利用電腦視覺技術以圓形資訊作物體定位之研究及其在運載工具自動化和虛擬實境之應用

Abstract

在電腦視覺領域中，物體定位是一個基本且重要的課題，此一問題即如何透過物體上的一組特徵來決定相機與物體之間的相對位置。由於在許多的日常的環境中圓形與直線是隨處可見的資訊而常被運用，因此在本論文中我們首先針對不同個數之圓形的情況，利用電腦視覺技術提出了四個結合圓形與直線資訊進行物體定位的方法：第一個方法為跟相機座標系統只有偏角的單圓定位、另外三個方法分別為同平面上之同半徑雙圓、同平面上之不同半徑雙圓及同平面上之兩同心圓的定位。本論文並實際應用此四個定位方法於運載工具自動化和虛擬實境等四種環境，分別概述如下： 對於在道路上行駛的車輛，在轉換車道時如何偵測旁車道上有無汽車；以及當有車輛自旁車道轉入本車道時，如何即時偵測該車輛；兩者對於輔助安全駕駛而言都是很重要的工作，然而以往之自動車研究很少對此一狀況進行探討。在本論文中我們提出兩個新方法來處理此一問題，第一個方法為利用旁車後輪鋼圈之圓形資訊來作旁車定位，並利用鋼圈所在平面之方向來預測旁車車頭之方向。同時針對影像中鋼圈邊界的橢圓特徵提出一穩定的影像偵測技巧。本論文之第二個方法為利用旁車前後兩輪鋼圈之圓形資訊偵測影像中車輪與地面接觸點，之後再求得該點之三度空間位置來作旁車定位。 針對已經求得的旁車定位資訊，我們推導出兩種預測旁車行進路線的方法。第一種為利用連續兩張影像所偵測之資訊來預測旁車行進路線；第二種則為利用在單一影像所偵測之空間資訊（例如車前後輪的方向及位置），來預測旁車行進路線。兩者皆可利用來達到偵測旁車及避碰之目的。 在空中運載工具的環境中，航空飛行員的負擔十分沉重，自動化或輔助飛行無疑是一大幫助。一般飛行自動化之研究大都只考慮以固定之標誌特徵來計算直昇機相對落地點之位置，而達到輔助飛行之目的。但是在直昇機降落的過程中，隨著高度的降低，直昇機上攝得之落地點標誌在影像中會愈變愈大以致原來選定之標誌特徵不再適用。本論文第三個方法針對此問題，提出一個階層化的新概念來作直昇機定位。此一方法隨著直昇機高度的改變，自動採用落地點標誌影像中更細部的圓形與直線之資訊，以求得更精確之直昇機相對落地點的位置，而達到輔助飛行之目的。 隨著虛擬實境及網路視訊會議應用的與日俱增，利用攝影機作為輸入裝置的需求也相對增加。本論文第四個方法針對此類需求，設計了一套快速校正照相機和電腦螢光幕之相對位置的方法。在此類應用中相機通常是架設在螢幕旁隨時偵測使用者的動作，因此螢幕並不在相機的視野中，本論文利用一個特別設計其上附有小型雷射光頭的眼鏡作為輔助工具。第一步先求得相機相對於眼鏡的位置(以二條平行線及二個圓為物體幾何特徵)，第二步利用螢幕上特定點相對眼鏡的位置和第一步結果可得螢幕上特定點關於相機的相對位置，最後求得相機相對於螢幕座標的相對位置。同時針對虛擬實境及人機互動的三個應用亦作詳細之數學公式推導。 最後，本論文所提的各個方法均具有公式解，且都已由實驗證明其正確性及可行性。

In the field of computer vision, one of the basic and important problems is object pose and location estimation. This kind of problem means to determine the relative position and orientation between a camera and a certain feature such as a point, a line, or a curve detected on an object. Among various kinds of features, circles are very popular in the man-made environments. In this study, one purpose is to explore feasible and effective object location approaches by using combinations of features of circles and lines. Another purpose is to create new computer vision applications, in which the features of lines and circles are utilized to solve the problem of object pose and location estimation. Four methods for object pose and localization of circles are proposed in this study. Then, their applications are discussed, including moving vehicle localization, moving vehicle trajectory prediction, automatic rotorcraft landing, and camera calibration with respect to a computer monitor for virtual reality applications. An approach to estimation of moving lateral vehicle locations for driving assistance using one-wheel shape information in single 2D vehicle images by 3D computer vision techniques is proposed. There are still very few studies on the detection and localization of lateral vehicles. In this study, the wheel shape information of the lateral vehicle is employed to locate the lateral vehicle, in which the wheel information comes from images captured with a camera mounted on the driver’s vehicle. Then, the trajectory of the moving lateral vehicle is predicted by using the temporal information of single wheel shapes in consecutive image pairs. In addition, another approach to estimation of lateral moving vehicle locations and prediction of their trajectories for driving assistance using two-wheel shape information in single 2D images by 3D computer vision techniques is also proposed. The radius of a wheel need not be known in advance. No special mark on the vehicle is required. The 3D relative positions of the contact points of wheels and ground with respect to the camera are obtained by the back-projection principle as the desired relative location of a lateral vehicle. Then, the trajectory of a lateral vehicle is predicted from the spatial information obtained in single images Automatic landing system can alleviate heavy burdens on aircraft pilots. In this study, a new hierarchical approach to landmark localization by computer vision techniques is proposed. Most of the existing related vision-based methods for aircraft landing consider only a situation, in which the entire landmark is in the field of view of the camera. The proposed method instead achieves aircraft location estimation by four hierarchical stages according to the aircraft flight altitudes. In different stages, the features of different parts of the landmark are utilized to estimate the landmark location. Finally, a new approach to camera calibration with respect to a computer monitor is proposed. The camera location parameters, including three position parameters and three orientation ones, obtained from the proposed calibration method are useful for many virtual reality applications in which a camera is mounted on or located near the monitor to “look at” a user in front of the monitor. The calibration problem is solved indirectly by introducing an auxiliary tool, a pair of specially-designed eye-glasses, with a laser pointer mounted on the nosepiece between the two glass frames. Three applications of the proposed method, namely, VR scene display, cursor control by the head direction, and face view rectification for preserving eye contact in video conferencing or Internet telephony, are also discussed, with their respective mathematics derived in detail. Experiments have also been conducted with good results to prove the correctness of the derived equations and the feasibility of all the proposed approaches.