基於Homography的自動車定位研究

Abstract

本論文的目標是探討利用電腦視覺技術來完成自動車在室內環境所做定位工作。其最主要的概念在於建立架設於自動車上的攝影機所擷取之影像與工作空間環境兩者的平面投影轉換關係，以求出自動車在空間中的正確位置。更具體來說，我們提供兩種定位的方法：1.對每一張所擷取的影像特徵反覆計算其對應於工作空間之平面轉換關係，以此來計算自動車在地圖中的位置。2.利用單一的平面轉換矩陣計算特徵點在空間中相對於自動車的運動向量，以此推估自動車位置。其中第一種方法，相較於早先我們所曾經使用的Cross-Ratio定位方法，兩者在使用最少特徵點數目（4點）之下有等價關係，但對於特殊情況的處理，實驗結果顯示本論文所提之方法，有著較優的處理效能以及精確性。而第二種方法的特色在於能夠對自動車較小範圍的運動提供簡單的定位，其精確性雖然不及第一種方法，卻仍能提供使用者一定程度的定位結果作為參考。

The goal of this study is to automatically calculate robot locations in an indoor environment using computer vision-based techniques. The main concept is based on homographies that model the geometric transformations between images captured by a camera and the workspace of a planar environment. Specifically, we propose two localization methods: (i) localization based on repeated computation of homographies using image features in each frame and (ii) localization based on a single planar transformation by evaluating relative motion of landmarks in the workspace w.r.t. the robot. For (i), one can show that it is equivalent to the localization method based cross-ratio proposed previously. If minimum number of image features are used. While (i) can be extended directly to take into account more image point features, extensions of the cross-ratio based method is not straightforward. On the other hand, experimental results show that (i) is more accurate than (ii); however (ii) is more simple and can provide useful localization results by locating robot positions in small range.