In order to strengthen the monitoring of the
elderly and reduce the safety risks caused by falls, a
video-based indoor fall detection algorithm using a dual
network structure is proposed. Firstly, for the recorded
video stream, we apply the fine-tuned YOLACT
network to extract the contours of the human body, and
then design a simple convolutional neural network to
distinguish the categories of different family activities
(including bending, standing, sitting and lying), and
finally make a fall decision. When a lying position is
detected on the floor region, it is considered as a fall.
Experiments show that the proposed algorithm can
successfully detect fall events in different indoor
scenarios, and have a low false detection rate on the
constructed data set.
Keywords : health care; YOLACT; convolutional neural network; gesture recognition; fall detection