透過八角形的多Kinect攝影裝置進行立體視訊監控

Abstract

本研究提出一個利用八角形多Kinect視訊裝置做三維視訊監控的系統。這個八角形的多Kinect視訊裝置由八台朝外拍攝的Kinect和一台朝下拍攝的Kinect組成，固定在建築物的天花板上，監控底下的室內環境，可以得到360o的環境場景。這個系統的功能包含三維人物影像建構、三維環境影像建構，以及在環境中人物活動的監視與顯示。為了要實作出這些功能，本研究提出了數個方法和策略。 首先，本研究採用使用Kinect攝取影像來建構三維影像的方法。接著，提出一個新的距離度量，亦即以顏色做過濾的距離權重相關係數(color-filtered distance-weighted correlation, CFDWC)，用以校正兩張連續拍攝的Kinect影像的相對關係，並據以建構出完整的三維人物影像。此外，為了要處理在所建構三維人物影像上所出現的一種模糊現象，本研究提出一尋找出最大的或最合適的人物部分的解決策略，並提出一基於膚色過濾和凱尼(Canny)邊緣檢測來分離人物四肢的處理方法。 其次，基於CFDWC，本研究亦提出一環境建構的方法，可用於疊合從八角形多Kinect攝影設備所攝得的背景影像，建立出360o的室內環境影像。本研究還提出一從各個視角觀看完整三維人物的方法，該方法使用人物特徵擷取及回歸分析技術來決定人物的三個方向軸，之後依照這三個方向軸展示完整的三維人物。 基於上述諸方法的實驗結果良好，證明其在實際應用上確實可行。

In this study, a system for 3D video surveillance by an octal-shaped multi-Kinect imaging device has been proposed. Affixed on the ceiling of a building for monitoring the indoor environment below, the octal-shaped multi-Kinect imaging device has a 360o view of the environment, composed of eight Kinect devices looking outward and one Kinect device looking downward. The system includes the functions of 3D human image construction, 3D environment image construction, and monitoring and displays of human activities in the environment. To implement such functions, several methods and strategies have been proposed. Firstly, a method for 3D image construction using Kinect images is adopted. Then, a method for calibration of the relevant transformation between every two consecutive Kinect images using a new distance measure, color-filtered distance-weighted correlation (DWC), is proposed for use in constructing a complete 3D human image. Furthermore, in order to handle the blurring phenomenon appearing in a completed 3D human image, a strategy for selecting typical image frames with largest or proper human body parts and a method for splitting human limbs in the images based on skin-color filtering and canny edge detection are proposed. Secondly, an environment construction method based on the color-filtered DWC measure is proposed, which can be used to merge the background images acquired by the octal-shaped multi-Kinect imaging device to build up a 360o indoor environment image. A method for viewing the complete 3D human image in the environment image from different perspective view is also proposed, which uses the techniques of human body detection and regression analysis to decide the three axes of the human in the 3D image and displays the complete 3D human image accordingly. Good experimental results are also shown, which prove the feasibility of the proposed system and methods for real applications.