Person Detection and Head Tracking to Detect Falls in Depth Maps

We present a system for fall detection in which the fall hypothesis, generated on the basis of accelerometric data, is validated by k-NN based classifier operating on depth features. We show that validation of the alarms in such a way leads to lower ratio of false alarms. We demonstrate the detection performance of the system using publicly available data. We discuss algorithms for person detection in images acquired by both a static and an active depth sensor. The head is modeled in 3D by an ellipsoid that is matched to point clouds, and which is also projected into 2D, where it is matched to edges in the depth maps.