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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Gerontology

Background:

• Falls pose a significant safety risk for the elderly population.
• Existing object detection models for fall detection are often computationally intensive, hindering their use on resource-limited edge devices.
• Lightweight models may sacrifice accuracy for reduced computational demands.

Purpose of the Study:

• To develop a lightweight and accurate fall detection model suitable for edge device deployment.
• To address the trade-off between computational complexity and detection accuracy in elderly fall detection systems.
• To propose a novel fall detection model, Lightweight Fall Detection YOLO (LFD-YOLO), based on the YOLOv5 architecture.

Main Methods:

• Proposed a novel lightweight feature extraction module, Cross Split RepGhost (CSRG), to minimize information loss.
• Integrated Efficient Multi-scale Attention (EMA) for enhanced focus on human pose.
• Developed a Weighted Fusion Pyramid Network (WFPN) with Group Shuffle Convolutions (GSConv) for efficient multi-scale feature fusion and reduced complexity.
• Introduced an Inner Weighted Intersection over Union (Inner-WIoU) loss function to improve convergence and generalization.
• Created a diverse Person Fall Detection Dataset (PFDD) and utilized the Falling Posture Image Dataset (FPID).

Main Results:

• LFD-YOLO demonstrated improved mean Average Precision (mAP0.5) by 1.5% and 1.7% on PFDD and FPID datasets compared to YOLOv5s.
• Achieved a reduction in parameters by 19.2% and calculations by 21.3% compared to YOLOv5s.
• Outperformed YOLOv8s by reducing parameters by 48.6% and calculations by 56.1%, while improving mAP0.5 by 0.3% and 0.5%.
• The model exhibits higher detection accuracy and lower computational complexity.

Conclusions:

• LFD-YOLO offers a promising solution for elderly fall detection, balancing accuracy and efficiency.
• The proposed lightweight architecture and novel modules are effective for deployment on resource-constrained edge devices.
• This research contributes to improving the safety and well-being of the elderly through advanced AI technology.