利用全向影像偵測室內環境消失點與估測旋轉量

Abstract

機器人或無人載具之自動導航越來越受重視，其與旋轉估測之正確性極為相關。全向相機相較傳統相機擁有更大的視野，可以抓取更多周遭環境中的視覺資訊，而使用全向影像中的垂直線資訊，供地面載具計算水平旋轉角度的方法已被提出。但在三維環境中，多種無人載具(如 UAV)並非僅行動於水平面上，故需提出三維空間中估測旋轉角度的方法，而環境中的消失點有助於判識機身的旋轉情況。本論文藉由全向相機所攝之全向影像，由中提取消失點以及估測旋轉情形。 本研究先以邊界偵測找出全向影像中的邊界點，並將這些邊界點匯集成連續線段，經由直線偵測與消失點提取之後，可得環境中多組線段對應之消失點的方向資訊，運用於後續旋轉估測分析。並藉由全向影像中的直線與現實空間中的直線對應關聯，發展全向相機校正模式，透過室內環境之多組全向影像試驗，與追蹤垂直線方法的結果進行探討。

Automatic navigation for mobile robots or unmanned vehicles becomes a more and more important ability, and this ability has a great deal to do with accurate rotation estimation. An omnidirectional camera, with a wider field of view compared with a traditional camera, can catch more visual information from the surroundings. The method using the information vertical lines in omnidirectional images is proposed to calculate two-dimensional rotation angles of the camera. But in three-dimensional environment, unmanned vehicles such as UAVs can move not only on a two-dimensional plane. It is necessary to propose methods to estimate three-dimensional rotation. Vanishing points in the environment would be the effective information for estimating the rotation of robots or unmanned vehicles. And in this thesis, the methods are proposed to solve the vanishing points extraction for omnidirectional images taken by an omnidirectional camera and to estimate rotation of the omnidirectional camera. For vanishing points extraction, we use edge detection for omnidirectional images to find out the boundary points and then link these points as line segments. After line detection and vanishing points extraction for these line segments, the directions information of vanishing points in the surroundings can be obtained and be applied for rotation estimation later. For camera calibration, we propose a method using lines in omnidirectional images to obtain camera intrinsic parameters, and the camera intrinsic parameters will be applied for line detection and vanishing points extraction later. Vanishing points extraction for omnidirectional images taken by the omnidirectional camera in the indoor environment would be applied to estimate the camera’s rotation. The results will be shown and compared with the results calculated by the method using vertical lines tracking.