Surface-based GICP


In this paper we present an extension of the Generalized ICP algorithm for the registration of point clouds for use in lidar-based SLAM applications. As opposed to the plane-to-plane cost function, which assumes that each point set is locally planar, we propose to incorporate additional information on the underlying surface into the GICP process. Doing so, we are able to deal better with the artefacts that are typically present in lidar point clouds, including an inhomogeneous and sparse point density, noise and missing data. Experiments on lidar sequences of the KITTI benchmark demonstrate that we are able to substantially reduce the positional error compared to the original GICP algorithm.

Proceedings of the 15th Conference on Computer and Robot Vision (CRV)