利用旋轉攝影機進行具時間軸之互動式環場影像建構

Abstract

這篇論文是討論如何應用記憶視覺技術去表現出背景的沿革。過去有許多的應用，例如數位保存以及觀光景點。一般而言，捕捉到的影像通常為有限的視野；要瀏覽寬網的景象，一個可行的方法是用有多功能及靜止功能的相機作各個不同角度的攝影。然而這個方法有許多的缺點。舉例而言，不同角度圖像間的關聯容易模糊，而且多功能相機的成本可能較高。近年來，搖攝相機廣泛的被應用，例如視覺監看以及視覺機器人進而擴大監看範圍。透過使用搖攝相機，各種的活動景象可以循環性的被構建起來，而且景象可以直接被組織化。要生成一個理想的影像背景而不被前景所影響，一個封閉性的背景模式常被用來預先進行影像偵測及移動物的消除。我們最後使用中位數過濾器創造出一個視覺效果更好的時空全景象。這個結果會透過一個可允許使用者參與景象互動的介面顯現。

This thesis discusses how to apply vision-based techniques to exhibit the evolution of a background scene through a period of time. It has many applications, such as digital preservation and tourist attractions. Ordinary captured images usually have a limited field of view. To browse a wide scene, one possible way is using multiple static cameras to monitor different viewing directions. However, with this kind of method there are certain drawbacks. For example, the relationships between different viewing directions could be confusing and the cost for multiple cameras could be costly. In recent years, panning cameras are widely used in many applications, such as visual surveillance and vision robot, to enlarge the monitoring area. With the use of a panning camera, a manifold panorama can be constructed in each round of the rotation, and a scene can be organized in an intuitional way. To generate an ideal background panorama without foreground influence, a block-based background modeling method is employed to pre-process the image for detecting and eliminating moving objects. We finalize by applying mean or median filter to create a space-time panorama for giving a better visual results. The results are demonstrated by an interactive user interface which allows users to be immersed with the panoramas.