₀,v₀) it is figure principal point coordinate, d_xAnd d_yRepresent respectively each pixel at transverse axis x and Physical coordinates on longitudinal axis y.

Then stereo calibration calculates spatially two video camera P₁And P₂Geometrical relationship, the rotation between i.e. two cameras Matrix R and translation matrix T.Select right camera as reference camera.Relation is as follows:

P₁=R* (P₂-T) (2)

The correction chart picture of row alignment is obtained finally by non-demarcation three-dimensional correction HartLey algorithm.Requirement binocular camera is adopted Collection image synchronization.

Step 2, Stereo matching obtain parallax.

By the match point between the camera view of binocular solid matching primitives left and right, obtain disparity map, select Gaussian Mixture Modeling method carries out background modeling to disparity map, eliminates the complex background interference to target detection.According to parallax, baseline and interior Ginseng, uses trigonometric calculations scene three-dimensional coordinate.Choose the world coordinate system for XOY face with ground, three-dimensional point is projected to Ground, projects to the number of three-dimensional point of certain pixel as the color value of this pixel, generates downward projection figure.

Step 3, downward projection figure cluster.

Probability density at x is f_h,k(x):

$f_{h,k}\left(x\right)=\underset{i=1}{\overset{n}{\Σ}}K\left(\right|\left|\frac{x-x_i}{h}\right|\left|\right)---\left(3\right)$

Wherein, K (x) is kernel function, and h is radius

F to be made_h,kX () is maximum, to f_h,kX () derivation obtainsWherein g (s)=-k'(s),

$\begin{array}{c}\▿f_{h,k}=\underset{i=1}{\overset{n}{\Σ}}(x-x_i)g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)\\ =\[\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)\]\[\frac{\underset{i=1}{\overset{n}{\Σ}}{x}_{i}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}{\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}-x\]\end{array}---\left(4\right)$

Order:

$m_{h,g}\left(x\right)=\frac{\underset{i=1}{\overset{n}{\Σ}}{x}_{i}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}{\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}-x---\left(5\right)$

MakeAnd if only if m_h,g(x)=0, the central coordinate of circle that must make new advances:

$x=\frac{\underset{i=1}{\overset{n}{\Σ}}{x}_{i}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}{\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}---\left(6\right)$

Due to the particularity of projection, only consider that the distance of pixel cannot obtain cluster result accurately, calculating probability During density, need to meet: the color value of (a) pixel is the most close with the color value of central pixel point, and probability density is the highest；(b) From the position of central point more close to pixel, probability density is the highest.Therefore, kernel function K is selected_h(x):

$K_h\left(x\right)=K\left(\right|\left|\frac{x^s-x_i^s}{h}\right|\left|\right)*K\left(\right|\left|\frac{x^r-x_i^r}{h}\right|\left|\right)---\left(7\right)$

After MeanShift cluster, each class represents a target.This result is projected to display in former right image final Testing result.

The invention has the beneficial effects as follows: first the method is demarcated binocular camera and obtained the correction chart picture of row alignment, then leads to Cross Stereo matching to obtain disparity map and carry out background modeling, calculate scene three-dimensional coordinate and generate downward projection figure, finally use MeanShift method carries out cluster to downward projection figure and obtains testing result.The present invention utilizes space three-dimensional information, effectively solves Determine the technical problem in the monocular vision such as interference of target occlusion, the change of scene light, shade and complex background, improve Accuracy of detection.

Below in conjunction with detailed description of the invention, the present invention is elaborated.

Detailed description of the invention

The present invention specifically comprises the following steps that based on the shelter target detection method under binocular stereo vision complex background

Step one, binocular camera are demarcated.

