Real-time Fall Detection using Key Joint Position Transition Time in the Elderly Activity Monitoring System

The aging population is rapidly increasing worldwide, leading to a growing concern for the well-being of elderly individuals. Falls among the elderly pose a significant threat to their health and independence, often resulting in serious injuries and reduced quality of life. To address this issue, accurate fall estimation methods are essential for timely intervention and effective preventive measures. By combining wearable sensors and machine learning algorithms can develop robust and reliable fall detection and prediction systems. However, it’s important to acknowledge that wearable sensors can come with certain drawbacks, primarily related to their cost, installation, and convenience. Compared to some high-end wearable sensor systems, setting up video cameras can be more affordable. Therefore, we propose a novel system for fall detection utilizing pose estimation framework (PoseNet) on the real-time camera. PoseNet detects the pose and estimates its key-joint position, the falling position or not using are classified using the time difference between the start and end of the pose change. K-Nearest Neighbor (KNN) is used as classifier. From the test, the resulting accuracy, precision, recall, and F-1 scores were 95%, 90%, 100%, and 93.23%, respectively. By providing a comprehensive monitoring solution, the proposed system enhances the safety and well-being of elderly individuals. Its ability to not only detect falls but also monitor daily activities and respond to emergencies immediately makes it a valuable tool for improving the quality of care and support for the elderly population.