Visual-based fall detection using histogram of oriented gradients of Poisson distance image

A novel visual-based fall detection method used as home-care for the elderly is presented in this note. Firstly, the binary silhouettes of moving human objects in a video sequence captured by a wall-mounted camera are detected, from which the so-called spatial-temporal motion accumulative image is created as the aggregation of human silhouettes evolution both in the image space and the time dimension. Then, Poisson distance image is acquired by solving the two-dimensional Poisson equations defined on the spatial-temporal accumulative image. What's more, the histogram of oriented gradients of Poisson distance image is collected as the newly-defined action descriptor. Finally, the action descriptors are fed into the nearest neighbour classifier to recognize daily actions, especially to detect the abnormal fall action. Experimental results on a home-brewed database exhibit that the proposed method can not only distinguish the fall action from others without exception, but also recognize all the daily actions with a high accuracy.