視訊中區塊遞移之研究

Abstract

本論文中實作了一個追蹤區塊遞移的演算法，不同於其他相關研究中使用移動向量來預測下一張影像中的區塊位置，我們讓每一張影像中的切割盡可能完整，因而能用較直觀的方式來找到連續影像之間區塊的對應關係。我們提出的方法是基於區塊擴增演算法的基礎，加上一些區塊合併的機制讓相似的區塊合在一起，並透過不同的區塊擴增結果及時間回饋的概念，希望過去好的結果能盡量被保留，讓區塊的大小及形狀更穩定。區塊遞移的判斷利用相同區塊在連續影像中的交集、相似性等容易理解的概念，能達到不錯的結果。最後，我們對區塊遞移的追蹤結果進行分析，以評估所提出的方法是否達到理想中的正確性及穩定性。

In this thesis, we implement an algorithm for tracking the region propagation in a video sequence. Unlike ordinary objects tracking approaches which detect motion vector of each image block to predict the position of a region in the next frame, we try to focus on reasonable region segmentation in each frame and then track the regions by some simple and starghtforward methods. Our segmentation method is based on region growing. To avoid over-segmentation, we use some criteria to merge similar regions. We try to integrate the different segmentation results for same image as well as, by adopting the concept of feedback, from consecutive images. Since developed algorithm can generate stable segmentation results, characteristics of regions’ intersection and similarity are used to track these regions of them. Finally, we analyze the tracking results to show that the proposed approach works reasonably well in terms of correctness and stability.