LFD-YOLO: a lightweight fall detection network with enhanced feature extraction and fusion.

Falls are one of the significant safety hazards for the elderly. Current object detection models for fall detection often suffer from high computational complexity, limiting their deployment on resource-constrained edge devices. Although lightweight models can reduce computational requirements, they typically compromise detection accuracy. To address these challenges, and considering the more lightweight architecture of YOLOv5 compared to other YOLO series models such as YOLOv8, we propose a lightweight fall detection model based on YOLOv5, named Lightweight Fall Detection YOLO (LFD-YOLO). Our method introduces a novel lightweight feature extraction module, Cross Split RepGhost (CSRG), which reduces information loss during feature map transmission. We also integrate an Efficient Multi-scale Attention (EMA) to enhance focus on the human pose. Moreover, we propose a Weighted Fusion Pyramid Network (WFPN) and utilize Group Shuffle Convolutions (GSConv) to reduce the model's computational complexity and improve the efficiency of multi-scale feature fusion. Additionally, we design an Inner Weighted Intersection over Union (Inner-WIoU) loss to accelerate model convergence and enhance generalization. We construct a Person Fall Detection Dataset (PFDD) dataset covering diverse scenarios. Experimental results on the PFDD and the publicly available Falling Posture Image Dataset (FPID) datasets show that, compared to YOLOv5s, LFD-YOLO improves mAP0.5 by 1.5% and 1.7%, respectively, while reducing the number of parameters and calculations by 19.2% and 21.3%. Furthermore, compared to YOLOv8s, LFD-YOLO reduces the number of parameters and calculations by 48.6% and 56.1%, respectively, while improving mAP0.5 by 0.3% and 0.5%. These results demonstrate that LFD-YOLO achieves higher detection accuracy and lower computational complexity, making it well-suited for fall detection tasks.