Detection of fall for the elderly in an indoor environment using a tri-axial accelerometer and Kinect depth data

Monitoring elderly people who are living alone is a crucial task as they are at great risk of fall occurrence. In this paper, we present a robust framework for fall detection that makes use of two different signals namely tri-axial data from an accelerometer and depth maps from a Kinect sensor. Our approach functions at two stages. At the first stage, the accelerometer data is continuously being monitored and is used to indicate fall whenever the sum vector magnitude of the tri-axial data crosses a specific threshold. This fall indication denotes a high probability of fall occurrence. To confirm this and to avoid false alarms, the depth maps of a predefined window length captured prior to the instant of fall indication are obtained and processed. We propose a new descriptor, Entropy of Depth Difference Gradient Map that acts as a discriminative descriptor in differentiating fall from other daily activities. Finally, fall confirmation is done by employing a sparse representation-based classifier using the extracted descriptors. To ascertain the proposed model, we have performed experimental analysis using a publicly available UR Fall Detection dataset and also using a Synthetic dataset. The experimental results clearly depict the superior performance of our model.