Opti-acoustic underwater SLAM using style-transferred multi-modal matching

Abstract

Underwater robot pose and odometry estimation has constraints that have to solve challenging sensor modalities. In this thesis, I propose the underwater vehicle odometry optimization and estimation method with neural network and opti-acoustic-inertial module configuration. Neural network-based style transfer enables feature detection and matching between optic and sonar images. Feature points are utilized for deriving the reprojection errors and extrinsic matrix between the sonar and camera. Robot pose estimation is accomplished with error factor minimization, and optimize the odometry with pose factor inserted graph simultaneous localization and mapping (SLAM). Also learning-based selective link proposal facilitates efficient SLAM processing. The complications came from sensor modality are overcame with style transfer method and opti-acoustic-inertial module configuration.