RSOS-Net: Real-Time Surface Obstacle Segmentation Network for Uncrewed Waterborne Vehicles

作者：Ning Wang, Yuan Feng, Lixin Tian, Yi Wei · 发表于：IEEE Transactions on Intelligent Transportation Systems · 年份：2025 · DOI：10.1109/tits.2025.3628677 · 被引用次数：3 · 研究领域：Underwater Vehicles and Communication Systems、Image Enhancement Techniques、Advanced Neural Network Applications

Due to water-surface reflection, wake and sun glitter, an uncrewed waterborne vehicle (UWV) faces a long-standing challenge in identifying water-surface obstacles especially with small-scale appearance. In this paper, inspired by the encoder-decoder architecture, a real-time surface obstacle segmentation network (RSOS-Net) is created to enable online surface-obstacle detection for a UWV. Primarily, the improved lightweight feature pyramid network structure is deployed to flexibly accommodate significant scale-variations and enhance focus on small obstacles, simultaneously. To address visual ambiguities caused by water-surface disturbances, the fast pyramid pooling module (FPPM) and attention-based feature fusion module (AFFM) are holistically devised within lightweight encoder and decoder, respectively. Accordingly, the FPPM is able to distinguish obstacles from sun glitters by capturing both local and global contextual information via cascaded pooling, while the AFFM can rule out reflections by virtue of channel-spatial attention mechanism augmenting detailed features and spatial locations. Results show that the RSOS-Net achieves an F1 score of 65.1% on the LaRS dataset, while the detection speed reaches 79.5 frames per second on an NVIDIA RTX 3060 platform. Notably, the RSOS-Net secured first place in the 3rd USV-based Embedded Obstacle Segmentation Challenge, with official results available athttps://macvi.org/workshop/macvi25/summary