透過KINECT影像做視訊監控應用上的立體環境建模與監視

Abstract

本研究設計一新的八角形9-KINECT視訊裝置，並提出一系列相關策略和方法，進行視訊監控上立體環境之建模及人物活動之追蹤。 首先，對於環境模型之建立，使用針孔成像原理將KINECT影像轉換成立體影像，進而使用幾何修正的方式，利用最小平方差橢圓曲面內差近似法，去修正立體影像的彎曲現象。接著，使用遞迴最近點(iterative closest point, ICP)演算法校正KINECT裝置間之空間相對關係。最後，使用校正出來的結果和立體影像，建立出室內環境的模型。 在使用八角形9-KINECT視訊裝置做人物追蹤方面，本研究首先進行人物偵測，使用的方法包括深度影像背景相減法、數學形態學操作和區域增長等技術。偵測到人物之後、進行人物活動追蹤時，會動態去改變KINECT裝置的仰角及處理KINECT之間的換手問題。 最後是建立人物之模型並擷取人物之特徵，應用於安全監視。對此，本研究藉由立體影像之序列，搭配距離權重相關係數(distance-weighted correlation, DWC)以及k維樹(k-d tree)之結構，建構出單一人物之立體模型，並從模型中擷取出人物特徵，如身高、體寬以及身體厚度，做為安全監視和事後觀看之用。 上述諸方法的實驗結果良好，證明在實際應用上該等方法確實可行。

In this study, several methods and strategies are proposed for 3D environment modeling and monitoring using an octagonal-shaped 9-KINECT imaging device for video surveillance. Firstly, an environment modeling method is proposed which, based on the pinhole camera model, converts KINECT images into 3D images. In the method, at first a new technique is employed to correct geometrically the bending phenomenon existing in constructed 3D images. The technique is based on the use of an MMSE paraboloid approximation scheme and a data interpolation scheme. Also, a technique is proposed to calibrate spatial relations between KINECT devices by the ICP algorithm. Finally, a technique using the calibration result and the constructed 3D images as inputs is proposed to construct the indoor environment model. Secondly, a human tracking method is proposed, by which human activities can be detected and tracked using the 9-KINECT imaging device. Specifically, a human detection process is conducted first, which includes the operations of background subtraction, mathematical morphology, and region growing. Then, during the human tracking process, the tilting devices of the KINECTs are used dynamically to track a human. The problem of handoff between KINECT devices, which occurs during the human tracking process, is also solved in this study. Finally, to extract the features of tracked humans for use in security monitoring, a human modeling method is proposed, in which sequences of 3D images constructed from KINECT images are integrated, using the distance-weighted correlation (DWC) measure and the K-d tree structure, to form a human model. From the model, human features like body height, width, and thickness may be extracted for use in security monitoring and off-line video search. Good experimental results are also shown, which prove the feasibility of the proposed methods for real applications.