Selecting Feature Detectors for Accurate Visual Odometry

This work analyzes the performances of different feature detectors/descriptors in the context of incremental path estimation from passive stereo vision (Visual Odometry). Several state-of-the-art approaches have been tested, including a fast Hessian-based feature detector/descriptor developed at INRIM. Tests on both synthetic image sequences and real data show that in this particular application our approach yields results of accuracy comparable to the others, while being substantially faster and much more reliable. Key–Words: Robot localisation, Stereo vision, Visual odometry, Feature descriptors, Feature detectors