Stein ICP for Uncertainty Estimation in Point Cloud Matching

Quantification of uncertainty in point cloud matching is critical many tasks such as pose estimation, sensor fusion, and grasping. Iterative closest (ICP) a commonly used estimation algorithm which provides estimate the transformation between two clouds. There are sources this process that may arise due to noise, ambiguous environment, initial condition, occlusion. However, for safety problems autonomous driving, not sufficient it does provide information about multiple solutions. Current probabilistic ICP methods usually do capture all unreliable estimates can have detrimental effect state or decision making use information. In work we propose new align clouds precisely ICP’s parameters. We develop Stein variational inference framework with gradient based optimization cost function. The method non-parametric transformation, model complex multi-modal distributions, be effectively parallelized on GPU. Experiments using 3D kinect data well sparse indoor/outdoor LiDAR show our capable efficiently producing accurate estimates.