Unmanned Surface Vehicles (USVs) require robust real-time visual perception to safely operate in dynamic maritime environments characterized by waves, reflections, illumination variations, and highly irregular object distributions. However, convention...
Unmanned Surface Vehicles (USVs) require robust real-time visual perception to safely operate in dynamic maritime environments characterized by waves, reflections, illumination variations, and highly irregular object distributions. However, conventional computer vision models—primarily optimized for terrestrial scenes or high-performance GPU servers—struggle to meet the strict latency, power, and thermal constraints of onboard Edge-AI platforms. This study presents a unified real-time maritime vision perception pipeline that integrates object detection and sea-region segmentation models optimized specifically for USV onboard operation. A large-scale maritime vision dataset consisting of 226,000 real-sea images was constructed using USVs, drones, and fixed cameras under diverse weather, illumination, and sea-state conditions. Based on this dataset, a maritime-adapted YOLOv7 detector was developed through anchor re-optimization, SIoU-based loss enhancement, class rebalancing, and maritime-specific augmentation techniques. For sea-region perception, an enhanced U-Net model incorporating AIN (Adaptive Illumination Normalization), RAA (Reflection-Aware Attention), and WSD (Wave-Sensitive Decoder) was designed to improve robustness against reflections and wave-induced distortions. To enable real-time inference on Jetson AGX Orin, Post-Training Quantization (PTQ) and a carefully curated 1,000-image calibration dataset were applied, achieving significant reductions in model size and inference latency. Both models were integrated into an RTSP-based unified perception pipeline, where detection and segmentation outputs are fused to produce navigable-path and maritime-scene understanding in real time. Experimental results demonstrate improved mAP, recall, and small-object detection performance, as well as stable segmentation under challenging maritime conditions. The INT8-optimized pipeline achieves 55 ms end-to-end processing latency, satisfying the real-time requirements of USV onboard systems. Overall, this work establishes a practical and efficient onboard Edge-AI maritime perception framework, providing a foundation for real-time autonomous navigation and anomaly awareness in future USV systems.