Coastal obstacles pose a threat to the safety of nearshore navigation vehicles. Currently, there is a lack of dedicated datasets for detecting such targets. Compared with ground platforms, Unmanned Aerial Vehicles (UAVs) have unique advantages in coastal monitoring. However, their limited onboard computational power makes it difficult to ensure high-precision real-time identification. To address this, this paper constructs a coastal obstacle dataset encompassing six categories (including Rail-obstacle, soldier, anti-tank pyramids, tank, wire entanglement and hedgehog) and proposes an improved YOLOv5s lightweight model for UAV-based coastal obstacle recognition. The model’s backbone network utilizes C3Ghost and GhostConv modules to reduce computational load. The neck network integrates DWConv and C3Ghost to compress the parameter count. Furthermore, Squeeze-and-Excitation (SE) and Coordinate Attention (CA) modules are incorporated to suppress interference caused by complex coastal environments during feature extraction. Experiments demonstrate that while maintaining only a slight decrease in mAP@0.5 accuracy on the self-built dataset, the model achieves a 52.26% reduction in FLOPs and a 52.58% decrease in parameters. This significantly enhances the UAV’s real-time perception capability in complex coastal environments, providing a reliable technical solution for coastal safety monitoring and the safe operation of nearshore navigation vehicles.