Monocular vision SLAM for large scale outdoor environment

We present an algorithm which can realize the monocular vision simultaneous localization and mapping(SLAM) for mobile robot in the large scale outdoor environment. Unused traditional mechanic odometer sensor information, we utilize the ideas of Structure From Motion(SFM) for step-to-step motion estimation reliably only with visual information. Scale Invariant Feature Transform(SIFT) image feature is used as natural landmark, and its 3D position is constructed directly through triangulation methodology after the scale of robot's translational motion is uniquely determined. Then in the Rao-Blackwellised particle filter framework, one concurrent Extended Filter(EKF) is used as proposal density function, and the associated landmark state is updated by one EKF filter independently in the corresponding landmark map of this particle. Finally, real experiments show that our method is feasible and robust, even against large translation and large rotation movements.