Real-time based human-fall detection from an indoor video surveillance

Abstract

In this paper, we present a human fall detection method from visual surveillance. In the first step, background subtraction is performed using improved GMM to find the foreground objects. In the second step, contour-based human template matching is applied to categorise the human or non-human object. It helps to detect fall incident by providing sudden change in generated score after matching. Height-width ratio and area of contour of an object is computed in the third step to decide whether the human shape is changed or not. In the fourth step, distance between top and mid centre of rectangle covering human is computed, if it is less than a certain threshold, then human fall is confirmed. Finally, if inactive pose of human is continued till 100 consecutive frames, then an alarm is generated to alert the people at home to provide treatment on time. Experiments have been performed on two different datasets having different usual and unusual fall incidents. Experimental results show that proposed system works well efficiently and effectively in real-time for recognising human fall.