Fall recognition using a three stream spatio temporal GCN model with adaptive feature aggregation

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Falls are a major health risk for the elderly, causing severe injuries and fatalities.
• Existing fall detection systems lack accuracy, robustness, and efficiency, and are sensitive to environmental factors.
• Accurate and efficient fall detection is crucial for preventing injuries and aiding elderly individuals.

Purpose of the Study:

• To propose a novel three-stream spatio-temporal feature-based human fall detection system.
• To address the limitations of existing systems, including accuracy, robustness, and computational complexity.
• To develop an efficient and generalizable fall detection solution for real-world applications.

Main Methods:

• Incorporation of joint skeleton-based and joint motion-based spatial and temporal Graph Convolutional Network (GCN) features.
• Utilizing residual connections and adaptive graph-based feature aggregation.
• Employing consecutive separable convolutional neural networks (Sep-TCN) to reduce computational complexity and model parameters.

Main Results:

• Achieved high accuracies: 99.68% (ImViA), 99.97% (Fall-UP), 99.47% (FU-Kinect), and 98.97% (UR-Fall).
• Demonstrated superior effectiveness and efficiency compared to existing fall detection systems.
• Showcased remarkable performance, highlighting the system's superiority, efficiency, and generalizability.

Conclusions:

• The proposed system offers significant advancements in healthcare and societal well-being through reliable fall detection.
• The novel approach effectively overcomes the limitations of previous fall detection methods.
• The system's high accuracy and efficiency make it a promising solution for real-world fall detection scenarios.