摘要
在以視覺為基礎的視訊安全監控領域內，背景模型技術是一個重要的元件，其可以協助系統進行背景的學習以及提供前景區域 的偵測結果。傳統上，背景模型技術大多使用在固定攝影機的拍攝條件下，其缺點為拍攝的範圍有所限制。近年來，動態攝影 機被廣泛使用在許多監控環境中以增加監控範圍，如銀行、機場、飯店大廳等，在動態攝影機拍攝狀態下進行前景物體偵測是 一個困難的問題，因為拍攝畫面中環境背景與前景物體皆處於運動狀態，針對此一問題，本論文提出了環場影像法以及視訊比 對法來進行動態攝影機之背景模型學習，這兩種方法採用不同的影像對位方式來估算攝影機的拍攝角度，並利用該角度之背景 模型來進行前景偵測與背景模型的更新。在偵測出前景區域後，我們使用以邊緣方向直方圖為基礎的行人偵測器與平均位移追 蹤演算法來偵測入侵人員及追蹤該人員。實驗結果顯示，本論文提出的動態攝影機之入侵人員偵測方法具有效率與可行 性。
並列摘要
Background model is a critical component in vision-based video security surveillance. It can help system to learn the background and provide foreground detection results. Typically, background model is employed under the use of stationary cameras. The drawback is that the field of view is restricted. To enlarge the monitoring area, dynamic cameras are widely used in recent years. For example, the surveillance systems in banks, airports, and lobbies. However, foreground detection under the use of dynamic cameras is a difficult problem, because the background and foreground are both in moving status in captured videos. To solve this problem, this thesis proposes two different background modeling methods, panoramic-based and video matching-based approaches, for dynamic cameras. These two methods adopt different image registration techniques to estimate camera viewing direction, and the corresponding background model can be used for foreground detection and model update. After foreground detection, HOG-based human detector and Mean-Shift tracking algorithm are both employed for human detection and tracking. From our experiment results, the proposed methods can provide efficient and feasible real-time intruders detection systems.