Robust Obstacle Detection for Unmanned Surface VehiclesŽust, Lojze (Avtor) Kristan, Matej (Mentor) maritime visionobstacle detectionsegmentationpanoptic segmentationUnmanned surface vehicles (USVs) offer significant potential for automating repetitive or hazardous maritime tasks, such as hazard inspection, bathymetry, environmental monitoring and public transport. Visual perception is central to enabling autonomous navigation in USVs, yet despite rapid progress in camera-based obstacle detection, domain-specific challenges continue to hinder real-world deployment. This thesis examines and addresses several of these challenges. First, we tackle the high annotation cost of segmentation-based obstacle detection models by introducing a lightweight annotation scheme and a corresponding training method, SLR, which effectively leverages such annotations. Models trained with SLR match and even exceed the performance of fully supervised models while requiring 20× less annotation effort. Next, we address the limited robustness of existing methods to water reflections by proposing a temporally aware model, WaSR-T, which exploits motion cues to distinguish objects from their reflections, resulting in more robust and temporally consistent predictions. We then advocate a shift toward joint instance and semantic segmentation (i.e. panoptic segmentation). To overcome the lack of suitable training data, we introduce LaRS, a large and diverse maritime dataset with high-quality panoptic annotations, and benchmark a wide range of semantic and panoptic segmentation methods to highlight remaining challenges. Finally, we present PanSR, a novel panoptic maritime segmentation method that addresses key architectural limitations of prior work, improving small-object detection, handling of crowded scenes, and instance separation. Overall, this thesis advances maritime obstacle detection through new methods, datasets, and benchmarks that collectively move the field toward panoptic, navigation-aware maritime perception.20262026-10-07 11:20:05Doktorsko delo/naloga189465sl