Adaptive Deep Learning for a Vision-based Fall Detection

Abstract

Fall is one of the main causes of severe accidents or even death especially for the elderly. Thus, it is imminent to prevent falls before they occur. In this paper, a vision-based system is adopted for fall detection exploiting novel self-adaptable deep machine learning strategies. The deep network is exploited to distinguished humans (foreground) from the background. Adaptation is necessary to tackle dynamic changes in the visual conditions (shadows, illumination, background changes) which are very often for a real-life environment. For the adaptable we are based on a decision mechanism that enable network retraining whenever the visual conditions are not proper for foreground/background separation. Then, a constraint minimization algorithm is activated to optimally estimate new network weights so that i) data from the current visual environment are trusted as much as possible while ii) a minimal degradation of the already existing network knowledge is accomplished. For the activation of the algorithm a set of new labeled data from the current environment is selected by constraining iterative motion information with a human face/body modeler. Experimental results and comparisons with non-adaptable deep network schemes or shallow non-linear classifier indicate the superior performance of the algorithm than other approaches.