A two-point algorithm for stereo visual odometry in open outdoor environments

This paper proposes a novel method to estimate relative poses for a calibrated stereo camera. Three corresponding points in 3D space are theoretically required to recover unconstraint motion which has six degrees of freedom. The proposed method solves this problem with only two 3D points by exploiting a common reference direction between poses. Two points are selected in accordance with the distance to the camera: one distant point is used for deriving a reference direction, and one near point is used for estimating accurate translation. The distance is computed by triangulation in stereo vision. The uncertainty of triangulation can be mitigated by the appropriate selection strategy. The experiments using synthetic and real data demonstrate the effectiveness and higher stability of the proposed method against image pixel noise.