Robust Real-Time Registration of RGB-D Images using Multi-Resolution Surfel Representations

Fast and robust registration of 3D scans is required in many approaches to perception in robotics such as pose tracking or simultaneous localization and mapping. We propose a novel efficient method to register RGB-D images. We convert the image content into a multi-resolution surfel representation and exploit the dense image neighborhood to construct such views at high frame-rates on a single CPU. Our approach registers views using an efficient and robust variant of the Iterative Closest Points algorithm. We evaluate our method on a recently published benchmark dataset and achieve results beyond the state-of-the-art. We also report on the successful public demonstration of our method at RoboCup 2011.