Research And Implementation Of Binocular Stereo Vision System

Vision is an important way for human beings to observe and recognize the world.Human beings obtains most of information by vision system.Simulation of the human vision system by machine has very important theoretical significance and practical value.Binocular stereo vision is a very popular research field in computer vision.Imitating human visual system,a typical binocular stereo vision system utilizes two cameras on different positions to capture a pair of images of a same scene simultaneously.Once the disparity between these two images is estimated,the depth from objects to cameras could be calculated according to the triangulation measure.A typical binocular stereo vision usually consists of four steps,i.e.,camera calibration,stereo rectification,matching and reconstruction.In this thesis,we mainly execute our work in two aspects,i.e.,1)according to the fundamental theory of binocular stereo vision technology,we set up a binocular stereo vision system platform that could realize distance measurement of near-field;2)proposing a novel matching algorithm for binocular stereo vision based on fuzzy clustering.We compared several algorithms of different procedures associated with binocular vision system,and choose the suitable ones to build the binocular platform.Camera calibration is the system's premises,we incorporate the Zhang's method into our platform as the algorithm of calibration since it possesses high precision and is easy to realize.Actually,our platform implemented single camera calibration and two cameras calibration.Through a series of experiments,we obtained the reasonable calibration configuration of parameters.Rectification could ease the difficulty level of stereo matching,and here Bouguet algorithm is selected since it mainly parameterized by calibration algorithm,meanwhile it possesses lower computational cost.Next,we implemented many stereo matching algorithms to compare their performance.Finally,we combined these algorithms with triangle measure to realize the entire workflow,and executed a series of systematic experiments.We also made a complete researching on the theorems and techniques of stereo matching.Inspired by fuzzy clustering,we proposed a novel fuzzy partition based stereo matching method for binocular stereo vision.We verified the performance of our methods in Middlebury database.Meanwhile we deployed it on our platform,and tested it on captured images.The comparison experiments show the proposed method is more robust against segmentation parameters,meanwhile,obtains better performance for the image with some specified phenomenon,e.g.,few texture,repeated texture.The hardware of our platform includes binocular stereo camera and calibration target.As for software,its GUI is design by using Qt and all algorithms are coded by C++ and are implemented based on Open CV library.Our binocular stereo vision platform could realize the distance measurement of object in near-field.