Prof. Reinhard Klette
Date and Time:
14:00-15:00, Sept 12 2018
Variants for Visual Odometry Estimation
Despite a variety of implementation details, proposed VO approaches basically share the same principle - a minimisation of a carefully chosen energy function. This talk also reviews four commonly adopted energy models including perspective, epipolar, rigid, and photometric alignments, and proposes a novel VO technique that unifies multiple objectives for outlier rejection and egomotion estimation to outperform mono-objective egomotion estimation. Experiments show an improvement above 50% is achievable by trading off 15% additional computational costs. - This talk informs about joint work with Dr. Hsiang-Jen Chien and further (former) students or colleagues.
Visual odometry (VO) has been extensively studied in the last decade. VO aims at the recovery of a camera trajectory from an image sequence. Stereo vision-based VO techniques solve the egomotion estimation problem by means of disparity-derived 3D scene structure. Typically, one of the two images is only used for disparity computation.
This talk discusses at first the development of a generic feature tracking framework extending the classical VO problem into a higher dimension, where the image data of both cameras are fully used. Six tracking topologies proposed in literature, namely linear, lookahead, stereo linear, parallel, circular and crosseye, are reviewed and evaluated. Based on the experimental results, we found benefits of taking right images into account through the feature tracking process, over the typical stereo VO implementation. The stereo-parallel configuration, which independently maintains feature tracking on each camera and have the tracked features integrated via a left-right matching, has achieved the most significant improvement of 30% over the conventional linear configuration.