視訊索引技術在視訊隨選系統與視訊圖書館系統中扮演相當重要的角色，但是龐大的視訊資料卻是造成瀏覽困難與搜尋不易的最主要原因。完整的視訊索引技術可提供的視訊影片索引方法及影片前處理，使得使用者可以快速搜尋與瀏覽影片。 視訊分段為製作視訊索引的第一個步驟，它可以將影片依鏡頭分割成數個視訊片段。我們使用色彩統計的方法來偵測場景變換與分割影片，其中我們使用了四種不同的色彩空間，共十六種色彩量化的方式來檢驗連續鏡頭中色彩統計的差異值。 當視訊分段完成之後，第二個步驟為關鍵畫面的萃取。關鍵畫面代表著此一視訊片段以供使用者查詢。然而，人們對於關鍵畫面的選擇是以劇情上的代表性為選取標準，所以我們提出使用鏡頭運鏡方式與視訊片段邊界為選擇關鍵畫面的標準；鏡頭運鏡的三個主要的動作分別為：panning、zooming、tilting。將每一個視訊片段偵測出其主要的動作，再依其動作決定關鍵畫面的選擇方式。我們使用光流場來決定鏡頭運鏡的方式，而以Lucas的方法來計算畫面中的光流場分佈。 Video Indexing plays an important role in video libraries and video on demand systems. However, the task of browsing or querying a huge amount of video data is difficult. They require that the source material should be first effectively indexed. The first step of video indexing is video segmentation, which partitions a video into individual camera shots. In this thesis, we use the histogram-based algorithm to test the difference metrics between adjacent frames and present scene change detection algorithm. Sixteen methods of color quantization for scene change detection are considered in 4 types of color spaces: RGB, HSI, HSV and YIQ. We also compare the performances of different color spaces and their color quantizations on two types of video sequence: film and education video. The second step is key frame extraction. In a shot, a key frame is selected to represent the shot. The key frame depends heavily on the perception of people. In our work, we combine the shot boundary approach and camera operation models to establish our key frame selection rules. The key frames are selected according to the camera operation models: panning-like, tilting-like and zooming-like sequences. Lucas and Kanade’s method is adopted for estimating optical flow in this step. Experimental results demonstrate the feasibility and effectiveness of our proposed video indexing system.
