Implementation and experimental verification of a 3D-to-2D Visual Odometry Algorithm for real-time measurements of a vehicle pose

This work describes the implementation of a software for real-time vehicle pose reconstruction using stereo visual odometry algorithms. We firstly guess motion by a linear 2D-to-3D method solving a PnP problem embedded within a RANSAC process to remove outliers. Then, a maximum likelihood motion estimation is performed minimizing a non-linear problem. GPU implementations of feature extraction and matching algorithms are used. We demonstrate an accuracy of better than 1% over 1350 mm of travel