An Edge-device Based Fast Fall Detection Using Spatio-temporal Optical Flow Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Elderly fall detection is crucial for healthcare, but traditional cloud-based video analysis incurs significant communication overhead.
• Real-time fall detection systems for seniors are needed to address this critical healthcare issue.
• Existing methods often struggle with efficiency and on-device implementation.

Purpose of the Study:

• To propose a fast, real-time fall detection system for the elderly optimized for edge devices.
• To reduce communication overhead associated with cloud-based video analysis.
• To develop a highly accurate and efficient fall detection solution for clinical monitoring.

Main Methods:

• A novel spatio-temporal optical flow model is developed for estimating motion object features.
• Features from objects and their optical flow fields are extracted and fused.
• A tensor-compressed model processes fused features for fall detection on edge devices.
• An object extractor identifies motion objects within video clips.

Main Results:

• The proposed system achieves high accuracy, reaching 96.23% and 99.37% on the Multicam and URFD datasets, respectively.
• Exceptional inference speed of 83.3 FPS and a 210.9× storage reduction were attained.
• Implementation on an AI acceleration core-based edge device reduced runtime by 9.21×.

Conclusions:

• The developed edge-based fall detection system offers a highly efficient and accurate solution for elderly monitoring.
• This technology significantly reduces communication overhead and enables real-time processing on resource-constrained devices.
• The system holds promise for future applications in clinical monitoring and elderly care.