在行動平台上的相機追蹤及其應用

Abstract

本論文提出了一個在智慧型手機上實作相機追蹤的系統這個系統可以估計出相機在3D空間中的位置。在機器人領域有一項發展多年的技術，是同時做機器人定位和場景繪製的技術，稱為SLAM。然而在電腦視覺領域，這類問題解決的方式通常是用移動中重建場景的技術，稱為SfM。然而這兩項技術都會遇到錯誤累積造成追蹤結果會逐漸飄移，為了克服這個缺點，找一個比較標準是必要的。在SfM中，某些相機位置會被當成固定的標準，其他所有的畫面對這些標準位置最對齊。雖然準確但是計算所花的資源跟時間都很多。另一方面，在SLAM系統則是選擇某些畫面當成關鍵面，用這些關鍵畫面對建立好的場景模型做相機位置的最佳化。此方式計算量比較小，但比較不準確，因為通常機器人的移動方式比手持的相機移動方式簡單很多。本論文提出的系統結合了上述兩者的優點，並平行化追蹤跟場景模型建立(包含SfM的步驟)兩件事情。透過簡化計算，將此系統開發在智慧型手機平台上。此項技術對現今流行熱門的擴充實境(AR)有非常大的幫助。

This thesis present a camera tracking system on mobile device. The system estimate 3D camera pose in an unknown scene. The simultaneous localization and mapping (SLAM) technique in robotics has been researched for many years, however this problem, in computer vision, is called structure from-motion (SfM). To overcome the drift problem caused by accumulating the computing error, it is necessary to create a global model as comparative criterion. In SfM, some image frames are fixed, and the bundle adjustment over whole video sequence is accurate but computational expensive. On the other hand, SLAM system maintains a global map and selects some frames as key frames for camera pose optimization, which is faster than SfM, but less accuracy when estimating a free moving camera since the motion of robot is limited. The proposed system takes advantages of both SLAM and SfM, and parallels the tracking and mapping (including SfM) procedure. By reducing the computational cost, it is capable to develop our system on mobile device with camera. This technique is useful for augmented reality (AR), which is promising and desired recently.