碩士 國立臺灣大學 電機工程學研究所 105 With the development of autonomous vehicle, obtaining the global position of autonomous vehicle becomes more and more significant. While GPS is widely used in global positioning, it suffers from a systematic noise which can be modeled as a bias. To obtain a more accurate localization result, a global map with recorded landmarks can be used. By calculating relative position between ego-vehicle and landmarks, GPS bias can be effectively corrected. Road markers are widely seen on structured road. Bright color and consistent shape make them suitable to be landmarks on global map. Hence, this thesis proposes a global positioning algorithm based on road marker matching and GPS bias correction. To achieve the goal, road marker detection and localization algorithm are proposed. For road maker detection, not only the road marker should be recognized, but the position and orientation should be detected. Two road marker detection methods are proposed. One is based on machine learning algorithm followed by correlation to find the orientation and position. The other one is based on template matching which searches road marker pattern on road surface. For localization, road marker matching is proposed. Different kinds of road markers are treated in different way based on their geometrical property. Moreover, the localization result is used to enhance road marker detection algorithm. By using the vehicle position and orientation, correctness and efficiency are both increased. The calculation time is reduced by 98.5% compared with an exhaustive searching method. Finally, the position information from road marker matching is combined with GPS and ORB-SLAM, respectively. Kalman filter or moving average are used to combined GPS with road marker matching. For ORB-SLAM, scale and global position can be recovered by road marker matching. Experiments on road marker detection and localization are performed with 6 videos in several conditions such as occlusion on road markers or illumination variance. It is shown that trajectories from proposed algorithm is much more reasonable than the original GPS trajectories. Furthermore, ORB-SLAM trajectory can be recovered from road marker matching in some situations.