利用空間高斯混合模型進行影像地標判定及輔助定位

Abstract

在電腦視覺的領域中，要讓機器人能夠具有對於場景的認知，往往都需要在環境空間中，標注許多人工的特徵，或是紀錄場景中，特殊的形狀或是顏色，來幫助機器人確認自己的所在位置。然而在圖形識別的領域中，有許多功能成熟且強大的演算法，對於資料分析而言，並不需要對於輸入資料有很大的限制，也可以達到很高的辨識結果。本論文中採用高斯混合模型(Gaussian Mixture Model, GMM)來描述空間中的場景，將場景概念化，這樣任意的場景皆可以用高斯混合模型來表達，而無需加入大量的人工場景，接著將所獲取的每個希望校正點的資料採用由修改後的Cyr and Kimia的結合演算法（combination algorithm）來減低及統合資料，當場景接近原本的當初建立的場景，經由最大可能性(Maximum Likely)大體上會呈現單調的特色，將輔助機器人修正回原本的位置，以達到協助機器人定位的目標。

To achieve the robot registration and localization using computer vision, artificial landmarks or specific shapes or specific colors in the image are usually used. However, there exist many robust algorithms in pattern recognition and image segmentation which do not require constraints on input data to achieve a good recognition performance. In other words, localization can be performed based on a general scene. In this thesis, we adopt Spatial Gaussian Mixture Model (GMM) in image segmentation to describe an image viewed by the robot in spatial domain without any artificial landmark. Secondly, we use a modified combination algorithm by Cyr and Kimia to combine similar data. It is found out that a monotonic relationship exists among the scene registered and its neighborhood in terms of distance. This phenomenon can be used to assist to localize the robot and this work demonstrate the feasibility by several experiments.