俱學習機制之車道線偵測與追踨方法

Abstract

近年來交通事故頻繁，大部分車禍發生的原因主要由於駕駛者分心、疲勞駕駛或不注意車況等不當駕駛行為所引起。因此，如何利用現代科技讓駕駛者能更安全地駕駛車輛就變得越來越受重視，其中最普遍的一種方法就是車道偏移系統（LDWS），靠著電腦快速運算，讓車輛可以認知本身所在的車道位置，並且分析判斷當下是否安全，若是已偏移車道時，可以發出警告提醒駕駛者車況，以避免車禍發生。而車道偏移系統最重要的部份就是車道偵測，本論文主要研究利用前置、單鏡頭行車紀錄器取得之車前路面影像，透過深度學習之分析方法來偵測車道資訊。 在我們提出的系統中，首先針對每一張影像進行影像前處理，包含影像縮小、灰階化、銳化與亮度平均化，再來使用邊緣偵測方法截取邊緣特徵值，以深度學習方法採用Haar特徵階層式分類器來偵測車道位置；最後採用移動向量的預測與紀錄方法來做簡單車道追踨，以簡化需要分析的影像範圍和資料量，達到加快分析速度並且改進分析結果之精確度。 實驗結果顯示，我們提出的深度學習車道偵測方法，特別在一些場景如夜晚、車道有物體陰影，或是複雜的車道標線下，皆能準確的偵測車道。

In recent years, traffic accidents are happening frequently. Most accidents occurred are due to improper driving behavior such as driver’s distraction, drowsy driving, or not focusing on traffic. Therefore, how to use modern technology to make drivers driving more safely becomes more and more important. The most common technology is Lane Departure Warning System (LDWS). By fast computer processing, the vehicle is able to recognize where the lane line position and knows whether it is safe now. If lane departure happens, the system can call driver’s attention back to traffic with warnings to prevent car accident. Lane detection is the most important part of LDWS. This thesis tries to detect lane lines through a learning analysis with captured images from a single-lens camera facing toward vehicle front. In our proposed system, it starts with image pre-processing such as images resizing, gray scaling, sharpness and equalization for every frame. Using edge detection to extract edge features and Haar-like feature cascade classifier with Adaboost to detect lane lines. Finally, to speed up analysis and improve the accuracy of analysis results by reducing the range and data amount using simple lane tracking with record and motion vector prediction. Experimental results show that our proposed learning-based lane detection method is able to detect the lane lines accurately, especially in some scenes, such as nighttime, roads with shadow of objects, or complex lane markings.