主從式影像監測系統之研製與生態監測應用

Abstract

影像監視系統已經廣泛的應用於日常生活之中，其中主從式影像監視系統可以同時提供寬廣的影像和高清晰的影像。主控端攝影機負責監視寬廣的影像範圍，並且透過電腦控制被控端攝影機 (旋轉可變焦攝影機) 取得高清晰的影像。主控端攝影機通常是具備廣角鏡頭或是魚眼鏡頭的攝影機，但是影像會有嚴重變形以及解析度過低的問題。因此，本研究提出一種新的主控端攝影機架構，由八顆網路攝影機組成環場攝影機組，應用影像縫合技術，同時提供寬廣的視野、微小變形且高解析度的影像。另外，我們建立新的幾何對應技術，將環場攝影機的座標轉換成旋轉可變焦攝影機的座標，據以控制旋轉可變焦式攝影機擷取清晰的目標影像。此系統進行影像目標追蹤時引入了互動式多運動模型，增進預測所追蹤目標位置的能力，提高了取得高解析度影像的成功率。所建置的系統實際應用於行人監測與生態池監測，可以記錄目標物的高解析度影像以及記錄追蹤目標的移動軌跡，其中環場攝影機解析度達到4390 × 587像素，速度可以達到每秒10張影像。幾何座標對應方法誤差約為0.5度，整體取得高清晰影像以及軌跡紀錄成功率約為70%，而高解析度影像以及目標的移動軌跡有助於目標的行為分析。

Video surveillance has been widely used in various applications. The master-slave imaging system architecture can provide both a large field of view (FOV) and high resolution images. A master camera is responsible for monitoring large FOV, object detection, and guiding the slave camera. A slave camera is usually a pan-tilt-zoom (PTZ) camera, which rotates and zooms in to acquire high resolution images of targeted objects. In traditional approach, a camera with wide angle or fish-eye lens is usually used as the master camera. However, such an approach is limited in applications requiring high resolution image. Instead, we propose a new kind of master camera, which is a panoramic camera set integrated with eight webcams. We employ the panorama technology to provide video images with large FOV, low image distortion and high resolution simultaneously. Moreover, we develop a new geometrical mapping method to achieve coordinate transformation between the panoramic camera set and the PTZ camera. We also apply the interactive multiple model to improve the estimation of the targeted object state, which facilitates the PTZ camera to center on the targeted object for proper image acquisition. The proposed system has been applied to pedestrian and ecological pool monitoring to test its performance. The developed master-slave imaging system provides high resolution images and the trajectory of the targeted objects. The panoramic camera is capable of acquiring video images of 4390 × 587 resolution at rate of 10 fps. The mapping error is around 0.5 degree. The overall tracking rate is about 70%. The high resolution images and the recorded trajectories are useful in further analyses of the behaviors of targeted objects.