Feature Extraction Network Architecture.

To solve the problem of reduced positioning accuracy caused by changes in scale, background and occlusion in port and dock video images, this research proposes an enhanced model combining YOLOv5s-DeepSORT, integrating target load recognition and trajectory tracking to improve adaptability to dock environments. The findings indicate that incorporating multi-scale convolution into YOLOv5s improved the robustness of multi-scale object detection, resulting in a 0.4% increase in mean Average Precision (mAP). Furthermore, the integration of an efficient pyramid segmentation attention (EPSA) network enhanced the accuracy of multi-scale feature fusion representation. The model’s mAP@0.5:0.95 increased by 1.2% following the introduction of EPSA. Finally, the original classification loss function was enhanced using a distributed sorting loss approach to mitigate the imbalance among loaded objects and the influence of background variations in the dock image sequence. This optimization led to a 3.1% improvement in multi-target tracking accuracy (MOTA). Experimental results on self-constructed datasets demonstrated an average accuracy of 90.9% and a detection accuracy of 92.2%, offering a valuable reference for target recognition and tracking in port and dock environments.