利用整合式單眼視覺之機器人同步自我定位及建立地圖系統實現大範圍之室內環境探索

Abstract

在大尺度環境下機器人同步建構地圖與自我定位(Simultaneously Localization and Mapping, SLAM)的應用中，常會遇到一個問題，即過多的地圖地標無可避免地會讓共同估測機器人位置與地圖地標的濾波器計算負擔太大。這主要是由於兩個原因所造成：一方面是在地圖地標的選擇機制上不夠扎實，導致在環境觀測過程中不必要的定位地標太多；而另ㄧ方面，則是濾波器的本身數學特性，導致計算負擔的增加。在本論文中，我們提出了一個結合加速強健特徵擷取(SURF Extraction)以及逆深度特徵初始化(Inverse Depth Initialization)的影像前端系統，來有效的選出強健的靜止地圖地標，用以提供定位及地圖資訊，並且在已知地圖再次觀測到的前提之下，有效達到大範圍的不確定縮減。此外，在後端濾波器的選擇上，我們將稀疏線性化資訊濾波器演算法延伸到影像感測器的應用。稀疏線性化資訊濾波器已被證實，在使用雷射實現SLAM時，可以有效的維持計算效能。最後，透過實驗以及模擬，我們證實了此系統的效能及可靠性。

In the application of root Simultaneous Localization and Mapping (SLAM) in a large scale environment, it remains a challenge to resolve the obstacle of the inevitable computational burden on the filtering scheme imposed by the excessive number of landmarks. This obstacle maily attributes to two facts: one is that the selection scheme is not sufficiently stringent, thus resulting in the inclusion of valueless localization landmarks during the environment observation process; the other is the mathematical characteristic of the filter, i.e. the computational complexity is proportional to the number of landmarks. In this thesis, we propose a visual front-end system integrating the speed-up robust feature extraction (SURF Extraction) and Inverse Depth Initialization to efficiently and effectively select robust static landmark for the information of localization and mapping and significantly reduce the uncertainty of the large exploration environment under the presumption of re-observation of the map. Furthermore, we extend the sparse linearization information filtering algorithm to the application of visual sensor. In the SLAM of laser, it has been proved the adoption of sparse linearization information filter effectively improve the computational efficiency. The performance and reliability is validated by the simulation and experiments.