Initially with the gridiron pattern scaling method of Zhang Zhengyou, about 20 gridiron pattern images of shooting respectively, in order to demarcate two The inner parameter M of individual camera₁, external parameter M₂The scaling board image calculation on ground it is placed on by shooting.Image coordinate (u, v) With world coordinates (X_w,Y_w,Z_w) homogeneous transformational relation as follows:

$\begin{array}{c}{z}_c\left[\begin{array}{c}u\\ v\\ 1\end{array}\right]=\left[\begin{array}{ccc}\frac{1}{d_x}& 0& u_0\\ 0& \frac{1}{d_y}& v_0\\ 0& 0& 1\end{array}\right]\left[\begin{array}{cccc}f& 0& 0& 0\\ 0& f& 0& 0\\ 0& 0& 1& 0\end{array}\right]\left[\begin{array}{cc}{R}_C& T_C\\ \stackrel{\→}{0}& 1\end{array}\right]\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]\\ =\left[\begin{array}{cccc}{f}_u& 0& u_0& 0\\ 0& f_v& v_0& 0\\ 0& 0& 1& 0\end{array}\right]\left[\begin{array}{cc}{R}_C& T_C\\ \stackrel{\→}{0}& 1\end{array}\right]\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]=M_1{M}_2\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]\end{array}---\left(1\right)$

Wherein f is camera focus, (u₀,v₀) it is figure principal point coordinate, d_xAnd d_yRepresent respectively each pixel at transverse axis x and Physical coordinates on longitudinal axis y, these parameters all can be obtained by camera calibration.

Then stereo calibration calculates spatially two video camera P₁And P₂Geometrical relationship, the rotation between i.e. two cameras Matrix R and translation matrix T.Select right camera as reference camera.Relation is as follows:

P₁=R* (P₂-T) (2)

The correction chart picture of row alignment is obtained finally by non-demarcation three-dimensional correction HartLey algorithm.Requirement binocular camera is adopted Collection image synchronization.

Step 2, Stereo matching obtain parallax.

By the match point between the camera view of binocular solid matching primitives left and right, obtain disparity map, select Gaussian Mixture Modeling method carries out background modeling to disparity map, to eliminate the complex background interference to target detection.According to parallax, baseline and interior Ginseng, uses trigonometric calculations scene three-dimensional coordinate.Choose the world coordinate system for XOY face with ground, three-dimensional point is projected to Ground, projects to the number of three-dimensional point of certain pixel as the color value of this pixel, generates downward projection figure.

Step 3, downward projection figure cluster.

Probability density at x is f_h,k(x):

$f_{h,k}\left(x\right)=\underset{i=1}{\overset{n}{\Σ}}K\left(\right|\left|\frac{x-x_i}{h}\right|\left|\right)---\left(3\right)$

Wherein K (x) is kernel function, and h is radius

F to be made_h,kX () is maximum, to f_h,kX () derivation obtainsWherein g (s)=-k'(s),

$\begin{array}{c}\▿f_{h,k}=\underset{i=1}{\overset{n}{\Σ}}(x-x_i)g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)\\ =\[\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)\]\[\frac{\underset{i=1}{\overset{n}{\Σ}}{x}_{i}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}{\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}-x\]\end{array}---\left(4\right)$

Order:

$m_{h,g}\left(x\right)=\frac{\underset{i=1}{\overset{n}{\Σ}}{x}_{i}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}{\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}-x---\left(5\right)$

MakeAnd if only if m_h,g(x)=0, it can be deduced that new central coordinate of circle:

$x=\frac{\underset{i=1}{\overset{n}{\Σ}}{x}_{i}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}{\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}---\left(6\right)$

Due to the particularity of projection, only consider that the distance of pixel cannot obtain cluster result accurately, calculating probability During density, need to meet: the color value of (a) pixel is the most close with the color value of central pixel point, and probability density is the highest；(b) From the position of central point more close to pixel, probability density is the highest.Therefore, kernel function K is selected_h(x):

$K_h\left(x\right)=K\left(\right|\left|\frac{x^s-x_i^s}{h}\right|\left|\right)*K\left(\right|\left|\frac{x^r-x_i^r}{h}\right|\left|\right)---\left(7\right)$

After MeanShift cluster, each class represents a target.This result is projected to display in former right image final Testing result.

