Reduction of Vision-Based Models for Fall Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Falls pose significant risks, necessitating early detection systems.
• Current fall detection technologies often require high computational resources, limiting real-time application on edge devices.
• Complex deep learning models struggle with resource-constrained environments, hindering immediate response capabilities.

Purpose of the Study:

• To develop computationally efficient models for fall detection using image data.
• To reduce the parameter size of deep learning models for fall detection.
• To enable real-time fall detection on devices with limited computational power.

Main Methods:

• Utilized image sequences (video frames) from open-source datasets for fall detection.
• Applied the Sparse Low Rank Method to reduce layers in Convolutional Neural Networks (CNNs).
• Incorporated Long Short-Term Memory (LSTM) layers to process temporal data sequences.

Main Results:

• Significantly reduced the parameter size of the fall detection models.
• Maintained acceptable performance levels despite model compression.
• Demonstrated the feasibility of efficient fall detection on resource-limited platforms.

Conclusions:

• Model compression techniques, like Sparse Low Rank Method and LSTM integration, are effective for fall detection.
• Reduced model complexity enables immediate, on-device fall detection, crucial for timely intervention.
• This approach enhances the practicality of fall detection systems in real-world scenarios.