Deep learning-based floating garbage detection for unmanned surface vehicles in complex water environments
Yuzhao Xie, Da Lian, Qiyang Ling, Saiyu Yuan, Xianglong Wei, Min Chen, Dantong Zhu, Xiangju Cheng · Journal of Hydroinformatics · 2026
ABSTRACT With the increasing problem of floating garbage, unmanned surface vehicles (USVs) have developed as a water surface cleaning technology. As the foundation of floating garbage cleaning for USVs, object detection faces challenges, including small objects, complex weather conditions, object occlusion, and the computational limitations of edge computing devices. This study proposes GBS-YOLO, an improved floating garbage detector for complex water environments. The network enhances target attention by introducing the GAM, promotes feature propagation and fusion through BiFPN, and improves model robustness using SIoU. It achieves model lightweighting through DepGraph structural pruning and channel-wise knowledge distillation. The model has been deployed on an independently developed USV. GBS-YOLO achieves a 4.8% AP50 gain on the original test set, with additional gains in the simulated rain and occlusion test sets. After lightweighting, the model's size is compressed to 6.9 M with only a 0.8% decrease in accuracy, and its inference speed is increased by 34%. It can meet the real-time detection requirements on embedded computing devices. The deployment results demonstrate the feasibility of real-time onboard inference on embedded computing devices, providing a technical basis for future unmanned surface vehicle applications in floating garbage detection and collection.