基於單眼視覺之機器人迴圈原點及同步定位與地圖建立

Abstract

本論文發展一套使用單眼攝影機來實現機器人同步定位與地圖建立(simultaneous localization and mapping,SLAM)的系統。利用SIFT演算法擷取環境特徵點及建立描述元，並結合擴展式卡門濾波器(Extended Kalman Filter,EKF) 完成移動式機器人之SLAM系統。本論文利用單眼攝影機擷取畫面的特性，設計去除比對錯誤的特徵點，並以實驗驗證其正確性。經由此方式，可以改善因比對錯誤所造成定位誤差，並且提升EKF的執行的穩定性。另外對迴圈原點之畫面比對，則使用建立參考畫面之方式，當機器人回到之前的位置，經過畫面比對位置差異的方法，可以改善迴圈原點的正確性。多次迴圈的實驗結果顯示本論文所提出的方法可達成機器人室內導航並建立特徵點位置地圖的功能。

In this thesis, we develop a system that uses a monocular camera to realize simultaneous localization and mapping(SLAM) of a mobile robot. The vSLAM system is based on Extended Kalman Filter (EKF) and uses scaling invariant feature transform (SIFT) algorithm for feature extraction. We propose a new method by using the characteristic of monocular camera to discard the outlier and improve the feature matching rate. The method also helps the stability of EKF algorithm to make more accurate robot localization more correctly. We save the reference images into a database and match the current image with reference images for improving the loop closing. The experimental results show that the proposed method effectively improves the feature matching and loop closing accuracy. A multi loop indoor navigation experiment reveals that the proposed localization algorithm can help robot to navigate in indoor environment and build the features map simultaneously.