基於電腦視覺之即時穩健的泛型障礙物與車道偵測行車系統

Abstract

近年來隨著交通問題日益嚴重，智慧型運輸系統(Intelligent transportation system, ITS)的相關研究愈來愈受到重視，其中智慧型車輛又是最有發展潛力的研究之一。而泛型障礙物與車道偵測系統是智慧車所需配備的最基本功能，能夠偵測出路面障礙物的位置與車道資訊，用以預警駕駛人注意或者提供車輛自動行駛所必需的道路資訊。 本文主要是利用影像處理與電腦視覺的技術去偵測路上障礙物與車道的位置。將兩支單色CCD攝影機分別上下地架設在車上，利用histogram-based的方法將上方攝影機所擷取出的道路影像做分類，以偵測出不同類別的交界所組成之近乎水平的邊線。所偵測出的邊線可能位於地面或者障礙物上，這兩種情況判斷的依據是藉由立體視覺的技術分別預估此邊線在下方影像中可能是地面的位置以及可能是障礙物的位置，然後量測與上方影像中之邊線的相關係數何者比較大來做判斷，因此可以鑑別出影像中障礙物與路面的部分。 而在車道偵測方面，使用一支單色CCD攝影機擷取道路影像，以偵測車道標線的位置。本文所發展出的車道偵測演算法是基於車道幾何模型的標線偵測方式，能夠提供一個穩健的偵測結果，並且適當地預估與縮小搜尋範圍，以降低搜尋時間而提高車道偵測的效率。最後並重建車道3-D幾何模型以修正道路傾斜度與寬度，因此本文所提出的演算法亦適用於非平坦的路面。 本文所發展出的車道偵測系統已經在快速道路與高速公路成功地實車驗證過，在2.6 GHz的PC平台上平均每張影像所需的偵測時間小於1 ms。此外，車道偵測系統亦結合方向盤控制器，做為無人駕駛智慧車的視覺系統，完成台灣第一台可以hand-free自動駕駛的智慧車TAIWAN iTS-1。TAIWAN iTS-1以時速90 km/hr與110 km/hr分別在東西向快速道路與國道3號高速公路順利地自動駕駛實車測試，並經過國外卓越計畫評審委員的評鑑與肯定，驗證了本文所提出的車道偵測系統的實用性與穩定性。

As the traffic is becoming more and more serious in most developed countries, a lot of researches about the intelligent transportation system (ITS) have been paid attention in recent years. Above all, one of the most promoting topics for the ITS applications is concerning the smart vehicles. The fundamental function of the smart vehicle is the generic obstacle and lane detection system, which can warn the driver or provide the road information for the unmanned vehicle. In this thesis the techniques of image processing and computer vision are applied to the detection system. Two monochromatic CCD cameras are mounted top and bottom on the vehicle, and the road image captured by the top camera is segmented by thresholding the histogram. After that, the quasi-horizontal boundaries formed by the interconnection of two different segments are detected in order, and each detected boundary could belong to either the ground or the obstacle. The criterion to distinguish between them is to predict the corresponding ground and obstacle boundaries in the bottom image by the stereo vision, and to compute the normalized correlation coefficients of the detected boundary in the top image with respect to the ground and obstacle boundaries in the bottom image respectively. The detected boundary in the top image belongs to the obstacle if the normalized correlation coefficient associated with the obstacle is larger than that associated with the ground. Thus the road image can be divided into the ground and obstacle parts. On the other hand, a single monochromatic CCD camera is used in the lane detection system to detect the lane markings. Based on the geometric lane model, the algorithm of lane detection proposed in this thesis can generate a robust result. Besides, the detection region of interest can be estimated to narrow the searching area and to reduce the computational load. Eventually, the 3-D lane geometry is reconstructed to update the road inclination and lane width. Therefore the proposed algorithm is available in the case of non-flat roads. The lane detection system proposed in this thesis has been successfully verified on the expressway and freeway. On the PC platform of 2.6-GHz CPU and 512-MB RAM, the average time of lane detection is less than 1 ms per frame. In addition, the lane detection system can be treated as the vision system of the automatic vehicle by integrating the controller of the steering wheel. This work has been implemented on the experimental car, TAIWAN iTS-1, running on the expressway and freeway with velocities of 90 km/hr and 110 km/hr respectively. TAIWAN iTS-1 is the first smart car in Taiwan capable of hand-free driving on the real road, which verifies the practicability and robustness of the proposed lane detection system.