Monocular Visual Inertial Odometry for Underwater Vehicle Navigation, Optimized on Embedded System

Abstract

Underwater robots are widely used in various operations below the surface of the sea. In many of these operations the precise localization and positioning is quite important, especially in cases that the robot must navigate between man-made installations. Although underwater vehicles can carry a number of different sensors, many of them can not be exploited for efficient navigation, due to costs, complexity, weight and inaccuracy. Most of robots feature at least one camera, a sensor that can be cheap, small, with low power requirements and that offers vast amount of information. The advances in computer vision and embedded systems have set cameras as an attractive mean for navigation. This combination allows computer vision as a viable alternative to provide data in real-time. This data can be used locally as a positioning source for the own robot control system, or to inform the navigation coordinates to an observer
or another system. Visual Odometry (VO) is the process of estimating the position and orientation of a robot based completely on data acquired by cameras. The camera-only system can be enhanced with the fusion of an Inertial Measurement Unit (IMU) to estimate motion even more accurately, resulting in Visual Inertial Odometry (VIO). This thesis researches various approaches for underwater navigation with edge computing and focuses on systems that utilize a camera and an IMU. A VIO system architecture is proposed, comprised by a machine vision camera, an IMU, a micro-controller and the embedded system Jetson Xavier from NVIDIA. Some of the most recent VO/VIO algorithms are evaluated on a number of underwater datasets and the impact of the IMU on their performance is assessed. Design issues are discussed and challenges related to the camera - IMU fusion are analyzed. The Visual-Inertial ORB-SLAM (ORB-SLAM + IMU) algorithm is deployed on the proposed VIO system and is optimized on the embedded GPU. The whole framework is evaluated on an artificial underwater structured environment which was created in an indoor tank. While the outcome of the algorithm on motion estimation is examined, the computational performance of the embedded system is profiled for various power modes as well. Results show that the proposed VIO system is able to estimate the underwater robot’s traversed trajectory in the tank with adequate accuracy and that can execute the Visual-Inertial ORB-SLAM in real-time with sufficient speed.