A multiple-stage classification of fall motions using kinect camera

Abstract

This paper proposes a model of fall detection using hybrid classification methods in video streaming. In particular, we are interested in a stream of data representing time sequential frames of fifteen body joint positions capturable by Kinect camera. A set of features is then extracted and fed into the designated multiple-stage classification. The first stage classifies a fall as a different event from normal activities of daily living (ADLs). The second stage is to classify types of fall once the fall was detected in the first stage, for aiding the diagnosis and treatment of a fall by a physician. We selected a number of reliable machine learning algorithms (MLP, SVM, and decision tree) in forming a hybrid model. Experimental results show that the first stage classifier can differentiate falls and ADLs with 99.98% accuracy and the second stage classifier can identify type of fall with 99.35% accuracy.