This article explains the background, workflow, and technical details of using the YOLOv5 one‑stage object detection algorithm to enable fast, accurate video analytics for urban management, covering data augmentation, backbone design, FPN‑PAN neck, and prediction output processing.
The inaugural Ship‑Sea Data Intelligent Application Innovation Competition, co‑hosted by Taihu Laboratory, Huawei and Wuxi authorities, showcased cutting‑edge AI techniques—such as multi‑scale training, TTA, knowledge distillation, and model pruning—to improve surface and underwater target detection for vessels, nets, buoys, and marine life, while offering transparent rankings, research funding, and a platform for advancing maritime AI.
The article reviews deep learning object detection solutions: small object detection via FPN and TDM, irregular shapes via deformable convolution, sample imbalance via focal loss and cascade methods, occlusion handling with Soft‑NMS and RRC, large‑batch training using MegDet, relationship modeling with Relation Networks, and network improvements such as DetNet, RefineDet, Pelee, and RFBNet.
