Vol 77 No 4 77406

Abstract

Enhancing the environmental perception of Unmanned Surface Vehicles (USVs) in complex waters is among the primary approaches to ensuring the safety of autonomous navigation. This paper proposes a robust stereo-vision system that integrates ship detection and localization for maritime scenarios. In terms of detection, GS-YOLO is proposed to address the problems of large model parameters and low detection accuracy for small target ships. Based on YOLO11n, the Global-Local Spatial Attention (GLSA) module, Bidirectional Feature Pyramid Network (BiFPN), and SIoU loss function are introduced to improve detection performance while maintaining model lightness. For localization, a cascaded filtering localization algorithm is proposed to address the instability of distance measurement caused by dynamic interference. The algorithm takes depth data generated by RAFT-Stereo as input and applies temporal smoothing by sequentially combining median filtering and Kalman filtering. This significantly enhances robustness against dynamic interference. Experimental results show that GS-YOLO achieves a mean average precision of 93.5 % on the Mcships dataset while reducing parameters by 17.8 % compared with YOLO11n. It achieves an optimal balance between detection accuracy and model lightweighting. Additionally, compared with traditional methods, the cascaded filtering localization algorithm significantly reduces positioning error. Within the range of 50 m, the standard deviation of the error is reduced by 89.3 %; within an 80 m range, the positioning error is maintained below 2.3 %. These results demonstrate that the proposed stereo-vision system provides accurate and stable perception data for the autonomous navigation of USVs.