Fall Detection Method for Infrared Videos Based on Spatial-Temporal Graph Convolutional Network.

Area of Science:

• Computer Science
• Biomedical Engineering
• Artificial Intelligence

Background:

• Timely fall detection is crucial for elderly health monitoring, especially for those living alone.
• Traditional visual sensor-based fall detection methods suffer from poor adaptability, privacy concerns, and low recognition accuracy.
• Existing systems struggle with diverse environmental conditions and individual movement variations.

Purpose of the Study:

• To propose and evaluate an infrared video-based fall detection method using spatial-temporal graph convolutional networks (ST-GCNs).
• To address the limitations of conventional visual sensors in elderly fall detection.
• To enhance the accuracy and robustness of fall recognition in real-world scenarios.

Main Methods:

• Utilized infrared videos for fall detection, preserving privacy.
• Employed fine-tuned AlphaPose to extract 2D human skeleton sequences.
• Developed a two-stream ST-GCN incorporating improved adjacency matrices and multi-scale temporal units, processing skeleton data in Cartesian and polar coordinates.

Main Results:

• Achieved a highest accuracy of 96% in identifying fall behaviors on a proprietary dataset.
• Demonstrated robust performance in detecting falls from both near-infrared and thermal-infrared videos.
• Optimized ST-GCN parameters (time window, network depth) for practical deployment, balancing accuracy and speed.

Conclusions:

• The proposed infrared video-based ST-GCN method offers a highly accurate and privacy-preserving solution for elderly fall detection.
• The enhanced ST-GCN architecture effectively captures spatial-temporal features for reliable fall recognition.
• This approach provides a promising advancement for remote health monitoring systems.