有全方位監控之多輛電腦視覺自動車的最佳安全巡邏之研究

Abstract

本論文提出利用多輛視覺型自動車執行安全巡邏工作的系統。我們假設巡邏環境的地板形狀是由矩形區塊所組成，並採用兩輛裝設有攝影機的自動車，以及兩部固定於天花板的魚眼鏡頭攝影機做為實驗平台。我們提出了一項獲取未知環境資訊的方法，可得到組成環境的所有矩形區域、導航自動車的轉折點、每組監控點間的距離與巡邏路徑。利用這些資訊可使自動車在航行時，不會與牆壁產生擦撞。另外，針對相機校正的問題，我們亦提出一項結合影像內插的點對應方法，利用多組二維影像與三維真實空間的對應點與影像內插法，去求得點在扭曲影像中的實際正確位置。此外，為了提高相機校正的精準度，我們也提出一項快速求取多組對應點的方法。藉由這些方法，我們可利用固定於天花板上的俯視攝影機來學習自動車相對於監控物品的位置與朝向，並利用此資訊來執行安全巡邏的工作。再者，我們也用俯視攝影機來定位與監控自動車的移動行為。而為了使得本系統有較好效益，我們則提出一項具有優化隨機與工作量平衡特性的路徑規劃方法，來分配巡邏路徑給各自動車，縮短完成每輪巡邏工作所花費的時間，進而提高環境中的安全程度。最後，由於本系統中所採用的自動車數量不只一台，所以我們提出了即時避碰的技巧，根據可能產生碰撞的路線狀況，自動產生可行駛的避碰路線。實驗結果證明我們所提出的這些方法是可行而且有效的。

A multiple vision-based vehicle system for security patrolling in an indoor environment, whose floor shape is composed of rectangular regions, is proposed. Two autonomous vehicles controllable by wireless communication and equipped with cameras, as well as two cameras with fish-eye lenses fixed on the ceiling, are used as a test bed. To acquire information of an unknown environment, an environment-information calculation method is proposed for obtaining all rectangular regions composing the floor shape of the environment, the turning points for navigation, all distances between monitored objects, and the patrolling paths. These data enable the vehicles to navigate without collisions with walls. Also, a point-correspondence technique integrated with an image interpolation method is proposed for camera calibration. By a technique of finding corresponding points in 2-D image and 3-D global spaces as well as an image interpolation method, the correct positions of interesting feature points can be obtained from the warped images captured by the cameras with fish-eye lenses. Besides, a faster point-correspondence technique is proposed to obtain abundant corresponding points that yield better calibration accuracy. With this camera calibration technique, the cameras on the ceiling can be utilized to learn the poses of the vehicles with respect to monitored objects. Also, the vehicles are taught where and in which direction to perform the security monitoring task, in which the position information is used to guide the vehicles. Additionally, the top-view cameras can also be utilized to locate the vehicles and monitor vehicle activities in the navigation phase. An optimal randomized and load-balanced path planning method is proposed as well, which requires shorter time to accomplish object monitoring in one session and provides higher degrees of patrolling security. Because the number of the vehicles used in this study is more than one, a real-time collision avoidance technique is also proposed. According to the state of path-intersecting, feasible alternative paths for the vehicles can be obtained. Good experimental results show the flexibility and feasibility of the proposed methods for the application of multiple-vehicle security patrolling.