3D SLAM for omnidirectional camera

This paper proposes a method for simultaneous localization and mapping using a hand-held omnidirectional camera traversing in a 3D environment with an unpredictable trajectory. Unlike most existing works, the method does not assume any motion model of the camera. The proposed method follows the extended Kalman filter (EKF) framework for which we propose an update process that takes into account many camera's poses estimated several steps prior to the current update. Each of these poses is used as a reference for approximating the current pose using a nonlinear least square computation. This update process is shown to efficiently avoids map divergence. The method is implemented and preliminary experimental results are presented.