Underwater Photogrammetry and Visual Odometry

Abstract

We propose an improved visual odometry approach that is adapted to low computational resources systems in an underwater environment. The aim is to guide underwater photogrammetry surveys in real time. The visual odometry relies on stereo image stream that is captured by an embedded system. An improved pose estimation procedure underlying fast stereo matching approach is followed by a semi-global bundle adjustment. Computed trajectory is maintained stochastically and a divergence measure is used for more realistic optimization zone selection. In particular, we propose a new approach to find an approximation of the uncertainty for each estimated relative pose based on machine learning manifesting on simulated data. This allows the user to find potential overlaps in the estimated trajectory for better drifts handling and loop closure. The evaluation of the proposed method demonstrates the gain in terms of computation time w.r.t. other approaches. The built system opens promising areas for further development and integration of embedded vision techniques.