Enhanced YOLOv8-based Lightweight Small Personnel Detection Algorithm for UAV Flood Emergency Rescue

This study proposes an enhanced lightweight small-target detection algorithm tailored for UAVbased flood emergency rescues, building upon YOLOv8. By introducing a Linear Deformable Convolution kernel and a redesigned bottleneck structure with partial convolution, the algorithm not only captures personnel target features of different scales and shapes more efficiently and achieves higher detection accuracy, but also reduces the number of model parameters. In addition, by improving the structure of the detection head and adding the ResNeXt-SENet fusion layer, the algorithm is able to suppress the interference of the complex background in emergency rescue scenarios and focus more on detecting small-targeted people, while improving the global information integration capability of the model, so that the algorithm is better applicable to different small-targeted detection datasets. Evaluation on custom flood-rescue datasets and VisDrone2019 demonstrates a significant increase in detection accuracy for small targets and reduction in the number of model parameters. The detection accuracy and model size also compare favorably with other state-of-the-art target detection algorithm models under the same experimental conditions, highlighting the suitability of the model for resource-constrained real-time UAV applications in challenging environments.