基於全向式影像之機器人同步定位與環境地圖建立

Abstract

本論文提出一使用全向式影像之機器人定位方法。以全向式攝影機為感測器，結合基於Extended Kalman Filter(EKF)之同時定位與環境地圖建立(Simultaneous Localization and Mapping, SLAM)演算法，讓機器人在移動的同時，能夠建立出環境特徵地圖並定位出機器人本身的位置。全向式攝影機具有360度的視角，除了能取得更多的環境特徵外，亦能增加持續追蹤到landmark的時間，讓SLAM的運作更為穩定。配合攝影機的特性，本論文參考Scale Invariant Feature Transform(SIFT)演算法發展出一有效之特徵點辨識演算法，用以辨識兩張相鄰的影像中相同的環境特徵點，此方法對於影像經過旋轉及大小縮放後依然擁有相當穩健的辨識。本論文提出一視覺參考點建立與轉換的策略，讓機器人進入新的環境時能建立新的參考點與地圖，當走回舊地區時則從資料庫中取回舊有的參考點資訊使用，減少參考點總數，降低EKF濾波器的運算負擔。論文中以實驗室之機器人進行導航實驗來驗證所提出之定位演算法，實驗結果顯示特徵點比對之正確率為90%，行經30公尺後之定位誤差為0.1公尺。實驗的結果證實機器人能依定位系統的幫助在走廊上長距離的移動，並且同時建立出走廊環境的特徵點地圖，達成機器人室內導航的功能。

This study investigates robot localization and mapping using omni-directional images. A method is proposed to use an omni-directional camera to realize simultaneous localization and mapping (SLAM) algorithm based on extended Kalman filter (EKF). Because of the 360° field of view, an omni-directional camera is suitable for simultaneous localization and mapping (SLAM) for detecting and tracking environmental features. A new algorithm is developed adopting scale invariant feature transform (SIFT) method to match features in environment between two images. This thesis also presents a switching method of visual reference scans. In this method, reference scan can be added to a database or switched automatically among reference scans. These scans can be used repeatedly to reduce the complexity of extended Kalman filter (EKF). Experiments results show that the matching rate of landmark features is 90%. A long range indoor navigation experiment revealed that the proposed localization algorithm can help robot to navigate in indoor environment and build the features map simultaneously.