| Vision based SLAM (Simultaneous Localization and Mapping) has attractedmuch attentionn in the field of mobile robot navigation; most of the research havefocused on the exaction of high-quality and robustfeatures, data association, andreal-time localizaition and mapping. This thesis aims to develop a new monocularceiling vision based SLAM appraoch by utilizing an improved SRUKF (Square RootUnscented Kalman Filter). The proposed approach can not only achieve highaccuracy but also reduce the computational complexity of standard SRUKFalgorithm.In order to decrease the influence of external noises to the exaction of fearure, astrategy of ceiling vision is utilized. Equiped with a monocular camera which islooking upward to the ceiling, the robot only need to processes salient features thefeatures contain only rotation and shear deformation, and no scale change occursbetween images. Therefore, the ceiling strategy can lower computational burden,having a high accuracy.Based on the analysis of traditional SLAM algorithms, this thesis adopt amethod of SRUKF-SLAM which is a linearization needless and numerical stablealgorithm. However, traditional SRUKF-SLAM has many drawbacks, and can oftenlead failures, like divergence of filter, numerical instability, high computationalburden, etc. To address these drawbacks, a modification is proposed to helpconstruct a related Cholesky factor instead of the traditional method, and this canhelp to overcome filter divergence. In additoin, optimization methods are used tosovle the strong relation problem between elements in getting features’ uncertainty.In addition, this thesis adopt submap method to get a global map in the SLAMalgorithm. In the process, submap-building trigger condition is set based on thethreshold of displacement difference and angle change difference. Finally,comparison experiments between traditional algorithm and the modified algorithmare performed to illustrate the effectiveness in ceiling vision based SRUKF-SLAM. |
PDF Full Text Request