AI-DRIVEN FALL DETECTION SYSTEM USING SKELETAL DATA

Abstract

empowering them to perform tasks such as object detection, facial recognition, and scene understanding. Leveraging these capabilities, this paper proposes a novel fall detection system that enhances accuracy and reliability. This study introduces an automatic fall detection system utilizing skeletal data extracted from video sequences by the Mediapipe platform. Designed for the elderly and individuals with limited mobility, it aims to prevent or mitigate serious injuries through timely detection. Using MediaPipe Pose, 33 body landmarks are identified per frame to generate spatial features for a deep learning model. The system is built and evaluated on four different neural network architectures, including RNN, LSTM, GRU, and BiLSTM. Among them, the BiLSTM model achieved the highest accuracy of 97.32%. The proposed system does not require wearable devices and can be flexibly deployed in many environments, such as homes, hospitals, or rehabilitation centers, contributing to improving the efficiency and safety of caring for the elderly and sick.