Using RGB Information to Improve NDT Distribution Generation and Registration Convergence

Abstract

Unmanned vehicles are becoming an inevitability in our society and with them comes the need for highly robust and accurate algorithms to perform their critical functions, such as localization and mapping. The proliferation of these robots into wide spread use requires a generalized, robust SLAM solution. This paper proposes an improved NDT algorithm, which is capable of performing robust, accurate localization and mapping in an broad spectrum of possible environments and with a multitude of different sensors. The method uses a color greedy cluster approach to cluster points and generate Gaussian distributions and then uses an exhaustive color weighted distribution to distribution cost function to optimize the scan alignment. With the addition of these key features to the NDT framework the method is capable of providing accurate results with minimal computation time. Evaluation is performed on both the Freiburg and Ford datasets to demonstrate a multitude of environments and shows robust registration throughout a wide range of environments and viewpoints.