基於類神經網路之即時車道障礙物偵測

Abstract

近年來世界各國投入了許多人力發展智慧型運輸系統，而障礙物偵測為智慧型運輸系統計畫內極為重要的一環。本篇論文利用架設於車上之單一影像擷取設備來取得車輛前方影像，接著利用影像邊緣資訊、透視轉換及類神經網路來完成障礙物偵測的目的。 我們利用邊緣資訊及顏色去找出影像中可能的車道線位置，將這些資訊投影到世界座標中去分類找出車道線，利用車道線的位置把搜尋障礙物的區域集中在左邊、右邊和前方三個車道上，並減少道路上的標誌對系統所造成的影響，不僅可以降低障礙物重疊時對偵測所造成的困擾，更可以增加系統的穩健性。之後在利用邊緣偵測、陰影特徵以及類神經網路來偵測出車道內的障礙物，找到了障礙物後，再利用線性的方式來進行追蹤，以減少偵測所需要的運算量。

Intelligent transportation system has been one of the important issues over the world for decades and obstacles detection plays an important role in it. This thesis implements the task of obstacle detection by the forward images captured by a single CCD mounted on the vehicle by using edge detection, perspective transformation and neural network. We find the probabilistic position of lanes in images by using edge detection and perspective transformation. We project this information to world coordinate system to find out the lanes. After that, we take the detected lane region as our searching region and reduce the effects of mark on the road. We not only decrease the effects of overlapping obstacles for the system but also increase the robustness of system by lane detection. We then detect the obstacles on the lane by using edge detection, shadow feature and neural network. After that, we track obstacle by using linear approach to reduce complex operation.