Elderly fall detection system based on multiple shape features and motion analysis

Abstract

This paper presents an intelligent video-based fall detection system. First, the silhouette of a person is extracted using a background subtraction technique, then a set of features is measured to define if a fall happened, for that a new technique is presented to estimate the head position, and a finite state machine (FSM) is used in the aim to compute the vertical velocity of the head. This algorithm is tested on the L2ei dataset where more than 2700 frames have been labelled in order to train three different classifiers. The results show that our system can predict the correct class with an accuracy that can reach up to 99.61% with a maximum global error of 1.5%.