A fall detection method in elevators based on Graph Convolutional Networks

Falling is one of the serious health risks and is an increasing concern in society. Since the elevator is a closed and unnoticed space, falling in the elevator is even more dangerous. In this paper, we propose a fall detection method in elevators based on graph convolution networks. Firstly, the human body bounding box is extracted by a target detection network, and it is fed into the pose estimation network to extract the human skeleton key points, then we stack key points of consecutive frames along the time dimension to form a skeleton spatio-temporal graph. Finally, we propose a spatiotemporal fall detection method based graph convolution network (Fall-GCN) to detect falling, and it consists of three layers of spatio-temporal graph convolution modules, which can extract spatial and temporal information of skeletal data. Furthermore, we design a partitioning strategy for the falling behavior to improve the spatial modeling ability of graph convolution, considering the special kinematic characteristics of the falling behavior. We performed experiments on the public Fall Detection Dataset i.e., Le2i and our self-designed Elevator Fall Detection Dataset i.e., Fall_Elevator. The results show that our method can effectively detect falling and has high recognition accuracy.

I. Introduction

As the main means of transportation in high-rise buildings, elevators are widely used around the world. However, while elevators bring convenience to people, they also bring safety risks. As shown in Fig. 1, since the elevator is a relatively closed place, it is difficult for people who fall in the elevator to be discovered in time, which will cause great harm to people's health. Nowadays, more and more cameras are installed in elevators, which can help detect abnormal situations. However, traditional monitoring methods require a lot of manpower. With the development of deep learning, intelligent monitoring solutions have attracted everyone's attention. It can intelligently analyze video clips, help us detect abnormal situations in the elevator in time, and minimize the impact of falls on people's lives and health.