CN103366158A - Three dimensional structure and color model-based monocular visual road face detection method - Google Patents

Abstract

The invention discloses a three dimensional structure and color model-based monocular visual road face detection method, comprising the steps of firstly extracting and matching feature points of a road face scene image shot by monocular vision, then calculating the motion parameters of a camera, reconstructing the three dimensional structure of the feature points according to the motion parameters, separating road face and non-road face feature points from the three dimensional structure, and finally establishing the color Gauss mixed model of the two types of feature points and splitting a road face area by an image-splitting method. The invention provides a whole effective road face detection scheme, and solves the effects of complex road face scenes such as lighting and zebra lines to a certain extent.

Description

Monocular vision pavement detection method based on three-dimensional structure and colour model

Technical field

The invention belongs to computer vision and image processing field, relate to a kind of pavement detection method in conjunction with three-dimensional structure and colour model.

Background technology

Traditional pavement detection method mainly is divided into based on the method for pixel characteristic with based on two kinds of the methods of structural information.Method based on pixel characteristic exists following shortcoming: can accurately not detect and have a large amount of illumination shades and the such complex road surface of zebra stripes, can not obtain the constructional depth information of scene simultaneously.Method based on structural information exists following shortcoming: can not effectively detect pavement edge, and only near the pavement detection unique point be had good testing result.

Compare the vehicular platform that binocular vision system is more suitable for moving based on the pavement detection system of monocular vision, this is because monocular vision utilizes successive frame information that scene structure is analyzed, and therefore binocular vision system is fit to static scene more owing to not utilizing the movable information of scene.Single camera vision system has reduced equipment requirement simultaneously, is convenient to install set up.

Summary of the invention

Goal of the invention: in order to address the above problem, the present invention proposes a kind of monocular vision pavement detection method based on three-dimensional structure and colour model.

Technical scheme: the monocular vision pavement detection method based on three-dimensional structure and colour model comprises the steps:

Feature point detection and coupling: adopt the Harris angular-point detection method to extract the unique point of scene image, and with the unique point of current frame image by the method for sparse optical flow coupling and the Image Feature Point Matching of next frame;

Calculate camera motion: concern the fundamental matrix that calculates between the two continuous frames image according to Feature Points Matching, and obtain camera motion by decomposing fundamental matrix;

Reconstruct unique point three-dimensional structure: camera motion adopts the three-dimensional structure of the method reconstruct unique point of Linear Triangular shape;

Extract the road surface characteristic point: adopt the stochastic sampling coherence method to extract road areal model parameter in the three-dimensional feature point cloud, unique point is divided into road surface characteristic point and non-road surface characteristic is put two classes;

Set up the color gauss hybrid models: set up two types Gaussian Mixture colour model according to described two category features point, the color gauss hybrid models of each class has five gaussian component, and the generation of these five gaussian component is to obtain by the K-means clustering method;

The method that employing figure cuts is cut apart the zone, road surface.

The present invention adopts technique scheme, has following beneficial effect: complete effective pavement detection scheme in the present invention proposes, the impact that has solved to a certain extent the complex road surface scenes such as illumination, zebra stripes.Pavement detection method based on monocular vision is compared the vehicular platform that binocular vision system is more suitable for moving, and this is because monocular vision utilizes successive frame information that scene structure is analyzed.Single camera vision system has reduced equipment requirement simultaneously, is convenient to install set up.

Description of drawings

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is stochastic sampling consistency detecting method match of the present invention road surface process flow diagram;

Fig. 3 is that the present invention adopts the color gauss hybrid models to cut apart zone, road surface process flow diagram.

Embodiment

Below in conjunction with specific embodiment, further illustrate the present invention, should understand these embodiment only is used for explanation the present invention and is not used in and limits the scope of the invention, after having read the present invention, those skilled in the art all fall within the application's claims limited range to the modification of the various equivalent form of values of the present invention.

Concrete steps such as Fig. 1 the method are:

(1) extracts scene image unique point and coupling

Feature point detection adopts the Harris angular-point detection method, this is the similarity based method between a kind of detection pixel of the principal curvatures based on the image local autocorrelation matrix, and its basic thought is to assert that it is angle point that the point of obvious gray scale derivative is arranged on two orthogonal directionss.The method of the sparse optical flow that Feature Points Matching is used is a kind of matching process of seeking the unique point optimum displacement, and its energy function is:

$E=\∫\underset{w}{\∫}{n}^2\left(X\right)w\left(X\right)\mathrm{dX}=\∫\underset{w}{\∫}{(I(X-d)-J\left(X\right))}^{2}w\left(X\right)\mathrm{dX}$

Wherein, the weight function of w (X) expression window, I, J represent the pixel value of former frame image and current frame image.

