RGB-D Fall Detection via Deep Residual Convolutional LSTM Networks

Abstract

The development of smart healthcare environments has witnessed impressive advancements exploiting the recent technological capabilities. Since falls are considered a major health concern especially among older adults, low-cost fall detection systems have become an indispensable component in these environments. This paper proposes an integrable, privacy preserving and efficient fall detection system from depth images acquired using a Kinect RGB-D sensor. The proposed system uses an end-to-end deep learning architecture composed of convolutional and recurrent neural networks to detect fall events. The deep convolutional network (ConvNet) analyses the human body and extracts visual features from input sequence frames. Fall events are detected via modeling complex temporal dependencies between subsequent frame features using Long-Shot-Term-Memory (LSTM) recurrent neural networks. Both models are combined and jointly trained in an end-to-end ConvLSTM architecture. This allows the model to learn visual representations and complex temporal dynamics of fall motions simultaneously. The proposed method has been validated on the public URFD fall detection dataset and compared with different approaches, including accelerometer based methods. We achieved a near unity sensitivity and specificity rates in detecting fall events.