Benchmarking Marine Litter Detection on USVs: A study on dataset and camera parameters choices for edge AI

Floating marine litter monitoring can be addressed with camera-based artificial intelligence (AI) and marine robotics. However, robust deployment in limited resource environments such as Unmanned Surface Vehicles (USVs) or coastal optical camera systems is still challenging. Foundation models and zero-shot detectors can still be computationally too expensive to run effectively on edge devices. Thus, the need to employ traditional lightweight object detectors, which, on the other hand, lack generalization performance. This work presents an experimental study of real-Time floating litter detection from an unmanned surface vehicle (USV) using a state-of the-Art lightweight object detector deployed on CPU hardware. The first contribution is a condensed comparative analysis of existing floating-litter datasets, individually and in combination. The second contribution is a systematic comparison of camera settings, including elevation, shutter control, wide dynamic range (WDR), and infrared mode. The paper provides practical guidelines for reproducible USV-based marine litter detection under multiple image acquisition conditions.