Vision Based Simultaneous Localization And Mapping For Robot

Localization and mapping are the two key problems needed to be solved for autonomous robot. But the localization approach only based on dead recking has accumulation error, and make the consistent characteristic of mapping been destroyed. Because the multiple observations have high degree of correlation, and we wish to use this information to constrain the error accumulated, thus obtain more accurate localization and landmarks' position. In the stochastic process framework, SLAM utilizes probabilistic inference technology to realize simultaneous localization and mapping estimation process. Meanwhile, vision information has the advantages of much quantity, low cost, minimal energy consumption and intuitive effects, which particularly suitable for the SLAM task of out-space planet exploration. In recent years, vision SLAM has been raised gradually, its applications has been expanded from indoor environment to outdoor environment, 3DOF estimation to 6DOF estimation, and more and more close to practical application. This dissertation will discuss the problem of how to use vision information to realize robot SLAM.The main research contents include:Firstly, the theoretical solution for SLAM problem based on Bayes filter is presented. In order to avoid the full-state estimation, Rao-Blackwellised factorization is used to decouple the the estimation process of robot and landmark state space, meanwhile a succinct proof of this factorization is put out; The basic theory of particle filter is introduced, and the basic particle filter algorithm is presented; The detailed realization process of FastSLAM algorithm is presented, which is an efficient approach to solve SLAM problem based on Rao-Blackwellised particle filter.Secondly, the extraction process of SIFT feature is discussed, and an approximate SIFT extraction method will be developed. Against the low efficiency characteristic of SIFT's global matching, an approximate nearest neighbors searching algorithm based on kd-tree will be developed. In order to extract sub-map effectively and maintain the feature map efficiently, an SIFT feature management method will be advanced.Thirdly, through using the stereo matched SIFT features as natural landmarks, one SLAM algorithm based only on binocular vision information will be developed. After putting out the framework of binocular vision SLAM, with the previous and present stereo matched image information, one robust visual odometry algorithm will be developed, which will be used as control inputs of SLAM process. Then in framework of FastSLAM 2.0, the specific realizations of drawing samples from proposal density, updating the landmark position, particle's weight computing are discussed detailedly.Fourthly, against the encountered feature number and initial motion uncertainty increased problems for monocular vision SLAM, in order to increase the localization accuracy and computational efficiency, a new proposal density with Gaussian Mixture Model (GMM) and density estimation process for Rao-Blackwellised particle filter will be developed. The posterior density of landmark position will be modeled as GMM, thus a new SLAM algorithm based on Rao-Blackwellised particle filter will be developed, in which only one GMM landmark map need to be maintained. Then all the realization process such as drawing samples from proposal density, posterior density estimation, particle's weight computation and landmark update are discussed detailedly.Lastly the conclusion and perspective are given at the end of the dissertation.