Improving Human Activity Recognition in Smart Homes
International Journal of E-Health and Medical Communications, 6(3), 19-37, July-September 2015 19
Improving Human Activity Recognition in Smart Homes M’Hamed Bilal Abidine, University of Sciences and Technology Houari Boumediene, Bab Ezzouar, Algeria Lamya Fergani, University of Sciences and Technology Houari Boumediene, Bab Ezzouar, Algeria Belkacem Fergani, University of Sciences and Technology Houari Boumediene, Bab Ezzouar, Algeria Anthony Fleury, Mines Douai, URIA, Douai, France and University of Lille, Lille, France
ABSTRACT Even if it is now simple and cheap to collect sensors information in a smart home environment, the main issue remains to infer high-level activities from these simple readings. The main contribution of this work is twofold. Firstly, the authors demonstrate the efficiency of a new procedure for learning Optimized CostSensitive Support Vector Machines (OCS-SVM) classifier based on the user inputs to appropriately tackle the problem of class imbalanced data. It uses a new criterion for the selection of the cost parameter attached to the training errors. Secondly, this method is assessed and compared with the Conditional Random Fields (CRF), Linear Discriminant Analysis (LDA), k-Nearest Neighbours (k-NN) and the traditional SVM. Several and various experimental results obtained on multiple real world human activity datasets using binary and ubiquitous sensors show that OCS-SVM outperforms the previous state-of-the-art classification approach. Keywords:
Activity Recognition, Class Imbalance Data, Cost Sensitive Learning, Support Vector Machines (SVM)
INTRODUCTION Recognition of Human Activities is one of the most important tasks in pervasive computing applications (Tentori, 2008; Stiefmeier, 2008; Tapia, 2004; Van Kasteren, 2008). It is a thriving research field. It aims to identify high-level activities such as cooking, brushing, dressing, bathing and so on, performed in a ubiquitous environment, with real-time response. Due to an increasing number of elderly people and especially the ones living alone, wide-spread deployment of sensors has become prominent in home environments. Smart homes (Van Kasteren, 2008; Fleury, 2010) are equipped with sensor networks used, after information fusion, to automatically recognize
DOI: 10.4018/IJEHMC.2015070102 Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
20 International Journal of E-Health and Medical Communications, 6(3), 19-37, July-September 2015
activities of the inhabitant and to detect medical emergencies or assist sick/elderly people at home, and thus enabling them to live longer on their own. Therefore, they enable intelligent control by residential occupants to various automations. In this approach, sensors can be attached to either a person (wearable) under observation or objects that constitute the environment. When a resident moves from one room to another or uses different objects that on which area sensor, a series of firings with the corresponding timestamps are generated, allowing to automatically detect which activity the resident is currently performing, its duration and what objects are used for this activity. In this work, we focus on unobtrusive long-term monitoring of daily activities of a single person at home, on a daily basis. To perform a correct assistance, a good understanding of the current activity performed by the dweller is important. Sensor data collected needs to be analysed using data mining and machine learning techniques (Van Kasteren, 2008; Fleury, 2010) to determine which activity is taking place. As for any pattern recognition task, the keys to successful Activity Recognition are: (i) appropriately design the feature representation of the sensor data; and (ii) design a suitable classifier to infer the activity. Supervised classification algorithms are trained with labelled samples to generate a model that will be used to classify activities performed by the user. The ubiquitous computing literature describes a wide variety of creatively applied classification approaches (Duda, 2001; Shawe-Taylor, 2004). The learning of such models is usually done in a supervised manner (human labelling) and requires a large annotated dataset recorded in different settings. State of the art methods used for such recognition can be divided into two main categories: the so-called generative models and the discriminative models (Duda, 2001; Shawe-Taylor, 2004). In this work, we are interested in the supervised discriminative models for its simplicity and good performances, associated with a fast prediction speed. Recently, the class imbalance problem has been recognized as a crucial problem in machine learning (Chawla, 2004; Chawla, 2010). A two-class dataset is said to be imbalanced (or skewed) when one of the classes (the minority class) is heavily under-represented in comparison to the other (the majority one) during the training process. The existing activity recognition algorithms suffer from imbalanced data as some activities occur more frequently than others. These differences in frequency may correspond to how often a particular activity is performed (e.g. sleeping is generally done once a day, while toileting is done several times a day), or to the number of time slices an activity takes up (e.g. a sleeping activity generally takes up considerably more time slices than a toileting activity). However, the learning system may have difficulties to learn the concept related to the minority class, and therefore, will not incorporate this class imbalance results in an evaluation, and that may lead to disastrous consequences for elderly person in real life systems. There have been several attempts to deal with the class imbalance problem (Chawla, 2002; Chen, 2005; Thai-Nghe, 2010). Traditionally, research on this topic has mainly focused on solutions both at the data processing level (called external methods) and the algorithmic level (called internal methods). At data level, these solutions can be: oversampling (in which new samples are created for the minority class), under sampling (where, the samples are eliminated for the majority class) or any combination of both. Chawla et al.(2002) proposed the Synthetic Minority Over-sampling Technique (SMOTE) to deal with the problem of overfitting caused by random over-sampling. At an algorithmic level, the solutions include adjusting the costs associated with misclassification so as to improve performance (Chen,2005;Veropoulos, 1999; Thai-Nghe, 2010;Abidine, 2013), adjusting the probability estimate at the tree leaf (when working with decision trees), adjusting the decision threshold, or replacing the discrimination algorithm (twoclass classification) by a recognition algorithm (one-class classification). Akbani et al. (2004) proposed a SMOTE algorithm with Different Costs (SDC). SDC conducts SMOTE oversampling on the minority class with different error costs. Wu et al. (2005) proposed the Kernel Boundary
Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
17 more pages are available in the full version of this document, which may be purchased using the "Add to Cart" button on the product's webpage: www.igi-global.com/article/improving-human-activityrecognition-in-smart-homes/133567?camid=4v1
This title is available in InfoSci-Journals, InfoSci-Journal Disciplines Medicine, Healthcare, and Life Science. Recommend this product to your librarian: www.igi-global.com/e-resources/libraryrecommendation/?id=2
Related Content Accelerating Biomedical Research through Semantic Web Services Artemis Chaleplioglou (2014). International Journal of Reliable and Quality EHealthcare (pp. 32-48).
