在棒球視訊中以物件爲基礎的視訊檢索與事件偵測

Object-Based Video Retrieval and Event Detection for Baseball Video

杜維昌(Wei-Chang Du) ； 李健銘(Jian-Ming Li) ； 張簡大敬(Da-Jing Zhang Jian)

《電腦學刊》 17卷1期 (2006/04) Pp. 37-64

摘要

隨著電腦處理技術與硬體設備的不斷提升，利用數位攝影機拍攝各種場景已不再是那麼遙不可及的事。但由於攝影機只是真實呈現當時場景的影像，缺乏一些高階語意上的資訊，因而仍需依賴人類的認知。爲了提升視訊的價值，如何從中自動擷取出一些高階意涵將是必要的過程。在此一研究中，我們以棒球球賽視訊資料爲研究標的，針對固定式監視器所拍攝的視訊資料中有效偵測出運動事件。首先，我們考量低解析監視畫面中多移動物件追蹤之情形，利用視訊資料中空間性與時間性特徵作爲前景與背景分離的基礎，並結合連續畫面間的匹配與補償，以期可以獲得穩定的追蹤成效。接著，對運動行爲與特徵設計相關表達與比對方法，以利於特徵的有效使用。爲了進一步擷取高階意涵，開發有關棒球球賽的事件模式，以提供在時間軸上得到較爲高階的內涵。

關鍵字

監視視訊 ； 棒球視訊 ； 多物件追蹤 ； 視訊檢索 ； 事件偵測

並列摘要

In recent few years, the convenience and low cost of obtaining and storing digital video make people eager to have an effective tool for utilizing these data. From the viewpoint of human vision, digital video just keeps track of frames. Since it lacks for semantic information, it is necessary to understand the digital content with human brain. To promote the scope and value of further applications, automatic generation of high-level semantics is necessary. For a general sport video, motion behavior and characteristic of moving objects is an essential feature. In this paper, our goal is to develop an automatic extraction of motion-based feature via a fixed camera. More precisely, we firstly develop the motion-based segmentation and tracking of moving objects. Then, we design the indexing techniques of motion information for further efficient utilization. To demonstrate the effect of extracting motion-based features, we will build a tested platform of motion-based retrieval system. Based on this platform, we will develop an event-based model of baseball video so as to provide high-level captions on the timeline axis.

並列關鍵字

Surveillance video ； Baseball video ； Object tracing ； Video retrieval ； Event detection