Research On Real-time Reconstruction Of Indoor Scene Based On Trinocular Stereo Vision Of Mobile Robot

For mobile robots,a basic reconstruction process includes steps such as environment perception,positioning and pose optimization of mobile robot,map generation and stitching,and relocalization.This thesis focuses on the following three parts: depth sensing,optimization,and relocalization.And a complete real-time scene reconstruction system for mobile robots is built.This thesis studies a high-precision depth sensing algorithm based on information fusion of a trinocular camera.Traditional stereo vision algorithms based on passive sensing often have a strong dependence on textures in the environments;Range sensors do not rely on textures,but their output resolution is usually low and often suffer from system errors.A novel joint depth filter is proposed to combine their advantages and make a depth fusion.Experimental results show that compared with pure depth cameras or binocular cameras,the proposed fusion algorithm outputs depth maps with less error and higher reliability,and is superior to other classic trinocular fusion algorithms.A visual odometry is built to locate and pose optimization for mobile robots in structured environments.Based on a state-of-the-art direct visual odometry,DSO,a more robust odometry is proposed in structured environments by introducing plane constraints,which is common in indoor scenes.The introduction of plane constraints greatly reduces the number of variables in the model and accelerates the convergence of the optimization process.Experimental results show that,compared with DSO,the improved odometry is not only more stable,but also the operating efficiency has been significantly improved.A novel hierarchical clustering matching algorithm used for fast relocalization is proposed.The algorithm uses the idea of hierarchal clustering,which removes those do not comply with local geometry consistency iteratively until convergence.Compared with conventional RANSAC algorithms,the proposed algorithm performs higher efficiency(approximating ())and accuracy with less accumulative errors on relocalization,especially in complex environments.Finally,a complete real-time reconstruction system for indoor scenes was built and experiments were performed on a mobile cart with a trinocular camera.Experiments show that compared with classic binocular reconstruction schemes and RGB-D reconstruction schemes,the scene reconstruction based on trinocular stereo vision proposed in this thesis performs better,especially on the adaptability to different scenes.