Objectives Aiming at the problems of easy switching of tracking ID （identity document） and low accuracy caused by complex scenes such as platform shaking and small target caused by long distance during the working process of cleaning ships for collecting floating garbage on the water surface, a multi-target tracking method based on improved YOLOv5-Byte was proposed.

Methods First, the Byte data association model was fused with the YOLOv5 detector to construct the multi-target tracking algorithm. Secondly, aiming at the problem that CIoU (complete intersection over union） in YOLOv5 is sensitive to small targets, the normalized Wasserstein distance metric is proposed for the boundary box Gaussian modeling. Then, the balance factor is introduced to adjust the contribution of CIoU and normalized Wasserstein distance measures to the loss function to adjust the sensitivity of the detector to small targets. Finally, IoU (intersection of Union）is introduced into the adjustable hyper-parameter in the form of a power exponent in the Byte data association model to reduce the risk of small targets being discarded due to low confidence.

Results The experimental results on the surface floating garbage dataset showed that the evaluation index IDF1 (identification F1 score) and MOTA (multiple object tracking accuracy) increased by 11.5% and 8.7% respectively, while the number of IDs (identity switches) decreased by 7 compared to the algorithm before improvement.

Conclusion The proposed algorithm achieves accurate tracking of multiple small targets on the surface of the water, providing a reference for the autonomous trash-collecting technology of intelligent cleaning ships.