CN104376567A

CN104376567A - Linear segmentation guided filtering (LSGF)-based stereo-matching method

Info

Publication number: CN104376567A
Application number: CN201410705013.9A
Authority: CN
Inventors: 刘怡光; 王晓峰
Original assignee: Sichuan University
Current assignee: Sichuan University
Priority date: 2014-12-01
Filing date: 2014-12-01
Publication date: 2015-02-25

Abstract

The invention relates to stereo matching methods in computer vision, provides an accurate linear segmentation guided filtering (LSGF)-based stereo-matching method used for overcoming the defects of low-resolution textures and border discontinuity when a guided filtering stereo-matching method is adopted and belongs to the field of computer application. The method comprises the steps that firstly, the accuracy of the cost measure of guided filtering is improved through introduction of the BT measure and gradients in two directions; secondly, in the image segmentation process, superfine segmentation is introduced to replace excessively fine segmentation; finally, image segmentation (Mean-shift algorithm) is used for conducting further refining treatment on parallaxes obtained through guided filtering. In combination with a linear plane fitting algorithm, in order to obtain a robust and accurate parallax effect, a partial derivative decomposition method is adopted to obtain linear plane fitting parameters, a segmented parallax set of segmented parallaxes which are stably calibrated is obtained, and refining is conducted on instable planes of instable parallaxes through the stable parallax set. The linear segmentation guided filtering method has the advantages that the stereo matching parallax accuracy is high, and edges can maintain a good smooth effect.

Description

A kind of solid matching method guiding filtering (LSGF) based on linear partition

Technical field

The present invention relates to computer vision and Stereo Matching Technology, particularly relate to a kind of solid matching method guiding filtering (LSGF) based on linear partition, belong to computer application field.

Background technology

Stereo matching is one of computer vision direction of current focus.Its target is in different viewpoint visual angles and different time places, in a width or a few width picture, find out corresponding point, and then obtains a dense degree of depth or parallax mapping graph.

Solid matching method common at present comprises partial approach and global approach.Global approach mainly comprises figure and cuts (GC) and belief propagation (BP) method.Global approach precision is higher, but speed is relatively slow, is difficult to meet apply in real time, accurately and rapidly, especially cannot meet in the fields higher to real-time such as three-dimensional reconstruction, three-dimensional video-frequency, robot navigations.Local algorithm mainly comprises bilateral filtering (BF), guides filtering (GF) etc.The usual speed of partial approach is fast, can meet real-time application, wherein guides filtering to be one of part filter that the fastest effect is best at present, has good edge preserving smoothing effect.

But guide filtering discontinuity zone edge and large area smooth region effect not ideal enough, there will be error hiding, therefore be improve Stereo matching precision further, do not affect speed simultaneously, we introduce linear partition and guide filtering, to improve the precision of Stereo matching further, especially at discontinuity zone.

Summary of the invention

The object of the invention be to provide a kind of accurately, effective linear partition guides the solid matching method of filtering, the solid matching method being intended to solve existing guiding filtering is in the defect on large-area low texture region and discontinuity zone border.

The present invention realizes by the following method: first for improving navigational figure precision, in conjunction with BT(Birchfield and Tomasi) estimate and two direction gradients (horizontal gradient and VG (vertical gradient)), color in estimating cost is respectively estimated and gradient is estimated and improved, to improve the precision guiding filtering; Then we are on the basis guiding filtering, introduce Iamge Segmentation (mean-shift algorithm), are found by lot of experiments, and meticulous segmentation still has greater probability to cross over discontinuity zone, therefore ultra-fine segmentation is introduced into alternative meticulous segmentation; In the ultra-fine segmentation obtained, implement to carry out linear plane matching to parallax, for obtaining robust, stable linear solution, we adopt linear partial differential coefficient decomposition method to obtain stable plane parameter, obtain plane parallax collection, finally with stable parallax collection refining point of instability, obtain linear partition and guide filtering (LSGF).

The inventive method concrete steps are as follows:

Step 1: adopt BT(Birchfield and Tomasi) estimate and two direction gradients, color in estimating the cost of navigational figure is respectively estimated and gradient is estimated and improved, to improve the precision guiding filtering. on the cost measure function basis guiding filtering, we improve cost function.For the gradient measure function improved, we adopt two direction gradients (horizontal direction and vertical direction), and in the color improved is estimated, we introduce BT and estimate.

Step 2: in Iamge Segmentation, ultra-fine segmentation is introduced into alternative meticulous segmentation; In usual Iamge Segmentation, meticulous segmentation obtains good effect, is widely used.But we find, meticulous segmentation still also exists larger probability and crosses over noncoherent boundary and region.So we to adopt ultra-fine segmentation that block is made as far as possible little.

