In this study, a 3D Normal Distributions Transform (NDT)-based pose correction framework for a mobile robot is proposed to prove its robustness under a considerable odometry uncertainty. Our proposed method consists of Extended Kalan Filter based on IMU and wheel odometry to estimate initial pose and NDT that corrects the initial guess via 3D scans. Then, the final pose correction is retrieved with Kalman gain using corrected poses from both EKF and NDT, followed by point cloud accumulation that generates the global map. Before examining our proposed method, quantitative comparisons of NDT over other 3D scan matching algorithms targeted for mobile robot pose correction were conducted. After that, the experiments are conducted for both indoor environments and outdoor environments. As verified experimentally, NDT is a) suitable for real-time pose correction over other 3D-scan matching algorithms b) robust even when loop closing is not available and c) shows better performance than EKF on pose correction under very imprecise raw odometry. Additionally, we analyzed the time-wise performance of NDT during registration over different voxel sizes
of a global map.