Fall Detection on a single Doppler Radar Sensor by using Convolutional Neural Networks

Falls are a major public health concern and the leading cause of accidental deaths among elders. Technologies capable of fast and accurate fall detection are crucial to provide immediate assistance and reduce complications associated with a fall. This article presents a new deep-learning approach for automated fall detection from signatures of a continuous-wave micro-Doppler radar. Unlike related methods, the proposed approach applies convolutional neural networks directly to raw time-frequency spectrograms to identify falls. It allows CNN to learn the fall features and classification boundaries without employing additional hardware or large data sets to distinguish abrupt and slow falls from non-fall actions. Experiments show that the approach achieves the fall classification accuracy of 95%.