Fast Stereo-based Visual Odometry for Rover Navigation

The object of visual odometry is the computation of the path of a rover from onboard passive vision data only. The approach presented here relies on the accumulation of ego-motion estimates obtained by stereo vision and bundle adjustment of tracked feature points. We also propose a new feature detector/descriptor, which is a simplified and faster form of other well known descriptors (SURF). For cyclic paths, a déjà vu mechanism allows further control over the accumulated error. Tests on real-world data show that our descriptors are effective for accurate path estimation, while being fast enough for use in tasks such as autonomous planetary exploration. Key–Words: Robot localisation, Stereo vision, SLAM, Hessian feature detectors