Multi-Person Fall Detection Using Data from MediaPipe with Deep Learning

Abstract

This research aims to develop a multi-person fall detection system using skeletal data extracted from MediaPipe, and to evaluate the performance of deep learning models, including CNN, LSTM, and a hybrid CNN-LSTM model. The proposed system detects multiple individuals within a single frame by using three-dimensional coordinates of 13 key body joints to classify five postures: standing, forward fall, backward fall, left-side fall, and right-side fall. Experiments were conducted to examine the effect of temporal sequence length on classification performance using two sequence lengths: 10 and 16 frames. The results show that the CNN-LSTM model achieves the best performance, particularly with a 16-frame sequence, reaching an accuracy of 98%. This is because the model effectively captures both spatial features and temporal dependencies of human movements, reducing misclassification among similar fall postures. These findings indicate that combining skeletal data with temporal-aware deep learning models is suitable for real-time fall monitoring and alert systems.