1: \begin{abstract}

2: In mobile robotics, scan matching of point clouds using Iterative Closest Point

3: (ICP) allows estimating   sensor displacements. It may prove important to assess

4: the associated uncertainty about the obtained rigid transformation, especially

5: for sensor fusion purposes. In this paper we propose a novel approach to 3D uncertainty of ICP  that accounts for all the sources of error as listed in

6: Censi's pioneering work \cite{censiAccurate2007}, namely wrong convergence, underconstrained situations,

7: and sensor noise. Our approach builds on two facts. First, the uncertainty about the ICP's output fully depends on the

8: initialization accuracy. Thus speaking of the covariance of ICP makes   sense only in relation to the  initialization uncertainty, which generally stems from odometry errors.  We capture this using the unscented transform, which also reflects

9: correlations between initial and final uncertainties. Then, assuming white

10: sensor noise leads to overoptimism as ICP is biased owing to e.g. calibration

11: biases, which we account for. Our solution is tested on  publicly available real

12: data ranging from structured to unstructured environments, where our algorithm

13: predicts consistent results with actual uncertainty, and compares   favorably

14: to previous methods.

15: \end{abstract}

16: