Fall detection with two cameras based on occupied areas

In the effort of supporting elderly people living alone, this paper describes a novel video-based system for detecting fall incidents. Widths of a same person are extracted from two cameras whose fields of view are relatively orthogonal, for estimating the occupied area. We divide the scene into many small patches. Sizes of a person moving through the scene are clustered and kept in the buffer of each patch in which the person is captured, so-called Local Empirical Templates (LET), for building spatial distributions of occupied areas of the person in walking or standing poses. We realize that occupied areas of lying-down and sitting person are proportional to that of LET, spotted in the same scene patch. Therefore, we normalize the height and the occupied area of a person estimated from the two cameras with respect to those of LET in the same scene patch, leading the generation of a promising feature space in which three human states of standing, sitting or bending, and lying down, are in separable regions. Fall incidents can be inferred from the time-series analysis of human state transition. The experimental results with 24 realistic video samples in Multiple cameras fall dataset (1) demonstrates high detection and low false alarm rates