使用一對雙環場攝影機成像系統對視訊監控車周圍的物體做監測之研究

Abstract

本論文提出了一個以視覺為基礎的視訊監控方法，該方法使用了兩組架設於監控車車頂的雙環場攝影機成像系統。首先，我們使用空間對應的方法建立起此環場攝影機成像系統的校正資訊表，稱之為全景對應表。基於此對應表以及兩個環場影像的對應點，我們提出了影像座標點與世界座標點之間的轉換方法。為了觀察監控車周遭的環境，我們另提出了建構上視圖的方法，以及將兩張上視圖合併成一張寬廣視角圖的技巧。此外，我們也使用一區域網路作兩台筆記型電腦的溝通管道，供傳輸移動指令，讓使用者能移動滑鼠即可建構出各種視角的透視圖。 另一方面，我們提出了一個自動偵測可疑人物並將其標記在上視圖的方法，這方法使用了一些影像處理的技巧，如距量保存影像二值化和灰階動態補償等。再者，利用一組雙環場攝影機成像系統所拍攝的影像組，我們能計算出可疑人物於立體空間中的距離及高度。如果監控車中的使用者想更直接的觀察可疑人物，他/她也能使用此系統而不需要走出車外即能觀察所建構出的對應透視圖。另外，我們也提出了一個自動偵測車輛的方法，在該法中，為了移除影像中的地板區域以及擷取出環場影像中的車體，我們使用了區域生長、組件標籤、圖像變換、模板匹配等影像處理技巧，並將之有效整合來獲得環場影像中車子的位置。最後，使用一組雙環場攝影機機成像系統所拍攝的影像組，我們可計算出車子位於真實世界座標中的位置。 良好的實驗結果顯示我們所提出視訊監控系統的可行性及應用靈活性。

Vision-based methods for video surveillance via the use of a pair of two-camera omni-directional imaging devices affixed on the roof of a video surveillance car are proposed. First, a space mapping method is used to construct the so-called pano-mapping tables of the pairs of two-camera omni-directional imaging devices. By the mapping tables and corresponding points of two omni-images, a method for converting the coordinates of the points between the image coordinate system and the world coordinate system is proposed. To see the environment around the video surveillance car, techniques for constructing top-view images and merging them into wider-area integrated ones are proposed. Also, a local network architecture for data communication between two laptop PCs, as well as a technique for constructing perspective-view images of any view direction decided by mouse clicks are proposed. Furthermore, a method for detecting a suspicious passer-by automatically and marking his/her position on a top-view image is proposed, which is based on image processing schemes of moment-preserving thresholding and dynamic grayscale offsetting. Moreover, the distance and height of a passer-by in 3D space is computed by image pairs taken with a two-camera omni-directional imaging device. If a user in the surveillance car wants to see a detected suspicious passer-by directly, he/she may use the system to generate a corresponding perspective-view image to inspect the suspicious passer-by without going out of the car. Additionally, a method of detecting a passing-by car automatically is proposed. To eliminate the ground region and capture the passing-by car shape in the omni-image, image processing techniques like region growing, component labeling, image transformation, template matching, etc. are used integrally and effectively to get the accurate position of the passing-by car in an omni-image. Finally, the position of the passing-by car in the real world is estimated as well using image pairs taken with a two-camera omni-directional imaging device. Good experimental results show the flexibility and feasibility of the proposed methods for the application of video surveillance.