A Skeleton-Based Deep Learning Framework for Gait Analysis and Fall Detection Using MediaPipe Pose

Abstract

This study proposes a skeleton-based human activity recognition framework that integrates MediaPipe Pose–based feature extraction with deep learning for intelligent healthcare applications, including gait analysis and fall detection. Skeletal time-series data from 33 keypoints are extracted from video streams and transformed into angle-based temporal features to characterize human motion patterns. Experimental results demonstrate that the proposed method effectively captures gait characteristics and achieves stable performance in both binary and multi-class fall detection tasks. In addition, an Internet of Things (IoT) architecture based on Raspberry Pi and ESP32 is developed to enable real-time fall detection and remote monitoring. The proposed system exhibits high computational efficiency and strong integration capability, making it suitable for smart home care and real-time safety monitoring applications.