整合多攝影機之三維監控系統

Abstract

本論文發展一套視訊監控系統，整合多支攝影機的畫面到三維模型上，並擷取影像中的移動物體加以立體化，呈現一個具空間感且可即時互動的監控畫面。常見的傳統監控系統是利用電視牆來顯示所有攝影機的畫面，監控人員往往很難了解每個畫面之間的空間對應關係，並且容易混淆，在情況緊急時，可能無法做出正確的判斷。透過我們的系統，監控人員可以任意選擇想要觀看的視點，且攝影機的空間關係已融入監控環境的三維模型。如此一來，同時監控多支攝影機不再是一件困難的事。 本系統可以分為兩部分，建立攝影機和三維模型的對應關係，和擷取移動物體加以立體化。首先，我們使用平面補釘模型化(planar patch modeling)建立三維模型，然後在監控場景中放入一些標記，量取標記之間的相對位置關係，再利用單應性矩陣(homography matrix)來表示攝影機畫面和監控場景的對應關係。利用這些資訊建立查詢表(lookup table)，來融合出三維的監控場景。接著，使用背景模型(background modeling)偵測移動物體，並使用告示板(billboarding)技術來立體化顯示此物體的影像。最後我們得到一個三維視覺化的整合監控系統。

In conventional surveillance systems, multiple screens are often required for displaying video from multiple cameras and may cause the difficulty of operators to keep track of targets due to the lack of special relationship among the screens. In this thesis, we develop an effective surveillance system with the 3-D environment model that integrates multiple scenes into one single comprehensive view. The system does not require accurate camera calibration and environment model construction with advanced equipments, and can provide a multiscale operating view for showing the status of the surveillance area. To integrate the monitored area with camera views, the 3-D environment is first manually constructed by planar patch modeling. To map video contents to the corresponding areas of the 3-D model, different homography transformations are estimated for every pairs of image regions in the video contents and corresponding areas in the 3-D model. The planar patches in the 3-D model are automatically divided into different sizes and numbers of the smaller patches are determined by the intensity differences between the image polygons obtained from the homography transformation and texture mapping. Lookup tables are built beforehand for accelerating the coordinate mapping. In monitored scenes, 3-D objects including pedestrians, are projected to the imaging sensor planes through perspective transformation. Therefore, all 3-D objects will appear flattened on the planes in the 3-D model after texture mapping. To overcome this problem, we use billboarding method to model the 3-D moving objects. First, each foreground object extracted from the scene by background modeling is mapped to a billboard, and then vertically aligned to the ground plane. In this way we can adjust the direction of foreground objects according to the viewing direction. The proposed system can provide operators the situational awareness of the monitored site, including activities of the tracking targets through a comprehensive 3-D view.