影像序列中網型物件之建立與追蹤

Abstract

在本文中，我們建構一架構來產生一影像序列的網型物件，而利用此網型 物件中節點的亮度，就可以重建回大略之影像。在此架構中，我們首先對影像作特徵點的偵測，在此我們用到的是影像曲面上高曲率的特徵點。接著會將這些特徵點連接成較長的線段，並利用 splitting的方式來找出一些節點及節點間的連線。再利用這些節點和節點之間的連線以及有限制條件的Delaunay三角化的演算法來產生此影像的網型物件來。我們也提出了一個階層式的架構來產生網型物件。利用此階層式的的結構，除了可以將影像依不同品質的需要產生不同複雜度的網型物件之外，我們更可以利用此階層式的架構，來幫助估測動態影像中網型物件的動態變化。

In this thesis, we propose a new way to generate the mesh objects of an image sequence. Based on the intensity information on the mesh nodes, we can roughly reconstruct the images back. Here, we first extract the high-curvature features in the images. Then we apply the splitting method to extract the representative nodes and the interconnection edges between these nodes. By using these nodes and edges, we apply a constrained Delaunay triangulation to generate the meshes of the images. We also propose a hierarchical way to generate mesh objects. With this method, meshes of different resolution can be generated to meet different quality requirements. Moreover, we can also use this structure to track the mesh object in the dynamic sequence.