An Efficient Augmented Reality Architecture by Binary Tree and Optical Flow

[[abstract]]擴充實境在近年來已慢慢受到重視，它被廣泛的應用於產品行銷、廣告、醫療影像、工業設計、教育、遊戲等領域。目前大部分的擴充實境系統仍局限在自然環境和時間兩項限制，本篇論文就是使用強健的局部特徵描述子和穩定的追蹤機制來克服這兩項限制，讓使用者可以在任何環境使用任何喜愛的照片或圖片，就可以與視訊中的3D資訊或物體進行即時的互動，以達到擴充實境的效果。 本篇論文將廣義的介紹擴充實境系統的類型和追蹤技術，主要的焦點會放在以電腦視覺為主的無標記符號追蹤技術上，我們將深入探討此種追蹤技術在2003年到2008年間六種相關的擴充實境系統，並歸納出一個共通的擴充實境架構。首先，我們提出一種有彈性的二元搜尋樹來改善Bastos所提出的二元碼分群機制，得以大幅減少特徵樣版比對的次數。再來，我們將追蹤機制加入到共通的擴充實境架構中，然後選取強健的光流演算法來追蹤特徵點，以加速整個擴充實境系統。我們所提出的二元搜尋樹方法應用在靜態影像實驗上，在比對700個特徵點時，只需要花費大約500毫秒的時間，而在即時視訊影像實驗，我們提出的光流追蹤架構分別應用在兩種不同的局部特徵描述子SIFT和SURF上，實驗的結果都被證明非常有效率，其中最好的結果還達到每秒三十一個影格的速度。[[abstract]]Augmented Reality has been valued in recent years. It’s applied extensively in produce sale, advertisement, medical image, industrial design, and education and games so on. Most of the Augmented Reality Systems are still limited in natural environment and time those two limitations. This thesis uses robust Local Invariant Feature Detectors and stable tracking technology to overcome these two limitations and makes users can use any loved pictures or photos in any environments therefore they can have real-time interactions with 3D information and objects in webcam to achieve the effect of Augmented Reality. This thesis will introduce widely the types of Augmented Reality and tracking technology. The main focus is Markerless Tracking technology of computer vision. We will explore further the six relative Augmented Reality Systems of this tracking technology from 2003 to 2008.and to sum up a common structure of Augmented Reality. First, we propose a flexible Binary Search Tree to Improve the Binary code cluster systems that was proposed by Bastos so it can decrease drastically template matching. Then, we put Tracking systems in the common Augmented Reality Architecture and select robust Optical Flow algorithm to track feature points in order to speed up the whole Augmented Reality System. To apply the Binary Search Tree that we proposed in Static Image experiment only cost approximately 500 ms in matching 700 feature points. And in Real-Time Video Image experiments, the Optical Flow tracking Architecture that we proposed apply separately in two different Local Invariant Feature Detectors: SIFT and SURF. The result of experiments is proved very effective. The best result achieves the speed of 31 frames/sec