Human Fall Detection Based on ResNet and LSTM Network in Surveillance Cameras

One of the biggest problems with an elderly living alone at home is that of falls. Without prompt assistance and medical treatment, failure to detect falls in a timely manner could lead to more serious injury or death. While there are existing solutions that requires the elderly to wear sensors such as accelerometers or gyroscopes on their bodies, this is inconvenient to them and many will resist putting on such sensors. There is therefore a need to look for a less intrusive solution such as using computer vision and deep learning. In this paper, we proposed a novel framework which combines the advantages of Residual neural network (ResNet) and Long-Short term memories network (LSTM) to perform fall detection. The feature maps can be extracted from each frames by the trained ResNet network which is used to identify important features. Those feature maps are then fed in to an LSTM model to detect the existence of a fall. Our approach achieved a 100% detection accuracy on the URFD dataset and a 97.3% detection accuracy on the FDD dataset, outperforming other state-of-the-art algorithms.