具有彈性的特徵點選擇策略應用於強化機器人之同時定位與建立地圖系統

Abstract

同時定位與建立地圖的能力是機器人達成其自主性並完成人們交付於工作的首要課題。由於影像感測器具有低成本、取得性高等優點，利用單眼攝影機實現以上能力成為最近幾年的熱門研究。這篇論文提出了一套新的影像特徵點擷取及選擇方法，基於結合了由下而上與由上而下兩種不同的視覺注意力模型並建造出一些感興趣區域。這些感興趣區域不僅可以減少影像中特徵點的數量，更降低了特徵點配對的時間。此方法也具備了選擇可以穩定追蹤的特徵點來當作目標物的能力，因此提高了同時定位與建立地圖系統的效能。 因為利用了單眼攝影機作為主要感測器，我們無法直接從影像上得到目標物的距離資訊。所以，我們需要一系列從不同的機器人位置取得的觀測量來克服此不足，也就是目標物初始化的問題。本研究搭配了增強型的卡爾曼濾波器架構，我們展示了利用較少量的特徵點也可以執行同時定位與建立地圖的程序。實驗的成果證實了我們提出策略的效能。

This thesis presents a novel approach for extracting and selecting visual interest points, Speeded Up Robust Features (SURF), for bearing-only SLAM in indoor environments. The algorithm is based on combining bottom-up and top-down visual attention to construct the regions of interest (ROIs). These ROIs reduce the number of features generated by SURF as well as the matching time. This method is also capable of selecting features as landmarks that can be matched reliably and thus increases the efficiency for SLAM. When using the monocular camera as our primary sensor, we cannot directly get the information of the distance to the landmarks. Hence, a set of measurements taken from different robot positions is needed to do landmark initialization. With the help of an extended Kalman filter (EKF) framework, we demonstrate that SLAM process can be executed with a lower number of features. Results from several real-world experiments verify the improvement of the proposed algorithm.