Robust Place Recognition and Loop Closing in Laser-Based SLAM for UGVs in Urban Environments

Abstract

Robust place recognition plays a key role for the long-term autonomy of unmanned ground vehicles (UGVs) working in indoor or outdoor environments. Although most of the state-of-the-art that approaches for place recognition are vision-based, visual sensors lack adaptability in environments with poor or dynamically changing illumination. In this paper, a 3-D-laser-based place recognition algorithm is proposed to accomplish loop closure detection for simultaneous localization and mapping. An image model named bearing angle (BA) is adopted to convert 3-D laser points to 2-D images, and then ORB features extracted from BA images are utilized to perform scene matching. Since the computational cost for matching a query BA image with all the BA images in a database is too high to meet the requirement of performing real-time place recognition, a visual bag of words approach is used to improve search efficiency. Furthermore, a speed normalization algorithm and a 3-D geometry-based verification algorithm are proposed to complete the proposed place recognition algorithm. Experiments were conducted on two self-developed UGV platforms to verify the performance of the proposed method.