UNDERWATER OBSTACLE DETECTION IN WIRELESS SENSOR NETWORKS USING YOLOV8S

Abstract

It is still a very challenging problem to accurately and in real time detect the obstacles in sonar images for Autonomous Underwater Vehicles (AUVs) and Underwater Wireless Sensor Networks (UWSNs), especially in an underwater environment with clutter and noise, where the traditional sonar detection method is weak in accuracy and robustness. The undersea object detection technology currently available can be used to detect undersea objects, but it is not good at performing detection in noise environments, low visibility environments and complex target structures. To overcome these challenges, the light and efficient deep learning underwater acoustic target detection framework based on YOLOv8s architecture is presented in this paper. The model is trained and tested on an underwater acoustic target detection (UATD) dataset consisting of 1127 labeled sonar images from 10 types of obstacles. To boost feature extraction and model generalization, transfer learning with COCO-pretrained weights, advanced data augmentation, and AdamW optimization are used. Experimental results showed that the proposed approach achieved a precision of 92.81% and a recall of 91.07% with the mean Average Precision (mAP@50) being 94.80%. It achieves an mAP@50 of 8.4% improvement over the YOLOv7 model and enables efficient training on a single NVIDIA Tesla T4 GPU in about an hour, making it a suitable model for real-time and scalable underwater detection applications.

Keywords : Underwater obstacle detection; YOLOv8s; sonar image classification; Underwater Wireless Sensor Networks (UWSN); UATD dataset; deep learning; object detection; autonomous underwater vehicles.