www.igi-global.com/article/accelerating-biomedical-research-throughsemantic-web-services/111464?camid=4v1a Reforming Medical Curriculum While Ignoring Examinations: A Commentary on Sood and Ananthakrishnan’s “Reforming Medical Curriculum in India in Recent Years: Conflicts of Political, Regulator, Educationist and Professional Natures and Strategies for their Resolution” Sabyasachi Sircar (2012). International Journal of User-Driven Healthcare (pp. 7678).
www.igi-global.com/article/reforming-medical-curriculum-whileignoring/64333?camid=4v1a Creating and Using the Knowledge Archive in the Internet Medical Consultant for Decision Support at the Point of Care Draško Nakic and Suzana Loškovska (2012). International Journal of E-Health and Medical Communications (pp. 72-85).
www.igi-global.com/article/creating-using-knowledge-archiveinternet/70010?camid=4v1a
Electronic Medical Records and Public Perceptions: A Deliberative Process Tarik Abdel-Monem, Mitchel N. Herian and Nancy Shank (2013). International Journal of Healthcare Information Systems and Informatics (pp. 38-57).
www.igi-global.com/article/electronic-medical-records-and-publicperceptions/93042?camid=4v1a
Improving Human Activity Recognition in Smart Homes M’Hamed Bilal Abidine, University of Sciences and Technology Houari Boumediene, Bab Ezzouar, Algeria Lamya Fergani, University of Sciences and Technology Houari Boumediene, Bab Ezzouar, Algeria Belkacem Fergani, University of Sciences and Technology Houari Boumediene, Bab Ezzouar, Algeria Anthony Fleury, Mines Douai, URIA, Douai, France and University of Lille, Lille, France
ABSTRACT Even if it is now simple and cheap to collect sensors information in a smart home environment, the main issue remains to infer high-level activities from these simple readings. The main contribution of this work is twofold. Firstly, the authors demonstrate the efficiency of a new procedure for learning Optimized CostSensitive Support Vector Machines (OCS-SVM) classifier based on the user inputs to appropriately tackle the problem of class imbalanced data. It uses a new criterion for the selection of the cost parameter attached to the training errors. Secondly, this method is assessed and compared with the Conditional Random Fields (CRF), Linear Discriminant Analysis (LDA), k-Nearest Neighbours (k-NN) and the traditional SVM. Several and various experimental results obtained on multiple real world human activity datasets using binary and ubiquitous sensors show that OCS-SVM outperforms the previous state-of-the-art classification approach. Keywords:
Activity Recognition, Class Imbalance Data, Cost Sensitive Learning, Support Vector Machines (SVM)
INTRODUCTION Recognition of Human Activities is one of the most important tasks in pervasive computing applications (Tentori, 2008; Stiefmeier, 2008; Tapia, 2004; Van Kasteren, 2008). It is a thriving research field. It aims to identify high-level activities such as cooking, brushing, dressing, bathing and so on, performed in a ubiquitous environment, with real-time response. Due to an increasing number of elderly people and especially the ones living alone, wide-spread deployment of sensors has become prominent in home environments. Smart homes (Van Kasteren, 2008; Fleury, 2010) are equipped with sensor networks used, after information fusion, to automatically recognize
DOI: 10.4018/IJEHMC.2015070102 Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
20 International Journal of E-Health and Medical Communications, 6(3), 19-37, July-September 2015
activities of the inhabitant and to detect medical emergencies or assist sick/elderly people at home, and thus enabling them to live longer on their own. Therefore, they enable intelligent control by residential occupants to various automations. In this approach, sensors can be attached to either a person (wearable) under observation or objects that constitute the environment. When a resident moves from one room to another or uses different objects that on which area sensor, a series of firings with the corresponding timestamps are generated, allowing to automatically detect which activity the resident is currently performing, its duration and what objects are used for this activity. In this work, we focus on unobtrusive long-term monitoring of daily activities of a single person at home, on a daily basis. To perform a correct assistance, a good understanding of the current activity performed by the dweller is important. Sensor data collected needs to be analysed using data mining and machine learning techniques (Van Kasteren, 2008; Fleury, 2010) to determine which activity is taking place. As for any pattern recognition task, the keys to successful Activity Recognition are: (i) appropriately design the feature representation of the sensor data; and (ii) design a suitable classifier to infer the activity. Supervised classification algorithms are trained with labelled samples to generate a model that will