Milto Miltiadou

3D motion capture of an Actor for Fall Detection and Health Monitoring

Thesis

Abstract

Many elderly people are hospitalised because they cannot autonomously live in their own homes due to the high risk of falls. Many researches attempted to solve the Fall Detection problem. The potential use of Kinect sensors may lead to a good markless Fall Detection System.

A scene was set and an actor was asked to performed actions, like sit on a chair and fall down. Using the iPi Studio and two Kinect sensors, .bvh motion capture files were generated. Those files are the input files of the system.

In that paper two classification methods are proposed: a simple Classification using Euclidean Distance and a Bayesian Inference with Gaussian Functions. Both methods use statistical analysis of 3D training data sets to recognise abnormal activities, when inactivity is observed.

Convolution, low level filtering and buffers that return the average actions are used to filter and improve the quality of the classification results.