Terrain adaptive detector selection for visual odometry in natural scenes

Pose estimation is one of the important tasks for mobile robots exploring in outdoor environments. Recently, visual odometry has received a lot of attention since its localization is accurate even with low-cost sensors. Furthermore, the technique is not affected by wheel slips, and it can be performed without external infrastructures and preliminary maps. While existing techniques successfully provide good localization of outdoor vehicles, possible failures are not yet fully examined in untextured terrains where feature tracking occasionally fails. This paper proposes an approach to detect interest points from a wide variety of terrains by adaptively selecting algorithms. Experiments show that the approach provides robust and fast interest point detection even in untextured natural scenes.