Development of Unscented Kalman filter algorithm for 3-D point based rigid registration

This paper proposes an effective algorithm for estimating the transformation parameters which align two data sets belonging to the rigid objects. This point based algorithm uses unscented Kalman filter for estimating the state vector of transformation which is a nonlinear function of translation and rotation. It is also assumed that the Gaussian noise is added to the fixed data set. In this paper, firstly we investigate the newest algorithm performance for point based rigid body registration using UKF and then we show the drawback of this algorithm in estimating of high range rotations. It is shown that the UKF algorithm is sensitive to selecting the appropriate initial sate vector. For solving limitation of the UKF algorithm, we propose an extended unscented kalman filter algorithm. It is shown that by exploiting this new algorithm which uses pre-registration, we can find an appropriate initial state vector. In this paper, we compare the results of EUKF algorithm with the previous UKF algorithm for different rotations. We demonstrate the effectiveness of the EUKF algorithm for estimation of high transformations between two data sets.