Self Adaptive background modeling for identifying persons´ falls

This paper presents a new scheme for detecting humans´ falls in highly dynamic house environments. The scheme distinguishes falls from other humans´ activities, like sitting, walking, lying, under (a) sudden and abrupt illumination changes (b) non-periodic/significant motions in the background (chairs, curtains, tables), (c) humans´ movements towards all possible directions across camera. In particular, we combine adaptive background models - able to capture slight modifications of the background patterns with motion-based algorithms that define with high confidence parts of an image that should be considered as foreground/background after a significant visual change. We adopt Gaussian Mixtures for the adaptive background modeling, while we propose hierarchical motion estimation algorithms implemented on selective descriptors. The algorithms are of real time and require single low cost cameras.