NICP: Dense normal based point cloud registration

In this paper we present a novel on-line method to recursively align point clouds.By considering each point together with the local features of the surface (normal and curvature), our method takes advantage of the 3D structure around the points for the determination of the data association between two clouds.The algorithm relies on a least squares formulation of the alignment problem, that minimizes an error metric depending on these surface characteristics.We named the approach Normal Iterative Closest Point (NICP in short).Extensive experiments on publicly available benchmark data show that NICP outperforms other state-of-the-art approaches. I. INTRODUCTIONRegistering two point clouds consists in finding the rotation and the translation that maximize the overlap between the two clouds.This problem is a crucial building block in several applications that solve more complex tasks.These include simultaneous localization and mapping (SLAM), object detection and recognition, augmented reality and many others.Point cloud registration has been addressed by many authors and an excellent overview is given by Pomerleau et al. in [?].Usually, this problem is solved by using some variant of the Iterative Closest Point (ICP) algorithm, originally proposed by Besl and McKey [?].Among all these variants, Generalized ICP (GICP) [?] showed to be one of the most effective and robust.The availability of affordable high speed RGB-D sensors, like the Microsoft Kinect or the Asus Xtion, enlarged the basin of potential users and resulted in an increased interest from researchers.Registration techniques that were previously designed for 3D laser scanners, when used in conjunction with these sensors, suffer from the different noise characteristics of the depth measurements.Furthermore, an RGB-D camera can deliver data up to 60Hz and such frame rate is not achievable by common 3D laser scanners, that typically require several seconds for a dense scan.RGB-D cameras led to the development of approaches that benefit from this high data rate like Kinect Fusion (KINFU) [?] or Dense Visual Odometry (DVO) [?].These methods, however, are not well suited to deal with 3D laser scanners data since they require a good guess of the current transformation that is available only at high frame rates.In this paper, we present a complete system to recursively align point clouds that combines and extends state-of-theart approaches in both domains to achieve enhanced speed