LiDAR-Based Localization System for Kidnapped Robots

Place recognition is the ability to recognize previously seen places in the world. When the environment is known, the robot may localise itself when it starts and whenever it needs to verify where it is. We propose a solution to implement place recognition capability using a geometric representation of the environment build with a LiDAR. Nowadays, large Field Of View LiDARs allows to get a dense and precise representation of the environment. Assuming that a 3D map of the environment is available, we propose to build a compressed codebook based on FPFH descriptors. The code book is then used to find the robot localization with one LiDAR image. The system has been tested on real and simulated measurements and shows promising results. The pose estimation can be thoroughly enhanced with point cloud registration methods.