Yolo-based power-efficient object detection on edge devices for USVs

Abstract

Advances in Artificial Intelligence, the Internet of Things, and Computer Vision have introduced challenges in minimizing resource usage-economically, energetically, and environmentally. This work presents a vision system for unmanned surface vehicles (USVs) focused on object detection and autonomous navigation. The system leverages hardware and software acceleration to enhance model performance while also evaluating energy efficiency. In this paper, we analyze the Ultralytics models across various platforms, including MYRIAD VPU, Intel CPUs/GPUs, and NVIDIA Jetson AGX Orin and Orin Nano. The results show that the Orin Nano is especially suitable for real-time detection, consuming less than 2 watts. To increase efficiency, optimization techniques, such as quantization and pruning, are performed. We also compare our models with related studies and assess YOLOv6 to YOLO11 in terms of inference time and FPS. YOLOv8-based models consistently deliver the best results, confirming their suitability for USV applications.