Synthetic 2D LIDAR for Precise Vehicle Localization in 3D Urban Environment Z. J. Chong 1 , B. Qin 1 , T. Bandyopadhyay 2 , M. H. Ang Jr. 1 , E. Frazzoli 3 , D. Rus 3 Abstract— This paper presents a precise localization algo- rithm for vehicles in 3D urban environment with only one 2D LIDAR and odometry information. A novel idea of synthetic 2D LIDAR is proposed to solve the localization problem on a virtual 2D plane. A Monte Carlo Localization scheme is adopted for vehicle position estimation, based on synthetic LIDAR measurements and odometry information. The accuracy and robustness of the proposed algorithm are demonstrated by performing real time localization in a 1.5 km driving test around the NUS campus area. I. I NTRODUCTION Mobile robot localization is the problem of determining the pose of a robot relative to a given map of the environment [28]. It is a fundamental requirement to realize vehicle auton- omy. This problem has been well solved for indoor robots on a planar surface. However, there are still many challenges to get an accurate, robust while low-cost approach for vehicle localization in an outdoor 3D urban environment. Together with localization comes its concomitant problem of mapping. A map is an abstract representation of the environment, which usually serves as a prior for robot localization. How to map the 3D environment and what representation to use are additional questions that must be addressed. This paper introduces a novel idea of synthetic LIDAR, which constructs synthetic 2D scan from 3D features, and solves the localiza- tion and mapping problem in a 2D manner. The algorithm is developed under the idea of minimal sensing, using only one tilted-down single-layer LIDAR, and odometry information. The fusion of Global Positioning System (GPS) and Iner- tial Navigation System (INS) to estimate vehicle position has been the most popular localization solution in recent years [18], [21], [4]. This solution works well in open areas; however, it is not suitable for dense urban environment where GPS signals severely suffer from satellite blockage and multipath propagation caused by high buildings [14]. Road-matching algorithms are then proposed to alleviate this problem, where a prior road map is used as either additional motion constraint or observation to update the localization estimation[6], [7]. While this solution achieves good global localization, it is not designed for precise estimation relative to the local environment, an ability that is highly desirable in many cases. 1 Z. J. Chong, B. Qin, M. H. Ang Jr. are with the National University of Singapore, Kent Ridge, Singapore {chongzj, baoxing.qin, mpeangh} at nus.edu.sg 2 T. Bandyopadhyay is with the Singapore-MIT Alliance for Research and Technology, Singapore tirtha at smart.mit.edu 3 E. Frazzoli and D. Rus are with the Massachusetts Institute of Technol- ogy, Cambridge, MA., USA frazzoli at mit.edu, rus at csail.mit.edu Map-aided algorithms are proposed for high precision localization using local features. In [9], single side curb features are extracted by a vertical LIDAR to build a boundary map to improve vehicle localization. This map is learned beforehand in the form of line segments. In [26], lane markers serve as local features, which are extracted from reflectivity values of LIDAR scans. A digital lane marker map is used as prior. The performance of the algorithm is similar to those in [9]. While these algorithms reduce lateral localization error considerably, they help little in the longitudinal direction. Levinson et al. in [16], [17] utilize road surface reflectivity for precise localization. A particle filter is used to localize the vehicle in real time with a 3D Velodyne LIDAR. The algorithm first analyses the laser range data, and extract those points cast on the ground. Then reflectivity measurements of these points are correlated to a map of ground reflectivity to update particle weights. One assumption underlying this algorithm is that road surfaces remain relatively constant, which may undermine the performance in some cases. Be- sides, the need for costly 3D LIDAR sensor limits its usage. Baldwin et al. in [2] utilizes accumulated laser sweeps as local features. The algorithm first generates a swath of laser data by accumulating 2D laser scans from a tilted- down LIDAR. Then the swathe is matched to a prior 3D survey by minimizing an objective function. This algorithm demonstrates its accuracy and robustness in GPS-denied areas. Although the algorithm proposed does not require an accurate 3D model of the environment, we argue that an accurate and consistent prior is always desired when the localization is integrated with other navigation functions. Similarly in [31], [15], a 3D point cloud of the environment is obtained by servoing a 2D LIDAR, and a reduced 2D feature is used to perform localization. This algorithm has been shown to work well in an indoor environment with a well structured ceiling features. In [5], a microwave radar sensor is used to perform SLAM. While the radar has the ability to “see through” obstacles, association of radar targets is a complex task and SLAM processing is done offline. This paper proposes a novel notion of synthetic LIDAR to solve the 3D localization problem in a 2D manner. The synthetic LIDAR is constructed in real time with interest points extracted from a 3D rolling window. The basic as- sumption is that many surfaces in the urban environment are rectilinear in the vertical direction. The interest points are extracted from the rectilinear surface, and then projected on a virtual horizontal plane to form a synthetic LIDAR. The synthetic LIDAR serves as a bridge between the real- 2013 IEEE International Conference on Robotics and Automation (ICRA) Karlsruhe, Germany, May 6-10, 2013 978-1-4673-5643-5/13/$31.00 ©2013 IEEE 1554