Lightweight UAV Image Drowning Detection Method Based on Improved YOLOv7*

Using Unmanned Aerial Vehicle (UAV) for searching for drowning individuals can enhance the efficiency of search and rescue operations. To tackle the challenges posed by the YOLOv7 model, including its numerous parameters and high computational costs, as well as the difficulties in UAV deployment, we propose a lightweight method for UAV image drowning detection, utilizing an enhanced version of YOLOv7. We replace the backbone in YOLOv7 with MobileNetV4 and reconstruct the SPPCSPC structure to decrease the model's parameter count and computational complexity. Additionally, we enhance detection accuracy by improving convolutional layers and modifying pooling structures. Additionally, we introduce a CA attention mechanism to mitigate the problem of limited feature information caused by the small area of human bodies exposed on water surfaces. Finally, we use Wise-IoU as the model loss function to mitigate the adverse effects of low-quality samples. Experimental results show that our improved algorithm only sacrifices 1.08% mAP, the number of parameters was reduced by 46.7% compared to YOLOv7, increases detection speed by 97.8%, and better balances detection speed with accuracy.