基於網格平面的線特徵與矩形幾何約束之單眼視覺里程計

Abstract

科技發展迅速，機器人已經可以獨自執行各種複雜的任務，不需要額外的人為參與，從而提升我們日常生活的便利性。為了讓機器人能夠自主執行任務，機器人必須能夠感知周遭的環境，以便在場景中移動與避開障礙物。 本論文提出一個單眼視覺里程計，用於機器人在都市環境的定位。透過線段偵測、合併、分群與排除異常值，偵測出畫面中建築物外牆上的水平與垂直結構線，並透過矩形幾何約束，從偵測到的線條在三維空間中建立網格平面，經由追蹤建立的網格平面來估測機器人的位置。為了讓本論文所提出的演算法能在即時條件下執行，加入了可靠性測試，透過重投影特徵將估測資料做降取樣處理，進而提升運算的效率。 本論文透過模擬、室內與室外的實驗驗證了演算法的效能，展示在實際的都市環境中定位的能力。

Nowadays, mobile robots can perform complex tasks independently, reducing the requirement for human involvement and thus significantly improving the convenience of our everyday lives. To perform tasks independently, mobile robots need to determine their location, sense its surroundings for navigation and avoid obstacles. This thesis proposes a monocular visual odometry for a mobile robot localization in urban environments. The proposed algorithm detects the horizontal and vertical structural lines on a building facade using image processing techniques including line segment detection, segments merging, clustering and outlier removal. Then, the rectangular geometric constraint is applied to form the grid mesh plane in 3-D space using the detected line features, and the mobile robot position is estimated by tracking the plane in camera images. For the proposed algorithm to operate in the real-time, a reliability testing is involved to enhance the calculation efficiency by downsampling data via feature reprojection. The proposed algorithm in this thesis has validated the performance through simulations and both indoor and outdoor experiments, demonstrating its capability for localization in noisy urban environments.