以環場視覺及擴增實境技術作室內停車之自動導引

Abstract

當一駕駛開車進入一個大型停車場時，要找到一個停車格常常不是件容易的事，有時候甚至要排隊一段時間才能等到空位，或須在停車場來回來尋找才能解決問題。 對此本研究提出一個擴增實境的系統來協助駕駛找到空車位，本系統能自動偵測空車位位置、規劃出導航路徑，並以擴增實境方式在行動裝置上指引駕駛前往空位處。本系統提供了四大功能來達到導航目的，分別是：車輛追踨、空車位偵測、路徑規劃以及擴增實境式導航。 為了偵測並追踨在停車場中的車輛，本研究提出三項方法，第一項使用一三維盒子形狀來找出車輛所在位置；第二項使用銀杏葉形狀的預測區域來改善車輛偵測效果；第三項是設定電子圍籬來決定追踨起點，並考慮車輛行進方向的連續性。最後將行進的軌跡標示在停車場地圖上，駕駛觀看此圖即可知道他目前在停車場內的位置。 為了要偵測停車場中的空車位，本研究提出的系統採用一套動態背景學習的方法，並再次使用三維盒子形狀來判斷一個停車格是否被車輛停放。接著，駕駛可以從系統找到的諸多空車位中選擇自己想要停的車格。在選擇停車格後，系統會應用戴克斯特拉（Dijkstra）演算法規劃出一條最短路徑，標示在停車場地圖上，指引駕駛從目前所在位置開到所選定的停車位。 為了要以擴增實境的方式來指引駕駛，本研究提出一把魚眼影像轉換成環場影像及駕駛視角影像的方法，駕駛只要從行動裝置上觀看標有箭號的擴增實境影像，並沿著箭號前進，就可到達事先所選定的停車位。 最後，本研究以實驗證明其結果良好，顯示出本研究所提系統確實可行。

When driving a car into a large parking lot, a driver always needs to find a parking space and sometimes has to stand in line for a long time before a parking space becomes available. After entering the parking lot, it might even be necessary to drive around the entire parking lot more than once before an empty space can be found out. To solve such a parking-space finding problem, an augmented reality (AR)-based guidance system is proposed in this study to help a driver to save parking time, which has the functions of both finding empty parking spaces automatically and displaying a planned path to guide the driver to an empty space on a user’s mobile-device screen in an AR way. The proposed system includes four major components, including car detection and tracking, parking-space detection, path planning, and AR-based guidance for car parking. In order to detect and track a car driven in the parking lot, a car localization method by using 3D bounding boxes is proposed to find the current position of the car. Furthermore, a car detection method by using a gingko-shaped prediction area is proposed to refine the car localization result. Finally, the concept of virtual fence is used to find a start point for the car detection process. The idea of tracking continuity is also applied to get the best tracking accuracy. And the trajectory of the car is drawn onto the environment map. In such a way, the driver is able to know where he/she is in the parking lot by looking at the map. In order to find empty parking spaces in the parking lot, a dynamic environment learning technique and a parking-space detection method based on the use of 3D bounding boxes are proposed, by which the system can find out the positions of all empty parking spaces among which the driver can select one as the destination for car parking. Also proposed is a new method for path planning which is based on the Dijsktra algorithm and yields a shortest path from the current location of the car to the selected empty parking space. In order to guide the driver, an integral method for generating panorama images and perspective-view images from fisheye images is proposed, by which the driver can get an AR-based guidance image with the planned navigation path drawn on it. By following the augmented navigation path in the image which is displayed on the mobile-device screen, the driver can finally get to the empty parking-space that he/she chooses. Good experimental results showing the feasibility of the proposed system and methods are also included.