考慮具有大範圍移動物體的晃動影片之穩定與修復

一般徒手拍攝的影片之中，或多或少會受外力的干擾而造成影片的不規則晃動，對於日益成熟的影片穩定系統雖有不錯的成果，但由於其適用的情況在於影片主題佔據畫面部分較小的情況之下方可得到較完美的結果。換言之，前景的區域大時，或是相對於前景，背景的場景不夠完整時，可能會造成影片在對齊的過程之中參考區域的誤判，造成影片的不連續。本篇論文針對此情況提出一新的修正技術，目的於確保使用者的拍攝主體頗大時，或是背景常景較不完整時，經過本系統作些適當的修正後也能有其穩定畫面的效果。本系統運用optical flow計算出連續兩個影格之間每個像素相對移動的垂直及水平方向變化量，進而以k-means clustering方式適當地分割出影格畫面中不同移動型態的區塊，繼而針對適切的區塊進行攝影路徑的擷取方可得到較為正確的路徑。藉由取得的攝影路徑，本系統利用些許直線來代替原本的攝影路定以達到畫面的平穩性。本系統針對transform matrix運用了QR decomposition 以幫助我們將位移以及角度旋轉問題分開進行改善，同時影格盡可能保持原有的大小比例。最後以內插第一對影格與最後一對影格間的角度變化以達到畫面轉動的連續性以及其穩定性。由以上的步驟，即可得出一段看起來較為舒適及穩定的影片，甚至像是利用腳架拍攝出來的結果。

關鍵字

影片穩定；影片修復；手震；晃；移動物體偵測；前景背景分離

並列摘要

This thesis presents an approach to post-processing casually captured videos to improve apparent camera movement. A lot of home videos have some problems about the artifacts like hand shaking when capturing without tripod. Video stabilization is an important technique to solve this problem. However, the technique does not work very well in some situations such as the larger foreground, incomplete background or other situations,etc. In this thesis,we propose a novel method of video stabilization to overcome the situations with the larger foreground, or the feature points amount of the background is less then of foreground and some camera motion such as zoom in and zoom out. This system applies a method, optical flow, to estimate the motion vector of all pixels between each pair frames. Then, we use K-means clustering to group the similar motion vectors of each frame. To select an adequate segment to estimate the global camera path of video could obtain one more correct global camera path. After motion vector segmentation and camera path estimation, we could stitch all of video frames to get a panorama and estimate the range of moving object we can recover by neighbor frames. Based on the background panorama and moving object recovered range, we could find some new paths which would lose information less. After the above operations, a full-stabilized video could be achieved.

並列關鍵字

home video ； moving object detection ； optical flow ； polyline ； video ROI ； background segmentation

參考文獻

[IKN98] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual

[GL07] Michael Gleicher and Feng Liu. Re-cinematography: improving the camera dynamics

[BA96] Michael J. Black and P. Anandan. The robust estimation of multiple motions:

[SBF00] Hedvig Sidenbladh, Michael J. Black, and David J. Fleet. Stochastic tracking of

[LKK03] Andrew Litvin, Janusz Konrad, and William C. Karl. Probabilistic video stabilization

國際替代計量

考慮具有大範圍移動物體的晃動影片之穩定與修復

全文下載

主題瀏覽