本論文提出分離更新式FastSLAM方法，分離FastSLAM過程中機器人狀態與地標狀態之更新，藉以改善SLAM過程中機器人狀態的預測，在高運動雜訊情形下能有較佳表現；同時亦探討單眼視覺式SLAM，使具備複雜運動模型的行動機器人系統在SLAM過程中對機器人狀態之估測可脫離運動平台之限制。研究內容包含FastSLAM理論之推導、以雷射測距儀為感測器之FastSLAM與分離更新式FastSLAM的實現、單眼視覺系統透視投影法與反深度特徵初始化的推導，反深度特徵初始化與FastSLAM方法之結合，單眼視覺FastSLAM系統之實現。 In this dissertation, separated-update FastSLAM (FastSLAM SU), a modified FastSLAM method is developed in order to improve the robot state prediction in high motion uncertainty situation by updating the robot and landmarks state separately. Meanwhile, for complex motion platform robots, a free moving camera SLAM is also discussed. This research contains introduction, simulation and implementation of FastSLAM and FastSLAM SU and combination of FastSLAM and monocular vision system with inverse depth feature initialization.
