Open Water Detection for Unmanned Surface Vehicles

Unmanned Surface Vehicles (USVs) require robust perception systems to operate autonomously and safely across diverse environments. Among these, the ability to detect navigable water surfaces is crucial, as it enables higher-level tasks such as path planning, obstacle detection, and classification. These perception systems can be developed using techniques from Computer Vision and Machine Learning. This thesis lays the groundwork for implementing water segmentation and environmental awareness on a real maritime USV (BlueBoat). Due to strict equipment constraints, the experiment aimed to evaluate the robustness of a pre-trained image classification model under varying conditions, including different locations, image quality, resolution, and other environmental challenges. The results showed promising performance in image-based water detection, suggesting that inference models can serve as cost-effective alternatives to traditional sensors. However, computational time emerged as a critical factor in selecting suitable hardware for real-time inference.