be used to classify activities performed by the user. The ubiquitous computing literature describes a wide variety of creatively applied classification approaches (Duda, 2001; Shawe-Taylor, 2004). The learning of such models is usually done in a supervised manner (human labelling) and requires a large annotated dataset recorded in different settings. State of the art methods used for such recognition can be divided into two main categories: the so-called generative models and the discriminative models (Duda, 2001; Shawe-Taylor, 2004). In this work, we are interested in the supervised discriminative models for its simplicity and good performances, associated with a fast prediction speed. Recently, the class imbalance problem has been recognized as a crucial problem in machine learning (Chawla, 2004; Chawla, 2010). A two-class dataset is said to be imbalanced (or skewed) when one of the classes (the minority class) is heavily under-represented in comparison to the other (the majority one) during the training process. The existing activity recognition algorithms suffer from imbalanced data as some activities occur more frequently than others. These differences in frequency may correspond to how often a particular activity is performed (e.g. sleeping is generally done once a day, while toileting is done several times a day), or to the number of time slices an activity takes up (e.g. a sleeping activity generally takes up considerably more time slices than a toileting activity). However, the learning system may have difficulties to learn the concept related to the minority class, and therefore, will not incorporate this class imbalance results in an evaluation, and that may lead to disastrous consequences for elderly person in real life systems. There have been several attempts to deal with the class imbalance problem (Chawla, 2002; Chen, 2005; Thai-Nghe, 2010). Traditionally, research on this topic has mainly focused on solutions both at the data processing level (called external methods) and the algorithmic level (called internal methods). At data level, these solutions can be: oversampling (in which new samples are created for the minority class), under sampling (where, the samples are eliminated for the majority class) or any combination of both. Chawla et al.(2002) proposed the Synthetic Minority Over-sampling Technique (SMOTE) to deal with the problem of overfitting caused by random over-sampling. At an algorithmic level, the solutions include adjusting the costs associated with misclassification so as to improve performance (Chen,2005;Veropoulos, 1999; Thai-Nghe, 2010;Abidine, 2013), adjusting the probability estimate at the tree leaf (when working with decision trees), adjusting the decision threshold, or replacing the discrimination algorithm (twoclass classification) by a recognition algorithm (one-class classification). Akbani et al. (2004) proposed a SMOTE algorithm with Different Costs (SDC). SDC conducts SMOTE oversampling on the minority class with different error costs. Wu et al. (2005) proposed the Kernel Boundary
Copyright © 2015, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
17 more pages are available in the full version of this document, which may be purchased using the "Add to Cart" button on the product's webpage: www.igi-global.com/article/improving-human-activityrecognition-in-smart-homes/133567?camid=4v1
This title is available in InfoSci-Journals, InfoSci-Journal Disciplines Medicine, Healthcare, and Life Science. Recommend this product to your librarian: www.igi-global.com/e-resources/libraryrecommendation/?id=2
Related Content Accelerating Biomedical Research through Semantic Web Services Artemis Chaleplioglou (2014). International Journal of Reliable and Quality EHealthcare (pp. 32-48).
www.igi-global.com/article/accelerating-biomedical-research-throughsemantic-web-services/111464?camid=4v1a Reforming Medical Curriculum While Ignoring Examinations: A Commentary on Sood and Ananthakrishnan’s “Reforming Medical Curriculum in India in Recent Years: Conflicts of Political, Regulator, Educationist and Professional Natures and Strategies for their Resolution” Sabyasachi Sircar (2012). International Journal of User-Driven Healthcare (pp. 7678).
www.igi-global.com/article/reforming-medical-curriculum-whileignoring/64333?camid=4v1a Creating and Using the Knowledge Archive in the Internet Medical Consultant for Decision Support at the Point of Care Draško Nakic and Suzana Loškovska (2012). International Journal of E-Health and Medical Communications (pp. 72-85).
www.igi-global.com/article/creating-using-knowledge-archiveinternet/70010?camid=4v1a
Electronic Medical Records and Public Perceptions: A Deliberative Process Tarik Abdel-Monem, Mitchel N. Herian and Nancy Shank (2013). International Journal of Healthcare Information Systems and Informatics (pp. 38-57).
www.igi-global.com/article/electronic-medical-records-and-publicperceptions/93042?camid=4v1a
