Hardware constrained fast maritime obstacle detection for autonomous vesselsTeršek, Matija (Avtor) Kristan, Matej (Mentor) Žust, Lojze (Komentor) semantic segmentationunmanned surface vehiclesmobile networkreal time performancelightweight network architectureMetaFormerPrecise detection of obstacles is needed for the successful navigation of autonomous vessels. Recent approaches use semantic segmentation to better generalize to unknown scenarios. However, the majority do not consider computational complexity and constraints, which makes proposed architectures undeployable on edge VPUs. In this thesis, we develop a novel architecture based on state-of-the-art water segmentation and refinement network WaSR. We explore various encoders and decoder modifications to propose a novel architecture. Based on WaSR, TopFormer, designed for mobile semantic segmentation, and an abstracted transformer architecture MetaFormer, for which we propose previously unused token mixers, we introduce WaSRFormer. It uses a TopFormer-like decoder with MetaFormers for speedup and good practices from WaSR to deal with diverse water features. On challenging MODS benchmark, WaSRFormer achieves 92.98% and 86.27% F1 score overall and inside the danger zone, respectively, with only 0.51% and 0.25% drop in F1 score compared to WaSR. On a modern laptop GPU it runs more than 10x faster (115.45 FPS) than WaSR (10.94 FPS). To emphasize the practical contribution, we deploy WaSRFormer on the embedded low-power hardware OAK-D. While WaSR cannot even fit into this hardware, WaSRFormer comfortably runs at 5.45 FPS.20222023-09-18 18:49:00Magistrsko delo/naloga150473sl