Claims (1)

1. one kind based on the shelter target detection method under binocular stereo vision complex background, it is characterised in that include following step Rapid:

Step one, binocular camera are demarcated；

Initially with Zhang Zhengyou gridiron pattern scaling method, shoot multiple gridiron pattern images respectively, in order to demarcate in two cameras Portion's parameter M₁, external parameter M₂The scaling board image calculation on ground it is placed on by shooting；Image coordinate (u, v) and world coordinates (X_w,Y_w,Z_w) homogeneous transformational relation as follows:

$\begin{array}{c}{z}_c\left[\begin{array}{c}u\\ v\\ 1\end{array}\right]=\left[\begin{array}{ccc}\frac{1}{d_x}& 0& u_0\\ 0& \frac{1}{d_y}& v_0\\ 0& 0& 1\end{array}\right]\left[\begin{array}{cccc}f& 0& 0& 0\\ 0& f& 0& 0\\ 0& 0& 1& 0\end{array}\right]\left[\begin{array}{cc}{R}_C& T_C\\ \stackrel{\→}{0}& 1\end{array}\right]\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]\\ =\left[\begin{array}{cccc}{f}_u& 0& u_0& 0\\ 0& f_v& v_0& 0\\ 0& 0& 1& 0\end{array}\right]\left[\begin{array}{cc}{R}_C& T_C\\ \stackrel{\→}{0}& 1\end{array}\right]\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]=M_1{M}_2\left[\begin{array}{c}{X}_w\\ Y_w\\ Z_w\\ 1\end{array}\right]\end{array}---\left(1\right)$

Wherein, f is camera focus, (u₀,v₀) it is figure principal point coordinate, d_xAnd d_yRepresent that each pixel is at transverse axis x and longitudinal axis y respectively On physical coordinates；

Then stereo calibration calculates spatially two video camera P₁And P₂Geometrical relationship, the spin matrix R between i.e. two cameras With translation matrix T；Select right camera as reference camera；Relation is as follows:

P₁=R* (P₂-T) (2)

The correction chart picture of row alignment is obtained finally by non-demarcation three-dimensional correction HartLey algorithm；Require binocular camera collection figure As synchronizing；

Step 2, Stereo matching obtain parallax；

By the match point between the camera view of binocular solid matching primitives left and right, obtain disparity map, select Gaussian Mixture modeling Method carries out background modeling to disparity map, eliminates the complex background interference to target detection；According to parallax, baseline and internal reference, adopt With trigonometric calculations scene three-dimensional coordinate；Choose the world coordinate system for XOY face with ground, three-dimensional point projected to ground, Project to the number of three-dimensional point of certain pixel as the color value of this pixel, generate downward projection figure；

Step 3, downward projection figure cluster；

Probability density at x is f_h,k(x):

$f_{h,k}\left(x\right)=\underset{i=1}{\overset{n}{\Σ}}K\left(\right|\left|\frac{x-x_i}{h}\right|\left|\right)---\left(3\right)$

Wherein, K (x) is kernel function, and h is radius

F to be made_h,kX () is maximum, to f_h,kX () derivation obtainsWherein g (s)=-k'(s),

$\begin{array}{c}\▿f_{h,k}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left(x-x_i\right)g\left(\left|\right|\frac{x-x_i}{h}|{|}^2\right)\\ =\[\underset{i=1}{\overset{n}{\mathrm{\Σ}}}g\left(\left|\right|\frac{x-x_i}{h}|{|}^2\right)\]\[\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_{i}g\left(\left|\right|\frac{x-x_i}{h}|{|}^2\right)}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}g\left(\left|\right|\frac{x-x_i}{h}|{|}^2\right)}-x\]\end{array}---\left(4\right)$

Order:

$m_{h,g}\left(x\right)=\frac{\underset{i=1}{\overset{n}{\Σ}}{x}_{i}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}{\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}-x---\left(5\right)$

MakeAnd if only if m_h,g(x)=0, the central coordinate of circle that must make new advances:

$x=\frac{\underset{i=1}{\overset{n}{\Σ}}{x}_{i}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}{\underset{i=1}{\overset{n}{\Σ}}g\left(\right|\left|\frac{x-x_i}{h}\right|{|}^2)}---\left(6\right)$

Due to the particularity of projection, only consider that the distance of pixel cannot obtain cluster result accurately, calculating probability density Time, need to meet: the color value of (a) pixel is the most close with the color value of central pixel point, and probability density is the highest；(b) from The pixel that the position of heart point is the nearest, probability density is the highest；Therefore, kernel function K is selected_h(x):

$K_h\left(x\right)=K\left(\right|\left|\frac{x^s-x_i^s}{h}\right|\left|\right)*K\left(\right|\left|\frac{x^r-x_i^r}{h}\right|\left|\right)---\left(7\right)$

After MeanShift cluster, each class represents a target；Project to this result former right image shows final inspection Survey result.