Abstract
In recent years, with the deep exploitation of marine resources and the development of maritime transportation, ship collision accidents occur frequently, which leads to the increasingly heavy task of maritime Search and Rescue (SAR). Unmanned Aerial Vehicles (UAVs) have the advantages of flexible maneuvering, robust adaptability and extensive monitoring, which have become an essential means and tool for emergency rescue of maritime accidents. However, the current UAVs-based drowning people detection technology has insufficient detection ability and low precision for small targets in high-altitude images. Moreover, limited by the load capacity, UAVs do not have enough computing power and storage space, resulting in the existing object detection algorithms based on deep learning cannot be directly deployed on UAVs. To solve the two issues mentioned above, this paper proposes a lightweight deep learning detection model based on YOLOv5s, which is used in the SAR task of drowning people of UAVs at sea. First, an extended small object detection layer is added to improve the detection effect of small objects, including the extraction of shallow features, a new feature fusion layer and one more prediction head. Then, the Ghost module and the C3Ghost module are used to replace the Conv module and the C3 module in YOLOv5s, which enable lightweight network improvements that make the model more suitable for deployment on UAVs. The experimental results indicate that the improved model can effectively identify the rescue targets in the marine casualty. Specifically, compared with the original YOLOv5s, the improved model mAP@0.5 value increased by 2.3% and the mAP@0.5:0.95 value increased by 1.1%. Meanwhile, the improved model meets the needs of the lightweight model. Specifically, compared with the original YOLOv5s, the parameters decreased by 44.9%, the model weight size compressed by 39.4%, and Floating Point Operations (FLOPs) reduced by 22.8%.