Large motion estimation for omnidirectional vision

We present a method for estimating large camera motions, including combinations of large rotations and displacements. We know that translation estimates are improved as the displacements between two camera positions increase. However, there is little work on motion estimation methods targeted at large motions since it is difficult to find correspondences between two planar perspective images under large motions. We can overcome the correspondence problem with omnidirectional cameras, which obtain a hemisphere or cylinder projection of the environment. We first develop a new motion estimation method for smooth omnidirectional camera motions. Our method incrementally improves estimates from incomplete information provided by a single feature correspondence, using an Implicit Extended Kalman Filter (IEKF). This general motion estimation method produces stable estimates from smooth camera motions, but it is not directly applicable to estimating large motions. We adapt the motion estimation method to large motions by using a novel Recursive Rotation Factorization (RRF) that removes the image motions due to rotation. Simulation results show that the RRF large motion estimates are accurate and robust.