Defall: A LSTM based Early Fall Detection Framework for Health Care

Falls among the elderly are a great concern for public health, frequently leading to injuries, restricted mobility, and diminished independence. Traditional fall detection systems are very accurate, dependent on wearable sensors, and suffer from high false-positive rates under dynamic conditions. Most of the existing solutions, based on vision and machine learning models, do not model the pre-fall behavior and are affected by environmental noise. To address these shortcomings, this paper puts forward an advanced fall detection system utilizing surveillance cameras. Falls are detected through pose estimation features derived from joint angles employing LSTM models for temporal analysis to improve response speed. Diffusion models are included to increase robustness against challenging situations like low light and occlusion. Testing this solution against a very broad dataset provides promising results on accuracy, precision, and recall, thus validating its applicability for real environments. The system also manages to reduce the false-positive rate, providing a reliable tool for elderly care environments.