主動式監控攝影機之變動偵測與動態追蹤

Abstract

在本論文當中, 提出了一套使用主動式攝影機的即時影像監控系統。這個系統能透過控制攝影機的左右及俯仰轉動，使所要追蹤的物體保持在影像的中心部分。首先，使用半全域移動補償(semi-global motion compensation)來產生移動邊緣圖(moving edge map)，此圖可以明確地指出移動像素點的位置。接下來，以像素為基礎的移動邊緣圖會被轉換成以方塊為基礎的移動物件圖(moving component map)。為了使移動追蹤的效果更穩定，先前影像的移動物件圖會被加以整合，來加強現在影像的移動物件圖。接下來，使用連接物件方法(connected component analysis)從移動物件圖中找出移動物體，而互相重疊的物體將被整合成單一的物體。此時分析出的各個物體都被視為可能的追蹤目標。為了進一步強化追蹤的效果，(Kalman filtering)被使用來實做一個簡單但有效的追蹤視窗(tracking window)。藉著量測每一個候選物體與追蹤視窗的距離遠近以及大小相似度，具有最大相似度的候選物體將被視為所要追蹤的物體。最後，根據追蹤到的物體的位置，此追蹤系統能自動控制攝影機進行左右及俯仰轉動，來持續保持追蹤物體始終落在擷取影像的中心位置附近。

In this thesis, we propose a real-time surveillance system on an active camera. This system can control the movement of an active camera, via panning and tilting, to keep the tracked object around the image center. In this system, a newly proposed scheme using semi-global motion compensation is employed to extract a moving edge map, which indicates the occurrence of moving pixels. This pixel-based moving edge map is then converted to a block-based moving component map. To increase the robustness of motion tracking, the moving component maps of previous frames are pieced together to underscore the current moving component map. Then, the connected component analysis is used to identify moving objects. Overlapped moving objects are combined together and the candidates of moving objects are extracted. To further improve the performance of motion tracking, the Kalman filtering technique is used to implement a simple but efficient prediction of the tracking window. By measuring the similarity between the location of each candidate and the location of the predicted window, the candidate with the highest similarity is selected as the target object. At last, according the location of the target object, the surveillance system controls the camera to pan and tilt to keep the object around the center of the captured scene.