Optimization of Unscented Kalman Filter Algorithm for 3-D Point Based Rigid Registration

This paper proposes an optimized algorithm for 3-D Point Based Rigid Registration. This algorithm uses an Unscented Kalman filter (UKF) for estimating the state vector of transformation, which can be interpreted as a nonlinear function of translation and rotation. In the previous work, we showed that the drawback of the UKF algorithm in estimating high range rotations is due to its sensitivity to initial state vector. To address this drawback, we proposed using a preregistration step to find an appropriate initial state vector [9]. In this paper we optimize our proposed algorithm to trade off running time with accuracy by selecting the initial state vectors out of a uniformly sampling and using a large error threshold for stopping the pre-registration stage. It is shown that by applying these strategies we can have an enhanced UKF algorithm, which can robustly estimate any rigid transformation with high accuracy and acceptable time consumption.