使用多KINECT攝影裝置進行立體視訊監控之人物建模與追蹤

Abstract

本研究使用可視角達360度的八角形9-KINECT視訊裝置，該裝置由八台KINECT裝置往外監測，一台KINECT裝置往下監測組成，且提出一系列針對3D人物建模以及環境監控的相關策略和方法，進行立體視訊環境之建模及人物活動之追蹤監控。 本研究使用八角形9-KINECT視訊裝置來偵測人物，追蹤人物，並且在多KIENCT之間進行攝影換手；透過追蹤收集到人物資訊之後，利用資訊重新建立人物模型；完成人物模型後，利用提出的方法自動偵測人物相關的身體資訊。 更詳細來說，本研究複習了由馬秉晨及蔡文祥博士所提出的3D影像重建方法，接著提出了使用DWC Measure並且利用Evolution strategy避免困於區域最佳解的方法得到好的環境校正結果，並且使用該校正結果重建出環境模型。 除此之外，本研究提出一個演算法針對彩色深度影像作背景學習，以及提出一個方法利用3D connected-component labeling還有背景學習做人物的偵測，並且利用人物偵測結果來進行人物的追蹤與攝影機換手。 本研究利用人物追蹤所得的資訊之後，進行人物的建模與身體資訊擷取，並且提出了一個利用DWC measure以及randomized Kd Tree所加速的人物建模演算法。本研究找出部分有用的幾何性質，並且使用於尋找人物的最佳Bounding Box，藉此計算人物身體資訊，如身高，身寬，以及身體厚度等資訊使用於3D視訊監控。 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 algorithms are proposed for 3D human modeling and environment monitoring via the use of KINECT images for 3D video surveillance. An octagonal multi-KINECT imaging device to monitor the indoor environment is adopted. The octagonal multi-KINECT imaging device has a 360-degree view of the indoor environment, composed of eight KINECT devices looking outward and one KINECT device looking downward. With the octagonal multi-KINECT imaging device used as a 3D video surveillance system, methods for detecting humans, tracking human activities, and conducting handoff processes between the nine KINECT devices are proposed. After collecting the human data from the tracking process, a method for human modeling is applied. With the human model completed, a method for human body feature extraction is carried out and the result is shown to users. In more detail, firstly a method for 3D image construction using KINECT images proposed by Ma and Tsai [11] is reviewed. A method for calibration between KINECT devices using the DWC measure based on the evolution strategy is proposed, which can avoid being trapped in the local minimum and get a good calibration result. Then, a method using the mentioned calibration results and the 3D images converted from KINECT images is proposed to construct an indoor environment model. Furthermore, an algorithm for background learning using RGBD images is proposed. Also proposed is a method for human detection based on a background learning scheme and a 3D connected-component labeling technique. Afterwards, a method for human tracking is proposed, which uses the result from the human detection and conducts dynamic tracking after solving the handoff problem between the nine KINECT devices. With the data saved during human tracking, human modeling and human body feature extraction are conducted. A method for building up the human model by using the DWC measure and the randomized K-d tree structure is proposed. Then, a method for finding the bounding box circumscribing the human model is proposed. Some useful geometric properties are also exploited for finding an optimal bounding box. Finally, the bounding box is measured to compute human body features like height, width, and thickness for use in video surveillance and other applications. Good experimental results are also shown, which prove the feasibility of the proposed methods for real applications.