Hybrid Architecture for Automatic Video-Based Fall Detection Using YOLOv11, MediaPipe Pose, and LSTM Networks

Falls represent one of the leading causes of injury and loss of autonomy among older adults worldwide. This work proposes a lightweight hybrid deep learning architecture for automatic fall detection, combining person detection with YOLOv11m, human pose estimation with MediaPipe, and temporal analysis using a long short-term memory network. Evaluated on the Le2i dataset, the model classified frames into normal activity, fall in progress, and person on the floor, achieving an overall accuracy of 99.23% and a weighted F1-score of 97.38%. The system matches or outperforms recent hybrid and transformer-based approaches while requiring lower computational resources, demonstrating its suitability for real-time embedded or home monitoring applications. Future work will focus on performance in uncontrolled environments and optimization for edge computing.