Step 3: linear partition guides filtering (LSGF). first, simple linear solution can produce more stable and efficient parallax effect usually, and can meet the needs of precision and stability, we adopt linear plane approximating method.We adopt partial derivative decomposition method to obtain linear plane three parameters, and concrete steps are as follows:

1) credible parallax is obtained.We adopt left and right consistent method (LRC) to obtain credible parallax, filter insincere parallax simultaneously.When the ratio η of parallax credible in each segmentation is greater than 0.6, we think that segmentation is stable.For the segmentation of stable parallax, can in the hope of three of a plane fitting component by believable parallax value;

2) horizontal component α _hsolve.Consider that partial derivative decomposition method obtains α _h, α _hcan ask the credible parallax value of horizontal direction in each stable segmentation, then sequence obtains intermediate value;

3) vertical component α v solves.Same method we consider that partial derivative decomposition method obtains α v, α v can ask the believable parallax of horizontal direction in each stable segmentation, and then sequence obtains intermediate value;

4) α d solves.Same method we can be used in and ask α d, then sequence obtains intermediate value;

5) after, we obtain the fit Plane parameter of credible block, obtain disparity plane collection Γ.

Then, our further refining disparity plane collection, obtain final parallax, concrete steps are as follows:

1) the insincere parallax value of refining.For incredible parallax, we adopt the plane parameter of believable parallax collection Γ to carry out refining to insincere point;

2) we concentrate each segmentation plane to remove the insincere block of refining with credible segmentation, and error minimum parameter can think the credible parallax value of this segmentation;

3) merged by the same disparity plane, new plane parameter can be obtained by plane fitting, finally obtains final parallax.

The accurate linear partition that the present invention proposes guides the sectional perspective matching process of filtering (LSGF), it has speed and the precision of good Stereo matching, also there is good edge to keep and smooth effect, especially at the border of discontinuity zone and low texture region simultaneously.

Accompanying drawing explanation

Fig. 1 is that the present invention is a kind of based on linear partition guiding filtering process flow diagram.

Claims

1. guide a sectional perspective matching process for filtering (LSGF) based on linear partition, it is characterized in that comprising the following steps:

Step 1: for improving navigational figure precision, in conjunction with BT(Birchfield and Tomasi) estimate and two direction gradients (horizontal gradient and VG (vertical gradient)), the color estimated the cost of navigational figure is respectively estimated and gradient is estimated and improved, to improve the precision guiding filtering;

Step 2: we are on the basis guiding filtering, introduce Iamge Segmentation (mean-shift algorithm) to parallax refining to improve parallax precision, found by lot of experiments, meticulous segmentation still has greater probability to cross over discontinuity zone, therefore ultra-fine segmentation is suggested the meticulous segmentation of replacement;

Step 3: in the ultra-fine segmentation obtained, linear plane matching is carried out to the parallax obtained, for obtaining robust, stable linear solution, we adopt partial differential coefficient decomposition method to obtain three stable linear plane parameters, obtain credible plane parallax collection, finally with the parallax of credible parallax collection refining point of instability simultaneously.

2. linear partition according to claim 1 guides filtering, it is characterized in that, in step 1, improves guiding the cost of filtering to estimate:

In color is estimated, we introduce BT and estimate, and improve the color that cost estimates and estimate; In gradient is estimated, we adopt two Fextures (horizontal direction and vertical direction) to improve gradient that cost estimates is estimated.

3. linear partition according to claim 1 guides filtering, it is characterized in that, in step 2, improves Iamge Segmentation (mean-shift algorithm):

Experimentally observe, and found by lot of experiments, meticulous segmentation still have greater probability cross over discontinuity zone, still propose with ultra-fine automatic Segmentation image; By verification experimental verification, obtain good segmentation effect.

4. linear partition according to claim 1 guides filtering, it is characterized in that, in step 3, improves guiding filtering and Iamge Segmentation:

In the ultra-fine segmentation obtained above, linear plane matching is carried out to the parallax guiding filtering to obtain, for obtaining robust, stable linear solution, we adopt partial differential coefficient decomposition method to obtain stable linear plane parameter, when the ratio η of parallax credible in each block is greater than 0.6, be stable point, obtain stability plane parallax collection Γ accordingly, then with the parallax of stable parallax point set refining point of instability, and the planar set of identical parameters is merged, and then re-start plane and fit, finally obtain stable disparity map.