Real-Time Acoustic Scene Recognition for Elderly Daily Routines Using Edge-Based Deep Learning.

Area of Science:

• Gerontology
• Computer Science
• Artificial Intelligence

Background:

• Growing global demand for intelligent monitoring systems in elderly living environments due to population aging.
• Limitations of traditional cloud-based acoustic systems, including data transmission delays and privacy concerns.

Purpose of the Study:

• To propose and evaluate an edge computing-integrated deep learning system for real-time acoustic scene recognition in elderly care.
• To enable continuous monitoring of elderly individuals' daily activities while mitigating privacy risks and transmission delays.

Main Methods:

• Development of a low-power edge device system with multiple microphones and wearable components.
• Implementation and optimization of four deep learning models (CNN, LSTM, BiLSTM, DNN) using model quantization techniques for edge constraints.
• Comparative performance analysis of the developed deep learning models.

Main Results:

• The Convolutional Neural Network (CNN) model achieved the highest accuracy at 98.5%.
• The optimized CNN model exhibited a fast inference time of 2.4 ms and low memory footprint (25.63 KB Flash, 5.15 KB RAM).
• Model quantization effectively reduced computational complexity and memory usage for edge deployment.

Conclusions:

• The proposed edge computing acoustic scene recognition system offers an efficient, reliable, and user-friendly solution for real-time monitoring in elderly care.
• Deep learning models, particularly CNNs optimized for edge devices, are suitable for intelligent acoustic scene analysis in resource-constrained environments.
• This technology addresses key challenges in elderly monitoring, enhancing safety and independence.