(2) calculate camera motion

Feature Points Matching in the step of the present invention (1) relation is explained with a kind of mathematical model-fundamental matrix F, and F satisfies X' ^TFX=0, wherein X' and X represent a pair of matching characteristic point in two width of cloth images.Fundamental matrix can be explained by the kinematic parameter of video camera:

F＝[P'C] _×P'P ⁺＝[Kt] _×KRK ^-1＝K ^-T[t] _×RK ^-1

Wherein, t is the camera translation vector, and R is the video camera rotation matrix, and K is the confidential reference items matrix of video camera.

(3) reconstruct unique point three-dimensional structure

Adopt the method reconstruct unique point three-dimensional structure of common Linear Triangular shape, the method is by two video camera projection matrixes of the relative motion between two two field picture video cameras structure, then according to the three-dimensional coordinate of these two projection matrix calculated characteristics points.

(4) the road surface characteristic point in the extraction three-dimensional structure

Adopt the conforming method of stochastic sampling to extract plane, place, road surface in the three-dimensional structure, the thought of this algorithm is to recycle that a small amount of random sampled point removes to set up model in the sample, then assesses the degree that whole sample space mates this model.Think when sample can mate this model that when part (even can be lower than 50%) in the sample these samples are interior point (inlier), all the other unmatched samples exclude as exterior point (outlier).Maximum optimization model of counting in the last middle selection of in these, putting.The robustness of this method is embodied in a large amount of exterior points that exist in the effective Rejection of samples, and the method for comparing least square is more suitable for processing the actual photographed image.

(5) set up the color gauss hybrid models

Put this 2 category feature point according to road surface characteristic point and non-road surface characteristic and set up respectively two kinds of color gauss hybrid models, mixed Gauss model can be expressed as:

$D=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\π}}_i{g}_i(x;{\mathrm{\μ}}_i,{\mathrm{\Σ}}_i)$

Wherein $g_i(x;{\mathrm{\μ}}_i,{\mathrm{\Σ}}_i)=\frac{1}{\sqrt{\left(2\mathrm{\π}\right)\left|{\mathrm{\Σ}}_i\right|}}\exp [-\frac{1}{2}{(x-{\mathrm{\μ}}_i)}^T{\mathrm{\Σ}}_i^{-1}(x-{\mathrm{\μ}}_i)]$ The Gaussian probability-density function of expression sample x, μ _iBe the average of this Gaussian density function, Σ _iBe sample covariance, and mixed Gauss model is exactly with this K Gauss model weighting π _iLinear superposition is got up, wherein π _iSatisfy The K value is 5 in the present invention, namely by the K-means clustering method this two classes sample on road surface and non-road surface is gathered into respectively five kinds of gaussian component classifications.

(6) cut apart the zone, road surface

Employing figure segmentation method comes the road surface zone in the split image, and figure segmentation method core concept is that image configuration is become weighted graph G=＜E, V＞and, the summit of V presentation graphs wherein, the limit of E presentation graphs.The weights size of limit E among the figure is relevant with gray scale, the Texture eigenvalue of pixel, so just image segmentation problem is converted to max-flow (max flow) minimal cut (min cut) problem of figure, thereby according to the routing algorithm of max-flow and minimal cut figure is divided into two class subgraphs realization image segmentation.The present invention adopts the method for iteration to reach segmentation effect the best, namely by not stopping to upgrade the road surface and this two classes color gauss hybrid models of non-road surface is optimized segmentation effect, reaches stable until the energy of figure segmentation method cuts.

Claims (1)

1. based on the monocular vision pavement detection method of three-dimensional structure and colour model, it is characterized in that, comprise the steps:

Feature point detection and coupling: adopt the Harris angular-point detection method to extract the unique point of scene image, and with the unique point of current frame image by the method for sparse optical flow coupling and the Image Feature Point Matching of next frame;

Calculate camera motion: concern the fundamental matrix that calculates between the two continuous frames image according to Feature Points Matching, and obtain camera motion by decomposing fundamental matrix;

Reconstruct unique point three-dimensional structure: camera motion adopts the three-dimensional structure of the method reconstruct unique point of Linear Triangular shape;

Extract the road surface characteristic point: adopt the stochastic sampling coherence method to extract road areal model parameter in the three-dimensional feature point cloud, unique point is divided into road surface characteristic point and non-road surface characteristic is put two classes;

Set up the color gauss hybrid models: set up two types Gaussian Mixture colour model according to described two category features point, the color gauss hybrid models of each class has five gaussian component, and the generation of these five gaussian component is to obtain by the K-means clustering method;

The method that employing figure cuts is cut apart the zone, road surface.

Images