Vision-Based Monitoring System for Elderly Care Using Deep Learning

Abstract

Introduction:
As the global elderly population grows, the need for supportive technologies that enable independent living while ensuring safety becomes increasingly critical. Falls are a major health concern for individuals aged 65 and older, often leading to severe injuries or fatalities. Traditional monitoring systems like wearables have limitations due to reliance on user compliance. Vision-based systems offer a non-intrusive, effective alternative by utilizing real-time video surveillance powered by deep learning.

Objectives:
The primary objective is to develop a real-time monitoring system that accurately detects abnormal activities such as falls or prolonged inactivity. The system aims to ensure high sensitivity and precision, protect user privacy, and enable timely intervention through alert mechanisms like the Telegram API.

Methods:
The proposed system integrates Convolutional Neural Networks (CNN) for spatial feature extraction and Bidirectional Long Short-Term Memory (BiLSTM) networks for temporal analysis. Video frames are preprocessed, passed through a CNN for feature generation, and then analyzed sequentially using BiLSTM to detect anomalies. Upon detection of abnormal behavior, alerts are sent in real-time to caregivers. The system is trained on publicly available datasets such as LifeSeniorProfile and ALMOND.

Results:
The system achieved 95% accuracy in fall detection and 92% in identifying prolonged inactivity. Precision and recall stood at 93% and 90%, respectively, resulting in an F1 score of 91.5%. Real-time alerts were successfully delivered within 3 seconds via Telegram, with caregivers reporting high satisfaction with the system's performance and usability.

Conclusions:
The vision-based monitoring system effectively enhances elderly care by providing accurate, real-time fall and inactivity detection. It enables independent living while ensuring safety and privacy through local processing and data encryption. Future enhancements may include integration with wearable vitals sensors and improved model generalization through diverse datasets.