Visual Perception of Underwater Robotic Swarms

Abstract

The field of swarm intelligence has gained significant popularity in recent years, particularly within the domain of robotics, where using a swarm of robots offers increased robustness and efficiency. However, it also has its challenges such as dependable power supply, accurate location estimation, and efficient communication underwater. This thesis investigates the visual perception of underwater swarm agents to eventually coordinate a swarm of autonomous underwater vehicles (AUVs) known as H-SURF (Heterogeneous Swarm of Underwater Robotic Fish). The research aims to address those challenges by developing a visual perception framework focusing on three key tasks: multi-object detection and tracking, relative localization, and heading estimation. Utilizing a Tracking by Detection scheme for Multi-Object Tracking, we conducted several comparative analyses to properly select the optimum Tracker and Detector. We optimized the YOLOv5n model for real-time multi-object detection and tracking on the HSURF custom dataset, achieving a balance between speed and accuracy suitable for deployment on Raspberry Pi 4 microprocessors. Key findings indicate that ByteTrack and OC-SORT provided the most reliable tracking performance, with ByteTrack excelling in speed and OC-SORT in handling occlusions. Relative localization was implemented using geometric projection and monocular depth estimation models. Heading estimation leveraged keypoint detection methods to accurately determine the yaw angle of neighboring agents, essential for effective swarm coordination. Experimental results for relative localization and heading estimation showed an average heading error of 17.2 degrees and an average depth error of 24.4 cm. These results indicate a moderate level of accuracy in the localization and heading estimation tasks, underscoring the need for continued refinement of the estimation algorithms to enhance their precision and reliability in varying environmental conditions. The optimized visual perception framework developed in this research enhances the operational effectiveness of underwater robotic swarms, offering promising applications in marine exploration, environmental monitoring, and seabed scanning. This thesis lays a solid foundation for future research in underwater swarm robotics, emphasizing the need for enhanced hardware, advanced algorithms, adaptive techniques for varying underwater conditions, and extensive field testing to ensure robustness and reliability in real-world applications.

Date of Award	20 Jul 2024
Original language	American English
Supervisor	Jorge Dias (Supervisor)