A Novel Global Localization Method Using 3D Laser Range Data in Large-Scale and Sparse Environments

In large-scale and sparse environments, such as farmlands, orchards, mines and electrical substations, global localization based on particle filter framework without any prior knowledge still remains a challenging problem. Some issues such as speeding up the convergence of particles and improving the convergence accuracy in similar scenes need to be addressed. This paper proposes a novel global localization method, which treats the global localization problem as place recognition and pose estimation problem. Specifically, we firstly utilize the random forests algorithm to train a classifier to predict whether two 3D LiDAR observations are from the same place. Then, the classifier is used to match the current observation with the prior map to estimate the possible initial pose of the robot. Finally, a multiple hypotheses particle filter algorithm is proposed to achieve final localization. Experimental scenes are selected in the indoor parking lot with high dynamic characteristics and two electrical substations with the characteristics of sparse and large-scale. The experimental results show that the proposed algorithm has faster convergence speed and higher accuracy.