部分特徵點更新機制於無標記擴增實境系統之應用

Abstract

增廣實境(Augmented Reality)是結合虛擬模型與實際環境影像的即時互動技術。 大部分的增廣實境系統,採用平面標記(planar markers)計算出攝影機的姿態,來達到虛擬模型與實際影像的結合。但平面標記缺乏彈性,直覺,以及擴充性的缺點,使得無標記增廣實境(Markerless Augmented Reality)越來越受到矚目與歡迎。在這篇論文中,我們提出了FinMars,來提供友善的使用者介面以及強健的無標記增廣實境系統。在此系統設計中,使用手指的姿態來當作直覺的使用者介面。手指能夠組成一個矩形的範圍,來標定出想要追蹤物體的範圍。之後,FinMars採用光流追蹤法(KLT)以及影像辨識演算法(SURF)來實現強健以及準確的無標記增廣實境系統,並且使用平行處理,來增加運算的效率。光流追蹤法負責即時追蹤目標,而光流追蹤法如果失去目標後,影像辨識演算法可以幫助重新辨識目標,再繼續追蹤。我們也設計了一套機制,來降低影像辨識的範圍,進而增加辨識的速度。除此之外,藉由切割興趣區域(region of interest)來達到部分特徵點(partial features)更新,能夠增加重新辨識的效率。 最後,我們比較FinMars,SURF,以及KLT在準確性,計算時間,部分更新的效能,以及物 體被遮蔽的狀況下,來證明FinMars有更佳的效果以及效能。

Augmented Reality (AR) is an innovative interactive technology which combines a virtual image and real environment in real time. Most AR systems adopt planar markers for camera pose estimation to register a virtual model with real world. Since planar markers lack of flexibility, intuition and scalability, markerless AR technology has become more and more popular and inevitable. In this paper, we propose a markerless AR system, named FinMars to provide a user-friendly and robust AR system. In our design, gesture-based human fingers are used to provide a friendly tangible user interface. The fingers form a rectangle region to indicate the location of the object to be tracked. Then, FinMars adopts optical tracking algorithm (KLT) and recognition algorithm (SURF) to realize a robust and accurate markerless AR system. The optical tracking algorithm is in charge of tracking the object in real-time. And the recognition algorithm helps to re-recognize the object when the tracking algorithm fails to track the object. We also design a mechanism that decrease detection area of recognition to reduce re-recognition time. Besides, in order to improve the performance of re-recognition, we design a partial update approach in FinMars. With such an approach, the region of interest is divided into smaller areas and only partial features are updated during tracking. Finally, we compare FinMars, SURF, and KLT in terms of accuii racy, computational time, performance of partial updates and occlusion to prove that FinMars has better results and performance, especially when occlusion occurs. iii