Research Of Stereo Matching For Binocular And Its Application

Binocular Stereo is an active topic in computer vision,which is widely applied to 3D measure,3D location,robotic navigation,and autonomous driving.Stereo matching is the key step and difficult problem of binocular stereo.The performance of matching algorithm is crucial to binocular stereo.Noises,occlusions,disparity discontinuity,low texture and repeated texture limited the accuracy of stereo matching.Some complex matching methods are time consuming,which limit the applications of binocular stereo system.To resolve these problems,the main contributions of this paper are:(1)The image pairs are divided into many image patches and the stack auto encoder is used to extract features within these image patches.The distance between features represents the matching cost of two image patches.These features weaken the effects of measure noises,lens distortion and lighting variation,which improve the reliability of the matching cost.(2)We combine belief propagation algorithm with guided filter to aggregate matching cost.Guided filter could preserve the edge information of objects and is highly efficient.The guided filter improves the performance of belief propagation algorithm within occlusion areas and disparity discontinuity areas.At the same time,the guided filter will not reduce the matching accuracy in disparity continuity areas.This combination improves the accurate of the whole algorithm.Another drawback of belief propagation algorithm is computation complexity,which makes the algorithm too slow to be implicated.We propose an adaptive method to speed up the algorithm.The iterations of message update are controlled by the entropy of each pixel’s belief.The proposed method can reduce a lot of redundancy computations and can accelerate the matching algorithm effectively.Furthermore,this method doesn’t degrade the accuracy of the matching algorithm and can be applied to other message passing algorithms.(3)We use a guided filter to approximate the message passing process of a fully connected Markov random field(FMRF).The FMRF take more neighbors into account and work better than local connected Markov random field(LMRF)in disparity discontinuity areas.But there are many noises in disparity map within low texture areas,we use belief propagation algorithm to rectify these errors.The cascade of FMRF and LMRF takes advantages of global matching algorithm and local matching algorithm,which perform better than each of them.(4)Local matching algorithms have many mistakes in low texture areas and a matching algorithm based on laser speckle is proposed in this paper to overcome this weakness.We propose a matching cost function based on the gradient direction.This matching cost is more reliable,which simplifies the matching cost aggregation.A confidence function is proposed to rectify disparities in occlusion areas and disparity discontinuity areas.The speckle based matching algorithm is applied to measure 3D information of a package in real time.Experiment results show that the measure errors are within ?5%.The measure error will below ?3% for deformable package.(5)Stereo matching algorithm limits the performance and application of binocular stereo vision system.The geometry of some product is simple,this trait makes us to propose a virtual binocular stereo vision system.This method uses a mirror to construct a virtual camera,and reconstruction the 3D information of corners and edges.The 3D information of the object is fitted by these corners and edges.To realize this method,we proposed a 3D calibration method,the parameters of the camera is calibrated with one image.Then the camera is reconstructed by the intersection of nonplanar strait lines or the intersection of plans.This method is efficient and accurate,which successfully applied to chip package test.