在視覺追蹤時，有時無法避免在影像擷取時受到鏡面反射(specular reflection)的影響。由於現有文獻很少探討此類問題對追蹤效能的影響，因此本論文提出基於共推論融合的強健視覺追蹤演算法，利用反射分離(reflection separation)獲取混合影像(mixed images)之中的照度影像(illumination image)，以估算無反射遮罩(non-reflection mask)，指出混合影像之中的無反射的區域，建立較準確目標物的色彩分布模型。為了避免連續畫面中反射影像隨相機動量而產生變動，因此在分離反射之前先計算相機動量，並補償相機動量於混合影像上。本論文的整體的追蹤系統採用粒子濾波器(particle filter)為基礎，將RGB色彩分布資訊及無反射遮罩所建立的[I、R-G、Y-B]色彩分布資訊，將兩者資訊交互影響與融合，最後利用最大化似然機率(maximum likelihood)優化每顆粒子的權重。實驗結果中顯示，本論文提出的追蹤演算法能有效的克服在行動相機的情況下，混合影像序列所受到相機動量的影響，因此能有效的提高追蹤準確性。;For visual tracking, mixed images cannot be avoided since the transmitted scene may be captured with specular reflections. There are few previous method tackling this important problem, thus this paper proposes a novel robust visual tracking method using co-inference fusion for mixed sequences. Based on the framework of particle filter with compensated motion model, this paper adopts the co-inference method to fuse two types of color measurements of the target. Although both measurements are observed from the same mixed image, one of them is built based on a non-reflection mask, constructed from the illumination image. The proposed scheme adopts reflection separation to derive an illumination image and a reflection image from mixed images before tracking. To satisfy the time-invariant assumption of a reflection image, camera motion is compensated on each mixed image before reflection separation. Finally, the weight of each particle is individually optimized using maximum likelihood. Experimental results show that the proposed scheme effectively improves the tracking accuracy on mixed sequence with camera motion.
