LAE-YOLO: A Lightweight Attention-Enhanced Detection Algorithm for Vision-Based Autonomous Rescue USV

Abstract

Water surface target detection for autonomous rescue USVs faces significant challenges due to complex lighting conditions, reflections, and small partially submerged targets, while edge deployment constraints demand low computational cost. To address these issues, this paper proposes a Lightweight Attention-Enhanced YOLOv11(LAE-YOLO) for vision-based autonomous rescue USV. Two modules are integrated into the baseline YOLOv11-nano: a Global Attention Module (GAM) to enhance channel-wise feature representation against water surface clutter, and a DySample-based upsampling layer to improve small-target feature extraction. The algorithm is deployed on a heterogeneous dual-core platform consisting of Raspberry Pi 5 for inference and STM32 for real-time control. Experiments on a self-built water surface dataset show that LAE-YOLO achieves 77.6% mAP@0.5, outperforming the baseline by 3.9%, while maintaining 8-9 FPS on Raspberry Pi 5. Field tests demonstrate autonomous cruise, personnel detection (>20 m), obstacle avoidance, and approximately 2.5 h endurance.