Abstract

Affordable light detection and ranging sensors are becoming available for tasks such as simultaneous localization and mapping (SLAM) in robotics and autonomous driving; however, these sensors offer less quality data of lower resolution that hinders the performance of registration methods. The deep learning based approaches seem to be sensitive to these data flaws. Specifically, a state-of-the-art deep learning-based approach failed to produce meaningful results after several attempts to carry out transfer learning over a dataset collected indoors with one such affordable sensors. The paper introduces a hybrid approach combining two well-established registration techniques, the iterative closest point algorithm and the normal distributions transform that achieves good performance on the SLAM task over the same dataset.