室內自動車航行之研究：攝影機校正、模式比對導航、及障礙物避碰

Abstract

Proposed in this dissertation are four new, feasible and effective approaches to several cintral problems for indoor autonomous vehicle navigation by computer vision, including camera calibration, model-based guidance,and collision avoidance. Camera calibration must be performed for a vehicle to conduct vision-based autonomous navigation. Two approaches to camera calibration using vanishing lines are proposed, both showing intuitive relationships between scene images and camera parameters. In the first approach, a hexagon is employed as the calibration target to generate a vanishing line from its projected image. It is shown that the orientation, position, and scaling information of the vanishing line are analytically related to the camera orientation parameters and the focal length. The second approach extends the concept of the first approach to the three-dimensional case, in which a rectangular parallelepiped is used as the calibration target to generate three vanishing lines from its projected image. Several attractive results are shown, including the facts: (1) the orthocenter of the triangle formed by the three vanishing lines is just the image plane center; (2) the area of the triangle implies the value of the focal length; and (3) the slopes of the three vanishing lines determine the camera orientation parameters. The camera position parameters are calibrated by simple geometric projection relationships in both methods. In addition, an autonomous vehicle should be able to locate itself in an environment in navigation sessions. A model-based guidance approach based on the longest common subsequence concept is proposed to guide a vehicle in an indoor corridor environment. A map of the corridor contour, describing the navigation environment and measured before navigation sessions, is used as the model for guidance. Two cameras mounted on the vehicle are used as the visual sensors. The wall baselines in the images taken form the two cameras are extracted and constitute the input pattern. The input pattern and the environment model are represented in terms of line segments in a two-dimensional floor-plane world. By encoding the line segments into one-dimensional strings, the best matching between the input pattern and the environment model is just the longest common subsequence of the strings. So no complicated image matching is needed and robust matching results can be obtained. Successful navigation sessions on an experimental vehicle are performed and confirm the effectiveness of the proposed approach. Collision avoidance is also indispensable for autonomous navigation because a vehicle may face unknown static obstacles or unexpected moving obstacles during navigation. A new approach to collision avoidance using a modified version of least-mearn-square-error (LMSE) pattern classification is proposed to plan a local safe path in each navigation session. The trajectory of each obstacle is predicted by a real-time LMSE estimation method. And the maneuvering board technique use in nautical navigation is employed to dynamically determine the vehicle speed. Good simulation results are presented to show the feasibility of the proposed approach. 本論文利用電腦視覺作為室內自動車航行的依據，研究的主題包括攝影機校正、模式 比對導航及障礙物避碰。本論文首先提出兩種利用消失線作攝影機校正的方法，證明 攝影機係數反應在影像平面上之幾何直覺意義。第一種校正方法利用一個平行六邊形 在影像平面上的投影產生一條消失線，並證明出消失線的方向、位置和比例關係決定 了攝影機的方向參數及焦距。第二種校正方法利用一個長方體當作校正目標，由其在 影像平面上的投影產生了三條消失線，可得到此三條消失線與影像平面中心點位置、 焦距及攝影機方向參數之幾何關係。而兩個校正方法均利用簡單的幾何投影關係來計 算攝影機位置參數。 針對自動車在行進中必須有能力決定本身在環境中方位之需求，本論文利用模式比對 技術進行自動車在室內走廊的導航。室內走廊的模式事先儲存在自動車□，在自動車 行進當中，首先利用攝影機攝取自動車前方的牆壁踢腳，經處理後得到輸入圖形。將 輸入圖形與走廊模式作比對，即可得出自動車的方位。因二度空間線段表示的輸入圖 形與走廊模式可以編碼成一度空間的字串，故他們的最佳比對即為其對應之字串的最 長共同字串，如此可以得到很好的比對結果且毋須作複雜的比較運算。由許多成功的 自動航行實驗實此方法可行。 自動車在航行中可能會遇到靜止或行進的障礙物，為此本論文利用即時最小平方誤差 估計法來預測障礙物的軌跡，並利用最小平方誤差圖形分類的觀念，決定一條避開障 礙物的路徑，另外自動車速度可以航海學上的技術作機動性的改變。模擬的實驗結果 證明了此方法的可行。