FADES: Behavioral detection of falls using body shapes from 3D joint data

Abstract

Many efforts have been made to design classification systems that can aid the protection of elderly in a home environment. In this work, we focus on an accident that is a great risk for seniors living alone, a fall. Specifically, we present FADES, which uses skeletal joint information collected from a 3D depth camera to accurately classify different types of falls facing various directions from a single camera and distinguish an actual fall versus a fall-like activity, even in the presence of partially occluding objects. The framework of FADES is designed using two different phases to classify the detection of a fall, a non-fall, or normal behavior. For the first phase, we use a classification method based on Support Vector Machine (SVM) to detect body shapes that appear during an interval of falling behavior. During the second phase, we aggregate the results of the first phase using a frequency-based method to determine the similarity between the behavior sequences trained for each of the behavior. Our system shows promising results that is comparable to state-of-the-art techniques such as Viterbi algorithm, revealing real time performance with latency of <45 ms and achieving the detection accuracy of 96.07% and 95.7% for falls and non-falls, respectively.