Sea Docking by Dual-eye Pose Estimation with Optimized Genetic Algorithm Parameters

摘要

Three-dimensional (3D) estimation of position and orientation (pose) using dynamic (successive) images input at video rates needs to be performed rapidly when the estimated pose is used for real-time feedback control. Single-camera 3D pose estimation has been studied thoroughly, but the estimated position accuracy in the camera depth of field has proven insufficient. Thus, docking systems for underwater vehicles with single-eye cameras have not reached practical application. The authors have proposed a new 3D pose estimation method with dual cameras that exploits the parallactic nature of stereoscopic vision to enable reliable 3D pose estimation in real time. We call this method the "real-time multi-step genetic algorithm (RM-GA)." However, optimization of the pose tracking performance has been left unchallenged despite the fact that improved tracking performance in the time domain would help improve performance and stability of the closed-loop feedback system, such as visual servoing of an underwater vehicle. This study focused on improving the dynamic performance of dual-eye real-time pose tracking by tuning RM-GA parameters and confirming optimization of the dynamical performance to estimate a target marker's pose in real time. Then, the effectiveness and practicality of the real-time 3D pose estimation system was confirmed by conducting a sea docking experiment using the optimum RM-GA parameters in an actual marine environment with turbidity.