Extraction Algorithm And Implementation Of 3D Texture-less Planar Structure In Stereo Vision

Accurate plane detection in man-made environment is of high interest for 3D reconstruction and robot navigation applications.However,texture-less plane detection is a tough challenge in computer vision.In this thesis,we deeply investigated the traditional stereo matching algorithms,line segments detecting methods and the principle of two-view geometric in computer vision,then proposed a line-based multi-plane fitting algorithm as well as a homography-based 3D reconstruction system.This system and algorithm are designated to estimate the depth of indoor texture-less scenes from stereo images,which avoids the ambiguity problem of points-based stereo matching.Our research mainly consists of the following parts:the initial disparity estimation,the line segments extraction,the visibility constrained segments graph construction,the line-based multi-plane fitting algorithm and the spatial planar structure segmentation based on Markov Random Field.(1)Initial disparity estimation and image line segments extraction.In this thesis,we combine the Sobel operator and the Hough Transform methods,which are sensitive to edge regions of the image,to perform matching cost calculation,such that the disparity of edge regions can be obtained quickly and precisely.Then the edge segmentation method based on region growing is used to obtain the image edge,laying the foundation for the planar structure fitting.(2)Line segments graph construction based on visibility constrain.In this thesis,we use the Delaunay triangulation algorithm,and combined with the proposed visibility constraint to innovatively define an adjacent topological relationship between line segments.The line segment graph model constructed according to the relationship can not only save computation time in finding adjacent line segments,but also avoid obtaining invalidate planar structure when using two remote invisible line segments to fit a plane.(3)Line-based multi-plane fitting algorithm.When using sequential RANSAC to fit multiple models from sparse features,each time a model instance has been fitted,the inlier will be deleted,resulting in missed detection of some planes.Since the line segments extracted are sparse feature,it is crucial to deal with the missed detection problem.The line-based multi-plane fitting algorithm proposed in this thesis effectively overcomes the missed detection problem caused by sequential RANS AC,and improve the accuracy of planar structure detection.(4)Special planar structure segmentation based on Markov Random Field.In this thesis,we design the data term and smooth term of the Markov Random Field model to perform image segmentation based on the image color,disparity,planar structure information etc.In this way,each segmentation corresponds to a plane and has a clear,smooth boundary.Experimental results demonstrate the high quality of our generated 3D models,especially for large-scale texture-less environment.