Predicting Human Actions Using a Hybrid of ReliefF Feature Selection and Kernel-Based Extreme Learning Machine

Introduction

Wearable sensors are becoming popular in many areas such as medical, security, entertainment and commercial fields. These sensors are exceedingly useful in providing reliable and accurate information about people's actions and behaviors thus, a safe and healthy living environment is provided (Mukhopadhyay, 2015). Thus, it is possible to develop a more successful technology that can track the walking of older people who are in search of a measure against falling. In addition, a cost-effective system can also be established to investigate the relationship between a person's health status and the total number of daily living activities. The initial step in this direction is to develop a system that can classify a walking data set into different everyday life activities (Gupta & Dallas, 2014).

Monitoring and classifying human activities from wearable sensors can provide valuable information about patient’s mobility outside the hospital environment. Research in this field has attracted a great deal of attention in recent years. An activity monitoring approach that works with ubiquitous technologies and can be applied in clinical populations will provide great benefit to the decision making process based on the evidence for people with movement handicap (Capela, Lemaire & Baddour, 2015). Smartphones and smartwatches provide a handy, wearable, useful, easy-to-use computing environment rich with storage and computing power. Most of HAR systems have been developed for smartphones. Some of these systems use internal sensors and some use external biological sensors (Incel, Kose & Ersoy, 2013; Pantelopoulos & Bourbakis, 2010). When measuring movement or position, gyroscopes and accelerometers are prevalent selections because these sensors are affordable, small size, and can easily attach to the body. Several smartwatches and smartphones include gyroscopes and accelerometers thus; they are perfect devices for tracking in rehabilitation or real life environments. Wearable device sensors are used to measure the quality of post stroke movement such as upper extremity movement or walking characteristics (Patel, Huges, Hester, Stein, Akay, Dy & Bonata, 2010; Mizuike, Ohgi & Morita, 2009). Activity levels are usually measured by the number of passing of a certain threshold of total acceleration using accelerometer or other sensor data worn on the body (Fulk & Sazonov, 2011; Steele, Belza, Cain & Warms, 2003). On the other hand, activity level analysis lacks of content-based information. A system that provides content-based info about a person's actions will be a special interest for researchers and healthcare professionals (Capela, Lemaire & Baddour, 2015).

General guidelines for HAR are preprocessing, segmentation, feature extraction, feature selection and classification. Although a large number of attributes can be extracted from a raw signal, the accuracy of the classifier is not necessarily increased because the redundancy of the features is not a class indicator. For this reason, the feature selection process is used to decrease the data dimension and forward valuable attributes to the classifier (Capela, Lemaire & Baddour, 2015).