Using Intensity and Geometric Models of 3-D Point Cloud for Landmark Detection and Vehicle Localization

• Main Authors: Ti Lan, 藍迪
• Other Authors: 連豊力
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/857x48

碩士 === 國立臺灣大學 === 電機工程學研究所 === 106 === In order to achieve full self-driving capability, localization is one of the basic function of future autonomous vehicle. Although GPS is widely used for localization, it suffers from bias generally. To reduce the bias, landmarks around the ego-vehicle can be used to enhance the localization performance. If the global position of landmarks (or the landmark map) are known, the vehicle pose can be estimated by finding the transformation between the detected landmarks and the landmarks in global coordinate. For landmark detection, intensity value and geometric feature are extracted from the 3-D point cloud captured by LiDAR. Based on the known model of landmark, a model-driven approach is used to estimate the pose of landmark in local coordinate. To reduce the model matching error, an optimization is performed after initial landmark pose estimation.For vehicle localization, the data association between the detected landmarks and the map are estimated based on the prediction vehicle pose. If two or more landmarks are available, vehicle pose can be estimated from the detected landmarks. In addition, to increase the smoothness of localization trajectory, Kalman filtering is used from both time update and measurement update. The experimental results show that the average localization bias of the proposed method with available ground truth could be reduced to 0.19m, which is lower than the bias of using GPS only (1.81m).