基於重疊時間切片影像之車道線追蹤與分類

Abstract

近年來行車紀錄器的使用越來越廣泛，而車輛輔助安全駕駛的系統也跟著逐漸發展起來。其中，車道線偵測在車輛輔助安全系統中是一項必需的元件。而在測試車道線偵測的系統好壞的真實數據(ground truth)，在標示的過程中經常需要花大量人力與時間。在這篇論文中，我們提出一個利用重疊的時間切片影像(time-sliced images)進行偵測、追蹤車道線與車道線種類分類來自動產生真實數據的系統。另一方面，同時會產生一個高效能的即時車道線偵測系統。首先，在前置處理中從影像裡取得數張灰階的切片影像。在找尋候選點步驟中，對切片影像做頂-帽轉換、平滑影像、尋找峰值，找出每張切片影像中的車道線候選點，並記錄在切片影像中。接著在串連與刪除候選點步驟中，將切片影像加入相對應的時間切片影像中，在時間切片影像裡候選點會與前段時間切片的候選點連結，並刪除可能是雜訊的候選點，最後得出連結好的候選點。接著將各時間切片影像中連結好的候選點以彼此等距的方式將點互相配對找出車道線，當前張影像中已經找到車道線時，將會利用時間切片影像上預測這些車道線在新畫面的可能位置，以進行進一步的追蹤。最後，在找到車道線後系統會將車道線做種類分類。 實驗中採用行車紀錄器所拍下的影片當作測試資料，對於四種車道線路況、三種行車狀況與三種天氣狀況做車道線偵測。四種車道線路況如下：一是車道線為兩邊都是虛線時，二是車道線為兩邊都是實線時，三是車道線為一邊是虛線一邊是實線時，四是車道線其中一邊是雙線時。三種行車狀況為，直線、彎道和換車道。三種天氣狀況為，晴天、陰天和雨天。我們對執行出來的結果及產生的問題進行討論。實驗結果顯示，我們提出的真實數據產生方法能將絕大部分的將車道線找出，並區分其車道線種類。除此之外，我們提出的即時車道線偵測系統也能準確的找出車道線。

In recent years, with the widely spread of the car video recorder, more and more Driver Assistance Systems for safety are developed. Lane detection is an essential component of Driver Assistance System. In this thesis, we propose a lane tracking and lane mark classification system utilizing the overlaying time-sliced images for ground truth generation and real-time driver assistance application. First, we generate time-sliced images from a video by accumulating the sliced data of the selection rows in a frame. From the time-sliced images processed by top-hat transform and smoothing, we obtain the lane line candidate points by finding the peaks in each row. Then we remove the noises and connect the candidate points in the time-sliced images, and the lane mark type of each candidate point is determined by the characteristic of the point. Through overlaying the time-sliced images, the lane can be detected by tracking the connected candidate points in different time-sliced images. For lane mark classification, we collect the information of all the candidate points in the lane and apply the major voting algorithm. The experiments are conducted with the video sequences of four lane line conditions, three road scenarios, and three kind of weather. The lane line conditions are categorized into (1) two dash lines, (2) two solid lines, (3) one solid line and one dash line, and (4) one double line and another line. The three road scenarios are (1) straight lanes, (2) curve lanes, and (3) lane changes. The three kind of weather are (1) sunny, (2) cloudy, and (3) rainy. The experimental results show that our proposed ground truth generation method can detect over 99% of the lanes in all the video sequences. In addition, the proposed real-time lane detection system can also achieve high accuracy for lane tracking and lane mark classification.