adriennewinter/Marine_Vision

MSc research project in the African Robotics Unit research group at the University of Cape Town # Marine Vision ## MSc (Research) Mechatronics Engineering dissertation. **_Author:_** Adrienne Winter. **_Supervisors:_** Robyn Verrinder, Professor Edward Boje and Doctor Paul Amayo. **_Research Group:_** African Robotics Unit (ARU) at the University of Cape Town (UCT). **_Funders:_** South African International Maritime Institute (SAIMI) and Oppenheimer Memorial Trust. This dissertation explores the use of underwater SLAM (Simultaneous Localization and Mapping) for visual reconstruction of ship hulls — a problem with major implications for the maritime industry, which carries the majority of global trade. The research focuses on camera-based methods for mapping in feature-sparse underwater environments in two main areas - dataset creation and visual odometry in the context of SLAM. ## Sensor Rig Construction A lack of well-documented underwater SLAM datasets was noted and a prototype multi-sensor rig was designed and tested for the creation of a dataset (stereo-inertial-pressure). The rig construction highlighted practical lessons in hardware robustness. The ROS workspace and 3D printing files are provided here. Image of the prototype sensor rig. ## ROS Data Synchronisation Package A C++ ROS data synchronisation package was created that synchronises up to 4 sensor topics using timestamps and preserves original sensor frequencies. The package can be found in the ROS workspace of the sensor rig here. Diagram of a feature depth estimation issue encountered when synchronisation is performed using timestamps created by different clocks. ## Underwater Camera Calibration The logistical difficulty of collecting good underwater camera calibration, necessary for producing high-quality SLAM datasets, was noted. An open-source calibration method that addresses this (the Pinax camera calibration model) was evaluated and contributed to. The Pinax calibration method was shown to significantly improve underwater stereo depth estimation accuracy (6.2% mea …