A dual-path scenario-guided framework for maritime object detection under heterogeneous sea conditions
Year: 2026
Type: article
Abstract: Maritime object detection is a critical component for Unmanned Surface Vehicles (USVs) yet faces dual challenges: target geometric deformation induced by complex sea states and the identification of sparse, small-scale objects in open waters. However, conventional “one-size-fits-all” optimization strategies often fail to simultaneously balance feature alignment in dynamic scenes with regression precision in static scenarios. To address this deficiency, we propose Hy-YOLO, a scenario-guided framework that offers specialized optimization paths. This framework first establishes a universal attention-aware baseline to ensure excellent scenario generalization. Building on this foundation, it introduces a dual-path optimization mechanism: (1) For complex dynamic environments, the Hy-YOLO-G (Geometric) variant integrates Deformable Conv (DCNv3) to geometrically align features of non-rigid targets; (2) For sparse small-object scenarios, the Hy-YOLO-L (Localization) variant employs a hybrid loss function combining SIoU and Focal EIoU to enforce strict geometric constraints and hard sample mining. Unlike fully autonomous black-box models, Hy-YOLO provides a deterministic engineering selection law based on maritime imaging characteristics. Extensive experiments on YOLOv8 and YOLOv13 baselines demonstrate that this scenario-guided strategy yields superior performance. Specifically, Hy-YOLO-G improves mAP@50 by 1.5 percentage points on complex datasets (WSODD), while Hy-YOLO-L achieves a 4.2 percentage point boost on small-object datasets (Flow). Furthermore, robustness tests under synthetic fog and glare conditions confirm the framework’s reliability. Ultimately, this work offers a flexible and efficient solution for the deployment of USVs under extreme and diverse maritime conditions.
Source: Frontiers in Robotics and AI
Authors: Yongze Li(Formal analysis, Methodology, Writing – original draft, Writing – review & editing), Hailiang Zhao(Conceptualization, Data curation, Supervision, Validation, Writing – review & editing), Yongyi Kong(Conceptualization, Data curation, Validation, Writing – review & editing), Liqiang Wei(Resources, Visualization, Writing – review & editing), Dongkui Wu(Writing – review & editing), Jiansen Zhao(Data curation, Writing – review & editing)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,353
OpenAlex ID: W7235698846
