Effective Feature-Based Downward-Facing Monocular Visual Odometry

Hoyong Lee, Hakjun Lee, Inveom Kwak, Chiwon Sung, Soohee Han

Abstract: To achieve accurate pose estimation for robots in industrial applications and services, this brief proposes an effective feature-based downward-facing monocular visual odometry technology that uses an affordable sensor system and a systematic optimization approach. To extract more effective features simply and efficiently from images of the ground, even for small mobile systems, the proposed visual odometry system is designed in a lightweight and cost-effective manner; we used an easily available LED, a single-channel time-of-flight (ToF) sensor, and a monocular camera. From the extracted features, the potentially irrelevant ones are removed in advance, using a masking algorithm and measured velocity. This enhances feature efficiency and reduces the computational burden. Finally, the optimal pose estimate is explicitly obtained by solving a nonconvex optimization problem, to make the best use of the features. The experiments’ results show that our proposed method improves feature tracking ability and pose estimation accuracy.