Feature-based global localization system for mobile robot using prior data and monocular camera

Abstract

As robotic technologies have progressively developed, a localization problem of robot has become increasingly important. Although GPS(Global Positioning System) is widely used for the localization problem, there are several drawbacks such as the shadow areas of GPS signal or indoor environment. Recently, localization methods without GPS have been actively studied to overcome these drawbacks. This paper proposes a novel global localization method that utilizes a monocular camera without using GPS. The most important part of a localization system without GPS is to match landmarks between a current data and prior data which contain global information. This paper estimates a relative position from a global position of prior data by image matching for the global localization problem. The prior data consists of Velodyne, GPS/IMU, and images. Although the proposed algorithm was only tested in specific area by using KITTI data set as the prior data, we predict the algorithm is more useful by offering vast prior data from Google’s unmanned ground vehicle. The proposed algorithm is able to be divided into global localization and MCL(Monte Carlo Localization) algorithm. The global localization algorithm is performed by matching feature descriptors of the prior images and the current image. The 2D coordinate features of prior images are translated to 3D coordinates from the point cloud of Velodyne data. The localization result is estimated by a camera extrinsic parameter using the matched 3D-2D feature pairs. The MCL algorithm produces robust results, even in difficult environments such as regions with many dynamic objects or few feature descriptors, and also improves the computation time for realtime applications. The proposed system is tested through 36.6km of distance in 8 data sets of KITTI data. The results demonstrate the effectiveness of the proposed system, as shown by the higher accuracy of the global localization algorithm and the robustness of the local localiz...