Machine Learning for LiDAR-Based Navigation System

This paper presents a robust 6-DOF relative navigation by combining the iterative closet point (ICP) registration algorithm and a noise-adaptive Kalman filter (AKF) in a closed-loop configuration together with measurements from a laser scanner and an inertial measurement unit (IMU). In this approach, the fine-alignment phase of the registration is integrated with the filter innovation step for estimation correction while the filter estimate propagation provides the coarse alignment needed to find the corresponding points at the beginning of ICP iteration cycle. The convergence of the ICP point matching is monitored by a fault-detection logic and the covariance associated with the ICP alignment error is estimated by a recursive algorithm. This ICP enhancement has proven to improve robustness and accuracy of the pose tracking performance and to automatically recover correct alignment whenever the tracking is lost. The Kalman filter estimator is designed so as to identify the required parameters such as IMU biases and location of the spacecraft center-of-mass (CoM).