摘 要 日益繁忙的社會，交通運輸事業越顯複雜，人們對於交通安全及便利的需求也越來越重視。智慧型運輸系統(ITS, Intelligent Transportation Systems)便是各國積極發展的專案。最近幾年，研究發展更是迅速；其中，與人的生命安全息息相關的安全駕駛，特別受到重視。本論文即是發展用於輔助安全駕駛上的即時模式化道路線及前車偵測。 本論文的主要研究在於如何快速並有效地偵測出道路影像中的道路線、前車所在的位置、及判斷車輛是否偏離車道。針對道路線偵測，我們利用人類視覺的特性，加強所搜尋的資訊，並提出了一個減少搜尋空間的方法，不但更正確也更快速的偵測出道路線，我們也提出多車道線偵測的方法。針對前車偵測，我們提出一個適應性的門檻值，來偵測前車位置。最後我們也利用透視幾何模型，求得目前車輛相對於車道中的橫向位置，判別目前是否偏離車道，以警示駕駛人偏離車道，避免發生危險。我們的方法能夠有效的克服天候變化及其他車輛對影像所造成的影響。 在實驗方面，我們在Win2000平台、P4 1.8GHz CPU、540MB RAM、影像解析度為320×240的環境下，測試6千多張影像，影像包含多種不同天氣及不同環境。在大部份情況下均能正確並即時的偵測出道路線、前車、及車輛有無偏離車道；執行速度高達每秒30張影像，平均一張影像只需花0.033秒，單張影像處理正確率超過98%。 Abstract People pay attention on the safe driving more and more. The research on intelligent transportation systems (ITS) is quickly developed in recent years. The safe driving is one of the important subjects in the ITS. In this thesis, we propose a real-time model-based method for lane and vehicle detection for safe driving system. Our goal is to detect lane markings and front vehicle, and then provide lane departure warning based on the road images efficiently and effectively. In the lane detection, we exploit the property of human vision to enhance the difference map’s information such that the result of the lane detection is more effectively, and then propose a method for reduction of searching space in order to improve the detection efficiency. Moreover, we propose a multi-lane detection method. In the front vehicle detection, we exploit lane’s location as a searching region and define two adaptive threshold values to detect the front vehicle. Finally, we also exploit lane’s location and camera optical direction to estimate lateral offset of the vehicle with respect to the detected lane markers. Then the lane departure alarm is triggered by the decision of the estimation algorithm. In experiments, six-thousand images were processed to evaluate the system performance. The images were captured in variant weather conditions and with various driving situations. The rate of lane detection is over 98% and the processing time is about 0.033 seconds on average.
