Unsupervised Fall Detection Approach Using Human Skeletons

Fall detection in the elderly persons is crucial to prevent serious medical hazards. Human intervention can overcome such accidents but require extreme vigilance on part of the medical staff. The proposed approach aims to detect human fall behaviour automatically through visual data and alerts the respective caretakers and authorities. Immediate necessary actions can then be taken to rectify the situation in case of injuries and rescue operation. We detect human fall in an unsupervised way by using normal visual data only. Based on human skeletons, a recurrent encoder decoder model learns the normal human behaviour. We analyse the results in light of varying thresholds and input data. We have evaluated the proposed approach on Le2i fall detection dataset. The results on Le2i fall detection dataset indicate an accuracy of 77.52% which is comparable with the state-of-the art approaches.