A New Approach of Initial Localization for Unmanned Vehicles Based on 3D Descriptor and Normal Distributions Transform

Abstract

With the development of technology, autonomous driving technique has attracted more and more attention. Localization is an important module in the navigation application of automatic driving system. At present, the automatic driving system is still based on high-precison map, so it is very important to identify the initial position of the vehicle on the map. In the outdoor environment, GNSS is the mature scheme. However, in the environment where high buildings exist, GNSS signals are usually interfered strongly by electromagnetic signals from these buildings, which is a great challenge to the automatic driving system. In this paper, a two-step localizaton algorithm is proposed based on the prior point cloud map. Firstly, the intensity scan context algorithm is used to roughly localize the initial pose of the vehicle. After that, the result of rough localization is used as the initial value of the normal distributions transform (NDT) algorithm. After NDT registration, the precise pose of the vehicle can be obtained. At the same time, an optimal scoring strategy are proposed to improve the robustness of the localization system. Finally, the algorithm is tested on the campus of Wuhan University of Technology.