Multi-Target Tracking Algorithm for Floating Garbage on Water Surfaces Based on Unmanned Surface Vessels

To address the challenge of high-precision real-time detection and tracking of small floating debris from an unmanned surface vehicle (USV) perspective under limited computational resources, this paper first constructs a multi-object tracking model, YOLOv5s-B, by integrating the YOLOv5 detector with the Byte data association algorithm, and then proposes an enhanced framework, YOLO-EENB. The proposed framework employs a lightweight EfficientNetV2 backbone to substantially reduce model complexity while preserving feature extraction capability; incorporates an Efficient Channel Attention (ECA) module to enhance feature representation with minimal computational overhead; and introduces a distance-aware loss function based on the Normalized Wasserstein Distance (NWD) to optimize small-object bounding-box regression accuracy.To evaluate the performance of the proposed method, comparative experiments were conducted on a custom floating debris dataset, benchmarking YOLO-EENB against the baseline YOLOv5s-B model. Experimental results demonstrate that YOLO-EENB achieves improvements of 7.5%, 10.1%, 12.2%, and 11.8% in IDF1, IDR, Recall, and MOTA metrics, respectively, while reducing FN, FP, and ID switches by 28.8%, 20%, and 7.1%. Moreover, the proposed model attains approximately 24.5% higher FPS, indicating superior real-time performance and computational efficiency. Finally, YOLO-EENB is deployed on a USV platform, achieving stable, continuous tracking without ID switches or target loss in practical scenarios. The proposed solution offers an efficient and feasible approach for intelligent water-surface cleaning systems operating under resource constraints.