在360度視訊的電腦視覺應用中，物件追蹤(object tracking)可縮小待處理之感興趣區域(region of interest)，而對於三百六十度視訊(360-degree video)的等距長方投影(equirectangular mapping projection, ERP)，由於影像扭曲，造成大多數現有的物件追蹤演算法準確率大幅降低。因此，本論文基於粒子濾波器(particle filter)，提出於球體域(sphere domain)預測目標物狀態，使運動模型(motion model)不須考量等距長方投影中幾何形變(geometric distortions)的問題。由於ERP為全景影像，因此需進行模數(modulus)處理，使得預測後粒子之位置避免落於ERP邊界之外。此外，由於Oriented Fast and Rotated BRIEF (ORB)特徵具有影像幾何失真(geometric distortions)之不變性，且計算複雜度較低，因此本論文採用ORB描述子(descriptor)以進行粒子濾波器之更正階段(correction stage)。實驗結果顯示，相較於快速L1追蹤器，本論文所提方案雖然於目標進入ERP影像邊界前的幾張畫面時，追蹤準確率略低於快速L1追蹤器，但是當目標進入ERP影像邊界後再出現在ERP影像另一側，或是當目標被遮蔽時，本論文所提方案追蹤準確率皆高於快速L1追蹤器。;For applications of 360-degree videos in computer vision, object tracking can reduce regions of interest to be processed. For equirectangular projection of 360-degree videos, the distorted images significantly decreases tracking accuracy. Hence, this thesis proposes to predict the target state in the sphere domain in the framework of particle filter. Accordingly, the motion model does not need to consider geometric distortions in the equirectangular mapping projection. Because images of equirectangular mapping projection are panorama images, the modulus operation avoids the positions of the predicted particles out of the boundary of the equirectangular mapping projection. In addition, since the Oriented Fast and Rotated BRIEF(ORB) feature is geometric distortions invariant and is featured with lower computation complexity, this thesis adopts the ORB descriptor at the correction stage of particle filter. Experimental results show that the proposed scheme outperforms the fast L1 tracker when the target appears at the other side of the equirectangular mapping projection after the target moves into the boundary or the target is occluded. However, the fast L1 tracker slightly outperforms the proposed scheme before the target moves into the boundary.
