Associative Classification based Human Activity Recognition and Fall Detection using Accelerometer

Associative Classification based Human Activity Recognition and Fall Detection using Accelerometer

C. Sweetlin Hemalatha (Department of Information Technology, Madras Institute of Technology, Anna University, Chennai, Tamil Nadu, India) and V. Vaidehi (Department of Information Technology, Madras Institute of Technology, Anna University, Chennai, Tamil Nadu, India)
Copyright: © 2013 |Pages: 18
DOI: 10.4018/jiit.2013070102
OnDemand PDF Download:
$37.50

Abstract

Human fall poses serious health risks especially among aged people. The rate of growth of elderly population to the total population is increasing every year. Besides causing injuries, fall may even lead to death if not attended immediately. This demands continuous monitoring of human movements and classifying normal low-level activities from abnormal event like fall. Most of the existing fall detection methods employ traditional classifiers such as decision trees, Bayesian Networks, Support Vector Machine etc. These classifiers may miss to cover certain hidden and interesting patterns in the data and thus suffer high false positives rates. Hence, there is a need for a classifier that considers the association between patterns while classifying the input instance. This paper presents a pattern mining based classification algorithm called Frequent Bit Pattern based Associative Classification (FBPAC) that distinguishes low-level human activities from fall. The proposed system utilizes single tri-axial accelerometer for capturing motion data. Empirical studies are conducted by collecting real data from tri-axial accelerometer. Experimental results show that within a time-sensitive sliding window of 10 seconds, the proposed algorithm achieves 99% accuracy for independent activity and 92% overall accuracy for activity sequence. The algorithm gives reasonable accuracy when tested in real time.
Article Preview

1. Introduction

Tele health care provides an effective solution for home based monitoring of health status of people compared to hospitalization. In this context, it is essential to continuously monitor human movements in order to detect abnormal event like fall especially in the case of home alone elderly people (Petry & Yager, 2012). Human fall not only causes physical injuries but also may lead to death if left unattended. This may develop psychological fear within oneself resulting in reduced confidence and independent living. In India, number of people in the age 60 and above is expected to increase to 100 million in 2013. Increasing elderly population every year to the total population and the health risks caused by falls among the age group of 60 and above, demands the need for a reliable and robust system for detecting fall.

Several techniques have been proposed for human activity recognition and fall detection. The great challenge lies in providing a health care solution with less invasive monitoring technologies and without hindering mobility of a person and at the same time not compromising the accuracy of interpreted health status. With rapid advancements in wireless sensor network technology (Tripathi et al., 2011), recognizing human activity for detecting falls has created a platform for research in pervasive and ubiquitous computing. Recently, many works on fall detection have been reported using wearable wireless sensors like accelerometer and gyroscope.

Also, mining sensor data streams pose great challenges in data mining as large amount of data are generated continuously with high speed in real time. Mining sensor data stream possesses different characteristics compared to traditional database model (Agrawal et al., 1993) such as (1) Each data element should be examined only once. (2) Though data gets generated continuously, memory usage for mining data streams is limited. (3) Each data element should be processed faster. (4) The outputs generated by online classifier algorithms should be instantly available when user requested. Su et al., (2011) have proposed a new classification algorithm for data steam based on lossy counting and landmark window model. Lin et al., (2005) presented a novel approach for mining frequent itemsets from data streams using time-sensitive sliding window model. An efficient window sliding technique for mining frequent itemsets over data streams was proposed by Li and Lee (2009).

Jinlong et al., (2004) presented a survey on frequent pattern mining in data streams in which the authors have analyzed the algorithms for frequent pattern mining in data streams from probabilistic bounds and deterministic bounds and proposed abstract reference architecture for Data Stream Management System (DSMS). Leung et al., (2006) proposed a novel tree data structure called Data Stream Tree (DSTree) for mining important regular patterns from data streams and discussed its effectiveness in terms of runtime through experiments. A new algorithm for mining frequent patterns in data streams, proposed by Feng et al., (2009) guaranteed no false negatives and thus improved mining quality.

Most of the existing fall detection methods are based on classifiers constructed using traditional methods such as decision trees, Bayesian Networks (Li, 2008) Neural Networks (Chen et al., 2010), Support Vector Machine (Kaiquan et al., 2011) etc. These classifiers may miss to cover certain hidden and interesting patterns in the data and thus suffer high false positives rates.

This paper aims to construct a classifier for recognizing low-level activities and detecting human fall based on mining frequent patterns in tri-axial accelerometer data streams. The proposed approach addresses the problem of mining accelerometer sensor streams using time-sensitive sliding window based pattern mining algorithm. A classifier model is built based on Associative Classification (AC) (Liu et al., 1998) that mines frequent bit patterns and extracts rules for recognizing human activities like sitting/standing, lying and walking with an ultimate aim to detect human fall events.

Complete Article List

Search this Journal:
Reset
Open Access Articles: Forthcoming
Volume 13: 4 Issues (2017)
Volume 12: 4 Issues (2016)
Volume 11: 4 Issues (2015)
Volume 10: 4 Issues (2014)
Volume 9: 4 Issues (2013)
Volume 8: 4 Issues (2012)
Volume 7: 4 Issues (2011)
Volume 6: 4 Issues (2010)
Volume 5: 4 Issues (2009)
Volume 4: 4 Issues (2008)
Volume 3: 4 Issues (2007)
Volume 2: 4 Issues (2006)
Volume 1: 4 Issues (2005)
View Complete Journal Contents Listing