結合固定式廣角攝影機與活動式局域攝影機之監控演算法與合作策略

Abstract

在監視系統方面，攝影機擁有越廣的涵蓋範圍，則越能保證其區域的安全性。所以近幾年，有許多專注於多攝影機監控系統的功能性與可行性的研究。多攝影機系統一般以規劃攝影機的位置去達到較大的涵蓋範圍，或是以利用確認物體特性的方式給予其監測優先權。 而在本篇研究描述了兩種主要的監控場景，其中之一為廣域的公共區域監控；而另外一種為擁有多站點的室內環境監控。本篇第一部分為針對廣域公共區域監視（百貨公司、機場、大賣場…等）所設計的架構。然而，當在進行廣域的監視時，受限於解析度，攝影機很難去擷取到較詳細的資訊。所以在本架構設計中引入了活動式攝影機，以維持監視所需的解析度並且同時具有較廣的監控視野。 本監視系統為結合固定式全域攝影機與活動式局域攝影機。而為了達到多目標的物體偵測跟追蹤的目的，在執行廣視野監控的攝影機中，提出一些影像處理的方法。此外，在多目標追蹤中，另一個重要的課題為維持每個追蹤目標的標籤，本篇研究提出質心軌跡法（the trajectory of the center of mass）去解決這個問題。 並且提出座標轉換模型以達到有效地融合兩種不同的攝影機。然而，在沒有進行資源分配的動作下，在監控程序中系統冗餘將會持續地增加。所以本系統設計也同時導入合作策略以降低系統冗餘。 另一方面，在室內多站點的監控環境下（教室、工廠作業線、辦公室…等），攝影機需要對多個觀測點進行監控。而為了更進一步在固定式全域攝影機之間交換正確的資訊，也提出了改進物體偵測正確度的演算法。 在固有偵測物體演算法中，利用背景更新率去解決背景持續變化的問題。然而因為其固定背景更新率，在某些狀況下偵測效能較差。使用較低的固定背景更新率時，靜態物體會因為短暫時間內來不及更新為背景而造成偵測誤差。而在使用高背景更新率時，會將移動物體也更新為背景而造成偵測失誤。在本篇研究中提出了基於適應背景更新率的物體偵測法（motion detection with the adaptive background updating (ABU)）以提升正確性。在論文的最後，則展示監控系統與演算的模擬與實驗結果。

In recent years, much research has been focused on functionality and feasibility of multi-camera surveillance system. This study describes two types of surveillance scenarios. One is the surveillance of public monitoring and the other is the surveillance of indoor environment with numerous stations. In the type of surveillance of public monitoring in wide area, the proposed architecture is designed. On the aspect of surveillance system, the wider coverage guarantees the security of area. However, it is difficult to gather detailed information using the wide field-of-view (FOV) sensor due to its limitation in resolutions. In order to maintain the desired image resolution and still have a wide FOV, this requires the use of active camera. Thus, the current architecture design combined fixed global-view camera and active focused-view camera to make use of their advantage, respectively. Furthermore, the methods to achieve multi-target object detection and tracking are proposed. In order to maintain the identity of a moving object, the trajectory of the center of mass (TCM) is proposed to accomplish this task of labeling. To coordinate two different sensors, the model of coordinate transformation is derived. Without the act of resource assignment, system redundancy is increased during the surveillance process. Hence, the system aims to reduce this redundancy by applying the cooperative strategy. In the scenario of indoor environment with numerous stations (e.g., classroom, assemble line in factory, office), multiple observation points are required for visual sensors. The method of monitoring multiple points is proposed to improve the correctness of observed points for further information transmission with other global-view cameras. The performance of motion detection algorithm is occasionally poor due to its fixed background updating rate. The problem with the fixed low background updating rate is that static object is not updated as the background since the transient time is short. However, with the fixed high background updating rate, the result of updating moving objects as the background is not desired. To improve the correctness of detecting result, motion detection with the adaptive background updating (ABU) is proposed. Finally, the experimental results of different scenarios are shown in this study.